Stanford's groundbreaking discovery: Sleep one night, and AI can calculate your remaining lifespan.
2026年1月，斯坦福医学院的研究团队在《自然医学》杂志上发表了一项重磅研究，他们开发了一种名为SleepFM的AI基础模型，仅通过一晚的多导睡眠图（PSG）记录，就能预测超过100种健康状况的风险，包括全因死亡率。这一发现让“睡一觉就能预知寿命”从科幻走向现实，引发全球医学界和公众的广泛关注。传统上，睡眠监测主要用于诊断睡眠呼吸暂停等即时问题，而SleepFM则将睡眠数据转化为长期健康预测的“水晶球”，揭示了睡眠作为人体“健康档案”的深层价值。
In January 2026, a research team from Stanford Medicine published a landmark study in Nature Medicine. They developed an AI foundation model called SleepFM that, using just one night's polysomnography (PSG) recordings, can predict the risk of over 100 health conditions, including all-cause mortality. This discovery transforms the idea of "predicting lifespan from one night's sleep" from science fiction into reality, sparking widespread attention in the global medical community and among the public. Traditionally, sleep monitoring was mainly used to diagnose immediate issues like sleep apnea, but SleepFM turns sleep data into a "crystal ball" for long-term health predictions, revealing sleep's profound value as a "health archive" for the human body.
SleepFM模型的训练数据规模惊人，涵盖了超过6.5万名患者的近60万小时睡眠记录，这些数据来自斯坦福睡眠医学中心，并与长达25年的电子健康记录相匹配。AI通过自监督学习“掌握了睡眠的语言”，不仅能准确进行睡眠分期和呼吸暂停检测，还能在无标签数据上实现高效迁移学习。研究显示，该模型对130多种疾病的预测C指数（一致性指数）达到0.75以上，其中全因死亡率的C指数为0.84，痴呆为0.85，帕金森病高达0.89，心肌梗死0.81，前列腺癌0.89，乳腺癌0.87。这些指标远超基于人口统计学或传统睡眠指标的基准模型，证明睡眠中隐藏着丰富的纵向健康信号。
SleepFM was trained on an enormous dataset covering nearly 600,000 hours of sleep recordings from over 65,000 patients at Stanford's sleep centers, matched with up to 25 years of electronic health records. The AI "learned the language of sleep" through self-supervised learning, enabling accurate sleep staging, apnea detection, and efficient transfer learning on unlabeled data. The study showed that the model achieved C-indexes of at least 0.75 for predicting over 130 conditions, with all-cause mortality at 0.84, dementia at 0.85, Parkinson's disease at 0.89, myocardial infarction at 0.81, prostate cancer at 0.89, and breast cancer at 0.87. These metrics significantly outperform baselines based on demographics or traditional sleep metrics, proving that sleep harbors rich longitudinal health signals.
这项研究的创新之处在于SleepFM是一个多模态基础模型，它同时处理脑电图（EEG）、心电图（ECG）、呼吸信号、眼动、腿动等多种生理数据，而非依赖单一指标。传统睡眠分析往往只关注呼吸暂停指数或睡眠效率，但SleepFM能捕捉到跨系统的不匹配信号，例如大脑处于睡眠状态而心脏却表现出“清醒”迹象，这种“脱节”往往是未来心血管或神经退行性疾病的早期预警。研究者詹姆斯·邹（James Zou）教授表示：“睡眠数据就像一种被低估的‘生物语言’，AI帮助我们破译了它，从而提前多年发现疾病风险。”
The innovation of this study lies in SleepFM being a multimodal foundation model that simultaneously processes various physiological data such as electroencephalogram (EEG), electrocardiogram (ECG), respiratory signals, eye movements, and leg movements, rather than relying on single metrics. Traditional sleep analysis often focuses only on apnea-hypopnea index or sleep efficiency, but SleepFM captures cross-system mismatch signals—for instance, the brain appearing asleep while the heart shows "awake" patterns. Such "disconnections" often serve as early warnings for future cardiovascular or neurodegenerative diseases. Professor James Zou noted: "Sleep data is like an underestimated 'biological language,' and AI helps us decipher it to detect disease risks years in advance."
Beyond prediction accuracy, SleepFM demonstrates strong generalization. It performed competitively with specialized models like U-Sleep and YASA on standard tasks (mean F1 scores of 0.70–0.78 for sleep staging) and showed robust transfer learning on external datasets such as the Sleep Heart Health Study. This suggests the model isn't overfitting to Stanford-specific data but truly learning universal patterns in human sleep physiology. For mortality prediction, it achieved an AUROC of 0.85, surpassing demographic baselines (0.78) and end-to-end PSG models, highlighting how subtle factors like arousal burden, REM sleep reduction, hypoxemia, and sleep fragmentation integrate into powerful prognostic signals.
睡眠与寿命的关联早已被科学研究证实。早在2022年就有斯坦福相关研究显示，通过深度学习从睡眠研究中估计的“睡眠年龄”误差每增加10年，全因死亡风险上升29%，心血管死亡风险上升40%。对于60岁人群，睡眠年龄高估或低估10年，可能对应寿命减少约8.7年。SleepFM在此基础上实现了质的飞跃，它不只是估计生物年龄，还直接关联具体疾病风险和剩余寿命概率。研究指出，睡眠碎片化、低睡眠效率和间歇性低氧血症等特征，是独立于已知风险因素的死亡预测因子。
The link between sleep and lifespan has long been confirmed by scientific research. As early as 2022, related Stanford studies showed that every 10-year increase in error from deep learning-estimated "sleep age" raised all-cause mortality risk by 29% and cardiovascular mortality by 40%. For a 60-year-old, a 10-year over- or under-estimation in sleep age could correspond to an approximately 8.7-year reduction in life expectancy. SleepFM builds on this with a qualitative leap: it not only estimates biological age but directly links to specific disease risks and probabilities of remaining lifespan. The study notes that features like sleep fragmentation, low sleep efficiency, and intermittent hypoxemia are independent predictors of mortality beyond known risk factors.
在中国，睡眠健康问题同样突出。根据国家卫健委数据，成年居民睡眠不足比例较高，城市白领群体中睡眠障碍发生率可达30%以上。随着老龄化加剧和慢性病负担增加，利用AI分析睡眠数据预测寿命和疾病风险，具有巨大的公共卫生意义。国内多家医院和科技企业已开始探索类似多导睡眠监测与AI结合的应用，例如将可穿戴设备数据接入大模型，实现日常睡眠健康管理。如果SleepFM的技术能本土化落地，或许能帮助数亿人提前干预，减少医疗资源浪费，推动“健康中国”战略。
In China, sleep health issues are equally prominent. According to data from the National Health Commission, a high proportion of adults experience insufficient sleep, with sleep disorder rates among urban white-collar workers reaching over 30%. With accelerating aging and increasing chronic disease burdens, using AI to analyze sleep data for lifespan and disease risk prediction holds enormous public health significance. Several domestic hospitals and tech companies have begun exploring applications combining polysomnography with AI, such as integrating wearable device data into large models for daily sleep health management. If SleepFM technology can be localized, it could help hundreds of millions of people intervene early, reduce medical resource waste, and advance the "Healthy China" strategy.
The model's ability to predict specific conditions is particularly impressive. For neurological disorders, it excels at myoneural disorders, developmental delays, and speech/language issues. In circulatory diseases, it identifies atherosclerosis and acute pulmonary heart disease with high precision. Endocrine and metabolic predictions include diabetes complications and diabetic retinopathy, while respiratory predictions cover insufficiency and failure. Pregnancy complications and mental health disorders also show strong signals, suggesting sleep as a window into systemic health across life stages. Importantly, no single signal drives predictions; instead, the AI integrates multifactorial patterns, making it more robust than rule-based clinical scoring.
这项发现的潜在应用场景广阔。在临床上，睡眠实验室的常规检查可直接输出个性化风险报告，帮助医生制定预防方案。例如，高死亡风险患者可被优先推荐生活方式干预、药物治疗或进一步筛查。在保险和企业健康管理领域，AI睡眠预测能支持精准定价和员工福利优化。当然，隐私保护和伦理审查是前提，数据必须匿名化处理，且预测结果仅作为参考而非诊断。未来，研究团队计划将模型扩展到可穿戴设备数据，实现“在家睡一觉就能知寿命”的便捷模式。
The potential applications of this discovery are vast. Clinically, routine sleep lab tests could directly output personalized risk reports to help doctors devise prevention plans. For instance, patients with high mortality risk could be prioritized for lifestyle interventions, medications, or further screenings. In insurance and corporate wellness management, AI sleep predictions could support precise pricing and employee benefit optimization. Of course, privacy protection and ethical review are prerequisites—data must be anonymized, and predictions serve only as references, not diagnoses. In the future, the team plans to extend the model to wearable device data, enabling the convenient scenario of "sleep at home one night and know your lifespan."
However, challenges remain. While SleepFM outperforms baselines, real-world deployment requires validation across diverse populations, including different ethnicities, ages, and comorbidities common in China and globally. Accuracy for certain conditions like chronic kidney disease or stroke (around 0.78-0.79 C-index) is good but not perfect, underscoring the need for multimodal integration with genetics, blood tests, and lifestyle data. Interpretability is another focus: researchers are using techniques like SHAP to explain which sleep patterns drive predictions, building clinician trust. Over-reliance on AI without human oversight could lead to anxiety or false reassurance, so balanced communication is essential.
睡眠质量的改善本身就能延长寿命。大量流行病学研究表明，每晚7-9小时高质量睡眠可降低心血管疾病、糖尿病和认知衰退风险。SleepFM的预测结果或许能成为强大动机，鼓励人们重视睡眠卫生：规律作息、避免咖啡因、保持卧室凉爽黑暗等简单措施。结合AI反馈的个性化建议，例如“你的REM睡眠不足，建议增加运动以提升深度睡眠”，将使预防医学从被动转向主动。在中国“银发经济”和数字健康浪潮下，这一技术或催生新型睡眠健康产业，包括智能床垫、AI睡眠教练App等。
Improving sleep quality itself can extend lifespan. Extensive epidemiological studies show that 7-9 hours of high-quality sleep nightly reduces risks of cardiovascular disease, diabetes, and cognitive decline. SleepFM's predictions may serve as a powerful motivator, encouraging people to prioritize sleep hygiene: regular schedules, avoiding caffeine, and keeping bedrooms cool and dark. Personalized AI feedback, such as "Your REM sleep is insufficient; suggest increasing exercise to boost deep sleep," shifts preventive medicine from passive to active. Amid China's "silver economy" and digital health wave, this technology could spawn new sleep health industries, including smart mattresses and AI sleep coach apps.
Looking ahead, SleepFM represents a broader trend: foundation models unlocking hidden value in routine medical data. Similar to how large language models revolutionized text, multimodal sleep models could transform polysomnography from a niche diagnostic tool into a scalable population health platform. In China, integration with "East Data West Computing" infrastructure and domestic large models could accelerate adoption, supporting national goals for chronic disease prevention and healthy aging. International collaboration will be key to refining the model across global datasets and addressing biases.
要充分发挥这一技术的潜力，需要多方协同。医疗机构应加大睡眠监测设备的普及和AI培训；科技企业可开发用户友好接口，让普通人通过家用设备获取初步洞见；监管部门则需制定数据安全和伦理指南，避免信息滥用。同时，公众教育不可或缺：睡眠不是“浪费时间”，而是投资未来的健康资本。研究还提醒我们，年轻时忽视睡眠，可能在中年或老年付出寿命代价。
To fully leverage this technology's potential requires multi-stakeholder collaboration. Medical institutions should expand普及 of sleep monitoring equipment and AI training; tech companies can develop user-friendly interfaces for preliminary insights via home devices; regulators must establish data security and ethical guidelines to prevent misuse. Public education is equally essential: sleep is not "wasted time" but an investment in future health capital. The research also reminds us that neglecting sleep in youth may exact a lifespan cost in middle or old age.
Ethical considerations loom large. Predicting mortality or serious diseases years ahead raises questions about psychological impact—how to deliver "bad news" from sleep data without causing undue stress? Should insurers access such predictions? How to ensure equity so that low-income or rural populations in China aren't left behind? Researchers emphasize that SleepFM is a risk stratification tool, not a fortune teller; lifestyle, genetics, and medical interventions can still alter trajectories. Transparent, compassionate communication frameworks will be crucial for responsible deployment.
这一斯坦福重磅发现，不仅是AI在医学领域的又一里程碑，更是对人类与睡眠关系的重新审视。睡上一觉，AI或许就能勾勒出剩余寿命的轮廓，但真正的智慧在于行动：珍惜每晚的休息，用科学守护健康。未来，随着技术迭代和数据积累，类似模型或能实现更精准的个性化寿命预测，甚至指导精准医学干预。让我们期待，AI与睡眠的结合，能帮助更多人活得更长、更健康。
This Stanford groundbreaking discovery is not only another milestone for AI in medicine but also a re-examination of humanity's relationship with sleep. Sleep one night, and AI might outline the contours of remaining lifespan, but true wisdom lies in action: cherish every night's rest and guard health with science. In the future, with technological iterations and data accumulation, similar models may achieve even more precise personalized lifespan predictions and guide precision medicine interventions. Let us look forward to AI combined with sleep helping more people live longer and healthier lives.
In summary股票配资大盘, SleepFM illustrates how one night's seemingly ordinary sleep encodes signals of decades-long health trajectories. By learning the "language of sleep," AI uncovers patterns invisible to the human eye—arousal indices, stage transitions, oxygen desaturations, and cross-signal dyssynchronies—that collectively forecast Parkinson's, cancers, heart failure, dementia, and mortality with remarkable accuracy. While not a definitive crystal ball, it offers an unprecedented proactive window into health. For individuals, it underscores the power of good sleep; for society, it points toward driven prevention that could ease healthcare burdens worldwide, including in China's vast population. Responsible innovation, rigorous validation, and ethical integration will determine whether this technology fulfills its promise of extending healthy lifespans.Stanford's groundbreaking discovery: Sleep one night, and AI can calculate your remaining lifespan.
2026年1月，斯坦福医学院的研究团队在《自然医学》杂志上发表了一项重磅研究，他们开发了一种名为SleepFM的AI基础模型，仅通过一晚的多导睡眠图（PSG）记录，就能预测超过100种健康状况的风险，包括全因死亡率。这一发现让“睡一觉就能预知寿命”从科幻走向现实，引发全球医学界和公众的广泛关注。传统上，睡眠监测主要用于诊断睡眠呼吸暂停等即时问题，而SleepFM则将睡眠数据转化为长期健康预测的“水晶球”，揭示了睡眠作为人体“健康档案”的深层价值。
In January 2026, a research team from Stanford Medicine published a landmark study in Nature Medicine. They developed an AI foundation model called SleepFM that, using just one night's polysomnography (PSG) recordings, can predict the risk of over 100 health conditions, including all-cause mortality. This discovery transforms the idea of "predicting lifespan from one night's sleep" from science fiction into reality, sparking widespread attention in the global medical community and among the public. Traditionally, sleep monitoring was mainly used to diagnose immediate issues like sleep apnea, but SleepFM turns sleep data into a "crystal ball" for long-term health predictions, revealing sleep's profound value as a "health archive" for the human body.
SleepFM模型的训练数据规模惊人，涵盖了超过6.5万名患者的近60万小时睡眠记录，这些数据来自斯坦福睡眠医学中心，并与长达25年的电子健康记录相匹配。AI通过自监督学习“掌握了睡眠的语言”，不仅能准确进行睡眠分期和呼吸暂停检测，还能在无标签数据上实现高效迁移学习。研究显示，该模型对130多种疾病的预测C指数（一致性指数）达到0.75以上，其中全因死亡率的C指数为0.84，痴呆为0.85，帕金森病高达0.89，心肌梗死0.81，前列腺癌0.89，乳腺癌0.87。这些指标远超基于人口统计学或传统睡眠指标的基准模型，证明睡眠中隐藏着丰富的纵向健康信号。
SleepFM was trained on an enormous dataset covering nearly 600,000 hours of sleep recordings from over 65,000 patients at Stanford's sleep centers, matched with up to 25 years of electronic health records. The AI "learned the language of sleep" through self-supervised learning, enabling accurate sleep staging, apnea detection, and efficient transfer learning on unlabeled data. The study showed that the model achieved C-indexes of at least 0.75 for predicting over 130 conditions, with all-cause mortality at 0.84, dementia at 0.85, Parkinson's disease at 0.89, myocardial infarction at 0.81, prostate cancer at 0.89, and breast cancer at 0.87. These metrics significantly outperform baselines based on demographics or traditional sleep metrics, proving that sleep harbors rich longitudinal health signals.
这项研究的创新之处在于SleepFM是一个多模态基础模型，它同时处理脑电图（EEG）、心电图（ECG）、呼吸信号、眼动、腿动等多种生理数据，而非依赖单一指标。传统睡眠分析往往只关注呼吸暂停指数或睡眠效率，但SleepFM能捕捉到跨系统的不匹配信号，例如大脑处于睡眠状态而心脏却表现出“清醒”迹象，这种“脱节”往往是未来心血管或神经退行性疾病的早期预警。研究者詹姆斯·邹（James Zou）教授表示：“睡眠数据就像一种被低估的‘生物语言’，AI帮助我们破译了它，从而提前多年发现疾病风险。”
The innovation of this study lies in SleepFM being a multimodal foundation model that simultaneously processes various physiological data such as electroencephalogram (EEG), electrocardiogram (ECG), respiratory signals, eye movements, and leg movements, rather than relying on single metrics. Traditional sleep analysis often focuses only on apnea-hypopnea index or sleep efficiency, but SleepFM captures cross-system mismatch signals—for instance, the brain appearing asleep while the heart shows "awake" patterns. Such "disconnections" often serve as early warnings for future cardiovascular or neurodegenerative diseases. Professor James Zou noted: "Sleep data is like an underestimated 'biological language,' and AI helps us decipher it to detect disease risks years in advance."
