利用可穿戴设备实现心血管-肾脏-代谢(CKM)风险画像,以支持个性化指导

qimuai 发布于 阅读:33 一手编译

利用可穿戴设备实现心血管-肾脏-代谢(CKM)风险画像,以支持个性化指导

内容来源:https://research.google/blog/using-wearables-to-enable-cardiovascular-kidney-metabolic-ckm-risk-profiling-for-personalized-coaching/

内容总结:

谷歌研究团队近日发布一项处于研究阶段的机器学习框架,可将可穿戴设备采集的被动信号转化为可干预、可指导的“健康年龄”,并利用动作条件世界模型模拟个性化12周“假设性”习惯轨迹,同时扩展至常规血液生物标志物,以追踪心血管—肾脏—代谢(CKM)健康的早期变化。

研究指出,心脏病、肾脏疾病和2型糖尿病等代谢紊乱并非孤立存在,而是构成相互关联的心血管—肾脏—代谢综合征。CKM的进展往往在症状出现前数年便已悄然开始,受运动、睡眠和心血管活动等日常习惯影响。尽早捕捉这些细微变化,是逆转趋势的最佳机会。然而,智能手表虽能轻松记录昨日活动,但从被动追踪迈向有意义的行动却困难得多:日常习惯当下如何影响CKM?若每天多睡30分钟或增加快走,未来三个月健康状况能改善多少?

年龄常被视为回答这些问题的一个因素,但仅凭实际年龄并不足够,因为同年出生两人的生理健康可能差异巨大。研究人员已开发出基于血液检测或可穿戴传感器的“生物时钟”,用于估算身体衰老快于或慢于日历年龄,但多数现有时钟并非为日常指导而设计:有的仅训练用于估算出生年份,依赖难以改变的特征;有的则是“黑箱”,无法展示改变某一具体习惯将如何影响未来健康。

在两篇论文《通过可指导健康年龄和反事实可穿戴轨迹实现个性化健康指导》及配套研究《利用可穿戴设备和血液生物标志物的心血管—肾脏—代谢健康年龄》中,团队提出一个旨在打通“追踪健康”与“主动改善健康”闭环的研究框架。

为使健康指标可用于指导,所有输入均应可连续追踪并能通过日常选择施加影响。研究团队未训练模型猜测出生年份,而是将“可指导健康年龄”(CHA)模型建立在已发表的流行病学研究之上,这些研究将日常行为与心血管代谢疾病风险相关联。CHA整合消费级可穿戴设备被动测量的四项核心信号——静息心率、每日步数、睡眠时长和每周活跃区间分钟数——并在可用时纳入VO2max、睡眠规律性和胰岛素抵抗等可选指标。由于活跃人群往往静息心率更低、睡眠更好,团队对信号间的重叠进行数学调整,使每种习惯对健康年龄差的贡献各自独立。负值(如−3年)意味着生理状况看似比实际年龄年轻,正值则提示加速衰老。

该模型在三个独立队列中得到验证。

仅了解健康年龄只能说明当下状态。将其转化为个性化方案需回答“假设性”(反事实)问题:若改变特定习惯,未来12周身体将发生何种变化?哪种方案能切实坚持?

将新的步数或睡眠目标简单输入静态计算器,忽略了身体作为互联系统的适应过程。若连续数周增加运动并改善睡眠,静息心率和心率变异性也会逐步改善,产生下游额外健康收益。

为捕捉这些动态,团队开发了“传感器世界模型”(SensorWM),这是一个动作条件世界模型,基于约10万名同意参与的Fitbit用户、共228万人周的去标识化数据预训练。给定两周基线可穿戴数据,SensorWM可预测在不同每日步数、睡眠和心肺运动方案下,20项生理指标未来12周的演变,表现优于TimesFM等通用时间序列基础模型,同时保留与糖尿病、高血压和胰岛素抵抗相关的临床模式。

团队将SensorWM与可行性约束优化器及语言模型指导代理结合,实现三项功能。

研究还开展盲法头对头评估,由五名认证生活方式教练和五名委员会认证的初级保健及生活方式医学医师对49名参与者画像(涵盖五种行为类型,共490对评审)进行比较,并在WEAR-ME全队列(n=2,780)中进行自动化评估。

由于健康年龄模型采用模块化设计,可轻松纳入常规临床实验室检测。在配套研究中,团队将框架扩展至跨单个器官系统的CKM综合征分析,将连续可穿戴设备与偶尔的血液生物标志物结合,后者单独使用会遗漏日常生活方式模式。

早期CKM衰退极少引起症状,偶尔的临床抽血会遗漏日常模式。在819名WEAR-ME参与者中评估的多模态CKM健康年龄模型,将连续可穿戴信号(静息心率、步数、活跃区间分钟数和睡眠)与三项常规血液生物标志物结合:胆固醇与HDL比值(心血管)、eGFR(肾脏滤过)和HOMA-IR(代谢胰岛素抵抗)。

