Development of a Machine Learning Model to Predict Outcomes and Cost After Cardiac Surgery

• Published in: The Annals of thoracic surgery Vol. 115; no. 6; pp. 1533 - 1542
• Main Authors: Zea-Vera, Rodrigo, Ryan, Christopher T., Navarro, Sergio M., Havelka, Jim, Wall, Matthew J., Coselli, Joseph S., Rosengart, Todd K., Chatterjee, Subhasis, Ghanta, Ravi K.
• Format: Journal Article
• Language: English
• Published: Netherlands Elsevier Inc 01.06.2023
• Subjects: Aged; Cardiac Surgical Procedures - adverse effects; Coronary Artery Bypass - methods; Humans; Machine Learning; Medicare; Thoracic Surgery; United States - epidemiology; United States; PROMM; STS; PROM; XGBoost; AUC-PR; ROC; CABG; ML
• ISSN: 0003-4975; 1552-6259
• DOI: 10.1016/j.athoracsur.2022.06.055

Machine learning (ML) algorithms may enhance outcomes prediction and help guide clinical decision making. This study aimed to develop and validate a ML model that predicts postoperative outcomes and costs after cardiac surgery. The Society of Thoracic Surgeons registry data from 4874 patients who underwent cardiac surgery (56% coronary artery bypass grafting, 42% valve surgery, 19% aortic surgery) at our institution were divided into training (80%) and testing (20%) datasets. The Extreme Gradient Boosting decision-tree ML algorithms were trained to predict three outcomes: operative mortality, major morbidity or mortality, and Medicare outlier high hospitalization cost. Algorithm performance was determined using accuracy, F1 score, and area under the precision-recall curve (AUC-PR). The ML algorithms were validated in index surgery cases with The Society of Thoracic Surgeons risk scores for mortality and major morbidities and with logistic regression and were then applied to nonindex cases. The ML algorithms with 25 input parameters predicted operative mortality (accuracy 95%; F1 0.31; AUC-PR 0.21), major morbidity or mortality (accuracy 71%, F1 0.47; AUC-PR 0.47), and high cost (accuracy 84%; F1 0.62; AUC-PR 0.65). Preoperative creatinine, complete blood count, patient height and weight, ventricular function, and liver dysfunction were important predictors for all outcomes. For patients undergoing nonindex cardiac operations, the ML model achieved an AUC-PR of 0.15 (95% CI, 0.05-0.32) for mortality and 0.59 (95% CI, 0.51-0.68) for major morbidity or mortality. The extreme gradient boosting ML algorithms can predict mortality, major morbidity, and high cost after cardiac surgery, including operations without established risk models. These ML algorithms may refine risk prediction after cardiac surgery for a wide range of procedures. [Display omitted]