Beyond prediction accuracy, SleepFM demonstrates strong generalization. It performed competitively with specialized models like U-Sleep and YASA on standard tasks (mean F1 scores of 0.70–0.78 for sleep staging) and showed robust transfer learning on external datasets such as the Sleep Heart Health Study. This suggests the model isn't overfitting to Stanford-specific data but truly learning universal patterns in human sleep physiology. For mortality prediction, it achieved an AUROC of 0.85, surpassing demographic baselines (0.78) and end-to-end PSG models, highlighting how subtle factors like arousal burden, REM sleep reduction, hypoxemia, and sleep fragmentation integrate into powerful prognostic signals.
睡眠与寿命的关联早已被科学研究证实。早在2022年就有斯坦福相关研究显示，通过深度学习从睡眠研究中估计的“睡眠年龄”误差每增加10年，全因死亡风险上升29%，心血管死亡风险上升40%。对于60岁人群，睡眠年龄高估或低估10年，可能对应寿命减少约8.7年。SleepFM在此基础上实现了质的飞跃，它不只是估计生物年龄，还直接关联具体疾病风险和剩余寿命概率。研究指出，睡眠碎片化、低睡眠效率和间歇性低氧血症等特征，是独立于已知风险因素的死亡预测因子。
The link between sleep and lifespan has long been confirmed by scientific research. As early as 2022, related Stanford studies showed that every 10-year increase in error from deep learning-estimated "sleep age" raised all-cause mortality risk by 29% and cardiovascular mortality by 40%. For a 60-year-old, a 10-year over- or under-estimation in sleep age could correspond to an approximately 8.7-year reduction in life expectancy. SleepFM builds on this with a qualitative leap: it not only estimates biological age but directly links to specific disease risks and probabilities of remaining lifespan. The study notes that features like sleep fragmentation, low sleep efficiency, and intermittent hypoxemia are independent predictors of mortality beyond known risk factors.
在中国，睡眠健康问题同样突出。根据国家卫健委数据，成年居民睡眠不足比例较高，城市白领群体中睡眠障碍发生率可达30%以上。随着老龄化加剧和慢性病负担增加，利用AI分析睡眠数据预测寿命和疾病风险，具有巨大的公共卫生意义。国内多家医院和科技企业已开始探索类似多导睡眠监测与AI结合的应用，例如将可穿戴设备数据接入大模型，实现日常睡眠健康管理。如果SleepFM的技术能本土化落地，或许能帮助数亿人提前干预，减少医疗资源浪费，推动“健康中国”战略。
In China, sleep health issues are equally prominent. According to data from the National Health Commission, a high proportion of adults experience insufficient sleep, with sleep disorder rates among urban white-collar workers reaching over 30%. With accelerating aging and increasing chronic disease burdens, using AI to analyze sleep data for lifespan and disease risk prediction holds enormous public health significance. Several domestic hospitals and tech companies have begun exploring applications combining polysomnography with AI, such as integrating wearable device data into large models for daily sleep health management. If SleepFM technology can be localized, it could help hundreds of millions of people intervene early, reduce medical resource waste, and advance the "Healthy China" strategy.
The model's ability to predict specific conditions is particularly impressive. For neurological disorders, it excels at myoneural disorders, developmental delays, and speech/language issues. In circulatory diseases, it identifies atherosclerosis and acute pulmonary heart disease with high precision. Endocrine and metabolic predictions include diabetes complications and diabetic retinopathy, while respiratory predictions cover insufficiency and failure. Pregnancy complications and mental health disorders also show strong signals, suggesting sleep as a window into systemic health across life stages. Importantly, no single signal drives predictions; instead, the AI integrates multifactorial patterns, making it more robust than rule-based clinical scoring.
这项发现的潜在应用场景广阔。在临床上，睡眠实验室的常规检查可直接输出个性化风险报告，帮助医生制定预防方案。例如，高死亡风险患者可被优先推荐生活方式干预、药物治疗或进一步筛查。在保险和企业健康管理领域，AI睡眠预测能支持精准定价和员工福利优化。当然，隐私保护和伦理审查是前提，数据必须匿名化处理，且预测结果仅作为参考而非诊断。未来，研究团队计划将模型扩展到可穿戴设备数据，实现“在家睡一觉就能知寿命”的便捷模式。
The potential applications of this discovery are vast. Clinically, routine sleep lab tests could directly output personalized risk reports to help doctors devise prevention plans. For instance, patients with high mortality risk could be prioritized for lifestyle interventions, medications, or further screenings. In insurance and corporate wellness management, AI sleep predictions could support precise pricing and employee benefit optimization. Of course, privacy protection and ethical review are prerequisites—data must be anonymized, and predictions serve only as references, not diagnoses. In the future, the team plans to extend the model to wearable device data, enabling the convenient scenario of "sleep at home one night and know your lifespan."
However, challenges remain. While SleepFM outperforms baselines, real-world deployment requires validation across diverse populations, including different ethnicities, ages, and comorbidities common in China and globally. Accuracy for certain conditions like chronic kidney disease or stroke (around 0.78-0.79 C-index) is good but not perfect, underscoring the need for multimodal integration with genetics, blood tests, and lifestyle data. Interpretability is another focus: researchers are using techniques like SHAP to explain which sleep patterns drive predictions, building clinician trust. Over-reliance on AI without human oversight could lead to anxiety or false reassurance, so balanced communication is essential.
睡眠质量的改善本身就能延长寿命。大量流行病学研究表明，每晚7-9小时高质量睡眠可降低心血管疾病、糖尿病和认知衰退风险。SleepFM的预测结果或许能成为强大动机，鼓励人们重视睡眠卫生：规律作息、避免咖啡因、保持卧室凉爽黑暗等简单措施。结合AI反馈的个性化建议，例如“你的REM睡眠不足，建议增加运动以提升深度睡眠”，将使预防医学从被动转向主动。在中国“银发经济”和数字健康浪潮下，这一技术或催生新型睡眠健康产业，包括智能床垫、AI睡眠教练App等。
Improving sleep quality itself can extend lifespan. Extensive epidemiological studies show that 7-9 hours of high-quality sleep nightly reduces risks of cardiovascular disease, diabetes, and cognitive decline. SleepFM's predictions may serve as a powerful motivator, encouraging people to prioritize sleep hygiene: regular schedules, avoiding caffeine, and keeping bedrooms cool and dark. Personalized AI feedback, such as "Your REM sleep is insufficient; suggest increasing exercise to boost deep sleep," shifts preventive medicine from passive to active. Amid China's "silver economy" and digital health wave, this technology could spawn new sleep health industries, including smart mattresses and AI sleep coach apps.
Looking ahead, SleepFM represents a broader trend: foundation models unlocking hidden value in routine medical data. Similar to how large language models revolutionized text, multimodal sleep models could transform polysomnography from a niche diagnostic tool into a scalable population health platform. In China, integration with "East Data West Computing" infrastructure and domestic large models could accelerate adoption, supporting national goals for chronic disease prevention and healthy aging. International collaboration will be key to refining the model across global datasets and addressing biases.
要充分发挥这一技术的潜力，需要多方协同。医疗机构应加大睡眠监测设备的普及和AI培训；科技企业可开发用户友好接口，让普通人通过家用设备获取初步洞见；监管部门则需制定数据安全和伦理指南，避免信息滥用。同时，公众教育不可或缺：睡眠不是“浪费时间”，而是投资未来的健康资本。研究还提醒我们，年轻时忽视睡眠，可能在中年或老年付出寿命代价。
To fully leverage this technology's potential requires multi-stakeholder collaboration. Medical institutions should expand普及 of sleep monitoring equipment and AI training; tech companies can develop user-friendly interfaces for preliminary insights via home devices; regulators must establish data security and ethical guidelines to prevent misuse. Public education is equally essential: sleep is not "wasted time" but an investment in future health capital. The research also reminds us that neglecting sleep in youth may exact a lifespan cost in middle or old age.
Ethical considerations loom large. Predicting mortality or serious diseases years ahead raises questions about psychological impact—how to deliver "bad news" from sleep data without causing undue stress? Should insurers access such predictions? How to ensure equity so that low-income or rural populations in China aren't left behind? Researchers emphasize that SleepFM is a risk stratification tool, not a fortune teller; lifestyle, genetics, and medical interventions can still alter trajectories. Transparent, compassionate communication frameworks will be crucial for responsible deployment.
这一斯坦福重磅发现，不仅是AI在医学领域的又一里程碑，更是对人类与睡眠关系的重新审视。睡上一觉，AI或许就能勾勒出剩余寿命的轮廓，但真正的智慧在于行动：珍惜每晚的休息，用科学守护健康。未来，随着技术迭代和数据积累，类似模型或能实现更精准的个性化寿命预测，甚至指导精准医学干预。让我们期待，AI与睡眠的结合，能帮助更多人活得更长、更健康。
This Stanford groundbreaking discovery is not only another milestone for AI in medicine but also a re-examination of humanity's relationship with sleep. Sleep one night, and AI might outline the contours of remaining lifespan, but true wisdom lies in action: cherish every night's rest and guard health with science. In the future, with technological iterations and data accumulation, similar models may achieve even more precise personalized lifespan predictions and guide precision medicine interventions. Let us look forward to AI combined with sleep helping more people live longer and healthier lives.
In summary, SleepFM illustrates how one night's seemingly ordinary sleep encodes signals of decades-long health trajectories. By learning the "language of sleep," AI uncovers patterns invisible to the human eye—arousal indices, stage transitions, oxygen desaturations, and cross-signal dyssynchronies—that collectively forecast Parkinson's, cancers, heart failure, dementia, and mortality with remarkable accuracy. While not a definitive crystal ball, it offers an unprecedented proactive window into health. For individuals, it underscores the power of good sleep; for society, it points toward driven prevention that could ease healthcare burdens worldwide, including in China's vast population. Responsible innovation, rigorous validation, and ethical integration will determine whether this technology fulfills its promise of extending healthy lifespans.Stanford's groundbreaking discovery: Sleep one night, and AI can calculate your remaining lifespan.
2026年1月，斯坦福医学院的研究团队在《自然医学》杂志上发表了一项重磅研究，他们开发了一种名为SleepFM的AI基础模型，仅通过一晚的多导睡眠图（PSG）记录，就能预测超过100种健康状况的风险，包括全因死亡率。这一发现让“睡一觉就能预知寿命”从科幻走向现实，引发全球医学界和公众的广泛关注。传统上，睡眠监测主要用于诊断睡眠呼吸暂停等即时问题，而SleepFM则将睡眠数据转化为长期健康预测的“水晶球”，揭示了睡眠作为人体“健康档案”的深层价值。
In January 2026, a research team from Stanford Medicine published a landmark study in Nature Medicine. They developed an AI foundation model called SleepFM that, using just one night's polysomnography (PSG) recordings, can predict the risk of over 100 health conditions, including all-cause mortality. This discovery transforms the idea of "predicting lifespan from one night's sleep" from science fiction into reality, sparking widespread attention in the global medical community and among the public. Traditionally, sleep monitoring was mainly used to diagnose immediate issues like sleep apnea, but SleepFM turns sleep data into a "crystal ball" for long-term health predictions, revealing sleep's profound value as a "health archive" for the human body.
SleepFM模型的训练数据规模惊人，涵盖了超过6.5万名患者的近60万小时睡眠记录，这些数据来自斯坦福睡眠医学中心，并与长达25年的电子健康记录相匹配。AI通过自监督学习“掌握了睡眠的语言”，不仅能准确进行睡眠分期和呼吸暂停检测，还能在无标签数据上实现高效迁移学习。研究显示，该模型对130多种疾病的预测C指数（一致性指数）达到0.75以上，其中全因死亡率的C指数为0.84，痴呆为0.85，帕金森病高达0.89，心肌梗死0.81，前列腺癌0.89，乳腺癌0.87。这些指标远超基于人口统计学或传统睡眠指标的基准模型，证明睡眠中隐藏着丰富的纵向健康信号。
SleepFM was trained on an enormous dataset covering nearly 600,000 hours of sleep recordings from over 65,000 patients at Stanford's sleep centers, matched with up to 25 years of electronic health records. The AI "learned the language of sleep" through self-supervised learning, enabling accurate sleep staging, apnea detection, and efficient transfer learning on unlabeled data. The study showed that the model achieved C-indexes of at least 0.75 for predicting over 130 conditions, with all-cause mortality at 0.84, dementia at 0.85, Parkinson's disease at 0.89, myocardial infarction at 0.81, prostate cancer at 0.89, and breast cancer at 0.87. These metrics significantly outperform baselines based on demographics or traditional sleep metrics, proving that sleep harbors rich longitudinal health signals.
这项研究的创新之处在于SleepFM是一个多模态基础模型，它同时处理脑电图（EEG）、心电图（ECG）、呼吸信号、眼动、腿动等多种生理数据，而非依赖单一指标。传统睡眠分析往往只关注呼吸暂停指数或睡眠效率，但SleepFM能捕捉到跨系统的不匹配信号，例如大脑处于睡眠状态而心脏却表现出“清醒”迹象，这种“脱节”往往是未来心血管或神经退行性疾病的早期预警。研究者詹姆斯·邹（James Zou）教授表示：“睡眠数据就像一种被低估的‘生物语言’，AI帮助我们破译了它，从而提前多年发现疾病风险。”
The innovation of this study lies in SleepFM being a multimodal foundation model that simultaneously processes various physiological data such as electroencephalogram (EEG), electrocardiogram (ECG), respiratory signals, eye movements, and leg movements, rather than relying on single metrics. Traditional sleep analysis often focuses only on apnea-hypopnea index or sleep efficiency, but SleepFM captures cross-system mismatch signals—for instance, the brain appearing asleep while the heart shows "awake" patterns. Such "disconnections" often serve as early warnings for future cardiovascular or neurodegenerative diseases. Professor James Zou noted: "Sleep data is like an underestimated 'biological language,' and AI helps us decipher it to detect disease risks years in advance."
Beyond prediction accuracy, SleepFM demonstrates strong generalization. It performed competitively with specialized models like U-Sleep and YASA on standard tasks (mean F1 scores of 0.70–0.78 for sleep staging) and showed robust transfer learning on external datasets such as the Sleep Heart Health Study. This suggests the model isn't overfitting to Stanford-specific data but truly learning universal patterns in human sleep physiology. For mortality prediction, it achieved an AUROC of 0.85, surpassing demographic baselines (0.78) and end-to-end PSG models, highlighting how subtle factors like arousal burden, REM sleep reduction, hypoxemia, and sleep fragmentation integrate into powerful prognostic signals.
睡眠与寿命的关联早已被科学研究证实。早在2022年就有斯坦福相关研究显示，通过深度学习从睡眠研究中估计的“睡眠年龄”误差每增加10年，全因死亡风险上升29%，心血管死亡风险上升40%。对于60岁人群，睡眠年龄高估或低估10年，可能对应寿命减少约8.7年。SleepFM在此基础上实现了质的飞跃，它不只是估计生物年龄，还直接关联具体疾病风险和剩余寿命概率。研究指出，睡眠碎片化、低睡眠效率和间歇性低氧血症等特征，是独立于已知风险因素的死亡预测因子。
The link between sleep and lifespan has long been confirmed by scientific research. As early as 2022, related Stanford studies showed that every 10-year increase in error from deep learning-estimated "sleep age" raised all-cause mortality risk by 29% and cardiovascular mortality by 40%. For a 60-year-old, a 10-year over- or under-estimation in sleep age could correspond to an approximately 8.7-year reduction in life expectancy. SleepFM builds on this with a qualitative leap: it not only estimates biological age but directly links to specific disease risks and probabilities of remaining lifespan. The study notes that features like sleep fragmentation, low sleep efficiency, and intermittent hypoxemia are independent predictors of mortality beyond known risk factors.
在中国，睡眠健康问题同样突出。根据国家卫健委数据，成年居民睡眠不足比例较高，城市白领群体中睡眠障碍发生率可达30%以上。随着老龄化加剧和慢性病负担增加，利用AI分析睡眠数据预测寿命和疾病风险，具有巨大的公共卫生意义。国内多家医院和科技企业已开始探索类似多导睡眠监测与AI结合的应用，例如将可穿戴设备数据接入大模型，实现日常睡眠健康管理。如果SleepFM的技术能本土化落地，或许能帮助数亿人提前干预，减少医疗资源浪费，推动“健康中国”战略。
In China, sleep health issues are equally prominent. According to data from the National Health Commission, a high proportion of adults experience insufficient sleep, with sleep disorder rates among urban white-collar workers reaching over 30%. With accelerating aging and increasing chronic disease burdens, using AI to analyze sleep data for lifespan and disease risk prediction holds enormous public health significance. Several domestic hospitals and tech companies have begun exploring applications combining polysomnography with AI, such as integrating wearable device data into large models for daily sleep health management. If SleepFM technology can be localized, it could help hundreds of millions of people intervene early, reduce medical resource waste, and advance the "Healthy China" strategy.
