Caro Seguel, Juan CarlosGalgano Cabrera, Giorgio Stefano Ian Michele2025-10-292025-10-292025https://repositorio.udec.cl/handle/11594/13308Physical inactivity remains one of the leading preventable causes of non-communicable chronic diseases and premature mortality worldwide. Digital health technologies offer scalable solutions to promote physical activity, yet their effectiveness depends on the system´s capability to adapt to users motivational and contextual states. Reinforcement learning (RL), particularly contextual bandit algorithms, provides a promising framework for such adaptive personalization. However, these algorithms face the cold start problem (CSP) which limits their initial performance due to the lack of user data. This study explores whether theory driven simulated data can mitigate the CSP in training RL systems for personalized physical activity recommendations. A scoping review of empirical studies (included n = 18) on the Integrated Behavior Change model was conducted to extract population-level parameters describing controlled motivation, autonomous motivation, attitude, subjective norms, perceived behavioral control, intention and need for autonomy. These parameters informed the generation of a synthetic dataset, simulating 2,000 virtual users. An e-greedy algorithm was trained using this synthetic dataset and compared to its training in a real-world pilot conducted through the mHealth web app Apptivate with 558 university students. Results indicated strong alignment between synthetic and real behavioral patterns, with both reproducing the expected correlations among IBC constructs. The CB trained on synthetic data improved adherence predictions by approximately 12 percentage points over random allocation and demonstrated convergence patterns similar to those observed in real data. These findings suggest that behaviorally informed synthetic data can provide a feasible pre training environment for adaptive algorithms, reducing dependency on early empirical data and preserving user privacy. Integrating behavioral theory into synthetic data generation constitutes a practical and ethically sound strategy to address the cold start problem in personalized mobile health interventions. Our approach bridges psychological modeling and machine learning, enhancing interpretability, scalability, and the methodological transparency of digital behavior change systems.enCC BY-NC-ND 4.0 DEED Attribution-NonCommercial-NoDerivs 4.0 InternationalReinforcement learningArtificial intelligenceApplication softwareThesisBuena SALUD