通过测量各指标偏离健康临床阈值的程度(归一化偏差评分),CKM模型揭示多器官衰老并非遵循单一模板:代谢与心血管压力常相关,但早期肾脏衰退常独立出现。此外,CKM健康年龄差与客观临床评分(改良生命必需8项 r=−0.58;代谢综合征严重程度 r=0.65)及参与者日常感受高度一致。

两项研究共同表明,消费级可穿戴设备与机器学习可帮助个人健康从“后视镜”转向“导航前方”。尽管评估基于回顾性研究队列离线进行,且这些模型为研究原型而非临床诊断工具,但它们为未来前瞻性试验奠定基础,以测试交互式“假设性”指导如何帮助人们建立持久习惯并改善长期健康。

本研究为谷歌研究与合作团队的联合工作。贡献者包括Zeinab Esmaeilpour、Simon A. Lee、Menglian Zhou、Marius Guerard、Sohrab Sidhu、Alexandru Solot、Hamid Palangi、Pramod Rudrapatna、Ming-Zher Poh、Max A. Xu、Daniel McDuff、Dimitrios Antos、Yuzhe Yang、Jeffrey N. Chiang(加州大学洛杉矶分校)、Anthony Z. Faranesh、Conor Heneghan、Lindsey Sunden、Mark Malhotra、Cathy Speed、Javier L. Prieto、Shwetak Patel、A. Ali Heydari和Ahmed A. Metwally。

中文翻译:

2026年11月11日
Zeinab Esmaeilpour 和 Ahmed A. Metwally,Google Research 研究科学家

我们推出了一套探索性的机器学习框架,可将被动采集的可穿戴设备信号转化为可干预、可指导的健康年龄,利用以行动为条件的世界模型模拟个性化的12周“假设”习惯轨迹,并进一步整合常规血液生物标志物,以追踪早期心血管-肾脏-代谢(CKM)健康状况。

心脏病、肾脏疾病和2型糖尿病等代谢紊乱并非孤立存在,它们共同构成一个相互关联的网络,即心血管-肾脏-代谢(CKM)综合征。CKM的进展往往在症状出现前数年便悄然开始,受运动、睡眠和心血管活动等日常习惯的影响。及早捕捉这些细微变化,是逆转趋势的最佳时机。然而,尽管智能手表可以轻松追踪我们昨天的活动,从被动追踪迈向有意义的行动却要困难得多:我的日常习惯此刻如何影响我的CKM?如果我多睡30分钟,或在日常中增加一次快走,未来三个月我的健康能改善多少?

年龄常被视为回答这些问题的一个因素,但仅凭实际年龄远远不够,因为同年出生的人生理健康状况可能天差地别。尽管研究人员已开发出基于血液检测或可穿戴传感器的“生物钟”,用以估算一个人的身体比其日历年龄衰老得更快还是更慢,但大多数现有生物钟并非为日常指导而设计。许多模型仅仅被训练来估算出生年份,依赖的是难以轻易改变的特征,或者充当黑箱,无法展示改变某一特定习惯将如何影响未来的健康。

在《基于可指导健康年龄与反事实可穿戴轨迹的个性化健康指导》及我们的配套研究《利用可穿戴设备与血液生物标志物评估心血管-肾脏-代谢健康年龄》中,我们推出了一套探索性框架,旨在打通追踪个人健康与主动改善健康之间的闭环。该框架提供以下内容:

要让健康指标真正有助于指导,每一项输入都应是可持续追踪、且能通过日常选择加以影响的。我们没有训练模型去猜测一个人的出生年份,而是将可指导健康年龄(CHA)模型建立在已发表的流行病学研究之上,这些研究将日常行为与心血管代谢疾病风险相关联。CHA整合了消费级可穿戴设备被动测量的四项核心信号——静息心率、每日步数、睡眠时长和每周活跃区间分钟数——以及在可获得时的可选指标,如最大摄氧量、睡眠作息规律性和胰岛素抵抗。由于活跃人群往往静息心率更低、睡眠更好,我们通过数学方法校正这些信号之间的重叠,使每种习惯对你健康年龄差的贡献各自独立——健康年龄差即你的CHA与实际年龄之间的差值。负值(例如−3岁)意味着个人的生理特征看起来比其实际年龄更年轻,而正值则提示加速衰老。

我们在三个独立队列中验证了该模型:

了解自己的健康年龄,能告诉你今天处于什么状态。要将这个数字转化为个性化方案,则需要回答一个“假设”(或反事实)问题:如果我改变特定习惯,未来12周我的身体会发生什么变化?哪种方案是我现实中能坚持的?

