Predicting cardiometabolic multimorbidity in Chinese older adults via machine learning

Study Subjects

The data for this study were derived from the CHARLS, a nationally representative longitudinal survey focusing on Chinese residents aged 45 years and older. CHARLS aims to collect comprehensive information on the social, economic, and health conditions of middle-aged and older adults in China, and it provides a high-quality public database for scientific research supported by the National Development Institute of Peking University. The baseline survey was conducted between 2011 and 2012, with follow-up waves carried out every two years thereafter^[14]. The sampling strategy employed a multistage probability proportional to size design, covering 150 counties and 450 communities across all 28 provincial-level regions in mainland China. Households and individuals were randomly selected within sampled communities. For the present analysis, we pooled data from four survey waves (2011, 2013, 2015, and 2018). The initial pool consisted of 25,538 observations. We excluded participants under the age of 45 (n = 5,819) and those with missing data on key CMM diagnoses in 2018 (n = 2,749). After applying these criteria, the final analytical sample comprised 16,970 participants. A flowchart detailing the sample selection process is provided in Supplementary Figure 1A. All CHARLS survey waves received ethical approval from the Peking University Institutional Review Board (IRB). The protocols for the main questionnaire and physical measurements were approved under IRB number IRB00001052-11015, and the biomarker collection was approved under IRB number IRB00001052-11014^[15]. All participants provided written informed consent.

Predictors of CMM

Based on established evidence and clinical relevance, a comprehensive set of 42 predictors was selected for training the ML algorithms. These predictors were categorized into three domains for clarity: self-reported questionnaire data, physical examination measurements, and blood-based biomarkers. The self-reported questionnaire encompassed: (1) demographic factors (sex, age, and region)^[16]; (2) lifestyle behaviors (night sleep duration, smoking status, and drinking status)^[17]; (3) subjective health assessments (presence of pain, recent weight change, recalled health status during childhood, and self-expectations of current health status)^[18]; (4) disease history and conditions, including mental health (depression), sensory impairment (hearing and vision), and various chronic conditions (dyslipidemia, liver disease, kidney disease, stomach or other digestive disease, memory-related disease, arthritis or rheumatism, menopause, and prostatic diseases); (5) living environment factors (house structure, primary energy sources for heating and cooking, and subjective assessment of room temperature)^[19]; and (6) socioeconomic factors (education level and self-rated standard of living)^[20,21]. Objective measurements included: (1) physical examination data: hand grip strength (as a marker of musculoskeletal health), waist circumference, and body mass index (BMI), all indicators of body composition and metabolic risk^[22]; and (2) blood biomarker data: a panel of tests capturing inflammation, renal function, and lipid metabolism. This panel included white blood cell count, hemoglobin, platelet count, C-reactive protein (CRP), creatinine, blood urea nitrogen (BUN), uric acid (UA), cystatin C, total cholesterol (TC), triglyceride (TG), high-density lipoprotein cholesterol (HDL-C), and low-density lipoprotein cholesterol (LDL-C)^[23,24]. The detailed distribution and measurement methods for all predictor variables are provided in Supplementary Table 1. Except for the self-rated standard of living and weight change variables, which were extracted from the 2013 Wave, all other predictors were derived from the 2015 Wave data to predict the CMM outcomes in the 2018 Wave.

Measurement of CMM

CMM was defined as the presence of two or more of the following cardiometabolic diseases - hypertension, diabetes, heart disease, or stroke - diagnosed by a physician between the 2015 predictor assessment and the 2018 follow-up wave. In this study, CMM status in 2018 was ascertained using self-reported physician-diagnosed disease information collected in the CHARLS questionnaire. The specific conditions considered were: hypertension, diabetes, heart disease (including coronary heart disease, heart failure, or other heart problems), and stroke^[25]. The presence of each condition was determined based on affirmative responses to questions such as “Have you been diagnosed with [condition] by a doctor?”. Participants who reported a confirmed diagnosis of two or more of the aforementioned conditions were classified as having CMM. Those who reported none or only one of these conditions were categorized as not having CMM. This operational definition aligns with established approaches for studying multimorbidity in epidemiological research and facilitates the identification of individuals bearing a higher burden of interrelated cardiometabolic conditions.

