VitalSync ECG

VitalSync ECG is a research model for classifying annotated single-lead ECG beats into four groups: N, S, V and F. It combines morphology from the current beat and two preceding beats with causal RR-interval history.

The selected artifact is intended for reproducible machine-learning research. It is not a clinical device, diagnostic model, live monitor or automatic alert system.

Each example is a float32 tensor with shape [batch, 1, 608]. The first 600 values contain three 200-sample beat crops: the two preceding annotated beats and the current annotated beat. The final eight values contain causal RR and rhythm-history features.

• The model expects annotated beat centers; automatic R-peak detection is outside the evaluated pipeline.
• All temporal context precedes or includes the current beat, avoiding future-beat leakage at inference time.
• Outputs are class scores for N, S, V and F groups, not calibrated disease probabilities.
• The N group includes normal, conduction and escape beats; it must not be interpreted as a healthy label.

A shared one-dimensional residual encoder processes each beat crop. The three morphology embeddings are fused with the eight RR-history features before the four-class prediction head. The selected configuration is recorded as context-v1.

Input [B, 1, 608]
  ├─ previous-2 beat [200] ─┐
  ├─ previous-1 beat [200] ─┼─ shared 1D ResNet encoder
  ├─ current beat    [200] ─┘
  └─ causal RR features [8]
              ↓
         feature fusion
              ↓
        N · S · V · F scores

The selected MIT-BIH + INCART run reached its best validation result at epoch 10. These are development metrics: validation was used repeatedly for selection, and the selected context-aware model was not scored on an independent test set.

Development uses annotated records from the MIT-BIH Arrhythmia Database and the St Petersburg INCART 12-lead Arrhythmia Database, transformed into the documented single-lead beat context. Dataset access, attribution and license terms remain governed by PhysioNet and the respective dataset records.

• Not validated for medical alerts, diagnosis, triage, live monitoring or clinical decision-making.
• Requires annotated beat centers and has not validated automatic beat detection or streaming behavior.
• The selected model has no independent test score; reported validation results are not deployment estimates.
• Minority-class performance remains weak, including no correctly detected F examples in the selected validation run.
• Sensor-domain transfer, including AD8232 hardware compatibility, remains unvalidated.