Skip to main content
Download PDF

Selection of the optimal channel configuration for implementing wearable EEG devices for the diagnosis of mild cognitive impairment

Alzheimer's Research & Therapy volume 14, Article number: 170 (2022) Cite this article

Abstract

Background

Early diagnosis of mild cognitive impairment (MCI) is essential for timely treatment planning. With recent advances in the wearable technology, interest has increasingly shifted toward computer-aided self-diagnosis of MCI using wearable electroencephalography (EEG) devices in daily life. However, no study so far has investigated the optimal electrode configurations for the efficient diagnosis of MCI while considering the design factors of wearable EEG devices. In this study, we aimed to determine the optimal channel configurations of wearable EEG devices for the computer-aided diagnosis of MCI.

Method

We employed an EEG dataset collected from 21 patients with MCI and 21 healthy control subjects. After evaluating the classification accuracies for all possible electrode configurations for the two-, four-, six-, and eight-electrode conditions using a support vector machine, the optimal electrode configurations that provide the highest diagnostic accuracy were suggested for each electrode condition.

Results

The highest classification accuracies of 74.04% ± 4.82, 82.43% ± 6.14, 86.28% ± 2.81, and 86.85% ± 4.97 were achieved for the optimal two-, four-, six-, and eight-electrode configurations, respectively, which demonstrated the possibility of precise machine-learning-based diagnosis of MCI with a limited number of EEG electrodes. Additionally, further simulations with the EEG dataset revealed that the optimal electrode configurations had significantly higher classification accuracies than commercial EEG devices with the same number of electrodes, which suggested the importance of electrode configuration optimization for wearable EEG devices based on clinical EEG datasets.

Conclusions

This study highlighted that the optimization of the electrode configuration, assuming the wearable EEG devices can potentially be utilized for daily life monitoring of MCI, is necessary to enhance the performance and portability.

Introduction

Mild cognitive impairment (MCI) is a psychiatric syndrome characterized by a cognitive decline greater than that expected based on an individual’s age and education level [1]. MCI is believed to be associated with several underlying causes, especially for Alzheimer’s disease (AD) [2]. As approximately 40–60% of patients with MCI have an underlying AD pathology [3, 4], early detection of MCI is important for postponing cognitive decline and preventing its conversion to AD [5]. In general, however, a patient’s visit to the hospital for the diagnosis of MCI is considerably delayed because MCI does not usually interfere with daily activities, and individual symptoms are heterogeneous depending on the etiology and cognitive reserve [6, 7]. Moreover, conventional methods for diagnosing MCI such as history recording, clinical questionnaires, and simple cognitive tests are susceptible to misdiagnosis particularly in the early stages of MCI [8]. The exploration of all cognitive domains and quantification of overall cognitive performance are essential for avoiding potential diagnostic errors. However, such a time-consuming and tedious diagnostic procedure might act as a barrier preventing the early detection of MCI by discouraging potential patients from visiting a medical clinic [7].

Various biomarkers have been suggested for a more quantitative and objective diagnosis of MCI such as those extracted from cerebrospinal fluid (CSF), magnetic resonance imaging (MRI), and electroencephalography (EEG) [7, 9,10,11]. Recently, biomarkers from EEG have drawn increased attention owing to their advantages of relatively shorter diagnostic time and higher cost-effectiveness [12,13,14,15]. The key characteristics of the EEG data of elderly patients with MCI compared with those of normal elderly people (without MCI) are reduced complexity, decrease in inter-regional synchronizations, and shifts in the power spectrum from high-frequency components (alpha, beta, and gamma bands) toward low-frequency components (delta and theta bands) [16, 17]. These EEG biomarkers have been widely employed for the computer-aided diagnosis of MCI based on machine learning (ML) algorithms, thus demonstrating the possibility of a reliably high-performance EEG-based diagnosis of MCI [17,18,19,20,21]. For example, Morabito et al. [20] achieved an accuracy of 85% in classifying MCI and healthy control subjects (HCs) using a convolutional neural network (CNN) model. Fiscon et al. [21] achieved a classification accuracy of 92% for the detection of MCI using a decision tree classifier with wavelet coefficients from the EEG data.

However, previous studies on machine learning-based MCI diagnoses with EEG have generally employed research-grade EEG devices that require the aid of a trained experimenter so that the wider utilization of EEG-based MCI diagnosis in places other than laboratory or clinical environments is prevented [22,23,24]. Recent developments in wearable EEG technology have made EEG devices portable and inexpensive, thereby expanding their application fields [24,25,26]. For example, wearable EEG devices have the potential to be used for the early detection of neurological diseases [27, 28]. With the recent developments of hydrogel electrode technologies, it is expected that wearable EEG devices would be widely employed for continuous, long-term EEG monitoring and daily-life diagnosis of neurological diseases [29]. In this regard, these devices could be effective tools in the primary screening of MCI, for which early diagnosis is generally difficult due to the delayed visit to the hospital.

