Development and validation of echocardiography-based machine-learning models to predict mortality.

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Imaging

Background:

• Echocardiography (echo) based machine learning (ML) models show potential for identifying patients at high risk of all-cause mortality.
• Risk stratification is crucial for managing patients with heart conditions.

Purpose of the Study:

• To develop and validate machine learning (ML) models using echocardiography (echo) to predict 1-, 3-, and 5-year all-cause mortality.
• To compare the performance of ML models against established risk scores and assess their correlation with patient quality of life.

Main Methods:

• Developed ResNet (deep learning) and CatBoost (gradient boosting) ML models using echo videos and measurements.
• Trained models on the Mackay dataset (Taiwan) and validated externally on the Alberta HEART dataset (Canada).
• Evaluated model performance overall, in subgroups (controls, at-risk, HFrEF, HFpEF), against the MAGGIC score, and correlated with Kansas City Cardiomyopathy Questionnaire (KCCQ) scores.

Main Results:

• ResNet and CatBoost models demonstrated high internal validation AUROCs (85-92%).
• In external validation, ResNet models (AUROCs 78-82%) outperformed CatBoost (73-78%) and showed comparable performance to the MAGGIC score.
• ML models accurately predicted higher mortality risk in heart failure subgroups and correlated with patient-reported quality of life.

Conclusions:

• Echocardiography-based ML models possess strong internal and external validity, demonstrating generalizability.
• These models correlate with patient quality of life and offer comparable or superior performance to existing risk scores.
• ML models can be utilized for automated, point-of-care risk stratification in cardiovascular care.