Edited by: Davide V. Moretti, IRCCS San Giovanni di Dio Fatebenefratelli, Italy
Reviewed by: Xiaoli Li, Beijing Normal University, China; Umberto Melia, Universitat Politecnica de Catalunya, Spain
*Correspondence: Tiago H. Falk, Institut National de la Recherche Scientifique, Centre Energie, Matériaux, Télécommunications, University of Quebec 800, Rue de la Gauchetire Ouest, Suite 6900 Montreal, QC H5A-1K6, Canada e-mail:
This article was submitted to the journal Frontiers in Aging Neuroscience.
This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
Over the last decade, electroencephalography (EEG) has emerged as a reliable tool for the diagnosis of cortical disorders such as Alzheimer's disease (AD). EEG signals, however, are susceptible to several artifacts, such as ocular, muscular, movement, and environmental. To overcome this limitation, existing diagnostic systems commonly depend on experienced clinicians to manually select artifact-free epochs from the collected multi-channel EEG data. Manual selection, however, is a tedious and time-consuming process, rendering the diagnostic system “semi-automated.” Notwithstanding, a number of EEG artifact removal algorithms have been proposed in the literature. The (dis)advantages of using such algorithms in automated AD diagnostic systems, however, have not been documented; this paper aims to fill this gap. Here, we investigate the effects of three state-of-the-art automated artifact removal (AAR) algorithms (both alone and in combination with each other) on AD diagnostic systems based on four different classes of EEG features, namely, spectral, amplitude modulation rate of change, coherence, and phase. The three AAR algorithms tested are statistical artifact rejection (SAR), blind source separation based on second order blind identification and canonical correlation analysis (BSS-SOBI-CCA), and wavelet enhanced independent component analysis (wICA). Experimental results based on 20-channel resting-awake EEG data collected from 59 participants (20 patients with mild AD, 15 with moderate-to-severe AD, and 24 age-matched healthy controls) showed the wICA algorithm alone outperforming other enhancement algorithm combinations across three tasks: diagnosis (control vs. mild vs. moderate), early detection (control vs. mild), and disease progression (mild vs. moderate), thus opening the doors for fully-automated systems that can assist clinicians with early detection of AD, as well as disease severity progression assessment.
Alzheimer's disease (AD) is a chronic neuro-degenerative disorder that has recently been ranked as the third most expensive disease and the sixth leading cause of death in the United States (Leifer,
Driven by these limitations, quantitative electroencephalography (qEEG, henceforth referred to as “EEG”) has emerged as a promising tool capable of assisting physicians in the diagnosis of AD (e.g., Jeong,
Notwithstanding, EEG signals are inherently noisy and susceptible to blink, eye movement, heartbeats, and cranial muscle artifacts, all of which are detrimental to AD diagnosis performance. To overcome this limitation, the majority of the published works have resorted to using artifact-free EEG segments (called epochs) which have been selected by expert clinicians via meticulous visual inspection. Such dependence on human experts, however, hinders the benefits of automated low-cost analysis, as well as introduces possible human biases/errors (Daly et al.,
Here, three AAR algorithms have been selected after careful screening of the literature for available state-of-the-art methods applicable to our data. The first method, termed statistical artifact rejection (SAR), utilizes statistical characteristics of the signals to make accept/reject decisions over EEG epochs (Delorme et al.,
Fifty-nine participants were recruited from the Behavioral and Cognitive Neurology Unit of the Department of Neurology and the Reference Center for Cognitive Disorders at the Hospital das Clinicas in São Paulo, Brazil (Kanda et al.,
Twenty-channel EEG signals were acquired with the participants awake, relaxed, and with their eyes closed for at least 8 min. The Braintech 3.0 instrumentation (EMSA Equipamentos Médicos INC., Brazil) was used with 12-bit resolution and 200 Hz sample rate parameters. Impedance was maintained below 10 kΩ and scalp electrodes were placed according to the international 10–20 system. Bi-auricular referential electrodes were attached, as recommended by the Brazilian Society of Clinical Neurophysiology and the American EEG Society. An infinite impulse response low-pass elliptic filter with a zero at 60 Hz was applied to eliminate power grid interference. Moreover, based on evidence of an interhemispheric disconnection with AD (Jeong,
Unprocessed signals (both per-electrode and bipolar) constitute what will, henceforth, be referred to as the “raw” EEG. The enhanced signals, in turn, will constitute the raw signals processed by the different AAR algorithms described in the next subsection. Lastly, the raw signals have also been visually inspected by two experienced clinicians to obtain several 8-s epochs free of eye blinking, drowsiness, muscle movements, or equipment-related artifacts. This manually-selected data will be used to develop a gold-standard diagnostic system with which the AAR algorithms will be benchmarked against.
