That does not define the scale that they're using. That's a typical hazard analysis risk matrix which has two axes which can be converted into a 4-point scale (Low, Medium, Serious, High). Importantly, to do a risk assessment in the style of IEEE 1789's you have to identify the specific Hazards that you're analyzing, which TFA does not claim to be doing in that table, instead speaking vaguely of "health risks". IEEE 1789 does not provide a mechanism for evaluating "health risks" without specifying exactly which risks are being evaluated.
You can see on page 27 how this is meant to be used: it should produce a per-hazard matrix.
You might be thinking of Figure 18 on page 29, which does identify Low-risk and No-effect regions by Modulation % and Frequency, but that also does not claim to identify high-risk regions, it just identifies the regions we can be highly confident are safe. And importantly, as a sibling comment notes, TFA's table actually contradicts the line on Figure 18, labeling several devices as higher than Low even when they're squarely within the Low-Risk and No-Effect zones.
They list the 'Xiaomi 15 Ultra' as having a 'Moderately High' health risk, and cite it as having a 2.16 kHz PWM frequency at 30-75% modulation depth.
The IEEE article has recommended practices that state:
8.1.2.3 Example 3: PWM dimming
Using Figure 20, the recommended practice for PWM dimming at 100% modulation depth is that the frequency satisfies f > 1.25 kHz. This can also be derived using Recommended Practice 1 and solving 100% = 0.08×fFlicker. This level of flicker could help minimize the visual distractions such as the phantom array effects.
Seems like even at 100% mod depth, >1.25 kHz is just fine.
Also, the article does not seem to distinguish between modulation at reduced brightness, which the IEEE article calls out specifically as something that is unlikely to cause issues. E.g., movie theaters using film all flicker at 48 Hz and nobody complains about that.
