Related Experiment Video
Updated: Jul 8, 2025
08:10
Estimating Bilateral Atrial Function by Cardiovascular Magnetic Resonance Feature Tracking in Patients with Paroxysmal Atrial Fibrillation
Published on: July 20, 2022
1.7K
Assessing the Generalizability of a Deep Learning-based Automated Atrial Fibrillation Algorithm
Summary
A novel hybrid deep learning model for automated atrial fibrillation (AF) detection shows acceptable generalizability from electrocardiogram (ECG) traces. Retraining the model significantly improved accuracy for clinical use.
Area of Science:
- Cardiology
- Artificial Intelligence
- Medical Informatics
Background:
- Automated detection of atrial fibrillation (AF) from electrocardiogram (ECG) is vital for post-stroke telemonitoring.
- Existing deep learning (DL) algorithms face challenges in generalizability across different datasets and devices.
- Accurate AF detection requires robust classification of normal sinus rhythm (NSR), AF, other rhythms (OR), and noisy (TN) ECG recordings.
Purpose of the Study:
- To develop and evaluate a hybrid deep learning (HDL) model for automated ECG classification.
- To assess the generalizability of the HDL model on a novel dataset collected using a handheld ECG device.
- To determine the impact of retraining on the model's performance and accuracy.
Main Methods:
- Developed a novel hybrid deep learning (HDL) model utilizing the PhysioNet/CinC Challenge 2017 dataset.
- Classified ECG recordings into four categories: NSR, AF, OR, and TN.
- Tested the pre-trained HDL model on 636 ECG samples from 102 outpatients and inpatients using a CONTEC PM10 Portable ECG Monitor.
- Retrained the HDL model on the newly collected dataset using 5-fold cross-validation.
Main Results:
- The HDL model achieved an average test F1-score of 0.892 on the PhysioNet/CinC Challenge 2017 dataset.
- On the newly collected dataset, the pre-trained HDL model achieved an average F1-score of 0.722 (AF: 0.905, NSR: 0.791, OR: 0.471, TN: 0.342).
- After retraining on the new dataset, the average F1-score significantly increased to 0.961.
Conclusions:
- The developed HDL-based algorithm demonstrates acceptable generalizability for AF detection from short-term single-lead ECG traces.
- Retraining the HDL model with new, specific dataset parameters substantially enhances its accuracy and clinical applicability.
- The findings support the use of DL models for AF detection in telemonitoring, with an emphasis on dataset-specific fine-tuning.
