Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

An Attention-Augmented VAE-BiLSTM Framework for Anomaly Detection in 12-Lead ECG Signals

Marc Garreta Basora, Mehmet Oguz Mulayim

TL;DR

The paper tackles unsupervised anomaly detection in multivariate 12-lead ECGs by comparing three autoencoder-based architectures: CAE, VAE-BiLSTM, and the novel VAE-BiLSTM-MHA. Training is performed solely on normal ECGs to learn the healthy morphology, with anomaly scores derived from reconstruction or ELBO-based criteria and thresholded to flag deviations. The VAE-BiLSTM-MHA model, incorporating lead-wise and multi-head attention, achieves the best performance (AUPRC ≈ 0.81 and recall ≈ 0.85 on CPSC 2018) and is paired with an interactive dashboard for anomaly localization, enhancing clinical interpretability. The study provides a public codebase and highlights limitations such as evaluation on a single benchmark and offline deployment, pointing to future work in real-time analysis and broader clinical data integration.

Abstract

Anomaly detection in 12-lead electrocardiograms (ECGs) is critical for identifying deviations associated with cardiovascular disease. This work presents a comparative analysis of three autoencoder-based architectures: convolutional autoencoder (CAE), variational autoencoder with bidirectional long short-term memory (VAE-BiLSTM), and VAE-BiLSTM with multi-head attention (VAE-BiLSTM-MHA), for unsupervised anomaly detection in ECGs. To the best of our knowledge, this study reports the first application of a VAE-BiLSTM-MHA architecture to ECG anomaly detection. All models are trained on normal ECG samples to reconstruct non-anomalous cardiac morphology and detect deviations indicative of disease. Using a unified preprocessing and evaluation pipeline on the public China Physiological Signal Challenge (CPSC) dataset, the attention-augmented VAE achieves the best performance, with an AUPRC of 0.81 and a recall of 0.85 on the held-out test set, outperforming the other architectures. To support clinical triage, this model is further integrated into an interactive dashboard that visualizes anomaly localization. In addition, a performance comparison with baseline models from the literature is provided.

An Attention-Augmented VAE-BiLSTM Framework for Anomaly Detection in 12-Lead ECG Signals

TL;DR

The paper tackles unsupervised anomaly detection in multivariate 12-lead ECGs by comparing three autoencoder-based architectures: CAE, VAE-BiLSTM, and the novel VAE-BiLSTM-MHA. Training is performed solely on normal ECGs to learn the healthy morphology, with anomaly scores derived from reconstruction or ELBO-based criteria and thresholded to flag deviations. The VAE-BiLSTM-MHA model, incorporating lead-wise and multi-head attention, achieves the best performance (AUPRC ≈ 0.81 and recall ≈ 0.85 on CPSC 2018) and is paired with an interactive dashboard for anomaly localization, enhancing clinical interpretability. The study provides a public codebase and highlights limitations such as evaluation on a single benchmark and offline deployment, pointing to future work in real-time analysis and broader clinical data integration.

Abstract

Anomaly detection in 12-lead electrocardiograms (ECGs) is critical for identifying deviations associated with cardiovascular disease. This work presents a comparative analysis of three autoencoder-based architectures: convolutional autoencoder (CAE), variational autoencoder with bidirectional long short-term memory (VAE-BiLSTM), and VAE-BiLSTM with multi-head attention (VAE-BiLSTM-MHA), for unsupervised anomaly detection in ECGs. To the best of our knowledge, this study reports the first application of a VAE-BiLSTM-MHA architecture to ECG anomaly detection. All models are trained on normal ECG samples to reconstruct non-anomalous cardiac morphology and detect deviations indicative of disease. Using a unified preprocessing and evaluation pipeline on the public China Physiological Signal Challenge (CPSC) dataset, the attention-augmented VAE achieves the best performance, with an AUPRC of 0.81 and a recall of 0.85 on the held-out test set, outperforming the other architectures. To support clinical triage, this model is further integrated into an interactive dashboard that visualizes anomaly localization. In addition, a performance comparison with baseline models from the literature is provided.

Paper Structure

This paper contains 20 sections, 11 equations, 11 figures, 7 tables.

Table of Contents

  1. Introduction
  2. State of the Art
  3. Methodology
  4. Problem Formulation
  5. Data Preprocessing
  6. Input Data
  7. Window Segmentation
  8. Filtering and Normalization
  9. Model Architectures
  10. Anomaly Detection Task
  11. Decision Rule
  12. Experiments
  13. Datasets
  14. Evaluation Metrics
  15. Precision, Recall, and F1 Score:
  16. ...and 5 more sections

Figures (11)

  • Figure 1: ECG morphology of two normal beats. Reproduced from Zhang et al. zhang2015robust with permission.
  • Figure 2: Approaches for Anomaly Detection (AD) in ECGs
  • Figure 3: Example of a raw 12-lead ECG sample from the MIMIC-IV ECG dataset. The red and black boxes show consecutive windows extracted for training.
  • Figure 4: Comparison of non-processed and preprocessed ECG signals
  • Figure 5: CAE architecture
  • ...and 6 more figures