Impact of Imputation Methods on Supervised Classification: A Multiclass Study on Patients with Parkinson's Disease and Subjects with Scans Without Evidence of Dopaminergic Deficit

International Neuroscience and Neuroimaging databases commonly face the missing data problem. Performance of supervised classification could be negatively affected by the presence of missing data since most Machine Learning (ML) algorithms are not built to work with them. Several imputation methods handle the information leakage, but the influence on classification performance is poorly investigated in neuroscience field. In this paper, we compared two different imputation approaches, MissForest (MF) and Mean Imputation (Imean), on a Parkinson's Progression Markers Initiative dataset, to evaluate the classification performance of three supervised tree-based algorithms: Random Forest, XGBoost and LightGBM. First, we amputed the complete dataset with increasing percentages of missing data (10%, 30%, 50% and 70%) by applying Missing Completely At Random. Then, we used Imean and MF approaches on amputed datasets. Finally, we compared their classification performance for each percentage of amputation using PyCaret tools. Although, the average error on single features was comparable for MF and Imean in each amputation percentage, the performance of supervised classifiers changed. Comparing the classification metrics of best model in the original dataset, Imean outperformed, highlighting overfitting. On the contrary, MF demonstrated the ability to retain accurate information making it most powerful for neuroscience application.