Table 2 Extracted features across studies (a summary of pre-processing strategies, feature extraction methods, feature selection, and top predictors across studies).
First author, year | Pre-processing strategy | EEG features | Feature extraction method | Feature selection method | Top features Top 10 features, if applicable |
|---|---|---|---|---|---|
Studies predicting response to neurostimulation therapy | |||||
Bailey [26] | Data down-sampled to 1000 Hz Second order Butterworth filtering with bandpass from 1 to 80 Hz and a band-stop filter 47–53 Hz Fast ICA used to manually select and remove eye blinks, movements, and remaining muscle artifacts. | Power spectral analysis connectivity analysis | Power spectral analysis - Morlet Wavelet transform to calculate power in the upper alpha band (10–12.5 Hz), theta band (4–8 Hz), and gamma band (30–45 Hz) - Average power calculated across the entire retention period with each frequency band and averaged over trials Connectivity analysis - Hanning taper time–frequency transform to determine instantaneous phase values for complex Fourier-spectra from 4 to 45 Hz with a 1 Hz resolution across a 3-oscillation sliding time window - Weighted phase lagged index (wPLI) calculated between each electrode - wPLI provides a value between 0 and 1 for each electrode pair at each frequency and time point | Not applicable | Statistically significant variables between responders and non-responders; authors did not report top features in the total model - Greater theta power at Fz in responders vs. non-responders (F1 = 8.577, p = 0.006) - No significant differences for alpha or gamma power, or theta-gamma coupling - Responders showed a non-significant pattern of less gamma connectivity than non-responders at baseline (p = 0.523), and greater gamma connectivity at week 1 (p = 0.0836). - Responders showed significantly more theta connectivity across baseline and week 1, with both interhemispheric fronto-parietal coupling and frontal and parietal interhemispheric coupling (overall p = 0.003). |
Bailey [26] | Same Procedure as Bailey [26] | Power spectral analysis Connectivity analysis Theta cordance analysis | Power and connectivity analyses follow the same procedure as Bailey 2017 Theta cordance analysis - Absolute power values for each epoch 1–80 Hz underwent a multi-taper fast Fourier frequency transformation with a Hanning taper - Absolute power averaged across neighboring electrode pairs - Relative power in reattributed absolute theta band calculated by dividing power in theta band by total power from 1 to 80 Hz - Subtracted half-maximal values from normalized absolute and relative power in theta band, and summed together for each electrode iAPF analysis - Individualized alpha peak frequency averaged across F3, Fz, and F4 electrodes - Multitaper fast Fourier frequency transformation - Gaussian distribution with least-squared error fitted to electrodes in 6–14 Hz range - Peaks of distribution selected from each electrode and averaged | Not applicable | Statistically significant variables between responders and non-responders; authors did not report top features in the total model - Greater theta connectivity in responders vs. non-responders (p = 0.0216, FDR p = 0.030). Responders showed atypical, elevated theta connectivity, while non-responders showed typical theta connectivity, which was comparable to controls. - No main effect of theta cordance, frontal-midline theta power, or alpha power. |
Corlier [28] | ICA-based FASTER algorithm Dominant alpha frequency peak determined for each subject (highest spectral peak within 7-13 Hz alpha range) | EEG functional connectivity measures (coherence, envelope correlation, and alpha band frequency) | Functional connectivity measures - Coherence: correlation of amplitude and phase - Envelope: correlation of amplitude - Alpha frequency band: similarity of the spectral waveform of the alpha band across regions | Elastic Net | Coherence & Envelope: Connections in the frontal to temporo-parietal nodes Alpha frequency band: Connections between the left frontal seeds (near stimulation site) and contralateral fronto-temporal locations EN models for coherence and envelope correlation showed a diffuse coupling pattern, while αSC showed a more focal connectivity. |
Erguzel [30] | Manually selected artifact-free EEG data with a minimum split-half reliability ratio of 0.95 and minimum test-retest reliability ratio of 0.90. FFT | EEG cordance (combines absolute and relative EEG power, and negative discordance values) | EEG cordance - Normalized power across electrode sites and frequency bands - Maximum absolute and relative power of each frequency band is calculated to derive normalized absolute and relative power - Half-maximal value is subtracted, absolute/relative normalized power is summed. | Genetic algorithm - adaptive heuristic search algorithm was applied to features of all selected channels to reduce the number of dimensions | Fp1, Fp2, F7, F8, and F3 in the theta frequency band |
