Identifying Structural and Functional Alteration Through Harmonization of Multisite Gray Matter and EEG Functional Dynamics Using Cross-Modal Analysis and Sparse ICA
DOI: Pending DOI
Synopsis
Alzheimer’s disease (AD) is associated with progressive structural degeneration and functional brain abnormalities. Extensive research has been carried out to identify neurobiological changes associated with AD using diverse neuroimaging modalities. Most of them relied on single-modality analysis and conventional ICA or predefined region-based approaches, which may fail to capture sparse, distributed, and biologically meaningful latent disease patterns. This study proposes a harmonized data-driven framework using multisite gray matter MRI and resting-state EEG data. MRI voxel features and EEG spectral features were independently harmonized and analyzed using Sparse ICA. Statistical analysis identified significant structural alterations in cerebellar, hippocampal, temporal, parietal, frontal, and subcortical regions together with frontal low-frequency EEG abnormalities associated with AD.
Background
Alzheimer’s disease (AD) is the most common neurodegenerative disorder and the leading cause of dementia worldwide1. Structural MRI studies have reported progressive gray matter atrophy involving hippocampal, entorhinal, temporal, parietal, and frontal regions2,3,4, while EEG studies demonstrated diffuse slowing, increased delta/theta activity, reduced alpha activity, and disrupted synchronization patterns associated with AD1,4. Previous studies also suggested that gray matter co-alteration patterns resemble intrinsic brain connectivity architectures, indicating that AD progression involves distributed neural systems rather than isolated regions5. Although machine learning and independent component analysis (ICA-based) approaches have shown promising performance for AD analysis6, most previous studies relied on single-modality analysis, conventional ICA, ROI-based methods, predefined connectivity measures, or single-network representations7. These approaches may not fully capture sparse and distributed latent structural and functional alterations underlying AD progression.
Recent studies emphasized the importance of statistical harmonization and domain adaptation techniques for reducing multisite variability and improving model generalizability across heterogeneous neuroimaging datasets8. Therefore, this study was performed to investigate latent disease-related structural and electrophysiological alterations using harmonized multisite gray matter MRI and resting-state EEG data. Sparse ICA was independently applied to MRI and EEG modalities after modality-specific harmonization and dimensionality reduction to identify hidden structural and functional brain patterns associated with AD.
Methods
Structural MRI and resting-state EEG data from the BrainLat dataset10 were analyzed independently using a harmonized data-driven framework. Gray matter maps were extracted from preprocessed T1-weighted MRI images, while EEG spectral power features were computed from 19 channels across delta, theta, alpha, and beta bands. To reduce multisite and scanner-related variability while preserving biologically meaningful information, a parametric empirical Bayes harmonization framework was applied independently to gray matter voxel features and EEG spectral features. Principal Component Analysis (PCA) and whitening were applied for dimensionality reduction and decomposition stability. Sparse Independent Component Analysis (Sparse ICA) was independently applied to MRI and EEG data to identify latent structural and functional brain patterns associated with Alzheimer’s disease. Subject-specific loading parameters were used for group-level statistical analysis between healthy controls (HC) and AD subjects using two-sample statistical tests with false discovery rate correction. Because MRI and EEG subjects were not identical across modalities, direct subject-wise fusion was not performed. Instead, cross-modal associations were investigated using normalized component-level statistical signatures derived independently from MRI and EEG components. The complete processing pipeline included MRI preprocessing and gray matter (GM) extraction, EEG preprocessing and spectral feature estimation, multi-site harmonization, PCA-based dimensionality reduction, Sparse ICA decomposition, statistical group comparison, and cross-modal statistical association analysis.
Results
After preprocessing and multi-metric quality control using SPM25, EEGLAB, and MNE-Python pipelines, the final dataset consisted of 100 healthy controls (HC) and 100 Alzheimer’s disease (AD) subjects for structural MRI analysis, and 32 HC and 32 AD subjects for EEG analysis. Structural MRI preprocessing included tissue segmentation, DARTEL normalization, modulation, smoothing, and gray matter (GM) extraction, while EEG preprocessing included filtering, artifact removal using ICA and ICLabel, bad channel correction, and spectral feature computation across delta, theta, alpha, and beta bands.
