Hu Beilei, Chen Xuan, Chen Tingyang, Xu Tong, Cao Yungang, Sun Jing, Chen Xuanyu, Chen Songfang, Chen Keyang
Department of Neurology, The Second Affiliated Hospital and Yuying Children's Hospital of Wenzhou Medical University, Wenzhou, China.
Wenzhou Key Laboratory of Neurogenetics, Wenzhou, China.
EClinicalMedicine. 2025 Jul 1;85:103331. doi: 10.1016/j.eclinm.2025.103331. eCollection 2025 Jul.
Obesity is a significant risk factor for stroke. However, body mass index is insufficient in assessing fat distribution and there is a need for a better indicator to predict stroke risk. Additionally, early detection and prognosis prediction for stroke and mortality are crucial for pre-emptive interventions. We examined to evaluate the utility of obesity-related indices in a stacked machine learning (ML) model by developing an in-silico quantitative marker (ISS) to predict stroke risk.
This is a prospective cohort study utilizing data from the China Health and Retirement Longitudinal Study (CHARLS) (2011-2018) and a health examination cohort in China (2017-2024), English Longitudinal Study of Ageing (ELSA) (2004-2014) in the UK. A total of 13,324 participants from CHARLS were included in the cross-sectional analysis. For model development and internal and external validation, 10,044 participants from CHARLS, 3698 from ELSA, and 6884 from the second affiliated hospital of Wenzhou medical university were included. Stacked ML models with optimal obesity indices to detect the risk of stroke were constructed. The predictive accuracy of the models was evaluated with the area under the receiver operating curve (ROC-AUC).
Triglyceride-Glucose-Body Mass Index (TyG-BMI) and TyG were two optimal predictors and outperformed BMI (AUC = 0.821) in the cross-sectional study. In the longitudinal cohort, the model with the highest AUC was the stacked ML model incorporating TyG-BMI, which achieved an AUC of 0.816 (95% CI: 0.807-0.824) in the training cohort and 0.833 (95% CI: 0.816-0.849) in the internal set for predicting stroke risk. For the external sets, the AUC was 0.803 (95% CI: 0.791-0.816) for the ELSA cohort and 0.805 (95% CI: 0.793-0.818) for the health examination cohort. The stacked ML model based on TyG-BMI showed the best performance with the highest F1 score (0.209:0.124:0.117), lowest Brier score (0.040:0.041:0.041) and model improvement (all NRI and IDI >0). The ISS score was significantly associated with stroke and stroke-related death, classifying individuals into low- and high-risk groups for death in the training cohort with and AUC of 0.891 (95% CI: 0.840, 0.935) and 0.879 (95% CI: 0.749, 0.979) for the internal validation sets.
The stacked ML model incorporating TyG-BMI effectively predicts stroke risk, with the ISS score demonstrating strong performance across diverse populations. Further research is needed to assess its applicability in broader cohorts.
None.
肥胖是中风的一个重要危险因素。然而，体重指数在评估脂肪分布方面并不充分，需要一个更好的指标来预测中风风险。此外，中风和死亡率的早期检测及预后预测对于预防性干预至关重要。我们通过开发一种计算机定量标志物（ISS）来预测中风风险，以评估肥胖相关指标在堆叠机器学习（ML）模型中的效用。
这是一项前瞻性队列研究，利用了中国健康与养老追踪调查（CHARLS）（2011 - 2018年）、中国一个健康检查队列（2017 - 2024年）以及英国老龄化纵向研究（ELSA）（2004 - 2014年）的数据。共有13324名来自CHARLS的参与者被纳入横断面分析。为了进行模型开发以及内部和外部验证，纳入了10044名来自CHARLS的参与者、3698名来自ELSA的参与者以及6884名来自温州医科大学附属第二医院的参与者。构建了具有最佳肥胖指标的堆叠ML模型以检测中风风险。通过受试者工作特征曲线下面积（ROC - AUC）评估模型的预测准确性。
在横断面研究中，甘油三酯 - 血糖 - 体重指数（TyG - BMI）和TyG是两个最佳预测指标，其表现优于体重指数（AUC = 0.821）。在纵向队列中，AUC最高的模型是纳入TyG - BMI的堆叠ML模型，在训练队列中预测中风风险时AUC为0.816（95%CI：0.807 - 0.824），在内部验证集中为0.833（95%CI：0.816 - 0.849）。对于外部数据集，ELSA队列的AUC为0.803（95%CI：0.791 - 0.816），健康检查队列的AUC为0.805（95%CI：0.793 - 0.818）。基于TyG - BMI的堆叠ML模型表现最佳，F1分数最高（0.209:0.124:0.117），布里尔分数最低（0.040:0.041:0.041）且模型有改善（所有净重新分类指数和综合判别改善指数均>0）。ISS分数与中风及中风相关死亡显著相关，在训练队列中将个体分为低风险和高风险死亡组，内部验证集的AUC分别为0.891（95%CI：0.840，0.935）和0.879（95%CI：0.749，0.979）。
纳入TyG - BMI的堆叠ML模型能有效预测中风风险，ISS分数在不同人群中表现出色。需要进一步研究以评估其在更广泛队列中的适用性。
无。
