Predicting individual differences in digital alcohol intervention effectiveness through multimodal data

Area of Science:

• Digital health interventions
• Behavioral science
• Machine learning in healthcare

Background:

• Digital interventions show variable effectiveness for behavior change, such as reducing alcohol consumption.
• Accurate prediction of individual intervention responders versus non-responders is difficult, with prior methods performing only slightly above chance.
• Existing predictive models for intervention effectiveness have limited accuracy (AUC ≈0.60).

Purpose of the Study:

• To develop and validate a novel multimodal approach for predicting the effectiveness of smartphone-delivered alcohol interventions.
• To integrate psychological, social network, and neural data for ex-ante prediction of intervention outcomes.
• To identify key indicators for early detection of non-responders in digital alcohol interventions for young adults.

Main Methods:

• Utilized random forest models integrating multimodal data: psychological assessments, social network data, and neural responses to alcohol cues.
• Applied the approach to smartphone-delivered interventions targeting psychological distancing in young adults across two studies (N=67 and N=114).
• Evaluated model performance using balanced accuracy and Area Under the Curve (AUC), comparing against clinical-utility thresholds.

Main Results:

• The multimodal approach achieved high predictive accuracy in Study 1 (balanced accuracy = 0.71, AUC = 0.87) and replicated in Study 2 (balanced accuracy = 0.68, AUC = 0.68).
• Model performance met clinical-utility thresholds, correctly classifying responders and non-responders 67% of the time.
• Intervention effectiveness was highest for individuals perceiving their peers as moderate but frequent drinkers, suggesting peer perception as a potential indicator.

Conclusions:

• A novel multimodal data integration approach significantly improves the prediction of individual effectiveness for digital alcohol interventions.
• Peer drinking perceptions emerged as a low-burden indicator for identifying non-responders in preventive alcohol interventions among young adults.
• The developed approach offers a promising foundation for adaptive tailoring of digital behavior change interventions in real-world settings.