A Hierarchical Fine-Grained Deep Learning Model for Automated Medical Coding

During a patient's visit, hospital staff record the patient's condition, including measurements, observations, diagnoses, and treatment. These records, known as doctor's notes, are usually kept in natural language and must be standardized for consistent communication during treatment and billing. Medical coding is the procedure of applying standardized, universal codes to doctor's notes for efficient and effective record keeping. Due to the complexity of healthcare and the breadth of natural languages, this has been a difficult task requiring significant training and resources. In this paper, we present a fine-grained, evidence-based hierarchical deep learning model for automated medical coding. Instead of treating the doctor's notes as a whole, the proposed method processes the coding of each diagnosis separately. For each diagnosis, the knowledge encoded in ontologies is used to extract semantically related sentences from the doctor's notes. Then, a hierarchical deep learning classifier processes the diagnosis and its semantically related sentences and generates the medical code prediction. This approach not only achieved promising results in our experiments, but also provided a high degree of explainability and trustworthiness in the results, which is a key factor in the adoption of medical solutions.