Personal regression versus cohort sharing AI
The report compares two ways of estimating the same 42-feature pressure relationship. Personal ridge fits each person’s slopes independently. Cohort partial pooling estimates a shared slope pattern and a personal deviation from it. Both retain personal pressure calibration; neither is an uncalibrated new-person predictor.
This differs from a fully pooled regression, which would force one common slope vector. Partial pooling can stabilize noisy individual estimates while allowing relationships to differ across people. It can also harm a person whose relationship differs from the shared pattern.
Implementation in this report
Inputs use each person’s training-only imputation, mean and SD. Pressure is centered and divided by that person’s calibration BP SD, separately for SBP and DBP. The fit minimizes squared normalized errors plus two squared-coefficient penalties: one on shared slopes and one on personal deviations. Predictions return to mmHg using personal calibration mean and SD.
Increasing the deviation penalty enforces more sharing. Increasing the shared-slope penalty suppresses the common profile and moves the solution toward independent personal slopes. Shared/deviation penalties are 100/100 for SBP and 100/300 for DBP.
The calibration splits compare shared penalties 3, 10, 30 and 100 with deviation penalties 30, 100, 300 and 1,000. Mean participant validation MAE chooses the pair. Comparing BP-SD normalization with mean-centering in mmHg uses the same validation scores; normalization is selected for both targets. Held-out pressure does not select these settings.
The common pattern is learned from the calibration records of the same cohort whose later beats are evaluated. The results therefore concern known, calibrated participants. They do not establish performance on a new person. This model is an extension implemented in this project, not a method attributed to Carlson et al.