Detection of Cognitive Decline using a Single-Channel EEG with an Interactive Assessment Tool

Cognitive decline remains highly underdiagnosed despite efforts to nd novel biomarkers for detection. EEG biomarkers based on machine learning may offer a noninvasive low-coast approach for identifying cognitive decline. However, most studies use multi-electrode systems which are less accessible. This study aims to evaluate the ability to extract cognitive decline biomarkers using a wearable single-channel EEG system with an interactive assessment tool. Methods This pilot study included data collection from 82 participants who performed a cognitive assessment while being recorded with a single-channel EEG system. Seniors in different clinical stages of cognitive decline (healthy to mild dementia) and young healthy participants were included. Seniors’ MMSE scores were used to allocate groups with cutoff scores of 24 and 27. Data analysis included correlation analysis as well as linear mixed model analysis with several EEG variables including frequency bands and three novel cognitive biomarkers previously extracted from a different dataset.

later used in classi cation models to identify subjects with cognitive impairments. The results contribute to the idea that a single channel EEG can be used as an e cient and convenient way for detecting cognitive decline. However, the risk of over tting the data in such classi cation studies should be addressed to ensure generalization capabilities, especially with a small sample size. Studies that use the same dataset for training as well as feature extraction [20]- [22], extend the risk of over tting. For generalization of the data, the features should be examined on different datasets and provide consistency in the results on new datasets. Furthermore, measuring the correlations of the extracted features with standard clinical measurements (like the MMSE score) or behavioral results of cognitive tests (like reaction times and accuracy) is highly valuable for further validation of such novel features.
In this study we evaluated the ability of an easy-to-use single-channel EEG system with an interactive cognitive assessment to potentially detect cognitive decline in the elderly population. The system uses auditory stimuli, and extracts biomarkers using harmonic analysis and machine learning methods from the EEG signal. The pre-extracted biomarkers used in this study were validated in a previous study performed on young healthy subjects [23]. This pilot study aims to evaluate the ability of the system to extract cognitive decline biomarkers, recognizing the importance of providing an accurate low-cost alternative for cognitive decline detection.
Sixty patients from the inpatient rehabilitation department at Dorot Geriatric Medical Center were recruited to this study. For the full demographic details see Table 1. The overall mean age was 77.55 (9.67) years old. There was a wide range of ages for each group with no signi cant age difference between the groups. Participants consisted of 47% females and 53% males. Among the patients, 82% were hospitalized for orthopedic rehabilitation and 18% due to various other causes. Among the patients who had surgery, an average of 27 (16.3) days passed since the surgery. Potential subjects were identi ed by the clinical staff during their admission to the inpatient rehabilitation department. All subjects were hospitalized at the center and were chosen based on inclusion criteria speci ed in the study protocol. The patients undergo a Mini Mental State Examination (MMSE) by an Occupational Therapist upon hospital administration and this score was used for screening patients to include those that have scores between 10-30. All subjects were also evaluated for their ability to hear, read, and understand instructions for discussion of Informed Consent Form (ICF) as well as for the auditory task. Patients that speak English, Hebrew, and Russian were provided with the appropriate ICF and auditory task in the language they could read and understand. All participants provided ICF according to the guidelines outlined in the Declaration of Helsinki. Patients that showed any verbal or non-verbal form of objection were not included in the study. Other exclusion criteria included MMSE score lower than 10, presence of several comorbidities, damage to integrity of scalp and/or skull and skin irritation in the facial and forehead area, signi cant hearing impairments, and history of drug abuse.
In total, 50 of the 60 recruited patients, completed the auditory task and their EEG data was used. 10 patients signed the ICF and were included in the overall patient count but were excluded from data analysis due to their desire to stop the study or technical problems in the recording. The participants were divided into 3 groups according to the MMSE scores: 17 patients with a score of [17][18][19][20][21][22][23]  Apparatus EEG device EEG recordings were performed using a single channel EEG system ("Aurora" by Neurosteer Inc). A 3electrode medical-grade patch was placed on the subject's forehead using dry gel for optimal signal transduction. The electrodes were located at Fp1 and Fp2 and a reference electrode at Fpz. EEG signal was ampli ed by a factor of a 100 and sampled at 500 Hz. Signal processing was done in the Neurosteer cloud, for further details see Appendix A.

