A hierarchical model for automatic Neoplasm ICD coding

Abstract

International Classification of Diseases (ICD) codes are used for different management activities in hospitals. Previous researches employed Machine Learning (ML) models for automatic coding to simplify the disease code assignation process; nevertheless, model performance was compromised due to problems with label imbalance and the high number of labels. In the present research, a Support Vector Machine (SVM) model for Neoplasm ICD coding was trained with a dataset previously treated by applying re-sampling methods to mitigate label imbalance issues and increase the model sensitivity. To mitigate the issue with the high number of labels, human body location information contained in the medical records and ICD code descriptions were employed to build a hierarchical model, which improved the performance of a base non-hierarchical model by up to 15%.