Back to 2024 Abstracts
A Deep Neural Network Helping To Answer: Who Can Safely Be Discharged After The Norwood?
John T. Kennedy, III, MD1, Kevin Kulshrestha, MD, MBE
2, Haleh C. Heydarian, MD
2, Garick D. Hill, MD
2, Michael Gaies, MD, MPH
2, Marco Ricci, MD, MBA
2, David G. Lehenbauer, MD
2, Awais Ashfaq, MD
2, David L. S. Morales, MD
2.
1University of Cincinnati Medical Center, Cincinnati, OH, USA,
2Cincinnati Children's Hospital Medical Center, Cincinnati, OH, USA.
Objective: Interstage mortality following Stage I palliation remains significant and many patients may benefit from hospitalization during this high-risk period. This study describes a deep neural network (DNN) to help identify these patients.
Methods: Utilizing data from the Single Ventricle Reconstruction trial, we developed a DNN model to predict mortality following ICU discharge. We used SHapley Additive exPlanations values to interpret the model and identify significant predictors. The model's variables consisted of a combination of features chosen by the authors and others selected via machine learning tools.
Results: A total of 476 patients were included. Mortality between ICU discharge and Stage II palliation was 13% (60/476). Our DNN was trained on 380 patients and tested on 96 patients using 23 variables (Figure, left). On the test set, the model showed 86% sensitivity, 79% specificity, NPV of 0.97, and PPV of 0.41. Overall accuracy was 80% with an ROC AUC of 0.82 (Figure, right). Key risk factors included shunt type, socioeconomic status, and post-operative ICU complications. The model's training and validation loss curves were similar, indicating no overfitting.
Conclusion: We developed a highly sensitive DNN that outperforms existing models and effectively screens high-risk patients during the interstage period who would benefit from further discussion prior to discharge. Conversely, low risk patients can be safely discharged with an expected mortality < 3%.
Back to 2024 Abstracts
