Comparing Feature Selection for Opioid Use Disorder Prediction

Predicting Opioid Use Disorder (OUD) from Electronic Health Records (EHR) is challenging due to the high dimensionality, sparsity, noise, and redundancy of diagnosis-related features. This research conducted a comparative analysis of five distinct feature selection methods: recurrence enrichment, NTK-motivated early gradient sensitivity, LightGBM-SHAP, Elastic Net, and large language model (LLM)-guided semantic selection. The goal was to assess how each method impacts predictive performance, resampling stability, and the ability to represent less common diagnosis codes. The study found that increasing the number of features generally improved performance, though with diminishing returns beyond a certain point. Among the methods tested, NTK sensitivity demonstrated the most favorable balance between predictive accuracy and stability. Interestingly, while LLM-guided semantic selection did not achieve the highest standalone performance, it provided valuable, clinically meaningful signals that complemented the other methods, highlighting the potential for hybrid approaches in medical predictive modeling.