TITLE:
Transformer-Based Models for Predicting Rapid Cycling in Bipolar Disorder: Integrative Analysis of Digital Phenotyping and Pharmacological Data
AUTHORS:
Rocco de Filippis, Abdullah Al Foysal
KEYWORDS:
Transformer, Bipolar Disorder, Rapid Cycling, Digital Phenotyping, Deep Learning, Synthetic Data, Time-Series Modelling, Mental Health AI
JOURNAL NAME:
Open Access Library Journal,
Vol.12 No.8,
August
20,
2025
ABSTRACT: Bipolar disorder is a severe psychiatric condition characterized by mood instability, and a specific subtype known as rapid cycling presents with frequent mood episodes, complicating diagnosis and treatment. Accurate identification of rapid cycling is critical for early intervention and personalized care strategies. This study explores the feasibility of using Transformer-based deep learning models to predict rapid cycling in bipolar disorder through an integrative analysis of synthetic multimodal data. A simulated dataset comprising 200 virtual patients was created to reflect 30 consecutive days of behavioural and clinical signals, including mood ratings, sleep duration, physical activity, and pharmacological dosage (lithium), combined with static features such as age, sex, body mass index (BMI), baseline mood score, and medication status. Seven comprehensive visualizations were developed to examine temporal trends, distributional properties, and correlations among features. The results show clear separability in mood dynamics between stable and rapid cycling patients, with mood variance emerging as a highly discriminative marker. While physical activity and dosage patterns reflected structured behaviours suitable for modelling, sleep and sex-related variables showed less predictive utility. The mood trajectory and variability plots particularly justified the selection of attention-based architectures like Transformers, which are adept at capturing long-range temporal dependencies. This research provides a foundational pipeline that simulates real-world longitudinal data and prepares it for Transformer model implementation. While the dataset is synthetic, the methodology replicates realistic digital phenotyping workflows. The study underscores the value of combining behavioural dynamics and pharmacological history within AI frameworks to support next-generation mental health diagnostics and personalized treatment planning in bipolar disorder.