Fine-Grained ICD Code Assignment Using Ontology-Based Classification

Assigning International Classification of Diseases (ICD) codes based on doctors' clinical diagnoses has historically been a difficult task performed by highly trained clinical coding experts. Recently, attempts have been made to use machine learning techniques at a coarse-grained level to automatically generate lists of medical codes from doctors' notes; however, the results are often difficult to interpret and validate. In this paper, we propose a fine-grained approach that focuses on one diagnosis at a time. We use ontology-based human knowledge to extract semantically related sentences from doctor's notes to support the use of deep learning for reliable training and classification. This fine-grained deep learning approach significantly reduces training load and improves scalability while providing users with a rationale for ICD code prediction. To demonstrate the effectiveness and advantages of our approach, we apply it to the MIMIC-III dataset and show how ICD-9 codes can be automatically assigned to clinical diagnoses.