EEG Recording and Auditory battery
The recording room was quiet and illuminated. The research assistant set up the sanitized system equipment (electrode patch, sensor, EEG monitor, clicker) and provided general instructions to the participants before starting the task. Then the electrode was placed on the subject's forehead and the recording was initiated. The participant was sitting during the assessment and heard instructions through a loudspeaker connected to the EEG monitor. The entire recording session typically lasted 20-30 minutes.
The cognitive assessment battery was pre-recorded and included a detection task as well as answering a series of true/false questions by pressing on a wireless clicker button. Further explanations for the task were kept at a minimum to avoid bias. A few minutes of baseline activity were recorded to ensure an accurate test. The auditory cognitive assessment lasted 18 minutes. Detection Task Figure 1 illustrates the detection task used in the study. In each block, participants were presented with a sequence of melodies (played by a violin, a trumpet, and a ute). The participants were given a clicker to respond to the stimuli. In the beginning of each block, auditory instructions indicated an instrument to which the participant responded by clicking once. The click response was only to "yes" trials, when the indicated instrument melody had played. The task included two di culty levels to test increasing cognitive load. In level 1, each melody was played for 3 seconds, and the same melody repeated throughout the entire block. The participant was asked to click once as fast as possible for each repetition of the melody. In level 2, the melodies were played for 1.5 seconds, and all three instruments appeared in the block. The participants were asked to click only for a speci c instrument within the block and ignore the rest of the melodies.

Construction of Brain Activity Features and Classi ers
The signal processing algorithm interprets the EEG data using a time/frequency wavelet-packet analysis [24], [25], which characterizes different components by their fundamental frequency and corresponding higher harmonics. In addition, standard spectral analysis of the signal was used to produce the classical EEG frequency bands.
The optimal number of different features was determined on a large dataset using unsupervised Machine Learning techniques (i.e. ML extracting features on unlabeled data). This analysis resulted in 121 Brain Activity Features (BAFs). The BAFs activity can be regarded as a series of vectors with 121 values that are sampled every second. Combinatorically, there are millions of linear and non-linear combinations that can be created out of these long vectors. Prior to this examination, several such combinations were computed using ML. These combinations were calculated from data of healthy subjects performing different tasks in different di culty levels. The labelled data was used to train linear and non-linear classi ers that differentiate between different tasks. The value of each feature was calculated once every second from a moving window of 4 seconds. Three of those biomarkers were described in a previous study on healthy subjects performing a well-validated cognitive task (i.e. n-back task) [23]. The biomarkers exhibited separation between different levels of cognitive load and therefore were the most relevant for the present study. The full technical speci cations regarding the construction of these features and the extractions of the biomarkers are provided in Appendix A.

Dependent variables Behavioral measurements
The behavioral dependent variables included mean response accuracy and mean reaction times (RT) per participant, for correct responses only.

Statistical analyses
Statistical analyses were performed on data from 50 senior participants and 22 young participants.

Pearson correlations
Based on previous studies [3] we hypothesized that the Reaction Times (RTs) in the cognitive detection task would be higher for participants with lower MMSE scores. This was tested by calculating Pearson correlation between mean RTs in detection levels 1 and 2, and the individual MMSE score of each participant. RTs correlations to the EEG variables were also calculated.
Another correlation analysis was performed to nd which of the EEG variables correlated with the (previously assigned) MMSE score of each participant. For this purpose, Pearson correlation was calculated using the mean activity of the EEG variables during the detection task and each individual MMSE score.

