Machine Learning Analysis of Predictors for Inhaled Nitric Oxide Therapy Administration Time Post Congenital Heart Disease Surgery: A Single-Center Observational Study

TL;DR

This study uses machine learning to identify factors predicting the need for long-term inhaled nitric oxide therapy in children after heart surgery.

Contribution

A machine learning model is developed and validated to predict long-term inhaled nitric oxide use in congenital heart disease patients post-surgery.

Findings

Four factors (CPB, in-out balance, aortic cross-clamp time, and lactate) strongly predict long-term inhaled nitric oxide use.

The ML model achieved perfect classification in both training and testing cohorts.

Identifying these predictors can improve postoperative management and prevent complications in congenital heart disease patients.

Abstract

Background

Congenital heart disease (CHD) is a structural deformity of the heart present at birth. Pulmonary hypertension (PH) may arise from increased blood flow to the lungs, persistent pulmonary arterial pressure elevation, or the use of cardiopulmonary bypass (CPB) during surgical repair. Inhaled nitric oxide (iNO) selectively reduces high blood pressure in the pulmonary vessels without lowering systemic blood pressure, making it useful for treating children with postoperative PH due to heart disease. However, reducing or stopping iNO can exacerbate postoperative PH and hypoxemia, necessitating long-term administration and careful tapering. This study aimed to evaluate, using machine learning (ML), factors that predict the need for long-term iNO administration after open heart surgery in CHD patients in the postoperative ICU, primarily for PH management.

Methods

We used an ML approach to establish an algorithm to predict 'patients with long-term use of iNO' and validate its accuracy in 34 pediatric postoperative open heart surgery patients who survived and were discharged from the ICU at Kagoshima University Hospital between April 2016 and March 2019. All patients were started on iNO therapy upon ICU admission. Overall, 16 features reflecting patient and surgical characteristics were utilized to predict the patients who needed iNO for over 168 hours using ML analysis with AutoGluon. The dataset was randomly classified into training and test cohorts, comprising 80% and 20% of the data, respectively. In the training cohort, the ML model was constructed using the important features selected by the decrease in Gini impurity and a synthetic oversampling technique. In the testing cohort, the prediction performance of the ML model was evaluated by calculating the area under the receiver operating characteristics curve (AUC) and accuracy.

Results

Among 28 patients in the training cohort, five needed iNO for over 168 hours; among six patients in the testing cohort, one needed iNO for over 168 hours.

CPB, aortic clamp time, in-out balance, and lactate were the four most important features for predicting the need for iNO for over 168 hours.

In the training cohorts, the ML model achieved perfect classification with an AUC of 1.00. In the testing cohort, the ML model also achieved perfect classification with an AUC of 1.00 and an accuracy of 1.00.

Conclusion

The ML approach identified that four factors (CPB, in-out balance, aortic cross-clamp time, and lactate) are strongly associated with the need for long-term iNO administration after open heart surgery in CHD patients. By understanding the outcomes of this study, we can more effectively manage iNO administration in postoperative open heart surgery in CHD patients with PH, potentially preventing the recurrence of postoperative PH and hypoxemia, thereby contributing to safer patient management.