RL-Driven Sequential Matched Filtering for ECG R-peak Detection

_{SMF sequentially applies matched filters (MFs) designed by a signal-aware neural network policy, and identifies ECG R-peaks at the final step. The policy is trained end-to-end using Reinforcement Learning (RL).}

---

Overview

This is the official implementation of the Sequential Matched Filter (SMF) algorithm, presented in:

Machine Intelligence on the Edge: Interpretable Cardiac Pattern Localisation Using Reinforcement Learning
Haozhe Tian, Qiyu Rao, Nina Moutonnet, Pietro Ferraro, Danilo Mandic
Machine Intelligence Research [arXiv]

SMF is a lightweight, interpretable ECG R-peak detector that uses matched filters (MFs) to iteratively refine signals. SMF bridges classical signal processing and reinforcement learning, offering an interpretable alternative to black-box deep models for ECG analysis. You can run a quick, cloud-based trial of SMF via Code Ocean. (If the link does not open directly, please copy and paste https://doi.org/10.24433/CO.8370367.v1 into your browser.)

---

Requirements

Tested on Ubuntu 22.04 with Python 3.8+. Set up the conda environment by running:

conda env create -f environment.yml
conda activate SMF

---

Usage

Quick Comparison of Pan-Tompkins, MF, and SMF

python compare.py

R-peak Detection with Pre-trained Models

A demonstration notebook test.ipynb is provided for running inference with trained checkpoints. The core is a validate function:

from agents.ppo import PPO
from env import SMF

# Initialise environment: 3 MF steps, template length 8
env = SMF(eps_len=3, template_len=8)

# Load pre-trained PPO agent
agent = PPO(env, ckpt_path='saved/eps_len_3/PPO_best.pt', use_tb=False)

# Run validation
ecg, sig, peaks, preds, TP, FP, FN = validate(env, agent)

Training from Scratch

Argument	Description
eps_len	Number of sequential MF steps
template_len	Length of each matched filter template
ckpt_path	Path to a saved model checkpoint

python train.py --agent ppo --eps_len 3 --template_len 8

---

Citation

If you find the code in this repository useful, please cite:

@article{tian2025machine,
  title={Machine Intelligence on the Edge: Interpretable Cardiac Pattern Localisation Using Reinforcement Learning},
  author={Tian, Haozhe and Rao, Qiyu and Moutonnet, Nina and Ferraro, Pietro and Mandic, Danilo},
  journal={arXiv preprint arXiv:2508.21652},
  year={2025}
}

If you use the PhysioNet/CinC 2017 dataset included in this repository, please also cite:

@inproceedings{clifford2017af,
  title={AF Classification from a Short Single Lead ECG Recording: The PhysioNet/Computing in Cardiology Challenge 2017},
  author={Clifford, Gari D and Liu, Chengyu and Moody, Benjamin and Lehman, Li-wei H and Silva, Ikaro and Li, Qiao and Johnson, Alistair E and Mark, Roger G},
  booktitle={2017 Computing in Cardiology (CinC)},
  pages={1--4},
  year={2017},
  organization={IEEE}
}