Correlation Constraints for Regression Models: Controlling Bias in Brain Age Prediction

In neuroimaging, the difference between chronological age and predicted brain age, also known as delta , has been proposed a pathology marker linked to range of phenotypes. Brain is estimated using regression, which involves frequently observed bias due negative correlation delta. prediction models, this can manifest an overprediction young brains underprediction for elderly ones. We show that be controlled by adding constraints model training procedure. develop analytical solution constrained optimization problem Linear, Ridge, Kernel Ridge regression. The optimal in least-squares sense i.e., there no other satisfies better fit. Analyses on PAC2019 competition data demonstrate approach produces unbiased predictive models with number advantages over existing approaches. Finally, we introduce regression toolboxes Python MATLAB implement our algorithm.