Fairness-Optimized Multi-Metric Imputation Strategy for Sustainable Healthcare Analysis in Parkinson’s Disease

Abstract. Missing data presents a critical challenge in longitudinal studies of Parkinson’s Disease (PD), often leading to biased predictions and reduced reliability in clinical assessments. Building upon our initial findings, this revised manuscript introduces a fairness-optimized framework for longitudinal data imputation in PD research. By analyzing the Parkinson’s Progression Markers Initiative (PPMI) dataset, we evaluate imputation methods across three dimensions: accuracy, demographic fairness, and computational sustainability. Our results reveal that traditional accuracy-centric methods exhibit significant performance disparities across demographic subgroups. Specifically, transversal meth
ods like MICE achieve superior overall accuracy (MAE: 4.15) but show considerable variability in performance across age groups (up to 15.2% difference). Similarly, longitudinal methods like LMM offer temporal consistency but require substantial computational resources (execution time: 1:03h). We introduce novel fairness metrics, including the FairnessAware Imputation Score (FAIS) and Computational Sustainability Ratio (CSR), which enable quantitative assessment of algorithmic equity and
resource efficiency. Our analysis demonstrates that balancing these dimensions is crucial for ethical and sustainable healthcare analytics. The highest FAIS value (0.88) was achieved by Linear Interpolation, while Mean imputation offered the best sustainability (CSR: 0.64). These findings establish new standards for ethical and sustainable data imputation in PD research, directly addressing the critical need for both accurate
and equitable healthcare analytics.