简单地将新的步数或睡眠目标输入静态计算器,忽略了身体作为一个互联系统如何适应变化。当你连续数周坚持多运动、睡得更好,你的静息心率和心率变异性也会逐渐改善,从而在下游产生额外的健康收益。

为捕捉这些动态变化,我们开发了传感器世界模型(SensorWM),这是一个以行动为条件的世界模型,基于约10万名知情同意的Fitbit用户、共计228万人周的去标识化数据进行预训练。给定两周的基线可穿戴数据,SensorWM可预测20项生理指标在不同每日步数、睡眠和心肺运动方案下未来12周的演变,表现优于TimesFM等通用时间序列基础模型,同时保留了与糖尿病、高血压和胰岛素抵抗相关的临床模式。

我们将SensorWM与一个受可行性约束的优化器和一个语言模型指导智能体相配合,实现三项功能:

将语言模型建立在世界模型模拟之上,真的比给完全相同的模型仅提供用户的可穿戴历史记录能产生更好的指导吗?

我们开展了一项盲法头对头研究,邀请五位认证生活方式教练和五位委员会认证的初级保健与生活方式医学医师,对涵盖五种行为画像的49份参与者档案进行评审(共计490对配对评审),同时在完整的WEAR-ME队列(n = 2,780)上进行了自动化评估。

由于我们的健康年龄模型采用模块化设计,可以轻松纳入可获得的常规临床实验室检测。在配套研究中,我们将框架扩展至跨单个器官系统刻画CKM综合征,将连续可穿戴数据与偶尔的血液生物标志物相结合——后者单独使用时无法捕捉日常生活方式模式。

早期CKM衰退很少引起症状,而偶尔的门诊抽血也无法反映日常生活方式模式。在819名WEAR-ME参与者中评估,我们的多模态CKM健康年龄模型将连续可穿戴信号(静息心率、步数、活跃区间分钟数和睡眠)与三项常规血液生物标志物相结合:胆固醇与高密度脂蛋白比值(心血管)、eGFR(肾脏滤过功能)和HOMA-IR(代谢性胰岛素抵抗)。

通过测量每项标志物偏离健康临床阈值的程度(归一化偏差评分),CKM模型揭示多器官衰老并不遵循单一模板:代谢和心血管压力往往相关,但早期肾脏衰退常常独立出现。此外,由此得出的CKM健康年龄差与客观临床评分(改良版生命必需8项 r = −0.58;代谢综合征严重程度 r = 0.65)以及参与者日常的健康感受均高度吻合。

这两项研究共同表明,消费级可穿戴设备与机器学习如何帮助个人健康从“看后视镜”转向“导航前路”。尽管我们的评估是基于回顾性研究队列的离线评估,且这些模型属于探索性研究原型而非临床诊断工具,但它们为未来前瞻性试验奠定了基础,以检验交互式“假设”指导如何帮助人们建立持久习惯、改善长期健康。

本文所述研究是Google Research与合作团队的联合工作。贡献者包括Zeinab Esmaeilpour、Simon A. Lee、Menglian Zhou、Marius Guerard、Sohrab Sidhu、Alexandru Solot、Hamid Palangi、Pramod Rudrapatna、Ming-Zher Poh、Max A. Xu、Daniel McDuff、Dimitrios Antos、Yuzhe Yang、Jeffrey N. Chiang(来自UCLA)、Anthony Z. Faranesh、Conor Heneghan、Lindsey Sunden、Mark Malhotra、Cathy Speed、Javier L. Prieto、Shwetak Patel、A. Ali Heydari 和 Ahmed A. Metwally。

英文来源:

