Can fairness tools reduce sex bias in Alzheimer’s AI? A 757-person study shows an internal improvement
In one ADNI cohort, a debiased logistic model improved balanced accuracy from 85.16% to 88.08% and moved equalized-odds difference closer to zero. The result has no external or prospective clinical validation.
The short version
- What happened
- The cross-sectional study used 757 ADNI participants: 600 with mild cognitive impairment and 157 with Alzheimer’s disease; 337 participants were women.
- Who it affects
- People assessed for cognitive decline need models that do not hide group-specific error patterns behind a single accuracy figure.
- What remains unproven
- Single cross-sectional ADNI cohort with one random train/test split and no external or prospective validation.
Keep in mind
Single cross-sectional ADNI cohort with one random train/test split and no external or prospective validation.
Only 157 participants had Alzheimer’s disease; the study classified AD versus MCI rather than predicting future progression or diagnosing a clinic population.
Sex was treated as a binary protected attribute; race, ethnicity, education, socioeconomic status and intersectional fairness were not evaluated.
The adversarial model was selected from 100 random seeds, increasing the need for an independent locked-model test.
LIME analysis of corrected predictions involved only eight people; feature attribution is not causal explanation.
No clinician decision, diagnostic timing, treatment, patient experience, harm or health outcome was measured.
Source analysis · 2 linked sources. Primary analysed source date: 6 November 2026. Earliest linked supporting source: 11 September 2026.
Read the full report · 8 min The evidence, methods, implications and limitations
The direct answer: fairness improved inside one test split
A peer-reviewed Scientific Reports study shows that selected fairness interventions can reduce sex-related differences in an Alzheimer’s disease classifier without lowering its performance inside the authors’ Alzheimer’s Disease Neuroimaging Initiative dataset. The clearest example paired logistic regression with a Disparate Impact Remover. Balanced accuracy rose from 85.16% to 88.08%, equalized-odds difference moved from -0.036 to -0.003, and eight previously misclassified participants were corrected. Those numbers describe one held-out split, not a clinical diagnostic trial.
The distinction matters because ‘fairer’ is not one universal property. The study reports several metrics that compare prediction rates and error patterns between women and men; some methods improved one measure while worsening another or overcorrecting. The authors selected two models offering favourable combinations of performance and parity. This is evidence that fairness should be measured and tuned explicitly during development. It is not evidence that the resulting model is ready to diagnose Alzheimer’s disease, improves care or works equitably across race, ethnicity, gender identity, education or socioeconomic status.[1][2]
Who was in the dataset—and what the classifier was asked to do
The cross-sectional cohort contained 757 ADNI participants aged 55 to 108. Six hundred had mild cognitive impairment, including 269 women and 331 men; 157 had Alzheimer’s disease, including 68 women and 89 men. The model’s task was to classify those two diagnostic groups. There were no cognitively unimpaired controls, no prediction of future conversion and no new clinical assessment. Women represented 337 participants and men 420, while the Alzheimer’s group was substantially smaller than the MCI group.
The starting data contained 129 variables: diagnosis and sex, demographic and genetic fields, Clinical Dementia Rating items, Functional Assessment Questionnaire items, Mini-Mental State Examination items and 52 MRI-derived brain-volume measures. Brain volumes were normalized by intracranial volume. Outcome-related proxy columns and constant variables were removed, missing numeric values were imputed using medians learned from the training data and the minority class was oversampled only within training. That workflow reduces leakage risk, but it cannot make ADNI representative of routine memory clinics.[1][2]
The distinction that matters
What the internal fairness result does—and does not—establish
The models were trained and tested inside one curated cohort. No clinic or second dataset was evaluated.
Cohort
757 ADNI participants
600 MCI; 157 Alzheimer’s disease
Split
70% train / 30% test
Five-fold tuning inside training
Logistic + DIR
88.08% balanced accuracy
Equalized-odds difference -0.003
Adversarial
88.91% balanced accuracy
Selected from 100 seeds
Missing evidence
No external cohort
No prospective diagnosis or patient outcome
Four classifiers and five different routes to mitigation
The researchers compared logistic regression, a polynomial support-vector classifier, random forest and XGBoost. Seventy per cent of participants were assigned to training and 30% to testing. Hyperparameters were selected by five-fold stratified cross-validation within training. They then applied five methods from IBM’s AI Fairness 360 toolkit: reweighing and Disparate Impact Remover before training; adversarial debiasing and prejudice-remover regularisation during training; and reject-option classification after training.
