12-Lead ECG Anomaly Detection

This repository provides all the necessary material, including the final report and its progress reports, as well as code for model training, evaluation, visualization, and dashboard usage.

About

The project focuses on the unsupervised detection of anomalies in 12-lead electrocardiograms (ECGs) autoencoder-based architectures. It explores and compares three models:

• Convolutional Autoencoder (CAE)
• Variational Autoencoder Bidirectional Long Short-Term Memory(VAE-BiLSTM)
• Variational Autoencoder Bidirectional Long Short-Term Memory with Multi-Head Attention (VAE-BiLSTM-MHA)

In addition to modeling, the project also includes an interactive dashboard designed for interpretable visualization of the results. This web application allows users to analyze ECG signals and interpret possible anomalous patterns within a user-friendly environment.

---

Academic context

This code is part of my Final Degree Project in Artificial Intelligence (Universitat Autònoma de Barcelona, Escola d’Enginyeria).

• Author: Marc Garreta Basora – @marcgarreta
• Supervisor: Dr. Mehmet Oguz Mulayim – @omulayim
• Degree: Artificial Intelligence bachelor's degree
• University: Universitat Autònoma de Barcelona (UAB)
• Report: see /docs/final_report.pdf for the final report.
• TFG Title: Anomaly Detection in Patient Vital Signs for Early Warning of Critical Health Events

---

Model Architectures

CAE

A lightweight convolutional autoencoder that detects anomalies via reconstruction error on raw ECG signals. The architecture is inspired from the convolutional structure presented by Zhang et al. [1].

VAE-BiLSTM

A bidirectional LSTM-based variational autoencoder that captures global ECG deviations without attention mechanisms. This model is primarily inspired by the anomaly detection approach (OmniAnomaly) of Fu et al. [3], with theoretical foundations based on the variational inference concepts introduced by Kingma and Welling [2].

VAE-BiLSTM-MHA

A variational autoencoder enhanced with both lead-wise and temporal multi-head attention for highlighting when and where anomalies occur. This novel approach is inspired by the architecture proposed by Correia et al. [4], which applies a similar structure for anomaly detection in multivariate time-series data, although in a different data domain than ECG signals.

---

Dashboard

This user-friendly web application allow users to:

• Upload ECG signals in standard formats (.npy or .dat/.hea).
• Select between different models for analysis:
  • Convolutional Autoencoder (CAE)
  • VAE-BiLSTM-MHA
• Visualize the full 12-lead ECG with anomaly detection results.
• Adjust configuration options, such as:
  • Choosing specific leads to visualize.
  • Modifying the anomaly threshold (not recommended, as it is precomputed during model validation).

For each lead, the interface provides:

• The original signal (black) vs. reconstructed signal (red).
• The anomaly score (orange).
• A lead-wise attention heat-map
• A point-wise reconstruction error (MSE) color bar.

This setup enables users to not only detect anomalies but also interpret when and where they occur within the ECG signal.

---

Results

---

Data sources

Before you begin, download the ECG datasets in case of training:

• MIMIC-IV ECG (restricted)

  • URL: https://physionet.org/content/mimic-iv-ecg/1.0/
  • Access requires an authenticated PhysioNet account and approval to the MIMIC project.

• PTB-XL (open)

  • URL: https://physionet.org/content/ptb-xl/1.0.3/

---

Quick start

# clone repo
git clone https://github.com/marcgarreta/12-lead-ECG-AD
cd 12-lead-ECG-AD

# create & activate environment
conda env create -f environment.yml
conda activate ecg-anomaly-detection

# pre-process mimic data -- data will be stored in /data/processed/mimic/
cd src/data_processing
python pre-process.py --dataset mimic --input-dir {path_of_mimic-iv-ecg-diagnostic-electrocardiogram-matched-subset-1.0} --clean-nans

# pre-process ptbxl data -- data will be stored in /data/processed/ptbxl/
cd src/data_processing
python pre-process.py --dataset ptbxl --input-dir {path_of_ptb-xl-a-large-publicly-available-electrocardiography-dataset-1.0.3}

# training
cd src
python training.py --dataset [ptbxl or mimic or both]

# evaluation
python evaluation_metrics.py

# visualization (to visualize defined samples from the test set)
python visualization_cae.py #if you want to plot the cae visualizations
python visualization_vae.py #if you want to plot the vae visualizations

# ui (to run the local web application)
cd ui
python app.py 

---

License

This project is licensed under the MIT License. See the LICENSE file for more details.

---

References

1. T. Zhang, Y. Zhang, B. Zhang, H. Zhou, Q. Xu, Y. Zhang, and C. Xu, Unsupervised Deep Anomaly Detection for Multi-Sensor Time-Series Signals, arXiv preprint arXiv:2107.12626, 2021. [Online]. Available: https://arxiv.org/abs/2107.12626

2. D. P. Kingma and M. Welling, Auto-encoding variational Bayes, arXiv preprint arXiv:1312.6114, 2013. [Online]. Available: https://arxiv.org/abs/1312.6114

3. C. Fu, D. Zhang, P. Huang, S. Liu, Y. Qiao, and H. Xu, Robust Anomaly Detection for Multivariate Time Series via Generative Adversarial Networks, in Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (KDD '19), 2019, pp. 2828–2837.

4. L. Correia, J.-C. Goos, P. Klein, T. Bäck, and A. V. Kononova, MA-VAE: Multi-head Attention-based Variational Autoencoder for Anomaly Detection in Multivariate Time Series, arXiv preprint arXiv:2309.02253, 2023. [Online]. Available: https://arxiv.org/abs/2309.02253

---