Keywords

1 Introduction

Recently, with the aging of the population in Japan, dementia patients are on the rise and it is becoming a big social problem. According to reports by the Ministry of Health, Labor and Welfare in Japan, the number of patients is expected to reach about 7.3 milion in 2025 [1]. With the progress of medicine, if we can find dementia early, we can delay treatment of progression. Therefore, early diagnosis of dementia is becoming more important from the viewpoint of maintaining patient’s QoL (Quality of Life) [2].

Dementia is mainly screened by performing brain image examination using MRI and SPECT, and neuropsychological examination such as MMSE (Mini-Mental State Examination), and comprehensively summarizing these results. However, the MRI examination has high running cost, and the neuropsychological examination has a problem that it is difficult to apply to people who understand Japanese insufficiently. Therefore, it is necessary to develop an inexpensive screening tool that can diagnose in a short time and research on technology.

So far research on event related potentials for dementia patients has been conducted in Japan. Morita showed that both the amplitude of P300 is significantly decreased and latency of P300 is significantly prolonged between the cognitive group and the dementia low risk group and the dementia group and healthy group at the visual event-related potential [3]. Therefore, it is considered that Event-Related Potential (ERP) measurement using character input BCI (Brain-Computer Interface) using P300 is possible even if dementia patients. Many papers that have studied dementia from the perspective of brain science have been posted. However, technological research to combine with brain information technology and diagnose at the same time with measurement is not conducted worldwide. Therefore, we aim to develop a screening tool for dementia using character input type BCI. If dementia can be screened using character input BCI, We can conduct inspections at low cost for introduction cost and running cost. We can also expect reduction in human cost by automating explanation of inspection. Also, while neuropsychological examination, which is a conventional method, takes about 1 h to perform, screening by character-input BCI can be performed in about 30 min.

Therefore, it is expected that the medical time can be shortened greatly. So far, character-input BCI was used for practical realization as a communication aid device for amyotrophic lateral sclerosis (ALS) patients, and researches on reducing the number of electrodes and reducing the time for inputting characters have been conducted based on patient needs [4, 5]. We focused on erroneous input in BCI. It is necessary that the subject has to concentrate high attention to input correct characters in the character-input BCI. On the other hand, it causes an input-error of far character when the subject is distractedness. Hence, we considered that the number of input error of far character may increase when the dementia patient uses the character-input BCI. In this research, in order to develop a dementia screening tool using character input type BCI, we calculated the weighted mean distance SEDV (Spelling - Error Distance Value) in consideration of the character estimation principle in character input type BCI. And we visualized and quantified distraction in the elderly.

2 The Character-Input BCI

2.1 Event-Related Potential P300

Event-related potential (ERP) is electric activity of the brain that occurs in response to stimuli such as light and sound and movement such as bending and stretching of a finger [6]. P300 is a positive potential associated with attention. Generally P300 is observed in the oddball task, but it is obtained with most tasks seeking some kind of mental judgment such as selection and understanding [7].

2.2 P300 Speller

P300 Speller is a character input type BCI proposed by Farwell, Donchin in 1988 [8]. The P300 Speller is a BCI that discriminates P300 at online and inputs characters by converting the discrimination result into switch information. Displays a dial on the screen and gives visual stimulus by blinking characters in rows or columns at random. The subject searches for the character intended for input and gazes at it. P300 is triggered when the gazing character (target character) flashes, characters are input by detecting P300 on-line (Fig. 1). In this study, P300 Speller do not use as a character input device, but it uses for visual stimulus presentation and ERP recording. Or later, when we mention BCI in this paper, we indicate the P300 Speller.

Fig. 1.
figure 1

Character estimate principle in BCI

Full size image

2.3 Measurement Principle of Distraction Using BCI

In this study, I thought that the degree of distraction of distraction can be known by the position of incorrect input of BCI. In character input BCI, if attention to characters is sufficiently high, P300 appears strongly in rows and columns including target characters, and P300 does not appear in other rows and columns (Fig. 1). As an erroneous input that may occur when attention to characters is high, there is a neighbor error. This is caused by erroneous reaction when characters adjacent to the target character blink [9]. On the other hand, if the attention is distracted, P300 of the row and column including the target character become weak, and P300 appears extensively depending on distraction. Therefore, BCI can not correctly detect P300, making input-error of far character more likely to occur (Fig. 2). Therefore, it is considered that the average distance of input-error represents the degree of distraction.

