AI-Powered Behavioral Health Monitoring in Telemedicine Platforms: A 72-Hour Early Warning Framework Using Passive Digital Biomarkers and LSTM Networks

Requiring patients to complete screening questionnaires periodically is a challenge for early detection of deterioration in behavioral health in telemedicine. This study suggests an Artificial Intelligence system that uses Passive Digital Biomarkers, derived from multimodal sensing from remote monitoring devices, to be integrated at the feature-level to predict behavioral health events 72 hours ahead of time. The data used were from the TIHM dataset with 56 patients and 5,096 patient day records, from which we engineered 15 features encompassing activity patterns, sleep architecture, and physiological signals. We benchmarked baseline machine learning models (Logistic Regression, Random Forest, Gradient Boosting) with a temporal Long Short-Term Memory (LSTM) network with an input window of 7 days. The model performance was assessed with area under the receiver operating characteristic curve (AUC-ROC), Brier score for calibration, and statistical comparison through a null-centered bootstrap approximation to DeLong's test. The LSTM model achieved significantly higher discrimination (AUC = 0.808) compared to the best baseline (Random Forest: AUC = 0.711; ΔAUC = −0.097, 95% CI [−0.142, −0.052], p < 0.001). The model gave a median warning lead time of 1.0 day with a tolerable false positive rate at a clinically calibrated threshold to achieve 90% sensitivity. Multimodal passive sensing, coupled with temporal deep learning, holds great promise for more precise and earlier behavioral health monitoring, but requires external validation with a larger sample size. The proposed framework sets a repeatable and privacy-friendly groundwork for proactive telemedicine interventions.