MITIGATING GENDER BIAS AND FAIRNESS IN DEMENTIA CLASSIFICATION USING BORDERLINE SMOTE AND ENSEMBLE MACHINE LEARNING MODELS

Abstract— Fairness issues in medical AI can lead to uneven diagnostic
performance across demographic groups. In MRI-based dementia
detection, skewed class and gender distributions often degrade equity.
Using an MRI dataset of 150 adults (60–96 years) labelled as
demented, non-demented or converted, we trained Logistic
Regression, SVM, Random Forest and Gradient Boosting models
under three data regimes: Original (imbalanced), SMOTE and
Borderline-SMOTE. We standardised features, performed stratified
70/30 train–test splits and tuned models with grid search and five fold
cross validation. Performance was evaluated via Accuracy, Precision,
Recall, F1-score and MCC, while fairness was assessed by reporting
metrics separately for male and female subgroups under an equal
opportunity notion focused on subgroup F1-scores. Performance
improved monotonically from Original to SMOTE to Borderline-
SMOTE. Ensemble methods showed the strongest gains after
Borderline-SMOTE (MCC > 0.85). The male–female F1-score gap
narrowed from about 7 percent (Original) to about 4 percent (SMOTE)
and to under 3 percent (Borderline-SMOTE), indicating progressive
bias mitigation while maintaining accuracy. Borderline-SMOTE, by
focusing synthesis near decision boundaries, outperformed standard
SMOTE, improving stability and reducing gender disparity. Coupling
targeted resampling with ensemble learning is a practical route to fairer
dementia classifiers without sacrificing predictive power.