A Dataset and Benchmarks for Atrial Fibrillation Detection from Electrocardiograms of Intensive Care Unit Patients

1 A Dataset and Benchmarks for Atrial Fibrillation Detection from Electrocardiograms of Intensive Care Unit Patients Sarah Nassar, Nooshin Maghsoodi, Sophia Mannina, Shamel Addas, Stephanie Sibley, Gabor Fichtinger, David Pichora, David Maslove, Purang Abolmaesumi, and Parvin Mousavi Abstract— Objective: Atrial fibrillation (AF) is the most common cardiac arrhythmia experienced by intensive care unit (ICU) patients and can cause adverse health effects. In this study, we publish a labelled ICU dataset and benchmarks for AF detection. Methods: We compared machine learning models across three data-driven artificial intelligence (AI) approaches: feature-based classifiers, deep learning (DL), and ECG foundation models (FMs). This comparison addresses a critical gap in the literature and aims to pinpoint which AI approach is best for accurate AF detection. Electrocardiograms (ECGs) from a Canadian ICU and the 2021 PhysioNet/Computing in Cardiology Challenge were used to conduct the experiments. Multiple training configurations were tested, ranging from zero-shot inference to transfer learning. Results: On average and across both datasets, ECG FMs performed best, followed by DL, then feature-based classifiers. The model that achieved the top F1 score on our ICU test set was ECG-FM through a transfer learning strategy (F1=0.89). Conclusion: This study demonstrates promising potential for using AI to build an automatic patient monitoring system. Significance: By publishing our labelled ICU dataset 1 and performance benchmarks, this work enables the research community to continue advancing the state-of-the-art in AF detection in the ICU environment. Index Terms— Atrial fibrillation, intensive care unit, electrocardiography, machine learning, deep learning, foun- dation models. I. INTRODUCTION A. Motivation Atrial fibrillation (AF) is the most common cardiac arrhythmia and can lead to negative health outcomes such as heart failure and stroke [1]. The prevalence of AF is higher in the intensive care unit (ICU) than in S. Nassar is with the Department of Electrical and Computer Engineering at Queen’s University, Kingston, ON, Canada (e-mail: sarah.nassar@queensu.ca). N. Maghsoodi and P. Mousavi are with the School of Computing at Queen’s University, Kingston, ON, Canada. S. Mannina and S. Addas are with the Smith School of Business at Queen’s University, Kingston, ON, Canada. S. Sibley is with the Departments of Emergency Medicine and Critical Care Medicine at Queen’s University, Kingston, ON, Canada. G. Fichtinger is with the School of Computing and the Department of Electrical and Computer Engineering at Queen’s University, Kingston, ON, Canada. D. Pichora is with the Department of Surgery at Queen’s University, Kingston, ON, Canada. D. Maslove is with the Departments of Medicine and Critical Care Medicine at Queen’s University, Kingston, ON, Canada. P. Abolmaesumi is with the Department of Electrical and Computer Engineering at the University of British Columbia, Vancouver, BC, Canada. 1LinkToBeAdded the general population at up to 15-20% or more [2]– [4]. AF is primarily diagnosed by visually inspecting the electrocardiogram (ECG) reading of a patient and iden- tifying morphological irregularities such as inconsistent intervals between R peaks and missing P waves [5]. AF management in ICU patients presents a unique challenge as these patients are at higher risk of rapid health deterioration. However, ICU patients are connected to bedside monitors that continuously capture their ECG readings, allowing for automatic monitoring of their cardiac rhythm. Therefore, it is of utmost importance to leverage this continuous telemetry data to allow for timely initiation of treatment strategies. Given the continuous nature of ECG data capture, the amount of time, training, and experience needed to interpret ECGs, and the urgency required to treat patients with AF as soon as possible, artificial intelligence (AI) can be used to develop an automatic detection algo- rithm that can continuously process a patient’s ECG and classify whether AF is occurring. To facilitate this downstream application of real-time patient monitoring, different AI technologies need to be compared to find the best approach suitable for the ICU context. B. Problem The body of literature in AI-powered ECG-based ar- rhythmia detection is rich, with many diverse approaches. By exploring existing literature, it can be noted that deep learning (DL) approaches are the most common [6]. Classical machine learning (ML) approaches, which re- quire hand-crafted features to be extracted from ECGs, are less common, even though the ECG signal modality has distinguishable features that clinicians use to iden- tify arrhythmia and that can be quantified. Additionally, classical ML is typically less resource-intensive than DL, which is an important consideration for real-time deployment contexts. Further, ECG foundation models (FMs) are an emerging trend [7], [8]. FMs are l

This content is AI-processed based on open access ArXiv data.