To increase the effectiveness of wearable EEG devices, it is necessary to increase the accessibility of the patients by making the devices portable and affordable [30]. Although attempts have been made to reduce the number of EEG channels while preserving the performance of the devices [31] and to investigate the optimal channel configurations that maximize the performance in brain-computer interfaces and biometric systems [32, 33], to the best of our knowledge, no study was performed to investigate the optimal electrode configurations for wearable EEG devices to maximize the accuracy of the machine learning-based diagnosis of MCI.

In this paper, we present a procedure to determine the optimal electrode configurations for wearable EEG devices to diagnose MCI, considering both the practicality and classification accuracy of wearable EEG devices. We suggest optimal channel configurations for two-, four-, six-, and eight-electrode conditions using a clinical EEG dataset collected from patients with MCI and healthy individuals. Furthermore, we demonstrate that the proposed optimal electrode configurations exhibit statistically higher performance in the diagnosis of MCI than commercial wearable EEG devices.

Methods and materials

Determination of optimal electrode configuration

To determine the optimal electrode configurations for different numbers of electrodes, we only considered configurations consisting of electrode pairs positioned symmetrically with respect to the midline (e.g., F7–F8 and FC5–FC6) and electrodes located on the midline (e.g., Fz and Pz), because it was thought that wearable EEG devices with bilateral symmetry are much easier to implement than configurations with randomly distributed electrodes [27]. In other words, the optimal electrode configurations were determined by comparing the classification accuracies of all possible electrode configurations that can be formed with the combinations of midline electrodes (Fz, Cz, Pz, and Oz) and symmetric electrode pairs (Fp1–Fp2, AF3–AF4, F7–F8, F3–F4, FC5–FC6, FC1–FC2, T7–T8, C3–C4, CP5–CP6, CP1–CP2, P7–P8, P3–P4, PO3–PO4, and O1–O2) when the number of electrodes was set to two, four, six, and eight. The number of all possible electrode combinations for the two-, four-, six-, and eight-electrode configurations were 20, 176, 924, and 3276, respectively.

The optimal electrode configuration was determined by selecting the electrode configuration that exhibited the highest accuracy among all possible channel combinations. After determining the optimal electrode configurations with two, four, six, and eight electrodes, the classification accuracies of the optimal electrode configurations were compared with those of commercial wearable EEG devices including FocusbandTM (T2 Green Pty Ltd.; Carrara, QLD, Australia), InsightTM (Emotiv Inc.; San Francisco, CA, USA), DSI-7TM (Wearable Sensing LLC; San Diego, CA, USA), ImecTM (Imec Inc.; Leuven, FB, Belgium), and EPOCTM (Emotiv Inc.; San Francisco, CA, USA). The electrode configurations of commercial devices are shown in Fig. 1.

Fig. 1
figure 1

Electrode configurations of commercial wearable EEG devices: a FocusbandTM (2ch), b InsightTM (4ch), c DSI-7TM (7ch), d ImecTM (8ch), and e EPOCTM (14ch)

Full size image

Participants

EEG data were acquired from 42 participants (21 patients with MCI and 21 healthy controls [HCs]) at the Inje University Ilsan Paik Hospital. The diagnosis of MCI was based on a clinical evaluation by trained psychiatrists who used the Structured Clinical Interview for DSM-IV (or V) Axis I Disorders (SCID-I) or the Mini International Neuropsychiatric Interview (MINI). None of the patients who participated in the study were pregnant or had any of other neurological or comorbid disorders, organic brain damage, or impairments in sensory or motor functions.

A total of 21 healthy participants were recruited from the local community. They did not satisfy the DSM-IV or V-based lifetime diagnostic criteria for any major psychiatric disorder as screened by the SCID-I Non-Patient Edition (SCID-NP) or MINI-based diagnostic criteria.

Demographic information, including sex, age, and education, was compared between the patients with MCI and HCs. A measurement of educational level is based on the duration of formal education. There were no statistically significant differences between the two groups in terms of demographic characteristics, including sex, age, and education. The detailed demographic information is presented in Table 1. All participants signed a written informed consent form approved by the Inje University Ilsan Paik Hospital Institutional Review Board (IRB no. 2018-12-012-013).

