Fig. 2: Model comparison in VITAL-AF Test Set.

Depicted are receiver operating characteristic curves for 2-year incident AF in the VITAL-AF test set according to three alternative approaches to 2-year AF risk estimation using deep learning of the 1 L ECG: a previously published 12-lead ECG-based convolutional neural network (CNN) trained using only single standard ECG leads (1 L ECG-AI, green), 1 L ECG-AI with additional fine-tuning within a subset of the VITAL-AF sample (1 L ECG-AI Fine-Tuned, purple), and a de novo CNN trained solely within a subset of the VITAL-AF data (1 L VITAL, magenta).