Linear Mixed Models (LMM)
The independent variables included task as a within-participants variable (including detection 1, detection 2, resting state tasks) and group as a between-participants variable. Groups were allocated as follows (see Figure 2): data of the senior participants was divided into three groups according to MMSE scores: Mild Dementia (MD) group with MMSE scores between 17-24, Mild Cognitive Impairment Risk (MCI-R) group with MMSE scores between 24-27, and Healthy Seniors with MMSE scores between 28-30. We used MMSE score cutoffs of 24 and 27 in allocating the groups as we are mostly interested in detecting cognitive decline as early as possible and found previous indications that a higher cutoff score would achieve optimal evaluations of diagnostic accuracy [7]. Furthermore, it was argued that educated individuals who score below 27 are at greater risk of being diagnosed with dementia [26]. The last group included in the analysis was consist of the 22 young healthy participants (Healthy young group). Overall, 4 groups were taken into the analysis.
Since both independent variables have more than two levels, indicator variables (aka dummy variables) were computed. For the group variable, Healthy seniors group was determined as the reference level, and the three remaining levels were computed as dummy variables: Healthy young (e.g., the difference between Healthy seniors and Healthy young), MCI-R (e.g., the difference between Healthy seniors and MCI-R), and MD (e.g., the difference between Healthy seniors and MD). Accordingly, a signi cant effect in one of the groups would mean that the activity of the Healthy seniors group is signi cantly different than the activity of said group. For the task variable, detection 1 (which represents low cognitive load level) was determined as the reference level, and the two remaining levels were computed as dummy variables: Detection 2 (e.g., the difference between detection 1 and detection 2 which represents high cognitive load level) and Resting state (e.g., the difference between detection 1 and resting state). A signi cant effect of one of the levels, would mean that the activity during detection 1 level was signi cantly different than the activity during said level.
An LMM analysis was performed on each EEG variable separately (A0, ST4, VC9, Delta, Theta, Alpha, Beta, and Gamma). Due to the interpretation complexity of the effects, the rst analysis included the two variables (group and task level) without their interactions. Signi cance level for this analysis was p<0.05.
For models that achieved signi cant effects in the task variable, follow-up models were calculated to examine the activity level differences within each group. Therefore, the interaction between task level and group was examined within the senior groups and the young healthy group separately. For these post-hoc models (measuring the simple effects of task within each study group), signi cance level resulted in the LMM model was corrected with Bonferroni correction to p<0.025 (in accordance with two comparisons).

Validation of behavioral task
Correlations between individual MMSE scores and participants' Reaction Times (RTs) in both levels of the detection task were signi cant, both for each level separately as well as the mean activation in the entire cognitive detection task (p<.01 for all, see Table 2). Additionally, mean RTs were calculated for each participant and each task. A0 activity signi cantly increased with slower participant RTs, while ST4 activity signi cantly decreased with slower RTs (see Table 3 and Figure 3).

Correlations between MMSE scores and EEG variables
Pearson r and p values of correlations for each EEG variable are presented in Table 4. The activity of A0 increased for lower MMSE score while the activity of ST4 decreased for lower MMSE score (p=.010 and p=.009, respectively, see Figure 4). All other EEG variables did not exhibit signi cant correlation with MMSE scores.

Mixed linear models results
The rst LMM analysis (see table 5), revealed signi cant differences between the groups in VC9, A0 and ST4 biomarkers' activity. Speci cally, VC9 activity was signi cantly lower in healthy young participants compared to healthy seniors (p=0.005), A0 showed signi cant differences between healthy seniors and all other groups (Healthy young, p<0.001, MCI-R, p=0.04 and MD p=0.011), and ST4 showed a signi cant difference between healthy seniors and MD group (p=0.011).
Follow-up LMM analysis revealed that in the young healthy group (see table 6), VC9 showed increased activity for detection 2 compared to detection 1, and decreased activity for resting state (corrected p=0.012 and p=0.008 respectively). A0 activity signi cantly increased in detection 2 relative to detection 1 (corrected p=0.004). No signi cant differences between task levels were found in any of the senior groups (for all means and SE see gure 5, and tables 5-7).
No signi cant differences were found between the groups in any of the frequency bands. However, Theta and Delta showed increased activity for detection 2 compared to detection 1 level (p=0.001, and p=0.006 respectively). Theta also exhibited a signi cant decrease in activity in resting state task compared to detection 1 (p=0.001).
Follow-up LMM analysis revealed that among the healthy young group, a signi cant increase in Theta activity in detection level 2 compared to detection level 1, as well as a signi cant decrease in resting state task (corrected p=0.006 and p=0.016 respectively). Delta activity in the healthy young group was signi cantly higher in detection level 2 relative to detection level 1 (corrected p=0.022). The difference in Delta power between detection 1 and resting state in the MCI-R group did not survive Bonferroni correction (corrected p=0.07). No other signi cant differences between task levels were found in any of the senior groups.

