Getting better outcomes: Using off-policy reinforcement learning to optimize just-in-time interventions targeting weight loss in overweight and obesity
Year: 2026
Type: preprint
Abstract: Overweight and obesity are major risk factors for death and disability, causing 35.8 million years lost due to premature mortality and disability worldwide (Islam et al., 2024). Importantly, current weight-loss interventions have poor outcomes, with only 20% of individuals with overweight or obesity successfully losing weight (Wing & Hill, 2001). One promising new treatment approach is the just-in-time adaptive intervention (JITAI), where patients are monitored in their daily lives using mobile technologies (Nahum-Shani et al., 2017). This data is used to identify when patients are at risk for unhealthy behaviors, and determine which intervention should be delivered to meet a patient’s immediate needs (Nahum-Shani et al., 2017). Increasingly, these decision rules are derived using reinforcement learning (RL) (Sutton & Barto, 2020). However, traditional RL approaches rely on real-time (i.e., online) data collection to learn decision rules for the actions being taken (i.e., on-policy), which can be costly and expose participants to suboptimal actions during early learning (Ghosh et al., 2024; Liao et al., 2020). Therefore, there is an interest in using previously gathered (i.e., offline) data to learn new decision rules that differ from those used during data collection (i.e., off-policy) (Uehara et al., 2022). This project will apply such methods to data from a previously conducted JITAI among individuals with overweight or obesity, with the aim of developing decision rules that maximize percentage weight loss after a combined behavioral weight-loss and JITAI treatment (Goldstein et al., 2021). The analysis will follow the predefined RL framework described in this preregistration, which was developed using data from 20% of the participants.
Institution: KU Leuven
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7131904754