Evaluating AI Methods for Pulse Oximetry: Performance, Clinical Accuracy, and Comprehensive Bias Analysis

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Signal Processing

Background:

• Blood oxygen saturation (SpO2) monitoring is critical in patient care.
• Traditional SpO2 estimation methods face limitations.
• Artificial intelligence (AI) offers potential for improved SpO2 analysis via photoplethysmography (PPG) signals.

Purpose of the Study:

• To systematically review AI techniques for SpO2 estimation from PPG signals.
• To identify leading AI models, features, datasets, and performance metrics.
• To assess bias and identify challenges in current AI-driven SpO2 monitoring.

Main Methods:

• Systematic literature review following PRISMA guidelines.
• Analysis of 183 references, with 26 studies meeting inclusion criteria.
• Evaluation of AI models (e.g., DNNs, Gaussian Processes), features, datasets, and bias using QUADAS-2 criteria.

Main Results:

• Linear regression and deep neural networks (DNNs) are common AI methods.
• Gaussian Process models demonstrated high accuracy (MAE as low as 0.57%, RMSE as low as 0.69%).
• Bias analysis revealed needs for better patient selection, reference standards, and SpO2 intervals.

Conclusions:

• AI, particularly Gaussian Process models, shows promise for accurate SpO2 estimation.
• Improvements in reference standards, datasets, and ethical considerations are needed for clinical translation.
• Integrating AI with physiological models can enhance SpO2 monitoring and patient care.