Spectrogram-Based Transfer Learning for Electrocardiogram (ECG) Artifact Detection in the ICU

Tuesday, March 24, 2026

11:15 AM - 12:15 PM Central Time

Location: Connections Central - RST 16

Tony K. Okeke, MS

Pennsylvania

Disclosure(s): Moberg Analytics: Employee (any industry) (Ongoing)
EM

Ethan Moyer, n/a

VP R&D
Moberg Analytics, Inc

Disclosure information not submitted.
KH

Karen Hirsch, MD

Professor of Neurology and Neurological Sciences
Department of Neurology, Stanford University

Disclosure information not submitted.
TM

Teresa May, DO

Director, NeuroCritical Care
Department of Critical Care Services, Neuroscience Institute, Maine Medical Center

Disclosure information not submitted.
ZH

Zihuai He, PhD

Assistant Professor
Quantitative Sciences Unit, Department of Neurology, Stanford University

Disclosure information not submitted.
JT

Jonathan Tam, MD

Assistant Professor of Critical Care Medicine and Emergency Medicine
Departments of Emergency Medicine, Critical Care Medicine and Neurology, University of Pittsburgh

Disclosure information not submitted.
LF

Laura Faiver, MD

Post-Doctoral Scholar
Departments of Emergency Medicine, Critical Care Medicine and Neurology, University of Pittsburgh

Disclosure information not submitted.
RM

Richard Moberg, n/a

CEO
Moberg Analytics, Inc

Disclosure information not submitted.
JE

Jonathan Elmer, MD, MS

Associate Professor of Emergency Medicine, Critical Care Medicine, and Neurology
Departments of Emergency Medicine, Critical Care Medicine and Neurology, University of Pittsburgh

Disclosure information not submitted.

Introduction: Continuous intensive care unit (ICU) ECG is frequently degraded by artifact, impairing clinical use and downstream statistical analysis. Prior quality control commonly relies on statistical signal quality indices (SSQIs), which can be inconsistent across datasets. We hypothesized that transfer learning on ECG spectrograms would outperform SSQI-based baselines for automated artifact detection at clinically relevant high-sensitivity operating points.

Methods: We analyzed 10,890 ten-second ECG segments from 65 post–cardiac arrest patients in the PRECICECAP study. Two physicians graded segment quality on a four-point scale, and labels were dichotomized to artifact versus non-artifact. Baselines were Random Forest and XGBoost trained on SSQI features. Deep learning models (ImageNet-initialized ResNet-18 and EfficientNet-B0) were fine-tuned on ECG spectrograms. We also evaluated simple ensembling and patient-specific fine-tuning (PS-FT) using 20 labeled segments per patient. Our primary objective was to maintain high sensitivity on a held-out test cohort. We first report test set performance (AU-ROC, accuracy, sensitivity) for the best performing model, then fix sensitivity at 95% using thresholds selected on the validation set and report test set specificity and accuracy at that operating point.

Results: Artifact prevalence was 30.1% across 6,963 monitoring hours. Deep learning substantially outperformed SSQI baselines. The fine-tuned EfficientNet-B0 achieved AU-ROC 0.970, accuracy 92.2%, sensitivity 85.8% (best SSQI: AU-ROC 0.848, accuracy 73.9%, sensitivity 77.4%). At the 95% sensitivity operating point, the model attained 72.2% specificity and 80.1% accuracy. PS-FT increased median specificity (from 91.9% to 96.0%) and accuracy (from 90.0% to 92.0%) with a decrease in sensitivity (from 77.8% to 76.1%). Classification errors clustered at the subjective boundary between ‘minimal artifact’ and ‘substantial artifact’, rather than at quality extremes.

Conclusions: A spectrogram-based transfer learning pipeline provides superior ECG artifact detection versus SSQI baselines and enables high-sensitivity curation of continuous ICU data. In parallel work using the same approach, we observed similarly strong discrimination for arterial blood pressure pulses, supporting cross-modality applicability.