Identifying healthy individuals with Alzheimer’s disease neuroimaging phenotypes
Poster No:
164
Submission Type:
Abstract Submission
Authors:
Tiago Azevedo1, Richard Bethlehem2, David Whiteside3, Nol Swaddiwudhipong3, James Rowe3, Pietro Lio4, Timothy Rittman3
Institutions:
1Department of Computer Science and Technology, University of Cambridge, Cambridge, United Kingdom, 2Autism Research Centre, Department of Psychiatry, University of Cambridge, Cambridge, United Kingdom, 3Department of Clinical Neurosciences and Cambridge University Hospitals NHS Trust, Cambridge, United Kingdom, 4Department of Computer Science and Technology, University of Cambridge, Cambridge, United Kingdom, Cambridge, United Kingdom
First Author:
Tiago Azevedo
Department of Computer Science and Technology, University of Cambridge
Cambridge, United Kingdom
Department of Computer Science and Technology, University of Cambridge
Cambridge, United Kingdom
Co-Author(s):
Richard Bethlehem
Autism Research Centre, Department of Psychiatry, University of Cambridge
Cambridge, United Kingdom
Autism Research Centre, Department of Psychiatry, University of Cambridge
Cambridge, United Kingdom
David Whiteside
Department of Clinical Neurosciences and Cambridge University Hospitals NHS Trust
Cambridge, United Kingdom
Department of Clinical Neurosciences and Cambridge University Hospitals NHS Trust
Cambridge, United Kingdom
Nol Swaddiwudhipong
Department of Clinical Neurosciences and Cambridge University Hospitals NHS Trust
Cambridge, United Kingdom
Department of Clinical Neurosciences and Cambridge University Hospitals NHS Trust
Cambridge, United Kingdom
James Rowe
Department of Clinical Neurosciences and Cambridge University Hospitals NHS Trust
Cambridge, United Kingdom
Department of Clinical Neurosciences and Cambridge University Hospitals NHS Trust
Cambridge, United Kingdom
Pietro Lio
Department of Computer Science and Technology, University of Cambridge, Cambridge, United Kingdom
Cambridge, United Kingdom
Department of Computer Science and Technology, University of Cambridge, Cambridge, United Kingdom
Cambridge, United Kingdom
Timothy Rittman
Department of Clinical Neurosciences and Cambridge University Hospitals NHS Trust
Cambridge, United Kingdom
Department of Clinical Neurosciences and Cambridge University Hospitals NHS Trust
Cambridge, United Kingdom
Introduction:
Challenges in timely and accurate Alzheimer's disease (AD) diagnosis have an impact on patients and impede clinical trial success. Noteworthy findings in previous cohorts suggest changes in structural neuroimaging decades before symptoms. Deep learning has been proved to be effective in addressing diverse neuroscientific challenges; however, despite some recent success in AD's classification, few studies have validated results in completely independent datasets. In this work, we identify a group of people at high risk of developing dementia in the healthy UK Biobank dataset, using Bayesian deep learning modelling techniques. We demonstrate that our approach can be applied to identify people at high risk of developing dementia.
Methods:
We preprocessed structural MRI (MPRAGE) scans across three cohorts. Firstly, from ADNI, we used baseline scan sessions with a diagnosis of AD (n=331) and Controls (n=405). Secondly, we used the NACC dataset for validation of which we used a total of 1706 scans from patients with an AD diagnosis, as well as controls (n=2824), and other disorders (n=679). Finally, we used 37104 scans from the UK Biobank.
We extracted regional cortical thicknesses and cortical volumes from 68 surface-based regions, as well as brainstem volume, and 9 other volume features per hemisphere, totalling 155 features per brain scan. To avoid data leakage, 155 deconfounding regression models (using age, estimated intracranial volume, and sex) were fitted only on the ADNI training set, and the corresponding learned statistics were later employed to deconfound all the other datasets.
As illustrated in Figure 1, we implemented a neural network with two hidden layers with empirically found hyperparameters giving stable learning curves. We used Monte Carlo dropout to approximate Bayesian inference by activating the dropout layers at inference time. Gal and Ghahramani showed that, after training, each forward pass in the network corresponds to a good approximation to sampling from the true posterior distribution, with very little added computational cost.
