Adaptive learning for depression and anxiety prediction using behavioral time series

Introduction : Depression and anxiety affect hundreds of millions of people globally and are typically assessed through episodic self-report instruments that fail to capture within-person temporal dynamics. The objective of this study is to determine whether an incrementally updated, person-specific machine learning model improves discriminative performance for depression and anxiety prediction as individual-level longitudinal data accumulate. Materials and methods : Data were drawn from 514 participant-waves in the GLOBEM dataset across three study cohorts (2019–2021), each contributing passive behavioral signals from smartphones and Fitbit devices alongside weekly Patient Health Questionnaire-4 (PHQ-4) assessments. Five binary classification tasks were defined using PHQ-4 (thresholds ≥3, ≥6, ≥9), Patient Health Questionnaire-2 (PHQ-2) (≥3), and Generalized Anxiety Disorder-2 (GAD-2) (≥3). An Adaptive Random Forest (ARF) was initialized under a leave-one-participant-wave-out framework and updated incrementally with each individual’s observations in temporal order. Performance was evaluated at each personalization step using the AUROC and area under the recall-precision curve (AUPRC) as primary metrics, with bootstrapped confidence intervals from 100 replicates per participant-wave. Statistical inference used a participant-wave-clustered bootstrap with 2000 replicates and Benjamini-Hochberg correction across all 100 comparisons, because days sharing a weekly label are not independent. Results : AUROC rose from near chance at baseline (0.50–0.55) to maxima of 0.88–0.93 across PHQ-4 thresholds, 0.92 for GAD-2, and 0.90 for PHQ-2. AUPRC reached 0.53–0.83 across tasks. Gains relative to the cohort-initialized baseline were significant under clustered inference at every step in every task, all 50 comparisons surviving correction, while 28 of 50 consecutive step-to-step gains were significant and none beyond step 8 reached significance. Static classifiers retrained on the same accumulating individual data plateaued at an AUROC of 0.65 to 0.69, and a cohort-only model without individual updating remained near chance (AUROC of about 0.54), suggesting that the gains were associated with the ARF’s incremental adaptive architecture, which includes adaptive windowing (ADWIN) drift detection. Conclusions : Incremental, person-specific modeling using ARF consistently improves depression and anxiety prediction beyond what static population-level classifiers achieve, with gains detectable after a single personalization step and sustained through the full observation window. Median-replacement perturbation attribution scores identify which behavioral signals most influenced each prediction.