Pulse Oximeter Bias: Unfixed 30-Year Flaw in AI-Driven Care

The promise of artificial intelligence in healthcare is often framed by its potential to revolutionize diagnostics, personalize treatment, and enhance patient outcomes. Yet, a critical examination of its deployment reveals a stark reality: AI can inherit and even amplify existing biases, leading to significant disparities in care. This is not a hypothetical concern for the future; it is a problem deeply embedded in technologies we rely on today, exemplified powerfully by the persistent issue of racial bias in pulse oximeters.

A Long-Standing Failure: The Pulse Oximeter’s Blind Spot

Pulse oximetry, a seemingly innocuous and ubiquitous medical device, measures blood oxygen saturation (SpO2) non-invasively. It is a cornerstone of patient monitoring, guiding critical decisions from supplemental oxygen administration to intubation. However, for decades, these devices have systematically overestimated oxygen saturation in individuals with darker skin pigmentation, leading to potentially dangerous undertriage and delayed interventions. This is not a new discovery; the phenomenon has been documented since at least 1990. The core of the problem lies in how pulse oximeters function. They transmit red and infrared light through tissue and measure the absorption of these wavelengths. Hemoglobin, the protein in red blood cells that carries oxygen, absorbs light differently depending on whether it is oxygenated or deoxygenated. Melanin, the pigment responsible for skin color, also absorbs light, particularly in the red and infrared spectrum. In individuals with higher melanin concentrations, this additional absorption can be misinterpreted by the device’s algorithms, leading to an artificially inflated SpO2 reading. Research has consistently highlighted the magnitude of this error. Studies have shown that pulse oximeters can have a 50% to 860% greater error rate in individuals with dark skin compared to those with light skin Peer-reviewed study on pulse oximeter racial bias. This substantial discrepancy means that a patient with darker skin could be experiencing dangerously low oxygen levels, yet their pulse oximeter reading might appear within a normal or only slightly reduced range, masking their true physiological state. Recent large studies, including one published in The BMJ in January 2026, continue to confirm that fingertip pulse oximeters give higher readings for patients with darker skin tones, with SpO2 readings, on average, 0.6-1.5 percentage points higher for darker skin tones, leading to increased false negative rates for detecting low oxygen levels.

The Regulatory Labyrinth: FDA 510(k) and SaMD Frameworks

The continued prevalence of this bias raises serious questions about the regulatory pathways governing medical devices, particularly in the context of the FDA’s oversight. Most medical devices, including pulse oximeters, gain market clearance through the FDA 510(k) Pathway. This pathway requires manufacturers to demonstrate that their new device is “substantially equivalent” to a predicate device already on the market. While intended to streamline innovation, this approach can inadvertently perpetuate existing flaws if the predicate device itself harbored biases or shortcomings that were not adequately addressed during its initial clearance. The FDA Center for Devices and Radiological Health (CDRH) is responsible for ensuring the safety and effectiveness of medical devices. Yet, the long-standing issue of pulse oximeter bias demonstrates that even FDA-cleared devices can have systematic racial bias. This underscores a critical gap in the regulatory framework: the need for more rigorous evaluation of algorithmic fairness and equity, particularly for devices that rely on light absorption or other physiological measurements that can be affected by demographic factors. In January 2025, the FDA released new draft guidance specifically addressing pulse oximeters, recommending increased testing pool sizes, diverse skin tone representation in studies, and objective skin tone measurement methods like colorimetry. The agency has also proposed creating a publicly available webpage to identify pulse oximeters that demonstrate comparable performance across skin pigmentations. With the advent of AI and machine learning in healthcare, the FDA has also developed the FDA SaMD Framework for Software as a Medical Device. While this framework aims to address the unique challenges of software-based medical devices, the pulse oximeter experience serves as a stark warning. If traditional hardware devices can embed such profound biases for decades, the potential for AI-driven SaMD to introduce or amplify similar disparities is immense if not proactively managed. The complexity of AI algorithms can make bias even harder to detect and mitigate without dedicated testing protocols. Building on this, the FDA finalized guidance on Predetermined Change Control Plans (PCCPs) for AI-enabled device software in December 2024, updated in August 2025, to manage iterative algorithm updates. Additionally, draft guidance on the Total Product Lifecycle Management of AI/ML SaMD, released in January 2025, outlines recommendations for addressing algorithmic bias and transparency throughout a device’s lifespan. The new Quality Management System Regulation (QMSR), effective February 2, 2026, further aligns U.S. requirements with international standards, emphasizing a risk-based approach to device quality.