Fig. 2.
figure 2

Measurement principle of distraction

Full size image

3 Neuropsychological Examination MMSE

MMSE (Mini-Mental State Examination) is a screening test for dementia that can be performed from 5 min to 10 min [10]. It can examine the functions of time, place orientation, immediate and delayed reproduction of 3 words, attention, calculation, nomenclature, repetition, language understanding, reading, writing and drawing functions. The upper limit is 30 points. In this study, subjects were classified into five groups of Questionably significant (L0) (27 to 30 points), Mild (L1) (24 to 26 points), Moderate1 (L2) (20 to 23 points), Moderate2 (L3) (16 to 19 points), and Severe (L4) (15 points or less) according to the scores of MMSE (Table 1).

Table 1. Subject grouping using MMSE scores
Full size table

4 Evaluation Method of BCI Character Estimation Result

4.1 Relative Position Plot of False Input

In evaluating character estimation results by BCI, we visualized the relative positional relationship between the target character and the estimated character of BCI (erroneous input plot). Let the position of the target character in the dial used be (0, 0). When the position of the character estimated by BCI is regarded as (Ca, Cb) that the position when “a” character to horizontal direction and “b” character to vertical direction from target character shifted, add 1 to the coordinates (Ca, Rb) of the plot of Fig. 3. The erroneous input plot of Fig. 2 is an example in the case where the target character is “Annkomochi”, and the estimated character of BCI is “Ifukamomo”. By performing addition for the total number of input characters \( N_{char} \) times, the relative positional relationship of erroneous input becomes clear. By dividing each coordinate of the erroneous input plot by “\( N_{char} \)’’, the relative frequency distribution of each erroneous input can be obtained [11].

Fig. 3.
figure 3

The input-error plot

Full size image

4.2 Spelling-Error Distance Value

SEDV (Spelling-Error Distance Value) is the weighted average distance taking into account the character estimation principle of BCI. Upon calculation of SEDV, weight of incorrect input of location of “a” character to horizontal direction and “b” character to vertical direction from target character was defined by Eq. (1). Here, \( D_{1} \) indicates a case where one of the line and the column of the target character matches the character estimated by the BCI. And \( D_{2} \) indicates a case where character estimated by the BCI and the row/column of the target character are different. In character estimation in BCI, it is only possible to enter correct characters only after the target character and the row/column of characters estimated by BCI match. Thus, matching one of the rows and columns means that the character estimation of BCI is half correct [9]. Therefore, we add processing that halves the weight. SEDV asked by multiplying the erroneous input plot \( (l) \) obtained in 4.1 by the weight \( (W) \) and dividing by the total number of characters input \( (N_{char} ) \) (Eq. (2)) [11]. In the Eq. (2), r represents the number of rows of the dial and c represents the number of columns of the dial. SEDV is statistically a value corresponding to the standard deviation. When it was plotted as shown in Fig. 3, the SEDV is 2.24 character.

$$ {\text{W}} = \left\{ {\begin{array}{*{20}c} {D_{1} \frac{1}{2}\sqrt {\left( {a^{2} + b^{2} } \right)} \left( {a = 0 \cup b = 0} \right)} \\ {D_{2} \sqrt {\left( {a^{2} + b^{2} } \right)} \left( {a \ne 0 \cap b \ne 0} \right)} \\ \end{array} } \right. $$
(1)
$$ {\text{SEDV}} = \frac{1}{{N_{char} }}\sum\limits_{{i = - \left( {r - 1} \right)}}^{r - 1} {\sum\limits_{{j = - \left( {c - 1} \right)}}^{c - 1} {\left( {l \times W} \right)} } $$
(2)

5 Experiment

5.1 Subject Information

Subjects were 24 persons (aged 80.5 ± 5.3 years old, MMSE 22.5 ± 5.0 points) in their 70’s to 90’s who were judged healthy or mild to severe dementia by specialists. The experiment was conducted with the approval of the ethics review committee of Tokyo Medical University (early diagnosis of dementia using brain computer interface (BCI), 2016-083). In addition, subject conducted a neuropsychological examination by a clinical psychologist on the same day as the experiment.

