Development and prospective evaluation of a machine learning model to predict serious cardiac outcomes among paediatric cardiac inpatients

Objectives Objectives were to develop a machine learning (ML) model based on electronic health record data to predict the risk of a serious cardiac outcome within the next 3 months among patients admitted to the cardiology service using retrospective data, and to evaluate the model prospectively in a silent trial (predictions not provided to clinicians). Methods and analysis Admissions between 2 June 2018 to 21 August 2023 (retrospective) and 10 May 2024 to 26 October 2024 (prospective) to the cardiology service were included. Data were a curated and validated source named SickKids Enterprise-wide Data in Azure Repository. Prediction time was the morning following admission. The label was a composite outcome consisting of ventricular assist device procedure, heart transplant waitlisting or death within 3 months. We trained models using L2-regularised logistic regression, LightGBM and XGBoost. Training cohorts include the target cohort and all inpatient admissions. Results The best-performing model in the retrospective phase was LightGBM trained on all inpatients. There were 51 571 admissions used for model development in the retrospective phase and 515 admissions in the prospective silent trial. The number of features in the final model was 7553. The area under the receiver operating characteristic curve was 0.88 (95% CI 0.88 to 0.89) for retrospective and 0.82 (95% CI 0.79 to 0.83) for prospective silent trial phases. Based on a threshold selected during the retrospective phase, silent trial positive and negative predictive values were 0.19 and 0.97, respectively. Conclusions We created an ML model to predict serious cardiac outcomes using a deployment-aware framework leveraging real-world data. Postdeployment evaluation will be an important future goal.