Expert Voices and Unheeded Warnings

Prominent voices in healthcare and AI have consistently highlighted these dangers. Dr. Ziad Obermeyer, a physician and researcher, has been a leading figure in exposing algorithmic bias in healthcare, including in areas where AI is used to allocate resources or predict health outcomes. His work, often published in prestigious journals like the New England Journal of Medicine (NEJM), has demonstrated how seemingly neutral algorithms can produce racially biased results, leading to inequities in access to care. His recent work includes a 2024 paper on race adjustments in clinical algorithms and testimony before the U.S. Congress in February 2024 regarding AI’s potential and pitfalls in healthcare. Similarly, Dr. Eric Topol, a renowned cardiologist and geneticist, has frequently emphasized the need for rigorous validation and ethical considerations in the deployment of AI in medicine. He has cautioned against the uncritical adoption of AI tools without a thorough understanding of their potential for harm, particularly in exacerbating health disparities. The persistent issue with pulse oximeters reinforces their warnings, showcasing a tangible, life-threatening example of how systemic bias can become embedded in medical technology and persist despite expert scrutiny. In May 2026, Dr. Topol highlighted the ‘paradox’ of medical AI implementation, noting the slow adoption of proven AI tools while unproven large language models are being used for care. Multiple device manufacturers have been implicated in producing pulse oximeters that exhibit this racial bias. While some manufacturers have acknowledged the issue, concrete and widespread corrective actions have been slow to materialize. However, some manufacturers, such as Masimo and Nonin Medical, have recently released FDA-cleared pulse oximeters claiming improved accuracy across diverse skin tones. This highlights a collective failure across the industry to prioritize equitable performance across diverse populations, even when the underlying scientific principles for bias are well-understood.

The Economic and Ethical Imperatives of Equitable AI

The economic implications of unaddressed algorithmic bias in devices like pulse oximeters are substantial. Health disparities, often exacerbated by such biases, lead to higher rates of chronic disease, increased hospitalizations, and greater healthcare utilization among marginalized populations. This translates into significant economic burdens on healthcare systems, payers, and society at large. Data from sources like the Kaiser Family Foundation and CMS.gov consistently illustrate the disproportionate impact of health inequities on covered lives and vulnerable populations, leading to avoidable costs and diminished productivity. Furthermore, the lack of equitable performance in medical devices erodes trust in the healthcare system, particularly among communities that have historically experienced discrimination. This distrust can lead to reduced engagement with healthcare providers, delayed care seeking, and poorer health outcomes, creating a vicious cycle that further entrenches disparities. For investors and policymakers, this represents not only an ethical failing but also a market inefficiency. Companies that fail to address these biases risk losing market share, facing regulatory scrutiny, and incurring reputational damage. Robust post-market surveillance, a critical component of the FDA SaMD Framework, is essential but must specifically include rigorous analysis of performance across diverse demographic groups to identify and rectify such biases proactively. The example of pulse oximeter racial bias serves as a potent reminder that the pursuit of technological advancement in healthcare must be inextricably linked with a commitment to equity. The lessons learned from this long-standing failure must inform the development and regulation of all future medical AI. Without a fundamental shift towards designing, testing, and deploying AI solutions that are fair and effective for all, the promise of AI in healthcare risks becoming another mechanism for perpetuating and deepening existing health disparities. Patient safety advocates, clinicians, and regulatory officers alike must demand accountability and systemic change to ensure that medical technology serves everyone, without exception.