← Back to All Projects
Research Study & Podium Presentation Clinical AI & Signals

Vitalis

Stratified Deep Learning for Rare Arrhythmia Detection on ECG

Convolutional deep learning architecture overcoming the long-tail problem in ECG classification to reliably catch rare supraventricular and fusion beats.

ECGArrhythmiaDeep LearningPumps & PipesSignal Processing
Launch Live Application ↗
Actual Application Interface · High-Resolution Capture vitalis.studio
Interface of Vitalis

The Clinical & Engineering Problem

Standard arrhythmia classifiers achieve deceptive ~98% accuracy by overfitting to dominant normal sinus rhythms while missing rare, life-threatening ventricular and supraventricular ectopy.

The Technical Approach

Applies stratified class resampling paired with non-linear time-warping and noise augmentation, enforcing feature separation across under-represented arrhythmia categories.

Engineering Architecture

  • 1D convolutional deep neural network trained on combined MIT-BIH and PTB diagnostic ECG datasets
  • Stratified balancing resolving extreme class imbalance between normal and ectopic morphology
  • Non-linear time-warp data augmentation preserving pathological intervals (PR, QRS, QT)
  • Interactive multi-lead rhythm explorer demonstrating classification confidence across beats

Documented Evidence & Verification

97.53%
Test Accuracy
Validated on held-out test split from MIT-BIH Arrhythmia benchmark
+67%
Macro-F1 Gain
F1-macro 0.9017 vs 0.5400 baseline on underrepresented rare classes
Top 4
Pumps & Pipes
Podium presentation selected among top 4 research innovations
Scientific Status & Regulatory Disclaimers

Trained on public retrospective benchmark databases (MIT-BIH, PTB). Requires prospective clinical trial validation on continuous multi-lead telemetry prior to medical deployment.

Open Working Workspace ↗ Inquire About This Project ← Back to All Projects