Prediction of Stress-Induced Substance Use via a Multimodal Data Acquisition Framework and an Ensemble Machine Learning Model

Abstract

The importance of understanding stress-related substance use episodes is a key research area, as it offers new insight into the association between stress and substance use behaviours. Existing methods are typically limited to analysing self-reported or isolated physiological signals, without providing real-time contextual analysis. This study proposes a multimodal approach for data acquisition and an ensemble machine learning technique to analyse how stress affects an individual’s substance use behaviour. The dataset was acquired using wearable sensors and a psychological stress assessment questionnaire with a craving intensity scale. A dataset consisting of 1,325 instances was acquired from 53 voluntary participants in certified recovery environments in North-East India. Furthermore, a fusion-based Ensemble Random Forest Machine Learning (ERFML) model is proposed to analyse an integrated dataset of physiological, psychological, and craving features. It has been observed from the experiments that the proposed model has higher prediction accuracy (AUC=0.95) for stress-induced substance use. Furthermore, a positive correlation between stress and craving was identified (r=.73). Likewise, heart rate and electrodermal activity features reflect physiological imbalances that take place during stress(r=0.60). Approximately 70% of craving instances were observed to co-occur with elevated stress levels within a defined temporal window, where stress and craving measurements were aligned at the same or preceding time step. The model is further validated using Leave-One-Subject-Out cross-validation to ensure subject-independent generalisability. This fusion-based model provides enhanced prediction of addiction vulnerability prediction using wearable biosignals and psychological assessments to manage relapse and recovery during treatment.