Statistical analysis

Statistical analyses and predictive modeling were conducted using R version 4.4.2. Continuous variables were summarized as medians with interquartile ranges (IQR), while categorical variables were presented as frequencies and percentages. All predictive modeling was performed using the H2O package (version 3.46.0.6) in R^[26]. Additional R packages used included missForest (version 1.5) for imputation, pROC (version 1.18.0) for DeLong’s tests, rms (version 6.7.0) for calibration metrics, and ggplot2 (version 3.4.2) for visualizations. We developed and compared several ML algorithms to predict CMM: the Generalized Linear Model (GLM) with regularization, serving as the benchmark logistic regression model^[27]; the Gradient Boosting Machine (GBM), which builds trees sequentially to minimize prediction errors through boosting^[28]; the Distributed Random Forest (DRF), an ensemble of decision trees trained on bootstrapped samples to enhance robustness and reduce overfitting^[29]; the Deep Learning (DL) model, employing a fully connected multilayer neural network to capture complex nonlinear patterns^[30]; and a Stacked Ensemble model (GBM-XGBoost-GLM-DL-DRF) (XGBoost = Extreme Gradient Boosting). The Stacked Ensemble model combined the predictions of the four base learners (GLM, GBM, DRF, DL) using a meta-learner to synthesize their strengths and improve overall predictive performance^[31]. Specifically, we employed logistic regression as the meta-learner, using the cross-validated predicted probabilities from the four base models as input features. The meta-learner was trained with 5-fold cross-validation on the training set to optimize its parameters, and the final output was the predicted probability of CMM. The study process is shown in Supplementary Figure 1B. Prior to model development, the dataset underwent several preprocessing steps. Missing values were imputed using the random forest-based multiple imputation algorithm (missForest) with five imputations; the first imputed dataset was used for analysis. Continuous predictor variables were standardized to have a mean of 0 and a standard deviation of 1 based on the means and standard deviations of the training set. Crucially, all preprocessing steps - including imputation, normalization, and over-sampling - were performed using only the training set, and the resulting parameters were then applied to the testing set. To ensure internal evaluation, the data were randomly split into a training set (80%) for model development and a testing set (20%) for final validation. Given the imbalanced distribution of the CMM outcome (ratio approximately 1:3.5), we applied the Synthetic Minority Over-sampling Technique (SMOTE) to the training data to mitigate potential bias and enhance the model’s sensitivity in identifying the minority class (CMM cases)^[32]. Hyperparameters for each algorithm were tuned using 5-fold cross-validation on the training set to optimize performance and prevent overfitting^[33]. A random search strategy was employed over a pre-specified hyperparameter grid. Model performance was evaluated on the independent testing set. The area under the receiver operating characteristic curve (AUC) served as the primary metric for assessing discriminative ability, interpreted as follows: 0.5-0.7 (poor), 0.7-0.8 (acceptable), 0.8-0.9 (good), and > 0.9 (excellent)^[34]. Calibration curves were plotted to visualize the agreement between predicted probabilities and observed outcomes^[35]. In addition to graphical assessment, we quantitatively evaluated calibration using the calibration intercept, calibration slope, and the Spiegelhalter Z-test. The Brier score was calculated to quantify the overall prediction error, with scores closer to 0 indicating better calibration^[36]. Conventionally, a Brier score below 0.25 is considered indicative of acceptable model performance. Additional performance metrics derived from the confusion matrix - including accuracy, precision, true positive rate (TPR), and F1-score (the harmonic mean of precision and recall) - were reported for a comprehensive assessment. Finally, variable importance analysis was performed for the best-performing models to identify and rank the key predictors contributing most significantly to CMM prediction.