November 11, 2026
Zeinab Esmaeilpour and Ahmed A. Metwally, Research Scientists, Google Research
We introduce an investigational machine learning framework that translates passive wearable signals into a modifiable coachable health age, uses an action-conditioned world model to simulate personalized 12-week "what-if" habit trajectories, and expands to routine blood biomarkers to track early cardiovascular-kidney-metabolic (CKM) health.
Heart disease, kidney conditions, and metabolic disorders like type 2 diabetes aren't isolated, they form an interconnected web known as cardiovascular-kidney-metabolic (CKM) syndrome. CKM progression often begins silently years before symptoms appear, shaped by everyday habits like movement, sleep, and cardiovascular activity. Catching these subtle shifts early offers the best chance to reverse the trend. Yet while smartwatches easily track what we did yesterday, moving from passive tracking to meaningful action is much harder: How do my daily habits affect my CKM right now? And if I sleep 30 minutes longer or add a brisk walk to my routine, how much will my health improve over the next three months?
Age is often seen as playing a part in answering these questions, but chronological age alone isn’t enough because two people born in the same year can have vastly different physiological health. While researchers have developed “biological clocks” from blood tests or wearable sensors to estimate whether someone's body is aging faster or slower than their calendar age, most existing clocks aren't built for everyday coaching. Many are trained simply to estimate the birth year, latching onto traits you can't easily change, or act as black boxes that can't show how modifying a specific habit will shift your future health.
In “Personalized health coaching via coachable health age and counterfactual wearable trajectories” and our companion study, “Cardiovascular-Kidney-Metabolic Health Age Using Wearables and Blood Biomarkers”, we introduce an investigational framework designed to close the loop between tracking a person’s health and actively improving it. The framework provides the following:
To make a health metric useful for coaching, every input should be something that can be tracked continuously and influenced through daily choices. Rather than training a model to guess a person’s birth year, we grounded our coachable health age (CHA) model in published epidemiological studies linking everyday behaviors to cardio-metabolic disease risk. CHA integrates four core signals passively measured by consumer wearables — resting heart rate, daily step count, sleep duration, and weekly Active Zone Minutes — alongside optional markers like VO2max, sleep schedule regularity, and insulin resistance, when available. Because active people often have lower resting heart rates and better sleep, we mathematically adjust for overlap across these signals so each habit contributes its distinct share to your health age delta, the difference between your CHA and chronological age. A negative delta (say, −3 years) means an individual's physiological profile appears younger than their chronological age, while a positive delta signals accelerated aging.
We validated this model across three independent cohorts:
Knowing your health age tells you where you stand today. Turning that number into a personalized plan requires answering a "what-if" (or counterfactual) question: What will happen to my body over the next 12 weeks if I change specific habits, and which plan can I realistically stick with?
Simply plugging new step or sleep targets into a static calculator misses how the body adapts as a connected system. When you consistently move more and sleep better over several weeks, your resting heart rate and heart rate variability gradually improve too, creating extra health benefits downstream.
To capture these dynamics, we developed Sensor World Model (SensorWM), an action-conditioned world model pre-trained on 2.28 million person-weeks of de-identified data from roughly 100,000 consenting Fitbit users. Given two weeks of baseline wearable data, SensorWM forecasts how 20 physiological markers will evolve over the next 12 weeks under different daily steps, sleep, and cardio routines, outperforming general-purpose time-series foundation models like TimesFM while preserving the clinical patterns associated with diabetes, hypertension, and insulin resistance.
We paired SensorWM with a feasibility-constrained optimizer and a language model coaching agent to do three things:
Does grounding a language model in world-model simulations actually produce better guidance than giving the exact same model a user's wearable history alone?
We ran a blinded, head-to-head study with five certified lifestyle coaches and five board-certified primary care and lifestyle medicine physicians across 49 participant profiles spanning five behavioral personas (490 paired reviews total), alongside automated evaluations across the full WEAR-ME cohort (n = 2,780).
Because our health age model is modular, it can easily incorporate routine clinical lab tests when available. In our companion study, we expanded the framework to profile CKM syndrome across individual organ systems, pairing continuous wearables with occasional blood biomarkers that alone miss day to day lifestyle patterns.
Early CKM decline rarely causes symptoms, and occasional clinic blood draws miss day-to-day lifestyle patterns. Evaluated in 819 WEAR-ME participants, our multimodal CKM health age model combines continuous wearable signals (resting heart rate, steps, active zone minutes, and sleep) with three routine blood biomarkers: cholesterol-to-HDL ratio (cardiovascular), eGFR (kidney filtration), and HOMA-IR (metabolic insulin resistance).
By measuring how far each marker departs from healthy clinical thresholds (a normalized deviance score), the CKM model revealed that multi-organ aging doesn't follow a single template: while metabolic and cardiovascular strain are often correlated, early kidney decline frequently appears independently. Moreover, the resulting CKM health age gap aligned closely with both objective clinical scores (modified Life's Essential 8 r = −0.58; metabolic syndrome severity r = 0.65) and how healthy participants felt day to day.
Together, these studies show how consumer wearables and machine learning can help shift personal health from looking in the rearview mirror to navigating the road ahead. While our evaluations were conducted offline using retrospective research cohorts, and these models are investigational research prototypes rather than clinical diagnostic tools, they lay the foundation for future prospective trials testing how interactive "what-if" coaching can help people build lasting habits and improve their long-term health.
The research described here is joint work across Google Research and partnering teams. Contributors include Zeinab Esmaeilpour, Simon A. Lee, Menglian Zhou, Marius Guerard, Sohrab Sidhu, Alexandru Solot, Hamid Palangi, Pramod Rudrapatna, Ming-Zher Poh, Max A. Xu, Daniel McDuff, Dimitrios Antos, Yuzhe Yang, Jeffrey N. Chiang (from UCLA), Anthony Z. Faranesh, Conor Heneghan, Lindsey Sunden, Mark Malhotra, Cathy Speed, Javier L. Prieto, Shwetak Patel, A. Ali Heydari, and Ahmed A. Metwally.

谷歌研究进展

文章目录


    扫描二维码,在手机上阅读