We extracted regional cortical thicknesses and cortical volumes from 68 surface-based regions, as well as brainstem volume, and 9 other volume features per hemisphere, totalling 155 features per brain scan. To avoid data leakage, 155 deconfounding regression models (using age, estimated intracranial volume, and sex) were fitted only on the ADNI training set, and the corresponding learned statistics were later employed to deconfound all the other datasets.
As illustrated in Figure 1, we implemented a neural network with two hidden layers with empirically found hyperparameters giving stable learning curves. We used Monte Carlo dropout to approximate Bayesian inference by activating the dropout layers at inference time. Gal and Ghahramani showed that, after training, each forward pass in the network corresponds to a good approximation to sampling from the true posterior distribution, with very little added computational cost.
Results:
Table 1 shows the performance of our model. We compared our model with a simple hippocampal volume measure, other machine learning models as well as to TPOT, a tool which tries to find the best possible pipeline for a specific dataset.
The highest performance was found in the ADNI test set, which is expected; however, performance in NACC was still comparable to previous literature. The NPV performance of our model was the best; this is important given our application in the UK Biobank, where the rate of AD is substantially lower than either ADNI or NACC and we want to avoid the risk of misclassifying healthy people. Our model was able to give a better balance of sensitivity/specificity; other models had higher specificity, but with low sensitivity.
We applied Bayesian regression models with age as a covariate to assess the clinical validity of the AD score. For the UK Biobank, there were 3.4% of people with a positive AD score and the group with a positive AD score was only slightly older than the AD score negative group (1.79 years). There was strong evidence of worse fluid intelligence in the AD score positive group (−0.35), and strong evidence that people in the AD group were more likely to report their overall health as 'poor' or 'fair' rather than 'good' or 'excellent'. There was some evidence of a difference in both diastolic blood pressure (1.12) and systolic blood pressure (2.29).
The highest performance was found in the ADNI test set, which is expected; however, performance in NACC was still comparable to previous literature. The NPV performance of our model was the best; this is important given our application in the UK Biobank, where the rate of AD is substantially lower than either ADNI or NACC and we want to avoid the risk of misclassifying healthy people. Our model was able to give a better balance of sensitivity/specificity; other models had higher specificity, but with low sensitivity.
We applied Bayesian regression models with age as a covariate to assess the clinical validity of the AD score. For the UK Biobank, there were 3.4% of people with a positive AD score and the group with a positive AD score was only slightly older than the AD score negative group (1.79 years). There was strong evidence of worse fluid intelligence in the AD score positive group (−0.35), and strong evidence that people in the AD group were more likely to report their overall health as 'poor' or 'fair' rather than 'good' or 'excellent'. There was some evidence of a difference in both diastolic blood pressure (1.12) and systolic blood pressure (2.29).
Conclusions:
By using a Bayesian deep learning network, we have identified an AD-like cohort in the UK Biobank, with a diagnosis of dementia or reported symptoms. For disease prevention, our results highlight smoking history, greater pack-year exposure, and hypertension as potentially modifiable risk factors. Bayesian statistics for group comparisons and regression models offered advantages in our work as they ultimately allow us to distinguish a small effect size from an incorrect parameter estimate.
Disorders of the Nervous System:
Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s) 1
Modeling and Analysis Methods:
Classification and Predictive Modeling 2
Keywords:
Data analysis
Degenerative Disease
Informatics
Machine Learning
Modeling
Open-Source Code
STRUCTURAL MRI
1|2Indicates the priority used for review
Provide references using author date format
Gal, Y. & Ghahramani, Z. Dropout as a Bayesian approximation: Representing model uncertainty in deep learning. In Balcan, M. F. & Weinberger, K. Q. (eds) Proc. 33rd International Conference on Machine Learning Research (PMLR, 2016), Vol. 48, 1050–1059. (ML Research Press, 2016).
Kinnunen, K. M. et al. Presymptomatic atrophy in autosomal dominant Alzheimer’s disease: a serial magnetic resonance imaging study. Alzheimer’s & Dementia 14, 43–53 (2018).
Abrol, A. et al. Deep learning encodes robust discriminative neuroimaging representations to outperform standard machine learning. Nature Communications 12, 353–380 (2021).