Thursday, November 20, 2025 at 1-2pm ET
This NISS-CANSSI Collaborative Data Science CoLab Webinar presented research on implementing AI and machine learning for atrial fibrillation detection in the ICU, with presenters from Queen's University discussing their interdisciplinary approach to developing context-aware solutions. The team's research focused on various computational methods for arrhythmia detection, including feature-based models, deep learning approaches, and self-supervised learning models for ECG data analysis. Their work culminated in developing an AI algorithm to detect atrial fibrillation in ICU patients, which includes conducting focus groups with healthcare providers and exploring digital nudges for implementation.
ECG Deep Learning in ICU
Joel introduced the NISS-CANSSI Collaborative Data Science CoLab Webinar Series presentation on deep learning with ECG data in the ICU, titled "Modeling to Actionable AI." He introduced the presenters, Dr. Parvin Mousavi and Dr. David Maslove, both from Queen's University, who specialize in computer-aided diagnosis and interventions, and precision medicine methods, respectively. Joel requested attendees to submit questions via the Q&A feature and mentioned that questions would be addressed at the end of the presentation.
AI Integration in ICU Monitoring
David Maslove, a clinician-scientist from Queen's University, introduced the research group and discussed their interdisciplinary research project focused on implementing AI and machine learning in the ICU environment. The team, composed of members from various departments including Medicine, Computing, Electrical and Computer Engineering, and the Smith School of Business, is addressing the complex challenges of integrating AI into the ICU workflow. Their specific use case involves monitoring for atrial fibrillation, a condition common in critically ill patients. David emphasized the importance of developing context-aware and domain-specific solutions that consider the ICU's unique challenges and the need for iterative improvement through plan-do-study-act cycles.
Atrial Fibrillation Prediction Tools
Stephanie Sibley, a clinician scientist at Queen's working in the ICU with David, discussed atrial fibrillation, an abnormal heart rhythm that affects critically ill patients. She explained its prevalence and impact on patient outcomes, highlighting the need for early prediction to prevent complications and reduce costs. Stephanie’s research focuses on atrial fibrillation outcomes and treatments, including a study using magnesium for prophylaxis. Both emphasized the importance of developing risk prediction tools to enrich patient samples for clinical trials, potentially reducing sample sizes and costs.
AFib Prediction with Deep Learning
Dr. Parvin Mousavi discussed computational approaches for arrhythmia detection, focusing on AFib prediction in the ICU setting. She highlighted the challenges of working with noisy ECG signals and limited expert-labeled data. Parvin outlined a methodology that progresses from feature-based machine learning models to deep learning methods, including the potential use of transfer learning and foundation models to improve prediction accuracy. She concluded by introducing Sarah Nassar, who would elaborate on some of these computational methods.
AI Models for Atrial Fibrillation Detection
Sarah Nassar presented her MASc research comparing three AI approaches for detecting atrial fibrillation using ECG data. She tested feature-based models, deep learning with convolutional neural networks (CNNs), and ECG foundation models across two datasets: an in-house ICU dataset and a public PhysioNet dataset. The ECG foundation models generally performed best, achieving an F1 score of 89% on the ICU test set after fine-tuning with transfer learning. Sarah plans to use the most accurate models to label vast amounts of unlabeled data for potential applications in atrial fibrillation prediction and forecasting.
Physiology-Aware ECG Model Development
Nooshin Maghsoodi presented her research on developing a physiology-aware self-supervised learning model for ECG data analysis. She explained the challenges of limited labeled ICU data and proposed using large amounts of unlabeled ECG data for model training. The model incorporates physiological ECG knowledge through contrastive learning and three strategies for positive and negative pair selection. Nooshin demonstrated the effectiveness of their approach by comparing it to other models on various ECG datasets, showing improved performance across all tested datasets.
AI-Driven AFib Detection System
The team presented their research on developing an AI algorithm to detect atrial fibrillation (AFib) in ICU patients, which will trigger alerts to healthcare providers. Sophia Mannina presented her research, which involved conducting focus groups with 21 ICU nurses and physicians at Kingston General Hospital to understand their needs for a new alert system, finding that providers want it to be distinct, accurate, and easy to notice. She discussed an approach using digital nudges, such as pop-up messages in electronic health records, to capture attention without overwhelming providers.
The team addressed questions about comparing their model to simpler methods, handling class imbalance, and ensuring no data leakage between training and testing sets. The project is funded by New Frontiers and Research Fund (NFRF), Canadian Institutes of Health Research (CIHR), and other organizations, with the goal of supporting ICU healthcare providers in preventing AFib through a clinically deployable AI system.
Thank You / Acknowledgements
The National Institute of Statistical Sciences (NISS) and CANSSI extend their sincere appreciation to our speakers—Dr. Parvin Mousavi, Dr. David Maslove, Dr. Stephanie Sibley, Sarah Nassar, Nooshin Maghsoodi, and Sophia Mannina—for sharing their leading-edge research and insight into AI-enabled atrial fibrillation detection in critical care settings. Their interdisciplinary contributions underscore the potential of data science to drive meaningful clinical impact. We also thank our moderator, Joel Dubin, for guiding the session and fostering an engaging discussion. Finally, heartfelt thanks to all attendees for joining us and supporting the NISS-CANSSI Collaborative Data Science CoLab Series.