Table 1 Demographic information of the MCI and HC groups
Full size table

Signal acquisition and preprocessing

The EEG signal was acquired at a rate of 1000 Hz using a SynAmps amplifier (Neuroscan, Compumedics USA, Charlotte, NC, USA) from 32 scalp electrodes (Fp1, Fp2, AF3, AF4, Fz, F3, F4, F7, F8, FC1, FC2, FC5, FC6, Cz, C3, C4, T7, T8, CP1, CP2, CP5, CP6, Pz, P3, P4, P7, P8, PO3, PO4, Oz, O1, and O2) according to the modified international 10–20 system. The impedance of each electrode was maintained below 5 kΩ throughout the experimental period. Ground and reference electrodes were placed on the forehead and mastoids.

The resting EEG was recorded for 4 min with eyes closed. The acquired EEG data were manually inspected to eliminate blocks contaminated by environmental or physiological noise. Then the EEG data were baseline-corrected by subtracting the average value for each channel and band-pass filtered at cut-off frequencies of 0.5 Hz and 50 Hz using a 6th-order zero-phase Butterworth infinite impulse response filter. Thereafter, the preprocessed resting EEG data were segmented into 5-s epochs without an overlap. We rejected epochs in which the maximal absolute potential value exceeded the threshold of 75 μV; therefore, the numbers of remaining epochs were different among subjects, with twenty being the minimal number of epochs. To equalize the number of epochs for each subject, twenty epochs were randomly selected for each subject.

Feature extraction

The absolute power spectrum density (APSD), relative power spectrum density (RPSD), differential asymmetry (DASM), rational asymmetry (RASM), phase-amplitude coupling (PAC), Shannon entropy (SE), Hjorth parameters (HP), Lyapunov exponent (LE), Hurst exponent (HE), and Kolmogorov complexity (KC) were extracted as candidate features for MCI diagnosis. These candidate features have been used as biomarkers in previous studies on EEG-based diagnosis of MCI or Alzheimer’s disease (AD) [17, 34,35,36,37]. The equations used to compute these features are summarized in Table 2, and their derivation process and detailed description are provided in the Supplementary Material.

Table 2 List of EEG features evaluated in this study
Full size table

Spectral features were calculated for each of the following nine sub-frequency bands: delta (δ, 1–4 Hz), theta (θ, 4–8 Hz), low alpha (αL, 8–10 Hz), high alpha (αH, 10–12 Hz), total alpha (α, 8–12 Hz), low beta (βL, 12–18 Hz), high beta (βH, 18–30 Hz), total beta (β, 12–30 Hz), and gamma (β, 30–50 Hz).

For the PAC, which evaluates the coherence between the low-frequency phase and high-frequency amplitude, the low-frequency signal was set as either δ or θ, while the high-frequency signal was set as one of the other seven sub-frequency bands (except δ or θ). The total number of candidate features extracted from the 32 channels was 1500 consisting of {9 (APSD) + 9 (RPSD) + 14 (PAC) + 3 (HP) + 1 (SE) + 1 (LE) + 1 (HE) + 1(KC)} × 32(channels) + {9 (DASM) + 9 (RASM)} × 14 (pairs). Each candidate feature was averaged over all twenty epochs per channel, resulting in 1500 averaged candidate features per participant.

Feature selection and classification

A support vector machine (SVM) classifier was employed using the statistics and machine learning toolbox in MATLAB 2018b (MathWorks, Natick, MA, USA). Leave-pair-out (LPO) cross-validation (CV) was conducted to evaluate the classification accuracy of the model [44]. In detail, the data of two participants (one patient with MCI and one from the HC group) were used as the test set, and the data of the remaining 40 participants (20 patients with MCI and 20 HCs) were used as the training set. Accuracy, sensitivity, and specificity were calculated and averaged over all possible combinations of the two participants (one MCI and one HC; 21 × 21 = 441).

In each iteration of LPO-CV, z-score normalization was applied to each feature of the training set [45]:

$$z=\frac{X-\overline{X}}{\sigma },$$
(1)

where \(\overline{X}\) and σ denote mean and standard deviation, respectively. The features of the test set were normalized using (1) with the \(\overline{X}\) and σ calculated from the training set. The optimal feature subset was then selected from the training set based on the rank order according to Fisher’s score, which is one of the most widely used filter methods for supervised feature selection [46]. The feature selection was performed with the candidate features extracted from the target electrode configuration. The maximum number of selected features was set to 15 to avoid potential overfitting. The “N-feature accuracy” was evaluated by averaging the results of all LPO-CV iterations for N features, where N represents the number of features used for the classification, which ranged from 1 to 15. The classification accuracy for each electrode configuration was determined as the highest accuracy among the “N-feature accuracies.” Note that the highest accuracy was achieved when the number of features was less than 10 in most cases.

