Adverse events after repair of tetralogy of Fallot: prediction models by machine learning of a retrospective cohort study in western China

Abstract

Background: The incidence of clinical adverse events after tetralogy of Fallot (TOF) repair remains high. This study was performed to explore risk factors for adverse events and develop a prediction model through machine learning (ML) to forecast the incidence of clinical adverse events after TOF repair.

Methods: A total of 281 participants who were treated with cardiopulmonary bypass (CPB) at our hospital from January 2002 to January 2022 were included in the study. Risk factors for adverse events were explored by composite and comprehensive analyses. Five artificial intelligence (AI) models were used for ML to build prediction models and screen out the model with the best performance in predicting adverse events.

Results: CPB time, differential pressure of the right ventricular outflow tract (RVOTDP or DP), and transannular patch repair were identified as the main risk factors for adverse events. The reference point for CPB time was 116.5 minutes and that for right ventricular (RV) outflow tract differential pressure was 70 mmHg. SPO₂ was a protective factor, with a reference point of 88%. By integrating the results for the training and validation cohorts, we confirmed that, among all models, the logistic regression (LR) model and Gaussian Naive Bayes (GNB) model were stable, showing good discrimination, calibration and clinical practicability. The dynamic nomogram can be used as a predictive tool for clinical application.

Conclusions: Differential pressure of the RV outflow tract, CPB time, and transannular patch repair are risk factors, and SPO₂ is a protective factor for adverse events after complete TOF repair. In this study, models developed by ML were established to predict the incidence of adverse events.

Keywords: Tetralogy of Fallot (TOF); adverse events; artificial intelligence (AI); machine learning (ML).

Figure 1

Flow chart of the study. PSM, propensity score matching; RCS, restricted cubic spline; ML, machine learning; DCA, decision curve analysis.

Figure 2

Coefficients of importance. RVOTDP or DP, differential pressure of the right ventricular outflow tract; CPB, cardiopulmonary bypass; TP repair, transannular patch repair; ACCT, Aortic cross-clamp time; EF, ejection fraction; PvO, opening of the pulmonary valve; VSD, ventricular septal defect; PV-leaflet, pulmonary valve leaflet; HCT, haematocrit; LVEDI, left ventricular end-diastolic volume index; Z-index, pulmonary valve annulus Z score; M-index, McGoon index.

Figure 3

Coefficients of the LASSO model: LASSO coefficient profiles of the 6 features. A coefficient profile plot was produced against the log λ sequence. LASSO, least absolute shrinkage and selection operator.

Figure 4

Venn diagram of the variables selected according to the three analytical methods mentioned above. CPB, cardiopulmonary bypass; RVOTDP or DP, differential pressure of the right ventricular outflow tract; TP repair, transannular patch repair.

Figure 5

RCS analysis of CPB time, DP and SPO₂ according to adverse events. The reference points (HR/OR =1) were 116.5 minutes for CPB time, 88% for SPO₂ and 70 mmHg for DP. OR, odds ratio; CI, confidence interval; RCS, restricted cubic spline; CPB, cardiopulmonary bypass; DP, differential pressure of the right ventricular outflow tract; HR, hazard ratio.

Figure 6

Trend analysis between different TP repair strategies and adverse events over SPO₂. TP, transannular patch.

Figure 7

Predicted ROC curve of adverse events in the training and testing sets of the five models. ROC, receiver operating characteristic; AUC, the area under the curve; CI, confidence interval; GNB, Gaussian Naive Bayes; MLP, multilayer perceptron; SVM, support vector machine; KNN, K-nearest neighbour.

Figure 8

Nomogram of the LR model for adverse events. To estimate the probability for an individual patient, the value of each factor is acquired on each variable axis; then, a line is drawn upwards to determine the point. The sum of these numbers is located on the total points axis, and a line is drawn downwards to the risk axis to determine the likelihood of an adverse event. Z-index, pulmonary valve annulus Z score; M-index, McGoon index; LVEDI, left ventricular end-diastolic volume index; PvO, opening of the pulmonary valve; RVOTDP or DP, differential pressure of the right ventricular outflow tract; VSD, ventricular septal defect; EF, ejection fraction; ACCT, aortic cross-clamp time; CPB, cardiopulmonary bypass; HCT, haematocrit; TP repair, transannular patch repair; PV-leaflet, pulmonary valve leaflet; LR, logistic regression.

Figure 9

DCA for the model combining SPO₂ and TP repair. The y-axis measures the net benefit. TP repair, transannular patch repair; DCA, decision curve analysis.