title	Pulsewatcher
emoji	❤️
colorFrom	red
colorTo	blue
sdk	docker
app_file	app.py
pinned	false

❤️ PulseWatcher — ECG Anomaly Detection

Cardiac arrhythmia detection using unsupervised deep learning. Trained exclusively on normal heartbeats — anything the model can't reconstruct is flagged as an anomaly.

---

🎯 What It Does

PulseWatcher monitors ECG signals and detects abnormal heartbeat patterns in real time. Unlike traditional classifiers, it never needs labeled anomaly data — it learns only what normal looks like, and flags anything that deviates.

---

🧠 How It Works

MIT-BIH ECG Data (48 real patient recordings) ↓ Signal Preprocessing + R-peak segmentation ↓ LSTM Autoencoder — trained ONLY on normal beats ↓ Reconstruction Error per beat ↓ Threshold = 95th percentile of normal errors ↓ High error → ANOMALY ALERT 🚨 ↓ Per-timestep explainability chart

📊 Results

Metric	Score
Precision	95.21%
Recall	44.64%
F1 Score	60.79%
ROC-AUC	0.8678
Training Loss	0.001584
Normal Beats	74,771
Anomalous Beats	33,308

ROC Curve

Confusion Matrix

Why is Recall 44%? Deliberate tradeoff. In cardiac screening, false positives cause unnecessary panic and expensive tests. We optimized for precision first — when PulseWatcher alerts, it is right 95% of the time. Recall is tunable by adjusting the detection threshold.

What does ROC-AUC 0.8678 mean? Random classifier = 0.50 | PulseWatcher = 0.8678 | Perfect = 1.00 Strong discriminative power — without ever seeing a single anomalous beat during training.

---

🏗️ Architecture

Model: LSTM Autoencoder ECG Signal (187 timesteps) ↓ LSTM Encoder (187 → 64 → 32 hidden units) ↓ LSTM Decoder (32 → 64 → 187 reconstruction) ↓ Reconstruction Error (MSE) ↓ Threshold (95th percentile) ↓ NORMAL ✅ or ANOMALY 🚨

Why LSTM? ECG is time-series data. LSTM has memory across time steps — it learns the temporal shape of a normal heartbeat (P-wave → QRS complex → T-wave). A regular neural network treats each timestep independently and cannot do this.

Why Autoencoder? In real clinical settings, anomalous heartbeats are rare. You cannot collect enough labeled anomaly data to train a classifier. The autoencoder learns only from normal data — making it the correct real-world approach.

Why not UCI Heart Disease dataset? UCI is tabular data with pre-extracted features. MIT-BIH is raw time-series ECG — the actual signal a cardiologist reads. Far more realistic and technically challenging.

---

🔍 Explainability

The dashboard shows per-timestep reconstruction error — exactly which part of the heartbeat the model struggled to reconstruct:

Region	Timesteps	Clinical Meaning
P-wave	0–40	Atrial depolarization
QRS Complex	60–100	Ventricular depolarization
T-wave	110–160	Ventricular repolarization

Red bars = timesteps the model could not reconstruct → where the anomaly lives.

---

📁 Project Structure

ecg-anomaly-detection/ ├── data/ │ ├── train.npy ← 59,816 normal beats for training │ ├── test.npy ← 14,955 normal beats for evaluation │ └── anomaly.npy ← 33,308 anomalous beats ├── models/ │ ├── lstm_autoencoder.pt ← Trained model weights │ └── threshold.npy ← Decision threshold ├── outputs/ │ ├── roc_curve.png ← ROC curve plot │ └── confusion_matrix.png ← Confusion matrix plot ├── src/ │ ├── preprocess.py ← Data download + beat segmentation │ ├── model.py ← LSTM Autoencoder architecture │ ├── train.py ← Training loop │ └── evaluate.py ← Metrics + plot generation ├── dashboard/ │ └── app.py ← Streamlit web dashboard ├── Dockerfile └── README.md

---

🗃️ Dataset

MIT-BIH Arrhythmia Database — PhysioNet

• 48 half-hour ECG recordings from real patients
• Sampled at 360 Hz
• Manually annotated by cardiologists
• Gold standard benchmark in published cardiology research since 1980

---

🚀 Run Locally

git clone https://github.com/shivams496/PulseWatcher.git
cd PulseWatcher

python -m venv venv
venv\Scripts\activate        # Windows
source venv/bin/activate     # Mac/Linux

pip install -r requirements.txt

python src/preprocess.py
python src/train.py
python -m src.evaluate
streamlit run dashboard/app.py

---

🛠️ Tech Stack

Tool	Purpose
PyTorch	LSTM Autoencoder
Streamlit	Web dashboard
Plotly	Interactive ECG charts
scikit-learn	Metrics + evaluation
wfdb	MIT-BIH data loading
Matplotlib	Plot generation

---

👨‍💻 Author

Shivam Singh — B.Tech CSE Final Year Built from scratch as Final Year Project — May 2026

Live Demo GitHub