As mentioned previously, three AAR algorithms are explored within this work and were chosen based on characteristics of our dataset; more specifically, on the electrode layout (international 10–20 system), relatively small number of electrodes (20), absence of electrooculographic (EOG) reference channels, and lack of data from alternate modalities (e.g., accelerometers or gyroscopes). In the subsections to follow, a brief summary of the three AAR algorithms is given, as well as a description of their implementations. References to literature with more detailed descriptions of the algorithms are provided, where appropriate, for the interested reader.
The SAR method utilizes thresholding on the statistical characteristics of the EEG signals to select epochs that appear to contain artifacts. The implementation of this method was done using the well-known EEGLAB toolbox for Matlab (Delorme and Makeig,
The BSS algorithm utilizes spatial filtering to remove ocular and muscular artifacts from EEG data without external references (e.g., EOG or accelerometer signals) (De Clercq et al.,
Wavelet analysis has been used in the past for EEG artifact detection (e.g., Achanccaray and Meggiolaro,
Here, we have tested the three above-mentioned AAR algorithms alone, as well as in cascade; more specifically, we have tested the SAR-BSS and SAR-wICA combinations. Overall, experimental results will be presented using the “raw” data (this will be henceforth refereed to as the “baseline”), the manually-selected artifact-free EEG data (henceforth referred to as the “gold-standard”), and the five “enhanced” EEG datasets (i.e., SAR, BSS, wICA, SAR-BSS, SAR-wICA). To maintain consistency with the gold-standard system, all datasets are segmented into several 8-s epochs.
Several EEG features have been proposed in the literature over the last decade and shown to accurately discriminate between healthy controls and AD patients. The effects of EEG artifacts on these features, however, are unknown, as are their effects on overall diagnostic performance. Here, we will pursue such an investigation and focus will be placed on four traditional EEG feature categories, namely, spectral power, magnitude square coherence, phase coherence/synchrony, and the recently-proposed EEG amplitude modulation rate-of-change. In the subsections to follow, a brief description of the features will be given. References to literature with more detailed descriptions of the features are provided, where appropriate, for the interested reader.
