Machine Learning-Based Prognostic Models for Mortality in Patients Receiving Implantable Cardioverter Defibrillators

Area of Science:

• Cardiology
• Medical Informatics
• Machine Learning

Background:

• Accurate prediction of clinical trajectory for patients with implantable cardioverter-defibrillators (ICDs) is crucial for patient management.
• Machine learning (ML) offers advanced capabilities for analyzing complex data patterns and providing personalized risk estimates, surpassing traditional statistical methods.

Purpose of the Study:

• To develop and validate machine learning models for predicting the clinical trajectory of patients with ICDs.
• To compare the performance of different ML models in predicting patient outcomes.
• To identify key predictors of clinical outcomes in ICD patients.

Main Methods:

• Retrospective study involving 3175 patients from four Chinese hospitals.
• Development and validation of six ML models, including multilayer perceptron (MLP).
• Model performance assessed using Area Under the Receiver Operating Characteristic Curve (AUROC); survival stratified using Kaplan-Meier curves and SHapley Additive exPlanations (SHAP).

Main Results:

• The MLP model achieved superior predictive accuracy with AUROC of 0.70 (internal) and 0.72 (external).
• MLP model effectively stratified patients into high-risk and low-risk groups (p < 0.001).
• Glomerular filtration rate (GFR) was identified as the most influential predictor among seven key factors.

Conclusions:

• The MLP model accurately predicts 3-year survival for ICD/CRT-D patients and stratifies them into distinct risk groups.
• Integration of MLP and SHAP provides explainable AI for individualized risk prediction, aiding clinical decisions and treatment optimization.