LeXT-Net: A lead-adaptive hierarchical network for 3-lead to 12-lead ECG reconstruction

Standard 12-lead electrocardiography provides spatial cardiac information that is not fully available from the small number of channels typically used in ambulatory and wearable monitoring, whereas acquisition of all 12 leads requires multiple electrodes and controlled anatomical placement. Estimating unobserved leads from reduced-lead recordings may therefore extend the information available from compact ECG systems. In this study, we propose LeXT-Net, a hierarchical lead-adaptive architecture for reconstructing 12-lead ECG signals from three-lead inputs. The model combines a shared temporal convolutional network backbone, lead-specific grouped temporal attention, independent U-Net decoding heads with different capacities for limb and precordial leads, and a composite objective incorporating waveform fidelity, Einthoven-Goldberger consistency, and precordial spatial regularization. Four predefined input configurations (EGV1, EGV2, EGV3, and QSI) were evaluated using all-lead (mAll), non-input (mOther), and independent-target (mInd) scopes. On PTB-XL, LeXT-Net achieved R 2 = 0.910, PCC = 0.952, SSIM = 0.840, RMSE = 0.041, and SNR = 27.48 dB under EGV3-mAll; under the stricter EGV3-mInd evaluation, R 2 = 0.786, PCC = 0.885, SSIM = 0.622, and SNR = 9.918 dB. Additional five-fold evaluations on the independent Chapman-Shaoxing and CPSC2018 databases yielded EGV3-mAll R 2 /PCC values of 0.887/0.939 and 0.912/0.953, respectively. These results indicate that LeXT-Net can reconstruct substantial signal-level information from three-lead inputs across multiple ECG databases, including targets that are neither directly observed nor algebraically derivable. The findings support further investigation of reduced-lead ECG reconstruction, while clinical equivalence to directly acquired 12-lead ECG and robustness to native wearable-device recordings remain to be established.