Supervised Machine Learning Applied to Automate Flash and Prolonged Capillary Refill Detection by Pulse Oximetry

Area of Science:

• Biomedical Engineering
• Clinical Informatics
• Artificial Intelligence in Medicine

Background:

• Capillary refill time (CRT) is a critical indicator of peripheral circulation.
• Manual CRT assessment can be subjective and time-consuming.
• Automated methods are needed for objective and rapid CRT evaluation.

Purpose of the Study:

• To develop and validate supervised machine learning (ML) models for automated detection of flash (<1.0 s) or prolonged (≥2.0 s) capillary refill time (CRT).
• To correlate ML-based CRT detection with clinician judgment using pulse oximeter plethysmography data.

Main Methods:

• Collected paired CRT measurements from 99 children and 30 adults using a modified pulse oximeter and clinician assessment.
• Developed supervised ML models (XGBoost, logistic regression, SVM) on 485 waveform pairs.
• Compared model performance using Area Under the Receiver Operating Curve (AUC) and external validation.

Main Results:

• XGBoost demonstrated superior performance for both flash (AUC 0.79) and prolonged CRT (AUC 0.77) detection compared to logistic regression and SVM.
• The best performing XGBoost model achieved an AUC of 0.88 in external validation.
• Feature importance analysis identified distinct waveform characteristics predictive of flash and prolonged CRT.

Conclusions:

• Supervised ML models effectively identify abnormal CRT from pulse oximeter waveforms.
• This novel application provides an automated and objective method for CRT assessment.
• The developed models correlate well with clinician judgment, offering a valuable clinical tool.