Deep Learning for Improved Risk Prediction in Surgical Outcomes.

Area of Science:

• Pediatric Cardiology
• Medical Informatics
• Machine Learning in Healthcare

Background:

• The Norwood procedure is critical for neonatal single ventricle congenital heart defects but has a high mortality rate.
• Existing risk prediction models lack patient specificity or only predict in-hospital mortality.

Purpose of the Study:

• To develop accurate, patient-specific predictive models for one-year postoperative mortality/transplantation and prolonged hospital stay after the Norwood procedure.
• To aid clinicians and families in preoperative decision-making.

Main Methods:

• Utilized the Pediatric Heart Network Single Ventricle Reconstruction trial dataset.
• Applied Markov Chain Monte-Carlo simulation for missing data imputation.
• Employed multiple machine learning models, including deep neural networks.

Main Results:

• Deep neural network model achieved 89±4% accuracy and 0.95±0.02 AUROC for mortality/transplantation prediction.
• Models predicted prolonged length of stay with 85±3% accuracy and 0.94±0.04 AUROC.
• High accuracy in predicting individual patient risk factors.

Conclusions:

• Developed accurate machine learning models for predicting critical outcomes in neonates undergoing the Norwood procedure.
• These predictive tools can significantly inform clinical and organizational decision-making processes.
• Patient-specific risk stratification can improve care for single ventricle heart defects.