From what I've read it simply takes the ratio of absorbance from the systole and diastole and considers that a stable non-linear estimator of %SpO2. The device

From what I've read it simply takes the ratio of absorbance from the systole and diastole and considers that a stable non-linear estimator of %SpO2. The device itself needs to be calibrated to get the non-linear fit in place, but each individual is assumed to be (in gross measure) similar enough to get meaning out of it. Of course, there's going to be huge variation across finger types, blood volumes, lighting conditions, placement. I wouldn't consider them to be accurate much more than those impedance-based body fat percentage computers. In both cases, huge discrepancies from healthy values are still telling despite small-scale noise.