Introduction
Atrial Fibrillation (AF) is the most prevalent heart rhythm disorder encountered in the Intensive Care Unit (ICU). In this high-acuity environment, patients are at constant risk of rapid physiological deterioration, making continuous electrocardiogram (ECG) monitoring a standard of care. However, the sheer volume of data generated by bedside monitors presents a significant challenge: interpreting these signals in real-time is labor-intensive and frequently leads to "alarm fatigue" among clinical staff due to high false-positive rates.
Automating AF detection in the ICU is complicated by two primary factors. First, existing public ECG datasets are largely derived from outpatient settings, which do not reflect the electrical and physical "noise" characteristic of an ICU. Second, there is a lack of consensus on which modeling approach—classical machine learning, standard deep learning, or the newer class of "Foundation Models"—is most effective for this specific clinical context. This research addresses these gaps by introducing a new ICU-specific dataset and conducting a comprehensive benchmark of 23 model variations.
The Kingston ICU Dataset
A central contribution of this work is the release of the Kingston ICU Dataset. This dataset consists of approximately 600 labeled 10-second, four-lead ECG segments collected from patients in Kingston, Ontario. Unlike broader datasets, this collection specifically captures the nuances of the ICU environment, providing a high-quality benchmark for evaluating how models perform in real-world, high-stakes clinical settings.
To supplement this ICU-specific data, the researchers also utilized the 2021 PhysioNet/CinC Challenge dataset, which contains over 70,000 12-lead ECGs. This larger corpus provided the "big data" necessary to train and pretrain complex deep learning architectures.
Methodology and Model Benchmarking
The study implemented a rigorous "four-way" training configuration to evaluate 23 different model variations across three distinct eras of artificial intelligence:
- Feature-Based Models (Classical ML): These models utilized 27 handcrafted features per lead, such as heart rate variability and entropy. Classifiers included XGBoost, LightGBM, and TabPFN.
- Deep Learning (DL): The team tested 1-D CNNs processing raw signals and 2-D CNNs processing signals converted into visual representations, such as spectrograms and recurrence plots.
- ECG Foundation Models (FMs): The study evaluated ECG-FM (a Transformer-based architecture) and ECGFounder (a CNN-based architecture). These models are pretrained on millions of heartbeats to develop a generalized understanding of cardiac signals before being fine-tuned for specific tasks.
These models were tested across four scenarios: zero-shot (no ICU-specific training), training only on small ICU data, training on massive public data, and transfer learning (pretraining on public data followed by fine-tuning on ICU data).
Key Insights
The Dominance of Foundation Models
Foundation Models proved to be the most robust performers. While standard Deep Learning models struggled when limited to the small ICU dataset, the FMs leveraged their "pre-existing knowledge" of cardiac signals to maintain high performance even with limited local data. The ECG-FM model, utilizing a transfer learning strategy, achieved the top F1 score of 0.89 on the ICU test set.
Technical Nuances in Signal Processing
For 2-D Deep Learning, the method of signal conversion was critical. The researchers found that converting ECGs into recurrence plots—a technique for visualizing periodic signals—significantly outperformed standard spectrograms. Additionally, in zero-shot settings for FMs, using a Sigmoid mapping for classification consistently outperformed Softmax, providing a vital technical detail for future deployments.
The Persistence of Classical ML
Despite the power of neural networks, classical models like KNN and XGBoost remained competitive. In scenarios with extremely limited data and low computational resources, these feature-based models often outperformed standard CNNs that lacked pretraining.
Clinical Implications and Results
The benchmark results suggest that automated systems are reaching a level of maturity suitable for clinical support. The top-performing models achieved a precision of up to 0.98. In an ICU setting, high precision is vital as it directly translates to fewer false alarms, mitigating the risk of alarm fatigue.
Furthermore, the models' F1 score of 0.89 compares favorably against human performance; previous studies have estimated average cardiologist F1 scores for AF detection to be between 0.67 and 0.83. This suggests that AI-driven detection can serve as a highly reliable safety net for clinicians.
Limitations and Future Directions
While the results are promising, the study's ICU-specific test set was relatively small, containing 50 positive AF samples. The current scope was also limited to binary classification (AF vs. Sinus Rhythm). Future research will focus on multi-class classification to distinguish AF from other arrhythmias like ventricular tachycardia and handling the intense physical noise of patient movement. The ultimate goal is the integration of these models into bedside monitors for real-time clinical trials and the development of predictive models to forecast AF before it occurs.
Summary updated