Artificial Intelligence‐Enhanced Electrocardiography for Predicting Paroxysmal Atrial Fibrillation From Sinus Rhythm: Impact of Data Integration Across Institutions and Devices

Figures

• 

  FIGURE 1. Data source information.

• 

  FIGURE 2. Flowchart of patient selection. AF, atrial fibrillation; ECG, electrocardiography; SR, sinus rhythm.

• 

  FIGURE 3. Convolutional neural network (CNN) analysis. The CNN architecture consisted of a temporal axis, which included the convolutional component (a) and the residual component (b), as well as a lead axis (c). We developed five models: Scratch models S1, S2, and S3 using Datasets 1 (from Tokai‐All), 2 (from Tokai‐Cardiol), and 3 (from Shinken), respectively; and fine‐tuned models F1 and F2, which were trained on Datasets 1 and 2, respectively, after pretraining on Dataset 3.

• 

  FIGURE 4. Model performance of scratch and fine‐tuning models in datasets with institutional variations. Area under the curve (AUC) values of the receiver operating characteristic (ROC) curve, sensitivity, and specificity are shown using color‐coded panels representing different value ranges. Results are presented separately for the three testing datasets (A1: Tokai‐All, A2: Tokai‐Cardiol, and A3: Shinken). For each dataset, the performance of five models is displayed: three scratch models (S1: Tokai‐All, S2: Tokai‐Cardiol, and S3: Shinken) and two fine‐tuned models (F1: Tokai‐All fine‐tuned on the pretrained Shinken model, and F2: Tokai‐Cardiol fine‐tuned on the pretrained Shinken model).

• 

  FIGURE 5. Model performance of scratch and fine‐tuning models in datasets with ECG‐system variations. Area under the curve (AUC) values of the receiver operating characteristic (ROC) curve, sensitivity, and specificity are shown using color‐coded panels representing different value ranges. Results are presented separately for the three testing datasets (B1: GE original, B2: GE‐NK intermediate, B3: NK feature). For each dataset, the performance of five models is displayed: three scratch models (S1: Tokai‐All, S2: Tokai‐Cardiol, and S3: Shinken) and two fine‐tuned models (F1: Tokai‐All fine‐tuned on the pretrained Shinken model, and F2: Tokai‐Cardiol fine‐tuned on the pretrained Shinken model).