Causal AI Recommendation System for Digital Mental Health: Bayesian Decision-Theoretic Analysis.

BACKGROUND

Digital mental health tools promise to enhance the reach and quality of care. Current tools often recommend content to individuals, typically using generic knowledge-based systems or predictive artificial intelligence (AI). However, predictive AI is problematic for interventional recommendations as cause-effect relationships can be confounded in observed data. Therefore, causal AI is required to compare future outcomes under different interventions.

OBJECTIVE

We aimed to develop a causal AI recommendation system that uses an individual's current presentation, their preferences, and the learned dynamics between domains to rank interventions.

METHODS

We frame the recommendation problem within a Bayesian decision-theoretic framework, whereby a preference ordering of decisions is estimated using the expected utility of outcomes under interventions. The causal processes are assumed to follow a structural causal model, where the posterior distribution of structural causal models is estimated using a Markov chain Monte Carlo method. Expected utilities under interventions are estimated using a do-operation, which estimates the effects of changing a variable on outcomes, while accounting for confounders. We apply our approach to rank domains relating to mental health and well-being as intervention targets for adults (n=619) who used the Innowell Fitness app between September 2021 to September 2023 and completed a questionnaire at 2 time points (1 wk-6 mo from baseline).

RESULTS

The causal AI recommendation system recommends intervention targets as a function of a user's baseline presentation, the causal effects of the intervention on itself and other domains, and the utility function. In our example, psychological distress was typically the optimal intervention target in complex cases where multiple domains were unhealthy at baseline, due to it affecting multiple domains with paths to personal functioning (probability [p] of path; ppath=86%), social support (ppath=92%), sleep (ppath=88%), and physical activity (ppath=86%). The probability of being the optimal intervention target was personal functioning (popt=30%), psychological distress (popt=29%), social support (popt=18%), nutrition (popt=9.6%), substance use (popt=6.7%), sleep (popt=4.5%), and physical activity (popt=2.2%).

CONCLUSIONS

This work illustrates the use of causality and decision-theoretic principles to personalize interventions in digital mental health tools.