okekeclean-ecg-v1

This repository contains the released ECG checkpoint for OkekeClean, an artifact detection pipeline for clinical ECG waveforms.

Model Description

The released ECG model is okekeclean-ecg-efficientnet_b0, an EfficientNet-B0 model with full fine-tuning on spectrograms derived from 10-second ECG windows.

Intended Use

This weight is intended for inference on ECG waveforms represented as a pandas.Series with a DatetimeIndex. The public okekeclean package segments the waveform into 10-second windows, resamples them to 500 Hz when needed, converts each window to a normalized 3-channel spectrogram image, and returns artifact probabilities or binary labels.

Inference code: https://github.com/moberg-analytics/oss-models

Files

• cnn-efficientnet_b0-full-best-v65-5ef1694d9ae9fab6c6e0023903f63997.pth

Architecture Details

Public model	Backbone	Input	Threshold
okekeclean-ecg-efficientnet_b0	EfficientNet-B0 (full FT)	10-second ECG spectrogram, 224x224, 3-channel	0.20802117884159088

The classifier head is replaced with Linear -> ReLU -> Dropout -> Linear. ECG uses the default 3-channel EfficientNet stem after spectrogram generation.

Performance

Held-out ECG test-set performance from the associated paper:

Model	Accuracy	Sensitivity	Specificity	AU-ROC
EfficientNet-B0 (Full FT)	0.922	0.858	0.950	0.970

The ECG paper reported a slightly stronger two-model ensemble (AU-ROC 0.972), but this release publishes the best practical single-model checkpoint for straightforward deployment.

Training Data

The ECG model was trained on data from the PRECICECAP study:

• 65 post-cardiac-arrest patients across seven enrollment sites
• 6,963 hours of continuous ECG monitoring
• 10,890 expert-reviewed 10-second segments
• 30.1% artifact prevalence

Single-lead ECG (lead II) was sampled at 500 Hz and transformed into spectrograms for transfer-learning-based classification.

Limitations and Biases

• The model is trained on single-lead ICU ECG from one study cohort and may not generalize to other leads, ambulatory devices, or substantially different noise patterns without recalibration.
• The deployed model optimizes a fixed operating threshold; downstream users may need to retune thresholds for different sensitivity or specificity targets.
• Errors were concentrated in ambiguous, moderate-artifact segments rather than clear quality extremes.

Citation

• Tony K. Okeke, Ethan Moyer, Karen G. Hirsch, Teresa L. May, Zihuai He, Jonathan Tam, Laura Faiver, Richard Moberg, Jonathan Elmer. Spectrogram-Based Transfer Learning for Electrocardiogram (ECG) Artifact Detection in the ICU. Selected as a Research Snapshot presentation at the Society of Critical Care Medicine (SCCM) Critical Care Congress, March 2026.
• Moberg Analytics OSS Models Repository

@inproceedings{okeke2026spectrogram,
  title        = {Spectrogram-Based Transfer Learning for Electrocardiogram ({ECG}) Artifact Detection in the {ICU}},
  author       = {Okeke, Tony K. and Moyer, Ethan and Hirsch, Karen G. and May, Teresa L. and He, Zihuai and Tam, Jonathan and Faiver, Laura and Moberg, Richard and Elmer, Jonathan},
  booktitle    = {Society of Critical Care Medicine (SCCM) Critical Care Congress},
  year         = {2026},
  month        = mar,
  note         = {Research Snapshot presentation}
}