Leveraging Large Language Models to Enhance Machine Learning Interpretability and Predictive Performance: A Case Study on Emergency Department Returns for Mental Health Patients

Objective: To evaluate whether integrating large language models (LLMs) with traditional machine learning approaches improves both the predictive accuracy and clinical interpretability of ED mental health returns risk models. Methods: This retrospective cohort study analyzed 42,464 ED visits for 27,904 unique mental health patients at an Academic Medical Center in the deep South of the United States between January 2018 and December 2022. Main Outcomes and Measures: Two primary outcomes were evaluated: (1) 30 days ED return prediction accuracy and (2) model interpretability through a novel retrieval-augmented generation (RAG) framework integrating SHAP (SHapley Additive exPlanations) values with contextual clinical knowledge. Results: The proposed machine learning interpretability framework, leveraging LLM, achieved 99% accuracy in translating complex model predictions into clinically relevant explanations. Integration of LLM-extracted features enhanced predictive performance, improving the XGBoost model area under the curve (AUC) from 0.73 to 0.76. The LLM-based feature extraction using 10-shot learning significantly outperformed traditional approaches, achieving an accuracy of 0.882 and an F1 score of 0.86 for chief complaint classification (compared to conventional methods with an accuracy range of 0.59 to 0.63) and demonstrating accuracy values ranging from 0.65 to 0.93 across multiple SDoH categories, underscoring its robust performance in extracting features from clinical notes. Conclusions and Relevance: Integrating LLMs with traditional machine learning models yielded modest but consistent improvements in ED return prediction accuracy while substantially enhancing model interpretability through automated, clinically relevant explanations. This approach offers a framework for translating complex predictive analytics into actionable clinical insights.