Siamese Neural Networks for Small Dataset Classification of Electrograms
Bram Hunt, Eugene Kwan, Derek James Dosdall, Rob MacLeod, Ravi Ranjan · 2021 Computing in Cardiology (CinC) · 2021
Objective: In this work, we aimed to isolate endocardial electrograms with highly distinguishable activations in a small atrial fibrillation (AF) dataset by leveraging Siamese neural networks. Methods: Unipolar endocardial electrograms were captured with a basket catheter from a paced canine model of persistent AF, and 1,006 unique electrograms were randomly sampled from the endocardial wall and we isolated those with highly distinguishable activations. We trained Siamese neural networks to compare pairs of samples and then to classify electrograms in the testing dataset. Results: Using a reference from the validation dataset. the Siamese neural networks achieved a weighted accuracy of 90.3% and an F1 score of 0.94 on the classification testing dataset. Conclusion: Even in an electrogram dataset with significant size constraints, we achieved high accuracy and improved over a conventional neural network classifier weighted accuracy of 71.4%.