Geographical Distribution of CMM

The geographic distribution of CMM prevalence among middle-aged and older adults across Chinese provinces is presented in Figure 1. Analysis revealed marked regional disparities. The highest prevalence was observed in Tianjin (46/96, 47.9%), Xinjiang Uygur Autonomous Region (30/70, 42.9%), Shanghai (20/49, 40.8%), Inner Mongolia Autonomous Region (305/753, 40.5%), and Heilongjiang (104/281, 37.0%). Conversely, the lowest prevalence was found in Guizhou (18/176, 10.2%), Fujian (60/451, 13.3%), Guangdong (104/796, 13.1%), Guangxi (76/544, 14.0%), and Chongqing (35/254, 13.8%). The remaining provinces exhibited moderate CMM prevalence rates, ranging from approximately 15% to 35%. Detailed prevalence data for each province are provided in Supplementary Table 2.

Figure 1. The prevalence of cardiometabolic multimorbidity (CMM) by province in China from the China Health and Retirement Longitudinal Study in 2018. This map was created based on the standard map with the approval number GS(2024)0650 downloaded from the Standard Map Service website of the Ministry of Natural Resources of the People’s Republic of China (http://bzdt.ch.mnr.gov.cn/). The base map has not been modified.

Characteristics of the study participants

The final analytic sample comprised 16,970 individuals aged 45 years and older from the CHARLS. Among them, 22.0% (n = 3,737) were identified as having CMM. The sample was divided into a training set (n = 13,577, 80%) for model development and a testing set (n = 3,393, 20%) for validation. Selected demographic and socioeconomic characteristics of the participants are presented in Table 1. The prevalence of CMM increased significantly with age, from 20.6% in the 45-55 years group to 42.7% in those aged 65 years or older. A higher proportion of females (56.4%) were affected by CMM compared to males (43.6%). Geographically, the central region had the highest proportion of participants with CMM (38.2% of all CMM cases), followed by the eastern and western regions. Participants with lower educational attainment (less than elementary school: 43.4% of CMM cases) and those reporting a poorer standard of living (poor: 14.2% of CMM cases) were more prevalent in the CMM group. The complete distribution of all predictor variables is provided in Supplementary Table 1.

CMM prediction

The performance of the trained models was evaluated on the independent testing dataset. The distribution of predictor variables across the training and testing sets is provided in Supplementary Table 3. The Stacked Ensemble model demonstrated the highest discriminative ability in internal validation, with an AUC of 0.755, outperforming the GLM (DeLong’s test P = 0.03) and DL (P = 0.03) models, while exhibiting comparable performance to GBM (P = 0.21) and DRF (P = 0.15). Among the single algorithms, GBM achieved the highest AUC (0.751), closely followed by DRF (0.748) and GLM (0.740). The DL model yielded an AUC of 0.732. Corresponding accuracy, precision, recall, and F1-score for each model are provided. Given the class imbalance in our dataset, the AUC was prioritized as the primary metric for model comparison as it provides a more robust assessment of the model’s ability to rank positive and negative instances. The Stacked Ensemble model also attained the highest TPR (0.711) and F1-score (0.497). Comprehensive performance metrics for all models evaluated on the testing set are summarized in Supplementary Table 4. Figure 2 illustrates the ROC curves.

Figure 2. Receiver operating characteristic curves of all cardiometabolic multimorbidity (CMM) prediction models on the training and testing datasets. AUC: Area under the curve; DL: deep learning; DRF: distributed random forest; GBM: gradient boosting machine; GLM: generalized linear model; ROC: receiver operating characteristic; Stacked Ensemble: GBM-XGBoost-GLM-DL-DRF.

Assessing the efficacy of ML models for CMM prediction

The calibration curves for all models, which plot predicted probabilities against the observed frequencies of CMM in the testing dataset, are presented in Figure 3. These curves visually assess the agreement between the models’ predictions and the actual outcomes. The search space and the final selected hyperparameters for each model are detailed in Supplementary Table 5. Quantitative calibration metrics are summarized in Supplementary Table 6 and displayed on each calibration curve facet. The Stacked Ensemble model demonstrated excellent calibration, with an intercept of -0.02 [95% confidence interval (CI): -0.15, 0.11], a slope of 0.96 (95%CI: 0.88, 1.04), and a non-significant Spiegelhalter Z-test (Z = 0.74, P = 0.46), indicating no evidence of systematic miscalibration. To quantify this calibration performance, the Brier score was calculated for each model. As shown, all models achieved Brier scores well below 0.25, confirming their good reliability in predicting CMM risk. The Stacked Ensemble, GBM, and DRF models demonstrated the best calibration, both with a Brier score of 0.153.

