Title: Predictive modeling approaches to personalized medicine: A comparison of regression-based methods
Authors: David van Klaveren - Erasmus MC University Medical Center (Netherlands) [presenting]
Abstract: The benefits and harms of medical treatments vary substantially between individual patients. Predictive modeling approaches to personalized medicine are designed to predict the benefit of one treatment over another for an individual patient. A regression model including many interactions between treatment and patient risk factors may be an obvious choice for predicting treatment benefit. However, simulated data will be used to show that including fewer treatment interactions often leads to better treatment benefit predictions. This will be further illustrated with the recently proposed Syntax Score II (SSII)-2020 which was developed to predict the difference in 10-year mortality when treating complex coronary artery disease patients with heart bypass surgery rather than coronary stenting. Cox regression was first used in the SYNTAX trial data $(n=1,800)$ to develop a prognostic index (PI) for mortality over a 10-year horizon consisting of 7 clinical predictors of mortality. Second, a Cox model was fitted which included the treatment, the PI and pre-specified treatment interactions with type of disease and with anatomical disease complexity. In contrast to its more flexible predecessor SSII-2013, SSII-2020 was well calibrated for treatment benefit at 10 years post-procedure, both at cross-validation in the same data and at external validation in new data.