An adaptive sliding-window based dynamic functional connectivity analysis for AD classification
Poster No:
2042
Submission Type:
Abstract Submission
Authors:
Pavithran Pattiam Giriprakash1, Filippo Cieri1, Xiaowei Zhuang1, Zhengshi Yang1, Dietmar Cordes1,2
Institutions:
1Cleveland Clinic Lou Ruvo Center for Brain Health, Las Vegas, NV, 2University of Colorado, Boulder, Boulder, CO
First Author:
Co-Author(s):
Dietmar Cordes
Cleveland Clinic Lou Ruvo Center for Brain Health|University of Colorado, Boulder
Las Vegas, NV|Boulder, CO
Cleveland Clinic Lou Ruvo Center for Brain Health|University of Colorado, Boulder
Las Vegas, NV|Boulder, CO
Introduction:
Dynamic functional connectivity (dFC) has been proven effective in quantifying the temporal dynamics of brain networks[1], for example, abnormal connectivity in the frontal and temporal cortices using resting-state fMRI (rs-fMRI) in Alzheimer's disease[2] (AD). However, the fixed window size used in the most widely implemented sliding window analysis for dFC estimation must be large enough for good low frequency resolution and small enough to capture the inherent dynamics[1]. We utilize an adaptive sliding window[3] derived from Empirical Mode Decomposition (EMD) that captures the local frequency characteristics. We also computed different dFC metrics for both adaptive and fixed window sizes and evaluated their ability to classify AD in a multiclass classification setting.
Methods:
rs-fMRI data was obtained from 53 cognitively normal (CN) (20 male;age:76.7± 6.2 years), 58 mild cognitively impaired (MCI) (31 male;age:76.1 ± 7.8 years) and 61 AD (33 male;age:77.1± 7.3 years) participants, all amyloid-β positive (standardized uptake value ratio[4], SUVR ≥ 1.1) from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database (adni.loni.usc.edu). rsfMRI acquisition parameters : TR/TE/resolution = 3000ms/30ms/3.3x3.3x3.3mm3, flip angle = 80, 48 slices and 140 timepoints. Data were minimally preprocessed and group Independent Component Analysis (ICA) based on FastICA[5] was implemented to obtain 100 independent components (ICs). Then 54 ICs spanning 9 major resting-state networks (auditory, cerebellum, cognitive control, default mode, frontal, subcortical, somatosensory, temporal, and visual) were shortlisted. An adaptive time dependent sliding window[6] based on the instantaneous period of the IMFs (frequency range between 0.01 Hz and the 0.16 Hz (Nyquist frequency)) from the Hilbert transform[7,8] was used. The Pearson's correlation matrix (54x54) calculated using this window was concatenated across time for each subject and then across all 3 groups. Dynamic brain states were then estimated by running a k-means clustering. Subsequently, we computed subject specific dFC metrics[9] and tested for group differences for EMD and fixed window sizes (10, 15, 20, 25 and 30 TR). We built a Randomforest classifier using these features (dFC metrics, dFC metrics + EMD metrics (log profiles of the instantaneous period and energy profiles)) to classify AD in a multiclass scenario with 3 classes (CN, MCI, and AD). Recursive feature elimination based on permutation feature importance was used for feature selection.
Results:
Three dynamic functional brain states were obtained from clustering for each of the window sizes (fixed and EMD) as shown in Figure 1A. Figure 1(B1-B3) shows the different dFC metrics obtained for the EMD approach. Only EMD showed significant group differences in the dFC metrics, specifically in the probability of transition from state 3 to state 1 (One way ANOVA : F statistic = 4.15, p-value = 0.02). No significant differences were observed in other metrics for all window sizes. Figure 2(A1) shows that the use of an adaptive window size results in better classification of AD (mean AUC-ROC= 0.65) compared to a constant window size. The significant feature identified in Figure 1(B1) had the highest feature importance in the classifier model, see Figure 2(A2). When dFC metrics were combined with EMD metrics, frontal and default mode networks showed the highest AUC-ROC (mean = 0.94) while somatosensory had the least value (mean = 0.73). The optimal feature set of this model consisted only of EMD metrics, primarily from IMF2 and IMF3 with peak frequencies at 0.025 Hz and 0.045 Hz, respectively.
Conclusions:
Our analysis shows that only a time-dependent window size based on the local frequency characteristics best captures the dynamics of the brain's intrinsic functions. Also, the estimated dFC metrics from this approach show better classifier performance for AD compared to fixed window sizes and the best performance when combined with the EMD-derived metrics.
Disorders of the Nervous System:
Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s)
Modeling and Analysis Methods:
Classification and Predictive Modeling
fMRI Connectivity and Network Modeling 2
Task-Independent and Resting-State Analysis 1
Novel Imaging Acquisition Methods:
BOLD fMRI
Keywords:
FUNCTIONAL MRI
Machine Learning
Multivariate
Other - Empirical Mode Decomposition (EMD)
1|2Indicates the priority used for review
Provide references using author date format
2. Gu, Yue, et al; and Alzheimer’s Disease Neuroimaging Initiative. 2020. “Abnormal Dynamic Functional Connectivity in Alzheimer’s Disease.” CNS Neuroscience & Therapeutics 26 (9): 962–71. https://doi.org/10.1111/cns.13387.
3. Zhuang, Xiaowei, et al.. 2020. “Single-Scale Time-Dependent Window-Sizes in Sliding-Window Dynamic Functional Connectivity Analysis: A Validation Study.” NeuroImage 220 (October): 117111. https://doi.org/10.1016/j.neuroimage.2020.117111.
4. Farrell, Michelle E., et al. 2021. “Defining the Lowest Threshold for Amyloid-PET to Predict Future Cognitive Decline and Amyloid Accumulation.” Neurology 96 (4): e619–31. https://doi.org/10.1212/wnl.0000000000011214.
5. Hyvärinen, Aapo. 1999. “Fast and Robust Fixed-Point Algorithms for Independent Component Analysis.” IEEE Transactions on Neural Networks 10 (3): 626–34. https://doi.org/10.1109/72.761722.
6. Cordes, Dietmar, et al. 2018. “Advances in Functional Magnetic Resonance Imaging Data Analysis Methods Using Empirical Mode Decomposition to Investigate Temporal Changes in Early Parkinson’s Disease.” Alzheimer’s & Dementia: Translational Research & Clinical Interventions 4 (1): 372–86. https://doi.org/10.1016/j.trci.2018.04.009.
7. Huang, Norden E., et al. 1998. “The Empirical Mode Decomposition and the Hilbert Spectrum for Nonlinear and Non-Stationary Time Series Analysis.” Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences 454 (1971): 903–95. https://doi.org/10.1098/rspa.1998.0193.
8. Flandrin, Patrick, Gabriel Rilling, and Paulo Gonçalvés. 2004. “Empirical Mode Decomposition as a Filter Bank.” IEEE Signal Processing Letters 11 (2): 112–14. https://doi.org/10.1109/lsp.2003.821662.
9. Damaraju, Eswar, et al. 2014. “Dynamic Functional Connectivity Analysis Reveals Transient States of Dysconnectivity in Schizophrenia.” NeuroImage: Clinical 5 (January): 298–308. https://doi.org/10.1016/j.nicl.2014.07.003.