Longitudinal Enrichment of Imaging Biomarker Representations for Improved Alzheimer's Disease Diagnosis
Longitudinal data is often available inconsistently across individuals resulting in ignoring of additionally available data. Alzheimer's Disease (AD) is a progressive disease that affects over 5 million patients in the US alone, and is the 6th leading cause of death. Early detection of AD can significantly improve or extend a patient's life so it is critical to use all available information about patients. We propose an unsupervised method to learn a consistent representation by utilizing inconsistent data through minimizing the ratio of $p$-Order Principal Components Analysis (PCA) and Locality Preserving Projections (LPP). Our method's representation can outperform the use of consistent data alone and does not require the use of complex tensor-specific approaches. We run experiments on patient data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI), which consists of inconsistent data, to predict patients' diagnosis.
PDF Abstract