A Comparative Study of Feature Selection Methods for EHR Diagnosis Codes in Opioid Use Disorder Prediction

2026-08-04

AI Summary

This study addresses the challenge of high-dimensional, sparse, and redundant diagnostic features in electronic health records (EHR) for predicting opioid use disorder (OUD). Within a unified preprocessing and evaluation framework, it systematically compares five feature selection methods: recurrence enrichment, neural tangent kernel (NTK)-inspired early gradient sensitivity, LightGBM-SHAP, elastic net, and large language model (LLM)-guided semantic selection. Notably, this work introduces NTK-based sensitivity and LLM-driven semantic guidance to EHR diagnostic code selection for the first time, and establishes a multidimensional evaluation framework encompassing predictive performance, resampling stability, and representation of rare diagnoses. Results show diminishing returns in performance with increasing feature counts; the NTK method achieves the best trade-off between accuracy and stability, while the LLM approach, though modest in standalone performance, yields clinically meaningful complementary signals.

Abstract

Feature selection is a critical step in electronic health record (EHR)-based predictive modeling, where input variables are often high-dimensional, sparse, noisy, and redundant. Large feature sets not only increase computational burden and overfitting risk, but also make model interpretation difficult, leading to limited usefulness in clinical settings. In this study, we focus on diagnosis-related features and compare five feature selection paradigms for opioid use disorder (OUD) prediction: recurrence enrichment, NTK-motivated early gradient sensitivity, LightGBM-SHAP, Elastic Net, and large language model (LLM)-guided semantic selection. We use a unified preprocessing and evaluation framework and assess each method by downstream predictive performance, resampling stability, and representation of infrequent diagnosis codes. Our results demonstrate that performance improves with larger feature budgets with diminishing returns beyond a moderate size. NTK sensitivity provides the best overall balance of accuracy and stability, and LLM-guided selection contributes complementary clinically meaningful signals despite lower standalone performance.

Problem

• feature selection
• electronic health records
• opioid use disorder
• diagnosis codes
• predictive modeling

Innovation

• feature selection
• opioid use disorder
• large language model
• NTK sensitivity
• EHR diagnosis codes