Unraveling Cocaine and Heroin Addiction Patterns: A Robust LLM Approach Using iRISA Elements to Analyze Short and Spontaneous Speech Samples

This work presents a robust method for characterizing addiction patterns in individuals with cocaine and heroin use disorders by analyzing short, spontaneous speech samples using a large language model (LLM) framework. This architecture is designed to identify key elements of the Impaired Response Inhibition and Salience Attribution (iRISA) theoretical model and use this framework to assess whether features related to iRISA can inform drug use metrics, without directly evaluating the models’ inference of iRISA elements against human gradings. Specifically, i RISA c aptures d isruptions i n self-regulation, inhibitory control, and the attribution of salience to drug-related versus non-drug-related stimuli. Our analysis revealed significant correlations between iRISA elements and substance use patterns, with these associations being more pronounced when individuals talk about the positive consequences of quitting drugs (PC) than the negative consequences of drug use (NC). Prediction models utilizing outputs from the iRISA LLM framework demonstrated that integrating features from both PC and NC significantly improved predictive accuracy, especially for variables such as days of abstinence (r up to 0.58), withdrawal symptoms (r up to 0.41), and dependence severity (r up to 0.42). These findings highlight the potential of this approach to provide in-depth, data-driven insights into addiction, bridging the gap between computational linguistics and clinical substance abuse research, with significant i mplications f or c linical interventions.