Explainable ICD Code Assignment Using Knowledge-Based Sentence Extraction and Deep Learning

Medical coding involves the assignment of standardized codes like those of the International Classification of Diseases (ICD) to patient records such as doctor’s notes. Traditionally, medical coding has been performed by trained professionals and has incurred significant costs. Recent research efforts have produced good classification results but often lack in explainability and trustability of the coding results. This paper introduces a novel fine-grained evidence-based approach for medical coding, which improves explainability and trustability by extracting text related to a given diagnosis based on existing ontologies. Then, the given diagnosis along with the extracted sentences are treated as a fine-grained data point for deep training and prediction. Since the approach tracks verifiable human knowledge, the extracted sentences based on the knowledge can be used as evidence for ICD code classification. To demonstrate the effectiveness and efficiency of the approach, we used two subsets of the Medical Information Mart for Intensive Care III (MIMIC-III) dataset for case studies. The experimental results show that the classifier outperforms existing approaches and has a strong ability to distinguish between the different uses of similar terminologies.