The pivotal process to quantify the frequency-domain properties of the EEG signal lies in the estimation of its power spectral density (PSD) function, which is commonly achieved via a discrete Fourier transform (Sörnmo and Laguna,
The magnitude square coherence (MSC), frequently referred to as “coherence,” is a measure of co-variance between two power spectra. In EEG studies, the MSC is used as a metric of synchrony in neural activity, which is an indicator of cortical connectivity (Thatcher et al.,
In our experiments, we compute both metrics for each of the five EEG frequency bands. Following the recent evidence of an interhemispheric disconnection with AD (Jeong,
Global field synchrony (GFS) measures the phase synchrony in a given frequency (or frequency band) for a set of
Amplitude modulation analysis has shown to be a valuable tool for bio-signal processing and analysis (Atlas and Shamma,
Computed features were grouped into four feature sets: spectral, modulation, coherence (MSC), and phase (phase coherence and phase synchrony). To explore the complementarity of the extracted features, combined feature sets were also investigated. Henceforth, we will refer to the “All” feature set as the set that combines all the extracted features and the “Spec-Mod” set as the set that combines the spectral and amplitude-modulation based features. This latter combined set is motivated by the recent results suggesting the complementary of the two feature domains for AD characterization (Fraga et al.,
As an additional EEG “cleaning” tool, we use epoch averaging in the feature domain as a way of improving the signal-to-noise ratio (SNR) of the extracted features. This procedure was recently shown to improve the clustering of amplitude modulation rate-of-change features, thus leading to higher diagnostic accuracies (Fraga et al.,
The machine learning and pattern recognition literature has presented a plethora of possible feature selection and classification algorithms which can be fine-tuned to specific applications and feature sets. For the experiments herein, however, we are interested in understanding the effects of AAR algorithms on different EEG feature sets and on overall diagnostic performance, and not the effects of different selection/classification algorithms and their internal parameters. As such, our experiments are based on a support vector machine (SVM) feature selection and classification algorithm that is widely used in the EEG-based AD diagnosis literature (Lehmann et al.,
In our experiments, 25% of the available data was randomly set aside for feature selection and the remaining 75% was used for classifier training/testing using 10-fold cross validation. Using disjoint sets for feature selection and classifier training reduces any unwanted biases in the reported performance figures. To remain inline with the existing EEG-based AD diagnostic literature, feature selection was used to sift out the 24 most relevant features for AD diagnosis. In this study, we investigate the effects of AAR on AD diagnostic performance using three classification tasks, namely, (a) Task 1:
