Cardiac Comorbidity Risk Score: Zero‐Burden Machine Learning to Improve Prediction of Postoperative Major Adverse Cardiac Events in Hip and Knee Arthroplasty

Figure 1. Out‐of‐sample performance of Cardiac Comorbidity Risk Score (CCoR) and Revised Cardiac Risk Index (N=445 391) to predict major adverse cardiac events in the 4 weeks after elective primary lower extremity arthroplasty. The validation cohort comprised the 50% of the study sample not used to develop CcoR. A and B, The receiver operating characteristics and precision‐recall curves, respectively, of the two risk scores by sex. The AUROC of the receiver operating characteristics curve and the precision‐recall curve reflect the risk scores’ performance in predicting major adverse cardiac events 4 weeks after elective primary total hip or knee joint replacement. CcoR shows a strong prognostic performance that substantially exceeds that of Revised Cardiac Risk Index in both men and women. C, The relative importance of phenotypes (comorbidity phenotypes) in CcoR estimation. The 20 top phenotypes by sex are shown in descending order of relative importance in estimating the CcoR. Notably, these importances do not merely sum the predictive contribution from the presence or absence of individual diagnostic codes, a contribution reflected in the comorbidity spectra illustrated in Figure 2.…

Figure 2. Comorbidity spectra. Disorders that increase the odds of a “true positive” vs a “true negative”; ie, these disorders, ranked here according to the log‐odds ratio in ascending order of frequency in “true posititive” versus “true negative” patients, are more likely in patients who are in the positive cohort. ICD‐10 indicates International Classification of Diseases, Tenth Revision. Panel A and B shows comorbidity spectra for the two sexes, which while similar, have important differences.

Figure 3. Cost model. A, The distribution of estimated cost for Cardiac Comorbidity Risk Score (CcoR) and Revised Cardiac Risk Index (RCRI) over the years for achieving positive predictive value between 10% and 20%. B, The cost trade‐off with sensitivity for 2020 and 2040 along with 95% CI. Note that the estimated cost of achieving a 40% sensitivity in the general population (which approximately corresponds to >6% risk with RCRI and reflects current practice) with CCoR is ≈750M US dollars per year currently, while with RCRI this increases to 1.5B US dollars per year. C, The relative boost in sensitivity achieved by CCoR over RCRI as a percentage for a fixed cost. D, The relative boost in positive predictive value achieved by CCoR over RCRI as a percentage for a fixed cost. Thus, if the yearly cost is fixed at 500M, then CCoR can deliver >150% more sensitivity, or almost 20% more positive predictive value up to around 2025. If the costs are set to what is estimated for current practice ($1478M), then CCoR can deliver a >50% improvement in either sensitivity or positive predictive value. The estimated number of surgeries per year, and values for the costs of additional testing on a positive…