Electrocardiogram-Based Deep Learning to Predict Mortality in Repaired Tetralogy of Fallot

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Artificial intelligence-enhanced electrocardiogram (AI-ECG) analysis shows potential for predicting mortality in adults with acquired cardiovascular diseases.
• Its utility in the repaired tetralogy of Fallot (rTOF) population is currently unexplored.

Purpose of the Study:

• To develop and externally validate an AI-ECG model for predicting 5-year mortality in patients with rTOF.
• To assess the AI-ECG model's performance against established clinical and imaging biomarkers.

Main Methods:

• A convolutional neural network was trained on ECGs from Boston Children's Hospital.
• The model underwent internal testing (Boston) and external validation (Toronto INDICATOR cohort).
• Performance was measured using area under the receiver operating (AUROC) and precision recall (AUPRC) curves.

Main Results:

• The model demonstrated robust performance in both internal testing (AUROC 0.83, AUPRC 0.18) and external validation (AUROC 0.81, AUPRC 0.21).
• AI-ECG prediction was comparable to the biventricular global function index and superior to QRS duration.
• AI-ECG prediction independently predicted mortality in multivariable analysis.

Conclusions:

• An externally validated AI-ECG model shows promise for predicting 5-year mortality in rTOF patients.
• This AI-ECG tool may enhance risk stratification when used in conjunction with imaging biomarkers.