P4‐593: Early Prediction of Cognitive Decline Based on Brain Mri Images Using a Deep Learning Survival Analysis Model
Year: 2019
Type: article
Abstract: Early recognition of cognitive decline could promote timely interventions to slow or halt dementia progression. To achieve this goal, we develop an end-to-end deep learning method to extract informative features from brain MRI images and build a prognostic model to predict rapidity of cognitive decline of individual elderly persons, including cognitively normal (CN) and mild cognitive impairment (MCI) subjects. This study used data from Alzheimer's Disease Neuroimaging Initiative (ADNI). Particularly, CN and MCI subjects of ADNI-1 were used as training data and subjects of ADNI-GO&2 were used as testing data (Table 1). A gray matter (GM) tissue density map of baseline MRI scan was computed for each subject, and was used as input in conjunction with baseline demographic and cognitive measures to build a prediction model of rapidity of cognitive decline in a survival (time-to-event) analysis framework (Fig. 1). The event of interest was defined as the increase of 1 unit in CDR-SOB score from baseline. The prediction model was built using deep convolutional neural works (CNNs) with residual connections and deep supervision, and informative brain regions were visualized using class activation mapping. The deep learning model was compared with Cox regression models built on GM volume measures of brain regions defined by Automated Anatomical Labeling (AAL) template and demographic and cognitive measures. The prediction performance was evaluated based on concordance-index (C-index) values. The deep learning model obtained C-index values of 0.667 and 0.748 on the testing CN cohort and MCI cohort respectively, significantly better than those built upon regional GM volume measures, demographic and cognitive measures, and their combination (Table 2). The deep learning prediction model also identified informative brain regions, mainly located in the lateral temporal, temporoparietal, and prefrontal cortices (Fig. 2). The predicted risk of cognitive decline obtained by the deep learning model also clustered the CN and MCI subjects into subgroups with significant differences in their rapidity of cognitive decline (Fig. 3).
Source: Alzheimer s & Dementia
Authors: Hongming Li, David A. Wolk, Yong Fan
Institution: University of Pennsylvania
Language: English
Subfield: Psychiatry and Mental health
Field: Medicine
Domain: Health Sciences
Open Access status: bronze
OpenAlex ID: W2980965489