Fairness-Aware Evaluation of Missing Data Solutions for Longitudinal Parkinson’s Disease Research

Longitudinal clinical datasets in neurodegenerative disease research face critical challenges from pervasive missingness and stringent privacy constraints. This study introduces a comprehensive fairness-aware evaluation of 12 imputation methods and 4 synthetic data generation techniques for Parkinson’s disease longitudinal research using the Parkinson’s Progression Markers Initiative dataset. Advanced methodologies were implemented, including HyperImpute ensemble optimization, variational deep embedding with recurrence, and conditional tabular generative adversarial networks, across 1,483 PPMI participants spanning clinical, demographic, and biomarker variables. The evaluation integrates pointwise accuracy, distributional fidelity via sliced Wasserstein distance, temporal consistency, clinical range validity, and stratified fairness analyses across demographic subgroups. Fairness disparities were quantified using relative error differentials computed as the ratio of subgroup-specific MAE differences to reference group MAE, with stratification across age, education, disease duration, and sex. Uncertainty quantification was performed via bootstrapping (n = 1000) for key metrics, with 95% confidence intervals reported where statistically significant differences emerged. HyperImpute achieved superior imputation performance with mean absolute error of 5.16 compared to baselines (5.19–5.57) and maintained the highest coefficient of determination (R2 = 0.260). CTGAN delivered optimal synthetic fidelity with sliced Wasserstein distance of 0.039 ± 0.012 versus 0.062–0.146 for alternatives. Systematic bias analysis identified 23% increased cognitive imputation errors in participants aged 70 years or older and 18% education-related reconstruction disparities. The small MAE improvement of HyperImpute over the LMM baseline was confirmed as statistically significant under bootstrap analysis (n = 1000), and we therefore interpret HyperImpute’s superiority as statistically robust but modest in magnitude. Sex-stratified analysis revealed 8% higher motor assessment imputation errors for female participants compared to males, consistent with known sex-based differences in PD presentation. The framework provides evidence-based guidance for selecting data completion strategies by missingness mechanisms, clinical objectives, resource constraints, and fairness requirements. Discussion of framework adaptability to irregular visit schedules and other neurodegenerative pathologies extends the translational scope beyond PPMI. Findings promote reproducible, equitable, and privacy-preserving innovation in neurodegenerative disease research through rigorous quantification of temporal consistency and demographic bias.