Adaptive ultra-lightweight 12-lead ECG reconstruction from limited leads via a vectorcardiogram-mediated cascade framework

Background The 12-lead electrocardiogram (ECG) is the clinical gold standard for cardiovascular diagnosis, but its application in wearable long-term monitoring is constrained by limited available leads. Existing reconstruction methods are limited by poor robustness of fixed lead schemes, high model complexity, and insufficient physiological interpretability. Methodology This work proposes a two-stage LR-MLP cascade framework for adaptive 12-lead ECG reconstruction. Vectorcardiogram (VCG) acts as the physiologically interpretable intermediate representation, and VCG refers to a three-dimensional orthogonal representation of cardiac electrical activity. A label-free adaptive lead selection mechanism based on the orthogonality deviation index (ODI) is designed. The MLP contains 207 trainable parameters, while the final ECG reconstruction path uses 219 active coefficients, comprising the MLP parameters and the 12 coefficients of the ODI-selected ridge-regression block. Results The model achieves average Pearson correlation coefficients (PCC) of 0.9145 and 0.8853 on healthy and myocardial infarction (MI) cohorts of the PTB-DB dataset, respectively. After lightweight fine-tuning on the PTB-XL dataset, the average PCC reaches 0.9218 and 0.9103 for healthy and MI subjects, respectively. Core electrocardiographic features and inter-individual variability are well preserved, and the mean regression tendency is alleviated. Conclusions The framework balances reconstruction accuracy, interpretability and ultra-lightweight property, providing a technical reference for wearable ECG monitoring. Further multi-center clinical validation is required for practical deployment.