Pulse Oximeters, Race, and ICU AI — How Algorithmic Bias Costs Lives in Critical Care

The 2020 study by Sjoding et al. in the New England Journal of Medicine quantified something that critical care nurses had long suspected: pulse oximeters — the small devices clipped to a patient’s finger to measure blood oxygen saturation — overestimate oxygen levels in patients with darker skin tones at a rate that is clinically significant. The paper documented a rate of occult hypoxemia (true low oxygen not detected by the device) nearly three times higher in Black patients than in white patients. In a hospital context where oxygen saturation guides titration of ventilator settings and supplemental oxygen delivery, this is not an academic finding.

How Pulse Oximeters Work and Where They Fail

Standard pulse oximeters measure the ratio of red to infrared light absorption across the fingertip. Oxygenated hemoglobin absorbs more infrared light; deoxygenated hemoglobin absorbs more red light. The ratio of these absorptions is calibrated against a reference curve established in clinical trials conducted predominantly in light-skinned subjects — a calibration gap that was recognized in the biomedical engineering literature decades before the Sjoding paper but not addressed in device standards or clinical guidelines.

Melanin, the pigment responsible for darker skin tones, absorbs light across both the red and infrared wavelengths used by pulse oximeters. This interference systematically biases readings upward — causing the device to report adequate oxygen saturation when the patient’s arterial blood gas measurement tells a different story. The Sjoding study used paired arterial blood gas and oximetry measurements from ICU patients and found occult hypoxemia (arterial oxygen saturation below 88% despite oximetry reading 92-96%) in 11.7% of Black patients versus 3.6% of white patients.

Why AI Compounds the Problem

This is not merely a pulse oximetry problem. It is a template for understanding how measurement bias propagates through AI systems trained on clinical data. AI algorithms used in the ICU — for sepsis prediction, ventilator management, early warning scoring, and clinical deterioration detection — are trained on electronic health record data that includes pulse oximetry readings as input features. If the oximetry values fed into these systems are systematically elevated for Black patients, the algorithm’s model of what “adequate” oxygenation looks like is calibrated incorrectly for that population.

A sepsis prediction model trained on data where Black patients’ true hypoxemia is masked by biased oximetry readings will learn a different decision boundary for that group — one that may fail to generate alerts at the appropriate threshold. The bias in the upstream measurement does not disappear when the data is used for model training; it is encoded into the model’s parameters and reproduced at prediction time.

A 2022 analysis in JAMA Internal Medicine examined this propagation directly, finding that an algorithm trained on oximetry-influenced clinical data assigned lower risk scores to Black patients who subsequently experienced adverse respiratory events — a measurable downstream consequence of upstream measurement bias.

• Occult hypoxemia rate: 11.7% in Black ICU patients versus 3.6% in white patients
• Bias originates from calibration studies conducted in predominantly light-skinned subjects
• AI models trained on biased oximetry data inherit and encode the measurement error
• 2022 JAMA Internal Medicine analysis confirmed lower AI risk scores for Black patients with adverse outcomes

Structural Remedies and Their Limits

The technical remedy for pulse oximetry bias is well defined: measure more wavelengths of light across a more representative sample during calibration, and recalibrate existing algorithms. Masimo’s rainbow SET technology uses eight wavelengths rather than two and has demonstrated improved performance across skin tones. The problem is that standard two-wavelength devices remain dominant in clinical settings globally, and there is no regulatory requirement compelling device manufacturers to conduct skin tone-stratified validation before clearance.

The FDA issued a safety communication on pulse oximetry limitations in 2021 and has since indicated that future guidance will address skin tone diversity in validation studies. But guidance is not a mandate, and the installed base of existing devices will not automatically become more accurate when new standards are published for future submissions.

The Broader Principle

The pulse oximetry case illustrates a general principle in healthcare AI ethics: when the data-generating process is biased, the AI system trained on that data is biased, and the population harmed by the original measurement bias bears a second harm from algorithmic decisions made on the basis of it. Auditing AI systems for disparate performance across demographic groups is necessary but not sufficient — it is also necessary to audit the quality and representativeness of the underlying measurement infrastructure.

Key Takeaway

Pulse oximetry bias in Black ICU patients is not just a hardware problem — it is a case study in how measurement inequity propagates through AI clinical decision systems, producing compounded harm that can only be addressed by fixing both the measurement and the models trained on it.