Statistical analysis

Statistical analysis was conducted to investigate the differences in demographic information between the MCI and HC groups and differences in the classification accuracies between the proposed optimal electrode configurations and those of commercial wearable EEG devices. The chi-square test [47] was conducted to test the difference in sex composition between the MCI and HC groups. A two-tailed Student’s t-test [48] was conducted to identify the differences in age and education between the MCI and HC groups as the normality of the data was confirmed by the one-sample Kolmogorov–Smirnov test [49]. In addition, the Bonferroni-corrected Wilcoxon signed-rank test [50, 51] was conducted to identify the difference between the classification accuracies for the proposed electrode configurations and those of the commercial wearable EEG devices as the normality of the data was not confirmed by the one-sample Kolmogorov–Smirnov test.

Results

The optimal electrode configurations that resulted in the highest classification accuracy among all possible channel combinations are presented in Table 3 and Fig. 2a. The optimal two-channel (hereafter denoted by Opt-2ch), four-channel (Opt-4ch), six-channel (Opt-6ch), and eight-channel (Opt-8ch) electrode configurations were “F3–F4,” “AF3–AF4–FC5–FC6,” “FC5–FC6–C3–C4–P7–P8,” and “F3–F4–FC5–FC6–C3–C4–P7–P8,” respectively. All optimal electrode configurations were composed of combinations of the following five electrode pairs: AF3–AF4, F3–F4, FC5–FC6, C3–C4, and P7–P8, which were in the frontal and parietal areas. Notably, Opt-8ch was a combination of Opt-2ch and Opt-6ch, and no midline electrode was included in the optimal electrode configuration.

Table 3 Channel combination and diagnostic performances of the optimal electrode configurations and those of five commercial wearable EEG devices
Full size table
Fig. 2
figure 2

a Proposed optimal electrode configurations consisting of two, four, six, and eight channels (from top to bottom). b Performance comparison between the proposed optimal electrode configurations and the electrode configurations of commercial wearable EEG devices. The x-axis of the bar graphs represents the electrode configurations. The error bars indicate standard error. * Bonferroni-corrected p < 0.05, ** Bonferroni-corrected p < 0.001, *** Bonferroni-corrected p < 0.0001

Full size image

The accuracy, sensitivity, and specificity of the machine learning-based MCI diagnosis using optimal electrode configurations are summarized in Table 3. The numbers of features that resulted in the highest accuracies were 2, 2, 6, and 7 for Opt-2ch, Opt-4ch, Opt-6ch, and Opt-8ch, respectively. Accuracies of 74.04% ± 4.82, 82.43% ± 6.14, 86.28% ± 2.81, and 86.85% ± 4.97 were achieved for Opt-2ch, Opt-4ch, Opt-6ch, and Opt-8ch, respectively, which demonstrated the possibility of precise machine learning-based diagnosis of MCI with a limited number of EEG electrodes. It can be seen from the table that the classification accuracy gradually increases as the number of electrodes increases; however, the increment in the accuracy is nearly saturated when the number of electrodes becomes six. Furthermore, while the sensitivity was approximately 20% lower than the specificity for Opt-2ch and Opt-4ch, the difference between the sensitivity and specificity was reduced to less than 1% for Opt-6ch and Opt-8ch.

The accuracy, sensitivity, and specificity of the machine learning-based MCI diagnosis using the electrode configurations of the five commercial wearable EEG devices are provided in Table 3. A statistical comparison of the classification accuracies between the four optimal electrode configurations and the five wearable EEG devices is shown in Fig. 2b. Opt-2ch statistically outperformed the electrode configurations of three commercial devices (FocusbandTM, InsightTM, and DSI-7TM, Bonferroni-corrected p < 0.001) that were composed of two, five, and seven electrodes, respectively. Opt-2ch did not show significant improvement compared to ImecTM (Bonferroni-corrected p = 0.94) and EPOCTM (Bonferroni-corrected p = 0.88), which are composed of eight and 14 electrodes, respectively. The other optimal electrode configurations (Opt-4ch, Opt-6ch, and Opt-8ch) statistically outperformed the electrode configurations of all the commercial wearable EEG devices considered in this study (Bonferroni-corrected p < 0.001 in all cases).

Figure 3 shows the 3D-rendered concept design of wearable EEG devices with the proposed optimal electrode configuration. 3D images were rendered using Maya 2022 (Autodesk Inc., San Rafael, CA, USA). The overall design concept was inspired by EPOCTM (Emotiv Inc.; San Francisco, CA, USA). In addition, the devices were assumed to have fingered EEG electrodes, which reflected the latest trend in electrode design for wearable EEG devices [24].

Fig. 3
figure 3

3D-rendered images of wearable EEG devices with the electrode configurations proposed in this study

Full size image

Discussion

In this study, we attempted to determine the optimal electrode configurations that could lead to the highest accuracy in the machine-learning-based diagnosis of MCI. The electrode configurations investigated in this study were composed of a small number of electrodes (two, four, six, and eight electrodes), assuming wearable EEG devices can potentially be utilized for daily life monitoring of MCI. The proposed optimal electrode configurations showed statistically higher accuracies than the electrode configurations of commercial wearable EEG devices, albeit with a smaller number of electrodes. For example, Opt-4ch showed a higher diagnostic accuracy than the electrode configurations of all five commercial wearable EEG devices tested in this study, which highlighted the importance of optimizing the electrode configurations in the design of wearable EEG devices considering the specific purpose of the wearable EEG devices.

