ecg-anomaly-maf

ECG Anomaly Detection — CNN + MAF

Unsupervised cardiac arrhythmia detection on ECG signals, built entirely from scratch in PyTorch.

A 1D CNN encoder compresses each heartbeat into noise-robust features. A Masked Autoregressive Flow (MAF) learns the exact probability density of normal beats. Any beat with low log p(x) is flagged as anomalous — no anomaly labels needed during training.

Live Demo

Try the model interactively in the browser: 👉 HF Spaces — Live Demo

Results

Evaluated on 21,892 test beats from the MIT-BIH Arrhythmia Database (all 5 classes). Trained on normal beats only — the model never sees any anomaly during training.

How to use

from huggingface_hub import hf_hub_download
import torch
import json

# Download weights and config
model_path  = hf_hub_download("Aibygaurav/ecg-anomaly-maf", "best_model.pt")
config_path = hf_hub_download("Aibygaurav/ecg-anomaly-maf", "train_config.json")

with open(config_path) as f:
    config = json.load(f)

# Load model (requires src/ from GitHub repo)
from src.hybrid_model import HybridECGModel

model = HybridECGModel(
    input_len   = 187,
    latent_dim  = config["latent_dim"],
    n_layers    = config["n_layers"],
    hidden_dims = config["hidden_dims"],
)
model.load_state_dict(torch.load(model_path, map_location="cpu", weights_only=True))
model.eval()

# Score a beat — higher = more normal, lower = anomaly
import torch
beat = torch.randn(1, 187)   # replace with a real normalised ECG beat
with torch.no_grad():
    log_prob = model.log_prob(beat).item()

print(f"log p(x) = {log_prob:.2f}")

Architecture

Input: 187 time steps (one heartbeat)
         ↓
  CNN Encoder
    Conv1D blocks → 64-dimensional feature vector
         ↓
  MAF (8 layers, alternating reversal)
    Each layer: MADE with hidden dims [512, 512]
         ↓
  z ~ N(0, I)   ← base distribution

Two-stage training:

1. CNN autoencoder pretraining with MSE reconstruction loss (prevents encoder collapse)
2. MAF density estimation on frozen CNN features (NLL loss, normal beats only)

Dataset

MIT-BIH Arrhythmia Database — preprocessed CSV version (Kaggle)

Built from scratch

No external normalising flow libraries. All components implemented manually:

• MaskedLinear — linear layer with binary autoregressive weight masks
• MADE — Masked Autoencoder for Distribution Estimation
• MAFLayer — single affine flow transformation with O(D) log-determinant
• MAF — stacked flow with exact log_prob and sample
• ECGEncoder — 1D CNN compressing 187-step waveform to 64 features

References

• MADE: Germain et al., 2015 — arxiv.org/abs/1502.03509
• MAF: Papamakarios et al., 2017 — arxiv.org/abs/1705.07057
• Dataset: Kachuee et al., 2018 — arxiv.org/abs/1805.00794

Links

• GitHub: aibygauravkumar/ecg-anomaly-maf
• Live Demo: HF Spaces