All participants were sorted into three groups after undergoing the Montreal Cognitive Assessment (MoCA) and the clinical dementia rating scale (CDR) with physician supervision. These groups consist of the NC group (n = 30), the VCIND group (n = 32), and the VD group (n = 21). For the Montreal Cognitive Assessment (MoCA) (Nasreddine et al., 2005), a cutoff value of < 26 indicates cognitive impairment. Additionally, one point is added to the raw MoCA score if the participant’s education level is < 12. All participants in the study were right-handed individuals of Han Chinese descent from northeastern China. Enrollment requirements for the NC group included: (1) absence of cognitive decline complaints; (2) a MoCA score of 26 or higher and a CDR score of 0; (3) no identification of symptoms or positive signs during the physical examination, and (4) no significant anomalies found in the head magnetic resonance imaging. Enrollment requirements for the VCIND group included (Sachdev et al., 2014): (1) a complaint or evidence of cognitive dysfunction from a knowledgeable source, with normal or slightly impaired ability to perform daily activities; (2) a MoCA score less than 26 and a CDR score equal to 0 or 0.5; and (3) an intracranial lesion visible on imaging, meeting VCI diagnostic criteria and clearly linked to cognitive decline. Enrollment requirements for the VD group included (Erkinjuntti, 1994): (1) significant cognitive impairment, reported by the patient or by an appropriate caregiver, affecting activities of daily living, (2) MoCA score < 22 and CDR score ≥ 1, and (3) cognitive impairment from VD confirmed by imaging evidence of intracranial pathology meeting diagnostic criteria and clearly associated with patient cognitive impairment. Exclusion criteria: (1) The participant has a history of heart or kidney disease, cancer, or other significant systemic illness. (2) The cognitive decline is unrelated to cerebrovascular disease. (3) The participant experiences progressive memory or cognitive decline without associated imaging changes. (4) The participant is unable to communicate due to severe impairment in hearing, vision, or speech; (5) The participant displays symptoms of depression and anxiety according to the Hamilton depression and anxiety inventory; (6) The participant has a history of mental illness or congenital developmental abnormalities; (7) The participant refuses or is unable to undergo brain MRI and EEG; (8) The participant experienced an acute cerebral infarction within the past three months.

All participants underwent a thorough neuropsychological evaluation, which assessed their verbal and visual situational memory, attention, executive function, visuospatial skills, and language proficiency. The evaluation utilized various specific tests, such as the Montreal Cognitive Assessment (MoCA), shape trails test (STT), shape trails test-A (STT-A), and shape trails test-B (STT-B). The STT consisted of two components: shape trails test-A (STT-A) and shape trails test-B (STT-B). All neuropsychological assessments were performed under the guidance of a specialist physician in neurology.

Electroencephalography (EEG) data were collected early in the morning, between 8:00 and 9:30 a.m., with patients awake, fasted, and in a quiet, closed-eye state. All EEG data were recorded on the same digital EEG system (NicoletOne™ EEG system, Natus Neurology Inc.), and a uniform signal acquisition standard was used to eliminate bias caused by different EEG equipment and parameters. Standardization of signal acquisition was used to eliminate bias caused by different EEG equipment and parameters. We collected data by placing a total of 16 electrodes (including Fp1, Fp2, F7, F8, F3, F4, C3, C4, T3, T4, T5, T6, P3, P4, O1, and O2) according to the international 10–20 standard lead system at a sampling rate of 250 Hz, with the input impedance set to Z > 100 MΩ, and collected the EEG signals for at least 30 min.

(1) We used the EEGLAB toolkit¹ based on matlab2019b² to localize electrodes, reject useless electrodes, and perform mean-based data re-referencing for all EEG data. (2) Select low frequency 1HZ high frequency 30HZ to filter and save the file; (3) Two or more EEG experts visually analyze the data and remove bad segments and artifacts; (4) Run the EEGLAB independent component analysis (ICA) toolbox to analyze the data for principal components and remove ICA-unusable components; (5) Extraction of EEG power spectra in each frequency band based on short-time Fourier transform. (6) Finally, 64 qEEG features (16 channels, 4 frequency bands) were extracted for each patient, and we performed multivariate analysis of covariance (MANCOVA) with age, sex, and education as covariates for the NC, VCIND, and VD groups, followed by multiple comparisons to control for error rates at the level of statistical significance (using Bonferroni-Holm correction), and after Bonferroni-Holm correction, post-hoc tests were performed and significance was determined at p < 0.05.

