Abstract 14957: Personalized Disease Networks in Predicting Cardiovascular Outcomes

Introduction: In epidemiological studies, outcomes are usually predicted using standard statistical techniques including logistic regression and Cox proportional hazard models. These models do not include information on temporal sequences of risk factors and the progression from risk factors to outcomes.

Hypothesis: We hypothesize that the use of personalized disease networks (PDN) will improve the prediction of outcomes.

Methods: We developed a novel method for building patient-level personalized disease networks for the prediction of cardiovascular outcomes. We built one network for each patient that reflects the sequence of occurrence of risk factors, procedures, and co-morbidities in predicting cardiovascular outcomes such as all-cause death or the development of heart failure (Figure displays a network for a given patient). The color brightness of each oval and associated number indicates the temporal order of these conditions. The network structure data were used to predict outcomes e.g. heart failure, death, etc.

Results: We analyzed a dataset of 4,293 African American patients with cardiovascular disease from the Myocardial Infarction Data Acquisition database, a statewide database including all non-federal admissions for myocardial infarction in the state of New Jersey. The ability of the network in predicting all-cause mortality was much better than that of a standard Cox model (R²=55% vs. 20% for the standard Cox model).

Conclusions: By incorporating temporal sequence information, personalized disease networks improve the prediction power for cardiovascular outcomes compared to standard statistical models.