AI-driven clustering and visualization of electrocardiogram signals to enhance screening for atrial fibrillation: The supermarket/hypermarket opportunistic screening for atrial fibrillation study

Figures

• 

  Figure 1. A: Displays the data in the PDF format directly collected from the medical device. B: Displays the digitized waveforms extracted from the PDFs that were used in the analysis. AF = atrial fibrillation.

• 

  Figure 2. Outlines the workflow to go from the original ECG in PDF format to the extracted features that will be used for model development. ECG = electrocardiogram.

• 

  Figure 3. Proposed artificial intelligence–based methodology that builds on the approach in 23. The modification to the approach is contained within the green dashed box and demonstrates how the magnification factors are generated from the GTM model, used to identify neighborhood constraints, which are then used to influence the hierarchical clustering applied to the reference vectors. GTM = generative topographic mapping.

• 

  Figure 4. A: Membership map representing the latent space generated from the generative topographic mapping model trained on the features extracted from the athlete’s ECG rhythm strips. B: A selection of reference map visualizations demonstrating how 4 of the variables used to train the generative topographic mapping affect the cluster distribution in the latent space. ECG = electrocardiogram.

• 

  Figure 5. Magnification map calculated from the trained generative topographic mapping. Light areas of this map correspond to areas of low distortion during mapping, with the darker areas relating to areas of high distortion during the mapping.

• 

  Figure 6. Derived subgroups of athletes using data extracted from their ECG snippets. A: Membership map with a uniform size for the microclusters to show the distribution of the macrocluster regions. B: The size of the microclusters in the membership map, dictated by the number of ECG snippets assigned to it. ECG = electrocardiogram.

• 

  Figure 7. The top 3 plots display the ECG diagnoses overlaid on the membership map. The bluer a region is, the larger the percentage of ECG snippets in each node that came from an ECG diagnosis with the outlined condition. AF = atrial fibrillation; ECG = electrocardiogram.

• 

  Figure 8. Confusion matrix showing the performance of our approach when predicting normal, possible AF, and AF electrocardiograms. AF = atrial fibrillation.

• 

  Figure 9. Displays the final output from our proposed methodology. Each ECG is projected onto the trained membership map to see where each snippet is assigned and how the model’s decision changes over time. A probability distribution is also produced alongside this projection to provide the user with the area of the membership map most responsible for generating that ECG. Alongside the projections, the actual PDF ECG is provided with the highlighted sections to allow the user to check the model’s cluster assignments. A: Shows the output for an ECG recorded from a participant diagnosed as normal. B: Shows the output for the ECG recorded on a participant diagnosed as having AF. AF = atrial fibrillation; ECG = electrocardiogram.