(Screen 1 - 5:30 PM Saturday) Using Machine Learning to Predict Transplant-Free Survival Among Neonates with Hypoplastic Left Heart Syndrome: An Analysis of the Single Ventricle Reconstruction and Extension Study Data

Find board locations here.

Abstract: Introduction. Despite advances in management of staged palliation for hypoplastic left heart syndrome (HLHS), associated mortality remains high and relationships between pre-Norwood Stage 1 palliation (S1P) factors and longer-term outcomes of a staged surgical approach to palliation remain elusive. Patients within the Single Ventricle Reconstruction (SVR) study and associated Extension Study (SVR2) represent a cohort of neonates undergoing staged palliation for HLHS. As machine learning (ML) has been shown to produce improved results relative to traditional survival analysis methods, we hypothesized that an ML-based precision medicine approach to survival analysis following S1P would prove to be a superior predictive tool.

Methods. Data from the SVR and SVR2 datasets were used to build models designed to predict a composite endpoint of all-cause mortality or transplant during the study period. Only features available prior to S1P were used. Univariate feature selection using mutual information was applied with feature sets ranging from 30 to 70 in increments of five were tested, along with all pre-S1P features. A 60:40 split of train:test data was used, and model selection was used with the full feature selection. Top performing models were selected for hyperparameter optimization against the subset of features. Bayesian optimization was used to select the best hyperparameters using the concordance index (CI). Model performance was assessed using the CI, Brier score and the time-dependent area under the receiver operating characteristic curve (AUC). Scikit-survival was used for all model building.

Results. Of 549 patients included in the original SVR trial, the study endpoint of all-cause mortality or transplantation was identified in 212 patients across both SVR and SVR2 datasets. Median follow up time was 6 years (0.35-6.06 years). A total of 74 features were considered for model inclusion. As shown in Figure 1A, the 65-feature model outperformed other models for the entirety of the prediction window (0.5 – 6 years), demonstrating the highest CI (0.64) and lowest Brier score (0.20). The model was also stable over time, with small performance loss from 0.5 years (AUC 0.675) to 6 years (AUC 0.68). Highest-value features for the 65-feature model are shown in Figure 1B, incorporating elements of preoperative care with anatomic, physiologic, and socioeconomic features.

Conclusion. ML-driven survival analysis represents an important approach to developing clinical decision support tools. This work illustrates a model useful for identifying the probability of all-cause mortality or transplantation for over 5 years following S1P, using data only available prior to S1P. The development of an adaptive model leveraging additional features associated with operative and interstage care to improve model fidelity is ongoing.