Figure 3. The calibration curves of all cardiometabolic multimorbidity (CMM) prediction models on the testing dataset. DL: Deep learning; DRF: distributed random forest; GBM: gradient boosting machine; GLM: generalized linear model; Stacked Ensemble: GBM-XGBoost-GLM-DL-DRF.

Determinants of CMM

The GBM and DRF models, which demonstrated the highest predictive performance among the single models, identified a consistent set of predictors for CMM in the variable importance analysis [Figure 4]. Dyslipidemia was the most important predictor in both models. Age, TG, HDL-C, waist circumference, self-expectations of health status, pain, cooking energy type, weight change, and cystatin C were also identified as key factors in both algorithms. The prominence of self-rated health expectations and pain suggests that patients’ subjective experiences and perceptions capture important dimensions of risk not fully reflected in objective clinical measures, highlighting the potential value of incorporating patient-reported outcomes into routine CMM screening. The inclusion of cooking energy type - a proxy for indoor air pollution - underscores the role of environmental exposures in cardiometabolic risk and points to potential interventions beyond the healthcare setting. These results highlight the central role of metabolic health, body composition, lifestyle factors, and self-perceived health in the development of CMM among aging adults in China.

Figure 4. Variable importance analysis performed by the distributed random forest (DRF) and gradient boosting machine (GBM) model. BMI: Body mass index; CRP: C-reactive protein; BUN: blood urea nitrogen.

Comparison with existing research results

Our stacked ensemble model achieved an AUC of 0.755, which falls within the acceptable range (0.7-0.8) and outperforms most prior CMM prediction models based on similar populations. For example, Zhu et al. employed ML algorithms - including logistic regression and Gaussian Naive Bayes - on both the CHARLS and Health and Retirement Study (HRS) cohorts to predict CMM and achieved an AUC of approximately 0.70^[12]. While informative, their models relied primarily on conventional algorithms and demonstrated limited improvement over baseline methods. In contrast, our ensemble model - leveraging gradient boosting, random forests, and other learners - achieved higher predictive accuracy while maintaining interpretability. Moreover, our model exhibited improved calibration, as evidenced by a lower Brier score (0.153), indicating a better match between predicted risks and observed outcomes. In another recent study, Mao et al. applied a network-based ML approach to cross-sectional National Health and Nutrition Examination Survey (NHANES) data and reported an exceptionally high AUC of 0.988^[11]. While this result appears impressive, it is important to note that their model was trained and tested on cross-sectional data, which may overestimate performance due to a lack of temporal separation between predictors and outcomes. Additionally, network-based models often entail high model complexity and reduced generalizability, especially when deployed in longitudinal or real-world clinical settings. In comparison, our model was trained on temporally structured, longitudinal data and validated on a temporally distinct test set, which more closely mirrors practical deployment scenarios.

Importantly, our model confirms and extends existing evidence regarding key risk factors for CMM. Consistent with prior studies, age, dyslipidemia, and central obesity (as measured by waist circumference) emerged as dominant predictors^[12,37]. However, our inclusion of TG, HDL-C, and cystatin C - biomarkers often underrepresented in traditional risk models - highlighted the potential utility of expanded biochemical profiling in improving prediction accuracy. For instance, the identification of cystatin C, a marker of renal function, as a top contributor aligns with recent studies suggesting its role in systemic inflammation and cardiovascular-renal metabolic syndromes (Li et al.), indicating a shared pathophysiological pathway across organ systems^[38]. Furthermore, our study uniquely emphasizes the predictive value of self-rated expectations of health, subjective pain reports, and cooking fuel types. These variables are often overlooked in clinical or epidemiological risk models yet showed high importance in our GBM and DRF models. This finding echoes insights (Ji et al.)^[39], while chronic pain reflects low-grade systemic inflammation, a known contributor to cardiometabolic risk (Xu et al.)^[40]. The inclusion of indoor air pollution proxies - such as cooking energy type - is especially relevant in the Chinese context, where solid fuel use remains prevalent in rural and peri-urban areas and has been linked to hypertension and metabolic disorders through oxidative stress pathways (Smith et al.)^[41]. The significance of such environmental factors underscores the importance of tailoring prediction models to local contexts.

