Development and multicenter validation of an artificial intelligence ECG model for ventricular remodeling in repaired tetralogy of Fallot

Figures

• 

  Figure 1. Study overview. A multicentre study to train and validate a deep learning electrocardiogram model to predict the abnormalities in biventricular size and function on cardiac MRI for patients with tetralogy of Fallot.

• 

  Figure 2. Model discrimination and calibration. Receiver-operating characteristic curves (left panel) and precision–recall curves (middle panel) to identify cardiac MRI-quantified volumetric and functional abnormalities in patients with repaired tetralogy of Fallot. Composite outcome reliability diagram (right panel). Calibrated predictions were aggregated across all sites. ECE, expected calibration error. P-value is Spieghalter’s Z test P-value in which a P < 0.05 suggests model miscalibration.

• 

  Figure 3. Decision curve analysis. Implementation of artificial intelligence analysis of electrocardiogram as a screening tool to detect adverse ventricular remodelling to aid in the timing of advanced imaging in tetralogy of Fallot. The net reduction in cardiac MRI (green line) is plotted against the current strategy of every patient undergoing cardiac MRI (red dashed line) and the strategy of no patient undergoing cardiac MRI (black dashed line). The well-performing subset of sites (A, B, C) is also separately plotted to examine the best-case net benefit (green dashed line).

• 

  Figure 4. Key subgroup performance. Key subgroup model performance in the aggregated external test cohort. Bonferroni-corrected Delong test P-value for the difference in area under the receiver-operating curve displayed on the right.

• 

  Figure 5. Median waveform analysis and saliency mapping. Median waveforms of the 10 highest (red line) and 10 lowest (green line) predicted risk for the primary outcome in independent patient electrocardiograms from Site A. Saliency mapping (blue shade) demarcates important prediction regions of the waveform.