文献速递医学人工智能与信息学流行病学与生物统计真实世界研究与卫生经济

机器学习衍生的衰老指数用于心肾代谢综合征风险分层与死亡预测

A machine learning-derived aging index for risk stratification and mortality prediction in cardiovascular-kidney-metabolic syndrome: A retrospective cohort study.
PubMed 文献 · PLoS Med · 2026-10-06

该回顾性队列研究基于NHANES 2005-2018的6,896名成人,利用两阶段机器学习框架开发了心肾代谢综合征特异性衰老指数CKMAI。CKMAI在预测全因死亡、心血管死亡和高危CKM状态方面均优于PhenoAge、KDM和CMI,并识别出6种衰老代谢表型,其中衰弱老年组死亡风险最高(HR 11.74)。

要点

  • CKMAI在3、5、10年全因死亡预测中AUC分别为0.893、0.907、0.890,优于现有衰老指数。
  • CKMAI与PhenoAge存在超相加交互作用(RERI 15.06),提示联合评估可增强风险识别。
  • 抑郁部分介导CKMAI与结局的关联,中介比例分别为5.6%、8.2%和11.1%。

核心结论

  • CKMAI是CKM综合征特异性的有效衰老指数,在预测死亡和高危CKM状态方面优于通用生物学年龄指标。
  • CKMAI与PhenoAge的超相加交互及抑郁的部分中介作用为CKM异质性提供了新机制见解。

研究设计与样本

  • 回顾性队列研究,纳入NHANES 2005-2018的6,896名成人,权重代表美国成年人口。
  • 采用两阶段机器学习框架评估超过100个候选生存模型,通过帕累托前沿优化选择最优模型。
  • 预测性能评估使用时间依赖AUC、NRI、IDI、校准和决策曲线分析。
  • 使用限制性立方样条和两段式回归检查非线性关系与阈值效应。
  • 通过相对超额风险交互(RERI)和归因比例(AP)量化相加交互作用。

主要发现

  • 帕累托最优简单Cox模型平均C指数为0.893。
  • 全因死亡时间依赖AUC:3年0.893(95% CI [0.851, 0.925]),5年0.907(95% CI [0.884, 0.927]),10年0.890(95% CI [0.868, 0.910])。
  • 心血管死亡时间依赖AUC:3年0.904(95% CI [0.840, 0.943]),5年0.937(95% CI [0.892, 0.960]),10年0.937(95% CI [0.900, 0.959])。
  • 高危CKM状态时间依赖AUC为0.790(95% CI [0.765, 0.816])。
  • NRI和IDI在所有时间点均显著(P < 0.05),校准良好(Greenwood-Nam-D'Agostino检验P > 0.05),DCA证实净获益优越。

关键数值与推荐

  • 阈值分析识别拐点:CKMAI 61.587(全因死亡)、60.732(心血管死亡)、34.092(高危CKM状态),低于各阈值时风险增加更陡峭。
  • CKMAI与PhenoAge的超相加交互:RERI 15.06(95% CI [8.14, 38.44])。
  • 聚类识别6种衰老代谢表型,衰弱老年组死亡风险最高(HR 11.74,95% CI [7.85, 17.57])。
  • 抑郁部分中介比例:全因死亡5.6%(95% CI 2.3%-9.0%),心血管死亡8.2%(95% CI 2.8%-22.0%),高危CKM状态11.1%(95% CI 6.7%-24.0%)。

临床怎么用

  • CKMAI可整合临床实验室指标、营养评估、生活方式和社会健康决定因素,用于CKM综合征风险分层。
  • CKMAI优于PhenoAge、KDM和CMI,可作为高危CKM状态和死亡风险的实用预测工具。
  • 识别衰弱老年表型有助于针对性预防干预。

