In-hospital mortality among patients with congenital heart disease undergoing noncardiac procedures: machine learning analysis of a multicentre study

Background Improved survival for children with congenital heart disease (CHD) has resulted in an increasing number presenting for noncardiac procedures. Several factors are associated with mortality in these patients, including severity of heart disease, preoperative inotropic support, preoperative ventilatory support, emergency procedure, and comorbidities. Prior literature has relied on multivariable regression analyses. The aim of this analysis was to utilise machine learning (ML) methodology to determine risk factors for in-hospital mortality after noncardiac procedures in paediatric patients with CHD. Methods Seven paediatric centres collected data between January 2021 and December 2022 involving patients <21 yr of age with CHD undergoing noncardiac procedures. Both the random forest and eXtreme Gradient boosting machine models were tested, with the optimal model chosen based on the area under the precision recall curve in the validation set. Results Of 5977 patients, 1.74% ( n= 104) had in-hospital mortality; among 9837 surgical encounters, 1.48% ( n= 146) were associated with in-hospital mortality. The top features contributing to mortality prediction in the eXtreme Gradient boost model were similar to defined factors in previous multivariable regression analyses: pre-procedural inotropic support; duration of anaesthesia; severe cardiac disease; prematurity; white race; emergent procedure; gender; inpatient procedure; black race; weekend procedure; and gastrointestinal, neurological, haematological, and respiratory comorbidities. A decision curve analysis shows benefit to understanding patients at increased risk of mortality when the probability of mortality is between 0% and 60%. Conclusions The risk factors shown in our dataset using ML methods are similar to those demonstrated in our prior work utilising logistic regression. Ongoing awareness of these mortality predictors is essential to appropriate care planning.