We investigated the most frequently selected features in the LPO-CV iterations to identify the EEG features that contributed the most to distinguishing between patients with MCI and HCs. Table 4 shows the four most frequently selected features and their mean values for the MCI and HC groups. The four features were δ DASM between FC5 and FC6, the HE in F3, the HE in C4, and PAC (θ-βH) in P8, which were selected at least 397 times out of 441 LPO-CV iterations (approximately 90% of all LPO-CV iterations) for Opt-8ch (see Supplementary Fig. 1). These EEG features have been found to be closely associated with the pathophysiology of MCI in previous studies.

Table 4 Comparisons of the most frequently selected feature values between the MCI and HC groups
Full size table

First, the value of δDASM between FC5 and FC6 in patients with MCI was statistically higher than that in the HCs. This result is in line with a previous study that reported that patients with MCI exhibited a significantly higher delta-band power in the left central area than HCs; however, this was not the case for the right central area [52]. Additionally, the HE in F3 and C4 was significantly higher in the MCI group than in the HC group in our study. These results are also in good agreement with those of previous studies. According to John et al. [53], the HE of patients with MCI or AD was higher than that of the HCs in whole-brain areas when the HE values were calculated from eyes-closed resting EEG data, as in our study. Lastly, the PAC (θ-βH) at P8 was lower in patients with MCI than in the HCs in our study. This result is also consistent with that of Poza et al. [54] reported that patients with mild AD exhibited lower PAC values in the posterior area than HCs.

Although the proposed optimal electrode configurations exhibited statistically higher performance than the electrode configurations of commercial wearable EEG devices, their sensitivities were reported to be approximately 20% lower than the specificities in cases of Opt-2ch and Opt-4ch. In general, well-balanced sensitivity and specificity values are important in designing diagnostic systems. However, as wearable EEG devices are expected to be employed as tools for the primary screening of MCI, an EEG device with higher diagnostic sensitivity is required. Therefore, these optimal electrode configurations could be alternatively determined by selecting the electrode combinations with the highest sensitivity with the overall accuracy maintained at an appropriate level. For example, an electrode pair “C3–C4” showed a classification accuracy of 69.16% ± 7.55, which is approximately 5%p lower than that of Opt-2ch; however, it showed a relatively higher sensitivity of 74.60% ± 11.30 and a moderate level of specificity (63.72% ± 9.58) (see Supplementary Table 1). In the case of the 4-channel configuration, “F3–F4–FC5–FC6” showed an approximately 8% higher sensitivity than Opt-4ch when the difference in accuracy between the two configurations was only 2.5%p (see Supplementary Table 2). If sensitivity is regarded as a more important factor than specificity in designing wearable EEG devices for the primary screening of MCI, the “C3–C4” and “F3–F4–FC5–FC6” configurations might be considered possible alternatives to Opt-2ch and Opt-4ch, respectively. The top 20 electrode configurations for six- and eight-electrode cases can also be found in Supplementary Tables 3 and 4.

The optimal electrode configurations for the diagnosis of MCI may vary depending on several factors, such as the participants and recording devices. Because most EEG features have large inter-subject and/or inter-device variability [55], an additional EEG dataset acquired from a larger number of participants using various EEG devices might be necessary in future studies to further generalize the proposed electrode configurations. It is expected that the optimal electrode configurations will be fine-tuned based on an additional EEG dataset. Additionally, we are planning to manufacture a wearable EEG device with the proposed optimal electrode configuration and investigate in a future study whether the device is effective enough to be used for the early screening of MCI.

Availability of data and materials

The data used in this study cannot be open to the public according to the Institutional Review Board approval.

References

  1. Gauthier S, et al. Mild cognitive impairment. Lancet. 2006;367(9518):1262–70.

    Article  PubMed  Google Scholar 

  2. Ieracitano C, et al. A novel multi-modal machine learning based approach for automatic classification of EEG recordings in dementia. Neural Netw. 2020;123:176–90.

    Article  PubMed  Google Scholar 

  3. World Health Organization. Meeting on the Implementation of the global action plan of the public health response on dementia 2017-2025: meeting report: 11-12 December 2017. Geneva: World Health Organization; 2018.

  4. Luck T, et al. Incidence of mild cognitive impairment: a systematic review. Dementia Geriatr Cognitive Disord. 2010;29(2):164–75.

