Generally, for determining normal range of any parameter, a reasonably large sample of apparently healthy population is tested for that parameter. The normal range is then taken to be 2.5th to 97.5th percentile value. Values above and below this range are categorized as abnormal (abnormally low or high).
For some parameters, e.g. eyesight, one side of range is better than normal and may be called super-normal and not abnormal.
Round, convenient values are often taken as limits of normal ranges and mild/moderate/severe categories, so that they can be easily remembered and applied in busy clinics. Some of the blood sugar, blood pressure and BMI cutoffs follow this principle.
For some parameters, prospective studies of outcome (life expectancy, morbidity and mortality) may show a particular range to be the best or optimal and such range may be used to determine normality. Yu Chen et al (http://www.bmj.com/content/347/bmj.f5446) found a U shaped association between BMI of Asians and cardiovascular deaths.
The risk is often continuous, but the values are categorized into groups so that odds ratio can be calculated. Odds ratio provides easy method to compare risk in different categories. Regarding BMI, the World Health Organization page (http://apps.who.int/bmi/index.jsp?introPage=intro_3.html) also mentions:
The health risks associated with increasing BMI are continuous and the
interpretation of BMI gradings in relation to risk may differ for
There are many who strongly oppose conversion of continuous data to categorical: http://biostat.mc.vanderbilt.edu/wiki/Main/CatContinuous