Thirty healthy participants (age 26 ± 3 years) from the National Cheng Kung University were blinded and randomly assigned into two groups (i.e., Alpha, n = 15; control, n = 15). All participants were right-handed and had no experience of taking NFT in the past. The two groups had no difference in the factors of gender (p = 0.75), age (p = 0.46), and education (p = 0.39). The entire experiment complied with guidelines and regulations in the Institutional Review Board of National Cheng Kung University Hospital. Informed consent was provided and signed for all participants.
Hardware architecture
The hardware architecture of our portable neurofeedback system (Fig. 1) primarily contained an EEG amplification board, a microcontroller module, and a smartphone. The EEG amplification board aimed to amplify brain activity. The microcontroller module was responsible for EEG data sampling and to control wireless transmission of a Bluetooth module. The smartphone was used to receive and calculate wireless EEG data as a visual feedback and to save data.
The present study used a single-channel EEG recording through Ag/AgCl electrodes. Based on previous neurofeedback studies [1, 17], we selected a C3 channel as an active lead with a reference over the contralateral mastoid area (M2) according to the 10–20 system [24]. A ground electrode was placed over the Fpz region. The EEG signal was amplified with a gain of 10,000 through an instrumentation amplifier (AD623, Analog Device, Texas) in combination with two non-inverting operational amplifiers (AD8538, Analog Device, Texas) within a frequency range of 0.15–50 Hz [25]. The amplified EEG was then positive-biased to an analog-to-digital converter (ADC) of the microcontroller.
The microcontroller module included a MSP430F5438 integrated chip, which embedded with a MSP430 microcontroller unit, 256 kB flash memory, 16 kB RAM, and other peripherals such as an 8-bit ADC and three 16-bit timers. The MSP430 digitized data through an embedded ADC with 128 Hz and transferred sampled EEG data to a Bluetooth module. Afterwards, the Bluetooth module transmitted the data to a smartphone. The core component of this Bluetooth module was a Nordic nRF8001 chip that integrated a fully compliant Bluetooth radio and link layer controller. Bluetooth is designed for short-range and low power wireless communication, and it is widely adopted in personal computers and consumer electronic devices, e.g., mobile phone or media player. The present study used the Bluetooth version 4.0, which aimed at applications in the fitness, healthcare and security areas because it provided lower cost, lower power consumption, and a comparable communication range than a traditional Bluetooth protocol [20, 23].
Software implementation
The software of the proposed training system contained two parts: control firmware on the microcontroller module and a training application on the smartphone. The EEG signal analysis device was able to pair with any Bluetooth-compatible mobile device with the training application installed. To reduce effort of porting the proposed system to other mobile devices, all the analysis and calculation in the proposed system was executed on the microcontroller module of the EEG signal analysis device. The firmware running in the microcontroller module performed EEG signal acquisition, data analysis, and wireless transmission. The application running on the smartphone provided a graphic user interface to configure the training procedure and displayed the real-time EEG feedback. The software components were described below.
Data analysis and wireless transmission of the NFT
The data analysis task fetched the 1-s sampled EEG data in the buffer and then performed fast Fourier transform (FFT) to calculate the power of the alpha rhythm. Both raw data and calculated data were transmitted immediately to the smartphone via Bluetooth communication. Figure 2 shows the flowchart of the firmware (left) of the microcontroller, including the main program for EEG acquisition, EEG analysis, and wireless transmission. The smartphone received and displayed the alpha power and the total success duration for 1-s alpha events. Participants saw all training performance in terms of changes of alpha power and alpha duration throughout the training sessions over a smartphone. In addition, information of EEG changes with regard to training number per day displayed on the smartphone at the end of each training session.
Figure 3 reveals timing diagram of the EEG signal acquisition, data analysis, and wireless transmission tasks running on the microcontroller unit (MCU). The timing was obtained by toggling an MCU I/O pin at the start and end of the task and measuring the duration via an ADC (USB-6009, National Instruments, TX). The data analysis task was performed every 128 sampling periods. The Bluetooth transmission was triggered immediately after the completion of the data analysis task. The execution time of the data analysis task and the latency for wireless transmission were 172 and 2.7 ms, respectively. Although the data analysis task spanned several sampling periods, sampled data collection always performed with a higher priority.
Training interface of the NFT
The training application running on the smartphone was utilized for displaying real-time feedback. When the application started, users entered the desired time length of the training procedure and set up a Bluetooth connection between the signal analysis device and the smartphone (Fig. 4a). The connection was set up by clicking the “Discover All Devices” button to search nearby Bluetooth devices. The identification and type of Bluetooth devices displayed in a list to allow a user to select a target for Bluetooth connection setup. Thus, the training procedure was ready. Thereafter, quality of EEG recording was ascertained when the amplitude was < 100 μVrms after properly adjusting electrode–electrolyte-scalp conjunction.
During the training procedure, the training application received real-time consecutive 1-s EEG data wirelessly and displayed the information of calculated alpha power on the screen of the smartphone. Figure 4b shows the information of success episode number, alpha power of the current episode, and changes of consecutive alpha powers. The blue bar of Fig. 4b reflects current alpha power. The waveform shown in the bottom panel represented consecutive alpha powers. The received EEG data was stored in the internal storage of the smartphone for future analysis. Users were able to terminate the training procedure before the end of the procedure by clicking the Disconnect button. At the end of each session, alpha powers and success number of 300-s training episodes displayed on the smartphone to allow user/researcher to develop or establish their strategy through trial-and-error learning [26].
