Functional connectivity in EEG: a multiclass classification approach for disorders of consciousness.

Characterizing functional connectivity (FC) in the human brain is crucial for understanding and supporting clinical decision making in disorders of consciousness. This study investigates FC using sliding window correlation (SWC) analysis of electroencephalogram (EEG) applied to three connectivity measures: phase-lag index (PLI) and weighted phase-lag index (wPLI), which quantify phase synchronization, and amplitude envelope correlation (AEC), which captures amplitude-based coactivation patterns between pairs of channels. SWC analysis is performed across the five canonical frequency bands (delta, theta, alpha, beta, gamma) of EEG data from four distinct groups: coma, unresponsive wakefulness syndrome, minimally conscious state, and healthy controls. The extracted SWC metrics, mean, reflecting the stability of connectivity, and standard deviation, indicating variability, are analyzed to discern FC differences at the group level. Multiclass classification is attempted using various models of artificial neural networks that include different multilayer perceptrons (MLP), recurrent neural networks, long-short-term memory networks, gated recurrent units, and a hybrid CNN-LSTM model that combines convolutional neural networks (CNN) and long-short-term memory network to validate the discriminative power of these FC features. The results show that MLP model 2 achieves a classification accuracy of 96.3% using AEC features obtained with a window length of 16s, highlighting the effectiveness of AEC. An evaluation of the model performance for different window sizes (16 to 20 s) shows that MLP model 2 consistently achieves high accuracy, ranging from 95.5% to 96.3%, using AEC features. When AEC and wPLI features are combined, the maximum accuracy increases to 96.9% for MLP model 2 and 96.7% for MLP model 3, with a window size of 17 seconds in both cases.