Change point detection of high-dimensional graphs for early MCI classification in fMRI

1959 

Abstract Submission 

Martin Ondrus¹, Ivor Cribben¹, Emily Olds¹

¹University of Alberta, Edmonton, Alberta

Martin Ondrus  
University of Alberta
Edmonton, Alberta

Ivor Cribben  
University of Alberta
Edmonton, Alberta

Emily Olds  
University of Alberta
Edmonton, Alberta

Mild Cognitive Impairment (MCI), is a collection of precursory stages to Alzheimer's Disease which are challenging to detect in a clinical setting [Hampel and Lista, 2016]. Functional connectivity (FC) is a method of quantifying the dependence [Biswal et al., 1995] between regions of interest (ROIs) and can be delineated into static (SFC) or dynamic (DFC), where distributions are assumed to be either constant or changing, respectively. By far the most common way of estimating DFC is through sliding windows [Shakil et al., 2016, Hindriks et al., 2016], but this technique has many shortcomings. Change point detection (CPD) is an alternative, and provides a data driven way of modelling changing FC patterns. In this work, we apply FaBiSearch [Ondrus et al., 2021] which is a novel CPD method of extracting high-dimensional, stable, DFC features for eMCI classification using non-negative matrix factorization [Ogawa et al., 1990] to model brain signals and clustering.

Data was obtained from the open source Alzheimer's Disease Neuroimaging Initative studies ADNI2 and ADNIGO [ADNI, 2023]. In these resting state functional magnetic resonance imaging (rs-fMRI) experiments, subjects were instructed to remain still and relaxed in the scanner. Subjects include 33 eMCI patients (mean age 72.3, 15M/18F) and 35 healthy controls (mean age 74.6, 14M/21F). Data were then pre-processed using a combination of SPM8 [SPM, 2023] and RESTplus [Jia et al., 2019] software packages.
For the static condition, the entire scanning session was used to estimated FC. For both wDFC and cpDFC, we describe the hyperparameters for each condition and the values we used in Figure 1. For each subject, we ordered change points based on the p-value obtained from the statistical test during the permutation procedure. Then, the first (cpDFC1) or first two (cpDFC2) change points were used and we estimated FC between each change point. A full summary of the methods is shown in Figure 1.
For each FC meaure above, we generated graph theoretic features (clustering coefficient, degree, degree assortativity, shortest path, local efficiency, and betweenness centrality). Sure independence screening (SIS) was used for feature selection [Saldana and Feng, 2018], and we used a linear support vector machine (SVM) as the classifier model. Training and testing was done using leave-one-out cross validation.

Change point detection with two change points (cpDFC2) has the best performance across almost all our four main measures of performance (accuracy; 76.47%, F1 score; 78.38%, sensitivity; 87.88%, and specificity; 65.71%). Compared to the best wDFC model (accuracy; 72.06% , F1 score; 74.67%, sensitivity; 84.85%, specificity; 60.00%), cpDFC2 only marginally outperformed it (p = 0.2496). However, compared to all wDFC models (accuracy; μ = 0.5852, σ = 0.0483), cpDFC2 had superior performance across all four main performance measures (accuracy; p = 3.82 × 10−6, F1 score; p = 1.98 × 10−5, sensitivity; p = 3.50 × 10−6, specificity; p = 1.13 × 10−6). SFC performed poorly, achieving similar to chance performance (51.47%). A comparison of these results is shown in Figure 2.

In this work, we compare the predictive ability and stability of SFC, wDFC, and cpDFC for an eMCI classification task using rs-fMRI. Firstly, we show that DFC outperforms SFC for classification of eMCI, which is a difficult task. We also show that multiple cpDFC is superior to wDFC based methods for disease classification. Further, we delineate several key advantages of cpDFC in comparison to wDFC that make it attractive as a way of capturing DFC beyond just the downstream classification task. Our results demonstrate the power of CPD models with FaBiSearch for a challenging eMCI/CN classification task, and suggest CPD is an important tool in the statistical tool box for analyzing fMRI data, and is valuable in uncovering hidden disease dynamics in rs-fMRI experiments.

Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s) ²

Classification and Predictive Modeling

Connectivity (eg. functional, effective, structural)

fMRI Connectivity and Network Modeling

Multivariate Approaches ¹

Computational Neuroscience

Data analysis

FUNCTIONAL MRI

Machine Learning

Modeling

Multivariate

Statistical Methods

Other - Time Series