Big Data/Machine Learning/AI
Integrating Machine Learning in Pre-Fontan Hemodynamic Studies for Enhanced Cardiovascular Outcome Prediction Tzu-Chun Chu* Tzu-Chun Chu Yanxu Yang Lazaros Kochilas Matt Oster Jessica Knight
Single ventricle congenital heart disease is a severe birth defect with significant attrition after the initial surgical repair. In advancing personalized care, assessing the impact of hemodynamics and patient characteristics on long-term outcomes is essential. Our study aims to identify key determinants of long-term survival after the Fontan procedure, the final stage of single ventricle palliation, while navigating the complexities of high dimensionality, non-linearity, heterogeneity, and issues with censored and missing data by utilizing survival machine learning (ML) models.
We queried the Pediatric Cardiac Care Consortium (PCCC), a large US-based multicenter registry of congenital heart interventions for patients undergoing the Fontan procedure from 1982-2011. Post-discharge deaths were assessed by matching with the National Death Index through 2022. Missing data were inputted using a tree-based method before model training. We fitted a random survival forest (RSF) model comprising 1000 survival trees, constructed through log-rank splitting. A permutation importance measure was used to identify the top 20 most important variables associated with long-term survival post Fontan hospital discharge.
The study involved 1,366 patients who underwent Fontan procedure (median age at Fontan = 3.1 yrs) and survived to hospital discharge of whom 172 of them died (median time to event=12.3 yrs). The out-of-bag (OOB) performance error was 0.334, and OOB ensemble mortality yielded a c-index of 0.666. The figure shows that patient baseline characteristics were predominantly chosen as important features.
Through advanced techniques, including a RSF model and permutation importance measure, we identified key determinants of late mortality post-Fontan. Further research will focus on comparing advanced survival ML models, like boosted Cox regression and support vector machines, to identify the most effective approach in predicting Fontan long-term survival in rich and complex data.
