AIM-AHEAD Research Spotlight Series: Showcasing Innovative Research Across the Consortium

Responsible AI for Opioid Use Disorder

Yanmin Gong, PhD
Texas A&M University
South Central Hub

Opioid use disorder (OUD) remains a major public health challenge in the United States, affecting more than six million Americans and contributing to high rates of overdose-related deaths. Although medications for opioid use disorder (MOUD), including buprenorphine, methadone, and naltrexone, are effective in reducing cravings, overdoses, and mortality, retention in treatment remains a persistent challenge. Through participation in the AIM-AHEAD Consortium Development Program, Dr. Yanmin Gong has explored how artificial intelligence and machine learning (AI/ML) methods can be used to better understand factors associated with buprenorphine treatment retention and discontinuation.

Study Design and Methods

Dr. Gong’s research draws on multiple national datasets to examine treatment retention among individuals receiving buprenorphine for OUD. One study used reimbursement claims data from IQVIA prescription records spanning 2006 to 2022 and focused on adults receiving buprenorphine treatment. Predictors included provider availability, mental health services, demographics, psychiatric diagnoses, healthcare utilization, and measures of social vulnerability. Statistical analyses incorporated multinomial logistic regression and stepwise model selection approaches to evaluate treatment duration across patient populations.

Additional analyses applied machine learning methods to examine patterns of treatment discontinuation across demographic, insurance, regional, and community-level factors. These approaches included classification models, regression methods, decision trees, and neural networks. Further work used state-level treatment discharge data from SAMHSA’s Treatment Episode Data Set – Discharges (TEDS-D) to develop multi-task learning models designed to identify both shared and state-specific patterns in MOUD dropout risk.

The research also evaluated fairness across several AI models used to predict treatment retention and premature discontinuation, including approaches intended to reduce performance differences across patient groups.

Key Findings

Across studies, analyses identified variation in treatment retention associated with insurance status, mental health service utilization, age, and community-level social vulnerability. Early medication adherence during the first months of treatment was also identified as an important predictor of long-term retention.

The work further highlighted geographic variation in MOUD access and treatment outcomes across states, reflecting differences in healthcare infrastructure, non-medical health factors, and state-level policies. Multi-task learning approaches achieved stronger predictive performance than both independent state-level models and pooled national models.

Fairness analyses additionally identified differences in model performance across patient populations. Several AI models showed variation in their predictions across demographic groups, including differences in treatment discontinuation and long-term retention estimates. Approaches designed to reduce performance variation across groups improved consistency in some cases, highlighting the importance of continued evaluation of AI models used in MOUD research.

Implications for Research and Clinical Practice

By integrating national claims data, state-level treatment records, and AI/ML approaches, this work contributes to broader efforts to better understand factors associated with MOUD retention and discontinuation. The findings also support ongoing research into how predictive models may be used to examine regional variation, subgroup differences, and treatment-related outcomes in opioid use disorder care.

This research further highlights the importance of evaluating fairness and explainability in healthcare AI applications, particularly when predictive models are applied across varying patient populations.

Recognition and Next Steps

Dr. Gong’s work has resulted in three publications accepted for presentation at leading health informatics and information systems conferences, including CHASE 2025, CHITA 2025, and ICIS 2025. An additional three manuscripts are currently under review at peer-reviewed journals, including the Journal of the American Medical Informatics Association (JAMIA) and JAMA Network Open. Ongoing efforts continue to examine AI/ML approaches for understanding treatment retention, discontinuation, and variation in opioid use disorder care.

DETERMINE: Diabetes prEdicTion through Responsible MachINe lEarning

Feifan Liu, PhD
University of Massachusetts Chan Medical School 
Northeast Hub

Type 2 diabetes (T2D) continues to be a major public health challenge, yet existing prevention approaches often rely on limited definitions of “prediabetes” and may not fully account for broader factors that influence health outcomes. Through participation in the AIM-AHEAD Consortium Development Program (Year 2), Dr. Liu has explored how artificial intelligence and machine learning (AI/ML) can support more accurate, interpretable, and generalizable approaches to T2D risk prediction.

Study Design and Methods

The DETERMINE project (Diabetes prEdicTion through Responsible MachINe lEarning) focused on developing an AI-powered multivariable risk prediction model for five-year T2D risk using 1 year of historical clinical data. The study incorporated a wide range of features, including demographics, diagnoses, medications, laboratory results, body mass index (BMI), blood pressure, and community-level health indicators. Model development also examined how social and community-level factors influenced model performance and fairness metrics.

The study included more than 582,000 patient records after cohort inclusion, exclusion, and feature engineering. Patients under age 18, individuals with pre-existing diabetes diagnoses, and patients using diabetic medications prior to the index visit were excluded from the final modeling dataset. Additional preprocessing steps standardized diagnoses, medications, and laboratory data to support model development.

Researchers evaluated multiple traditional machine learning and deep learning approaches, including CatBoost, XGBoost, Random Forest, Logistic Regression, TabTransformer, ResNet, and MambaTab models. Performance was assessed using measures such as sensitivity, specificity, positive predictive value, and area under the ROC curve (AUROC), alongside fairness-focused evaluation metrics.

Key Findings

Analyses demonstrated that incorporating social and community-level factors influenced both model performance and fairness-related measures. Community-level variables, including income, education, poverty, and unemployment indicators, were associated with differences in model performance across several evaluation metrics.

The work also emphasized the importance of responsible AI development practices in healthcare prediction models. In addition to technical evaluation, the project incorporated guidance from a Community Advisory Board (CAB) composed of healthcare workers, clinicians, government representatives, and diabetes prevention experts. Feedback from the CAB informed the selection of cohort criteria, medication-based exclusions, validation approaches, and laboratory codes during model development.

Implications for Research and Clinical Practice

By integrating clinical and community-level data into AI/ML-based prediction models, this work contributes to ongoing efforts to better understand factors associated with future T2D risk. The findings also support broader research into how responsible and interpretable AI approaches may inform ongoing development of T2D risk prediction models while accounting for factors that influence healthcare outcomes across different settings.

The project further highlights the role of community engagement in healthcare AI research, demonstrating how stakeholder input can inform model design and evaluation.

Recognition and Next Steps

Dr. Liu’s work through the DETERMINE project has contributed to seven papers and abstracts, multiple extramural grant applications, and a funded R01 award in 2024. Additional efforts include a submitted state funding application to the Massachusetts Life Sciences Center and a subsequent R01 application submitted in January 2026.