Erguzel [29] | Band-pass filter with 0.15–30 Hz frequency FFT used to calculate absolute and relative power in each of two non-overlapping frequency bands (Delta—1–4 Hz, theta—4–8 Hz) | EEG cordance (combines absolute and relative EEG power, and negative discordance values) | EEG cordance - Normalized power across electrode sites and frequency bands - Maximum absolute and relative power of each frequency band is calculated to derive normalized absolute and relative power - Half-maximal value is subtracted, absolute/relative normalized power is summed. | ANN | NA |
Erguzel [31] | Band-pass filter with 0.15–30 Hz frequency Manually selected artifact-free EEG data (at least 2 min) FFT | EEG cordance (combines absolute and relative EEG power, and negative discordance values) | EEG cordance analyses follow the same procedure as Erguzel 2014 | Not applicable | Feature set was composed of frequency bands for six frontal electrodes (Fp1, Fp2, F3, F4, F7 and F8) |
Hasanzadeh [33] | Sampling frequency 500 Hz Bandpass FIR filter (1–42 Hz) ICA to remove noisy data MARA to label noisy ICs Visually inspected to eliminate remaining artifacts | 21 features in four categories (nonlinear, PSDl, spectral, and cordance) | Nonlinear features - LZC: Complexity measure of time series to estimate scholastic and chaotic behavior of time series - KFD: Algorithm for computing fractal dimension, a measure of self-similarity of a time series based on number of patterns repetitions Power spectral density - Delta (1–4 Hz)—Beta (12–30 Hz) by Welch method with a non-overlapped window, 500 samples in length - Average power computed for frequencies in each band Spectrum features - Method that quantifies the degree of phase coupling between components of a signal Cordance - measure of complexity of system based on chaos and time delay reconstruction theory | mRMR | - Nonlinear (LZC, KFD, CD)—80.4% accuracy - Power (D, T, A, B) - 91.3% accuracy - Spectrum (BispSL, Bisp2M, and BispEn in all bands)—84.8% accuracy - Cordance (Fr, Pre, Fr)—76.1% accuracy - All—87% accuracy |
Studies predicting response to pharmacological treatment | |||||
Cao [34] | Real-time artifact removal algorithm based on CCA, feature extraction, and a GMM used to improve signal quality | Power spectral analysis EEG Alpha Asymmetry EEG Theta Cordance | Power spectral analysis - 256-point FFT using Welch’s method - 10 min spans of data with 256-point moving window at 128-point overlap - Absolute and relative power of four prefrontal channels from delta (1–3.5 Hz), theta (4–7.5 Hz), lower alpha (8–10 Hz) and upper alpha (10.5–12 Hz) bands. EEG alpha asymmetry - mid-prefrontal (Fp1/Fp2) and mid-lateral (AF7/AF8) hemispheric asymmetry index to establish a relative measure of the difference in EEG (lower and upper) alpha power between the right and left forehead areas. EEG theta cordance - Combines information from both absolute and relative powers in the EEG theta band | p-value: measured using the Wilcoxon rank-sum test with a significant p-value < 0.05. | 0.5 mg/kg dose - AF7 theta—p = 0.042 - Fp2 theta—p = 0.035 0.2 mg/kg dose - Fp1 theta—p = 0.038 - Fp2 theta—p = 0.042 |
Cooks [35] | Artifact-free epochs selected following rejection of muscle, electrocardiographic, and drowsiness artifacts. | Power spectral analysis ATR Relative combined theta and alpha power | Power spectral analysis - Calculated using consecutive two-second epochs of eyes-closed rest, by averaging values calculated separately for each channel in each epoch Relative combined theta and alpha power - Non-linear weighted combination of relative combined theta and alpha power (3–112 Hz), alpha1 power (8.5–12 Hz) and alpha2 absolute power (9–11.5 Hz) | Relative combined theta and alpha power was scaled to a range from 0 to 100; a cut-off score of ≥46.2 was selected | NA |
Jaworska [37] | Bandpass filters 0.1–80 Hz 100 s of artifact-free data subjected to a FFT ln-transformed prior to analyses to ensure normality (Minimizes influence of extreme values) | eLORETA analysis Theta Cordance | eLORETA analysis - estimates neural activity as current density based on MNI-152 template, creating a low-resolution activation image Theta cordance - Values from prefrontal electrodes (Fp1, Fp2) at baseline and week 1 | Tree-based feature selection kernel PCA | eLORETA features were most important, comprising 17 delta, 20 theta, 14 alpha1, 20 alpha2, and 17 beta EEG features. Delta Power at week 1 at T8 followed by power at Cp6 Theta Baseline power at Fp2 and week 1 power at Fc2 Alpha1 Baseline power at F7/8 Alpha2 Baseline power at P8 and week 1 power at O1 Beta Baseline power at T7 and week 21 power at Fz |
