Normal Reference Limits

Our goal is to create normal reference limits for neurological parameters such a nerve conduction velocity,
using a growing clinical database with measured patient data where some are likely to be abnormal and some normal.
Based on experience, about 5% of the data are abnormal.
The parameters are dependent on demographic information such as age, height, BMI and gender.
The plan was to:
- Stratify data into all combinations of demographic data.
- For each stratum, apply a Gaussian fit based on the central part and exclude data outside the fit.
- Perform multiple regression analysis on the pooled normal data inside the Gaussian fit from all strata.
With so many demographics combinations we end up with thousands of strata, if age resolution is 1 year, and height resolution is 1 centimeter.
To obtain 120-500 data per stratum, will require millions of data.
A typical clinic records only 20 values per parameter per day. I.e. our plan fails due to lack of data.
Is there a way to create a multiple regression model without stratification, thus limit amount of required data?