Machine Learning for Missing Data Imputation in Alzheimer's Research: Predicting Medial Temporal Lobe Flexibility

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Alzheimer's disease (AD) pathology begins years before symptoms, necessitating early detection biomarkers.
• Medial temporal lobe (MTL) network flexibility, a measure of brain connectivity, is an early indicator of AD-related decline.
• Cognitive, genetic, and biochemical markers may predict MTL dynamic flexibility.

Purpose of the Study:

• To predict medial temporal lobe (MTL) dynamic flexibility using multimodal data in an aging cohort.
• To evaluate advanced machine learning and imputation methods for handling high missing data rates in AD research.

Main Methods:

• Utilized data from 656 participants, including cognitive, genetic, and blood biomarkers.
• Evaluated four missing data handling methods (case deletion, MICE, MissForest, GAIN) and five regression models (Ridge, k-NN, SVR, regression trees, ANN).
• Optimized hyperparameters via grid search and assessed model performance using MAE, RMSE, and runtime with cross-validation.

Main Results:

• Identified 25.86% missing values, with only 6.40% complete cases after listwise deletion.
• MissForest imputation combined with Random Forest regression achieved the best performance (MAE = 0.083), a 54.7% improvement over case deletion.
• MissForest significantly outperformed GAIN and MICE (p < 0.001), while GAIN was the fastest imputation method.

Conclusions:

• Robust imputation strategies are crucial for maximizing data utility and model reliability in studies with substantial missing data.
• Machine learning models, particularly with advanced imputation, can effectively predict MTL dynamic flexibility.
• Further research incorporating neuroimaging is needed to refine predictive models for clinical applications in AD.