A Reinforcement Learning Approach for Health-Behavioural Recommendations to Reduce Cancer Risk

In the past years, the adoption of specific lifestyle choices, e.g., healthy food, reduced smoke and alcohol consumption, has been revealed to be a crucial tool to reduce cancer mortality in several studies. However, digital health interventions to make people adopt such behaviours require personalization to ensure long-term engagement and effectiveness for specific users. Indeed, their design is challenged by the variability in users’ capabilities, learning patterns, and fatigue dynamics. In the literature, the development of such systems has been held back by the scarcity of available longitudinal, individual-level health behaviour datasets, which do not allow the use of classical reinforcement learning (RL) techniques for learning an effective personalized intervention strategy. In this work, we tackle the intervention recommendation problem using a meta-RL approach to provide personalized intervention suggestions to users and Model-Agnostic Meta-Learning (MAML) with Actor Critic (AC) policies to enable rapid policy adaptation to new users from minimal interaction data. We conduct empirical studies on cross-task adaptation showing that our approach adapts with limited data per user and outperforms both the chosen baselines.