Interpretable machine learning for rare events: an inverse probability weighted application to understanding...

Large administrative datasets enable investigation of rare outcomes, but accurate individual-level prediction remains challenging. Emerging evidence suggests that glucagon-like peptide-1 receptor agonists (GLP-1 RAs), used to treat diabetes and obesity, may reduce alcohol craving and intake. We examined predictors of incident alcohol-related disorder (ARD) using inverse probability-weighted (IPW) machine learning models, focusing on GLP-1 RA use.

We conducted a retrospective cohort study using the MarketScan Multi-State Medicaid Database (N = 96 460). Type 2 diabetic adults aged 18-64 years with GLP-1 RA prescription coverage (2022-2023) and no ARD in 2022 were included. Baseline variables included demographics, social risk factors (Z-codes), comorbidities, insurance characteristics, and GLP-1 RA use. Propensity scores were used to derive stabilized IPW weights to adjust for confounding in GLP-1 RA exposure. An Extreme Gradient Boosting model with IPW was trained using a 70/30 train-test split with five-fold cross-validation and probability calibration. Performance was evaluated using Area Under the Receiver Operating Characteristics Curve (AUROC), recall, and precision. Model interpretability was assessed using SHapley Additive exPlanations (SHAP).

GLP-1 RAs were used by 21.7% of individuals; 2.1% (N = 2004) developed incident ARD. The model demonstrated modest discrimination (AUROC = 0.67). Recall and precision were 0.28 and 0.05, respectively. SHAP analysis ranked GLP-1 RA use as the sixth most influential predictor and showed an association with lower predicted ARD risk.

Predicting rare outcomes remains difficult, particularly for identifying positive cases. Interpretable machine learning identified GLP-1 RA use as an important feature associated with lower predicted ARD risk, warranting further investigation.