局限性

  • 缺乏地理独立且具有完整死亡随访的队列验证。
  • NHANES内部时间验证和中国医院队列(n=261)初步外部验证支持泛化性,但仍需进一步验证。

以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。

原文

BACKGROUND: The cardiovascular‑kidney‑metabolic (CKM) syndrome shows substantial heterogeneity in progression, yet existing biological age indices are not tailored to CKM pathways. We developed an aging index (CKMAI) using machine learning and evaluated its predictive performance for mortality and high‑risk CKM status. METHODS AND FINDINGS: We conducted a retrospective cohort study of 6,896 adults from the National Health and Nutrition Examination Survey (NHANES) 2005-2018, weighted for the US adult population. A two‑stage machine learning framework evaluated over 100 candidate survival models; the optimal model was selected via Pareto front optimization. Predictive performance was assessed using time‑dependent area under the curve (AUC), net reclassification improvement (NRI), integrated discrimination improvement (IDI), calibration, and decision curve analysis (DCA). Nonlinear relationships and threshold effects were examined with restricted cubic splines and two‑piecewise regression. Additive interactions were quantified by relative excess risk due to interaction (RERI) and attributable proportion (AP). Unsupervised clustering identified aging‑metabolic phenotypes. Weighted bootstrap mediation analysis assessed the mediating role of depression. The Pareto‑optimal simple Cox model achieved a mean C‑index of 0.893. CKMAI consistently outperformed PhenoAge, KDM, and the cardiometabolic index (CMI) across all outcomes. For all‑cause mortality, CKMAI's time‑dependent AUCs were 0.893 (95% CI [0.851, 0.925]) at 3 years, 0.907 (95% CI [0.884, 0.927]) at 5 years, and 0.890 (95% CI [0.868, 0.910]) at 10 years; for cardiovascular mortality, 0.904 (95% CI [0.840, 0.943]), 0.937 (95% CI [0.892, 0.960]), and 0.937 (95% CI [0.900, 0.959]); for high‑risk CKM status, 0.790 (95% CI [0.765, 0.816]). NRI and IDI were significant at all time points (P < 0.05 for all), calibration was good (Greenwood‑Nam‑D'Agostino test P > 0.05), and DCA confirmed superior net benefit. Nonlinear associations were found for all three outcomes. Threshold analysis identified inflection points at CKMAI 61.587 (all‑cause mortality), 60.732 (cardiovascular mortality), and 34.092 (high‑risk CKM status), with steeper risk increases below each threshold. CKMAI exhibited super‑additive interactions with PhenoAge (RERI 15.06, 95% CI [8.14, 38.44]). Clustering revealed six distinct aging‑metabolic phenotypes; the frail elderly cluster had the highest mortality risk (hazard ratio [HR] 11.74, 95% CI [7.85, 17.57]). Depression partially mediated the CKMAI‑outcome associations (proportions mediated: 5.6% [95% CI 2.3%-9.0%] for all‑cause mortality, 8.2% [95% CI 2.8%-22.0%] for cardiovascular mortality, 11.1% [95% CI 6.7%-24.0%] for high‑risk CKM status). Findings were robust across prespecified subgroups and multiple sensitivity analyses. A key limitation is the lack of validation in geographically independent cohorts with complete mortality follow-up; however, temporal validation within NHANES and preliminary external validation in a hospital based Chinese cohort (n = 261) supported the generalizability of the CKMAI for identifying high risk CKM status. CONCLUSIONS: CKMAI is a valid, CKM‑specific aging index that outperforms universal biological age measures in predicting mortality and identifying high‑risk CKM status. Its super‑additive interactions with PhenoAge and the partial mediation by depression offer new mechanistic insights into CKM heterogeneity. By integrating a diverse set of clinical laboratory measures, nutritional assessments, lifestyle factors, and social determinants of health, CKMAI provides a practical, implementable tool for risk stratification and targeted prevention in CKM syndrome.

来源:PLoS Med,查看原文。版权归原作者所有。

引用

  1. Zhu Z, Ren K, Wang D, et al. A machine learning-derived aging index for risk stratification and mortality prediction in cardiovascular-kidney-metabolic syndrome: A retrospective cohort study. PLoS Med. 2026. PMID: 42837317. doi:10.1371/journal.pmed.1005078

供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。