Leveraging hierarchy in medical codes for predictive modeling

ICD-9 codes are among the most important patient information recorded in electronic health records. They have been shown to be useful for predictive modeling of different adverse outcomes in patients, including diabetes and heart failure. An important characteristic of ICD-9 codes is the hierarchical relationships among different codes. Nevertheless, the most common feature representation used to incorporate ICD-9 codes in predictive models disregards the structural relationships. In this paper, we explore different methods to leverage the hierarchical structure in ICD-9 codes with the goal of improving performance of predictive models. We compare methods that leverage hierarchy by 1) incorporating the information during feature construction, 2) using a learning algorithm that addresses the structure in the ICD-9 codes when building a model, or 3) doing both. We propose and evaluate a novel feature engineering approach to leverage hierarchy, while simultaneously reducing feature dimensionality. Our experiments indicate that significant improvement in predictive performance can be achieved by properly exploiting ICD-9 hierarchy. Using two clinical tasks: predicting chronic kidney disease progression (Task-CKD), and predicting incident heart failure (Task-HF), we show that methods that use hierarchy outperform the conventional approach in F-score (0.44 vs 0.36 for Task-HF and 0.40 vs 0.37 for Task- CKD) and relative risk (4.6 vs 3.3 for Task-HF and 5.9 vs 3.8 for Task-CKD).