LeXT-Net: Source Code and Pretrained Model Weights for 3-Lead to 12-Lead ECG Reconstruction
Year: 2026
Type: software
Abstract: LeXT-Net is a lead-adaptive hierarchical deep learning framework for reconstructing standard 12-lead electrocardiograms (ECGs) from three-lead ECG inputs. The architecture combines a shared temporal convolutional network (TCN) backbone, lead-specific grouped temporal attention, lead-adaptive decoder branches, and physiologically informed loss terms derived from inter-lead relationships. This record provides the source code, training and evaluation utilities, reproducibility documentation, and pretrained LeXT-Net model weights associated with the manuscript “LeXT-Net: A Lead-Adaptive Hierarchical Network for 3-Lead to 12-Lead ECG Reconstruction.” Four three-lead input configuration are supported: EGV1 (I, II, V1), EGV2 (I, II, V2), EGV3 (I, II, V3), and QSI. Pretrained LeXT-Net v2 checkpoints are provided for three publicly available 12-lead ECG datasets: PTB-XL, Chapman-Shaoxing, and CPSC2018. For each dataset, four input strategies and five cross-validation folds are included, yielding 60 pretrained checkpoints in total. A SHA-256 manifest is supplied for checkpoint provenance and integrity verification. The repository supports both full five-fold training and selected single-fold training, evaluation of pretrained checkpoints without retraining, lead-wise and aggregate performance reporting, reference-versus-reconstruction visualization, single-record inference, and adaptation to compatible external 12-lead ECG datasets through a documented manifest-based interface. Compatible external 12-lead ECG datasets are resampled and standardized to the study format of 100 Hz and 10-s recordings before generation of the corresponding three-lead model inputs. The original PTB-XL, Chapman-Shaoxing, and CPSC2018 datasets are not redistributed in this Zenodo record. Users should obtain the source datasets from their original providers and comply with the corresponding dataset licenses and terms of use. Small example datasets, when distributed through the GitHub repository, are subject to the redistribution permissions of their respective source datasets. Training uses per-record, per-lead z-score normalization. Accordingly, physical-scale error evaluation of reconstructed target leads is performed retrospectively when the reference 12-lead ECG is available; single-record three-lead-only inference produces reconstructed signals in the normalized model output space.
Institutions: Batman University, Şırnak University
Language: English
Topic: ECG Monitoring and Analysis
Subfield: Cardiology and Cardiovascular Medicine
Field: Medicine
Domain: Health Sciences
Open Access status: green
OpenAlex ID: W7224488196