Sex was the protected attribute, with men labelled privileged and women unprivileged for the fairness calculations. Disparate impact compares the rate of favourable predictions between groups and is closest to parity at one. Statistical-parity, average-odds and equalized-odds differences are closest to parity at zero. These metrics answer different questions, and none says whether a diagnosis is clinically correct for an individual. Treating sex as binary also reflects the fields available in ADNI and does not analyse gender identity or people outside those two recorded categories.[1][2]
The strongest reported models do not tell one simple story
Baseline balanced accuracy ranged from 85.16% to 88.32%. The logistic baseline already had disparate impact of 0.997, statistical-parity difference of -0.017 and equalized-odds difference of -0.036. With Disparate Impact Remover, balanced accuracy reached 88.08%, disparate impact 1.029, statistical-parity difference 0.010 and equalized-odds difference -0.003. It reclassified eight people correctly—three women and five men—compared with the original logistic model.
An adversarially debiased model reached balanced accuracy of 88.91%, disparate impact 0.997 and statistical-parity difference -0.001. But it was selected as the best combination from 100 random seeds, which creates an additional model-selection step that needs external confirmation. A random forest with Disparate Impact Remover achieved 89.34% balanced accuracy and disparate impact 0.986, yet its equalized-odds difference stayed at -0.194. Choosing a system therefore requires stating which disparity and error cost matter clinically, not announcing that bias has been solved.[1][2]
Explainability found familiar clinical features—but does not validate causation
For the eight people corrected by the logistic mitigation, the team used LIME to compare local feature attributions before and after debiasing. Caregiving ability and left Brodmann area 36 atrophy became important in both sexes, while retirement status and an APOE genotype also appeared in both groups. Depression remained important only for women. With only four corrected Alzheimer’s cases in each sex subgroup, these local patterns are exploratory and cannot establish a stable sex-specific diagnostic pathway.
For the adversarial model, SHAP ranked total MMSE score, Functional Assessment Questionnaire score, age and medial temporal lobe volume among the strongest contributors. Their directions are clinically plausible: lower cognitive scores and smaller medial temporal volumes aligned with Alzheimer’s classification. Plausibility is useful for checking gross anomalies, but SHAP and LIME explain the fitted model, not disease biology or causal mechanisms. Correlated features can divide attribution in unstable ways, and an explanation cannot rescue a model that fails in another hospital.[1][2]
Why the cohort limits the equity claim
ADNI is a deeply characterised research cohort with standardised imaging and assessment, but it is not a population sample. Participants tend to have relatively high educational attainment and limited ethnic diversity. The study used one random 70/30 split rather than an untouched external cohort, and no confidence intervals are shown for all headline model comparisons. The sex distribution is fairly balanced overall, yet balanced representation does not guarantee balanced errors. Conversely, parity in this split does not guarantee fairness where prevalence, referral, imaging or access differ.
The analysis also focuses on sex in isolation. It does not test whether performance differs jointly by sex and race, ethnicity, language, education, income, disability or geography. It classifies established MCI versus Alzheimer’s disease rather than detecting overlooked early disease, predicting progression or measuring delayed diagnosis. The study reports no clinician interaction, threshold-selection process, false-negative consequences, treatment, distress or outcome. Those gaps are decisive for patients: an internally fair score is not the same as an equitable diagnostic service.[1][2]
What clinical developers should do next
The practical contribution is a workflow: evaluate several group and error metrics, compare mitigation methods, preserve held-out testing, inspect feature attribution and document trade-offs rather than relying on overall accuracy. Developers should preregister target disparities and acceptable error costs before model selection, repeat the analysis across random splits and report uncertainty. They should then validate a locked model in independent cohorts such as AIBL, NACC or OASIS and in the clinics where it would be used.
Prospective evaluation should ask whether the tool changes referrals, diagnostic timing, clinician disagreement, patient communication and access to care, not only whether parity metrics improve. The study’s code is public. ADNI was funded by US public agencies and numerous nonprofit and industry contributors; one author also reports German and European public funding, and the authors declare no competing interests. Confidence would rise with external, intersectional and prospective validation. It would fall if the selected fairness advantage disappears under a new split or trades fewer errors in one group for clinically serious errors elsewhere.[1][2]