Notably, our study found a relatively modest AUC for DL (AUC = 0.732) compared to ensemble methods. This aligns with the broader literature, which has highlighted that in tabular health data with limited sample sizes and strong domain-specific structure, traditional ensemble learners often outperform deep neural networks (Miletic et al.)^[42]. Our ensemble strategy, which synthesized the strengths of diverse learners, ultimately enhanced predictive performance while maintaining interpretability through feature importance analyses. Another novel aspect of our work is the regional heterogeneity in CMM prevalence. While regional variation in chronic disease burden has been previously reported in China (Su et al.)^[43], our data-driven visualization reveals specific provinces - such as Tianjin, Inner Mongolia, and Xinjiang - with disproportionately high CMM prevalence. This spatial pattern may reflect differences in healthcare access, urbanization level, lifestyle, air pollution, and ethnic diversity, suggesting that region-specific risk models or public health strategies may be warranted^[44]. While most Western studies rely on electronic health records or clinical trial cohorts, our use of CHARLS - a community-based, longitudinal dataset - provides greater population-level generalizability. Moreover, by incorporating socioeconomic and environmental dimensions (e.g., education, heating fuels), we bridge the gap between clinical risk modeling and social determinants of health, which are particularly salient in aging societies facing urban-rural disparities such as China^[45].

Beyond these comparisons, our study offers three distinct and substantive advances. First, we provide the first empirical quantification of the incremental predictive value of psychosocial, environmental, and subjective health indicators for CMM in a large, nationally representative Chinese cohort. Second, this study represents the first systematic benchmarking of five diverse ML algorithms - including stacked ensemble - for CMM prediction in China’s aging population, establishing ensemble tree-based methods as the current optimal approach for this task. Third, we present the first province-level prevalence map of CMM among middle-aged and older Chinese adults, revealing marked geographic disparities and providing an actionable evidence base for regionally targeted public health interventions. Collectively, these contributions move beyond incremental application of ML to fundamentally expand the conceptual scope of CMM risk prediction by integrating psychosocial, environmental, and geospatial dimensions.

Strengths and limitations

A major strength of our study is the integration of rich multidimensional data from CHARLS, allowing us to examine not only biological and clinical predictors but also behavioral, psychosocial, and environmental factors. The use of ML approaches enabled the capture of complex, nonlinear interactions among predictors, which traditional regression models may fail to uncover. The ensemble model’s superior performance highlights the value of combining diverse learning strategies for robust risk prediction. Additionally, our study addresses the issue of class imbalance through SMOTE, ensuring that minority-class instances (CMM cases) are adequately learned by the model. This step is crucial given the typically lower prevalence of multimorbidity relative to non-cases, and it enhances the model’s practical utility in identifying high-risk individuals.

Nonetheless, some limitations merit attention. First, the CMM outcome was based on self-reported physician diagnoses, which may be affected by recall or reporting bias. Although previous CHARLS-based validation studies support the accuracy of self-reported chronic conditions, the absence of objective clinical or laboratory confirmation limits the precision of the outcome definition. Second, most predictors were obtained from a single time point (2013 or 2015), while CMM status was assessed in 2018. This cross-sectional approach precludes capturing time-varying risk dynamics. Future work should explore longitudinal modeling frameworks that incorporate changes in risk profiles over time. Third, although we addressed class imbalance through SMOTE and used internal testing for validation, the model has not yet been externally validated in independent cohorts - either in China or internationally. Without such external testing, the generalizability and transportability of our prediction tool remain uncertain. Fourth, while some interpretability was introduced via variable importance ranking, ML models - especially ensemble and DL methods - are often criticized for their “black box” nature. Clinicians may require further interpretive tools or simplified models before real-world adoption. Fifth, potential residual confounding may persist due to unmeasured variables such as dietary patterns, environmental toxins, social support networks, or medication use. These were not captured in the CHARLS dataset and should be considered in future iterations. Sixth, we also note that the AUCs on the training set were higher than those on the test set, which is partly attributable to the application of SMOTE on the training data; nevertheless, the test set AUCs remained at an acceptable level and the relative performance of the models was consistent. Finally, the observational design of our study prevents any causal inference. While associations were identified, they should not be construed as evidence of causality without further mechanistic or interventional studies.