In order to assess diagnosis performance, classification accuracy is used as a performance metric. Moreover, for the two 2-class problems described above, diagnosis sensitivity and specificity are also used. Throughout the remainder of this paper we will assess the impact of AAR on AD classification by measuring the performance gains obtained relative to the baseline (i.e., using the “raw" EEG data). The relative performance gain is given by:
Table
Baseline (%) | 73.2 | 68.4 | 60.1 | 45.7 | 72.3 | 73.5 |
SAR | 1.3 | −3.6 | 0.2 | 1.8 | 2.5 | −0.8 |
SAR-BSS | −5.9 | −10.6 | −6.2 | −12.2 | −1.0 | −3.7 |
SAR-wICA | −0.8 | −3.0 | 7.6 | 2.6 | 4.5 | 2.5 |
BSS | −4.0 | −4.6 | −6.5 | −12.2 | −6.6 | −7.4 |
wICA | 3.3 | 2.9 | 11.5 | 5.5 | 8.4 | 3.8 |
2 | 83.6 | 86.3 | 80.5 | 79.6 | 82.9 | 75.7 | 73.3 | 76.1 | 70.0 | 64.9 | 78.4 | 48.7 | 83.0 | 84.3 | 81.3 | 82.6 | 85.4 | 79.2 |
3 | 89.4 | 91.3 | 86.8 | 85.1 | 89.5 | 79.3 | 78.5 | 81.9 | 74.0 | 69.4 | 84.9 | 48.6 | 89.2 | 92.2 | 85.2 | 88.6 | 90.9 | 85.5 |
SAR | −0.3 | −3.0 | 2.9 | 3.4 | 1.9 | 5.2 | 2.8 | 3.3 | 2.2 | 3.7 | −0.9 | 11.6 | 2.2 | 1.8 | 2.7 | 2.5 | −1.0 | 6.7 |
SAR-BSS | −2.1 | −5.5 | 2.0 | −2.9 | −1.8 | −4.5 | −0.3 | 1.6 | −3.0 | −2.3 | −0.4 | −6.3 | −2.3 | −2.9 | −1.5 | −0.6 | −1.3 | 0.3 |
SAR-wICA | 4.3 | 3.2 | 5.7 | −2.0 | −3.2 | −0.3 | 1.9 | 3.8 | −0.8 | −1.5 | 0.6 | −5.6 | 4.6 | 4.1 | 5.1 | 3.6 | 2.8 | 4.5 |
BSS | −4.6 | −7.2 | −1.3 | −6.2 | −5.4 | −7.3 | −2.2 | 0.9 | −6.4 | −4.7 | 0.2 | −15.7 | −4.0 | −3.6 | −4.6 | −1.9 | −4.8 | 1.7 |
wICA | 6.7 | 5.1 | 8.8 | 3.2 | 3.7 | 2.4 | 0.9 | 4.3 | −3.9 | 4.5 | −1.8 | 14.6 | 8.7 | 8.8 | 8.5 | 7.7 | 4.8 | 11.2 |
SAR | 3.1 | 2.9 | 3.3 | 1.5 | 1.0 | 2.3 | −1.5 | −2.1 | −0.6 | 2.6 | 0.6 | 6.9 | 2.2 | −0.4 | 5.7 | 2.7 | 2.2 | 3.4 |
SAR-BSS | −3.8 | −2.3 | −6.0 | −5.8 | −6.0 | −5.5 | −2.7 | −1.1 | −5.2 | −2.8 | 5.3 | −28.2 | −0.9 | −2.4 | 1.2 | −2.2 | −1.0 | −3.9 |
SAR-wICA | 1.0 | 1.9 | −0.3 | 0.0 | 0.7 | −1.2 | 4.3 | 0.5 | 9.4 | 2.2 | 0.3 | 6.3 | 3.2 | 2.0 | 5.0 | 3.9 | 2.6 | 5.6 |
BSS | −5.2 | −4.8 | −5.8 | −7.4 | −5.1 | −11.0 | −2.1 | 3.7 | −12.0 | −4.8 | 3.7 | −32.1 | −8.1 | −8.4 | −7.7 | −3.9 | −3.8 | −4.0 |
wICA | 2.1 | 3.4 | 0.2 | 3.8 | 4.2 | 3.1 | 9.3 | 7.5 | 11.8 | 5.0 | 2.4 | 10.4 | 7.4 | 4.8 | 10.8 | 4.7 | 4.5 | 4.9 |
1 | PZ_alpha_pwr* | PZ_alpha_pwr* | P3_P4_delta_pwr |
2 | C3_C4_delta_pwr | P3_alpha_pwr* | O1_O2_theta_cohe_pha |
3 | P3_P4_delta_pwr | O1_O2_theta_pwr* | C3_alpha_pwr |
4 | P3_alpha_pwr* | T3_T4_delta_pwr | F4_delta_pwr |
5 | P3_P4_delta_m-delta | F7_delta_pwr | T4_delta_pwr |
6 | FP1_FP2_beta_cohe_mag* | C3_C4_beta_m-beta | T3_T4_beta_pwr* |
7 | P3_P4_delta_cohe_mag* | F3_delta_pwr | T5_beta_pwr* |
8 | T3_T4_delta_pwr | O1_O2_delta_m-delta | OZ_beta_pwr |
9 | P3_delta_pwr | O1_O2_beta_cohe_mag* | FP1_FP2_beta_cohe_mag* |
10 | O1_alpha_pwr* | FP1_FP2_delta_cohe_mag* | FZ_beta_m-alpha |
11 | T4_theta_pwr* | FP1_delta_pwr | F3_beta_m-beta |
12 | T3_delta_pwr | T3_delta_m-delta | T5_theta_pwr* |
13 | T5_beta_pwr* | C3_delta_m-delta | T3_alpha_pwr* |
14 | O1_O2_theta_pwr* | P4_alpha_pwr* | T5_T6_delta_cohe_mag* |
15 | F8_beta_pwr | O1_alpha_pwr* | C4_delta_pwr |
16 | CZ_beta_pwr | T5_beta_pwr* | C3_C4_delta_cohe_mag* |
17 | T4_theta_m-theta* | CZ_beta_pwr | O1_O2_beta_m-theta |
18 | C3_C4_beta_m-beta | F8_beta_pwr | P3_P4_delta_m-delta |
19 | F7_beta_pwr | T3_T4_beta_m-alpha | F3_F4_beta_m-beta |
20 | C3_beta_pwr | T3_T4_beta_cohe_mag* | T3_T4_delta_cohe_mag* |
21 | F3_delta_pwr | F7_F8_beta_cohe_mag* | P4_beta_m-alpha |
