TITLE:
A Multimodal Deep Learning Framework for Early Detection, Mood State Classification, and Episode Prediction in Bipolar Disorder
AUTHORS:
Rocco de Filippis, Abdullah Al Foysal
KEYWORDS:
Bipolar Disorder, Deep Learning, Multimodal AI, TCAN, Graph Neural Network, Bayesian Uncertainty, Affective Computing, EHR NLP, Mood Classification, Episode Prediction, Computational Psychiatry, Wearable Bio Signals
JOURNAL NAME:
Open Access Library Journal,
Vol.13 No.5,
May
29,
2026
ABSTRACT: Bipolar disorder (BD) affects approximately 45 million individuals worldwide and is characterized by recurrent episodes of mania, hypomania, and depression, with an average diagnostic delay exceeding seven years from symptom onset. Existing clinical tools are fundamentally reactive, episodic-assessment-based, and ill-equipped to capture the dynamic, multimodal nature of affective instability resulting in suboptimal pharmacological management, high relapse rates, and substantial disability-adjusted life years. We present BD-Net, a unified multimodal deep learning framework integrating 1) a Temporal Convolutional Attention Network (TCAN) for wearable bio signal analysis, 2) BD-BERT, a domain-adaptive transformer pre-trained on 3.2 million psychiatric clinical notes, 3) a Graph Attention Network (GAT-GNN) modelling inter-episode longitudinal dependencies, and 4) a Bayesian deep ensemble providing calibrated uncertainty estimates. BD-Net was trained and validated on a prospective federated cohort of 2847 participants monitored continuously for 18 months, comprising over 140 million bio signal samples and 94,000 clinical encounters. BD-Net achieves 91.3% mood state classification accuracy (AUC = 0.961, Macro F1 = 0.887), outperforming all 14 evaluated baselines. Manic episode prediction yields 88.7% sensitivity and 90.1% specificity with a mean lead time of 4.2 days. The Bayesian layer produces Expected Calibration Error (ECE) = 0.031. In prospective clinical simulation (n = 50 BD-I patients, 6 months), BD-Net reduced false hospitalization recommendations by 34.2% relative to standard screening protocols. BD-Net demonstrates that principled multimodal fusion, longitudinal temporal modelling, and Bayesian uncertainty quantification can deliver clinically meaningful, generalizable predictions for bipolar disorder establishing a new methodological benchmark for computational psychiatry and providing a framework extensible to other affective and neurodevelopmental disorders.Subject AreasPsychiatry & Psychology