TITLE:
A Multimodal Deep Learning Framework for Continuous Mood Monitoring and Episode Prediction in Bipolar Disorder
AUTHORS:
Rocco de Filippis, Abdullah Al Foysal
KEYWORDS:
Digital Phenotyping, Bipolar Disorder, Mood Instability, Smartphone Sensing, Deep Learning, Temporal Convolutional Network, Speech Analysis, GPS Mobility, Bayesian Uncertainty, Ecological Momentary Assessment, Circadian Rhythm, Affective Computing
JOURNAL NAME:
Open Access Library Journal,
Vol.13 No.5,
May
29,
2026
ABSTRACT: Bipolar disorder does not announce itself with a clean clinical signal. It builds in sleep that fragments days before a manic break, in accelerating movement recorded through a wrist sensor, in speech that picks up tempo before the patient notices anything has changed. Capturing these signals passively and continuously is the promise of digital phenotyping. Delivering on it requires machine learning architectures capable of integrating heterogeneous, irregularly sampled, high-dimensional sensor streams with the contextual knowledge of self-reported mood and circadian biology. We introduce MoodSense-Net, an end-to-end multimodal deep learning framework that fuses smartphone accelerometery, sleep metrics, GPS mobility traces, speech acoustics, and ecological momentary assessment (EMA) data to continuously monitor and predict mood instability in bipolar disorder. The model integrates a Multi-Scale Temporal Convolutional Network (MS-TCN) for accelerometery, a Bidirectional LSTM for sleep staging, a Rhythm CNN for GPS circadian patterns, a Speech-BERT module for acoustic analysis, and a cross-modal transformer fusion layer with a Bayesian deep ensemble output for uncertainty-calibrated predictions. Trained and validated on a prospective cohort of 1847 participants monitored continuously for 12 months encompassing over 26 million sensor samples and 312,000 EMA responses, MoodSense-Net achieves 5-class mood state classification accuracy of 92.7%, AUC-ROC of 0.963, and macro-F1 of 0.891. Episode onset prediction at the 7-day horizon yields sensitivity of 89.1% and specificity of 90.3% for manic episodes, with a mean prediction lead time of 5.1 ± 1.9 days. The Bayesian ensemble achieves Expected Calibration Error (ECE) of 0.028. MoodSense-Net establishes a new methodological benchmark for passive monitoring in computational psychiatry, providing a validated, deployable architecture for continuous bipolar mood instability surveillance.