Alzheimer’s Disease Psychosis-related Brain Network: A Deep Learning and Explainable AI Framework
Poster No:
155
Submission Type:
Abstract Submission
Authors:
Nha Nguyen1, Jesus Gomar2, Jack Truong3, János Barbero4, Patrick Do5, Andrea Rommal6, David Eidelberg2, Jeremy Koppel2, An Vo2
Institutions:
1Albert Einstein College of Medicine, Bronx, NY, 2The Feinstein Institutes for Medical Research, Manhasset, NY, 3Adelphi University, Garden City, NY, 4Donald and Barbara Zucker School of Medicine at Hofstra/Northwell, Hempstead,, NY, 5University of Massachusetts Amherst, Amherst, MA, 6Donald and Barbara Zucker School of Medicine at Hofstra/Northwell, Hempstead, NY
First Author:
Co-Author(s):
Introduction:
Brain network analysis has been used successfully to identify and characterize network patterns in Alzheimer's disease (AD), Parkinson's disease (PD), and most recently, Dementia with Lewy Bodies (DLB) (6). Deep learning has shown efficacy in assessing AD dementia (7, 8). Evidence suggests that the emergence of psychosis in AD, manifested by delusional beliefs and/or hallucinatory experiences, represents a distinct pathophysiologic subtype with a unique clinical course distinguishable from non-psychotic AD (1, 5). In this study, we aimed to determine the existence of an AD psychosis network (ADPN), distinguishing those who develop psychosis during the course of AD from those who do not.
Methods:
We studied 88 AD patients (n=174 scans) who developed psychosis during the study (AD+P, 75.0 ± 7.5 years), 174 AD patients without psychosis (AD−P, 74.5 ± 8.6 years) and 174 cognitively normal (NC, 74.4 ± 5.8 years) participants with no signs of depression, mild cognitive impairment, or dementia. FDG PET scans were retrieved from the ADNI (https://adni.loni.usc.edu). To assess psychotic symptoms, the first 2 items (delusions and hallucinations) of the 12-item Neuropsychiatric Inventory were used, following consensus criteria for psychosis in dementia (1). FDG PET scans were registered to a standard Montreal Neurological Institute (MNI)-based PET template, smoothed with an isotropic Gaussian kernel (8 mm), and intensity normalized to the mean of the cerebellum using the FMRIB (http://www.fmrib.ox.ac.uk/fsl/). Initially, we employed a 3D deep residual neural network (3) to identify and validate the ADPN using Deep Learning Toolbox in Matlab 2023a. The classifier was trained using 142 AD+P and 142 NC scans and subsequently tested on a dataset consisting of 32 AD+P and 32 NC scans. We utilized an explainable deep learning technique (9) to compute expression scores used to construct an ADPN-based classifier for predicting psychosis in AD, incorporating a support vector machine (SVM). Classifier performance metrics of classifiers were compared to a conventional classifier, achieved through 95 FDG PET regions of interest based on the AAL atlas (10) and SVM. Significant difference between two groups was computed using Student's t test with Bonferroni correction.
Results:
The ADPN classifier exhibited higher accuracy (96.9%) compared to the conventional approach (92.2%) (Fig. 1). The ADPN was characterized by significant differences in activation maps between NC and AD+P in the frontal cortex, insula, amygdala, hippocampus, parahippocampal gyrus, cingulate cortex, parietal and temporal cortices. The ADPN exhibited significant differences in expression scores between AD+P and NC (P<0.05). Notably, the AD+P demonstrated regions (Fig. 1) with significantly elevated expression scores compared to the NC. In additional, the ADPN-based classifer achieved superior accuracy (77%) and sensitivity (86.2%) compared to the conventional classifier (68.4% and 67.8%) (Fig. 2) in distinguishing between AD+P and AD−P. When compared with the AD−P subjects, the AD+P exhibited key regions with significantly higher expression score (P<0.05) (Fig. 2). The regions specific to the difference between the two groups included the frontal cortex (superior and middle), the cingulate cortex (anterior and middle), the primary auditory cortex (Hesch's and superior temporal), the inferior parietal cortex, and the sensorimotor cortex (precentral, postcentral and SMA).
Conclusions:
The ADPN, identified by the 3D ResNet101, revealed significant differences in activation maps between the AD+P and both NC and AD−P. The findings are consistent with previous PET studies (2, 4) conducted in this disorder. Leveraging explainable AI enhanced our understanding and trust in the results generated by deep learning.
Disorders of the Nervous System:
Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s) 1
Modeling and Analysis Methods:
PET Modeling and Analysis 2
Keywords:
Machine Learning
Positron Emission Tomography (PET)
Other - explainable deep learning
1|2Indicates the priority used for review
Provide references using author date format
2. Gomar, JJ (2022), ‘Increased retention of tau PET ligand [(18)F]-AV1451 in Alzheimer's Disease Psychosis’, Translational Psychiatry, vol. 12, no. 1, pp. 82.
3. He, K (2016), ‘Deep residual learning for image recognition’, Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 770-778.
4. Koppel, J (2014), ‘Alzheimer's Disease Neuroimaging I. Psychosis in Alzheimer's disease is associated with frontal metabolic impairment and accelerated decline in working memory: findings from the Alzheimer's Disease Neuroimaging Initiative’, American Journal of Geriatric Psychiatry, vol. 22, no. 7, pp. 698-707.
5. Koppel , J (2014), ‘Optimal treatment of Alzheimer's disease psychosis: challenges and solutions’, Neuropsychiatric Disease and Treatment, vol. 10, no. , pp. 2253-62.
6. Perovnik, M (2023), ‘Functional brain networks in the evaluation of patients with neurodegenerative disorders’, Nature Reviews Neurology, vol. 19, no. 2, pp. 73-90.
7. Puente-Castro, A (2020), ‘Automatic assessment of Alzheimer’s disease diagnosis based on deep learning techniques’, Computers in Biology and Medicine, vol. 120, pp. 103764.
8. Qiu, S (2022), ‘Multimodal deep learning for Alzheimer’s disease dementia assessment’, Nature Communications, vol. 13, no. 1, pp. 3404.
9. Selvaraju, R.R., (2020), ‘Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization’, International Journal of Computer Vision, vol. 128, pp. 336–359.
10. Tzourio-Mazoyer, N. (2002), ‘Automated anatomical labeling of activations in SPM using a macroscopic anatomical parcellation of the MNI MRI single-subject brain’, NeuroImage, vol. 15, no. 1, pp. 273-289.