okekeclean-abp-v1

This repository contains the released ABP checkpoints for OkekeClean, an artifact detection pipeline for arterial blood pressure waveforms.

Model Description

The primary released ABP model is okekeclean-abp-ensemble, a weighted ensemble of:

• ResNet-18 with full fine-tuning on 2D grayscale pulse images
• EfficientNet-B0 with shallow fine-tuning on the same pulse-image representation

The checkpoint repo also includes the two individual component weights so the public okekeclean package can expose them as standalone models.

Intended Use

These weights are intended for inference on continuous ABP waveforms sampled at 125 Hz. The public inference package segments the waveform into pulse-to-pulse intervals, converts each pulse into a 224x224 single-channel image, and returns artifact probabilities or binary labels.

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

Files

• cnn-resnet18-full-best-v2-6895bffcf59e0153fbae93fe7f3e0b93.pth
• cnn-efficientnet_b0-head_only-best-v3-ab40c448fd91102b747244f445e05137.pth

Architecture Details

Public model	Components	Threshold
okekeclean-abp-ensemble	ResNet-18 (full FT, weight 0.4) + EfficientNet-B0 (shallow FT, weight 0.6)	0.184
okekeclean-abp-resnet18	ResNet-18 (full FT)	0.17785164713859558
okekeclean-abp-efficientnet_b0	EfficientNet-B0 (shallow FT)	0.04994076117873192

Both CNN backbones replace the classifier head with Linear -> ReLU -> Dropout -> Linear, and the first convolution is adapted to single-channel pulse images.

Performance

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

Model	Accuracy	Sensitivity	Specificity	AU-ROC
Ensemble 1	0.795	0.952	0.730	0.958
ResNet-18 (Full FT)	0.851	0.915	0.824	0.951
EfficientNet-B0 (Shallow FT)	0.898	0.799	0.939	0.945

Training Data

The ABP models were trained on waveforms from the PRECICECAP study:

• 63 ICU patients across seven sites
• 5,707 total ABP monitoring hours
• 10,005 final pulse-level examples after segmentation review
• Approximately 30% artifact prevalence

Each pulse was annotated on a 4-point ordinal scale by expert physician reviewers and then binarized into artifact vs non-artifact classes.

Limitations and Biases

• Training data comes from a single clinical study population and may not represent all monitor vendors, line setups, or patient populations.
• The released operating threshold prioritizes high sensitivity, which lowers specificity relative to some standalone models.
• Borderline pulses near the distinction between slight artifact and substantial artifact remain the primary source of error.

Citation

• Tony K. Okeke, Manil Shrestha, Ethan Moyer, Karen G. Hirsch, Teresa L May, Zihuai He, Richard Moberg, Jonathan Elmer. Transfer Learning for Artifact Detection in ICU-Collected Arterial Blood Pressure Waveforms. Abstract presented at Neurocritical Care Society (NCS) Annual Meeting, Montreal, QC, Canada. September 2025.
• Moberg Analytics OSS Models Repository

@inproceedings{okeke2025transfer,
  title        = {Transfer Learning for Artifact Detection in {ICU}-Collected Arterial Blood Pressure Waveforms},
  author       = {Okeke, Tony K. and Shrestha, Manil and Moyer, Ethan and Hirsch, Karen G. and May, Teresa L. and He, Zihuai and Moberg, Richard and Elmer, Jonathan},
  booktitle    = {Neurocritical Care Society (NCS) Annual Meeting},
  year         = {2025},
  month        = sep,
  address      = {Montreal, QC, Canada},
  note         = {Abstract presentation}
}