Application of deep learning and novel survival models to predict MCI-to-AD dementia progression

项目总结/文摘

Project summary/Abstract Alzheimer's disease (AD) is a common and costly neurodegenerative disease that is characterized by a long pre-clinical stage, including a prodromal stage of AD also referred to as mild cognitive impairment (MCI). Many, but not all, MCI patients progress to AD dementia at varying rates. Among MCI patients, late stage MCI patients progress to AD faster than early stage MCI patients: a faster annual cognitive decline with loss of memory. As potential disease modifying drugs are tested for their ability to delay AD dementia, it becomes critical to have tools that can better accurately predict MCI-to-AD dementia conversion. This would allow selection of cohorts most likely to decline during the study period, maximizing the ability to detect a drug/placebo difference. The proposed project will respond to PA-20-200: NIH Small Research Grant Program (Parent R03 Clinical Trial Not Allowed). In Aim 1, we will develop new deep survival models to predict MCI-to- AD dementia conversion using baseline measures, by using data from the AD Neuroimaging Initiative (ADNI) study. We will use data from the NIH funded Center for Neurodegeneration and Translational Neuroscience (CNTN) as the test data. The majority of the existing deep survival models were developed for right censored data, but MCI-to-AD dementia conversion is interval censored. When interval censored data are analyzed by using the methods developed for right censored data, the survival rates are always over-estimated that leads to the delay in AD dementia diagnosis. We will develop separate prediction models for early stage MCI and late stage MCI with biomarkers from cerebrospinal fluid (CSF), positron emission tomography (PET), magnetic resonance imaging (MRI), and clinical measures. Recently, several new biomarkers have been discovered for AD that are of interest to this study. These include plasma phosphorylated-tau181 (p-tau181), p-tau217, and the ratio of amyloid-β 42 and amyloid-β 40, and glial fibrillary acidic protein (GFAP). In progressive disorders like AD, most clinical events are very strongly correlated with the dynamics of the disease. In Aim 2, we will develop novel survival models for interval-censored data with time-varying longitudinal biomarker data. Built on our developed penalized survival model for interval censored data using baseline measures, we propose to extend that model to leverage longitudinal biomarker data to produce more accurate predictions about future conversion. Biomarkers along with clinical and demographic features were shown to improve the model performances for right censored data. We expect that the new survival models will be able to improve model prediction for interval censored data as compared to state-of-the-art models. This project will develop optimal deep survival models to predict MCI-to-AD dementia conversion for each MCI subgroup. The results of this project will provide important understanding of how each feature contributes to prediction of MCI-to-AD dementia conversion.