22 | OZ_delta_pwr | FZ_beta_m-alpha | F3_F4_alpha_pwr |
23 | FZ_beta_m-alpha | T5_T6_theta_pwr* | FP1_theta_pwr* |
24 | C3_alpha_pwr* | F3_alpha_pwr* | O1_alpha_pwr |
Spectral power | 18 (7) | 14 (8) | 13 (5) |
Modulation | 4 (1) | 6 (0) | 6 (0) |
Coherence | 2 (2) | 4 (4) | 4 (4) |
Phase | 0 (0) | 0 (0) | 1 (0) |
Frontal | 5 (1) | 8 (3) | 7 (2) |
Central | 5 (1) | 3 (0) | 3 (1) |
Temporal | 5 (3) | 6 (3) | 7 (6) |
Parietal | 6 (3) | 3 (3) | 3 (0) |
Occipital | 3 (2) | 4 (3) | 4 (0) |
Delta | 9 (1) | 8 (1) | 8 (3) |
Theta | 3 (3) | 2 (2) | 3 (2) |
Alpha | 4 (4) | 5 (5) | 4 (1) |
Beta | 8 (2) | 9 (4) | 9 (3) |
1 | O1_O2_theta_pwr | O1_O2_theta_pwr | CZ_beta_pwr |
2 | P3_P4_theta_pwr | PZ_delta_pwr | P4_alpha_m-theta |
3 | T5_theta_m-theta | CZ_beta_m-theta | P3_P4_delta_pwr |
4 | F7_F8_alpha_cohe_pha | FP2_beta_pwr | F7_alpha_m-delta |
5 | T3_theta_m-delta | FP1_beta_m-beta | O1_O2_theta_cohe_pha |
6 | P3_P4_delta_pwr | O1_O2_alpha_pwr | T3_theta_pwr |
7 | PZ_alpha_pwr | O1_O2_beta_cohe_pha | OZ_beta_m-alpha |
8 | O1_O2_alpha_pwr | F7_F8_alpha_cohe_pha | P3_P4_theta_m-theta |
9 | C4_alpha_m-delta | T6_delta_m-delta | P3_P4_beta_m-alpha |
10 | FP2_beta_pwr | FP1_delta_pwr | O1_O2_theta_m-theta |
11 | T3_T4_alpha_m-theta | OZ_beta_m-beta | T4_theta_pwr |
12 | T5_T6_beta_m-delta | O1_O2_beta_m-theta | T6_theta_m-theta |
13 | T6_beta_m-delta | T3_T4_beta_m-alpha | P3_P4_beta_m-beta |
14 | T4_theta_pwr | F7_F8_beta_m-beta | C3_C4_alpha_cohe_mag |
15 | O1_O2_alpha_m-theta | PZ_alpha_pwr | P3_P4_beta_pwr |
16 | O1_delta_pwr | OZ_beta_pwr | P3_P4_theta_m-delta |
17 | P3_P4_beta_m-theta | C4_delta_m-delta | T5_T6_alpha_cohe_mag |
18 | T3_theta_pwr | CZ_beta_m-alpha | F7_F8_alpha_cohe_mag |
19 | OZ_beta_pwr | F4_theta_m-delta | P4_beta_m-beta |
20 | F3_F4_theta_pwr | F3_F4_delta_cohe_mag | T5_T6_delta_cohe_mag |
21 | T6_delta_pwr | FP1_FP2_beta_cohe_mag | T3_T4_theta_cohe_mag |
22 | C4_delta_m-delta | P3_P4_delta_cohe_mag | FP1_theta_m-delta |
23 | T3_T4_beta_m-beta | T5_beta_pwr | T3_theta_m-delta |
24 | PZ_delta_pwr | FZ_delta_pwr | C3_C4_delta_cohe_pha |
Spectral power | 13 | 9 | 5 |
Modulation | 10 | 10 | 12 |
Coherence | 0 | 3 | 5 |
Phase | 1 | 2 | 2 |
Frontal | 3 | 9 | 3 |
Central | 2 | 3 | 3 |
Temporal | 9 | 3 | 7 |
Parietal | 5 | 3 | 8 |
Occipital | 5 | 6 | 3 |
Delta | 5 | 7 | 3 |
Theta | 7 | 2 | 10 |
Alpha | 6 | 3 | 5 |
Beta | 6 | 12 | 6 |
Interhemispheric | 11 | 10 | 14 |
The list of top-selected features shown in Table
Moreover, when discriminating between the three classes, features from the temporal and parietal regions showed to be important across the two scenarios. For the
As for frequency bands, in the wICA scenario, delta and beta band features corresponded to roughly 70% of the selected features for each of the three tasks, followed by alpha band features (15%), thus corroborating previous studies that show the slowing of the EEG with AD (e.g., Coben et al.,
In order to characterize the effects of the wICA algorithm on the distribution and statistics of the salient features, we utilize a so-called distribution overlap metric which measures the amount of overlap between the histogram of a particular feature before and after wICA AAR. The metric is normalized to lie between 0−100% with 0 and 100% overlap values suggesting complete change and no change in feature statistics post-AAR, respectively. For simplicity, Table
