Publications

Parameter-Efficient 12-Lead ECG Reconstruction from a Single Lead

The 28th International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI 2025), Acceptance rate 29%, Daejeon, Korea, pp. 431–441, 2025, doi: 10.1007/978-3-032-04937-7_41.

Junseok Lee, Yeonho Yoo, Jinkyu Kim, Dosun Lim, Gyeongsik Yang, Chuck Yoo

Download PDF

Abstract

With the rise of wearable IoT devices such as smartwatches and smart rings, ECG signals have become more accessible and made cardiovascular monitoring a reality. However, analyzing the ECG signals for complex conditions, such as bundle branch blocks and myocardial infarction, requires multi-lead ECG data. Although various deep learning models for ECG reconstruction have been proposed, they are computationally expensive and unsuitable on resource-constrained wearable IoT devices. To address this challenge, we propose mEcgNet, a parameter-efficient model for reconstructing 12-lead ECG signals from a single lead. mEcgNet introduces a modular deep learning architecture for parameter efficiency and separates the single lead-I signal into multiple frequency segments to improve accuracy. Our experiments demonstrate that mEcgNet significantly reduces the number of parameters and inference time by ∼23.1× and ∼5.4×, respectively, compared to existing state-of-the-art models. Furthermore, it reduces the reconstruction error by ∼22.1%, demonstrating its high accuracy and efficiency.

Paper summary

Reconstruct more leads with a smaller model

Separating an ECG into frequency bands helps a compact model reconstruct twelve leads from one measured lead.

A single ECG lead is easier to collect, but it does not directly provide the twelve views used in a standard ECG. mEcgNet reconstructs those views while reducing the model size and computation required by earlier reconstruction networks.

Give each frequency band an appropriate module

The input lead is separated into low-, medium- and high-frequency signals using Fourier transforms. Three sequential reconstruction modules process those signals, passing earlier outputs to later modules. Smaller modules handle the first two bands; a larger module handles the high-frequency component.

The model assigns capacity according to each signal component.

Frequency-separated signals pass through sequential reconstruction modules to produce twelve ECG leads.View full-size figure
Frequency-separated signals pass through sequential reconstruction modules to produce twelve ECG leads.

Compare reconstruction error and inference cost

Experiments use PTB-XL and Chapman with five-fold cross-validation. On PTB-XL, mean squared reconstruction error is 0.113 for mEcgNet, compared with 0.120 for EKGAN and 0.145 for ECGrecover in the authors’ reproduced evaluation.

The reconstruction model has 1.13 million parameters. CPU-only inference on a Jetson Orin Nano takes 17.99 milliseconds per ECG, compared with 1.4× and 5.4× longer times for the two baselines respectively.

Mean squared reconstruction error on PTB-XL; lower is better.
ModelMSE
mEcgNet0.113
EKGAN0.120
ECGrecover0.145

Key takeaways

  1. Different ECG frequency components need different model capacity. Frequency separation gives reconstruction a way to reduce parameters without treating every part of the signal identically.
  2. Reconstruction is cumulative. Passing earlier outputs into later stages connects the frequency-specific estimates into a complete signal.
  3. Compactness matters alongside fidelity. Reconstruction error, computation and inference time jointly describe the practical value of the smaller model.