5.2 BCI Parameters

The dial was randomly blinked in units of rows and columns using a 6 × 10 Hiragana dial. Five or six characters of hiragana were gazed per experiment, and the experiment was conducted three or four times. The number of blinks (F) in each row and each column was set to 5 times (10 blinks per character). SOA (Stimulus Onset Asynchrony) was based on 210 ms (character lighting time 120 ms, character light off time 90 ms).

When subjects were difficult to recognize blinking, they were adjusted appropriately, and the maximum was 300 ms (character lighting time 170 ms, character lighting time 130 ms). Also, the target character search time S [s] is 6 s. The time required for a single experiment ET [s] can be obtained by the Eq. (3). In this study, the time required for one experiment was up to 198 s (SOA = 210 [ms], r = 6 [row], c = 10 [column], F = 10 [Times], S = 6 [s], N char  = 5 [characters]).

$$ {\text{ET}}\left[ {\text{s}} \right] = \left( {\frac{SOA}{1000} \times \left( {r + c} \right) \times F + S} \right) \times N_{char} $$
(3)

5.3 BCI System

The system diagram of the BCI used in this study is shown in Fig. 4. Electroencephalograms were measured using an active electrode (LADY bird electrode manufactured by g.tec), collected using an electrode box (g.SAHARAbox manufactured by g.tec), and amplified the signal using a bio-amplifier (g.USBamp manufactured by g.tec) and then incorporated into a PC. MATLAB 2012a was used for brain wave recording, stimulus presentation, and analysis processing. The electrodes were placed at eight location, Fz, Cz, P3, P4, Pz, O1, O2 and Oz, as defined by the international 10–20 system. Also, the reference electrode was placed on the back of the right earlobe and the ground electrode was placed on the forehead. The brain wave was derived using the monopolar induction method.

Fig. 4.
figure 4

BCI system configuration

Full size image

5.4 Teaching

Before starting the experiment, we made the subject check the characters that are displayed thin on the dial and whether the characters are blinking. After confirming that subjects can see the blinking of characters, we taught the subjects using the actual experiment screen as follows. “Look at the shining green characters. Next, various rows and columns will flash, but please only look at the characters that was shining green. Last, please tell me the number of the count that shined.” When judging from the state of the subject that the task contents of BCI were not understood, we taught to subjects until they understood the BCI task. During the experiment, if the experimenter judged that the subject was not able to search for the target character, it pointed to the target character before the character flashing began. We made the subjects report the number of blinks after the end of presentation of the blinking stimulus. Here, the number of times the subjects were not counted or counted unnecessarily was recorded as the number of erroneous reactions. It shows appearance of the character gaze experiment using BCI at Fig. 5. As shown in Fig. 5, blinking of characters was given to subjects as visual stimuli, and the subjects aim attention to blinking of characters. At this time, P300 appears when the gazing character blinks.

Fig. 5.
figure 5

Appearance of the character gazing experiment.

Full size image

5.5 P300 Discrimination Rate

In this study, we examined whether it is possible to diagnose patients with dementia who can not accurately perform the counting task by calculating the correlation between the P300 discrimination rate and the number of erroneous reactions in the counting task. When the target character flashes, the number of times P300 could be properly discriminated is divided by the total number of stimuli and the value expressed in percentage is set as P300 discrimination rate (Eq. (4)).

$$ {\text{P}}300\,{\text{discrimination}}\,{\text{rate}} = \frac{collect\,classification}{all\,data} \times 100\left[ \% \right] $$
(4)

6 Result

6.1 BCI Character Gaze Test

Figure 6 is a graph of SEDV in each subject group. From Fig. 6, there was a difference of 0.56 characters in SEDV between L1 group and L2 group, 1.81 characters in SEDV between L3 group and L4 group. As a result of analysis of variance, there was no significant difference, but as a result of multiple comparison, SEDV between L0 group and L4 group, SEDV between L1 group and L4 group showed a significant tendency difference.

Fig. 6.
figure 6

SEDV between each level of cognitive decline

Full size image

6.2 Relationship Between Erroneous Reaction Frequency and P300 Discrimination Rate

Correlation between the number of erroneous responses in the counting task and the P300 discrimination rate was determined, and the relationship between the two was examined. From Fig. 7, a significant weak negative correlation was obtained with a correlation coefficient r of −0.38 (p < 0.01). Therefore, it is indicated that the P300 discrimination rate tended to decrease as the counting task could not be performed.