Following multisite harmonization, PCA-based dimensionality reduction, whitening, and Sparse ICA decomposition, ten independent components were estimated separately for GM and EEG modalities. Group differences between HC and AD subjects were evaluated using two-sample statistical tests with false discovery rate (FDR) correction. The statistical threshold was set at corrected p < 0.0001 for GM and corrected p < 0.05 for EEG. Figures 1(a–b) illustrate the GM component loading differences and signed statistical significance scores between HC and AD groups, demonstrating substantial group separation across several components.
The GM analysis revealed strong disease-related structural alterations, where nine components (C1 and C3–C10) survived FDR correction. Components C1 and C8 demonstrated the strongest statistical significance scores, followed by C10, C3, C9, C4, C5, C6, and C7. Sparse ICA identified several biologically meaningful latent structural patterns associated with AD. Components C1, C8, C9, and C10 were primarily associated with cerebellar regions, including Cerebellum VI and Crus I/II subdivisions. Although partially overlapping, Sparse ICA separated these cerebellar regions into distinct localized latent components, suggesting multiple independent cerebellar degeneration patterns in AD. Components C5 and C6 revealed limbic and subcortical alterations involving the hippocampus, parahippocampal gyri, insula, caudate nuclei, thalamus, and frontal-parietal regions. Both components demonstrated stronger structural expression in HC subjects than AD subjects (HC > AD). Components C3 and C7 demonstrated temporal-parietal alterations involving inferior temporal, middle temporal, supramarginal, precuneus, and parietal cortices. Component C4 involved occipital and visual-association regions including calcarine cortex and cuneus. Most significant GM components showed reduced structural expression in AD subjects relative to HC subjects.
The EEG analysis revealed five significant functional components (C1, C3, C4, C5, and C8) after FDR correction. The largest EEG statistical significance score was observed in Component C1, followed by Components C8 and C3. EEG spectral topographic analysis (Fig. 1) demonstrated highly consistent low-frequency spatial patterns across significant components. Delta-band alterations were particularly prominent within frontal and frontotemporal regions. Frontal pole and lateral frontal channels, including Fp1, Fp2, F7, F4, and F8, consistently demonstrated opposite contribution patterns relative to frontal midline regions such as F3 and Fz. Similar spectral-topographic organizations across multiple EEG components suggest reproducible low-frequency electrophysiological abnormalities associated with AD. Compared with HC subjects, AD subjects demonstrated reduced expression of these frontal delta-band oscillatory patterns.
Cross-modal statistical profile comparisons between MRI and EEG components demonstrated partially similar structural-functional alteration trajectories across several components, suggesting potential relationships between latent gray matter degeneration patterns and electrophysiological dysfunction associated with Alzheimer’s disease.
Discussion
The proposed framework differentiated AD and healthy control (HC) groups and revealed disease-related patterns involving cerebellar, temporal, hippocampal, parietal, frontal, and subcortical regions, supporting the usefulness of harmonized data-driven decomposition methods for investigating distributed neurodegenerative processes in AD.
One important finding was the separation of overlapping cerebellar regions into multiple spatially localized latent components. Components C1, C8, C9, and C10 demonstrated partially overlapping but distinct Cerebellum VI and Crus I/II organizations, suggesting multiple partially independent cerebellar degeneration processes in AD. These findings are consistent with our previous studies11 and further demonstrate the ability of Sparse ICA to separate overlapping structural patterns more effectively than conventional ICA methods.
The identified temporal, hippocampal, parietal, and precuneus alterations are highly consistent with known AD pathology. EEG analysis additionally revealed reproducible low-frequency frontal and frontotemporal abnormalities, particularly involving delta-band oscillatory dynamics, consistent with previous EEG studies4,12,13.
From a clinical and research perspective, the proposed framework may support future biomarker discovery, disease monitoring, and large-scale multisite neuroimaging analysis. Future studies with larger datasets, longitudinal analysis, and multimodal shared-weight matrix computation may further improve robustness and clinical applicability for Alzheimer’s disease assessment.
Disclosures
This research was supported by the CSTEM Mini-grant Program at Clayton State University, Morrow, GA. The authors report no commercial financial relationships or conflicts of interest related to this work.
