Related Experiment Videos
Clinical risk stratification of stroke caregiver distress: an interpretable nonlinear model of caregiving-load
Yingjie Zheng1, Shailing Ma1, Xiaohui Liu1
1School of Nursing, Ningxia Medical University, Yinchuan, China.
Background:
Family caregivers of stroke survivors frequently experience psychological distress. Social support and active coping may be protective, but their associations can vary across levels of caregiving load. Interpretable machine learning may help characterize such nonlinear patterns without assuming a constant linear association.
Methods:
This multicenter cross-sectional study included 506 family caregivers. High psychological distress was defined as a Hospital Anxiety and Depression Scale (HADS) total score ≥15. We trained a machine-learning model using extreme gradient boosting (XGBoost) with six fixed analytic predictor domains: role burnout, perceived social support, active coping, daily sleep-duration category, daily caregiving-time category, and income level. Model performance was evaluated in a stratified 30% held-out test set, with bootstrap confidence intervals and nested cross-validation. TreeSHAP was used for model attribution and interaction exploration.
Results:
Among 506 caregivers, 178 (35.2%) met the high-distress criterion. Most caregivers were aged 18-59 years (85.4%), and 60.3% were women. XGBoost achieved a test-set AUC of 0.890 (95% CI: 0.830-0.937); the mean nested five-fold cross-validated AUC was 0.891 (SD 0.024). Random forest and support vector machine models showed comparable discrimination. Role burnout had the largest mean absolute SHAP value, followed by active coping and perceived social support. The three recorded sleep-duration categories ( ≤ 5, 6-8, and ≥9 h) and caregiving-time categories (6-8, 9-16, and ≥17 h) showed category-level differences in model attribution; these boundaries were questionnaire-defined and were not estimated as continuous thresholds. Interaction plots suggested that the associations of support and coping varied across load categories.
Conclusions:
An interpretable nonlinear model identified combinations of role burnout, psychosocial resources, and caregiving-load categories associated with high psychological distress. The category-level and interaction findings are hypothesis-generating rather than causal or diagnostic and require prospective external validation before clinical threshold use.
