AutoLeadX: A Unified Framework for Arbitrary Reduced-Lead to Standard 12-Lead ECG Reconstruction

The 12-lead electrocardiogram (ECG) remains the clinical gold standard for diagnosing cardiac abnormalities, yet mobile and wearable devices often employ reduced-lead configurations with fewer electrodes and variable placements. Existing reconstruction algorithms are tailored to specific setups, limiting their robustness and generalizability. We propose AutoLeadX, the first unified framework that reconstructs 12-lead ECGs from arbitrary subsets of leads, approximated in practice by randomly sampling subsets from the standard 12-lead ECGs. AutoLeadX introduces three key innovations: (1) a lead aggregator that adapts to varying input sizes, (2) a localization-aware loss that improves spatial discrimination by guiding the model to infer lead orientations directly from waveform characteristics, and (3) a shared decoder with lead tokens for efficient and flexible reconstruction. Experiments on three widely adopted ECG datasets demonstrate that AutoLeadX achieves state-of-the-art reconstruction performance across single and multi-lead settings while retaining diagnostic utility comparable to original 12-lead recordings. These results establish AutoLeadX as a practical foundation for reliable ECG acquisition in mobile, wearable, and home-monitoring applications.