Developing and evaluating machine learning models in the prediction of recurrent sudden cardiac arrest
Related Experiment Videos
Paulomi Gohel1, Michael Daidone1, Tanmay A Gokhale1
1Department of Medicine, Division of Cardiology and the Heart and Vascular Institute at the University of Pittsburgh Medical Center, Pittsburgh, PA, United States.
Background:
Sudden cardiac arrest (SCA) is a major killer of adults in the United States and has a significant recurrence rate. Predicting the risk of recurrent SCA has implications to patients' counseling and management.
Objective:
The goal of this study is to develop and evaluate machine learning (ML) models to predict recurrent SCA in this high-risk population.
Methods:
Patients (n = 5,286) whose index arrest was coded as a primary diagnosis were selected and followed to the endpoint of second SCA. The cohort was divided into training (n = 3,382, 3.8% event rate), internal validation (n = 846, 3.8% event rate), and holdout test (1,058, 3.8% event rate) sets. External validation was performed using a cohort of 1,776 patients (2.3% event rate). Multiple ML discriminators were tested using model area under the receiver operating characteristic curve (AUROC), sensitivity, specificity, positive (PPV) and negative (NPV) predictive values, and F1 scores. The model with highest AUROC was evaluated on a holdout test set and was externally validated.
Results:
The gradient boosting (GB) model outperformed other models with AUROC of 0.93 in the internal validation set. It also achieved high discrimination in the holdout (AUROC=0.92), and external validation (AUROC = 0.90, sensitivity = 0.92, specificity = 0.83, F1 score = 0.19) sets. Shapley additive explanations (SHAP) demonstrated that a diagnostic code of shockable rhythm (ventricular tachycardia), evidence of structural heart disease, conduction abnormalities, and initial blood work results at the time of the index arrest were main contributors to model prediction. Low (NPV ≥ 99%), intermediate, and high (PPV≥25%) risk groups defined from the internal validation set were applied to the holdout and external validation sets with strong prediction of time to recurrent SCA (p = 0.003).
Conclusion:
Our ML prediction model demonstrates very good and stable discrimination in the prediction of recurrent SCA based on explainable clinical factors. External validation results confirm the generalizability of our findings. This model could support risk stratification and counseling for survivors of SCA.