The model's ability to predict specific conditions is particularly impressive. For neurological disorders, it excels at myoneural disorders, developmental delays, and speech/language issues. In circulatory diseases, it identifies atherosclerosis and acute pulmonary heart disease with high precision. Endocrine and metabolic predictions include diabetes complications and diabetic retinopathy, while respiratory predictions cover insufficiency and failure. Pregnancy complications and mental health disorders also show strong signals, suggesting sleep as a window into systemic health across life stages. Importantly, no single signal drives predictions; instead, the AI integrates multifactorial patterns, making it more robust than rule-based clinical scoring.
这项发现的潜在应用场景广阔。在临床上，睡眠实验室的常规检查可直接输出个性化风险报告，帮助医生制定预防方案。例如，高死亡风险患者可被优先推荐生活方式干预、药物治疗或进一步筛查。在保险和企业健康管理领域，AI睡眠预测能支持精准定价和员工福利优化。当然，隐私保护和伦理审查是前提，数据必须匿名化处理，且预测结果仅作为参考而非诊断。未来，研究团队计划将模型扩展到可穿戴设备数据，实现“在家睡一觉就能知寿命”的便捷模式。
The potential applications of this discovery are vast. Clinically, routine sleep lab tests could directly output personalized risk reports to help doctors devise prevention plans. For instance, patients with high mortality risk could be prioritized for lifestyle interventions, medications, or further screenings. In insurance and corporate wellness management, AI sleep predictions could support precise pricing and employee benefit optimization. Of course, privacy protection and ethical review are prerequisites—data must be anonymized, and predictions serve only as references, not diagnoses. In the future, the team plans to extend the model to wearable device data, enabling the convenient scenario of "sleep at home one night and know your lifespan."
However, challenges remain. While SleepFM outperforms baselines, real-world deployment requires validation across diverse populations, including different ethnicities, ages, and comorbidities common in China and globally. Accuracy for certain conditions like chronic kidney disease or stroke (around 0.78-0.79 C-index) is good but not perfect, underscoring the need for multimodal integration with genetics, blood tests, and lifestyle data. Interpretability is another focus: researchers are using techniques like SHAP to explain which sleep patterns drive predictions, building clinician trust. Over-reliance on AI without human oversight could lead to anxiety or false reassurance, so balanced communication is essential.
睡眠质量的改善本身就能延长寿命。大量流行病学研究表明，每晚7-9小时高质量睡眠可降低心血管疾病、糖尿病和认知衰退风险。SleepFM的预测结果或许能成为强大动机，鼓励人们重视睡眠卫生：规律作息、避免咖啡因、保持卧室凉爽黑暗等简单措施。结合AI反馈的个性化建议，例如“你的REM睡眠不足，建议增加运动以提升深度睡眠”，将使预防医学从被动转向主动。在中国“银发经济”和数字健康浪潮下，这一技术或催生新型睡眠健康产业，包括智能床垫、AI睡眠教练App等。
Improving sleep quality itself can extend lifespan. Extensive epidemiological studies show that 7-9 hours of high-quality sleep nightly reduces risks of cardiovascular disease, diabetes, and cognitive decline. SleepFM's predictions may serve as a powerful motivator, encouraging people to prioritize sleep hygiene: regular schedules, avoiding caffeine, and keeping bedrooms cool and dark. Personalized AI feedback, such as "Your REM sleep is insufficient; suggest increasing exercise to boost deep sleep," shifts preventive medicine from passive to active. Amid China's "silver economy" and digital health wave, this technology could spawn new sleep health industries, including smart mattresses and AI sleep coach apps.
Looking ahead, SleepFM represents a broader trend: foundation models unlocking hidden value in routine medical data. Similar to how large language models revolutionized text, multimodal sleep models could transform polysomnography from a niche diagnostic tool into a scalable population health platform. In China, integration with "East Data West Computing" infrastructure and domestic large models could accelerate adoption, supporting national goals for chronic disease prevention and healthy aging. International collaboration will be key to refining the model across global datasets and addressing biases.
要充分发挥这一技术的潜力，需要多方协同。医疗机构应加大睡眠监测设备的普及和AI培训；科技企业可开发用户友好接口，让普通人通过家用设备获取初步洞见；监管部门则需制定数据安全和伦理指南，避免信息滥用。同时，公众教育不可或缺：睡眠不是“浪费时间”，而是投资未来的健康资本。研究还提醒我们，年轻时忽视睡眠，可能在中年或老年付出寿命代价。
To fully leverage this technology's potential requires multi-stakeholder collaboration. Medical institutions should expand普及 of sleep monitoring equipment and AI training; tech companies can develop user-friendly interfaces for preliminary insights via home devices; regulators must establish data security and ethical guidelines to prevent misuse. Public education is equally essential: sleep is not "wasted time" but an investment in future health capital. The research also reminds us that neglecting sleep in youth may exact a lifespan cost in middle or old age.
Ethical considerations loom large. Predicting mortality or serious diseases years ahead raises questions about psychological impact—how to deliver "bad news" from sleep data without causing undue stress? Should insurers access such predictions? How to ensure equity so that low-income or rural populations in China aren't left behind? Researchers emphasize that SleepFM is a risk stratification tool, not a fortune teller; lifestyle, genetics, and medical interventions can still alter trajectories. Transparent, compassionate communication frameworks will be crucial for responsible deployment.
这一斯坦福重磅发现，不仅是AI在医学领域的又一里程碑，更是对人类与睡眠关系的重新审视。睡上一觉，AI或许就能勾勒出剩余寿命的轮廓，但真正的智慧在于行动：珍惜每晚的休息，用科学守护健康。未来，随着技术迭代和数据积累，类似模型或能实现更精准的个性化寿命预测，甚至指导精准医学干预。让我们期待，AI与睡眠的结合，能帮助更多人活得更长、更健康。
This Stanford groundbreaking discovery is not only another milestone for AI in medicine but also a re-examination of humanity's relationship with sleep. Sleep one night, and AI might outline the contours of remaining lifespan, but true wisdom lies in action: cherish every night's rest and guard health with science. In the future, with technological iterations and data accumulation, similar models may achieve even more precise personalized lifespan predictions and guide precision medicine interventions. Let us look forward to AI combined with sleep helping more people live longer and healthier lives.
In summary, SleepFM illustrates how one night's seemingly ordinary sleep encodes signals of decades-long health trajectories. By learning the "language of sleep," AI uncovers patterns invisible to the human eye—arousal indices, stage transitions, oxygen desaturations, and cross-signal dyssynchronies—that collectively forecast Parkinson's, cancers, heart failure, dementia, and mortality with remarkable accuracy. While not a definitive crystal ball, it offers an unprecedented proactive window into health. For individuals, it underscores the power of good sleep; for society, it points toward driven prevention that could ease healthcare burdens worldwide, including in China's vast population. Responsible innovation, rigorous validation, and ethical integration will determine whether this technology fulfills its promise of extending healthy lifespans.Stanford's groundbreaking discovery: Sleep one night, and AI can calculate your remaining lifespan.
2026年1月，斯坦福医学院的研究团队在《自然医学》杂志上发表了一项重磅研究，他们开发了一种名为SleepFM的AI基础模型，仅通过一晚的多导睡眠图（PSG）记录，就能预测超过100种健康状况的风险，包括全因死亡率。这一发现让“睡一觉就能预知寿命”从科幻走向现实，引发全球医学界和公众的广泛关注。传统上，睡眠监测主要用于诊断睡眠呼吸暂停等即时问题，而SleepFM则将睡眠数据转化为长期健康预测的“水晶球”，揭示了睡眠作为人体“健康档案”的深层价值。
In January 2026, a research team from Stanford Medicine published a landmark study in Nature Medicine. They developed an AI foundation model called SleepFM that, using just one night's polysomnography (PSG) recordings, can predict the risk of over 100 health conditions, including all-cause mortality. This discovery transforms the idea of "predicting lifespan from one night's sleep" from science fiction into reality, sparking widespread attention in the global medical community and among the public. Traditionally, sleep monitoring was mainly used to diagnose immediate issues like sleep apnea, but SleepFM turns sleep data into a "crystal ball" for long-term health predictions, revealing sleep's profound value as a "health archive" for the human body.
SleepFM模型的训练数据规模惊人，涵盖了超过6.5万名患者的近60万小时睡眠记录，这些数据来自斯坦福睡眠医学中心，并与长达25年的电子健康记录相匹配。AI通过自监督学习“掌握了睡眠的语言”，不仅能准确进行睡眠分期和呼吸暂停检测，还能在无标签数据上实现高效迁移学习。研究显示，该模型对130多种疾病的预测C指数（一致性指数）达到0.75以上，其中全因死亡率的C指数为0.84，痴呆为0.85，帕金森病高达0.89，心肌梗死0.81，前列腺癌0.89，乳腺癌0.87。这些指标远超基于人口统计学或传统睡眠指标的基准模型，证明睡眠中隐藏着丰富的纵向健康信号。
SleepFM was trained on an enormous dataset covering nearly 600,000 hours of sleep recordings from over 65,000 patients at Stanford's sleep centers, matched with up to 25 years of electronic health records. The AI "learned the language of sleep" through self-supervised learning, enabling accurate sleep staging, apnea detection, and efficient transfer learning on unlabeled data. The study showed that the model achieved C-indexes of at least 0.75 for predicting over 130 conditions, with all-cause mortality at 0.84, dementia at 0.85, Parkinson's disease at 0.89, myocardial infarction at 0.81, prostate cancer at 0.89, and breast cancer at 0.87. These metrics significantly outperform baselines based on demographics or traditional sleep metrics, proving that sleep harbors rich longitudinal health signals.
这项研究的创新之处在于SleepFM是一个多模态基础模型，它同时处理脑电图（EEG）、心电图（ECG）、呼吸信号、眼动、腿动等多种生理数据，而非依赖单一指标。传统睡眠分析往往只关注呼吸暂停指数或睡眠效率，但SleepFM能捕捉到跨系统的不匹配信号，例如大脑处于睡眠状态而心脏却表现出“清醒”迹象，这种“脱节”往往是未来心血管或神经退行性疾病的早期预警。研究者詹姆斯·邹（James Zou）教授表示：“睡眠数据就像一种被低估的‘生物语言’，AI帮助我们破译了它，从而提前多年发现疾病风险。”
The innovation of this study lies in SleepFM being a multimodal foundation model that simultaneously processes various physiological data such as electroencephalogram (EEG), electrocardiogram (ECG), respiratory signals, eye movements, and leg movements, rather than relying on single metrics. Traditional sleep analysis often focuses only on apnea-hypopnea index or sleep efficiency, but SleepFM captures cross-system mismatch signals—for instance, the brain appearing asleep while the heart shows "awake" patterns. Such "disconnections" often serve as early warnings for future cardiovascular or neurodegenerative diseases. Professor James Zou noted: "Sleep data is like an underestimated 'biological language,' and AI helps us decipher it to detect disease risks years in advance."
Beyond prediction accuracy, SleepFM demonstrates strong generalization. It performed competitively with specialized models like U-Sleep and YASA on standard tasks (mean F1 scores of 0.70–0.78 for sleep staging) and showed robust transfer learning on external datasets such as the Sleep Heart Health Study. This suggests the model isn't overfitting to Stanford-specific data but truly learning universal patterns in human sleep physiology. For mortality prediction, it achieved an AUROC of 0.85, surpassing demographic baselines (0.78) and end-to-end PSG models, highlighting how subtle factors like arousal burden, REM sleep reduction, hypoxemia, and sleep fragmentation integrate into powerful prognostic signals.
睡眠与寿命的关联早已被科学研究证实。早在2022年就有斯坦福相关研究显示，通过深度学习从睡眠研究中估计的“睡眠年龄”误差每增加10年，全因死亡风险上升29%，心血管死亡风险上升40%。对于60岁人群，睡眠年龄高估或低估10年，可能对应寿命减少约8.7年。SleepFM在此基础上实现了质的飞跃，它不只是估计生物年龄，还直接关联具体疾病风险和剩余寿命概率。研究指出，睡眠碎片化、低睡眠效率和间歇性低氧血症等特征，是独立于已知风险因素的死亡预测因子。
The link between sleep and lifespan has long been confirmed by scientific research. As early as 2022, related Stanford studies showed that every 10-year increase in error from deep learning-estimated "sleep age" raised all-cause mortality risk by 29% and cardiovascular mortality by 40%. For a 60-year-old, a 10-year over- or under-estimation in sleep age could correspond to an approximately 8.7-year reduction in life expectancy. SleepFM builds on this with a qualitative leap: it not only estimates biological age but directly links to specific disease risks and probabilities of remaining lifespan. The study notes that features like sleep fragmentation, low sleep efficiency, and intermittent hypoxemia are independent predictors of mortality beyond known risk factors.
在中国，睡眠健康问题同样突出。根据国家卫健委数据，成年居民睡眠不足比例较高，城市白领群体中睡眠障碍发生率可达30%以上。随着老龄化加剧和慢性病负担增加，利用AI分析睡眠数据预测寿命和疾病风险，具有巨大的公共卫生意义。国内多家医院和科技企业已开始探索类似多导睡眠监测与AI结合的应用，例如将可穿戴设备数据接入大模型，实现日常睡眠健康管理。如果SleepFM的技术能本土化落地，或许能帮助数亿人提前干预，减少医疗资源浪费，推动“健康中国”战略。
In China, sleep health issues are equally prominent. According to data from the National Health Commission, a high proportion of adults experience insufficient sleep, with sleep disorder rates among urban white-collar workers reaching over 30%. With accelerating aging and increasing chronic disease burdens, using AI to analyze sleep data for lifespan and disease risk prediction holds enormous public health significance. Several domestic hospitals and tech companies have begun exploring applications combining polysomnography with AI, such as integrating wearable device data into large models for daily sleep health management. If SleepFM technology can be localized, it could help hundreds of millions of people intervene early, reduce medical resource waste, and advance the "Healthy China" strategy.
The model's ability to predict specific conditions is particularly impressive. For neurological disorders, it excels at myoneural disorders, developmental delays, and speech/language issues. In circulatory diseases, it identifies atherosclerosis and acute pulmonary heart disease with high precision. Endocrine and metabolic predictions include diabetes complications and diabetic retinopathy, while respiratory predictions cover insufficiency and failure. Pregnancy complications and mental health disorders also show strong signals, suggesting sleep as a window into systemic health across life stages. Importantly, no single signal drives predictions; instead, the AI integrates multifactorial patterns, making it more robust than rule-based clinical scoring.
这项发现的潜在应用场景广阔。在临床上，睡眠实验室的常规检查可直接输出个性化风险报告，帮助医生制定预防方案。例如，高死亡风险患者可被优先推荐生活方式干预、药物治疗或进一步筛查。在保险和企业健康管理领域，AI睡眠预测能支持精准定价和员工福利优化。当然，隐私保护和伦理审查是前提，数据必须匿名化处理，且预测结果仅作为参考而非诊断。未来，研究团队计划将模型扩展到可穿戴设备数据，实现“在家睡一觉就能知寿命”的便捷模式。
The potential applications of this discovery are vast. Clinically, routine sleep lab tests could directly output personalized risk reports to help doctors devise prevention plans. For instance, patients with high mortality risk could be prioritized for lifestyle interventions, medications, or further screenings. In insurance and corporate wellness management, AI sleep predictions could support precise pricing and employee benefit optimization. Of course, privacy protection and ethical review are prerequisites—data must be anonymized, and predictions serve only as references, not diagnoses. In the future, the team plans to extend the model to wearable device data, enabling the convenient scenario of "sleep at home one night and know your lifespan."
However, challenges remain. While SleepFM outperforms baselines, real-world deployment requires validation across diverse populations, including different ethnicities, ages, and comorbidities common in China and globally. Accuracy for certain conditions like chronic kidney disease or stroke (around 0.78-0.79 C-index) is good but not perfect, underscoring the need for multimodal integration with genetics, blood tests, and lifestyle data. Interpretability is another focus: researchers are using techniques like SHAP to explain which sleep patterns drive predictions, building clinician trust. Over-reliance on AI without human oversight could lead to anxiety or false reassurance, so balanced communication is essential.
睡眠质量的改善本身就能延长寿命。大量流行病学研究表明，每晚7-9小时高质量睡眠可降低心血管疾病、糖尿病和认知衰退风险。SleepFM的预测结果或许能成为强大动机，鼓励人们重视睡眠卫生：规律作息、避免咖啡因、保持卧室凉爽黑暗等简单措施。结合AI反馈的个性化建议，例如“你的REM睡眠不足，建议增加运动以提升深度睡眠”，将使预防医学从被动转向主动。在中国“银发经济”和数字健康浪潮下，这一技术或催生新型睡眠健康产业，包括智能床垫、AI睡眠教练App等。
Improving sleep quality itself can extend lifespan. Extensive epidemiological studies show that 7-9 hours of high-quality sleep nightly reduces risks of cardiovascular disease, diabetes, and cognitive decline. SleepFM's predictions may serve as a powerful motivator, encouraging people to prioritize sleep hygiene: regular schedules, avoiding caffeine, and keeping bedrooms cool and dark. Personalized AI feedback, such as "Your REM sleep is insufficient; suggest increasing exercise to boost deep sleep," shifts preventive medicine from passive to active. Amid China's "silver economy" and digital health wave, this technology could spawn new sleep health industries, including smart mattresses and AI sleep coach apps.