References
[1] Zheng X, Wang B, Liu H, Wu W, Sun J, Fang W, Jiang R, Hu Y, Jin C, Wei X and Chen SS-C (2023) Diagnosis of Alzheimer’s disease via resting-state EEG: integration of spectrum, complexity, and synchronization signal features. Front. Aging Neurosci. 15:1288295. doi: 10.3389/fnagi.2023.1288295
[2] Karas GB, Scheltens P, Rombouts SA, Visser PJ, van Schijndel RA, Fox NC, Barkhof F. Global and local gray matter loss in mild cognitive impairment and Alzheimer's disease. Neuroimage. 2004 Oct;23(2):708-16. doi: 10.1016/j.neuroimage.2004.07.006. PMID: 15488420.
[3] Möller C, Vrenken H, Jiskoot L, Versteeg A, Barkhof F, Scheltens P, van der Flier WM. Different patterns of gray matter atrophy in early- and late-onset Alzheimer's disease. Neurobiol Aging. 2013 Aug;34(8):2014-22. doi: 10.1016/j.neurobiolaging.2013.02.013. Epub 2013 Apr 3. PMID: 23561509.
[4] Aoki, Y., Takahashi, R., Suzuki, Y. et al. EEG resting-state networks in Alzheimer’s disease associated with clinical symptoms. Sci Rep 13, 3964 (2023). https://doi.org/10.1038/s41598-023-30075-3.
[5] Cauda F, Nani A, Manuello J, Premi E, Palermo S, Tatu K, Duca S, Fox PT, Costa T. Brain structural alterations are distributed following functional, anatomic and genetic connectivity. Brain. 2018 Nov 1;141(11):3211-3232. doi: 10.1093/brain/awy252. PMID: 30346490; PMCID: PMC6202577.
[6] Liu, J., Wu, F., and Wang, J. (2020). Enhancing the feature representation of multi-modal MRI data by combining multi-view information for MCI classification. Neurocomputing 400, 322–332. doi: 10.1016/j.neucom.2020.03.006.
[7] Wu Z, Peng Y, Hong M and Zhang Y (2021) Gray Matter Deterioration Pattern During Alzheimer's Disease Progression: A Regions-of-Interest Based Surface Morphometry Study. Front. Aging Neurosci. 13:593898. doi: 10.3389/fnagi.2021.593898
[8] Mellot, Apolline & Collas, Antoine & Rodrigues, Pedro & Engemann, Denis & Gramfort, Alexandre. (2023). Harmonizing and aligning M/EEG datasets with covariance-based techniques to enhance predictive regression modeling. Imaging Neuroscience. 1. 1-23. 10.1162/imag_a_00040.
[9] Johnson WE, Li C, Rabinovic A. Adjusting batch effects in microarray expression data using empirical Bayes methods. Biostatistics. 2007 Jan;8(1):118-27. doi: 10.1093/biostatistics/kxj037. Epub 2006 Apr 21. PMID: 16632515.
[10] Prado, P., Medel, V., Gonzalez-Gomez, R. et al. The BrainLat project, a multimodal neuroimaging dataset of neurodegeneration from underrepresented backgrounds. Sci Data 10, 889 (2023). https://doi.org/10.1038/s41597-023-02806-8
[11] Khalilullah KMI, Agcaoglu O, Sui J, Adali T, Duda M, Calhoun VD. Multimodal fusion of multiple rest fMRI networks and MRI gray matter via parallel multilink joint ICA reveals highly significant function/structure coupling in Alzheimer's disease. Hum Brain Mapp. 2023 Oct 15;44(15):5167-5179. doi: 10.1002/hbm.26456. Epub 2023 Aug 22. PMID: 37605825; PMCID: PMC10502647.
[12] Paitel ER, Otteman CBD, Polking MC, Licht HJ and Nielson KA (2025) Functional and effective EEG connectivity patterns in Alzheimer’s disease and mild cognitive impairment: a systematic review. Front. Aging Neurosci. 17:1496235. doi: 10.3389/fnagi.2025.1496235
[13] Iqbal, S., Nisar, H. & Yeap, K.H. Quantitative EEG signatures of power and functional connectivity alterations in Alzheimer’s disease and frontotemporal dementia. Sci Rep 16, 12158 (2026). https://doi.org/10.1038/s41598-026-42452-9