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

TL;DR

A systematic review of 15 studies found that AI models predicting addiction treatment outcomes consistently identify housing instability, psychiatric comorbidity, employment status, craving, and stress as influential factors, but existing models lack external validation and remain too preliminary for clinical decision-making.

What the Study Found

Researchers conducted a systematic review of peer-reviewed studies published between January 2020 and October 2025 to assess how artificial intelligence and machine learning models incorporate psychosocial, behavioral, and social-structural factors when predicting addiction treatment outcomes. The team searched PubMed, Scopus, and Web of Science and identified 15 studies that met rigorous inclusion criteria: they had to apply AI or machine learning to addiction treatment outcomes and explicitly include psychosocial, behavioral, or social-structural predictors.

The 15 studies employed diverse methodologies, reflecting the range of data sources available in addiction research. Some used electronic health records (EHR), administrative claims data, and program-level clinical datasets. Others leveraged ecological momentary assessment (EMA) and digital phenotyping, which capture real-time behavioral and contextual information through mobile devices. A few employed natural language processing and large language model approaches to extract psychosocial information from clinical notes. One study used causal machine learning to analyze randomized controlled trial data, offering a pathway toward causal inference rather than mere correlation.

Across all modalities, a consistent set of predictors emerged as influential for treatment dropout, relapse, overdose risk, and poor engagement. Housing instability appeared repeatedly as a strong predictor, alongside psychiatric comorbidity, employment status, craving intensity, stress levels, legal involvement, prior overdose history, treatment history, medication adherence, and neighborhood disadvantage. Notably, these psychosocial and social-structural variables often added prognostic value beyond medication-related variables, diagnostic categories, and routinely available clinical data. This finding underscores that social context and psychological state are not peripheral to addiction treatment prediction: they are central. EMA and digital phenotyping studies demonstrated the highest short-term predictive accuracy for near-term risk identification, whereas structured EHR-, administrative-, and claims-based models achieved moderate but potentially clinically actionable performance.

However, the authors identified substantial methodological gaps. Across the 15 studies, external validation was limited: most models were tested on the same population from which they were developed, raising questions about transportability to new settings or populations. Few studies formally assessed calibration (whether predicted probabilities match actual event rates), fairness (whether models perform equitably across demographic groups), or reproducibility through open-source code and detailed documentation. These omissions matter because an AI model with high predictive accuracy in one clinic may perform poorly or perpetuate bias when deployed elsewhere.

What This Means for You

If you work in addiction treatment, clinical research, or policy, this review delivers a clear message: the field has identified what matters for predicting treatment outcomes, but the tools themselves are not yet ready for real-world use without substantial additional work. Here's what practitioners should take from this evidence:

Psychosocial assessment is non-negotiable. The consistency with which housing, employment, stress, craving, and psychiatric comorbidity predicted outcomes across diverse study designs and populations suggests these dimensions belong in every treatment planning discussion. They are not luxuries or "nice-to-haves" but core drivers of who will remain engaged and who will relapse. This aligns with evidence-based addiction care frameworks that already emphasize comprehensive psychosocial assessment.

Digital tools show promise for near-term monitoring. EMA and digital phenotyping methods that collect data through mobile applications and wearables showed higher predictive accuracy for short-term risk, particularly for imminent relapse or overdose risk. If such tools are deployed, they should be viewed as real-time risk-flagging systems that prompt direct clinician intervention, not automated decision-makers.

Be skeptical of AI models claiming broad applicability. Until a model has been validated in multiple independent sites with transparent reporting of performance across demographic subgroups, it should be considered preliminary. The current evidence base does not support using any existing model to make clinical decisions without local validation and implementation evaluation.

Co-design is essential. The authors emphasize that future AI tools should be developed alongside clinicians and individuals with lived experience of addiction and recovery. This approach helps ensure that tools address real clinical questions and reflect the values and priorities of the communities they aim to serve, rather than optimizing for metrics that may not align with genuine clinical needs.

Study Details

Sources

Systematic review: Integrating Psychosocial Factors into Artificial Intelligence Models for Predicting Addiction Treatment Outcomes. European Addiction Research. PubMed ID: 42319872