PulsePath

A deep learning system that predicts supraventricular arrhythmias (SVA) 12 seconds before onset in patients with Wolff-Parkinson-White (WPW) Syndrome, using a hybrid GRU-Transformer chain model trained on real clinical ECG data.

Background

WPW Syndrome affects up to 24 million people worldwide. While many remain asymptomatic, approximately half will experience dangerous arrhythmias, often with little to no warning. Existing solutions are either reactive (Apple Watch), invasive (implantable loop recorders), or prohibitively expensive (up to $40,000). PulsePath's model provides a non-invasive, predictive alternative.

How It Works

The model uses a chain pipeline that runs two sequential classifiers on 12-second ECG windows:

1. Model 1 (Normal vs. Abnormal): filters out normal sinus rhythm using a GRU + Self-Attention architecture
2. Model 2 (SVA vs. Pre-SVA): classifies abnormal signals as either active SVA or pre-SVA, using the same architecture enhanced with Layer Normalization

This chain approach lets each model specialize in its own feature space, outperforming a single 3-class classifier.

Results

Metric	Value
Accuracy	95.26%
F1 Macro-Average	91.99%
Pre-SVA Sensitivity	90.03%
False Positive Rate	4.74%
AUROC	97.32%

Repository Structure

File	Description
downloading_ecg.ipynb	Downloads the MIT-BIH SVDB dataset from PhysioNet and segments it into labeled 12-second windows
train_test_split.ipynb	Creates and saves a stratified train/test split from the processed dataset
model1_normal_vs_abnormal.ipynb	Trains Model 1 (normal vs. abnormal)
model2_sva_vs_presva.ipynb	Trains Model 2 (SVA vs. pre-SVA)
chain_model.ipynb	Evaluates the full chain model pipeline and generates the confusion matrix

Running the Notebooks

These notebooks are designed to run on Google Colaboratory with a GPU runtime (we used an A100). To get started:

1. Run downloading_ecg.ipynb to generate ecg_data.npz
2. Run train_test_split.ipynb to generate train_test_split.npz
3. Train both models using model1_normal_vs_abnormal.ipynb and model2_sva_vs_presva.ipynb
4. Evaluate the full pipeline with chain_model.ipynb

Dataset

ECG data sourced from the MIT-BIH Supraventricular Arrhythmia Database (SVDB) via PhysioNet, containing 78 records from 66 patients with documented SVA episodes.

Credits

Built by Ziya Ahmad and Netra Khot for the 2025-26 Synopsys Science Fair.