Looking ahead, SleepFM represents a broader trend: foundation models unlocking hidden value in routine medical data. Similar to how large language models revolutionized text, multimodal sleep models could transform polysomnography from a niche diagnostic tool into a scalable population health platform. In China, integration with "East Data West Computing" infrastructure and domestic large models could accelerate adoption, supporting national goals for chronic disease prevention and healthy aging. International collaboration will be key to refining the model across global datasets and addressing biases.
要充分发挥这一技术的潜力，需要多方协同。医疗机构应加大睡眠监测设备的普及和AI培训；科技企业可开发用户友好接口，让普通人通过家用设备获取初步洞见；监管部门则需制定数据安全和伦理指南，避免信息滥用。同时，公众教育不可或缺：睡眠不是“浪费时间”，而是投资未来的健康资本。研究还提醒我们，年轻时忽视睡眠，可能在中年或老年付出寿命代价。
To fully leverage this technology's potential requires multi-stakeholder collaboration. Medical institutions should expand普及 of sleep monitoring equipment and AI training; tech companies can develop user-friendly interfaces for preliminary insights via home devices; regulators must establish data security and ethical guidelines to prevent misuse. Public education is equally essential: sleep is not "wasted time" but an investment in future health capital. The research also reminds us that neglecting sleep in youth may exact a lifespan cost in middle or old age.
Ethical considerations loom large. Predicting mortality or serious diseases years ahead raises questions about psychological impact—how to deliver "bad news" from sleep data without causing undue stress? Should insurers access such predictions? How to ensure equity so that low-income or rural populations in China aren't left behind? Researchers emphasize that SleepFM is a risk stratification tool, not a fortune teller; lifestyle, genetics, and medical interventions can still alter trajectories. Transparent, compassionate communication frameworks will be crucial for responsible deployment.
这一斯坦福重磅发现，不仅是AI在医学领域的又一里程碑，更是对人类与睡眠关系的重新审视。睡上一觉，AI或许就能勾勒出剩余寿命的轮廓，但真正的智慧在于行动：珍惜每晚的休息，用科学守护健康。未来，随着技术迭代和数据积累，类似模型或能实现更精准的个性化寿命预测，甚至指导精准医学干预。让我们期待，AI与睡眠的结合，能帮助更多人活得更长、更健康。
This Stanford groundbreaking discovery is not only another milestone for AI in medicine but also a re-examination of humanity's relationship with sleep. Sleep one night, and AI might outline the contours of remaining lifespan, but true wisdom lies in action: cherish every night's rest and guard health with science. In the future, with technological iterations and data accumulation, similar models may achieve even more precise personalized lifespan predictions and guide precision medicine interventions. Let us look forward to AI combined with sleep helping more people live longer and healthier lives.
In summary, SleepFM illustrates how one night's seemingly ordinary sleep encodes signals of decades-long health trajectories. By learning the "language of sleep," AI uncovers patterns invisible to the human eye—arousal indices, stage transitions, oxygen desaturations, and cross-signal dyssynchronies—that collectively forecast Parkinson's, cancers, heart failure, dementia, and mortality with remarkable accuracy. While not a definitive crystal ball, it offers an unprecedented proactive window into health. For individuals, it underscores the power of good sleep; for society, it points toward driven prevention that could ease healthcare burdens worldwide, including in China's vast population. Responsible innovation, rigorous validation, and ethical integration will determine whether this technology fulfills its promise of extending healthy lifespans.Stanford's groundbreaking discovery: Sleep one night, and AI can calculate your remaining lifespan.
2026年1月，斯坦福医学院的研究团队在《自然医学》杂志上发表了一项重磅研究，他们开发了一种名为SleepFM的AI基础模型，仅通过一晚的多导睡眠图（PSG）记录，就能预测超过100种健康状况的风险，包括全因死亡率。这一发现让“睡一觉就能预知寿命”从科幻走向现实，引发全球医学界和公众的广泛关注。传统上，睡眠监测主要用于诊断睡眠呼吸暂停等即时问题，而SleepFM则将睡眠数据转化为长期健康预测的“水晶球”，揭示了睡眠作为人体“健康档案”的深层价值。
In January 2026, a research team from Stanford Medicine published a landmark study in Nature Medicine. They developed an AI foundation model called SleepFM that, using just one night's polysomnography (PSG) recordings, can predict the risk of over 100 health conditions, including all-cause mortality. This discovery transforms the idea of "predicting lifespan from one night's sleep" from science fiction into reality, sparking widespread attention in the global medical community and among the public. Traditionally, sleep monitoring was mainly used to diagnose immediate issues like sleep apnea, but SleepFM turns sleep data into a "crystal ball" for long-term health predictions, revealing sleep's profound value as a "health archive" for the human body.
SleepFM模型的训练数据规模惊人，涵盖了超过6.5万名患者的近60万小时睡眠记录，这些数据来自斯坦福睡眠医学中心，并与长达25年的电子健康记录相匹配。AI通过自监督学习“掌握了睡眠的语言”，不仅能准确进行睡眠分期和呼吸暂停检测，还能在无标签数据上实现高效迁移学习。研究显示，该模型对130多种疾病的预测C指数（一致性指数）达到0.75以上，其中全因死亡率的C指数为0.84，痴呆为0.85，帕金森病高达0.89，心肌梗死0.81，前列腺癌0.89，乳腺癌0.87。这些指标远超基于人口统计学或传统睡眠指标的基准模型，证明睡眠中隐藏着丰富的纵向健康信号。
SleepFM was trained on an enormous dataset covering nearly 600,000 hours of sleep recordings from over 65,000 patients at Stanford's sleep centers, matched with up to 25 years of electronic health records. The AI "learned the language of sleep" through self-supervised learning, enabling accurate sleep staging, apnea detection, and efficient transfer learning on unlabeled data. The study showed that the model achieved C-indexes of at least 0.75 for predicting over 130 conditions, with all-cause mortality at 0.84, dementia at 0.85, Parkinson's disease at 0.89, myocardial infarction at 0.81, prostate cancer at 0.89, and breast cancer at 0.87. These metrics significantly outperform baselines based on demographics or traditional sleep metrics, proving that sleep harbors rich longitudinal health signals.
这项研究的创新之处在于SleepFM是一个多模态基础模型，它同时处理脑电图（EEG）、心电图（ECG）、呼吸信号、眼动、腿动等多种生理数据，而非依赖单一指标。传统睡眠分析往往只关注呼吸暂停指数或睡眠效率，但SleepFM能捕捉到跨系统的不匹配信号，例如大脑处于睡眠状态而心脏却表现出“清醒”迹象，这种“脱节”往往是未来心血管或神经退行性疾病的早期预警。研究者詹姆斯·邹（James Zou）教授表示：“睡眠数据就像一种被低估的‘生物语言’，AI帮助我们破译了它，从而提前多年发现疾病风险。”
The innovation of this study lies in SleepFM being a multimodal foundation model that simultaneously processes various physiological data such as electroencephalogram (EEG), electrocardiogram (ECG), respiratory signals, eye movements, and leg movements, rather than relying on single metrics. Traditional sleep analysis often focuses only on apnea-hypopnea index or sleep efficiency, but SleepFM captures cross-system mismatch signals—for instance, the brain appearing asleep while the heart shows "awake" patterns. Such "disconnections" often serve as early warnings for future cardiovascular or neurodegenerative diseases. Professor James Zou noted: "Sleep data is like an underestimated 'biological language,' and AI helps us decipher it to detect disease risks years in advance."
Beyond prediction accuracy, SleepFM demonstrates strong generalization. It performed competitively with specialized models like U-Sleep and YASA on standard tasks (mean F1 scores of 0.70–0.78 for sleep staging) and showed robust transfer learning on external datasets such as the Sleep Heart Health Study. This suggests the model isn't overfitting to Stanford-specific data but truly learning universal patterns in human sleep physiology. For mortality prediction, it achieved an AUROC of 0.85, surpassing demographic baselines (0.78) and end-to-end PSG models, highlighting how subtle factors like arousal burden, REM sleep reduction, hypoxemia, and sleep fragmentation integrate into powerful prognostic signals.
睡眠与寿命的关联早已被科学研究证实。早在2022年就有斯坦福相关研究显示，通过深度学习从睡眠研究中估计的“睡眠年龄”误差每增加10年，全因死亡风险上升29%，心血管死亡风险上升40%。对于60岁人群，睡眠年龄高估或低估10年，可能对应寿命减少约8.7年。SleepFM在此基础上实现了质的飞跃，它不只是估计生物年龄，还直接关联具体疾病风险和剩余寿命概率。研究指出，睡眠碎片化、低睡眠效率和间歇性低氧血症等特征，是独立于已知风险因素的死亡预测因子。
The link between sleep and lifespan has long been confirmed by scientific research. As early as 2022, related Stanford studies showed that every 10-year increase in error from deep learning-estimated "sleep age" raised all-cause mortality risk by 29% and cardiovascular mortality by 40%. For a 60-year-old, a 10-year over- or under-estimation in sleep age could correspond to an approximately 8.7-year reduction in life expectancy. SleepFM builds on this with a qualitative leap: it not only estimates biological age but directly links to specific disease risks and probabilities of remaining lifespan. The study notes that features like sleep fragmentation, low sleep efficiency, and intermittent hypoxemia are independent predictors of mortality beyond known risk factors.
在中国，睡眠健康问题同样突出。根据国家卫健委数据，成年居民睡眠不足比例较高，城市白领群体中睡眠障碍发生率可达30%以上。随着老龄化加剧和慢性病负担增加，利用AI分析睡眠数据预测寿命和疾病风险，具有巨大的公共卫生意义。国内多家医院和科技企业已开始探索类似多导睡眠监测与AI结合的应用，例如将可穿戴设备数据接入大模型，实现日常睡眠健康管理。如果SleepFM的技术能本土化落地，或许能帮助数亿人提前干预，减少医疗资源浪费，推动“健康中国”战略。
In China, sleep health issues are equally prominent. According to data from the National Health Commission, a high proportion of adults experience insufficient sleep, with sleep disorder rates among urban white-collar workers reaching over 30%. With accelerating aging and increasing chronic disease burdens, using AI to analyze sleep data for lifespan and disease risk prediction holds enormous public health significance. Several domestic hospitals and tech companies have begun exploring applications combining polysomnography with AI, such as integrating wearable device data into large models for daily sleep health management. If SleepFM technology can be localized, it could help hundreds of millions of people intervene early, reduce medical resource waste, and advance the "Healthy China" strategy.
The model's ability to predict specific conditions is particularly impressive. For neurological disorders, it excels at myoneural disorders, developmental delays, and speech/language issues. In circulatory diseases, it identifies atherosclerosis and acute pulmonary heart disease with high precision. Endocrine and metabolic predictions include diabetes complications and diabetic retinopathy, while respiratory predictions cover insufficiency and failure. Pregnancy complications and mental health disorders also show strong signals, suggesting sleep as a window into systemic health across life stages. Importantly, no single signal drives predictions; instead, the AI integrates multifactorial patterns, making it more robust than rule-based clinical scoring.
这项发现的潜在应用场景广阔。在临床上，睡眠实验室的常规检查可直接输出个性化风险报告，帮助医生制定预防方案。例如，高死亡风险患者可被优先推荐生活方式干预、药物治疗或进一步筛查。在保险和企业健康管理领域，AI睡眠预测能支持精准定价和员工福利优化。当然，隐私保护和伦理审查是前提，数据必须匿名化处理，且预测结果仅作为参考而非诊断。未来，研究团队计划将模型扩展到可穿戴设备数据，实现“在家睡一觉就能知寿命”的便捷模式。
The potential applications of this discovery are vast. Clinically, routine sleep lab tests could directly output personalized risk reports to help doctors devise prevention plans. For instance, patients with high mortality risk could be prioritized for lifestyle interventions, medications, or further screenings. In insurance and corporate wellness management, AI sleep predictions could support precise pricing and employee benefit optimization. Of course, privacy protection and ethical review are prerequisites—data must be anonymized, and predictions serve only as references, not diagnoses. In the future, the team plans to extend the model to wearable device data, enabling the convenient scenario of "sleep at home one night and know your lifespan."
However, challenges remain. While SleepFM outperforms baselines, real-world deployment requires validation across diverse populations, including different ethnicities, ages, and comorbidities common in China and globally. Accuracy for certain conditions like chronic kidney disease or stroke (around 0.78-0.79 C-index) is good but not perfect, underscoring the need for multimodal integration with genetics, blood tests, and lifestyle data. Interpretability is another focus: researchers are using techniques like SHAP to explain which sleep patterns drive predictions, building clinician trust. Over-reliance on AI without human oversight could lead to anxiety or false reassurance, so balanced communication is essential.
睡眠质量的改善本身就能延长寿命。大量流行病学研究表明，每晚7-9小时高质量睡眠可降低心血管疾病、糖尿病和认知衰退风险。SleepFM的预测结果或许能成为强大动机，鼓励人们重视睡眠卫生：规律作息、避免咖啡因、保持卧室凉爽黑暗等简单措施。结合AI反馈的个性化建议，例如“你的REM睡眠不足，建议增加运动以提升深度睡眠”，将使预防医学从被动转向主动。在中国“银发经济”和数字健康浪潮下，这一技术或催生新型睡眠健康产业，包括智能床垫、AI睡眠教练App等。
Improving sleep quality itself can extend lifespan. Extensive epidemiological studies show that 7-9 hours of high-quality sleep nightly reduces risks of cardiovascular disease, diabetes, and cognitive decline. SleepFM's predictions may serve as a powerful motivator, encouraging people to prioritize sleep hygiene: regular schedules, avoiding caffeine, and keeping bedrooms cool and dark. Personalized AI feedback, such as "Your REM sleep is insufficient; suggest increasing exercise to boost deep sleep," shifts preventive medicine from passive to active. Amid China's "silver economy" and digital health wave, this technology could spawn new sleep health industries, including smart mattresses and AI sleep coach apps.
Looking ahead, SleepFM represents a broader trend: foundation models unlocking hidden value in routine medical data. Similar to how large language models revolutionized text, multimodal sleep models could transform polysomnography from a niche diagnostic tool into a scalable population health platform. In China, integration with "East Data West Computing" infrastructure and domestic large models could accelerate adoption, supporting national goals for chronic disease prevention and healthy aging. International collaboration will be key to refining the model across global datasets and addressing biases.
要充分发挥这一技术的潜力，需要多方协同。医疗机构应加大睡眠监测设备的普及和AI培训；科技企业可开发用户友好接口，让普通人通过家用设备获取初步洞见；监管部门则需制定数据安全和伦理指南，避免信息滥用。同时，公众教育不可或缺：睡眠不是“浪费时间”，而是投资未来的健康资本。研究还提醒我们，年轻时忽视睡眠，可能在中年或老年付出寿命代价。
To fully leverage this technology's potential requires multi-stakeholder collaboration. Medical institutions should expand普及 of sleep monitoring equipment and AI training; tech companies can develop user-friendly interfaces for preliminary insights via home devices; regulators must establish data security and ethical guidelines to prevent misuse. Public education is equally essential: sleep is not "wasted time" but an investment in future health capital. The research also reminds us that neglecting sleep in youth may exact a lifespan cost in middle or old age.
Ethical considerations loom large. Predicting mortality or serious diseases years ahead raises questions about psychological impact—how to deliver "bad news" from sleep data without causing undue stress? Should insurers access such predictions? How to ensure equity so that low-income or rural populations in China aren't left behind? Researchers emphasize that SleepFM is a risk stratification tool, not a fortune teller; lifestyle, genetics, and medical interventions can still alter trajectories. Transparent, compassionate communication frameworks will be crucial for responsible deployment.
这一斯坦福重磅发现，不仅是AI在医学领域的又一里程碑，更是对人类与睡眠关系的重新审视。睡上一觉，AI或许就能勾勒出剩余寿命的轮廓，但真正的智慧在于行动：珍惜每晚的休息，用科学守护健康。未来，随着技术迭代和数据积累，类似模型或能实现更精准的个性化寿命预测，甚至指导精准医学干预。让我们期待，AI与睡眠的结合，能帮助更多人活得更长、更健康。
This Stanford groundbreaking discovery is not only another milestone for AI in medicine but also a re-examination of humanity's relationship with sleep. Sleep one night, and AI might outline the contours of remaining lifespan, but true wisdom lies in action: cherish every night's rest and guard health with science. In the future, with technological iterations and data accumulation, similar models may achieve even more precise personalized lifespan predictions and guide precision medicine interventions. Let us look forward to AI combined with sleep helping more people live longer and healthier lives.
In summary, SleepFM illustrates how one night's seemingly ordinary sleep encodes signals of decades-long health trajectories. By learning the "language of sleep," AI uncovers patterns invisible to the human eye—arousal indices, stage transitions, oxygen desaturations, and cross-signal dyssynchronies—that collectively forecast Parkinson's, cancers, heart failure, dementia, and mortality with remarkable accuracy. While not a definitive crystal ball, it offers an unprecedented proactive window into health. For individuals, it underscores the power of good sleep; for society, it points toward driven prevention that could ease healthcare burdens worldwide, including in China's vast population. Responsible innovation, rigorous validation, and ethical integration will determine whether this technology fulfills its promise of extending healthy lifespans.Stanford's groundbreaking discovery: Sleep one night, and AI can calculate your remaining lifespan.
