Machine Learning Workflow for Correlating Anxiety and Stress: A SHAP-Based Multimodal Analysis

Many studies employ black-box Machine Learning (ML) models to classify stress and anxiety without examining the underlying biological and phys iological relevance. In this study, we developed an ML workflow based on Shap ley values (SHAP) to interpret black-box models. This approach enables model agnostic visualization of complex relationships between features and predictions while facilitating the explanation of individual predictions, which is essential in clinical practice. To demonstrate the workflow, Random Forest (RF), eXtreme Gradient Boosting (XGBoost), Decision Tree (CART), and Random izedSearchCV-optimized models were trained on the Multilevel Monitoring of Activity and Sleep in Healthy People (MMASH) dataset. Participants were sub grouped into high and low state anxiety groups, where heart rate was predicted within each subgroup. SHAP analysis identified activity type as the most influ ential feature distinguishing anxiety states. Additionally, stress-inducing versus rest activity classification was performed using features including RMSSD, heart rate, and sleep-related measures, with heart rate emerging as the most significant attribute. The RF and XGBoost classifiers achieved an area under the ROC curve (AUC-ROC) exceeding 0.998 for stress-versus-rest classification. Key features influencing anxiety and stress classifications included heart rate, sleep efficiency, melatonin levels before sleep, and cortisol levels after sleep. The results highlight a direct correlation between stress and anxiety, emphasizing the potential of mul timodal data integration for clinical assessments and personalized interventions.