ECG Anomaly Detection

Overview

This project detects anomalies in 12-lead ECG signals by learning what a normal heartbeat looks like. A Variational Autoencoder is trained on FMM-filtered ECGs; reconstruction error and latent-space scores flag deviations, and an XGBoost classifier turns those scores into a decision with a calibrated threshold. It's trained and evaluated on two large public datasets (PTB-XL and Chapman-Shaoxing) with a full reproducible pipeline from preprocessing through evaluation on unseen data.

Highlights

• VAE learns a latent representation of normal ECGs; anomalies fall where reconstruction and latent scores diverge.
• FMM signal filtering as preprocessing, with an exploratory study of its effect on the waveforms.
• XGBoost decision head on the VAE scores, with tuned hyperparameters and a calibrated detection threshold.

Report