Safe Reinforcement Learning for Personalized Digital Interventions in Obsessive-Compulsive Disorder (OCD)

Obsessive-Compulsive Disorder (OCD) is a prevalent and disabling psychiatric condition characterized by intrusive thoughts (obsessions) and repetitive behaviors (compulsions). While standard therapies like Exposure and Response Prevention (ERP) and Cognitive Behavioral Therapy (CBT) are effective, their utility is often limited by a lack of real-time personalization, high patient dropout rates, and rigid, one-size-fits-all protocols. To address these critical limitations, this paper introduces a novel framework for Safe Reinforcement Learning (Safe RL) applied to digital psychiatric interventions for OCD. We formally model the therapy process as a Partially Observable Markov Decision Process (POMDP), capturing a patient's latent psychological state from a rich, multimodal data stream, including self-reports (SUDS scores, behavioral logs) and physiological markers (heart rate variability from wearables). An intelligent agent, using a Bi-LSTM network with an attention mechanism, dynamically selects from a range of micro-interventions to optimize for long-term therapeutic outcomes. We employ a Safe Proximal Policy Optimization (Safe-PPO) algorithm to ensure strict adherence to clinically derived safety constraints, such as limiting the intensity of ERP exposures and preventing rapid anxiety escalation. We evaluate our framework in a ۴-week cross-over pilot study with N=۱۵ participants, comparing our Safe RL agent to a static, rule-based system. Results demonstrate a clinically meaningful improvement in symptom reduction (AY-BOCS of -۵.۳ vs -۲.۱), a significant increase in patient adherence (+۱۵%), and a near-zero rate of safety violations. This work not only provides the first empirical evidence for the feasibility of Safe RL in OCD psychotherapy but also offers a robust, scalable, and clinically validated blueprint for a new generation of adaptive AI-driven interventions in digital psychiatry.

کلیدواژه ها:

Reinforcement Learning ، Obsessive-Compulsive Disorder (OCD) ، Digital Psychiatry ، Personalized Therapy ، POMDP ، Bi-LSTM