Title : Explainable machine learning model for early predicting mortality in children after congenital heart surgery
Abstract:
Background: Congenital heart disease (CHD) remains one of the leading causes of infant and postneonatal mortality. Early recognition and successful management of critically ill CHD patients is a difficult challenge considering the complexity of the anatomy and physiopathology of congenital heart diseases. This study aimed to construct and validate an early mortality prediction model in CHD patients who underwent congenital heart surgery to guide management strategies.
Methods: In a single-center retrospective study, pediatric CHD patients who underwent congenital heart surgery were eligible. The primary outcome was operative mortality, that is, any death during hospitalization within 30 days of surgery. Competing risk models with independent sample t-test, Mann–Whitney U test, Chi-square test, and Recursive Feature Elimination with cross-validation were used to identify and quantify clinical characteristics acquired within the first 24 hours of admittance in CICU associated with early predicting mortality. Afterward, Shapley Additive exPlanations (SHAP) method to intuitively interpret how each variable is attributed to the current precision model. Early mortality prediction models were implemented using chosen characteristics by seven machine learning algorithms and tuned with 5-fold cross-validations. The final model was chosen by Delong’s test by comparing its performance with other competing risk models. Finally, its performance was validated in an independent external cohort and explained by SHAP methods to show how this model works in a specific patient.
Results: Overall, 5756 pediatric CHD patients met eligibility criteria and were used as a development cohort. Risk predictors included arterial systolic pressure, blood platelet count, lactic acid, creatinine, alveolar-arterial gradient, and total intake per hour, of which the alveolar-arterial gradient is the most crucial variable. Seven early mortality prediction models were developed, and all of them adequately discriminated between patients with risks of mortality with the area under the curve (AUC) 0.888, 0.881, 0.897, 0.953, 0.924, 0.947, and 0.946, respectively. The model using the random forest (RF) method showed the best performance and demonstrated good agreement between the development and validation cohorts (AUC 0.953 and 0.95, respectively).
Conclusion: This study provides an explainable and validated early mortality prediction model with >80% prediction accuracy and incorporates risk factors that are meaningful to predict poor prognosis. An individualized early mortality prediction model has the potential to improve the alertness of physicians and assist with management improvement.
Audience Take Away Notes
- Explain how the audience will be able to use what they learn?
- How will this help the audience in their job?
- Is this research that other faculty could use to expand their research or teaching?
- Does this provide a practical solution to a problem that could simplify or make a designer’s job more efficient?
- Will it improve the accuracy of a design, or provide new information to assist in a design problem?
- List all other benefits.
- The observation in pediatric patients with congenital heart disease should not be based on notable symptoms, but should incorporate all age-appropriate, evidence-based risk factors to thorough evaluation.
- Machine learning methods could provide more subjective and better prediction mortality performance than the conventional statistical approach. Meanwhile, the use of Shapley Additive Explanations can easily rationalize the prediction performance of each variable in complex machine learning models and make models more interpretable by the user.
- The early mortality prediction model can be implemented in a specific patient as an evaluation tool to help guide physicians in decision-making with different management strategies.
