Impact of heart failure on reoperation in adult congenital heart disease: An innovative machine learning model

Abstract

ObjectivesThe study objectives were to evaluate the association between preoperative heart failure and reoperative cardiac surgical outcomes in adult congenital heart disease and to develop a risk model for postoperative morbidity/mortality. MethodsSingle-institution retrospective cohort study of adult patients with congenital heart disease undergoing reoperative cardiac surgery between January 1, 2010, and March 30, 2022. Heart failure defined clinically as preoperative diuretic use and either New York Heart Association Class II to IV or systemic ventricular ejection fraction less than 40%. Composite outcome included operative mortality, mechanical circulatory support, dialysis, unplanned noncardiac reoperation, persistent neurologic deficit, and cardiac arrest. Multivariable logistic regression and machine learning analysis using gradient boosting technology were performed. Shapley statistics determined feature influence, or impact, on model output. ResultsPreoperative heart failure was present in 376 of 1011 patients (37%); those patients had longer postoperative length of stay (6 [5-8] vs 5 [4-7] days, P < .001), increased postoperative mechanical circulatory support (21/376 [6%] vs 16/635 [3%], P = .015), and decreased long-term survival (84% [80%-89%] vs 90% [86%-93%]) at 10 years (P = .002). A 7-feature machine learning risk model for the composite outcome achieved higher area under the curve (0.76) than logistic regression, and ejection fraction was most influential (highest mean |Shapley value|). Additional risk factors for the composite outcome included age, number of prior cardiopulmonary bypass operations, urgent/emergency procedure, and functionally univentricular physiology. ConclusionsHeart failure is common among adult patients with congenital heart disease undergoing cardiac reoperation and associated with longer length of stay, increased postoperative mechanical circulatory support, and decreased long-term survival. Machine learning yields a novel 7-feature risk model for postoperative morbidity/mortality, in which ejection fraction was the most influential.