Early Detection of Depression Using Wearable Sensor Data: A Comparative Study of Machine Learning Models and Explainable AI Techniques

Depression screening relies largely on periodic self-report instruments, which can miss gradual behavioural change between assessments. Wearable devices continuously record sleep, physical activity and heart rate, yet published studies differ widely in data type, labelling and validation rigour, and evidence on participant-independent generalisation remains limited. This paper presents a complete, pre-specified study design for early depression-risk screening support from wearable-derived data, with a leakagecontrolled comparative framework for classical machine learning, deep learning and statistical time-series baselines, and a multi-method explainable-AI analysis (SHAP, LIME, permutation importance, partial dependence). We audit four public or access-controlled datasets (Depresjon, GLOBEM, RADAR-MDD, StudentLife), document which required modalities each contains, and assign them to separate, non-pooled experiments. We define sleep, activity, physiological and circadian features mathematically, specify nine experiments (A–I) covering modality ablation, model comparison, cross-subject and temporal generalisation, and robustness to missing or noisy data, and describe calibration, cluster-bootstrap confidence intervals and multiplicity-corrected significance testing. No experiments have been executed for this manuscript: all result tables are marked templates, and every quantitative statement about prior work is attributed to its source. The framework targets research screening support, not clinical diagnosis.