Predicting individual differences in digital alcohol intervention effectiveness through multimodal data

Digital interventions can change behaviors like alcohol use, but effectiveness varies widely across individuals. Accurately identifying non-responders-i.e., those least (vs. most) likely to change their behavior-before intervention delivery is difficult. Individual intervention effectiveness predictions from prior studies perform only slightly above chance (e.g., AUC ≈0.60; balanced accuracy ≈0.60). We present a novel approach integrating multimodal data across theory-driven domains-including psychological assessments, social network data, and neural responses to alcohol cues-to make ex-ante predictions about the effectiveness of smartphone-delivered alcohol interventions targeting psychological distancing in young adults (Study 1: N = 67; Study 2: N = 114). Demonstrating the feasibility of this approach, random forest models predicted individual differences in intervention effectiveness (Study 1: balanced accuracy = 0.71, 95% CI: 0.69-0.73, p = .020; AUC = 0.87, 95% CI: 0.85-0.88, p = .020) and replicated in a an external test sample (Study 2, balanced accuracy = 0.68; AUC = 0.68, 95% CI: 0.54-0.82), meeting clinical-utility thresholds from prior digital health studies (balanced accuracy = 0.67; correctly classifying (non)responders 67% of the time). Interventions were most effective for participants who perceived their peers as moderate but frequent drinkers. Peer drinking perceptions may serve as a low-burden indicator to support early identification of non-responders in preventive alcohol interventions among young adults. Future work can apply and extend the multimodal approach developed here for adaptive tailoring of digital behavior change interventions in real-world settings.