A Deep Learning Based Approach to Automate Clinical Coding of Electronic Health Records

The medical records in different electronic formats, such as handwritten notes, diagnosis summaries, lab reports, electronic pdfs, etc., contain valuable information that can be used for various medical purposes. These health records are currently coded manually or semi-automated to assign clinical codes (ICD-codes) for clinical research and analytics. This process is very time-consuming, expensive, and error-prone. This paper presents a method for automated clinical coding of electronic health records (EHRs) given the patient diagnosis summary and other medical-related documents. The presented method uses natural language processing (NLP) techniques, which capture knowledge from the free-text diagnosis descriptions, do the text matching and semantic mapping, and translate diagnosis descriptions into clinical codes. We develop one baseline Word2vec and cosine similarity hybrid model, a transformer encoder model, and a BERT (Bidirectional Encoder Representations from Transformers) model for the automated clinical coding. The presented models are evaluated using a publicly available Medical Information Mart for Intensive Care III (MIMIC-III) dataset. The used dataset consists of various patient diagnosis descriptions and corresponding ICD-9 codes. The experimental results show that the presented BlueBERT based automated clinical coding model produced an AUC (area under ROC curve) value of 98.9% for the top-10 ICD codes prediction. On the full MIMIC-III dataset, the transformer model produced an accuracy of 76.8%, a precision of 61.02%, a recall of 47.22%, a f1-score of 53.2%, and an AUC value of 92.1%. The hybrid baseline model and another used transformer encoder model also showed promising results.