2026年1月，斯坦福医学院的研究团队在《自然医学》杂志上发表了一项重磅研究，他们开发了一种名为SleepFM的AI基础模型，仅通过一晚的多导睡眠图（PSG）记录，就能预测超过100种健康状况的风险，包括全因死亡率。这一发现让“睡一觉就能预知寿命”从科幻走向现实，引发全球医学界和公众的广泛关注。传统上，睡眠监测主要用于诊断睡眠呼吸暂停等即时问题，而SleepFM则将睡眠数据转化为长期健康预测的“水晶球”，揭示了睡眠作为人体“健康档案”的深层价值。
In January 2026, a research team from Stanford Medicine published a landmark study in Nature Medicine. They developed an AI foundation model called SleepFM that, using just one night's polysomnography (PSG) recordings, can predict the risk of over 100 health conditions, including all-cause mortality. This discovery transforms the idea of "predicting lifespan from one night's sleep" from science fiction into reality, sparking widespread attention in the global medical community and among the public. Traditionally, sleep monitoring was mainly used to diagnose immediate issues like sleep apnea, but SleepFM turns sleep data into a "crystal ball" for long-term health predictions, revealing sleep's profound value as a "health archive" for the human body.
SleepFM模型的训练数据规模惊人，涵盖了超过6.5万名患者的近60万小时睡眠记录，这些数据来自斯坦福睡眠医学中心，并与长达25年的电子健康记录相匹配。AI通过自监督学习“掌握了睡眠的语言”，不仅能准确进行睡眠分期和呼吸暂停检测，还能在无标签数据上实现高效迁移学习。研究显示，该模型对130多种疾病的预测C指数（一致性指数）达到0.75以上，其中全因死亡率的C指数为0.84，痴呆为0.85，帕金森病高达0.89，心肌梗死0.81，前列腺癌0.89，乳腺癌0.87。这些指标远超基于人口统计学或传统睡眠指标的基准模型，证明睡眠中隐藏着丰富的纵向健康信号。
SleepFM was trained on an enormous dataset covering nearly 600,000 hours of sleep recordings from over 65,000 patients at Stanford's sleep centers, matched with up to 25 years of electronic health records. The AI "learned the language of sleep" through self-supervised learning, enabling accurate sleep staging, apnea detection, and efficient transfer learning on unlabeled data. The study showed that the model achieved C-indexes of at least 0.75 for predicting over 130 conditions, with all-cause mortality at 0.84, dementia at 0.85, Parkinson's disease at 0.89, myocardial infarction at 0.81, prostate cancer at 0.89, and breast cancer at 0.87. These metrics significantly outperform baselines based on demographics or traditional sleep metrics, proving that sleep harbors rich longitudinal health signals.
这项研究的创新之处在于SleepFM是一个多模态基础模型，它同时处理脑电图（EEG）、心电图（ECG）、呼吸信号、眼动、腿动等多种生理数据，而非依赖单一指标。传统睡眠分析往往只关注呼吸暂停指数或睡眠效率，但SleepFM能捕捉到跨系统的不匹配信号，例如大脑处于睡眠状态而心脏却表现出“清醒”迹象，这种“脱节”往往是未来心血管或神经退行性疾病的早期预警。研究者詹姆斯·邹（James Zou）教授表示：“睡眠数据就像一种被低估的‘生物语言’，AI帮助我们破译了它，从而提前多年发现疾病风险。”
The innovation of this study lies in SleepFM being a multimodal foundation model that simultaneously processes various physiological data such as electroencephalogram (EEG), electrocardiogram (ECG), respiratory signals, eye movements, and leg movements, rather than relying on single metrics. Traditional sleep analysis often focuses only on apnea-hypopnea index or sleep efficiency, but SleepFM captures cross-system mismatch signals—for instance, the brain appearing asleep while the heart shows "awake" patterns. Such "disconnections" often serve as early warnings for future cardiovascular or neurodegenerative diseases. Professor James Zou noted: "Sleep data is like an underestimated 'biological language,' and AI helps us decipher it to detect disease risks years in advance."
Beyond prediction accuracy, SleepFM demonstrates strong generalization. It performed competitively with specialized models like U-Sleep and YASA on standard tasks (mean F1 scores of 0.70–0.78 for sleep staging) and showed robust transfer learning on external datasets such as the Sleep Heart Health Study. This suggests the model isn't overfitting to Stanford-specific data but truly learning universal patterns in human sleep physiology. For mortality prediction, it achieved an AUROC of 0.85, surpassing demographic baselines (0.78) and end-to-end PSG models, highlighting how subtle factors like arousal burden, REM sleep reduction, hypoxemia, and sleep fragmentation integrate into powerful prognostic signals.
睡眠与寿命的关联早已被科学研究证实。早在2022年就有斯坦福相关研究显示，通过深度学习从睡眠研究中估计的“睡眠年龄”误差每增加10年，全因死亡风险上升29%，心血管死亡风险上升40%。对于60岁人群，睡眠年龄高估或低估10年，可能对应寿命减少约8.7年。SleepFM在此基础上实现了质的飞跃，它不只是估计生物年龄，还直接关联具体疾病风险和剩余寿命概率。研究指出，睡眠碎片化、低睡眠效率和间歇性低氧血症等特征，是独立于已知风险因素的死亡预测因子。
The link between sleep and lifespan has long been confirmed by scientific research. As early as 2022, related Stanford studies showed that every 10-year increase in error from deep learning-estimated "sleep age" raised all-cause mortality risk by 29% and cardiovascular mortality by 40%. For a 60-year-old, a 10-year over- or under-estimation in sleep age could correspond to an approximately 8.7-year reduction in life expectancy. SleepFM builds on this with a qualitative leap: it not only estimates biological age but directly links to specific disease risks and probabilities of remaining lifespan. The study notes that features like sleep fragmentation, low sleep efficiency, and intermittent hypoxemia are independent predictors of mortality beyond known risk factors.
在中国，睡眠健康问题同样突出。根据国家卫健委数据，成年居民睡眠不足比例较高，城市白领群体中睡眠障碍发生率可达30%以上。随着老龄化加剧和慢性病负担增加，利用AI分析睡眠数据预测寿命和疾病风险，具有巨大的公共卫生意义。国内多家医院和科技企业已开始探索类似多导睡眠监测与AI结合的应用，例如将可穿戴设备数据接入大模型，实现日常睡眠健康管理。如果SleepFM的技术能本土化落地，或许能帮助数亿人提前干预，减少医疗资源浪费，推动“健康中国”战略。
In China, sleep health issues are equally prominent. According to data from the National Health Commission, a high proportion of adults experience insufficient sleep, with sleep disorder rates among urban white-collar workers reaching over 30%. With accelerating aging and increasing chronic disease burdens, using AI to analyze sleep data for lifespan and disease risk prediction holds enormous public health significance. Several domestic hospitals and tech companies have begun exploring applications combining polysomnography with AI, such as integrating wearable device data into large models for daily sleep health management. If SleepFM technology can be localized, it could help hundreds of millions of people intervene early, reduce medical resource waste, and advance the "Healthy China" strategy.
The model's ability to predict specific conditions is particularly impressive. For neurological disorders, it excels at myoneural disorders, developmental delays, and speech/language issues. In circulatory diseases, it identifies atherosclerosis and acute pulmonary heart disease with high precision. Endocrine and metabolic predictions include diabetes complications and diabetic retinopathy, while respiratory predictions cover insufficiency and failure. Pregnancy complications and mental health disorders also show strong signals, suggesting sleep as a window into systemic health across life stages. Importantly, no single signal drives predictions; instead, the AI integrates multifactorial patterns, making it more robust than rule-based clinical scoring.
这项发现的潜在应用场景广阔。在临床上，睡眠实验室的常规检查可直接输出个性化风险报告，帮助医生制定预防方案。例如，高死亡风险患者可被优先推荐生活方式干预、药物治疗或进一步筛查。在保险和企业健康管理领域，AI睡眠预测能支持精准定价和员工福利优化。当然，隐私保护和伦理审查是前提，数据必须匿名化处理，且预测结果仅作为参考而非诊断。未来，研究团队计划将模型扩展到可穿戴设备数据，实现“在家睡一觉就能知寿命”的便捷模式。
The potential applications of this discovery are vast. Clinically, routine sleep lab tests could directly output personalized risk reports to help doctors devise prevention plans. For instance, patients with high mortality risk could be prioritized for lifestyle interventions, medications, or further screenings. In insurance and corporate wellness management, AI sleep predictions could support precise pricing and employee benefit optimization. Of course, privacy protection and ethical review are prerequisites—data must be anonymized, and predictions serve only as references, not diagnoses. In the future, the team plans to extend the model to wearable device data, enabling the convenient scenario of "sleep at home one night and know your lifespan."
However, challenges remain. While SleepFM outperforms baselines, real-world deployment requires validation across diverse populations, including different ethnicities, ages, and comorbidities common in China and globally. Accuracy for certain conditions like chronic kidney disease or stroke (around 0.78-0.79 C-index) is good but not perfect, underscoring the need for multimodal integration with genetics, blood tests, and lifestyle data. Interpretability is another focus: researchers are using techniques like SHAP to explain which sleep patterns drive predictions, building clinician trust. Over-reliance on AI without human oversight could lead to anxiety or false reassurance, so balanced communication is essential.
睡眠质量的改善本身就能延长寿命。大量流行病学研究表明，每晚7-9小时高质量睡眠可降低心血管疾病、糖尿病和认知衰退风险。SleepFM的预测结果或许能成为强大动机，鼓励人们重视睡眠卫生：规律作息、避免咖啡因、保持卧室凉爽黑暗等简单措施。结合AI反馈的个性化建议，例如“你的REM睡眠不足，建议增加运动以提升深度睡眠”，将使预防医学从被动转向主动。在中国“银发经济”和数字健康浪潮下，这一技术或催生新型睡眠健康产业，包括智能床垫、AI睡眠教练App等。
Improving sleep quality itself can extend lifespan. Extensive epidemiological studies show that 7-9 hours of high-quality sleep nightly reduces risks of cardiovascular disease, diabetes, and cognitive decline. SleepFM's predictions may serve as a powerful motivator, encouraging people to prioritize sleep hygiene: regular schedules, avoiding caffeine, and keeping bedrooms cool and dark. Personalized AI feedback, such as "Your REM sleep is insufficient; suggest increasing exercise to boost deep sleep," shifts preventive medicine from passive to active. Amid China's "silver economy" and digital health wave, this technology could spawn new sleep health industries, including smart mattresses and AI sleep coach apps.
Looking ahead, SleepFM represents a broader trend: foundation models unlocking hidden value in routine medical data. Similar to how large language models revolutionized text, multimodal sleep models could transform polysomnography from a niche diagnostic tool into a scalable population health platform. In China, integration with "East Data West Computing" infrastructure and domestic large models could accelerate adoption, supporting national goals for chronic disease prevention and healthy aging. International collaboration will be key to refining the model across global datasets and addressing biases.
要充分发挥这一技术的潜力，需要多方协同。医疗机构应加大睡眠监测设备的普及和AI培训；科技企业可开发用户友好接口，让普通人通过家用设备获取初步洞见；监管部门则需制定数据安全和伦理指南，避免信息滥用。同时，公众教育不可或缺：睡眠不是“浪费时间”，而是投资未来的健康资本。研究还提醒我们，年轻时忽视睡眠，可能在中年或老年付出寿命代价。
To fully leverage this technology's potential requires multi-stakeholder collaboration. Medical institutions should expand普及 of sleep monitoring equipment and AI training; tech companies can develop user-friendly interfaces for preliminary insights via home devices; regulators must establish data security and ethical guidelines to prevent misuse. Public education is equally essential: sleep is not "wasted time" but an investment in future health capital. The research also reminds us that neglecting sleep in youth may exact a lifespan cost in middle or old age.
Ethical considerations loom large. Predicting mortality or serious diseases years ahead raises questions about psychological impact—how to deliver "bad news" from sleep data without causing undue stress? Should insurers access such predictions? How to ensure equity so that low-income or rural populations in China aren't left behind? Researchers emphasize that SleepFM is a risk stratification tool, not a fortune teller; lifestyle, genetics, and medical interventions can still alter trajectories. Transparent, compassionate communication frameworks will be crucial for responsible deployment.
这一斯坦福重磅发现，不仅是AI在医学领域的又一里程碑，更是对人类与睡眠关系的重新审视。睡上一觉，AI或许就能勾勒出剩余寿命的轮廓，但真正的智慧在于行动：珍惜每晚的休息，用科学守护健康。未来，随着技术迭代和数据积累，类似模型或能实现更精准的个性化寿命预测，甚至指导精准医学干预。让我们期待，AI与睡眠的结合，能帮助更多人活得更长、更健康。
This Stanford groundbreaking discovery is not only another milestone for AI in medicine but also a re-examination of humanity's relationship with sleep. Sleep one night, and AI might outline the contours of remaining lifespan, but true wisdom lies in action: cherish every night's rest and guard health with science. In the future, with technological iterations and data accumulation, similar models may achieve even more precise personalized lifespan predictions and guide precision medicine interventions. Let us look forward to AI combined with sleep helping more people live longer and healthier lives.
In summary, SleepFM illustrates how one night's seemingly ordinary sleep encodes signals of decades-long health trajectories. By learning the "language of sleep," AI uncovers patterns invisible to the human eye—arousal indices, stage transitions, oxygen desaturations, and cross-signal dyssynchronies—that collectively forecast Parkinson's, cancers, heart failure, dementia, and mortality with remarkable accuracy. While not a definitive crystal ball, it offers an unprecedented proactive window into health. For individuals, it underscores the power of good sleep; for society, it points toward driven prevention that could ease healthcare burdens worldwide, including in China's vast population. Responsible innovation, rigorous validation, and ethical integration will determine whether this technology fulfills its promise of extending healthy lifespans.Stanford's groundbreaking discovery: Sleep one night, and AI can calculate your remaining lifespan.
2026年1月，斯坦福医学院的研究团队在《自然医学》杂志上发表了一项重磅研究，他们开发了一种名为SleepFM的AI基础模型，仅通过一晚的多导睡眠图（PSG）记录，就能预测超过100种健康状况的风险，包括全因死亡率。这一发现让“睡一觉就能预知寿命”从科幻走向现实，引发全球医学界和公众的广泛关注。传统上，睡眠监测主要用于诊断睡眠呼吸暂停等即时问题，而SleepFM则将睡眠数据转化为长期健康预测的“水晶球”，揭示了睡眠作为人体“健康档案”的深层价值。
In January 2026, a research team from Stanford Medicine published a landmark study in Nature Medicine. They developed an AI foundation model called SleepFM that, using just one night's polysomnography (PSG) recordings, can predict the risk of over 100 health conditions, including all-cause mortality. This discovery transforms the idea of "predicting lifespan from one night's sleep" from science fiction into reality, sparking widespread attention in the global medical community and among the public. Traditionally, sleep monitoring was mainly used to diagnose immediate issues like sleep apnea, but SleepFM turns sleep data into a "crystal ball" for long-term health predictions, revealing sleep's profound value as a "health archive" for the human body.
SleepFM模型的训练数据规模惊人，涵盖了超过6.5万名患者的近60万小时睡眠记录，这些数据来自斯坦福睡眠医学中心，并与长达25年的电子健康记录相匹配。AI通过自监督学习“掌握了睡眠的语言”，不仅能准确进行睡眠分期和呼吸暂停检测，还能在无标签数据上实现高效迁移学习。研究显示，该模型对130多种疾病的预测C指数（一致性指数）达到0.75以上，其中全因死亡率的C指数为0.84，痴呆为0.85，帕金森病高达0.89，心肌梗死0.81，前列腺癌0.89，乳腺癌0.87。这些指标远超基于人口统计学或传统睡眠指标的基准模型，证明睡眠中隐藏着丰富的纵向健康信号。
SleepFM was trained on an enormous dataset covering nearly 600,000 hours of sleep recordings from over 65,000 patients at Stanford's sleep centers, matched with up to 25 years of electronic health records. The AI "learned the language of sleep" through self-supervised learning, enabling accurate sleep staging, apnea detection, and efficient transfer learning on unlabeled data. The study showed that the model achieved C-indexes of at least 0.75 for predicting over 130 conditions, with all-cause mortality at 0.84, dementia at 0.85, Parkinson's disease at 0.89, myocardial infarction at 0.81, prostate cancer at 0.89, and breast cancer at 0.87. These metrics significantly outperform baselines based on demographics or traditional sleep metrics, proving that sleep harbors rich longitudinal health signals.
这项研究的创新之处在于SleepFM是一个多模态基础模型，它同时处理脑电图（EEG）、心电图（ECG）、呼吸信号、眼动、腿动等多种生理数据，而非依赖单一指标。传统睡眠分析往往只关注呼吸暂停指数或睡眠效率，但SleepFM能捕捉到跨系统的不匹配信号，例如大脑处于睡眠状态而心脏却表现出“清醒”迹象，这种“脱节”往往是未来心血管或神经退行性疾病的早期预警。研究者詹姆斯·邹（James Zou）教授表示：“睡眠数据就像一种被低估的‘生物语言’，AI帮助我们破译了它，从而提前多年发现疾病风险。”
The innovation of this study lies in SleepFM being a multimodal foundation model that simultaneously processes various physiological data such as electroencephalogram (EEG), electrocardiogram (ECG), respiratory signals, eye movements, and leg movements, rather than relying on single metrics. Traditional sleep analysis often focuses only on apnea-hypopnea index or sleep efficiency, but SleepFM captures cross-system mismatch signals—for instance, the brain appearing asleep while the heart shows "awake" patterns. Such "disconnections" often serve as early warnings for future cardiovascular or neurodegenerative diseases. Professor James Zou noted: "Sleep data is like an underestimated 'biological language,' and AI helps us decipher it to detect disease risks years in advance."