Discussion
Cognitive decline remains highly underdiagnosed [28]. Improving the detection rate in the community to allow early intervention is therefore imperative. The aim of this study was to evaluate the ability of a single-channel EEG system with an interactive assessment tool to detect cognitive decline with correlation to known assessment methods. We demonstrate that a short evaluation which can be selfadministered, together with automatically extracted objective EEG features from a wearable system with an easy set-up, may provide an assessment method for cognitive decline. 50 seniors and 22 young controls completed a short and interactive cognitive assessment battery. Classical EEG frequency bands were used in the analysis of the data as well as pre-de ned Machine Learning (ML) features. ML applied on raw EEG signals is increasingly being examined for detection of cognitive deterioration. The biomarkers that are extracted using ML approaches show accurate separation between healthy and cognitively impaired populations [17], [18], [29]. Our approach utilizes advanced wavelet-packet analysis [30], [31] as pre-processing to ML. The biomarkers used here were calculated on a different dataset to avoid the risks associated with classi cation studies such as over tting [32]. This is unlike other studies that use classi ers trained and tested via cross validation on the same dataset [10], [20], [21]. Speci cally, the pre-extracted biomarkers used here, ST4 and VC9, were previously considered in a study performed on young healthy subjects. Results showed a correlation of VC9 to working memory load and correlation of ST4 to individual performance [23].
A novel interactive cognitive assessment based on auditory stimuli with three cognitive load levels (high, low and rest) was used to probe different cognitive states. Individual response performance in the tasks (RT) was correlated to the MMSE score which further validates the cognitive assessment tool. Importantly, individual response performance also signi cantly correlated with A0 and ST4 biomarkers. These results support previous ndings that recording EEG during active engagement in cognitive and auditory tasks offer distinct features and may lead to better discrimination power of brain states [33]. To continue this notion, we used an auditory assessment battery with musical stimuli. It was previously shown that musical stimuli elicit stronger activity than using visual cues such as digits and characters [34].
Results further show that A0 and ST4 signi cantly correlated with individual MMSE scores. To get a clearer separation between cognitive stages, we divided the participants to groups according to the MMSE scores. In allocating the groups, we used the common cutoff score of 24 to divide between lowfunctioning (MMSE<24) and high-functioning seniors. However, we divided the high-functioning group further using a cutoff score of 27, to get a notion of possible separability between cognitive states in high-functioning seniors.
Results demonstrated the ability of A0 and ST4 to signi cantly differentiate between groups of seniors with high vs. low MMSE scores. Speci cally, ST4 separated between the healthy seniors and low-MMSE seniors (MD group) with the common cutoff score of 24, comparable to previous reports in this eld [16], [20], [21]. Conversely, A0 showed differences between the healthy seniors and all other groups. A0 results suggest detection of more delicate differences between healthy seniors and seniors that are considered healthy to-date, but at a greater risk for developing cognitive decline (with MMSE scores below 27 but above 24). These results may indicate a different cognitive functionality between entirely healthy seniors and seniors that score lower in the initial screening test but are not considered as suffering from cognitive decline. This result contributes to the debate in the literature over cognitive functionality of patients with scores below 27 [8], [26]. Furthermore, A0 and VC9 results suggest different cognitive patterns between healthy young participants and healthy seniors, consistent with reports from previous studies [35]. With regards to the commonly used EEG frequency bands, Theta and Delta activity increased with higher cognitive load, separating between the high and low cognitive load levels in the young healthy group. Theta also showed separation between low cognitive load and resting state. This is in line with previous reports of frontal Theta and Delta increase during cognitively demanding tasks [36]- [38]. This difference was not present in the senior population, supporting the notion that Theta and Delta may be indicative to cognitive decline and serve as a predictor of deterioration, consistent with previous ndings [10], [11]. Finally, A0 showed separation between high and low cognitive load levels within the cognitive task, further supporting the ability of A0 to differentiate between cognitive states. VC9 biomarker showed differences between all task levels, with a signi cant increase between rest and cognitive load levels within the cognitive task. This supports previous work validating VC9 as a working memory load biomarker in the healthy population [23]. Together, these results provide an initial indication of the ability of the proposed system to assess cognitive states and detect cognitive decline in the elderly population.
