This study adopted a cross-sectional design, which is suited for exploring associations between risk factors and functional outcomes in stroke survivors. Cross-sectional studies are advantageous for identifying potential predictors of post-stroke disability, providing a basis for future longitudinal investigations. Data were obtained from wave 5 of the China Health and Retirement Longitudinal Study (CHARLS). The dataset is publicly accessible via the official CHARLS website (http://CHARLS.pku.edu.cn). The study adhered to ethical norms and received approval from the Biomedical Ethics Committee of Peking University, China (IRB00001052-11,015) (7). All procedures followed the principles outlined in the Declaration of Helsinki, and informed consent was obtained from all participants. In this study, patients and the public were not involved in the design, conduct, reporting, or dissemination plans of the research. This study adhered to the Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD) (8).

In the CHARLS database, stroke status was determined through self-reported responses to the question: “Have you ever been diagnosed with a stroke by a doctor?” This self-reported approach may lead to potential misclassification, which was considered in the study's limitations. ADL in two main categories: Basic ADL (BADL), which includes six essential tasks: dressing, bathing, eating, transferring (getting in and out of bed), toileting, and controlling urination and defecation. Instrumental ADL (IADL), which involves more complex activities, such as managing housework, cooking, shopping, financial management, and medication adherence. We used a questionnaire-based survey method to assess ADL dysfunction. If a stroke survivor was unable to independently complete any of the tasks listed under BADL or IADL, they were classified as having ADL dysfunction. Inclusion criteria: History of stroke; Ability to cooperate in completing the ADL screening. Exclusion criteria: No history of stroke or uncertain stroke diagnosis; Missing key variables; Pre-existing ADL dysfunction before stroke. Initially, 1,381 individuals with a self-reported history of stroke were identified. After excluding individuals with more than 30% missing data, the final analysis included 1,131 stroke survivors.

Predictor selection was based on prior literature and clinical expertise (9–11). Although stroke characteristics (e.g., lesion location, infarct size) influence prognosis, these variables were not recorded in CHARLS. Instead, we examined demographic, behavioral, health, and socioeconomic factors available in the dataset. Basic factors: age, gender, residence (urban/rural), education level, marital status, and life satisfaction (five-point Likert scale). Behavioral factors: sleep duration, smoking, alcohol consumption, social activity participation (eight categories), and physical activity levels (light, moderate, heavy). Total energy expenditure from physical activity was calculated using metabolic equivalent (MET) scores. Health status and medical conditions: self-rated health (five levels), hypertension, diabetes, cancer, cardiac disease, mental disorders, and the 10-item Center for Epidemiologic Studies Depression Scale (CESD-10). Family and economic factors: household size, financial support from children/parents, and number of surviving children.

All statistical analyses were conducted using R software. Continuous variables were reported as medians and interquartile ranges, while categorical variables were presented as proportions. Between-group comparisons were performed using the Wilcoxon rank-sum test for continuous data and the Chi-square test or Fisher's exact test for categorical data. To develop the ADL dysfunction prediction model, the dataset was randomly split (6:4) into a training set (n = 678) and a testing set (n = 453). We applied the least absolute shrinkage and selection operator (LASSO) regression to identify key predictors while addressing multicollinearity. Optimal tuning parameters (λ) were selected via ten-fold cross-validation. Selected variables were incorporated into a multivariate logistic regression model, with predictors retained at P < 0.05. Model performance was assessed using the area under the receiver operating characteristic (ROC) curve (AUC) for discrimination, calibration plots for agreement between predicted and observed outcomes, and decision curve analysis (DCA) for clinical utility. SHapley Additive exPlanations (SHAP) values were computed to interpret predictor importance.

The study flow chart is presented in Figure 1.

A total of 1,131 stroke survivors were included in this study. The demographic and clinical characteristics of the participants are summarized in Table 1. The cohort consisted of 473 male participants (41.8%) and 658 female participants (58.2%), with an average age of 67 years. Among these stroke survivors, 57.3% experienced difficulties with ADL. Several factors showed significant differences (p < 0.05) between stroke survivors with normal and impaired ADL function, including age, gender, education level, life satisfaction, CESD-10 score, and health status factors (e.g., hypertension, lung disease, arthritis). Additionally, factors such as social activity, sleep duration, pain count, and physical activity levels were also significantly associated with ADL dysfunction.

LASSO regression was applied to identify the best predictors for ADL dysfunction, with predictors selected based on 10-fold cross-validation. The 11 significant variables identified included gender, age, hip fracture, CESD-10 score, memory disorder, pain areas, and levels of physical activity (Figure 2). These variables were then used in a multivariate logistic regression model, which selected the following significant predictors (P < 0.05): age, CESD-10 score, memory disorder, self-rated health, pain count, and heavy physical activity. The resulting predictive model was visualized through a nomogram, which allows for the quantitative assessment of ADL dysfunction risk in stroke survivors (Figure 3).

The predictive model's performance was assessed using the AUC. In the training set, the AUC value was 0.804 (95% CI: 0.772–0.837), and in the testing set, it was 0.779 (95% CI: 0.736–0.821), indicating good discriminatory power (Figures 4A, B). The nomogram's calibration curves (Figures 4C, D) showed alignment between predicted and observed probabilities of ADL dysfunction, confirming the model's accuracy and reliability. Clinical validity was assessed using DCA, shown in Figures 4E, F. The DCA demonstrated that the prediction model provided net benefits compared to the two extreme scenarios, suggesting its clinical utility in predicting ADL dysfunction.

SHAP values were calculated for six key variables in the model. The global importance plot and swarm plot (Figures 5A, B) demonstrated the predictive importance of these variables across the dataset. To further understand their impact at the individual level, waterfall and force plots (Figures 5C, D) were used to visualize the contribution of each variable to the model's predictions in selected samples, highlighting the practical significance of these variables in specific cases.