    Article  Google Scholar 

  5. Chen W, Wang H. Mild cognitive impairment: a concept useful for early detection and intervention of dementia. Shanghai Arch Psychiatry. 2013;25(2):119.

    PubMed  PubMed Central  Google Scholar 

  6. Pedrosa H, et al. Functional evaluation distinguishes MCI patients from healthy elderly people—the ADCS/MCI/ADL scale. J Nutr Health Aging. 2010;14(8):703–9.

    Article  CAS  PubMed  Google Scholar 

  7. Sabbagh MN, et al. Early detection of mild cognitive impairment (MCI) in primary care. J Prev Alzheimer's Dis. 2020;7(3):165–70.

    CAS  Google Scholar 

  8. Edmonds EC, et al. “Missed” mild cognitive impairment: High false-negative error rate based on conventional diagnostic criteria. J Alzheimer’s Dis. 2016;52(2):685–91.

    Article  Google Scholar 

  9. Roberts R, Knopman DS. Classification and epidemiology of MCI. Clin Geriatr Med. 2013;29(4):753–72.

    Article  PubMed  Google Scholar 

  10. Stomrud E, et al. Slowing of EEG correlates with CSF biomarkers and reduced cognitive speed in elderly with normal cognition over 4 years. Neurobiol Aging. 2010;31(2):215–23.

    Article  CAS  PubMed  Google Scholar 

  11. Vemuri P, et al. MRI and CSF biomarkers in normal, MCI, and AD subjects: predicting future clinical change. Neurology. 2009;73(4):294–301.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  12. Stam C, et al. EEG synchronization in mild cognitive impairment and Alzheimer's disease. Acta Neurol Scand. 2003;108(2):90–6.

    Article  CAS  PubMed  Google Scholar 

  13. König T, et al. Decreased EEG synchronization in Alzheimer’s disease and mild cognitive impairment. Neurobiol Aging. 2005;26(2):165–71.

    Article  Google Scholar 

  14. Baker M, et al. EEG patterns in mild cognitive impairment (MCI) patients. Open Neuroimaging J. 2008;2:52.

    Article  Google Scholar 

  15. Sharma N, et al. EEG and cognitive biomarkers based mild cognitive impairment diagnosis. IRBM. 2019;40(2):113–21.

    Article  Google Scholar 

  16. Poil S-S, et al. Integrative EEG biomarkers predict progression to Alzheimer's disease at the MCI stage. Front Aging Neurosci. 2013;5:58.

    Article  PubMed  PubMed Central  Google Scholar 

  17. Cassani, Raymundo, et al. Systematic review on resting-state EEG for Alzheimer's disease diagnosis and progression assessment. Dis Markers. 2018;2018:5174815.

  18. Trambaiolli LR, et al. Improving Alzheimer's disease diagnosis with machine learning techniques. Clin EEG Neurosci. 2011;42(3):160–5.

    Article  PubMed  Google Scholar 

  19. Kashefpoor M, Rabbani H, Barekatain M. Automatic diagnosis of mild cognitive impairment using electroencephalogram spectral features. J Med Sign Sensors. 2016;6(1):25.

    Article  Google Scholar 

  20. Morabito, Francesco Carlo, et al.  Deep convolutional neural networks for classification of mild cognitive impaired and Alzheimer's disease patients from scalp EEG recordings. In: 2016 IEEE 2nd International Forum on Research and Technologies for Society and Industry Leveraging a better tomorrow (RTSI). IEEE; 2016. p. 1-6.

  21. Fiscon G, et al. Combining EEG signal processing with supervised methods for Alzheimer’s patients classification. BMC Med Inform Decis Making. 2018;18(1):1–10.

    Article  Google Scholar 

  22. Casson AJ, et al. Wearable electroencephalography. IEEE Eng Med Biol Magazine. 2010;29(3):44–56.

    Article  Google Scholar 

  23. Duvinage M, et al. A P300-based quantitative comparison between the Emotiv Epoc headset and a medical EEG device. Biomed Eng. 2012;765(1):2012–764.

    Google Scholar 

  24. Casson AJ. Wearable EEG and beyond. Biomed Eng letters. 2019;9(1):53–71.

    Article  Google Scholar 

  25. Debener S, et al. How about taking a low-cost, small, and wireless EEG for a walk? Psychophysiology. 2012;49(11):1617–21.

    Article  PubMed  Google Scholar 

  26. Mihajlović V, et al. Wearable, wireless EEG solutions in daily life applications: what are we missing? IEEE J Biomed Health Inform. 2014;19(1):6–21.

    Article  PubMed  Google Scholar 

  27. Gaidar V, Sudakov O. Design of wearable EEG device for seizures early detection. Int J Electron Telecommun. 2021;67(2):187–92.