Beyond prediction accuracy, SleepFM demonstrates strong generalization. It performed competitively with specialized models like U-Sleep and YASA on standard tasks (mean F1 scores of 0.70–0.78 for sleep staging) and showed robust transfer learning on external datasets such as the Sleep Heart Health Study. This suggests the model isn't overfitting to Stanford-specific data but truly learning universal patterns in human sleep physiology. For mortality prediction, it achieved an AUROC of 0.85, surpassing demographic baselines (0.78) and end-to-end PSG models, highlighting how subtle factors like arousal burden, REM sleep reduction, hypoxemia, and sleep fragmentation integrate into powerful prognostic signals.
睡眠与寿命的关联早已被科学研究证实。早在2022年就有斯坦福相关研究显示，通过深度学习从睡眠研究中估计的“睡眠年龄”误差每增加10年，全因死亡风险上升29%，心血管死亡风险上升40%。对于60岁人群，睡眠年龄高估或低估10年，可能对应寿命减少约8.7年。SleepFM在此基础上实现了质的飞跃，它不只是估计生物年龄，还直接关联具体疾病风险和剩余寿命概率。研究指出，睡眠碎片化、低睡眠效率和间歇性低氧血症等特征，是独立于已知风险因素的死亡预测因子。
The link between sleep and lifespan has long been confirmed by scientific research. As early as 2022, related Stanford studies showed that every 10-year increase in error from deep learning-estimated "sleep age" raised all-cause mortality risk by 29% and cardiovascular mortality by 40%. For a 60-year-old, a 10-year over- or under-estimation in sleep age could correspond to an approximately 8.7-year reduction in life expectancy. SleepFM builds on this with a qualitative leap: it not only estimates biological age but directly links to specific disease risks and probabilities of remaining lifespan. The study notes that features like sleep fragmentation, low sleep efficiency, and intermittent hypoxemia are independent predictors of mortality beyond known risk factors.
在中国，睡眠健康问题同样突出。根据国家卫健委数据，成年居民睡眠不足比例较高，城市白领群体中睡眠障碍发生率可达30%以上。随着老龄化加剧和慢性病负担增加，利用AI分析睡眠数据预测寿命和疾病风险，具有巨大的公共卫生意义。国内多家医院和科技企业已开始探索类似多导睡眠监测与AI结合的应用，例如将可穿戴设备数据接入大模型，实现日常睡眠健康管理。如果SleepFM的技术能本土化落地，或许能帮助数亿人提前干预，减少医疗资源浪费，推动“健康中国”战略。
In China, sleep health issues are equally prominent. According to data from the National Health Commission, a high proportion of adults experience insufficient sleep, with sleep disorder rates among urban white-collar workers reaching over 30%. With accelerating aging and increasing chronic disease burdens, using AI to analyze sleep data for lifespan and disease risk prediction holds enormous public health significance. Several domestic hospitals and tech companies have begun exploring applications combining polysomnography with AI, such as integrating wearable device data into large models for daily sleep health management. If SleepFM technology can be localized, it could help hundreds of millions of people intervene early, reduce medical resource waste, and advance the "Healthy China" strategy.
The model's ability to predict specific conditions is particularly impressive. For neurological disorders, it excels at myoneural disorders, developmental delays, and speech/language issues. In circulatory diseases, it identifies atherosclerosis and acute pulmonary heart disease with high precision. Endocrine and metabolic predictions include diabetes complications and diabetic retinopathy, while respiratory predictions cover insufficiency and failure. Pregnancy complications and mental health disorders also show strong signals, suggesting sleep as a window into systemic health across life stages. Importantly, no single signal drives predictions; instead, the AI integrates multifactorial patterns, making it more robust than rule-based clinical scoring.
这项发现的潜在应用场景广阔。在临床上，睡眠实验室的常规检查可直接输出个性化风险报告，帮助医生制定预防方案。例如，高死亡风险患者可被优先推荐生活方式干预、药物治疗或进一步筛查。在保险和企业健康管理领域，AI睡眠预测能支持精准定价和员工福利优化。当然，隐私保护和伦理审查是前提，数据必须匿名化处理，且预测结果仅作为参考而非诊断。未来，研究团队计划将模型扩展到可穿戴设备数据，实现“在家睡一觉就能知寿命”的便捷模式。
The potential applications of this discovery are vast. Clinically, routine sleep lab tests could directly output personalized risk reports to help doctors devise prevention plans. For instance, patients with high mortality risk could be prioritized for lifestyle interventions, medications, or further screenings. In insurance and corporate wellness management, AI sleep predictions could support precise pricing and employee benefit optimization. Of course, privacy protection and ethical review are prerequisites—data must be anonymized, and predictions serve only as references, not diagnoses. In the future, the team plans to extend the model to wearable device data, enabling the convenient scenario of "sleep at home one night and know your lifespan."
However, challenges remain. While SleepFM outperforms baselines, real-world deployment requires validation across diverse populations, including different ethnicities, ages, and comorbidities common in China and globally. Accuracy for certain conditions like chronic kidney disease or stroke (around 0.78-0.79 C-index) is good but not perfect, underscoring the need for multimodal integration with genetics, blood tests, and lifestyle data. Interpretability is another focus: researchers are using techniques like SHAP to explain which sleep patterns drive predictions, building clinician trust. Over-reliance on AI without human oversight could lead to anxiety or false reassurance, so balanced communication is essential.
睡眠质量的改善本身就能延长寿命。大量流行病学研究表明，每晚7-9小时高质量睡眠可降低心血管疾病、糖尿病和认知衰退风险。SleepFM的预测结果或许能成为强大动机，鼓励人们重视睡眠卫生：规律作息、避免咖啡因、保持卧室凉爽黑暗等简单措施。结合AI反馈的个性化建议，例如“你的REM睡眠不足，建议增加运动以提升深度睡眠”，将使预防医学从被动转向主动。在中国“银发经济”和数字健康浪潮下，这一技术或催生新型睡眠健康产业，包括智能床垫、AI睡眠教练App等。
Improving sleep quality itself can extend lifespan. Extensive epidemiological studies show that 7-9 hours of high-quality sleep nightly reduces risks of cardiovascular disease, diabetes, and cognitive decline. SleepFM's predictions may serve as a powerful motivator, encouraging people to prioritize sleep hygiene: regular schedules, avoiding caffeine, and keeping bedrooms cool and dark. Personalized AI feedback, such as "Your REM sleep is insufficient; suggest increasing exercise to boost deep sleep," shifts preventive medicine from passive to active. Amid China's "silver economy" and digital health wave, this technology could spawn new sleep health industries, including smart mattresses and AI sleep coach apps.
Looking ahead, SleepFM represents a broader trend: foundation models unlocking hidden value in routine medical data. Similar to how large language models revolutionized text, multimodal sleep models could transform polysomnography from a niche diagnostic tool into a scalable population health platform. In China, integration with "East Data West Computing" infrastructure and domestic large models could accelerate adoption, supporting national goals for chronic disease prevention and healthy aging. International collaboration will be key to refining the model across global datasets and addressing biases.
要充分发挥这一技术的潜力，需要多方协同。医疗机构应加大睡眠监测设备的普及和AI培训；科技企业可开发用户友好接口，让普通人通过家用设备获取初步洞见；监管部门则需制定数据安全和伦理指南，避免信息滥用。同时，公众教育不可或缺：睡眠不是“浪费时间”，而是投资未来的健康资本。研究还提醒我们，年轻时忽视睡眠，可能在中年或老年付出寿命代价。
To fully leverage this technology's potential requires multi-stakeholder collaboration. Medical institutions should expand普及 of sleep monitoring equipment and AI training; tech companies can develop user-friendly interfaces for preliminary insights via home devices; regulators must establish data security and ethical guidelines to prevent misuse. Public education is equally essential: sleep is not "wasted time" but an investment in future health capital. The research also reminds us that neglecting sleep in youth may exact a lifespan cost in middle or old age.
Ethical considerations loom large. Predicting mortality or serious diseases years ahead raises questions about psychological impact—how to deliver "bad news" from sleep data without causing undue stress? Should insurers access such predictions? How to ensure equity so that low-income or rural populations in China aren't left behind? Researchers emphasize that SleepFM is a risk stratification tool, not a fortune teller; lifestyle, genetics, and medical interventions can still alter trajectories. Transparent, compassionate communication frameworks will be crucial for responsible deployment.
这一斯坦福重磅发现，不仅是AI在医学领域的又一里程碑，更是对人类与睡眠关系的重新审视。睡上一觉，AI或许就能勾勒出剩余寿命的轮廓，但真正的智慧在于行动：珍惜每晚的休息，用科学守护健康。未来，随着技术迭代和数据积累，类似模型或能实现更精准的个性化寿命预测，甚至指导精准医学干预。让我们期待，AI与睡眠的结合，能帮助更多人活得更长、更健康。
This Stanford groundbreaking discovery is not only another milestone for AI in medicine but also a re-examination of humanity's relationship with sleep. Sleep one night, and AI might outline the contours of remaining lifespan, but true wisdom lies in action: cherish every night's rest and guard health with science. In the future, with technological iterations and data accumulation, similar models may achieve even more precise personalized lifespan predictions and guide precision medicine interventions. Let us look forward to AI combined with sleep helping more people live longer and healthier lives.
In summary, SleepFM illustrates how one night's seemingly ordinary sleep encodes signals of decades-long health trajectories. By learning the "language of sleep," AI uncovers patterns invisible to the human eye—arousal indices, stage transitions, oxygen desaturations, and cross-signal dyssynchronies—that collectively forecast Parkinson's, cancers, heart failure, dementia, and mortality with remarkable accuracy. While not a definitive crystal ball, it offers an unprecedented proactive window into health. For individuals, it underscores the power of good sleep; for society, it points toward driven prevention that could ease healthcare burdens worldwide, including in China's vast population. Responsible innovation, rigorous validation, and ethical integration will determine whether this technology fulfills its promise of extending healthy lifespans.Stanford's groundbreaking discovery: Sleep one night, and AI can calculate your remaining lifespan.
2026年1月，斯坦福医学院的研究团队在《自然医学》杂志上发表了一项重磅研究，他们开发了一种名为SleepFM的AI基础模型，仅通过一晚的多导睡眠图（PSG）记录，就能预测超过100种健康状况的风险，包括全因死亡率。这一发现让“睡一觉就能预知寿命”从科幻走向现实，引发全球医学界和公众的广泛关注。传统上，睡眠监测主要用于诊断睡眠呼吸暂停等即时问题，而SleepFM则将睡眠数据转化为长期健康预测的“水晶球”，揭示了睡眠作为人体“健康档案”的深层价值。
In January 2026, a research team from Stanford Medicine published a landmark study in Nature Medicine. They developed an AI foundation model called SleepFM that, using just one night's polysomnography (PSG) recordings, can predict the risk of over 100 health conditions, including all-cause mortality. This discovery transforms the idea of "predicting lifespan from one night's sleep" from science fiction into reality, sparking widespread attention in the global medical community and among the public. Traditionally, sleep monitoring was mainly used to diagnose immediate issues like sleep apnea, but SleepFM turns sleep data into a "crystal ball" for long-term health predictions, revealing sleep's profound value as a "health archive" for the human body.
SleepFM模型的训练数据规模惊人，涵盖了超过6.5万名患者的近60万小时睡眠记录，这些数据来自斯坦福睡眠医学中心，并与长达25年的电子健康记录相匹配。AI通过自监督学习“掌握了睡眠的语言”，不仅能准确进行睡眠分期和呼吸暂停检测，还能在无标签数据上实现高效迁移学习。研究显示，该模型对130多种疾病的预测C指数（一致性指数）达到0.75以上，其中全因死亡率的C指数为0.84，痴呆为0.85，帕金森病高达0.89，心肌梗死0.81，前列腺癌0.89，乳腺癌0.87。这些指标远超基于人口统计学或传统睡眠指标的基准模型，证明睡眠中隐藏着丰富的纵向健康信号。
SleepFM was trained on an enormous dataset covering nearly 600,000 hours of sleep recordings from over 65,000 patients at Stanford's sleep centers, matched with up to 25 years of electronic health records. The AI "learned the language of sleep" through self-supervised learning, enabling accurate sleep staging, apnea detection, and efficient transfer learning on unlabeled data. The study showed that the model achieved C-indexes of at least 0.75 for predicting over 130 conditions, with all-cause mortality at 0.84, dementia at 0.85, Parkinson's disease at 0.89, myocardial infarction at 0.81, prostate cancer at 0.89, and breast cancer at 0.87. These metrics significantly outperform baselines based on demographics or traditional sleep metrics, proving that sleep harbors rich longitudinal health signals.
这项研究的创新之处在于SleepFM是一个多模态基础模型，它同时处理脑电图（EEG）、心电图（ECG）、呼吸信号、眼动、腿动等多种生理数据，而非依赖单一指标。传统睡眠分析往往只关注呼吸暂停指数或睡眠效率，但SleepFM能捕捉z7.mkxky.cn|v8.mkxky.cn|hx.mkxky.cn|fv.mkxky.cn|a8.mkxky.cn|ay.mkxky.cn|qk.mkxky.cn|v4.mkxky.cn|gf.mkxky.cn|lj.mkxky.cn|tj.mkxky.cn|4l.mkxky.cn|jv.mkxky.cn|sd.mkxky.cn|sl.mkxky.cn|aq.mkxky.cn|1e.mkxky.cn|ai.mkxky.cn|www.mkxky.cn|mkxky.cn到跨系统的不匹配信号，例如大脑处于睡眠状态而心脏却表现出“清醒”迹象，这种“脱节”往往是未来心血管或神经退行性疾病的早期预警。研究者詹姆斯·邹（James Zou）教授表示：“睡眠数据就像一种被低估的‘生物语言’，AI帮助我们破译了它，从而提前多年发现疾病风险。”
The innovation of this study lies in SleepFM being a multimodal foundation model that simultaneously processes various physiological data such as electroencephalogram (EEG), electrocardiogram (ECG), respiratory signals, eye movements, and leg movements, rather than relying on single metrics. Traditional sleep analysis often focuses only on apnea-hypopnea index or sleep efficiency, but SleepFM captures cross-system mismatch signals—for instance, the brain appearing asleep while the heart shows "awake" patterns. Such "disconnections" often serve as early warnings for future cardiovascular or neurodegenerative diseases. Professor James Zou noted: "Sleep data is like an underestimated 'biological language,' and AI helps us decipher it to detect disease risks years in advance."
Beyond prediction accuracy, SleepFM demonstrates strong generalization. It performed competitively with specialized models like U-Sleep and YASA on standard tasks (mean F1 scores of 0.70–0.78 for sleep staging) and showed robust transfer learning on external datasets such as the Sleep Heart Health Study. This suggests the model isn't overfitting to Stanford-specific data but truly learning universal patterns in human sleep physiology. For mortality prediction, it achieved an AUROC of 0.85, surpassing demographic baselines (0.78) and end-to-end PSG models, highlighting how subtle factors like arousal burden, REM sleep reduction, hypoxemia, and sleep fragmentation integrate into powerful prognostic signals.
睡眠与寿命的关联早已被科学研究证实。早在2022年就有斯坦福相关研究显示，通过深度学习从睡眠研究中估计的“睡眠年龄”误差每增加10年，全因死亡风险上升29%，心血管死亡风险上升40%。对于60岁人群，睡眠年龄高估或低估10年，可能对应寿命减少约8.7年。SleepFM在此基础上实现了质的飞跃，它不只是估计生物年龄，还直接关联具体疾病风险和剩余寿命概率。研究指出，睡眠碎片化、低睡眠效率和间歇性低氧血症等特征，是独立于已知风险因素的死亡预测因子。
The link between sleep and lifespan has long been confirmed by scientific research. As early as 2022, related Stanford studies showed that every 10-year increase in error from deep learning-estimated "sleep age" raised all-cause mortality risk by 29% and cardiovascular mortality by 40%. For a 60-year-old, a 10-year over- or under-estimation in sleep age could correspond to an approximately 8.7-year reduction in life expectancy. SleepFM builds on this with a qualitative leap: it not only estimates biological age but directly links to specific disease risks and probabilities of remaining lifespan. The study notes that features like sleep fragmentation, low sleep efficiency, and intermittent hypoxemia are independent predictors of mortality beyond known risk factors.
在中国，睡眠健康问题同样突出。根据国家卫健委数据，成年居民睡眠不足比例较高，城市白领群体中睡眠障碍发生率可达30%以上。随着老龄化加剧和慢性病负担增加，利用AI分析睡眠数据预测寿命和疾病风险，具有巨大的公共卫生意义。国内多家医院和科技企业已开始探索类似多导睡眠监测与AI结合的应用，例如将可穿戴设备数据接入大模型，实现日常睡眠健康管理。如果SleepFM的技术能本土化落地，或许能帮助数亿人提前干预，减少医疗资源浪费，推动“健康中国”战略。
In China, sleep health issues are equally prominent. According to data from the National Health Commission, a high proportion of adults experience insufficient sleep, with sleep disorder rates among urban white-collar workers reaching over 30%. With accelerating aging and increasing chronic disease burdens, using AI to analyze sleep data for lifespan and disease risk prediction holds enormous public health significance. Several domestic hospitals and tech companies have begun exploring applications combining polysomnography with AI, such as integrating wearable device data into large models for daily sleep health management. If SleepFM technology can be localized, it could help hundreds of millions of people intervene early, reduce medical resource waste, and advance the "Healthy China" strategy.