Mumtaz [38] | Bandpass filters 0.1–70 Hz EEG data collected during 5 min eyes open, and 5 min eyes closed - 3-stimulus visual Oddball task used 50 Hz notch filter used to suppress power line noise | Wavelet coefficients in the delta and theta frequency range | Wavelet coefficients - involves a window function to capture both low and high-frequency components of the signal | Rank-based feature selection according to their relevance to class labels minimum redundancy and maximum relevance | Top EEG features: Fp2—delta frequency C3—theta frequency F7—delta frequency F3—delta frequency F7—theta frequency T4—theta frequency F8—theta frequency F4—delta frequency Fz—delta frequency F4—delta frequency C4—delta frequency F8—theta frequency T4—delta frequency P3—theta frequency |
Rajpurkar [39] | Raw EEG signal was filtered using a band-pass filter with 0.15 - 30 Hz frequency prior to artifact removal FFT | Relative and absolute band power Frontal alpha asymmetry Occipital asymmetry Ratio of beta/alpha band power Ratio of theta/alpha band power | Relative/absolute power as described above Frontal alpha asymmetry - difference in alpha bandpower between O2 and O1 Occipital beta asymmetry - difference in beta bandpower between O2 and O1 ratio of beta/alpha and theta/alpha band power - Calculated for each electrode Feature selection: Decision tree weight in LightGBM | Gradient boosted feature selection | Top EEG features: 1. T7-T3 alpha absolute ratio 2. T7-T3 beta absolute ratio 3. F7 gamma relative 4. Fp2 delta relative 5. F3 alpha absolute 6. Fp2 theta absolute 7. P4 alpha absolute 8. T7-T3 beta relative ratio 9. F7 beta relative |
Salle [36] | Data was filtered (0.1–30 Hz), ocular-corrected, and inspected for artifacts (voltages ±μV, faulty channels, drift) Minimum of 100 s of artifact-free data was required for participant inclusion | Theta Cordance (Prefrontal— Fp1, Fp2 MRF—Fz, Fp2, F4, F8) | EEG theta cordance Combines information from both absolute and relative powers in the EEG theta band | NA | Top EEG features: Change in prefrontal theta cordance (Fp1 + Fp2) = 81% accuracy Change in MRF theta cordance (Fz, Fp2, F4, F8) = 74% accuracy |
Wu [40] | 60 Hz AC line noise artifact removed using CleanLine - Non-physiological slow drifts in EEG recordings were removed using 0.01 Hz high-pass filter - Spectrally filtered EEG data were re-referenced to common average - Bad channels were rejected based on thresholding spatial correlations among channels - Subjects with more than 20% bad channels were discarded - Rejected channels were interpolated from EEG of adjacent channels via spherical spline interpolation - Remaining artifacts were removed using ICA - EEG data re-referenced to common average | SELSER Channel-level alpha band power Theta Coherence Band power features of latent signals extracted with ICA or PCA | Alpha band power and theta coherence as described above SELSER - spatial filter transforms multi-channel EEG data into a single latent signal, where the power is used as a feature - model fitting is done under a sparse constraint on the number of spatial filters, which reduces dimensionality Latent signals extracted with ICA or PCA - eigenvalues of the covariance matrix to reduce dimensionality | SELSER | Best performance using SELSER on alpha frequency range eyes-open rsEEG data (feature importance was not reported) |
Zhdanov [41] | 0.05–100 Hz bandpass filter Filtering performed using 2nd order Butterworth filters applied to the data in forward and reverse direction, to eliminate phase distortion Data pre-processed with EEGLAB toolbox Channels contaminated by large sporadic artifact were identified by human analyst and deleted EEG data bandpass filtered 1–80 Hz Notch-filtered at 60 Hz | Electrode-level spectral features Source-level spectral features Multiscale-entropy-based features Microstate-based features | Electrode-level spectral features - EEGLAB function spectopo to obtain power spectrum - log-transformed absolute power obtained for each channel - For each pair, absolute power at left electrode divided by right, resulting in 25 features for each band Source-level spectral features eLORETA algorithm as implemented by LORETA-KEY software Following regions selected on basis of prior literature: ACC, rACC, and mOFC Multiscale-entropy-based features - Quantifies variability of time series by estimating predictability of amplitude patterns across a time series - Two consecutive data points were used for data matching, and points were considered to match if their absolute amplitude difference was <15% of the standard deviation of the time series. Microstate-based features - Implemented using CARTOOL - average duration: average amount of time a microstate class remains stable when it appears (in ms) - frequency: occurrence of each microstate class per second - coverage: % of recording covered by each microstate class | Unpaired 2-tailed t test | MSE asymmetry features—C3/C4 (baseline) MSE asymmetry features—FC3/FC4 (baseline) MSE asymmetry features—T7/T8 (week 2) MSE asymmetry features—CP3/CP4 (week 2) Electrode-level spectral asymmetry—P3/P4 alpha low (baseline) Electrode-level spectral asymmetry —T7/TP8 theta (week 2) Electrode-level spectral asymmetry —F7/F8 beta mid (week 2) Source-level spectral features—alpha high ACC, rACC (week 2) |
