Perhaps I am missing something in your problem statement, however linear regressions such as that done by fitlm is a least-squares approach and entirely deterministic, not ‘training’. There is nothing heuristic about it. You should get the same result (within the limits of floating-point calculation precision) every time you run it using the same data. If you use subsets of the entire data set, you will of course get different regression coefficients depending on the data you use, and they may be different (I would expect them to be different) from the values using the entire data set. Also, if you use the predict function to calculate the ‘regressionLine’ values, it will also supply the confidence limits on the regression.
With respect to the systolic, diastolic, and mean blood pressure, note that the mean blood pressure (averaged over the entire cardiac cycle) is approximately one-third the difference of the systolic and diastolic pressures, not their arithmetic mean, so —
SBP = 120;
DBP = 80;
meanBP = (SBP - DBP)/3 + DBP
not —
avgBP = mean([SBP DBP])
.