Implications for practice and research

Future efforts integrating the demonstrated predictive value of non-traditional factors - such as self-rated health expectations, chronic pain, and household air pollution proxies - highlight the need to adopt a more holistic approach to chronic disease risk assessment in primary care and public health settings. These findings suggest that effective CMM prevention must extend beyond clinical risk factor modification to address psychosocial well-being and environmental exposures. For example, interventions could combine routine screening for depression and chronic pain with referral pathways to mental health and pain management services, while simultaneously promoting access to clean cooking fuels through community-level programs or policy initiatives. Incorporating psychosocial and environmental factors into screening tools may improve early detection of individuals at risk of CMM. Second, the successful application of ML models in a nationally representative sample suggests that such tools could be feasibly integrated into digital health platforms or community screening programs in China. By identifying individuals at highest risk and the specific drivers of that risk, these models could enable truly personalized, multi-domain intervention packages that address medical, psychological, and environmental needs concurrently - moving beyond the concept of “systemic lifestyle intervention” to a more precise and actionable framework of “multi-level risk‐tailored prevention”. With appropriate validation and refinement, our model could support resource allocation by identifying high-risk groups for targeted interventions, particularly in underserved or high-prevalence regions identified by our geographic analysis. Third, our results underscore the importance of intersectoral collaboration in addressing multimorbidity. Given the influence of living conditions and health perceptions on disease risk, interventions that combine medical care with social support, environmental regulation, and health education may yield greater effectiveness than siloed clinical approaches. Fourth, we acknowledge that some predictors (e.g., dyslipidemia, kidney and liver disease) are conceptually linked to CMM components. However, our goal was risk prediction - not causal inference. Their inclusion is justified by temporal precedence (measured 3 years prior) and their value in short-term risk stratification. The model is intended for individuals with existing metabolic or renal abnormalities, rather than for lifelong primary prevention. Fifth, a key limitation is our reliance on internal validation only. Given the cultural, environmental, and genetic heterogeneity across Chinese regions - as reflected in the regional disparities in CMM prevalence - the model’s generalizability remains uncertain. Future studies must validate it in independent, geographically diverse populations. We also acknowledge that some predictors may capture overlapping physiological constructs; however, the ensemble tree-based algorithms used in this study are inherently robust to multicollinearity, and our primary focus is on predictive accuracy rather than estimation of independent effects. Finally, our study opens avenues for future mechanistic research. Additionally, investigating gene-environment interactions or integrating imaging/genomic data into ML models could further enhance prediction. Given the global trend of population aging, international comparative studies using similar modeling frameworks could elucidate cross-cultural patterns and determinants of multimorbidity.

In conclusion, in this study, we developed and validated a robust ML model to predict CMM among China’s aging population using data from the nationally representative CHARLS cohort. By integrating multidimensional predictors - including clinical biomarkers, physical measurements, self-reported health status, and environmental factors - our ensemble learning approach achieved favorable predictive performance and highlighted both established and novel determinants of CMM risk. The findings emphasize the critical roles of metabolic, renal, and other factors in chronic disease risk prediction and provide actionable insights for public health screening, early prevention, and resource allocation in aging societies. Future studies should focus on external validation across diverse populations, particularly in southern and western China, incorporation of dynamic longitudinal data, and integration into digital health tools for real-world clinical application.