    Google Scholar 

  28. Samuel N, et al. Consumer-grade electroencephalography devices as potential tools for early detection of brain tumors. BMC Medicine. 2021;19(1):1–3.

    Article  Google Scholar 

  29. Hsieh, Ju-Chun, et al. Design of hydrogel-based wearable EEG electrodes for medical applications. J Materials Chem B. 2022;10:7260–80.

  30. Ingolfsson, Thorir Mar, et al. Towards long-term non-invasive monitoring for epilepsy via wearable eeg devices. In: 2021 IEEE Biomedical Circuits and Systems Conference (BioCAS). IEEE; 2021. p. 01–04.

  31. Arvaneh, Mahnaz, et al. Robust EEG channel selection across sessions in brain-computer interface involving stroke patients. In: The 2012 International Joint Conference on Neural Networks (IJCNN). IEEE; 2012. p. 1–6.

  32. Park S, Han C-H, Im C-H. Design of wearable EEG devices specialized for passive brain–computer interface applications. Sensors. 2020;20(16):4572.

    Article  PubMed Central  Google Scholar 

  33. Abdullah, Muhammad Kamil, et al. Analysis of effective channel placement for an EEG-based biometric system. In: 2010 IEEE EMBS Conference on Biomedical Engineering and Sciences (IECBES). IEEE; 2010. p. 303–306.

  34. McBride JC, et al. Spectral and complexity analysis of scalp EEG characteristics for mild cognitive impairment and early Alzheimer's disease. Comp Methods Prog Biomed. 2014;114(2):153–63.

    Article  Google Scholar 

  35. Dimitriadis SI, et al. A novel biomarker of amnestic MCI based on dynamic cross-frequency coupling patterns during cognitive brain responses. Front Neurosci. 2015;9:350.

    Article  PubMed  PubMed Central  Google Scholar 

  36. Houmani N, Dreyfus G, Vialatte FB. Epoch-based entropy for early screening of Alzheimer’s disease. Int J Neural Syst. 2015;25(08):1550032.

    Article  CAS  PubMed  Google Scholar 

  37. Garn H, et al. Differential diagnosis between patients with probable Alzheimer’s disease, Parkinson’s disease dementia, or dementia with Lewy bodies and frontotemporal dementia, behavioral variant, using quantitative electroencephalographic features. J Neural Trans. 2017;124(5):569–81.

    Article  CAS  Google Scholar 

  38. Osipova D, Hermes D, Jensen O. Gamma power is phase-locked to posterior alpha activity. PloS One. 2008;3(12):e3990.

    Article  PubMed  PubMed Central  Google Scholar 

  39. Shannon CE. A mathematical theory of communication. Bell Syst Tech J. 1948;27(3):379–423.

    Article  Google Scholar 

  40. Hjorth B. The physical significance of time domain descriptors in EEG analysis. Electroencephalography Clin Neurophysiol. 1973;34(3):321–5.

    Article  CAS  Google Scholar 

  41. Rosenstein MT, Collins JJ, De Luca CJ. A practical method for calculating largest Lyapunov exponents from small data sets. Physica D Nonlinear Phenomena. 1993;65(1-2):117–34.

    Article  Google Scholar 

  42. Feder J. Fractals. New York: Plenum Press; 2013.

  43. Kolmogorov AN. Three approaches to the definition of the concept “quantity of information”. Probl Peredachi Informatsii. 1965;1(1):3–11.

    Google Scholar 

  44. Smith GC, et al. Correcting for optimistic prediction in small data sets. Am J Epidemiol. 2014;180(3):318–24.

    Article  PubMed  PubMed Central  Google Scholar 

  45. Jain A, Nandakumar K, Ross A. Score normalization in multimodal biometric systems. Pattern Recogn. 2005;38(12):2270–85.

    Article  Google Scholar 

  46. Gu Q, Li Z, Han J. Generalized fisher score for feature selection. arXiv preprint arXiv:1202.3725. 2012.

  47. McHugh ML. The chi-square test of independence. Biochem Med. 2013;23(2):143–9.

    Article  CAS  Google Scholar 

  48. Hsu H, Lachenbruch PA. Lachenbruch. Paired t test. Wiley StatsRef: statistics reference online. 2014.

  49. Lilliefors HW. On the Kolmogorov-Smirnov test for normality with mean and variance unknown. J Am Stat Assoc. 1967;62(318):399–402.

    Article  Google Scholar 

  50. Woolson RF. Wilcoxon signed-rank test. Wiley encyclopedia of clinical trials. 2007. p. 1–3.

  51. Bland JM, Altman DG. Multiple significance tests: the Bonferroni method. Bmj. 1995;310(6973):170.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  52. Kwak YT. Quantitative EEG findings in different stages of Alzheimer’s disease. J Clin Neurophysiol. 2006;23(5):457–62.

