Multi-disease Predictive Analytics: A Clinical Knowledge-aware Approach

AbstractMulti-Disease Predictive Analytics (MDPA) models simultaneously predict the risks of multiple diseases in patients and are valuable in early diagnoses. Patients tend to have multiple diseases simultaneously or develop multiple complications over time, and MDPA models can learn and effectively utilize such correlations between diseases. Data from large-scale Electronic Health Records (EHR) can be used through Multi-Label Learning (MLL) methods to develop MDPA models. However, data-driven approaches for MDPA face the challenge of data imbalance, because rare diseases tend to have much less data than common diseases. Insufficient data for rare diseases makes it difficult to leverage correlations with other diseases. These correlations are studied and recorded in biomedical literature but are rarely utilized in predictive analytics. This article presents a novel method called Knowledge-Aware Approach (KAA) that learns clinical correlations from the rapidly growing body of clinical knowledge. KAA can be combined with any data-driven MLL model for MDPA to refine the predictions of the model. Our extensive experiments, on real EHR data, show that the use of KAA improves the predictive performance of commonly used MDPA models, particularly for rare diseases. KAA is also found to be superior to existing general approaches of combining clinical knowledge with data-driven models. Further, a counterfactual analysis shows the efficacy of KAA in improving physicians’ ability to prescribe preventive treatments.