Functional gradient boosting predicts trajectories of cognitive impairment from brain morphometry
Poster No:
1374
Submission Type:
Abstract Submission
Authors:
Bruno Hebling Vieira1, Camille Elleaume2, Nicolas Langer1
Institutions:
1University of Zurich, Zurich, Switzerland, 2University of Zurich, Zurich, Zurich
First Author:
Co-Author(s):
Introduction:
The literature has shown that quantifiable MRI-markers of brain atrophy are highly predictive for future classification between cognitively normal (CN) individuals and Alzheimer's disease dementia (AD)¹ patients. This holds significant promise for prognostic applications, particularly in the recruitment phase of clinical trials targeting pre-clinical and subclinical stages of dementia².
However, cognitive impairment occurs along a continuum, not as discrete diagnostic categories. Our previous work has shown that structural MRI adds information to a model predicting the yearly rate of change in CDR-SOB and MMSE³. While approximating the trajectory of cognitive impairment as a linear decline is valid for small intervals, it breaks down at long intervals, since the actual evolution follows a non-linear trajectory⁴. Due to the sparsity of longitudinal data, accurately estimating non-linear trajectories for individual participants becomes a challenging task.
To test the effectiveness of different trajectory classes in the prediction of future cognitive impairment we present a method that can accommodate (non-)linear trajectories independent of the number of samples per participant in a cohort. Our approach involves estimating a versatile class of 'functions' aimed at best approximating the trajectories of participants' scores from features such as brain imaging features and risk factors. We employed a large-scale dataset (ADNI) to train three distinct models for predicting CDR-SOB trajectories using regional brain morphometry. We compared the performance of a linear functional, an inverse cloglog functional, and a direct regression that incorporates time as a feature.
However, cognitive impairment occurs along a continuum, not as discrete diagnostic categories. Our previous work has shown that structural MRI adds information to a model predicting the yearly rate of change in CDR-SOB and MMSE³. While approximating the trajectory of cognitive impairment as a linear decline is valid for small intervals, it breaks down at long intervals, since the actual evolution follows a non-linear trajectory⁴. Due to the sparsity of longitudinal data, accurately estimating non-linear trajectories for individual participants becomes a challenging task.
To test the effectiveness of different trajectory classes in the prediction of future cognitive impairment we present a method that can accommodate (non-)linear trajectories independent of the number of samples per participant in a cohort. Our approach involves estimating a versatile class of 'functions' aimed at best approximating the trajectories of participants' scores from features such as brain imaging features and risk factors. We employed a large-scale dataset (ADNI) to train three distinct models for predicting CDR-SOB trajectories using regional brain morphometry. We compared the performance of a linear functional, an inverse cloglog functional, and a direct regression that incorporates time as a feature.
Methods:
Features for modelling CDR-SOB trajectories were risk factors (i.e., age, sex, educational attainment, APOE2 and APOE4 allele counts), confounders (i.e., MRI magnetic field strength), covariates (i.e., the time between the input and output sessions in the direct regression model) and imaging derived phenotypes (IDPs). IDPs were obtained from Freesurfer, including cortical thickness and surface area from 148 regions in the Destrieux atlas, and volume from 46 non-neocortical regions. Data from ADNI (1, 2, GO, 3) were used for training and validation. We split ADNI into a training set comprising data of 80% of the subjects and a validation set containing the remaining 20% subjects. External validation relied on the independent dataset OASIS (3,4)⁵˒⁶ (see Figures 1.A-B).
We estimated a linear trajectory model, outputting two parameters only, and two non-linear trajectory models, including an inverse cloglog functional with a Beta likelihood, and a direct non-parametrized regression functional (see Figure 1.C-E).
We estimated a linear trajectory model, outputting two parameters only, and two non-linear trajectory models, including an inverse cloglog functional with a Beta likelihood, and a direct non-parametrized regression functional (see Figure 1.C-E).
Results:
A ll three models display a similar level of performance in ADNI and generalize to OASIS (see Figure 2.A-B,D). The models are overtly conservative, however, and tend towards flat predictions and underestimation of CDR-SOB in the case of AD due to regression-to-the-mean, which is visible in Figure 2.C. The MAE increases in longer time windows, with the best models reaching approximately 3 for AD patients, and 1 for CN individuals (see Figure 2.D).
Explicit parametrization of predictions can be used to estimate rates of decline, age of onset and phase transition, and integrate prior knowledge, e.g., the inverse cloglog function is bounded between 0 and 18, as is the CDR-SOB. Direct regression can also incorporate monotonicity constraints in features, but in our experiments, it did not result in better models. Future research will investigate other parametrizations as well as other likelihood functions.
Explicit parametrization of predictions can be used to estimate rates of decline, age of onset and phase transition, and integrate prior knowledge, e.g., the inverse cloglog function is bounded between 0 and 18, as is the CDR-SOB. Direct regression can also incorporate monotonicity constraints in features, but in our experiments, it did not result in better models. Future research will investigate other parametrizations as well as other likelihood functions.
·Figure 1
·Figure 2
Conclusions:
Our approach enables the prediction of (non-)linear trajectories of cognitive impairment, independent of the number of data points per subject. Non-linear trajectory models did not outperform the linear functional model in the prediction of CDR-SOB in the future from IDPs.
Disorders of the Nervous System:
Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s) 2
Lifespan Development:
Aging
Modeling and Analysis Methods:
Classification and Predictive Modeling 1
Keywords:
ADULTS
Aging
Cognition
Data analysis
Degenerative Disease
Machine Learning
Modeling
Morphometrics
MRI
STRUCTURAL MRI
1|2Indicates the priority used for review
Provide references using author date format
2. Tam, A. (2022), "Prediction of Cognitive Decline for Enrichment of Alzheimer’s Disease Clinical Trials". The Journal of Prevention of Alzheimer's Disease, 9, 400–409.
3. Vieira, B. H. (2022), "Predicting future cognitive decline from non-brain and multimodal brain imaging data in healthy and pathological aging". Neurobiology of Aging ,118, 55–65.
4. Delor, I. (2013), "Modeling Alzheimer’s Disease Progression Using Disease Onset Time and Disease Trajectory Concepts Applied to CDR-SOB Scores From ADNI". CPT: Pharmacometrics & Systems Pharmacology, 2, 78.
5. Koenig, L. N. (2020), "Select Atrophied Regions in Alzheimer disease (SARA): An improved volumetric model for identifying Alzheimer disease dementia". NeuroImage: Clinical, 26, 102248.
6. LaMontagne, P. J., (2019). "OASIS-3: Longitudinal Neuroimaging, Clinical, and Cognitive Dataset for Normal Aging and Alzheimer Disease". medRxiv preprint.