VitalSync ECG

This study evaluates whether causal beat and RR history can support four-class annotated ECG beat classification. The model encodes the current beat and two preceding beats with a shared one-dimensional residual network, then fuses those embeddings with eight rhythm features before predicting N, S, V or F.

The selected MIT-BIH + INCART run reached validation macro-F1 0.445308 and accuracy 0.734426. These are model-development results, not independent clinical or deployment estimates.

A single ECG beat can be morphologically ambiguous. VitalSync tests a strictly causal representation that adds recent beat morphology and RR timing without looking ahead to future beats.

• Three 200-sample beat crops represent previous-2, previous-1 and current beats.
• Eight causal RR and rhythm features complement morphology.
• A shared encoder keeps feature extraction consistent across the three beat positions.
• Patient-separated development splits reduce direct patient overlap between training and validation.

Tensor: float32 [batch, 1, 608]

0:200    previous-2 annotated beat crop
200:400  previous-1 annotated beat crop
400:600  current annotated beat crop
600:608  eight causal RR-history features

Output: four class scores [N, S, V, F]

This contract makes the evaluated boundary explicit. The work does not include an automatic R-peak detector, raw continuous-stream ingestion or calibrated clinical probabilities.

The narrow macro-F1 spread and persistent F-class failures matter: adding context and INCART data did not remove the core class-imbalance and confusion problems.

The experiments support a reproducible research baseline for studying causal context in annotated beat classification. They do not support claims of reliable arrhythmia monitoring, clinical diagnosis or real-world alert performance.

• Validation was repeatedly consulted during model and epoch selection.
• The selected context-aware model has test_evaluated=false and no independent test score.
• N is a heterogeneous beat group, not a synonym for healthy.
• Output scores are not calibrated disease probabilities.
• External sensor and population transfer have not been established.

The MIT-BIH Arrhythmia Database is described by Moody and Mark, IEEE Engineering in Medicine and Biology Magazine 20(3), 45-50 (2001), dataset DOI 10.13026/C2F305. PhysioNet is described by Goldberger et al., Circulation 101(23), e215-e220 (2000).

The St Petersburg INCART 12-lead Arrhythmia Database is available under DOI 10.13026/C2V88N and the Open Data Commons Attribution License v1.0. Reusers must follow the source datasets' attribution and license requirements.

The public repository contains the training and evaluation code, split manifests, selected-run artifacts, validation reports and the approximately 5.3 MB TorchScript export needed to inspect and reproduce the documented model contract.