Magnetic resonance imaging (MRI) data acquisition for all subjects was performed on a Philips Achieva 3.0T MRI machine, using an 8-channel head coil, performing routine cranial transverse T1WI sequence scans. Scanning parameters: FOV = 256 mm × 256 mm², slice thickness = 1 mm, GAP = 0, number of slices = 192, TR/TE/TI = 7/3.2/1,100 ms, 7° flip angle, matrix = 256 × 256.

We used MRIcron, SPM12 and DPABI software package to analyze the NMR data and calculate the gray matter volume of the whole brain voxel. All of the above were run on MatLab (R2019b). The main steps were as follows: (1) MRIcron software was used to convert the MRI data DICOM files of all participants into NIfTI files; (2) the NIfTI files were imported into CAT12 in SPM12 for segmentation; (3) quality checks were performed, and the segmented gray matter image was smoothed; (4) the smoothed data were imported into DPABI for statistical analysis and image presentation (all gray matter structures were partitioned using Anatomical Automatic Labeling). (5) Using DPABI, the significant brain regions obtained from the Voxel-based morphometry (VBM) analysis were set as regions of interest (ROI), and the gray matter volumes of the ROI were obtained.

In this study, a support vector machine (SVM) model is constructed and the algorithm consists of two main steps: training of the SVM classifier and evaluation of the model. We divided the 83 samples into a training set and a test set at a ratio of 8:2 to ensure the generalization performance of the model. In the model of the NC group with the VCIND group, 24 NC participants and 26 patients of the VCIND group were randomly selected as the training set to build the SVM model, and the remaining 12 were used as the test set; in the NC group and VD group model, 24 NC participants and 17 VD group patients were selected at random as the training set to build the SVM model, and the remaining 10 were used as the test set; and in the VCIND group and VD group model, 26 VCIND patients and 17 VD patients were selected at random as the training set, and the remaining 10 patients were used as the test set. In this study, we performed quantitative electroencephalogram (qEEG) power spectrum analysis and VBM analysis on the test sets of the NC group vs. VCIND group, the NC group vs. VD group, and the VCIND group vs. VD group, and we selected statistically significant (p < 0.05) data obtained from two-way comparisons as categorical features (qEEG power spectra and volume of brain area corresponding to gray matter atrophy).

Then the LibSVM toolbox³ in MATLAB is used for support vector machine classification, the model has two key parameters: the kernel function and the regularization parameter, to optimize the model’s performance, we chose the radial basis function as the kernel function and used the grid search method in quintuple cross-validation to determine the regularization parameter. We used the model on the training set to predict the diagnostic results on the test set and evaluated the predictive ability of the model using the receiver operating characteristic (ROC) curve and the area under the ROC curve (AUC-ROC). This approach helps to prevent overfitting of the model on the training set and thus allows for a more accurate assessment of the generalization ability of the model.

SPSS version 21 was used to analyze all clinical and demographic data between groups. Count data are presented as case numbers (proportion), and Fisher’s exact test was used to analyze participant demographics. Normally distributed data are presented as mean ± standard deviation, whereas non-normally distributed data are presented as M (Q1, Q3). Normally distributed data were analyzed by one-way analysis of variance (ANOVA), and the least significant difference was used for post-hoc testing. Non-normally distributed data were tested using the Kruskal-Wallis H test, an independent-samples non-parametric test (p < 0.05 was regarded as statistically significant). All MRI image data were analyzed using DPABI. Age, sex, education level, and total brain volume were used as covariates, and the permutation test was used to correct for multiple comparisons to analyze gray matter atrophy changes in participants.

The study included 83 subjects, and Table 1 displays their demographic characteristics. The three patient groups differed significantly in age, education level, history of hypertension, and level of cognitive impairment (P < 0.05).

