Explainable AI for Clinical Text-to-ICD-10 Mapping Using Biomedical Language Models

Abstract

Medical coding is very important for billing, record keeping, and data analysis in health care organizations. But doing the by manually takes a long time and can lead to mistakes. In this, an automated medical coding system that helps in suggesting correct ICD-10 codes from clinical text. To understand medical terms, the approach combines biomedical models BioBERT and PubMedBERT with natural language processing methods. Before the start of transforming into semantic representations, the input clinical text is cleaned and normalized. It helps in collecting key terms as well as the text's meaning. A similarity-based retrieval technique with cosine similarity and FAISS for faster searching is used to find the most relevant codes. To help the users in comparing and choosing the best option, the system gives a list of the top matching ICD-10 codes along with similarity scores. To understand the results easily, it has an interactive interface and an easy-to-understand explanation structure. A Strong results was noticed in the experiments which did on the ICD-10 dataset. Relevant codes are typically ranked very high, as model shows 94% MRR score for top predictions and approximately 95% Recall when considered the top 5 results. Overall, it shows how AI can increase accuracy and decrease manual work in medical coding while maintaining a clear and user-friendly process. In the future, the system could be extended to accommodate additional coding standards, such as CPT and HCPCS.