Pseudo-Relevance Feedback with Deep Learning for Automated ICD-10 Coding

Abstract

Thailand's decentralized healthcare system faces ongoing difficulties with ICD-10 coding, including staffing shortages and coding errors that affect reimbursement processes and clinical operations. We developed a deep learning framework that combines pseudo-relevance feedback (PRF) with the Rocchio algorithm to improve how clinical notes are mapped to ICD10 codes. Our approach was tested using the MIMIC-IV-Note dataset, where we achieved a micro F1-score of 59.5 and Precision@8 of 70.8 after applying iterative PRF improvements. We also built a web-based tool that provides real-time coding predictions through an easy-to-use interface, automated data processing, and API connections that work with existing hospital systems. The results demonstrate that our approach provides hospitals with an effective tool for more accurate medical record coding while lightening the workload for clinical staff. Because the system integrates smoothly with existing healthcare IT systems, it presents a realistic option for enhancing coding efficiency in actual clinical environments.