ECG Benchmark

Author's implementation of A Comprehensive Benchmark for Electrocardiogram Time-Series

[📄ACM MM 2025] [📄ARXIV] [⭐CODE] [📂DATA]

Introduction

Electrocardiogram (ECG) is crucial for assessing cardiac health and diagnosing various diseases. Given its time-series format, ECG data is often incorporated into pre-training datasets for large-scale time-series model training. However, existing studies often overlook its unique characteristics and specialized downstream applications, which differ significantly from other time-series data, leading to an incomplete understanding of its properties. In this paper, we present an in-depth investigation of ECG signals and establish a comprehensive benchmark, which includes (1) categorizing its downstream applications into four distinct evaluation tasks, (2) identifying limitations in traditional evaluation metrics for ECG analysis, and introducing a novel metric; (3) benchmarking SOTA time-series models and proposing a new architecture. Extensive experiments demonstrate that our proposed benchmark is comprehensive and robust. The results validate the effectiveness of the proposed metric and model architecture, which establish a solid foundation for advancing research in ECG signal analysis.

Updates

🚩 News (2025.12) Release evaluation code and our PSSM [🧭Quick Start].

🚩 News (2025.12) Release benchmark [📂DATA].

🚩 News (2025.7) Accepted by [📄ACM MM 2025].

TODO

• Release pre-training data and code
• Release large time-series models pre-trained in ECG

Quick Start

Step 0. Prepare Environment

Download [📂DATA] to Data folder, download [🤖FFD] model (for evaluation) to checkpoints folder.

pip install -r requirements.txt

Step 1. Run evaluation

python val_run.py \
    model=PSSM \
    data=ADFECGDB

Evaluate Your Custom Model

The ECG-Benchmark framework is shown below. You can follow the [🧭Tutorial] to evaluate your own model.

Benchmark

[Google Drive] [Quark Drive]

We define a set of evaluations designed for the broad range of medical applications of ECG, which include four downstream tasks:

1️⃣ Classification for disease diagnosis and event prediction,

2️⃣ Detection for key waveform (e.g., P-wave) localization,

3️⃣ Forecasting for ECG dynamics prediction,

4️⃣ Generation for maternal-fetal ECG separation.

These tasks provide a comprehensive framework for evaluation, and the following details the datasets used for each task in ECG-Benchmark. [Google Drive] [Quark Drive] have released the processed data used for evaluation.

Task	Name	Describe	# Sample	# Subject	Channel	Frequency
Classification	CPSC2018 [📕Paper]	7	205111	5847	12	500Hz
Classification	CPSC2021 [📕Paper]	3	330032	785	2	200Hz
Classification	AF [📕Paper]	2	8400	35	7	125Hz
Classification	NIFEADB [📕Paper]	2	21293	26	6	1000Hz
Classification	RFAA [📕Paper]	4	38640	315	4	250Hz
Classification	SPB [📕Paper]	6	34560	8	12	257Hz
Detection	FEPL [📕Paper]	Fetal QRS	37080	20	10	500Hz
Detection	CPSC2020 [📕Paper]	T-wave	175782	10	1	400Hz
Detection	MITDB [📕Paper]	QRS	20810	71	2	360Hz
Detection	MITPDB [📕Paper]	P-wave	4269	12	2	360Hz
Detection	NFE [📕Paper]	Fetal QRS	1200	25	4	1000Hz
Detection	SVDB [📕Paper]	QRS	56160	78	2	128Hz
Forecast	CPSC2019 [📕Paper]		2000	2000	1	500Hz
Forecast	RDBH [📕Paper]		22200	198	7	360Hz
Forecast	MIMICSub [📕Paper]		5552708	1210	8	125Hz
Forecast	NFE [📕Paper]		5467	125	4	1000Hz
Forecast	DALIA [📕Paper]		25887	15	2	100Hz
Forecast	PTB [📕Paper]		130455	516	12	1000Hz
Generation	ADFECGDB [📕Paper]	Maternal ECG	1200	5	5	1000Hz
Generation	FEPL [📕Paper]	Maternal ECG	7800	10	10	500Hz
Generation	BIDMC [📕Paper]	PPG	4992	52	7	125Hz
Generation	MITDB [📕Paper]	Noise ECG	116196	69	2	360Hz
Generation	PTBXL [📕Paper]	Noise ECG	262044	21837	12	500Hz
Generation	SST [📕Paper]	PCG	6760	338	9	1000Hz

Brothers link

• [💼 ECG-Benchmark]
• [🤖 ECG-Agent]
• [📖 A Comprehensive Benchmark for Electrocardiogram Time-Series]

Citation

If you find this repo helpful, please cite our paper.

@inproceedings{tang2025comprehensive,
  title={A Comprehensive Benchmark for Electrocardiogram Time-Series},
  author={Tang, Zhijiang and Qi, Jiaxin and Zheng, Yuhua and Huang, Jianqiang},
  booktitle={Proceedings of the 33rd ACM International Conference on Multimedia},
  pages={6490--6499},
  year={2025}
}

One More Thing