The model's ability to predict specific conditions is particularly impressive. For neurological disorders, it excels at myoneural disorders, developmental delays, and speech/language issues. In circulatory diseases, it identifies atherosclerosis and acute pulmonary heart disease with high precision. Endocrine and metabolic predictions include diabetes complications and diabetic retinopathy, while respiratory predictions cover insufficiency and failure. Pregnancy complications and mental health disorders also show strong signals, suggesting sleep as a window into systemic health across life stages. Importantly, no single signal drives predictions; instead, the AI integrates multifactorial patterns, making it more robust than rule-based clinical scoring.
这项发现的潜在应用场景广阔。在临床上，睡眠实验室的常规检查可直接输出个性化风险报告，帮助医生制定预防方案。例如，高死亡风险患者可被优先推荐生活方式干预、药物治疗或进一步筛查。在保险和企业健康管理领域，AI睡眠预测能支持精准定价和员工福利优化。当然，隐私保护和伦理审查是前提，数据必须匿名化处理，且预测结果仅作为参考而非诊断。未来，研究团队计划将模型扩展到可穿戴设备数据，实现“在家睡一觉就能知寿命”的便捷模式。
The potential applications of this discovery are vast. Clinically, routine sleep lab tests could directly output personalized risk reports to help doctors devise prevention plans. For instance, patients with high mortality risk could be prioritized for lifestyle interventions, medications, or further screenings. In insurance and corporate wellness management, AI sleep predictions could support precise pricing and employee benefit optimization. Of course, privacy protection and ethical review are prerequisites—data must be anonymized, and predictions serve only as references, not diagnoses. In the future, the team plans to extend the model to wearable device data, enabling the convenient scenario of "sleep at home one night and know your lifespan."
However, challenges remain. While SleepFM outperforms baselines, real-world deployment requires validation across diverse populations, including different ethnicities, ages, and comorbidities common in China and globally. Accuracy for certain conditions like chronic kidney disease or stroke (around 0.78-0.79 C-index) is good but not perfect, underscoring the need for multimodal integration with genetics, blood tests, and lifestyle data. Interpretability is another focus: researchers are using techniques like SHAP to explain which sleep patterns drive predictions, building clinician trust. Over-reliance on AI without human oversight could lead to anxiety or false reassurance, so balanced communication is essential.
睡眠质量的改善本身就能延长寿命。大量流行病学研究表明，每晚7-9小时高质量睡眠可降低心血管疾病、糖尿病和认知衰退风险。SleepFM的预测结果或许能成为强大动机，鼓励人们重视睡眠卫生：规律作息、避免咖啡因、保持卧室凉爽黑暗等简单措施。结合AI反馈的个性化建议，例如“你的REM睡眠不足，建议增加运动以提升深度睡眠”，将使预防医学从被动转向主动。在中国“银发经济”和数字健康浪潮下，这一技术或催生新型睡眠健康产业，包括智能床垫、AI睡眠教练App等。
Improving sleep quality itself can extend lifespan. Extensive epidemiological studies show that 7-9 hours of high-quality sleep nightly reduces risks of cardiovascular disease, diabetes, and cognitive decline. SleepFM's predictions may serve as a powerful motivator, encouraging people to prioritize sleep hygiene: regular schedules, avoiding caffeine, and keeping bedrooms cool and dark. Personalized AI feedback, such as "Your REM sleep is insufficient; suggest increasing exercise to boost deep sleep," shifts preventive medicine from passive to active. Amid China's "silver economy" and digital health wave, this technology could spawn new sleep health industries, including smart mattresses and AI sleep coach apps.
Looking ahead, SleepFM represents a broader trend: foundation models unlocking hidden value in routine medical data. Similar to how large language models revolutionized text, multimodal sleep models could transform polysomnography from a niche diagnostic tool into a scalable population health platform. In China, integration with "East Data West Computing" infrastructure and domestic large models could accelerate adoption, supporting national goals for chronic disease prevention and healthy aging. International collaboration will be key to refining the model across global datasets and addressing biases.
要充分发挥这一技术的潜力，需要多方协同。医疗机构应加大睡眠监测设备的普及和AI培训；科技企业可开发用户友好接口，让普通人通过家用设备获取初步洞见；监管部门则需制定数据安全和伦理指南，避免信息滥用。同时，公众教育不可或缺：睡眠不是“浪费时间”，而是投资未来的健康资本。研究还提醒我们，年轻时忽视睡眠，可能在中年或老年付出寿命代价。
To fully leverage this technology's potential requires multi-stakeholder collaboration. Medical institutions should expand普及 of sleep monitoring equipment and AI training; tech companies can develop user-friendly interfaces for preliminary insights via home devices; regulators must establish data security and ethical guidelines to prevent misuse. Public education is equally essential: sleep is not "wasted time" but an investment in future health capital. The research also reminds us that neglecting sleep in youth may exact a lifespan cost in middle or old age.
Ethical considerations loom large. Predicting mortality or serious diseases years ahead raises questions about psychological impact—how to deliver "bad news" from sleep data without causing undue stress? Should insurers access such predictions? How to ensure equity so that low-income or rural populations in China aren't left behind? Researchers emphasize that SleepFM is a risk stratification tool, not a fortune teller; lifestyle, genetics, and medical interventions can still alter trajectories. Transparent, compassionate communication frameworks will be crucial for responsible deployment.
这一斯坦福重磅发现，不仅是AI在医学领域的又一里程碑，更是对人类与睡眠关系的重新审视。睡上一觉，AI或许就能勾勒出剩余寿命的轮廓，但真正的智慧在于行动：珍惜每晚的休息，用科学守护健康。未来，随着技术迭代和数据积累，类似模型或能实现更精准的个性化寿命预测，甚至指导精准医学干预。让我们期待，AI与睡眠的结合，能帮助更多人活得更长、更健康。
This Stanford groundbreaking discovery is not only another milestone for AI in medicine but also a re-examination of humanity's relationship with sleep. Sleep one night, and AI might outline the contours of remaining lifespan, but true wisdom lies in action: cherish every night's rest and guard health with science. In the future, with technological iterations and data accumulation, similar models may achieve even more precise personalized lifespan predictions and guide precision medicine interventions. Let us look forward to AI combined with sleep helping more people live longer and healthier lives.
In summary, SleepFM illustrates how one night's seemingly ordinary sleep encodes signals of decades-long health trajectories. By learning the "language of sleep," AI uncovers patterns invisible to the human eye—arousal indices, stage transitions, oxygen desaturations, and cross-signal dyssynchronies—that collectively forecast Parkinson's, cancers, heart failure, dementia, and mortality with remarkable accuracy. While not a definitive crystal ball, it offers an unprecedented proactive window into health. For individuals, it underscores the power of good sleep; for society, it points toward driven prevention that could ease healthcare burdens worldwide, including in China's vast population. Responsible innovation, rigorous validation, and ethical integration will determine whether this technology fulfills its promise of extending healthy lifespans.Stanford's groundbreaking discovery: Sleep one night, and AI can calculate your remaining lifespan.
2026年1月，斯坦福医学院的研究团队在《自然医学》杂志上发表了一项重磅研究，他们开发了一种名为SleepFM的AI基础模型，仅通过一晚的多导睡眠图（PSG）记录，就能预测超过100种健康状况的风险，包括全因死亡率。这一发现让“睡一觉就能预知寿命”从科幻走向现实，引发全球医学界和公众的广泛关注。传统上，睡眠监测主要用于诊断睡眠呼吸暂停等即时问题，而SleepFM则将睡眠数据转化为长期健康预测的“水晶球”，揭示了睡眠作为人体“健康档案”的深层价值。
In January 2026, a research team from Stanford Medicine published a landmark study in Nature Medicine. They developed an AI foundation model called SleepFM that, using just one night's polysomnography (PSG) recordings, can predict the risk of over 100 health conditions, including all-cause mortality. This discovery transforms the idea of "predicting lifespan from one night's sleep" from science fiction into reality, sparking widespread attention in the global medical community and among the public. Traditionally, sleep monitoring was mainly used to diagnose immediate issues like sleep apnea, but SleepFM turns sleep data into a "crystal ball" for long-term health predictions, revealing sleep's profound value as a "health archive" for the human body.
SleepFM模型的训练数据规模惊人，涵盖了超过6.5万名患者的近60万小时睡眠记录，这些数据来自斯坦福睡眠医学中心，并与长达25年的电子健康记录相匹配。AI通过自监督学习“掌握了睡眠的语言”，不仅能准确进行睡眠分期和呼吸暂停检测，还能在无标签数据上实现高效迁移学习。研究显示，该模型对130多种疾病的预测C指数（一致性指数）达到0.75以上，其中全因死亡率的C指数为0.84，痴呆为0.85，帕金森病高达0.89，心肌梗死0.81，前列腺癌0.89，乳腺癌0.87。这些指标远超基于人口统计学或传统睡眠指标的基准模型，证明睡眠中隐藏着丰富的纵向健康信号。
SleepFM was trained on an enormous dataset covering nearly 600,000 hours of sleep recordings from over 65,000 patients at Stanford's sleep centers, matched with up to 25 years of electronic health records. The AI "learned the language of sleep" through self-supervised learning, enabling accurate sleep staging, apnea detection, and efficient transfer learning on unlabeled data. The study showed that the model achieved C-indexes of at least 0.75 for predicting over 130 conditions, with all-cause mortality at 0.84, dementia at 0.85, Parkinson's disease at 0.89, myocardial infarction at 0.81, prostate cancer at 0.89, and breast cancer at 0.87. These metrics significantly outperform baselines based on demographics or traditional sleep metrics, proving that sleep harbors rich longitudinal health signals.
这项研究的创新之处在于SleepFM是一个多模态基础模型，它同时处理脑电图（EEG）、心电图（ECG）、呼吸信号、眼动、腿动等多种生理数据，而非依赖单一指标。传统睡眠分析往往只关注呼吸暂停指数或睡眠效率，但SleepFM能捕捉到跨系统的不匹配信号，例如大脑处于睡眠状态而心脏却表现出“清醒”迹象，这种“脱节”往往是未来心血管或神经退行性疾病的早期预警。研究者詹姆斯·邹（James Zou）教授表示：“睡眠数据就像一种被低估的‘生物语言’，AI帮助我们破译了它，从而提前多年发现疾病风险。”
The innovation of this study lies in SleepFM being a multimodal foundation model that simultaneously processes various physiological data such as electroencephalogram (EEG), electrocardiogram (ECG), respiratory signals, eye movements, and leg movements, rather than relying on single metrics. Traditional sleep analysis often focuses only on apnea-hypopnea index or sleep efficiency, but SleepFM captures cross-system mismatch signals—for instance, the brain appearing asleep while the heart shows "awake" patterns. Such "disconnections" often serve as early warnings for future cardiovascular or neurodegenerative diseases. Professor James Zou noted: "Sleep data is like an underestimated 'biological language,' and AI helps us decipher it to detect disease risks years in advance."
Beyond prediction accuracy, SleepFM demonstrates strong generalization. It performed competitively with specialized models like U-Sleep and YASA on standard tasks (mean F1 scores of 0.70–0.78 for sleep staging) and showed robust transfer learning on external datasets such as the Sleep Heart Health Study. This suggests the model isn't overfitting to Stanford-specific data but truly learning universal patterns in human sleep physiology. For mortality prediction, it achieved an AUROC of 0.85, surpassing demographic baselines (0.78) and end-to-end PSG models, highlighting how subtle factors like arousal burden, REM sleep reduction, hypoxemia, and sleep fragmentation integrate into powerful prognostic signals.
睡眠与寿命的关联早已被科学研究证实。早在2022年就有斯坦福相关研究显示，通过深度学习从睡眠研究中估计的“睡眠年龄”误差每增加10年，全因死亡风险上升29%，心血管死亡风险上升40%。对于60岁人群，睡眠年龄高估或低估10年，可能对应寿命减少约8.7年。SleepFM在此基础上实现了质的飞跃，它不只是估计生物年龄，还直接关联具体疾病风险和剩余寿命概率。研究指出，睡眠碎片化、低睡眠效率和间歇性低氧血症等特征，是独立于已知风险因素的死亡预测因子。
The link between sleep and lifespan has long been confirmed by scientific research. As early as 2022, related Stanford studies showed that every 10-year increase in error from deep learning-estimated "sleep age" raised all-cause mortality risk by 29% and cardiovascular mortality by 40%. For a 60-year-old, a 10-year over- or under-estimation in sleep age could correspond to an approximately 8.7-year reduction in life expectancy. SleepFM builds on this with a qualitative leap: it not only estimates biological age but directly links to specific disease risks and probabilities of remaining lifespan. The study notes that features like sleep fragmentation, low sleep efficiency, and intermittent hypoxemia are independent predictors of mortality beyond known risk factors.
在中国，睡眠健康问题同样突出。根据国家卫健委数据，成年居民睡眠不足比例较高，城市白领群体中睡眠障碍发生率可达30%以上。随着老龄化加剧和慢性病负担增加，利用AI分析睡眠数据预测寿命和疾病风险，具有巨大的公共卫生意义。国内多家医院和科技企业已开始探索类似多导睡眠监测与AI结合的应用，例如将可穿戴设备数据接入大模型，实现日常睡眠健康管理。如果SleepFM的技术能本土化落地，或许能帮助数亿人提前干预，减少医疗资源浪费，推动“健康中国”战略。
In China, sleep health issues are equally prominent. According to data from the National Health Commission, a high proportion of adults experience insufficient sleep, with sleep disorder rates among urban white-collar workers reaching over 30%. With accelerating aging and increasing chronic disease burdens, using AI to analyze sleep data for lifespan and disease risk prediction holds enormous public health significance. Several domestic hospitals and tech companies have begun exploring applications combining polysomnography with AI, such as integrating wearable device data into large models for daily sleep health management. If SleepFM technology can be localized, it could help hundreds of millions of people intervene early, reduce medical resource waste, and advance the "Healthy China" strategy.
The model's ability to predict specific conditions is particularly impressive. For neurological disorders, it excels at myoneural disorders, developmental delays, and speech/language issues. In circulatory diseases, it identifies atherosclerosis and acute pulmonary heart disease with high precision. Endocrine and metabolic predictions include diabetes complications and diabetic retinopathy, while respiratory predictions cover insufficiency and failure. Pregnancy complications and mental health disorders also show strong signals, suggesting sleep as a window into systemic health across life stages. Importantly, no single signal drives predictions; instead, the AI integrates multifactorial patterns, making it more robust than rule-based clinical scoring.
这项发现的潜在应用场景广阔。在临床上，睡眠实验室的常规检查可直接输出个性化风险报告，帮助医生制定预防方案。例如，高死亡风险患者可被优先推荐生活方式干预、药物治疗或进一步筛查。在保险和企业健康管理领域，AI睡眠预测能支持精准定价和员工福利优化。当然，隐私保护和伦理审查是前提，数据必须匿名化处理，且预测结果仅作为参考而非诊断。未来，研究团队计划将模型扩展到可穿戴设备数据，实现“在家睡一觉就能知寿命”的便捷模式。
The potential applications of this discovery are vast. Clinically, routine sleep lab tests could directly output personalized risk reports to help doctors devise prevention plans. For instance, patients with high mortality risk could be prioritized for lifestyle interventions, medications, or further screenings. In insurance and corporate wellness management, AI sleep predictions could support precise pricing and employee benefit optimization. Of course, privacy protection and ethical review are prerequisites—data must be anonymized, and predictions serve only as references, not diagnoses. In the future, the team plans to extend the model to wearable device data, enabling the convenient scenario of "sleep at home one night and know your lifespan."
However, challenges remain. While SleepFM outperforms baselines, real-world deployment requires validation across diverse populations, including different ethnicities, ages, and comorbidities common in China and globally. Accuracy for certain conditions like chronic kidney disease or stroke (around 0.78-0.79 C-index) is good but not perfect, underscoring the need for multimodal integration with genetics, blood tests, and lifestyle data. Interpretability is another focus: researchers are using techniques like SHAP to explain which sleep patterns drive predictions, building clinician trust. Over-reliance on AI without human oversight could lead to anxiety or false reassurance, so balanced communication is essential.
睡眠质量的改善本身就能延长寿命。大量流行病学研究表明，每晚7-9小时高质量睡眠可降低心血管疾病、糖尿病和认知衰退风险。SleepFM的预测结果或许能成为强大动机，鼓励人们重视睡眠卫生：规律作息、避免咖啡因、保持卧室凉爽黑暗等简单措施。结合AI反馈的个性化建议，例如“你的REM睡眠不足，建议增加运动以提升深度睡眠”，将使预防医学从被动转向主动。在中国“银发经济”和数字健康浪潮下，这一技术或催生新型睡眠健康产业，包括智能床垫、AI睡眠教练App等。
Improving sleep quality itself can extend lifespan. Extensive epidemiological studies show that 7-9 hours of high-quality sleep nightly reduces risks of cardiovascular disease, diabetes, and cognitive decline. SleepFM's predictions may serve as a powerful motivator, encouraging people to prioritize sleep hygiene: regular schedules, avoiding caffeine, and keeping bedrooms cool and dark. Personalized AI feedback, such as "Your REM sleep is insufficient; suggest increasing exercise to boost deep sleep," shifts preventive medicine from passive to active. Amid China's "silver economy" and digital health wave, this technology could spawn new sleep health industries, including smart mattresses and AI sleep coach apps.
