Explainable Machine Learning for Risk Prediction of Reduced Quality of Life in Hypertension.

Background: Individuals with hypertension are at risk to reduced quality of life (QoL). Explainable machine learning (ML) can be used for domain-specific risk stratification and prioritization of modifiable determinants of low QoL. Objective: To train ML classifier algorithms for QoL risk stratification in hypertension, where meaningful determinants were explored through Shapley additive explanations (SHAP). Methods: Data from hypertensive individuals (n = 534) completed WHOQOL BREF, Quick Physical Activity Rating, Morisky Medication Adherence Scale 8, and standardized questionnaires for acceptance and knowledge were analyzed utilizing Decision Tree, Gradient Boosting, XGBoost, AdaBoost, Random Forest, and Naive Bayes ML classifiers. The trained ML algorithms were evaluated using stratified 10-fold cross-validation, where the stability was examined using rank-based metrics. SHAP were applied to the gradient boosting, as the most stable model. Results: For physical QoL, Random Forest (AUC 0.850; sensitivity 0.835; specificity 0.738) and Gradient Boosting (AUC 0.850; sensitivity 0.801; specificity 0.764) showed good reduced QoL identification. For psychological domain, best classifications were obtained from Gradient Boosting performed best (AUC 0.833; sensitivity 0.818; specificity 0.651) and XGBoost (AUC 0.831; sensitivity 0.824; specificity 0.660), with the former observed as the most stable SHAP analysis identified acceptance and medication adherence as the dominant shared drivers of risk across both QoL domains. Physical QoL risk was further influenced by physical activity-related factors, whereas Psychological QoL risk showed additional contributions from age and educational attainment. Conclusion: Ensemble tree-based classifiers, particularly Gradient Boosting, had the most optimal performance in discriminating reduced and good QoL. Acceptance and medication adherence are the most influential shared drivers of risk, while physical activity, age, and educational attainment contributed to domain-specific heterogeneity.