Age bias and calibration in deep learning Alzheimer's disease classification from structural MRI

Source article: Improving the fairness and robustness of deep learning classification models through bias mitigation and post hoc calibration

The goal of this study was to enhance the fairness and clinical applicability of deep learning models for classifying Alzheimer's disease from structural MRI scans. To achieve this, we addressed two key challenges: age-related bias and unreliable probability estimates. Specifically, we investigated in-processing debiasing strategies to achieve age-invariant predictions while preserving diagnostic accuracy. We also systematically evaluated post hoc calibration methods to improve the reliability, interpretability,…

By the Truvace deskOctober 20, 20262 min read1 sourcePeer-reviewedTRV-2026-1958· cite & verify

Improving the fairness and robustness of deep learning classification models through bias mitigation and post hoc calibration — Conditions pour IA digne de confiance conditions requirements for Trustworthy AI by Lamiot. CC BY 4.0 · https://creativecommons.org/licenses/by/4.0

Trace impact readingContested

P 72

Both sides are scored from claims and sources, not community votes.

G 71

In brief

Researchers tested ways to make deep learning models that classify Alzheimer's disease from structural MRI less dependent on patient age and better at reporting trustworthy probabilities. They compared two in-processing debiasing strategies in a shared 3D convolutional network and four post hoc calibration methods, using scans from three cohorts: ADNI, AIBL, and DELCODE.

The work matters because a diagnostic model that leans on age or gives poorly calibrated probabilities is harder to trust in clinical settings. The results favor feature residualization over adversarial training and nonparametric calibration, but the reported accuracy is moderate and the abstract does not show how the methods perform in routine clinical practice.

Main points

PMDN, a GLM-based feature residualization method, gave the best overall performance, with balanced accuracy of 0.71 and a Brier score of about 0.22.
PMDN reached near-zero correlation between predictions and age, while CF-Net's adversarial debiasing was only partial and unstable.
Isotonic regression and Bayesian binning produced the lowest calibration errors, with MCE of about 0.21 to 0.24.
PMDN was more consistent than CF-Net, particularly in clinically relevant subgroups.

The gain

A GLM-based residualization method (PMDN) removed age-related bias in a 3D CNN for Alzheimer's MRI classification while keeping balanced accuracy at 0.71 and improving calibration.

The problem

The adversarial debiasing approach (CF-Net) achieved only partial and unstable age-bias removal, and models with similar classification performance differed significantly in calibration, stability, and fairness.

The rundown

The study used structural MRI data from three independent cohorts, ADNI, AIBL, and DELCODE, processed through a standardized pipeline. Both debiasing strategies were implemented within one shared 3D CNN framework, which allowed a direct comparison.

Post hoc calibration used four regression methods: logistic, probit, isotonic, and Bayesian binning. Models were evaluated with classification metrics as well as calibration-specific and fairness metrics.

The authors report that PMDN showed substantially reduced variability compared with baseline and CF-Net. They conclude that calibration further improves the reliability of predicted probabilities, with nonparametric methods performing best.

What this doesn’t fix

Models achieved similar classification performance and moderate balanced accuracy (0.71), and CF-Net debiasing was unstable, so the benefits come mainly from fairness and calibration rather than from better diagnostic accuracy.