Fig. 7.
figure 7

Erroneous reaction frequency and P300 discrimination rate

Full size image

7 Discussion

7.1 Possibility of Dementia Screening by BCI

As shown in Fig. 6, the SEDV increased as dementia progressed, indicating a tendency to distraction. The SEDV difference between the Questionably significant (L0) and the Mild (L1) is 0.15 characters, and it is considered that screening by SEDV is difficult for L0 group and L1 group. However, there was a difference of 0.41 characters between the L0 group and the L2 group, and 0.56 characters between the L1 group and the L2 group, so It may be able to screening by SEDV for L0 and L2, L1 and L2. However, as a result of multiple comparison, there was no significant difference between L0 group and L2 group, L1 group and L2 group.

The reason for this may be that the variance of the SEDV for each subject is large, therefore for reduce the variation, it is considered that it is necessary to increase the number of subjects and examine them in the future. Also, there is a significant trend difference of 2.24 characters for SEDV between L0 and L4 and 2.39 characters for SEDV between L1 and L4, therefore screening is considered possible.

7.2 Comparison Between Young Healthy Subjects and Dementia Patients by SEDV

As a result, experiments similar to this study in 12 young healthy people in their 20s to 30s, SEDV was 0.90 characters [11]. Therefore, as a result of comparison with SEDV of mild dementia group, dementia group, moderate dementia group, and severe dementia group shown in Fig. 6, the mild dementia group was 1.11 characters, the dementia group was 2.09 characters, We obtained 1.63 characters for the moderate dementia group and 3.14 characters for the severe dementia group, with SEDV differing by more than 1.00 characters in each group. From this, it can be said that there is a big difference in cognitive function among healthy young people and dementia patients.

Similarly, as compared with SEDV of the healthy elderly group shown in Fig. 6, since there is a difference of 1.06 characters, it is considered that there is a difference in cognitive functions in healthy elderly people. At the same time, since aging is one of the causes of the decrease in the attention of elderly people diagnosed as healthy according to the score of MMSE, it was suggested that cognitive function may decrease with aging as well. However, as there are only 5 members of healthy elderly people, there is also a need to increase the number of subjects and review as well as dementia patients.

7.3 P300 Discrimination Rate by LDA and Number of Erroneous Responses in Counting Task

The correlation coefficient r was −0.36 in the erroneous response in the counting task and the P300 discrimination rate, and a weak negative correlation was observed. From this fact, even if the degree of achievement of the counting task is low, there is a possibility that accurate diagnosis cannot be made but the result of character input is not greatly adversely affected. “Double flash” can be cited as a cause of poor performance of counting tasks. Double flash is a phenomenon in which target characters shine continuously. It is expected that even young healthy people will have difficulty reacting and the priority of solution will be high. As a solution to this problem, it is necessary to create an algorithm to present the stimulus so as not to cause double flash.

8 Conclusion

In this study, the erroneous input distance value SEDV was obtained for the development of dementia screening tool using character-input BCI, and the distraction of the elderly including dementia was quantified. And we examined the relationship between the number of incorrect responses in the counting task and the P300 discrimination rate and examined whether it is possible to diagnose dementia using the P300 Speller even for patients who cannot achieve the counting task.

As a result, SEDV increased as dementia progressed, indicating that the attention tended to be distracted.

Regarding the number of false responses in the counting task and the P300 discrimination rate, a significant weak negative correlation was observed with a correlation coefficient r = −0.36. It was also found that there was no major impact, so it was shown that diagnosis of dementia is possible even if the achievement rate of the counting task is bad. However, since a weak negative correlation appears, it is not always that the results of the counting task do not have a bad influence on the diagnosis of dementia, so we indicated that we improve the UI of the dial and develop an algorithm to suppress the double flash.

As a future subject, it is necessary to continue the experiment and to confirm the tendency of distraction that has been clarified in this research.

In addition, focusing on the character estimation result of each subject, it is necessary to consider whether screening is possible by comprehensive judgment using P 300 latency or SEDV.