Work

SAMVT-Net: Source-Aware Multi-View Token Learning for Heterogeneous ECG Classification

Year: 2026
Type: preprint
Abstract: ECG classification requires representations that accommodate differences in signal duration, lead configuration, and label structure. We present SAMVT-Net, a compact supervised model built around physiological tokens and explicit source identity. The representation organizes the physiological information available to each task: heartbeat inputs combine waveform morphology with measured inter-beat timing and auxiliary spectral summaries, while record-level inputs use morphology and spectrum. Learnable lead identifiers inform attention-based fusion of these summaries. Task-specific readouts use cosine-similarity prototypes for single-label prediction and sigmoid outputs for co-occurring diagnoses, while retaining the same tokenization and fusion interface. Evaluation covers heartbeat classification on the MIT-BIH Arrhythmia Database and record-level diagnosis on PTB-XL, including its official patient-separated folds. Controlled comparisons associate the main multi-view gain with morphology–rhythm complementarity and show the clearest effect of lead identity in Transformer fusion. A separate decision-head study identifies the contribution of prototype classification within the chosen imbalance-aware training configuration. An auxiliary-token control supports omitting the zero-input token from the record-level fusion sequence. Matched-split reimplementations, representation geometry, graded input perturbations, and inference profiling further characterize predictive performance and execution cost. The results support source-aware physiological token learning as a compact approach to the evaluated ECG tasks, with component effects that depend on the input, backbone, and prediction objective.
Language: English
Field: Medicine
Open Access status: green
OpenAlex ID: W7225917027
Cites: 36
Cited by:
Related to: 10