Atrial Fibrillation Detection from Electrocardiograms of Intensive Care Unit Patients: A Comparison of Artificial Intelligence Approaches

Abstract

Atrial fibrillation (AF) is the most common cardiac arrhythmia experienced by intensive care unit (ICU) patients and can cause adverse health effects. Prompt and accurate detection of AF in ICU patients is crucial for positive health outcomes. In my research, artificial intelligence (AI) models were trained to automatically process electrocardiogram (ECG) signals captured by bedside monitors in the ICU and detect AF with improved performance over previous research. ICU data from the Kingston General Hospital (KGH) as well as multi-site data from the public 2021 PhysioNet/Computing in Cardiology (CinC) Challenge were used to compare various model types, including feature-based classifiers, deep learning convolutional neural networks (CNNs), and ECG foundation models. The feature-based classifiers included classical machine learning algorithms and a tabular foundation model named TabPFN v2. The CNNs included the Goodfellow CNN and Stanford CNN, which accept raw one-dimensional (1-D) signals, as well as ResNet-18 and Inception v3, which accept two-dimensional (2-D) images of plotted ECG signals or signal-to-image transformations including spectrograms, scaleograms, and recurrence plots. The pre-trained ECG foundation models were ECG-FM and ECGFounder. Four training configurations were used. First, the ECG foundation models were tested in a zero-shot setting, which required no additional training. Then, separate models were trained with either a small portion of the ICU data (n=298) or the larger public dataset (n=70,363). Finally, the deep learning models and foundation models trained with the public dataset were further fine-tuned with the small ICU train set in a transfer learning strategy. On the ICU test set, the best F1 score achieved was 0.89 by ECG-FM with transfer learning. On the public PhysioNet dataset, the best F1 score achieved was 0.97 by a zero-shot version of ECGFounder. This research demonstrated that AF detection in the ICU with a fine-tuned ECG foundation model shows promising potential to be leveraged in building an automatic and accurate patient monitoring system. The top performing models trained can also be used to support large-scale non-expert ECG labelling for further use in AF forecasting research.