Subject-Aware Detection of Depression from Wrist Actigraphy Using Per-Subject Distributional Aggregation and Interpretable Biomarker Analysis

Depression is a widespread mental health disorder whose early diagnosis is often hindered by the reliance on subjective clinical assessments. This study proposes a subject-aware machine learning framework for detecting depression using wrist-worn actigraphy data, enabling continuous and objective monitoring of daily activity patterns. Unlike many previous studies that evaluate models at the day or window level, the proposed approach employs strict leave-one-subject-out (LOSO) cross-validation, ensuring that no data from the same individual appear in both the training and testing sets and thereby preventing information leakage. The framework extracts time-series features with the Time-Series Feature Extraction Library (TSFEL) from daily actigraphy signals and aggregates them into subject-level representations using distributional statistics. Classical circadian rhythm biomarkers, including interdaily stability (IS), intradaily variability (IV), relative amplitude (RA), cosinor parameters, and activity fragmentation measures, were also investigated to improve interpretability. Seven feature selection techniques and fifteen machine learning classifiers were systematically evaluated. Model selection is performed entirely inside each training partition. The feature selector, the classifier and the number of retained features are chosen together by inner cross-validation over 840 candidate configurations, so the held-out subject contributes nothing to the choice of pipeline. Averaged over five random seeds the nested protocol classifies 45.4 of 55 subjects correctly (82.5%; range 80.0–85.5% across seeds; mean Cohen’s κ = 0.637; and mean sensitivity 74.8%, mean specificity 88.1%). For the median run, 45 of 55, the Wilson 95% confidence interval is 69.7–89.8%. The estimate is above the 58.2% majority-class rate (exact McNemar, p = 0.003), and above the 73% reported by the previously published subject-disjoint evaluation of the same 55-subject cohort; per-subject predictions are not available for that study, so the latter comparison is unpaired. Recording length is a confound in this cohort—used alone as a single feature under the same protocol it reaches 72.7%—and at n = 55 the framework is not separable from that baseline (exact McNemar p = 0.09 to 0.42 across seeds; achieved power 0.40). The single best configuration, Information Gain feature selection with a Bagging classifier, reaches 92.7%. That value is the maximum over a grid of 105 selector–classifier pairs and we report it as such, not as an out-of-sample estimate. Although incorporating circadian biomarkers did not improve classification accuracy (92.7% vs. 90.9%), one bout-fragmentation biomarker, the within-subject mean of daily bout count, was retained in every cross-validation fold; it is, however, a rescaling of the daily Zero Crossing Rate already present in the signature, so the retained features summarize a single daily quantity rather than several independent ones. On this benchmark, subject-level machine learning on wearable actigraphy classifies well above chance under an evaluation protocol that matches the clinical screening setting, but this cohort is too small to establish that the learned representation adds to recording length alone. All results come from one single-site cohort of 55 subjects, and the framework has not been evaluated on an independent dataset. The findings are therefore a reference point for Depresjon, not evidence of generalization to other populations.