Age-bias mitigation and post hoc calibration improve deep learning MRI Alzheimer's classification

A deep learning model with generalized linear model feature residualization (PMDN) classified Alzheimer disease on brain MRI with balanced accuracy 0.71 and near-zero age bias; post hoc isotonic or Bayesian binning calibration reached max calibration error 0.21–0.24.

• Three independent structural MRI cohorts (ADNI, AIBL, DELCODE) were used; the sample size was not reported.
• PMDN removed age-related bias with near-zero age prediction correlation, while the adversarial CF-Net approach showed only partial and unstable debiasing.
• Nonparametric post hoc calibration methods (isotonic regression and Bayesian binning) yielded the lowest calibration errors, with MCE about 0.21–0.24.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.