This study by Haredasht et al. investigated whether machine learning models could predict treatment retention in patients receiving buprenorphine-naloxone (BUP-NAL) for opioid use disorder, with a focus on incorporating features extracted from unstructured clinical notes using large language models. The researchers used electronic health record data from Stanford Health Care for model development and the NeuroBlu behavioral health database for external validation, training both classification models (Logistic Regression, Random Forest, XGBoost) and survival models (CoxPH, Random Survival Forest, Survival XGBoost) on 206 features, 193 from structured data and 13 psychosocial features extracted from free-text notes using their CLinical Entity Augmented Retrieval (CLEAR) pipeline. XGBoost achieved the highest classification performance with an ROC-AUC of 0.65, while Random Survival Forest and Survival XGBoost reached C-index values around 0.65 in time-to-event analysis. Incorporating LLM-derived features improved model performance across all architectures, with the largest gains in simpler models like Logistic Regression, and SHAP analysis identified features such as chronic pain, liver disease, and major depression as key predictors. The authors acknowledge that while these models demonstrated moderate predictive performance with statistically significant improvements from including unstructured data features, the overall discriminative power remains imperfect.