It was shown that cognitive decline is associated with high-risk for developing dementia, and Alzheimer's disease in particular [2], and that cognitive decline may be detected several years before dementia onset with known validated tools [39]. Dementia is gradually recognized as one of the most signi cant medical challenges of the future. So much so, that it has already reached epidemic proportions, with prevalence roughly doubling every ve years over the age of 65 [40]. This rate is expected to increase unless therapeutic approaches are found to prevent or stop disease progression [41]. Since Alzheimer's Disease (AD) is the most prevalent form of dementia, responsible for about 60-70% of cases [42], it remains the focus of clinical trials. To date, most clinical trials that include a disease-modifying treatment, fail to demonstrate clinical bene ts in symptomatic AD patients. This could be explained by the late intervention that occurs after neuropathological processes have already resulted in substantial brain damage [43]. Hence, the discovery of predictive biomarkers for preclinical or early clinical stages such as cognitive decline is imperative [44].
The ability to differentiate between cognitive states with the novel biomarkers shown here relies only on a single EEG channel and a short assessment battery, unlike most studies attempting to assess cognitive decline with multichannel EEG systems [17], [45], [46]. It has been argued that the long setup time of multichannel EEG systems may cause fatigue, stress, or even change mental states, affecting EEG patterns and subsequently study outcomes [22]. This suggests that cognitive decline evaluation using a wearable single-channel EEG with a quick setup time may not only make the assessment more affordable and accessible, but also potentially reduce the effects of the pretest time on the results. Using a single EEG channel was previously shown to be effective in detection of cognitive decline [20], however, here we demonstrate results using biomarkers that were extracted on an independent dataset to avoid an over tting of the data. The extracted cognitive biomarkers from the system offered here may potentially enable detection in the community in much earlier stages before dementia symptoms arise.
Limitations While this pilot study shows initial promising results, more work is needed. Speci cally, additional studies should include a longer testing period to quantify the variability within subjects and potentially increase the predictive power. Also, larger cohorts of patients that are quanti ed by extensive screening methods would offer an opportunity to get more sensitive separation between earlier stages of cognitive decline using the suggested tool, and potentially reduce the subjective nature of the MMSE. Moreover, a longitudinal study could assess cognitive state in asymptomatic elderly patients and follow participants' cognition over an extended period of time, validating the predictive power of the biomarkers.

Conclusions
In conclusion, this pilot study successfully demonstrated the ability to assess cognitive decline, using a wearable single-channel EEG system and novel cognitive biomarkers which correlate to well-validated clinical measurements for detection of cognitive decline. Using a low-cost approach and objective biomarkers, it may provide consistency in assessment across patients and between medical facilities clear of tester bias. Furthermore, due to a short set-up time and interactive cloud-based assessment, this tool has the potential to be used on a large scale in every clinic to detect deterioration before clinical symptoms emerge. Future studies should explore potential usefulness of this tool in characterizing changes in EEG patterns of cognitive decline over time, for early detection of cognitive decline to potentially allow earlier intervention.

Declarations
Ethics approval and consent to participate Ethical approval for this study was granted by the Ethics Committee (EC) of Dorot Geriatric Medical Center on July 01, 2019. All participants provided informed consent according to the guidelines outlined in the Declaration of Helsinki and singed the Informed Consent Form (ICF). Participants were provided with the appropriate ICF in the language they could read and understand.

Consent for publication
Not applicable, no individual person's data or patient information is published in this manuscript.

Availability of data and materials
The datasets generated and analyzed during the current study are not publicly available due to patient privacy but may be available from the corresponding author under restrictions.