Deep Transfer Learning for Detection of Atrial Fibrillation Using Holter ECG Color Maps

This study aims to explore the potential of deep transfer learning of pretrained image recognition deep neural networks (DNN) in a new diagnostic modality of atrial fibrillation (AF) using Holter ECG color maps.These maps consist of stacked color bars, transforming the ECG waveform amplitude of sequential heartbeats over 30 s into a color code, similar to ECHOView TM (Schiller AG, Switzerland).Using 13,536 training,6,624 validation and 11,400 test ECHOView color maps evenly split between AF and non-AF cases from the IRIDIA-AF Holter ECG monitoring database, we retrained, fine-tuned, and tested 18 image classification DNNs.Test accuracy of retrained DNNs (93.1-95.8%)was improved after fine-tunning (96.3-97.6%).Retrained EfficientNetV2B1 and finetuned MobileNetV2(=1.4) were both accurate and fast.ECHOView colormap images, composed of stacked heartbeat amplitudes, provide an interpretable visual modality for pretrained image classification networks that can be effectively retrained and fine-tuned to identify short AF episodes in Holter-ECG recordings.