From Tables
Interestingly, for Task 3 involving AD1 and AD2 patients, the wICA-AAR system outperformed the gold standard, achieving accuracy, sensitivity, and specificity values of 96.3, 96.9, and 95.5%, respectively. The gold standard, in turn, obtained values 92.8, 97.3, and 86.7%, respectively. It is suspected that this improved performance was obtained due to information harnessed from the frontal electrodes, which were often selected by the wICA-processed data and not from the manually-selected data. Frontal electrodes are susceptible to eye-related artifacts and are likely often discarded by human experts. Notwithstanding, the frontal region has been shown in classical studies to be severely affected by disease progression (Mann et al.,
Moreover, from Tables
Lastly, we explored the gains obtained with feature averaging as a simple SNR improvement tool. For Task 1, the accuracy gains relative to the baseline obtained with only feature averaging (i.e., raw EEG data without AAR) were of 3.3, 4.9, 3.4, and 1.9% for the spectral, amplitude modulation, coherence, and phase feature sets, respectively. For Task 2, in turn, these relative accuracy gains were of 1.5, 1.1, 2.6, 2.2%, respectively. Lastly, for Task 3 the relative gains were 3, 0.8, 2.4, and 2% respectively. As can be seen, simple feature averaging (Fraga et al.,
The three enhancement algorithms explored here represented the state-of-the-art applicable to the constraints imposed by our available database, such as small number of channels (20), limited amount of data per participant, and lack of EOG reference channels. For future studies without these limitations, alternate AAR algorithms can be explored. For example, for studies involving EEG with over 64 channels and EOG, the ADJUST (Automatic EEG artifact Detection based on the Joint Use of Spatial and Temporal features) (Mognon et al.,
The last decade has seen a rise in the development of EEG-based tools to assist clinicians with AD diagnosis. This paper has evaluated the effects of different state-of-the-art AAR algorithms on diagnosis performance; AAR algorithms were tested both alone and in tandem. Experimental results showed the wavelet enhanced ICA (wICA) AAR algorithm outperforming all other algorithms across four investigated feature sets (spectral, amplitude modulate rate-of-change, coherence, phase), as well as two combined feature sets (“All” and “Spectral-modulation”). In a disease progression monitoring task (Task 3), the automated system was shown to outperform a diagnostic system trained on artifact-free data processed by human experts. Such findings suggest that the discard of useful discriminatory information can be avoided if AAR algorithms are used. Ultimately, it is hoped that such fully-automated diagnostic tools be used to assist clinicians not only with early diagnostics, but also with disease progression monitoring and assessment.
The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
This work was funded by the Natural Sciences and Engineering Research Council of Canada (NSERC) and the Foundation for Research Support of the State of São Paulo (FAPESP). The authors would also like to thank Dr. Justin Dauwels for sharing his scripts to calculate the global field synchrony feature.