Search

Machine learning algorithms are increasingly used in the clinical literature, claiming advantages over logistic regression. However, they are generally designed to maximize the area under the receiver operating characteristic curve. While area under the receiver operating characteristic curve and other measures of accuracy are commonly reported for evaluating binary prediction problems, these metrics can be misleading. We aim to give clinical and machine learning researchers a realistic medical example of the dangers of relying on a single measure of discriminatory performance to evaluate binary prediction questions. Prediction of medical complications after surgery is a frequent but challenging task because many post-surgery outcomes are rare. We predicted post-surgery mortality among patients in a clinical registry who received at least one aortic valve replacement. Estimation incorporated multiple evaluation metrics and algorithms typically regarded as performing well with rare outcomes, as well as an ensemble and a new extension of the lasso for multiple unordered treatments. Results demonstrated high accuracy for all algorithms with moderate measures of cross-validated area under the receiver operating characteristic curve. False positive rates were < 1%, however, true positive rates were < 7%, even when paired with a 100% positive predictive value, and graphical representations of calibration were poor. Similar results were seen in simulations, with the addition of high area under the receiver operating characteristic curve ( > 90%) accompanying low true positive rates. Clinical studies should not primarily report only area under the receiver operating characteristic curve or accuracy.

Objectives: STAT Mortality Categories (developed 2009) stratify congenital heart surgery procedures into groups of increasing mortality risk to characterize case mix of congenital heart surgery providers. This update of the STAT Mortality Score and Categories is empirically based for all procedures and reflects contemporary outcomes., Methods: Cardiovascular surgical operations in the Society of Thoracic Surgeons Congenital Heart Surgery Database (January 1, 2010 - June 30, 2017) were analyzed. In this STAT 2020 Update of the STAT Mortality Score and Categories, the risk associated with a specific combination of procedures was estimated under the assumption that risk is determined by the highest risk individual component procedure. Operations composed of multiple component procedures were eligible for unique STAT Scores when the statistically estimated mortality risk differed from that of the highest risk component procedure. Bayesian modeling accounted for small denominators. Risk estimates were rescaled to STAT 2020 Scores between 0.1 and 5.0. STAT 2020 Category assignment was designed to minimize within-category variation and maximize between-category variation., Results: Among 161,351 operations at 110 centers (19,090 distinct procedure combinations), 235 types of single or multiple component operations received unique STAT 2020 Scores. Assignment to Categories resulted in the following distribution: STAT 2020 Category 1 includes 59 procedure codes with model-based estimated mortality 0.2% to 1.3%; Category 2 includes 73 procedure codes with mortality estimates 1.4% to 2.9%; Category 3 includes 46 procedure codes with mortality estimates 3.0% to 6.8%; Category 4 includes 37 procedure codes with mortality estimates 6.9% to 13.0%; and Category 5 includes 17 procedure codes with mortality estimates 13.5% to 38.7%. The number of procedure codes with empirically derived Scores has grown by 58% (235 in STAT 2020 vs 148 in STAT 2009). Of the 148 procedure codes with empirically derived Scores in 2009, approximately one-half have changed STAT Category relative to 2009 metrics. The New STAT 2020 Scores and Categories demonstrated good discrimination for predicting mortality in an independent validation sample (July 1, 2017-June 30, 2019; sample size 46,933 operations at 108 centers) with C-statistic = 0.791 for STAT 2020 Score and 0.779 for STAT 2020 Category., Conclusions: The updated STAT metrics reflect contemporary practice and outcomes. New empirically based STAT 2020 Scores and Category designations are assigned to a larger set of procedure codes, while accounting for risk associated with multiple component operations. Updating STAT metrics based on contemporary outcomes facilitates accurate assessment of case mix.