Select Page

As steps toward mitigating racial bias, Hewlett-Packard’s engineers marshaled a range of more inclusive approaches to oximetry. The instrument’s baseline calibrations were set by working with a “carefully selected” group, including 248 Black volunteers—which is, notably, 246 more nonwhite people than the FDA currently suggests for pre-market testing of the oximeters in hospitals today. Most importantly, the device could be personally adjusted for every individual. There was an option to squeeze a small droplet of blood from the wearer’s ear to scan the blood alone using spectrophotometry. This measurement, which helped discern exactly how much light was being absorbed by an individual’s skin and tissues, allowed the physician to personalize light level calibrations and optimize the device’s accuracy.

The oximeter could also account for circulation idiosyncrasies. Unlike modern pulse oxes that are tested only on healthy people, Hewlett-Packard’s device was designed to work for people who may be sick. The sensor was not made for the fingertip, for instance, because then the device wouldn’t work as well for patients with common health conditions such as shock, sepsis, and certain chronic illnesses. Instead, Hewlett-Packard placed its sensor on the top curve of the ear, one of the last parts of the body to be impacted by circulation issues during illness. This choice helped prevent building ableism into oxygen measures, while also avoiding gender disparities due to poor device fit. While ear oximeters still exist in specialty niches, by far the most common models in ERs and homes today are nonadjustable and built to fit the “average” geometry of a man’s finger, at times producing suboptimal readings for all others that may well compound with other errors.

Despite these achievements, when the personal computing market exploded in the ’80s, Hewlett-Packard shifted its focus and stepped back from medical equipment shortly before it released a long-planned miniature version of its oximeter. But Kryger still describes their larger device as “the best oximeter ever made.” His lab’s publications from that time show that the HP oximeters were in a range of ways more accurate than the pulse oximeters that soon came to take their place. They were referred to in clinical studies as the noninvasive “gold standard” by which early pulse oxes were tested, because Hewlett-Packard oximeter readings more closely matched the invasive arterial blood gas tests.

As the pandemic has painfully reminded us, the consequences of such inaccuracies can be devastating. Because today’s hospital oximeters are not made with capacity for personalization, they can inadvertently feed flawed data not only to doctors but also to other machines. Oximeter numbers provide key inputs to a range of computing systems, including the algorithms guiding ICU triage and certain insurance reimbursements. They are also on closed-loop algorithms with many ventilators—and when fed error-ridden inputs, such devices may not be able to optimize as effectively. Having these conversations now is crucial: As part of AI’s growing role in health care, a wide range of noninvasive sensors are being developed with the pulse oximeter as their model. Some, like certain optical sensors for sepsis or blood glucose, may already be at your local hospital or present in your home. Without care, a coming generation of optical color sensors could easily reproduce the unequal errors for which pulse oximetry is now known across many other areas of medicine.

We tend to assume that technology will unfold with a kind of linear progress, and that useful features or key questions will be built into future models. The history of devices often gets written later as if this had always been the case—that alternative approaches didn’t succeed because they were inferior. But like any history, it is useful to ask who wrote it and what’s left out.