Abstract:Detection of AD at the earlier symptomatic stage of mild cognitive impairment (MCI) is important for effective therapeutic intervention. Because MCI is a research construct that has heterogeneous etiologies, it is an important task to differentiate if an individual with MCI has early ongoing AD pathology (a.k.a. MCI due to AD). This can be accomplished by building a machine learning classifier to integrate images of different but complementary modalities together with clinical variables. The challenge is that multi-modality images are not universally available for all patients due to accessibility constraints (cost, insurance coverage, etc). In a previous study, we developed a transfer learning (TL) model capable of training a classifier using incomplete multi-modality images collected at baseline. In the present study, we include longitudinal multi-modality images in the TL model to further improve the classification accuracy. 214 MCI patients from ADNI were included. We downloaded MRI, FDG-PET, and amyloid-PET images (as three modalities) at baseline and a follow up. 97 patients were found to be Aβ+ at the follow up, an indicator of MCI due to AD. The patients were divided into four sub-cohorts: MRI only; MRI & FDG-PET; MRI & amyloid-PET; all three modalities available. We extracted commonly used imaging features including 3, 4, and 6 features from MRI, FDG-PET, and amyloid-PET, respectively. Clinical variables such as age, APOE 4 status, MMSE and CDR were also included. We applied TL to integrate the incomplete longitudinal multi-modality image features and clinical variables to build a classifier to differentiate MCI due to AD at the follow up. TL achieved 0.92 cross-validated AUC, 0.90 sensitivity, and 0.84 specificity. This is a significant improvement over using baseline images only (p<0.001). As a competing method, separate modeling of each sub-cohort by logistic regression achieved 0.86 AUC, 0.83 sensitivity, and 0.83 specificity, which were significantly outperformed by TL in AUC (p<0.001) and sensitivity (p<0.001). We proposed a novel TL model to accurately detect MCI due to AD for patients with varying availability of imaging modalities. This capability benefited a broader patient population especially those susceptible to resource or accessibility constraints.

ADDRESSING THE GAP OF IMAGING ACCESSIBILITY IN EARLY DETECTION OF AD BY A NOVEL TRANSFER LEARNING MODEL: A LONGITUDINAL STUDY