RL-AMHA: A Reinforcement Learning-Based Adaptive System for Personalized Mental Health Education and Intervention

Effective mental health education requires personalized recommendations and intervention strategies that align with an individual's psychological condition and learning preferences. This study introduces a Reinforcement Learning-Based Adaptive Mental Health Advisor (RL-AMHA) that selects instructional content and self-care interventions using a PPO agent and actor-critic neural architecture. The agent uses self-reports, in-app behavior, and contextual cues from a mobile mental health platform. It was trained on 1,200 anonymized user sessions using log data and validated questionnaire scores, after ethics approval and informed consent. The experiment contrasts RL-AMHA with a static rule-based curriculum and a supervised recommendation model that predicts content based on past clicks. All models share the same content pool and interaction constraints; PPO hyperparameters (learning rate, discount factor, clipping range, batch size) were tweaked on a validation split, and 500 episodes were trained until the average episodic reward converged. RL-AMHA improves engagement rate from 75.2% to 88.1% (+2.9%), user satisfaction from 4.21 to 4.85 (+0.64 on a 5-point scale), and weekly self-care activity frequency from 9.4 to 13.1 (+39%) compared to the baseline. Additionally, it enhances stress-reduction scores by 18.6% and increases continuous engagement time from 21.3 to 29.0 minutes. With minimal implementation costs, RL-AMHA demonstrates scalability, adaptability, and effectiveness for long-term psychological support across mobile health, e-learning, and clinical decision-support environments.