Patterning of Dynamic Brain States Distinguish Parkinson’s Patients with Mild Cognitive Impairments

297 

Abstract Submission 

Abigail Eubank¹, Aaron Kemp¹, James Galvin², Fred Prior¹, Linda Larson-Prior¹

¹University of Arkansas for Medical Sciences, Little Rock, AR, ²University of Miami, Miami, FL

Abigail Eubank  
University of Arkansas for Medical Sciences
Little Rock, AR

Aaron Kemp  
University of Arkansas for Medical Sciences
Little Rock, AR

James Galvin  
University of Miami
Miami, FL

Fred Prior  
University of Arkansas for Medical Sciences
Little Rock, AR

Linda Larson-Prior  
University of Arkansas for Medical Sciences
Little Rock, AR

Parkinson's disease (PD) is commonly associated with motor impairments, however it also causes cognitive impairments which can be classified in a range of stages from PD with normal cognition (PD-NC), PD with mild cognitive impairment (PD-MCI) (Litvan et al. 2011; 2012), to PD with dementia (Saredakis et al. 2019). Previous studies have found dynamic, time-varying measures of functional network connectivity (dFNC) derived from resting-state functional magnetic resonance imaging (rs-fMRI), may distinguish patients with PD-NC from those with PD-MCI (Díez-Cirarda et al. 2018; Jinhee Kim et al. 2017). We explored the use of a dichotomic pattern mining technique, Seq2pat (Wang et al. 2022), to determine whether sequential patterns in the ordering of dFNC states could accurately distinguish healthy control (HC) subjects from those with PD and distinguish PD-NC from PD-MCI.

Seven minutes of resting, eyes-closed fMRI data were collected on a Siemens 3T TRIO scanner from 33 individuals with PD and 22 HC subjects at the Center for Biomedical Imaging at New York University following IRB approval and informed consent. Patients with PD were categorized into PD-MCI or PD-NC subgroups, by clinical consensus. The rs-fMRI data were preprocessed using the FMRIB Software Library (FSL; Jenkinson et al, 2012), including brain extraction, B0-unwarping, slice-time and motion correction, registration to standard space (MNI 152), and Independent Component Analyses for the Automatic Removal of Motion Artifacts (ICA-AROMA; Pruim et al, 2015). Three dFNC states were derived with the Group ICA for fMRI Toolbox (GIFT; Allen et al., 2014), as determined by the elbow criterion. Sequences of letters indicating the transition across states were then used as input to a Python-based sequential pattern mining method, Seq2pat (Wang et al, 2022). One-hot encoding of sequential patterns yielded feature vectors that were then used to train Random Forrest (RF) classifiers using 80% of the data for training and 20% for testing. Separate classifiers were trained to distinguish HC from PD groups, and PD-NC from PD-MCI subgroups, and the results for each were averaged across 100 iterations.

Statistical comparisons between the 3 state matrices returned from GIFT for each group (Figure 1), revealed a reduction in overall network inter-connectivity of the frontal parietal (FP) and executive control (EC) networks for both PD cohorts versus HC in State 1, while State 2 displays higher intra-connectivity within the default mode network (DMN) of the PD-MCI cohort relative to both the HC and PD-NC cohorts, and reduced intra-connectivity of the sensorimotor (SM) network in both PD-NC and PD-MCI groups relative to HC in State 3. The Seq2pat method identified 683 sequential patterns, including 227 that were unique to the HC subjects, 118 that were unique to the PD patients, and 338 that both groups displayed (Figure 2). The mean (standard deviation) values for accuracy and area under the receiver operating curve for the RF classifier trained to distinguish HC from PD was 0.71 (0.11) and 0.65 (0.12), respectively. For the RF classifier trained to distinguish PD-NC from PD-MCI, these values were 0.62 (0.15) and 0.65 (0.14), respectively. A multi-class classifier trained to distinguish all three groups yielded mean accuracy of 0.49 (0.12) and mean AUC of 0.52 (0.14).

We believe that this exploratory analysis highlights the potential utility of Seq2pat to detect sequential patterns in the temporal ordering of dynamic brain states, as detected using GIFT. While the results of the classifiers show only modest accuracy, the fact that they were all above chance level indicates that the sequential patterning of derived brain states warrants further investigation as possible indicators of abnormality in the dynamic, temporal organization of neural network activity among people with PD, particularly those with cognitive impairments.

Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s) ¹

Classification and Predictive Modeling

fMRI Connectivity and Network Modeling ²

BOLD fMRI

Aging

Cognition

Data analysis

Degenerative Disease

FUNCTIONAL MRI

Machine Learning

Movement Disorder