Bridging the Gap: Missing Data Imputation Methods and Their Effect on Dementia Classification Performance.

Area of Science:

• Neuroscience and neuroimaging research.
• Clinical applications in neurodegenerative diseases.

Background:

• Missing data is prevalent in neuroscience studies, particularly for Alzheimer's Disease (AD) and Mild Cognitive Impairment (MCI).
• Inaccurate handling of missing data can negatively impact machine learning (ML) model performance and interpretability.

Purpose of the Study:

• To systematically compare the effects of five data imputation methods on ML classification accuracy.
• To evaluate imputation strategies using multimodal data from the Alzheimer's Disease Neuroimaging Initiative (ADNI).

Main Methods:

• Analysis of clinical, cognitive, and neuroimaging data from ADNI participants with MCI or AD.
• Application of five imputation techniques: mean, median, k-Nearest Neighbors (kNNs), Multiple Imputation by Chained Equations (MICE), and missForest (MF).
• Classification using Random Forest (RF), Logistic Regression (LR), and Support Vector Machine (SVM) models, with performance assessed via McNemar's test.

Main Results:

• Multiple Imputation by Chained Equations (MICE) achieved the highest accuracy for RF (0.76) and LR (0.81).
• Support Vector Machine (SVM) performed best with median imputation (0.81).
• Significant differences in classification performance were observed between RF and other models (p < 0.01).

Conclusions:

• The selection of an imputation method significantly influences classification accuracy in neurodegenerative disease research.
• Tailoring imputation strategies to specific data characteristics and chosen classifiers is vital for robust predictive modeling.