Integrating Psychosocial Factors into Artificial Intelligence Models for Predicting Addiction Treatment Outcomes: A Systematic Review.

Abstract

Introduction: The integration of artificial intelligence (AI) into addiction research has expanded rapidly, yet it remains unclear how psychosocial, behavioral, and social-structural determinants are incorporated into predictive models of addiction treatment outcomes. Because recovery is strongly shaped by psychological, social, and environmental context, assessing how AI approaches operationalize these dimensions is essential for developing clinically meaningful and equitable tools.

Methods: We conducted a systematic review of peer-reviewed studies published from January 1, 2020 to October 30, 2025 across PubMed, Scopus, and Web of Science. Eligible studies applied AI or machine learning (ML) models to addiction treatment outcomes and explicitly included psychosocial, behavioral, or social-structural predictors. Two reviewers independently screened studies, extracted data, and evaluated methodological quality using Joanna Briggs Institute and Cochrane Risk of Bias 2 domain structures. The protocol was prospectively registered on the Open Science Framework and in PROSPERO.

Results: Fifteen studies met inclusion criteria, including electronic health record (EHR), administrative-, claims-based, program-level clinical models and psychosocial assessment datasets, digital phenotyping/ecological momentary assessment (EMA) studies, natural language processing/large language model approaches, and one causal ML analysis of randomized controlled trial data. Across modalities, models consistently identified housing instability, psychiatric comorbidity, employment status, craving, stress, legal involvement, prior overdose, treatment history, medication adherence, and neighborhood disadvantage as influential predictors of treatment dropout, discontinuation, overdose risk, relapse, or poor engagement - often adding prognostic value beyond medication-related, diagnostic, and routinely available clinical variables. EMA and digital phenotyping showed the highest short-term predictive accuracy for near-term risk prediction, whereas structured EHR-, administrative-, claims-based, and program-level clinical models achieved moderate but clinically actionable performance. Methodological quality was moderate overall, with limited external validation and infrequent assessment of calibration, fairness, transportability, or reproducibility practices.

Conclusion: Current evidence indicates that psychosocial, behavioral, and social-structural determinants are central to AI-based prediction of addiction treatment outcomes. Although findings are promising, existing models remain preliminary and should not yet guide clinical decisions without external validation and implementation evaluation. Future work should prioritize multisite validation, transparent reporting, fairness evaluation, and co-development with clinicians and individuals with lived experience to ensure that AI tools strengthen person-centered and equitable addiction care.