Tensor-based methodology for constructing short-term high-dynamic microvolt-level T-wave alternans simulated ECG database

Microvolt-level T-wave alternans (TWA) is crucial for predicting arrhythmias and sudden cardiac death, but detection faces challenges from weak signals, noise interference, and insufficient training data. Existing databases have limited volume, poor realism, and cannot simulate short-term high-dynamic characteristics, restricting deep learning algorithm development. This paper proposes a tensor-based method for constructing short-term high-dynamic microvolt-level TWA synthetic ECG databases. Using third-order tensors (time × beat × channel), the method extracts real TWA signals from PTB database X, Y, Z leads and maintains spatial correlations via principal component analysis. Dynamic parameterization generates variant TWA components with 0-100μV amplitude fluctuations, simulating short-term high-dynamic characteristics. Multiple noise types create complex environments (-30dB to 5dB SNR), generating 38,400 TWA tensors. Experimental validation shows the database significantly outperforms existing datasets in simulating real TWA signal characteristics and spatial correlations, providing superior training resources for algorithm optimization and promoting clinical translation of deep learning-based TWA detection technology.