Logo image
Back
Evaluating artifact detection algorithms for the arterial blood pressure waveform acquired from the intensive care unit: A PRECICECAP informatics approach
Journal article   Peer reviewed

Evaluating artifact detection algorithms for the arterial blood pressure waveform acquired from the intensive care unit: A PRECICECAP informatics approach

Tony K. Okeke, Manil Shrestha, Ethan Moyer, Karen G. Hirsch, Teresa L. May, Zihuai He, Jonathan Tam, Laura Faiver, Richard Moberg and Jonathan Elmer
Biomedical signal processing and control, v 124, 110434
15 Sep 2026
DOI: https://doi.org/10.1016/j.bspc.2026.110434
Featured in Collection :   Drexel's Newest Publications

Abstract

Arterial blood pressure Artifact detection Deep learning Transfer learning Machine Learning
This study aims to develop and evaluate automated methods for detecting artifacts in continuous arterial blood pressure waveforms from intensive care unit monitoring to improve data quality for clinical interpretation and machine learning applications. We analyzed data from a prospective, multicenter cohort of 63 intensive care unit patients resuscitated from cardiac arrest, comprising 5707 hours of continuous monitoring and 10,005 individually annotated pulses. Using a two-stage annotation process, research assistants first identified artifact regions, then clinical experts reviewed individual pulses using a 4-point ordinal scale. We implemented and compared six model families: rule-based heuristics, handcrafted feature engineering with traditional machine learning, self-supervised reconstruction, image-based deep learning, and transfer learning with pre-trained image classifiers. We also evaluated ensemble methods and patient-specific fine-tuning. On an independent test set of 1493 pulses from 10 patients, deep learning approaches outperformed traditional methods. EfficientNet-B0 with shallow fine-tuning achieved 89.8% accuracy (79.9% sensitivity, 93.9% specificity), while ensemble models achieved 95.2% sensitivity at 73.0% specificity with AU-ROC of 0.958. Patient-specific fine-tuning with 20 labeled examples improved median specificity. In a downstream forecasting task, artifact filtering reduced prediction error by 21% (MAE) and 59% (RMSE). This systematic comparison demonstrates that transfer learning and ensemble methods achieve high performance for automated artifact detection in arterial blood pressure waveforms. The downstream experiment confirms that artifact filtering materially improves subsequent model performance. This work establishes validated methods for automated quality control of ICU waveform data, a critical prerequisite for reliable AI/ML model development.

Metrics

1 Record Views

Details

Logo image