Pulse oximeter error at Johns Hopkins reversed direction over five years of ICU data

An oximeter shines light through the skin at two wavelengths, one absorbed by oxygenated blood and the other by deoxygenated blood. It works out the oxygenated share by comparing how much of each is absorbed ^[11]. The Hopkins authors cannot say why that calculation gave different answers in different years. "One of the biggest problems with pulse oximeters is that they are this black box. I have no idea why we found what we found in this paper, and why it's changing year over year," said Ashraf Fawzy, a pulmonologist and critical care physician at Hopkins and an author on the paper ^[6].

Masimo, which made the oximeters in the study, said it was "unaware of any sensor, software, or manufacturing changes during the study period evaluated in the publications that would explain the reported results" ^[7]. The company listed other things that affect a reading: the type of sensor, where it is placed, signal quality and perfusion ^[8]. "The study's available data do not establish the cause of the reported shift from overestimation to underestimation," the spokesperson said ^[9].

On the design, Masimo has a point. A retrospective review of clinical records can show that the gap between oximeter and blood gas moved ^[2][5]. On its own, it cannot separate a change in the device from a change in sensors, placement or patients. Each factor on Masimo's list could shift over five years of intensive care practice with the hardware untouched ^[8].

STAT's report does not give the size of the error, the number of patients or paired readings, or the oxygen level above which underestimation begins. Those figures would show whether the drift is a fraction of a percentage point or large enough to change a decision at the bedside.

Earlier studies found oximeters overestimating oxygen on darker skin, making patients of color look healthier than they were ^[14]. EquiOx found underestimation, a result its authors and outside commentators struggled to justify ^[13]. I think a moving bias would let both be accurate, each measured at a different time. That requires the Hopkins drift to have happened at other hospitals too. EquiOx was one of the first studies to take readings prospectively and compare them with arterial blood gas tests ^[13].

Thomas Valley, a pulmonologist at University of Colorado Anschutz who worked on a 2020 New England Journal of Medicine paper on the devices' racial bias, described the clinical problem with a clock ^[15]. "If you have a broken clock and you know how it's going to be five minutes slow every time, you can adjust your routines to take care of that," he said ^[16]. It is now unclear whether that clock runs fast or slow ^[17]. "That's a huge problem, given how important these devices are," he added ^[18].

If the bias were fixed, physicians could adjust for it ^[20]. At Hopkins, the adjustment needed at higher oxygen levels changed sign between the start and the end of the review ^[21]. In my view, that turns skin-tone correction into a recurring local measurement, run on a hospital's own devices against its own blood gas results.