Machine learning: assessing neurovascular signals in the prefrontal cortex with non-invasive bimodal electro-optical neuroimaging in opiate addiction

Figure 1. Study design. (A) Flow diagram for the study. (B) Workflow of study design. Details in Materials and Methods.

Figure 2. SL as a global and local connectivity by EEG. (A) The SL matrices, using grand-averaged values across the HD patients and the control participants are showed for each frequency band. Alpha bands are classified into two: lower alpha (8–10 Hz) and upper alpha (10–13 Hz). The number of EEG electrodes is 64, resulting in the 64 × 64 square matrix whose elements serve the average strength of SL values across the whole subjects between a pair of EEG electrodes. (B) Global functional connectivity indexed by mean SLs of HD patients was reduced (p = 0.015, permutation test) in delta-, theta-, and lower-alpha-bands. (C) 2D visualization of the connectivity in lower alpha based on the synchronization for a 64 EEG electrode cap. The dot (black) represents a node (i.e. EEG electrode). The segment (green) represents an edge (i.e. connectivity when the threshold is above 0.8). (D) Clustered connections from Cluster-based Permutation Tests (CBPT) between groups for the SL index of lower alpha band as a local connectivity. The nodes consisted of the PFC and occipitoparietal cortex comprised reduced synchronization in lower-alpha band of HD patients (p < 0.05, uncorrected for multiplicity).

Figure 3. Blood oxygenation and metabolism as PFC connectivity by fNIRS. (A) Connectivity matrices of the HD patients and the control participants. The number of fNIRS channels is 14, resulting in the 14 × 14 square matrix whose elements represent the average correlation coefficients across the whole subjects between a pair of fNIRS channels. (B) Brain network patterns of the HDs and CGs. Only the topmost 5% with correlation values greater than 0.78 are shown in the figures. The nodes (green: FP; blue, dlPFC; channels) are numbered by channel, and the weighted edges (gray) are displayed. Color bars indicate the correlation coefficient values. (C) Degree strength with the group average of the correlation coefficients in PFC of the HD patients (green) and control subjects (blue). Error bars indicate the standard error of the mean across the HDs and CGs. Compared with the controls, the HD patients show significantly greater connectivity strength in all the channels covering the PFC in terms of HbO.

Figure 4. Correlation between feature extracted from EEG electrodes and fNIRS channels covered the PFC by multimodal source power comodulation analysis (mSPoC). (A) Correlations of lower alpha rhythm and HbO of resting-state features obtained by mSPoC. (B) Component pairs of EEG electrodes and their corresponding fNIRS channels. (C) Configuration of EEG electrodes and fNIRS channels.

Figure 5. Resting-state activation pattern of the coupled EEG and fNIRS component. (A) Whole-brain sensor-space activation pattern of the EEG mSPoC component in lower alpha oscillations between groups. Color bar indicates the amplitude (i.e., band power) of the lower alpha frequency band in µV. (B) HbO-based fNIRS activation pattern of the corresponding mSPoC component in the PFC. These components were identified based on resting-state comodulation of amplitude dynamics in the EEG and the time course of HbO dynamics in the fNIRS. Color bar indicates the change in concentration of HbO (i.e., [HbOf]) in µM. (C) Estimate of the source-space pattern of the EEG component based on the sensor-space pattern after EEG-fNIRS fusion. Color bar indicates the optimized feature value of estimated components from the extract EEG and fNIRS source data.