Figure 1

Overview of the TMS-EEG data processing for classifying AD and HC participants; (A) TMS-EEG data from both AD and HC participants are taken as input. (B) Prior to extracting features, the data is pre-processed to eliminate unwanted interference as well as TMS related artifacts, eye blinks or movement. (C) Several features are extracted from the TMS-EEG evoked potentials including descriptive statistics, Hjorth parameters and AUC of LMFP or GMFP, as well as peak values in early windows of the TEP (25–40 ms, 45–80 ms, 85–150 ms and 160–250 ms). (D) All features are extracted per individual trial from each recorded channel for each subject. (E) Ten regions of interest (ROI) are defined: FL frontal left, FR frontal right, CL central left, CR central right, CPL central-parietal left, CPR central-parietal right, POL parietal-occipital left, POR parietal-occipital right, TL temporal left, TR temporal right. The electrodes from each individual ROI are depicted with different colors. (F) Feature values are averaged over the entire electrode set (global) and on particular ROIs. (G) Different combinations of features are fed into a random forest classifier. (H) As a byproduct of classification, an importance value is extracted for each input feature. (I) The model is validated in a leave-one-out cross-validation scenario. (J) Finally, the metrics used for evaluating the classification performance are accuracy, sensitivity, specificity and F1 score. Max maximum amplitude, Min minimum amplitude, S skew, K kurtosis, E energy, hA Hjorth Complexity, hM Hjorth Mobility, hC Hjorth Complexity, P1-4 peaks from defined windows (25–40 ms, 45–80 ms, 85–150 ms, 160–250 ms), AUC area under the curve for LMFP and GMFP.