    Article  Google Scholar 

  53. John TN, Puthankattil SD, Menon R. Analysis of long range dependence in the EEG signals of Alzheimer patients. Cogn Neurodynam. 2018;12(2):183–99.

    Article  Google Scholar 

  54. Poza J, et al. Phase-amplitude coupling analysis of spontaneous EEG activity in Alzheimer's disease. In: 2017 39th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC). IEEE; 2017. p. 2259–2262.

  55. Takeuchi I, et al. Transdermal delivery of estradiol-loaded PLGA nanoparticles using iontophoresis for treatment of osteoporosis. Bio Med Mater Eng. 2016;27(5):475–83.

    Article  CAS  Google Scholar 

Download references

Funding

This work was supported in part by a National Research Foundation of Korea (NRF) grant funded by the Korean government (MSIT) (No. NRF-2019R1A2C2086593) and a Korea Medical Device Development Fund grant funded by the Korean government (No. 202013B10).

Author information

Authors and Affiliations

  1. School of Electronic Engineering, Hanyang University, Seoul, Republic of Korea

    Kyeonggu Lee, Kang-Min Choi, Seonghun Park & Chang-Hwan Im

  2. Department of Psychiatry, Inje University Ilsan Paik Hospital, Goyang, Republic of Korea

    Seung-Hwan Lee

  3. Clinical Emotion and Cognition Research Laboratory, Inje University Ilsan Paik Hospital, Goyang, Republic of Korea

    Seung-Hwan Lee

  4. School of Biomedical Engineering, Hanyang University, Seoul, Republic of Korea

    Chang-Hwan Im

Authors
  1. Kyeonggu Lee
    View author publications

    You can also search for this author in PubMed Google Scholar

  2. Kang-Min Choi
    View author publications

    You can also search for this author in PubMed Google Scholar

  3. Seonghun Park
    View author publications

    You can also search for this author in PubMed Google Scholar

  4. Seung-Hwan Lee
    View author publications

    You can also search for this author in PubMed Google Scholar

  5. Chang-Hwan Im
    View author publications

    You can also search for this author in PubMed Google Scholar

Contributions

C. H. Im conceived the study concept and designed the study. S. H. Lee provided the data. K. G. Lee constructed the machine learning designs, conducted statistical analyses, and drafted the manuscript. K.G. Lee and K.M. Choi contributed to the preprocessing of data. S. H. Park and K. G. Lee extracted the features from the data. All the authors have revised the manuscript accordingly. The author(s) read and approved the final manuscript.

Corresponding author

Correspondence to Chang-Hwan Im.

Ethics declarations

Ethics approval and consent to participate

The study protocol was approved by the Inje University Ilsan Paik Hospital Institutional Review Board (IRB no. 2018-12-012-013) and the study was conducted in accordance with the Declaration of Helsinki. All participants signed a written informed consent form before their participation in the study.

Consent for publication

Not applicable

Competing interests

The authors declare that they have no competing interests.

Additional information

Publisher’s Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary Information

Additional file 1: Supplementary Table 1.

Performances of the top 10 electrode configurations composed of two electrodes. The table lists the calculated accuracies, sensitivities, and specificities for the top 10 electrode configurations with the highest accuracies. These values were used as performance measures. Supplementary Table 2. Performances of the top 10 electrode configurations composed of four electrodes. The table lists the calculated accuracies, sensitivities, and specificities for the top 10 electrode configurations with the highest accuracies. These values were used as performance measures. Supplementary Table 3. Performances of the top 20 electrode configurations composed of six electrodes. The table lists the calculated accuracies, sensitivities, and specificities for the top 20 electrode configurations with the highest accuracies. These values were used as performance measures. Supplementary Table 4. Performances of the top 20 electrode configurations composed of eight electrodes. The table lists the calculated accuracies, sensitivities, and specificities for the top 20 electrode configurations with the highest accuracies. These values were used as performance measures. Supplementary Figure 1. List of the most frequently selected features. The features were most frequently selected in all CV iterations for the number of features that yielded the highest accuracy in each optimal electrode configuration. From top to bottom, the feature list of the Opt-2ch, Opt-4ch, Opt6ch, and Opt-8ch configurations are described. The x-axis is “oftenness,” which is calculated by dividing the feature selection times by all CV iteration times (441).

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Lee, K., Choi, KM., Park, S. et al. Selection of the optimal channel configuration for implementing wearable EEG devices for the diagnosis of mild cognitive impairment. Alz Res Therapy 14, 170 (2022). https://doi.org/10.1186/s13195-022-01115-3

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI: https://doi.org/10.1186/s13195-022-01115-3

Keywords

Alzheimer's Research & Therapy

ISSN: 1758-9193