Limited utility of consumer wearable APIs for predicting substance use and craving, despite model personalization

Abstract We explored the feasibility of predicting, using individually-tailored artificial intelligence (AI) models, substance use and craving from Fitbit data collected in free-living conditions. Most substances examined acutely raise heart rate, and craving is known to often involve autonomic arousal. Because heart rate responds to stress, exertion, and substance use differently between individuals, we sought to study in an exploratory experiment whether individualized artificial intelligence (AI) models could learn person-specific patterns that distinguish substance-related heart rate changes from other arousal when conditioned on current activity level. We also examined whether per-participant data coverage and participant adherence to the data collection protocol could explain gaps in performance. Sixteen participants contributed 47 participant–substance–task prediction scenarios (25 substance use, 22 craving) spanning six substances. For each scenario, we trained a model to predict ecological momentary assessment responses, comparing global supervised, global self-supervised, and personalized self-supervised modeling approaches. Discrimination was modest (mean AUROC 0.569–0.617) for all substances and for both craving and use prediction, and differences in data coverage and participant adherence did not explain performance gaps. These findings suggest that heart rate and step count from a consumer wristband carry limited substance-specific signal in isolation, and thus digital intervention research may gain more traction from targeting better-characterized antecedents such as stress than from predicting use or craving directly.