Risk Prediction of Long-term Outcomes and Utilization After Percutaneous Coronary Intervention

Tuesday, June 25, 2019
Exhibit Hall C (Marriott Wardman Park Hotel)

Presenter: Samuel Savitz

Co-Authors: Sue Hee Sung; Thomas Leong; Matthew Solomon; Alan Go

Background: Risk calculators for clinical outcomes and utilization have become increasingly common tools to assist with patient management. We aimed to improve a risk calculator for predicting long-term outcomes and utilization after percutaneous coronary intervention (PCI) within a large, diverse community-based population. Potential strategies for improving the risk calculator included: 1) adding predictors that have not typically been used in clinical risk calculators, and 2) using alternative machine learning approaches.

Methods: The original risk calculator used data from 2008-2012 for patients who received PCI within Kaiser Permanente Northern California, a large, integrated healthcare delivery system. The calculator predicted 1-year outcomes for death, coronary heart disease (CHD)-related hospitalization and emergency room (ER) visits, and repeat revascularization. The original model used Cox proportional hazards (Cox PH) regression with backwards selection. Novel predictors that were considered included: 1) area-based socioeconomic status (SES), 2) access to care, and 3) patient behavior-related measures. The contribution of these variables was evaluated by examining change in the C-index and net reclassification index (NRI). Alternative modeling approaches included: 1) Cox PH with lasso regression, 2) Cox PH with ridge regression, 3) Cox PH with elastic net regression, and 4) conditional inference trees. The improvement in predictive discrimination with different modeling approaches was assessed using changes in the C-index.

Results: The NRI results showed small, significant improvements from adding the access to care measures, but no improvement for the other domains. However, there was a consistent pattern across all three domains of only small, non-significant improvement in the C-index after including the addition potential predictors. Alternative methodological approaches also resulted in only small, non-significant improvement in the C-index for any approach or outcome.

Conclusion: SES, access to care and behavior-related variables and alternative analytic approaches did not materially improve discrimination of an existing risk calculator for PCI patients. Variables that measure regional medical supply and SES may be less important predictors in the context of an integrated care delivery system. Improving risk prediction may still be possible by identifying additional variables more related to patient clinical characteristics or implementing more advanced modeling approaches such as neural networks.