To reduce confounding, we used patient age, sex, education, and intracranial volume as covariates and corrected for multiple comparisons using the permutation test to analyze changes in gray matter atrophy between participants. In our study, VBM analysis showed significant differences only in gray matter in the Putamen_L, Caudate_L, and Thalamus_R regions when comparing the NC group with the VCIND group (as shown in Figure 2 and Table 2), but differences in more extensive gray matter atrophy were seen when comparing the VCIND group with the VD group and the NC group with the VD group. In the comparison between the VCIND group and the VD group, it showed atrophy in 18 relevant brain regions including Fusiform_L, Cerebelum_6_L, Cerebelum_4_5_R, Fusiform_R, Lingual_R, Cerebelum_4_5_L, Cerebelum_6_R, and so on (as shown in Figure 3 and Table 2); while the comparison between NC and VD groups showed atrophy of 25 brain regions including Thalamus_L, Thalamus_R, Fusiform_L, Olfactory_L, Cerebelum_6_L, Cerebelum_4_5_R, Fusiform_R, and Lingual_R (as shown in Figure 4 and Table 2).

In our experiment, the VCIND group exhibited an increase in theta power in the O2 region and a decrease in beta power in the O1-O2 region compared to the NC group. Three EEG power spectra were statistically significant: the T3-O1 (Alpha1/Alpha2), O2 (Beta1), and O2 (Theta) waves. As for the comparison between the VD and VCIND groups, in addition to the power changes in the O1-O2 region, it also showed an increase in theta in the frontal and parietal lobes and a higher delta power in the F3-F4 region, where the theta difference was most pronounced, and there were a total of four power spectral features with statistically significant differences, namely F4 (delta), O1-O2 (theta), T4 (beta2), and O2 (theta/gamma);The comparison of the NC and VD groups revealed significant differences in six power spectra: P3 (Theta), F3 (Delta), O1 (Alpha2), P4 (Theta/gamma), O2 (Alpha1/Alpha2), and O1-O2 (Beta1). Refer to Figure 5 and Table 2 for more details.

Based on the results of the VBM analysis, we labeled voxels suggesting statistical significance, calculated the volume of brain regions in their areas, and analyzed them (see Table 2 for details). We believe that since some features are useless, irrelevant or redundant for classification, too many features can lead to “overfitting,” so eliminating useless features and selecting meaningful ones not only simplifies the classification model, but also improves the classification accuracy. In the training set, we selected 3 power spectral features and 2 sMRI features that were statistically significant in the NC group compared with the VCIND group (see Table 2 for details, the upper right corner is labeled as a); 4 power spectral features and 11 sMRI features that were statistically significant in the VCIND group compared with the VD group (see Table 2 for details, the upper right corner is labeled as b); 6 power spectral features and 16 sMRI features that were statistically significant in the NC group compared with the VD group (see Table 2 for details, the upper right corner is labeled as c).

Our study shows that sMRI has better classification ability than qEEG in distinguishing VD from cognitively normal people. The area under the ROC curve of the sMRI-based support vector machine learning model is AUC = 0.90, and the area under the ROC curve of the machine learning classification model based on the qEEG features is AUC = 0.82. The “composite marker” model that combines sMRI and qEEG achieves the best classification results, with an area under the ROC curve of AUC = 0.98 (Table 3 and Figures 6A–C). When distinguishing between VD and VCIND populations, sMRI demonstrated better classification ability than qEEG. The machine learning model based on sMRI had an area under the ROC curve of AUC = 0.80, while the machine learning classification model based on qEEG features had an AUC of only 0.64. The composite marker model, which combined sMRI and qEEG, achieved optimal classification results with an ROC curve AUC of 0.92 (Table 3 and Figures 6D–F). When using only sMRI or qEEG features to differentiate between VCIND patients and NC, both methods had poor classification ability, with an area under the ROC curve of 0.56 for sMRI features and 0.54 for qEEG features. However, the “composite marker” model, which combines both sMRI and qEEG features, achieved relatively good classification ability with an area under the ROC curve of 0.72 (Table 3 and Figures 6G–I).