P3‐075: MODELING FUNCTIONAL CONNECTIVITY CHANGES IN LATE ONSET ALZHEIMER'S DISEASE USING DEEP LEARNING

Abstract

Alzheimer's disease (AD) is known to disrupt resting state functional connectivity (rs-fc) in various brain networks. Predicting functional changes could allow for targeted treatment and improved clinical outcomes for individuals with late onset Alzheimer's disease (LOAD). We propose an ensemble deep learning approach to predict rs-fc changes that occur due to AD. 34 AD participants (CDR > 0) and 33 cognitively normal healthy controls (CDR 0) with at least 2 separate imaging sessions from the Washington University Alzheimer's Disease Research Center (ADRC) were evaluated. Resting state fMRI over 91 aggregated regions of interest (ROIs) was determined for each participant at each time point. Imaging data was combined with genetic information (APOE e4), CDR status, and demographic variables and served as feature vectors for training (Table 1). A deep feed forward artificial neural network (ANN) was trained for each region using 80% of the data with the remaining 20% used for testing. The mean squared error (MSE) was calculated for each ROI in each test case matrix and averaged across all the test cases to evaluate the overall performance of the model. The average MSE for predicting changes in FC was 0.01, indicating the model performed well over all ROIs. Figure 1 plots the actual versus predicted values over all ROIs for the test data for a 4–5 year prediction (R2 = .75). Figure 2 depicts the error histogram for the test data for 1–5 year predictions. Most errors were clustered around zero, indicating little to no difference in the actual and predicted values.