From the study linked [https://link.springer.com/article/10.1007/s40279-024-02066-5] (with acronym meanings added inline for readability): "The lowest pulse arrival time standard deviation (PAT σ) at 2.0 times the standard deviation corresponded to 88.4% of the Region of Practical Equivalence (ROPE) for SDNN (Standard Deviation of Normal-to-Normal intervals), and 21.4% for RMSSD (Root Mean Square of Successive Differences). As the standard deviation of PAT increases, the equivalence between photoplethysmography-derived "heart rate variability" (PRV) and electrocardiography-derived heart rate variability (HRV) decreases for both SDNN and RMSSD. The width of the highest density interval (HDI), which encompasses 95% of the posterior distribution, increases with increasing PAT σ. This increase occurs at a higher rate for RMSSD than for SDNN."

So for "how accurate?", grossly and irresponsibly oversimplifying, the Apple method is roughly 90% accurate, and devices using the inferior method are roughly 20% accurate.

I would be really interested to know of any devices not locked down to Apple ecosystem that also use this approach, if anyone has any insight.