Title: Construction of risk scores for cardiac surgical patients: Dos and don’ts

Biography:

Siew-Pang Chan has completed his PhD in Medical Decision Analysis. Currently he is an Assistant Director at the Department of Medicine, Yong Loo Lin School of Medicine, National University of Singapore. He is also an Honorary Fellow of the College of Science, Health & Engineering, La Trobe University, where he taught from 2010 to 2013. He strongly advocates the applications of Bayesian techniques and structural equation models in biomedical research and has published over 120 papers in peer-reviewed journals.

Abstract

Constructing risk scores for cardiac surgical patients commands a high level of intellectual attention, fueled in part by its multi-disciplinary nature and the ever-emerging evidence from cardiac research. The search for an “ideal” model is pursued as a clinical and methodological undertaking, but the recent developments in statistics and data science have not been appropriately infused into the endeavor. While the celebrated EuroSCORE II, STS and ACEF scores continue to serve the needs of the scientific community, it is timey to re-examine the underlying methodological issues and to shed light on the pitfalls of the current practice. A risk score is only reasonable, useful and comprehensive if it is perioperative in nature, as the occurrence of death (Mortality) could be explained by demographics (Demo), preoperative (PreOp), intraoperative (IntraOp) and postoperative (PostOp) factors acting individually and jointly, directly and indirectly. The structural equation model (SEM) is thus proposed for risk-score construction, in view of the perioperative nature and the complexity in data structures. The decision trees could also be applied for model selection, in terms of identification of relevant predictors and variable discretization. The pitfalls of the conventional methodology, based on logistic regression for estimation and prediction, Hosmer-Lemeshow test for goodness of fit and c-statistics for assessment of predictive accuracy, are also discussed.