Autoencoder imputation of missing heterogeneous data for Alzheimer's disease classification

Area of Science:

• Artificial Intelligence in Medicine
• Neuroscience Data Analysis
• Biomedical Informatics

Background:

• Missing data is a major obstacle in Alzheimer's disease (AD) diagnosis and research.
• Existing data imputation methods for AD data have limitations, especially with deep learning.
• Heterogeneous AD datasets (tau-PET, MRI, genetics, clinical assessments) present unique imputation challenges.

Purpose of the Study:

• To evaluate denoising autoencoder (DAE) effectiveness for imputing missing key features in comprehensive AD datasets.
• To assess DAE performance in handling extreme missingness (≥40%) in AD-related features.
• To integrate DAE-extracted latent features with traditional features for improved AD classification.

Main Methods:

• Utilized a denoising autoencoder for imputing missing data in heterogeneous AD datasets.
• Focused on key AD progression-dependent features like maternal history of AD, APOE ε4 alleles, and Clinical Dementia Rating.
• Employed random forest classification with 10-fold cross-validation on imputed and feature-selected datasets.

Main Results:

• Imputed datasets demonstrated robust Alzheimer's disease predictive performance (accuracy: 79%-85%; precision: 71%-85%) across various missingness levels.
• High recall values were achieved even with 40% missing data.
• Feature-selected datasets, including autoencoder-derived features, outperformed the original complete dataset in classification scores.

Conclusions:

• Denoising autoencoders are effective and robust for imputing crucial missing information in Alzheimer's disease data.
• This imputation strategy enhances the reliability of AI-based clinical decision support systems for AD prediction.
• The study highlights the potential of deep learning for handling complex, incomplete biomedical data.