<p>Standard 12-lead electrocardiography (ECG) remains the clinical gold standard for diagnosing cardiac disorders, particularly myocardial infarction (MI). However, electrode-count constraints in portable and wearable ECG systems make simultaneous acquisition of all 12 leads impractical, thereby challenging the preservation of diagnostic fidelity. This study aims to reconstruct physiologically coherent and clinically meaningful 12-lead ECGs from a limited 3-lead input configuration. To this end, we developed a WaveUNet-based encoder–decoder architecture and systematically extended it with alternative anti-aliasing and resampling strategies, including BlurPool, AAPool, AAStride, NearestConv, and PixelShuffle. The models were trained and evaluated on the PTB-XL dataset using ten different three-lead input combinations. Reconstruction quality was assessed using signal-level metrics, including RMSE, MAE, R², and PRD. In addition, the diagnostic utility of the reconstructed 12-lead signals was quantitatively examined through MI versus non-MI classification. The experimental results demonstrate that the baseline WaveUNet provides strong and stable reconstruction performance, while anti-aliasing-based variants yield additional improvements, particularly for input combinations containing precordial leads. The highest signal-level accuracy was achieved with the I–II–V2 and I–II–V3 combinations, reaching R² values of approximately 0.86–0.87. Lead-wise analyses further revealed that inclusion of at least one chest lead is critical for accurate reconstruction of precordial outputs. In diagnostic evaluation, the AAStride variant delivered the most balanced MI classification performance, achieving 84.3% accuracy, an F1-score of 0.720, and an ROC-AUC of 0.908. Overall, the findings indicate that anti-aliased WaveUNet-based 3-to-12 lead ECG reconstruction can provide clinically meaningful morphological and diagnostic consistency for low-electrode wearable and portable ECG systems.</p> Graphical Abstract <p></p>