[Paper Review] A Deep Bayesian Neural Network for Cardiac Arrhythmia Classification with Rejection from ECG Recordings

Wenrui Zhang, Xinxin Di+4 more|arXiv (Cornell University)|Feb 26, 2022

ECG Monitoring and Analysis4 citations

TL;DR

This paper proposes a deep Bayesian neural network with Monte Carlo dropout to classify cardiac arrhythmias from 12-lead ECGs while enabling prediction rejection based on combined data and model uncertainty. By rejecting predictions with high uncertainty, the method improves classification accuracy on accepted cases, demonstrating that uncertainty estimation enhances clinical reliability and supports safer deployment of deep learning in ECG interpretation.

ABSTRACT

With the development of deep learning-based methods, automated classification of electrocardiograms (ECGs) has recently gained much attention. Although the effectiveness of deep neural networks has been encouraging, the lack of information given by the outputs restricts clinicians' reexamination. If the uncertainty estimation comes along with the classification results, cardiologists can pay more attention to "uncertain" cases. Our study aims to classify ECGs with rejection based on data uncertainty and model uncertainty. We perform experiments on a real-world 12-lead ECG dataset. First, we estimate uncertainties using the Monte Carlo dropout for each classification prediction, based on our Bayesian neural network. Then, we accept predictions with uncertainty under a given threshold and provide "uncertain" cases for clinicians. Furthermore, we perform a simulation experiment using varying thresholds. Finally, with the help of a clinician, we conduct case studies to explain the results of large uncertainties and incorrect predictions with small uncertainties. The results show that correct predictions are more likely to have smaller uncertainties, and the performance on accepted predictions improves as the accepting ratio decreases (i.e. more rejections). Case studies also help explain why rejection can improve the performance. Our study helps neural networks produce more accurate results and provide information on uncertainties to better assist clinicians in the diagnosis process. It can also enable deep-learning-based ECG interpretation in clinical implementation.

Motivation & Objective

• To address the clinical limitation of deep learning models in ECG interpretation, which provide no confidence estimates for predictions.
• To improve diagnostic reliability by incorporating uncertainty estimation into automated ECG classification.
• To enable prediction rejection—excluding uncertain cases from automated output—so clinicians can focus on high-risk or ambiguous cases.
• To evaluate whether uncertainty-based rejection improves performance on accepted predictions compared to full acceptance.
• To provide interpretability by linking high uncertainty to clinically challenging cases, such as label disagreement or noisy signals.

Proposed method

• A 61-layer deep Bayesian convolutional neural network is trained on a real-world 12-lead ECG dataset from the 2018 CPSC challenge.
• Monte Carlo dropout is applied during inference to estimate model uncertainty by sampling multiple forward passes.
• Data uncertainty is estimated from the variance of predicted class probabilities across dropout samples.
• Total uncertainty is computed as the sum of data uncertainty and model uncertainty for each prediction.
• A rejection threshold is applied: predictions with total uncertainty above the threshold are rejected and not classified.
• The method is evaluated across varying rejection thresholds, and case studies are conducted with a cardiologist to interpret high-uncertainty and misclassified cases.

Experimental results

Research questions

• RQ1 Can uncertainty estimation from Bayesian deep learning improve the reliability of automated ECG arrhythmia classification?
• RQ2 Does rejecting predictions with high uncertainty lead to improved performance on the remaining accepted predictions?
• RQ3 What are the clinical causes of high uncertainty, such as label disagreement or signal noise, and how do they affect model predictions?
• RQ4 How does the performance of the model change as the rejection ratio increases?
• RQ5 Can uncertainty-based rejection support safer clinical deployment by prioritizing difficult or ambiguous cases for expert review?

Key findings

• Correctly classified samples consistently exhibited significantly lower uncertainty (p < 0.01) than incorrectly classified samples, confirming that uncertainty correlates with prediction reliability.
• As the rejection threshold decreased (i.e., more predictions were rejected), the precision of accepted predictions increased across all arrhythmia types, demonstrating improved performance on retained cases.
• Rejected samples were more likely to be misclassified, indicating that uncertainty effectively identifies high-risk or ambiguous cases for clinical review.
• Case studies revealed that high data uncertainty often arose from noisy ECG signals or label disagreement among experts, especially in mixed or borderline cases.
• Incorrect predictions with low uncertainty were primarily due to inconsistent or incomplete labels, such as cases labeled for one arrhythmia when multiple co-occur.
• The method enables dynamic thresholding based on disease severity, allowing higher safety margins for life-threatening arrhythmias.