OASIS-SB: A Sex-Balanced, Distribution-Preserving, Synthetic Phenotypic Dataset for Bias-Resilient Clinical Prediction

Sex bias remains a pervasive issue in computational neuroscience, influencing both dataset composition and model predictions. Leveraging the Open Access Series of Imaging Studies (OASIS), we created a new, cross-sectional, sex-balanced dataset to examine how such bias persists through statistical and machine learning pipelines. The longitudinal OASIS-2 dataset (150 participants, 373 MRI sessions) was first converted into cross-sectional form, and missing values were imputed using a Random Forest model (ROC-AUC = 0.9263). Dementia diagnosis was binarized (0 = person without dementia, 1 = person with dementia).

 An XGBoost classifier trained on this processed OASIS-2 data achieved an accuracy of 0.9759 ± 0.01% and an AUC of 0.9706, and was subsequently used to predict dementia status for all OASIS-1 participants. This resulted in a unified OASIS-1 and OASIS-2 dataset containing 588 subjects. To address the female-to-male sample imbalance, a K-Nearest Neighbors–based Synthetic Minority Oversampling Technique (KNN-SMOTE) was applied to oversample male cases, achieving a 1:1 sex ratio while preserving the female distribution.

The final dataset contains 712 unique cross-sectional records with demographic, cognitive, and volumetric features including:
Age, Sex, EDUC, SES, MMSE, CDR, eTIV, nWBV, ASF, Group.

Pearson correlation analyses confirmed that sex–feature relationships remained statistically equivalent to those observed in the original OASIS-2 data, indicating that the pipeline preserved real-world sex-linked structure despite class balancing. The resulting synthetic dataset, OASIS-SB, provides a larger, harmonized, and bias-transparent foundation for future cross-sectional dementia research and fairness-aware studies in neuroimaging.

This dataset is intended for reproducible research in neuroimaging, dementia classification, and fairness-aware machine learning. Users should cite the original OASIS publications (Marcus et al., 2007; Marcus et al., 2010) and this dataset release when used in derivative work.