Electronic-record race: Black
threshold question
two lights, one proprietary curve, one consequential estimate
The Number Under the Fingernail
A pulse oximeter does not count oxygen molecules. It estimates arterial saturation from changing red and infrared light, then passes their ratio through a device-specific calibration. The number can be wrong. The error can cross a threshold, and the crossings in published cohorts are not evenly distributed.
Not a personal prediction. This page recomputes frequencies in published study records. It cannot tell you your arterial saturation, assess a device, or advise what to do with a reading.
layer one: the published threshold
A reassuring display, a lower arterial value
Sjoding and colleagues asked a precise cohort question. Among paired measurements where the pulse oximeter read from 92% through 96%, how often was directly measured arterial saturation below 88%? Choose either hospital cohort. Drag the displayed reading. The exact-reading sample size changes where the paper publishes it; the probability does not, because only the full-band event counts were published.
SpO2 display band
Paired SaO2 crosses the paper's 88% line
Electronic-record race: White
Risk ratio, Black / White
Risk difference
The bars are a two-bin distribution: below 88%, in red, and at least 88%, in dark green. Every length comes from a published numerator and denominator. Nothing about one person's skin, device, blood flow, or illness is inferred from it.
The wording matters. These are paired measurements classified by electronic-record race, not people assigned a melanin value. Pairs were within ten minutes in the retrospective study, not necessarily simultaneous. One patient could contribute several pairs.
mean-error question
What was the average SpO2 minus SaO2 across a controlled range?
layer two: the dismissal that survives layer one
Race is not pigment. An average is not a tail.
A Black-versus-White retrospective comparison cannot isolate melanin. A small mean bias can also hide large errors in both directions. So do not pool the records. Sort them. Ask which population was measured, how pigmentation was represented, whether perfusion was constrained, which device was present, and which error statistic answered the question.
Three error axes, none substitutable
Bias
mean(SpO2 − SaO2)Signed errors can cancel. A mean near zero does not say the tails are small.
ARMS
sqrt(mean((SpO2 − SaO2)²))Large errors count strongly, in either direction. It is not a missed-threshold frequency.
Threshold crossing
P(SaO2 < 88 | SpO2 92..96)This asks about one clinical region. Move either boundary and the estimand changes.
Khanna's controlled TruSignal record reports whole-range ARMS of 1.71 points for its dark pigment group and 1.64 for its light group, with a 0.08-point difference in mean bias. Sjoding's Michigan threshold frequencies are 11.75% and 3.56% by electronic-record race. Those are not competing answers to one calculation. The filter map keeps their different designs, instruments, saturation ranges, group definitions, and estimands attached.
Now add low perfusion. Gudelunas reports aggregate missed-diagnosis frequencies of 21.1%, 8.2%, and 1.1% for dark, medium, and light pigment groups under low-perfusion conditions in a controlled laboratory study of two devices. Add critical illness: Fawzy's 12-person ICU pilot reports higher ARMS in its darkly pigmented group. Add a much larger but high-saturation ICU sample: the in-press EquiOx record reports negative median bias overall and only two occult events. Direction itself can change with probe site and sample.
The check, with teeth
- RUNRecomputing assertions in this browser.
Free choices: the page follows Sjoding's thresholds of SaO2 below 88% and SpO2 from 92% through 96%; uses unadjusted count-based rates; defines error everywhere as SpO2 minus SaO2; and uses a descriptive 95% Wilson interval for visible pair counts.
Uncertainties: repeated pairs are not independent people; exact-reading event counts are unavailable; the Figure 1 bin totals do not equal the text denominators; electronic-record race is not pigment; several records are single-device or manufacturer-funded; aggregate rows cannot be re-derived; co-oximetry is a comparator, not an error-free oracle; and the 2026 EquiOx record was still an in-press pre-proof on 2026-08-25.
Refusal: when filters leave no compatible source record, the map says so and computes nothing. The verifier sweeps every filter combination and asserts that this path is visible.