Alzheimer’s Disease Detection Using Combined EEG Source Connectivity and Microstate Features

Figures

• 

  Figure 1. Algorithmic framework of SMMS. (a) wPLI features extraction: Separate the signal into distinct frequency bands, with sLORETA used for ESL, and extract average intensity from the wPLI matrix. (b) Microstate features extraction: The EEG’s global field power was clustered into five microstates using k-means, followed by the extraction of features. (c) Classification: Features were fused, selected with ReliefF, and classified using k-NN.

• 

  Figure 2. Average EEG microstate templates of the AD, FTD, and HC groups together with the Koenig template. Red and blue colors represent positive and negative scalp potentials, respectively.

• 

  Figure 3. Averaged wPLI connectivity matrices across different frequency bands. Warmer colors indicate stronger functional connectivity, while cooler colors indicate weaker connectivity. The blue lines mark the boundary of each brain region, and the color bar represents the wPLI connectivity strength.

• 

  Figure 4. Microstate parameter box diagrams for AD, FTD, and HC groups: (a) duration of microstates, (b) coverage of microstates, and (c) occurrence of microstates. Group color designations are yellow for AD, blue for FTD, and purple for HC. Boxplots represent the median (red center line), the 25th and 75th percentiles (box bounds), and the 1.5 × IQR range (whiskers), with outliers displayed individually as separate points (∗ p <0.0001).

• 

  Figure 5. State transition matrices for the diagnostic groups: (a–c) represent state transition probability matrices for the AD, FTD, and HC groups, respectively. Panels (d–f) depict pairwise group differences (the former minus the latter). For descriptive visualization rather than inferential testing, cells with numerical differences greater than +1.0% or lower than −1.0% are highlighted in red and green, respectively.

• 

  Figure 6. Correlation between microstate parameters and wPLI in the AD group. (a) Duration of microstate C was positively correlated with the average alpha wPLI (PCC=0.4664, p=0.0041), (b) Coverage of microstate B was negatively correlated with the average alpha wPLI (PCC=−0.4651, p=0.0043), (c) Coverage of microstate C was positively correlated with the average alpha wPLI (PCC=0.4137, p=0.0121).

• 

  Figure 7. Classification accuracy as a function of the number of selected ReliefF features under subject-independent cross-validation.