Explainable machine learning for predicting opioid-related aberrant behavior: A multimodal approach using clinical text and structured data.

PURPOSE The rising misuse of opioids, overdose deaths, and opioid use disorder (OUD) associated with chronic pain treatment present a significant public health challenge, continuing to fuel the opioid epidemic. Since opioid-related aberrant behaviours (ORABs) are early indicators of potential misuse, it is crucial to develop advanced predictive models for safe opioid management.

METHODS This study introduces an explainable machine learning framework to predict confirmed aberrant behaviours by integrating clinical text with structured data from the Opioid-Related Aberrant Behavior Detection Dataset (ODD). A multimodal, clinically applicable, and explainable predictive model is developed in this study. The study used GloVe (Global Vectors for Word Representation) embeddings and ClinicalBERT contextual embeddings for EHR text, applied Synthetic Minority Oversampling Technique (SMOTE) for data balancing, and trained various machine learning algorithms.

RESULTS The performance of models was assessed using multiple evaluation metrics. Additionally, SHAP (Shapley Additive exPlanations) was employed to address explainability concerns and to assess feature importance via an ablation study. Results demonstrated that Opioid Risk Ensemble achieved an AUROC of 96.0 % and an accuracy of 98.8 %, indicating that opioids, benzodiazepine prescriptions, and factors related to the central nervous system are significant predictors, as confirmed by SHAP analysis. Furthermore, the Opioid Risk Neural Network, applied with ClinicalBERT embeddings, achieved an AUROC of 98.75 % and an accuracy of 98.47 % on the held-out test set, with SHAP interpretation providing insights into the most influential clinical note terms.

CONCLUSION The multimodal, explainable AI approach will be valuable for modern healthcare decision-making, supporting risk prediction for ORABs.