Looking ahead, SleepFM represents a broader trend: foundation models unlocking hidden value in routine medical data. Similar to how large language models revolutionized text, multimodal sleep models could transform polysomnography from a niche diagnostic tool into a scalable population health platform. In China, integration with "East Data West Computing" infrastructure and domestic large models could accelerate adoption, supporting national goals for chronic disease prevention and healthy aging. International collaboration will be key to refining the model across global datasets and addressing biases.
要充分发挥这一技术的潜力，需要多方协同。医疗机构应加大睡眠监测设备的普及和AI培训；科技企业可开发用户友好接口，让普通人通过家用设备获取初步洞见；监管部门则需制定数据安全和伦理指南，避免信息滥用。同时，公众教育不可或缺：睡眠不是“浪费时间”，而是投资未来的健康资本。研究还提醒我们，年轻时忽视睡眠，可能在中年或老年付出寿命代价。
To fully leverage this technology's potential requires multi-stakeholder collaboration. Medical institutions should expand普及 of sleep monitoring equipment and AI training; tech companies can develop user-friendly interfaces for preliminary insights via home devices; regulators must establish data security and ethical guidelines to prevent misuse. Public education is equally essential: sleep is not "wasted time" but an investment in future health capital. The research also reminds us that neglecting sleep in youth may exact a lifespan cost in middle or old age.
Ethical considerations loom large. Predicting mortality or serious diseases years ahead raises questions about psychological impact—how to deliver "bad news" from sleep data without causing undue stress? Should insurers access such predictions? How to ensure equity so that low-income or rural populations in China aren't left behind? Researchers emphasize that SleepFM is a risk stratification tool, not a fortune teller; lifestyle, genetics, and medical interventions can still alter trajectories. Transparent, compassionate communication frameworks will be crucial for responsible deployment.
这一斯坦福重磅发现，不仅是AI在医学领域的又一里程碑，更是对人类与睡眠关系的重新审视。睡上一觉，AI或许就能勾勒出剩余寿命的轮廓，但真正的智慧在于行动：珍惜每晚的休息，用科学守护健康。未来，随着技术迭代和数据积累，类似模型或能实现更精准的个性化寿命预测，甚至指导精准医学干预。让我们期待，AI与睡眠的结合，能帮助更多人活得更长、更健康。
This Stanford groundbreaking discovery is not only another milestone for AI in medicine but also a re-examination of humanity's relationship with sleep. Sleep one night, and AI might outline the contours of remaining lifespan, but true wisdom lies in action: cherish every night's rest and guard health with science. In the future, with technological iterations and data accumulation, similar models may achieve even more precise personalized lifespan predictions and guide precision medicine interventions. Let us look forward to AI combined with sleep helping more people live longer and healthier lives.
In summary, SleepFM illustrates how one night's seemingly ordinary sleep encodes signals of decades-long health trajectories. By learning the "language of sleep," AI uncovers patterns invisible to the human eye—arousal indices, stage transitions, oxygen desaturations, and cross-signal dyssynchronies—that collectively forecast Parkinson's, cancers, heart failure, dementia, and mortality with remarkable accuracy. While not a definitive crystal ball, it offers an unprecedented proactive window into health. For individuals, it underscores the power of good sleep; for society, it points toward driven prevention that could ease healthcare burdens worldwide, including in China's vast population. Responsible innovation, rigorous validation, and ethical integration will determine whether this technology fulfills its promise of extending healthy lifespans.Stanford's groundbreaking discovery: Sleep one night, and AI can calculate your remaining lifespan.
2026年1月，斯坦福医学院的研究团队在《自然医学》杂志上发表了一项重磅研究，他们开发了一种名为SleepFM的AI基础模型，仅通过一晚的多导睡眠图（PSG）记录，就能预测超过100种健康状况的风险，包括全因死亡率。这一发现让“睡一觉就能预知寿命”从科幻走向现实，引发全球医学界和公众的广泛关注。传统上，睡眠监测主要用于诊断睡眠呼吸暂停等即时问题，而SleepFM则将睡眠数据转化为长期健康预测的“水晶球”，揭示了睡眠作为人体“健康档案”的深层价值。
In January 2026, a research team from Stanford Medicine published a landmark study in Nature Medicine. They developed an AI foundation model called SleepFM that, using just one night's polysomnography (PSG) recordings, can predict the risk of over 100 health conditions, including all-cause mortality. This discovery transforms the idea of "predicting lifespan from one night's sleep" from science fiction into reality, sparking widespread attention in the global medical community and among the public. Traditionally, sleep monitoring was mainly used to diagnose immediate issues like sleep apnea, but SleepFM turns sleep data into a "crystal ball" for long-term health predictions, revealing sleep's profound value as a "health archive" for the human body.
SleepFM模型的训练数据规模惊人，涵盖了超过6.5万名患者的近60万小时睡眠记录，这些数据来自斯坦福睡眠医学中心，并与长达25年的电子健康记录相匹配。AI通过自监督学习“掌握了睡眠的语言”，不仅能准确进行睡眠分期和呼吸暂停检测，还能在无标签数据上实现高效迁移学习。研究显示，该模型对130多种疾病的预测C指数（一致性指数）达到0.75以上，其中全因死亡率的C指数为0.84，痴呆为0.85，帕金森病高达0.89，心肌梗死0.81，前列腺癌0.89，乳腺癌0.87。这些指标远超基于人口统计学或传统睡眠指标的基准模型，证明睡眠中隐藏着丰富的纵向健康信号。
SleepFM was trained on an enormous dataset covering nearly 600,000 hours of sleep recordings from over 65,000 patients at Stanford's sleep centers, matched with up to 25 years of electronic health records. The AI "learned the language of sleep" through self-supervised learning, enabling accurate sleep staging, apnea detection, and efficient transfer learning on unlabeled data. The study showed that the model achieved C-indexes of at least 0.75 for predicting over 130 conditions, with all-cause mortality at 0.84, dementia at 0.85, Parkinson's disease at 0.89, myocardial infarction at 0.81, prostate cancer at 0.89, and breast cancer at 0.87. These metrics significantly outperform baselines based on demographics or traditional sleep metrics, proving that sleep harbors rich longitudinal health signals.
这项研究的创新之处在于SleepFM是一个多模态基础模型，它同时处理脑电图（EEG）、心电图（ECG）、呼吸信号、眼动、腿动等多种生理数据，而非依赖单一指标。传统睡眠分析往往只关注呼吸暂停指数或睡眠效率，但SleepFM能捕捉到跨系统的不匹配信号，例如大脑处于睡眠状态而心脏却表现出“清醒”迹象，这种“脱节”往往是未来心血管或神经退行性疾病的早期预警。研究者詹姆斯·邹（James Zou）教授表示：“睡眠数据就像一种被低估的‘生物语言’，AI帮助我们破译了它，从而提前多年发现疾病风险。”
The innovation of this study lies in SleepFM being a multimodal foundation model that simultaneously processes various physiological data such as electroencephalogram (EEG), electrocardiogram (ECG), respiratory signals, eye movements, and leg movements, rather than relying on single metrics. Traditional sleep analysis often focuses only on apnea-hypopnea index or sleep efficiency, but SleepFM captures cross-system mismatch signals—for instance, the brain appearing asleep while the heart shows "awake" patterns. Such "disconnections" often serve as early warnings for future cardiovascular or neurodegenerative diseases. Professor James Zou noted: "Sleep data is like an underestimated 'biological language,' and AI helps us decipher it to detect disease risks years in advance."
Beyond prediction accuracy, SleepFM demonstrates strong generalization. It performed competitively with specialized models like U-Sleep and YASA on standard tasks (mean F1 scores of 0.70–0.78 for sleep staging) and showed robust transfer learning on external datasets such as the Sleep Heart Health Study. This suggests the model isn't overfitting to Stanford-specific data but truly learning universal patterns in human sleep physiology. For mortality prediction, it achieved an AUROC of 0.85, surpassing demographic baselines (0.78) and end-to-end PSG models, highlighting how subtle factors like arousal burden, REM sleep reduction, hypoxemia, and sleep fragmentation integrate into powerful prognostic signals.
睡眠与寿命的关联早已被科学研究证实。早在2022年就有斯坦福相关研究显示，通过深度学习从睡眠研究中估计的“睡眠年龄”误差每增加10年，全因死亡风险上升29%，心血管死亡风险上升40%。对于60岁人群，睡眠年龄高估或低估10年，可能对应寿命减少约8.7年。SleepFM在此基础上实现了质的飞跃，它不只是估计生物年龄，还直接关联具体疾病风险和剩余寿命概率。研究指出，睡眠碎片化、低睡眠效率和间歇性低氧血症等特征，是独立于已知风险因素的死亡预测因子。
The link between sleep and lifespan has long been confirmed by scientific research. As early as 2022, related Stanford studies showed that every 10-year increase in error from deep learning-estimated "sleep age" raised all-cause mortality risk by 29% and cardiovascular mortality by 40%. For a 60-year-old, a 10-year over- or under-estimation in sleep age could correspond to an approximately 8.7-year reduction in life expectancy. SleepFM builds on this with a qualitative leap: it not only estimates biological age but directly links to specific disease risks and probabilities of remaining lifespan. The study notes that features like sleep fragmentation, low sleep efficiency, and intermittent hypoxemia are independent predictors of mortality beyond known risk factors.
在中国，睡眠健康问题同样突出。根据国家卫健委数据，成年居民睡眠不足比例较高，城市白领群体中睡眠障碍发生率可达30%以上。随着老龄化加剧和慢性病负担增加，利用AI分析睡眠数据预测寿命和疾病风险，具有巨大的公共卫生意义。国内多家医院和科技企业已开始探索类似多导睡眠监测与AI结合的应用，例如将可穿戴设备数据接入大模型，实现日常睡眠健康管理。如果SleepFM的技术能本土化落地，或许能帮助数亿人提前干预，减少医疗资源浪费，推动“健康中国”战略。
In China, sleep health issues are equally prominent. According to data from the National Health Commission, a high proportion of adults experience insufficient sleep, with sleep disorder rates among urban white-collar workers reaching over 30%. With accelerating aging and increasing chronic disease burdens, using AI to analyze sleep data for lifespan and disease risk prediction holds enormous public health significance. Several domestic hospitals and tech companies have begun exploring applications combining polysomnography with AI, such as integrating wearable device data into large models for daily sleep health management. If SleepFM technology can be localized, it could help hundreds of millions of people intervene early, reduce medical resource waste, and advance the "Healthy China" strategy.
The model's ability to predict specific conditions is particularly impressive. For neurological disorders, it excels at myoneural disorders, developmental delays, and speech/language issues. In circulatory diseases, it identifies atherosclerosis and acute pulmonary heart disease with high precision. Endocrine and metabolic predictions include diabetes complications and diabetic retinopathy, while respiratory predictions cover insufficiency and failure. Pregnancy complications and mental health disorders also show strong signals, suggesting sleep as a window into systemic health across life stages. Importantly, no single signal drives predictions; instead, the AI integrates multifactorial patterns, making it more robust than rule-based clinical scoring.
这项发现的潜在应用场景广阔。在临床上，睡眠实验室的常规检查可直接输出个性化风险报告，帮助医生制定预防方案。例如，高死亡风险患者可被优先推荐生活方式干预、药物治疗或进一步筛查。在保险和企业健康管理领域，AI睡眠预测能支持精准定价和员工福利优化。当然，隐私保护和伦理审查是前提，数据必须匿名化处理，且预测结果仅作为参考而非诊断。未来，研究团队计划将模型扩展到可穿戴设备数据，实现“在家睡一觉就能知寿命”的便捷模式。
The potential applications of this discovery are vast. Clinically, routine sleep lab tests could directly output personalized risk reports to help doctors devise prevention plans. For instance, patients with high mortality risk could be prioritized for lifestyle interventions, medications, or further screenings. In insurance and corporate wellness management, AI sleep predictions could support precise pricing and employee benefit optimization. Of course, privacy protection and ethical review are prerequisites—data must be anonymized, and predictions serve only as references, not diagnoses. In the future, the team plans to extend the model to wearable device data, enabling the convenient scenario of "sleep at home one night and know your lifespan."
However, challenges remain. While SleepFM outperforms baselines, real-world deployment requires validation across diverse populations, including different ethnicities, ages, and comorbidities common in China and globally. Accuracy for certain conditions like chronic kidney disease or stroke (around 0.78-0.79 C-index) is good but not perfect, underscoring the need for multimodal integration with genetics, blood tests, and lifestyle data. Interpretability is another focus: researchers are using techniques like SHAP to explain which sleep patterns drive predictions, building clinician trust. Over-reliance on AI without human oversight could lead to anxiety or false reassurance, so balanced communication is essential.
睡眠质量的改善本身就能延长寿命。大量流行病学研究表明，每晚7-9小时高质量睡眠可降低心血管疾病、糖尿病和认知衰退风险。SleepFM的预测结果或许能成为强大动机，鼓励人们重视睡眠卫生：规律作息、避免咖啡因、保持卧室凉爽黑暗等简单措施。结合AI反馈的个性化建议，例如“你的REM睡眠不足，建议增加运动以提升深度睡眠”，将使预防医学从被动转向主动。在中国“银发经济”和数字健康浪潮下，这一技术或催生新型睡眠健康产业，包括智能床垫、AI睡眠教练App等。
Improving sleep quality itself can extend lifespan. Extensive epidemiological studies show that 7-9 hours of high-quality sleep nightly reduces risks of cardiovascular disease, diabetes, and cognitive decline. SleepFM's predictions may serve as a powerful motivator, encouraging people to prioritize sleep hygiene: regular schedules, avoiding caffeine, and keeping bedrooms cool and dark. Personalized AI feedback, such as "Your REM sleep is insufficient; suggest increasing exercise to boost deep sleep," shifts preventive medicine from passive to active. Amid China's "silver economy" and digital health wave, this technology could spawn new sleep health industries, including smart mattresses and AI sleep coach apps.
Looking ahead, SleepFM represents a broader trend: foundation models unlocking hidden value in routine medical data. Similar to how large language models revolutionized text, multimodal sleep models could transform polysomnography from a niche diagnostic tool into a scalable population health platform. In China, integration with "East Data West Computing" infrastructure and domestic large models could accelerate adoption, supporting national goals for chronic disease prevention and healthy aging. International collaboration will be key to refining the model across global datasets and addressing biases.
要充分发挥这一技术的潜力，需要多方协同。医疗机构应加大睡眠监测设备的普及和AI培训；科技企业可开发用户友好接口，让普通人通过家用设备获取初步洞见；监管部门则需制定数据安全和伦理指南，避免信息滥用。同时，公众教育不可或缺：睡眠不是“浪费时间”，而是投资未来的健康资本。研究还提醒我们，年轻时忽视睡眠，可能在中年或老年付出寿命代价。
To fully leverage this technology's potential requires multi-stakeholder collaboration. Medical institutions should expand普及 of sleep monitoring equipment and AI training; tech companies can develop user-friendly interfaces for preliminary insights via home devices; regulators must establish data security and ethical guidelines to prevent misuse. Public education is equally essential: sleep is not "wasted time" but an investment in future health capital. The research also reminds us that neglecting sleep in youth may exact a lifespan cost in middle or old age.
Ethical considerations loom large. Predicting mortality or serious diseases years ahead raises questions about psychological impact—how to deliver "bad news" from sleep data without causing undue stress? Should insurers access such predictions? How to ensure equity so that low-income or rural populations in China aren't left behind? Researchers emphasize that SleepFM is a risk stratification tool, not a fortune teller; lifestyle, genetics, and medical interventions can still alter trajectories. Transparent, compassionate communication frameworks will be crucial for responsible deployment.
这一斯坦福重磅发现，不仅是AI在医学领域的又一里程碑，更是对人类与睡眠关系的重新审视。睡上一觉，AI或许就能勾勒出剩余寿命的轮廓，但真正的智慧在于行动：珍惜每晚的休息，用科学守护健康。未来，随着技术迭代和数据积累，类似模型或能实现更精准的个性化寿命预测，甚至指导精准医学干预。让我们期待，AI与睡眠的结合，能帮助更多人活得更长、更健康。
This Stanford groundbreaking discovery is not only another milestone for AI in medicine but also a re-examination of humanity's relationship with sleep. Sleep one night, and AI might outline the contours of remaining lifespan, but true wisdom lies in action: cherish every night's rest and guard health with science. In the future, with technological iterations and data accumulation, similar models may achieve even more precise personalized lifespan predictions and guide precision medicine interventions. Let us look forward to AI combined with sleep helping more people live longer and healthier lives.
In summary, SleepFM illustrates how one night's seemingly ordinary sleep encodes signals of decades-long health trajectories. By learning the "language of sleep," AI uncovers patterns invisible to the human eye—arousal indices, stage transitions, oxygen desaturations, and cross-signal dyssynchronies—that collectively forecast Parkinson's, cancers, heart failure, dementia, and mortality with remarkable accuracy. While not a definitive crystal ball, it offers an unprecedented proactive window into health. For individuals, it underscores the power of good sleep; for society, it points toward driven prevention that could ease healthcare burdens worldwide, including in China's vast population. Responsible innovation, rigorous validation, and ethical integration will determine whether this technology fulfills its promise of extending healthy lifespans.
仁信优配提示：文章来自网络，不代表本站观点。
