Combining knowledge and data driven insights for identifying risk factors using electronic health records

The ability to identify the risk factors related to an adverse condition, e.g., heart failures (HF) diagnosis, is very important for improving care quality and reducing cost. Existing approaches for risk factor identification are either knowledge driven (from guidelines or literatures) or data driven (from observational data). No existing method provides a model to effectively combine expert knowledge with data driven insight for risk factor identification.

We present a systematic approach to enhance known knowledge-based risk factors with additional potential risk factors derived from data. The core of our approach is a sparse regression model with regularization terms that correspond to both knowledge and data driven risk factors.

The approach is validated using a large dataset containing 4,644 heart failure cases and 45,981 controls. The outpatient electronic health records (EHRs) for these patients include diagnosis, medication, lab results from 2003-2010. We demonstrate that the proposed method can identify complementary risk factors that are not in the existing known factors and can better predict the onset of HF. We quantitatively compare different sets of risk factors in the context of predicting onset of HF using the performance metric, the Area Under the ROC Curve (AUC). The combined risk factors between knowledge and data significantly outperform knowledge-based risk factors alone. Furthermore, those additional risk factors are confirmed to be clinically meaningful by a cardiologist.

We present a systematic framework for combining knowledge and data driven insights for risk factor identification. We demonstrate the power of this framework in the context of predicting onset of HF, where our approach can successfully identify intuitive and predictive risk factors beyond a set of known HF risk factors.