AI Reads Between the Lines to Predict Opioid Treatment Success

Keeping patients on life-saving medication for opioid use disorder is a critical challenge in modern medicine. A new study has demonstrated that Artificial Intelligence can significantly improve predictions of which patients will stick with buprenorphine-naloxone therapy for at least six months.

Standard Electronic Health Records (EHRs) act like strict forms, capturing structured data such as dates and dosages but missing the nuance of a consultation. To bridge this gap, researchers utilised Large Language Models (LLMs)—sophisticated algorithms designed to understand and generate human text—to mine unstructured clinical notes. They extracted 13 specific clinical and psychosocial features, including mentions of chronic pain, liver disease, and major depression.

When these ‘hidden’ features were fed into machine learning models like XGBoost, accuracy in predicting treatment retention improved across the board. The study achieved a classification performance score (ROC-AUC) of 0.65, with the text-derived data providing a notable boost to simpler models. This approach highlights valuable risk factors that check-box medical records often obscure. The team has since developed an interactive web tool, allowing clinicians to utilise these personalised risk profiles in real-time to better support their patients’ recovery journeys.