System assessment
The present study used a 3.7 V, 1000 mAH, Li-ion battery (HYB, China) for the EEG signal analysis device. Current consumption of the EEG signal analysis device was measured using a 6-1/2-digit Digital Multimeter (USB-4065, National Instruments). Operation duration of a Li-ion battery was defined under a free running test until the system ran out of power. The timestamp testing data in the smartphone indicated battery life of the EEG signal analysis device.
Experimental procedure
To verify the effect of the proposed system on memory, three-stage experiment (i.e., pretest, training, and posttest) was designed. The pretest and posttest of three cognitive tests were performed immediately before and after the training stage. During the training stage, the 1-channel EEG signal (C3-M2) was utilized. Subjects in the Alpha group received the projection of alpha power (8–12 Hz) on the screen of a smartphone. The control group received various randomly selected 4-Hz bandwidth in the range of 7–20 Hz for every 1-s event, which was used in our previous study [26].
At the beginning, brain activity was recorded and analyzed to assess its noise level, including artifacts of eye blink or muscle contraction, etc. To reduce possible artifact signals, each subject was reminded before the training [26]: (1) avoiding frequent eyes blink; (2) eyes closure or fall asleep was informed as an invalid strategy; (3) avoiding body’s movement or shaking/nodding head; (4) avoiding too much facial expression intentionally. A digital camera was used to rule out the effects of these behavioral artifacts.
Twelve training sessions were performed within 3 weeks (Fig. 5). Four sessions were performed per week. A session contained 5 blocks, and each block took 5 min. In the beginning of a training, a 2-min baseline EEG was recorded followed by 1-min rest. Thereafter, a 5-min training block followed by a 1-min resting period was performed. Subjects used the proposed system and attempted to increase activities of particular rhythms shown on the screen of a smartphone.
In an NFT, participant can see the instantaneous information of the 1-s power of a selected bandwidth and the waveform of all consecutive 1-s powers of a selected bandwidth. The instantaneous power was expressed in a horizontal bar (Fig. 4b). If an EEG power increased, the bar moves to right side. Otherwise, an EEG power decreased, the bar moved to left side. Participants were instructed to move the bar to the rightmost position and to maintain the bar as long as possible.
During the 1-min rest period between two blocks, we tried to help participants to develop a good strategy using the information of consecutive 1-s power information (the bottom panel of Fig. 4b). For example, we pointed out timestamps with higher power in the training block and asked participants to recall the strategy they used. During the inter-block rest, we encouraged participants to try their best to move/control the bar. Although the control group seemed to be uneasy with controlling their brain activities, they reported no difficulty and frustration during the training.
Evaluation of cognitive function
The cognitive function was evaluated through the backward digit span test, word pair test, and Mini-Mental State Examination (MMSE). MMSE was used to evaluate possible cognitive impairment. MMSE was able to evaluate various cognitive abilities, such as orientation to time and space, recall, language, attention, calculation, etc. The MMSE score ranged from 0 to 30 points. A MMSE score greater than or equal to 25 points indicated normal cognition [27]. Participant was excluded if MMSE < 25 in this study.
The backward digit span task [28] is a measure of working memory and contains phases of practice and test. In the practice phase, subjects were instructed to familiarize themselves with the processes of the test. In the test phase, thirty trials were performed. At the beginning of each trial, the subject was asked to focus on a cross symbol on the monitor. A series of digits (4–8 randomly) were displayed after the cross disappeared, and each digit lasted for one second. The subject answered the digits in a reverse order on an answer sheet at the end of each trial. Each digit in the correct place had one point. There were a total of 180 digits in the 30 trials, thus the maximum score was 180 points.
The word-pair test [26, 29] was composed of two phases, learning and retrieval phases. In the learning phase, the monitor displayed a cross for 3500 ms followed by a pair of Chinese words for 1500 ms. Thereafter, a white screen was displayed for 5000 ms before the next start. Eighty word pairs were used in the word-pair test. Subjects had a 30-min break between the learning and retrieval phases. In the beginning of the retrieval phase, a cross was displayed for 3500 ms to make the subject focus on the monitor, followed by a priming word for 6500 ms. Subjects had to pronounce the paired word within 6500 ms. Each correct answer was worth 1 point. The maximum score was 80.
Data analysis
In an NFT, EEG was transferred into a power spectrum using FFT with a Hamming window. Power of the alpha bandwidth or particular bandwidth was obtained by summation of selected bandwidth in the power spectrum. Thereafter, the power was projected to a horizontal bar to indicate current status of EEG (Fig. 4b). To further illustrate time–frequency characteristics of various activities, such as cortical activity of the C3 or Fp1 lead, electrooculogram (EOG), or electromyogram (EMG), a short-time FFT with a Hamming window was performed with 50% data overlapping.
There were two indexes used to assess the training progression of EEG throughout 12 training sessions: the mean alpha power ratio and total duration of successful alpha events [26]. Alpha power ratio is defined by the power of 8–12 Hz normalized by averaged 8- to 12-Hz power of all 1-s baseline EEGs as shown below.
$${\text{Alpha power ratio}} = \frac{Alpha\,power}{Baseline\,alpha\,power}$$
If alpha power ratio of 1-s EEG was higher than 1.2, thus the 1-s EEG segment was considered as a successful event. All successful 1-s events within a session were cumulated as an index of the total duration of successful alpha events. Moreover, alpha power ratios of all successful 1-s EEGs within a session was averaged to obtain an index of the mean alpha power ratio. The mean alpha power ratio throughout 12 sessions was used to reflect dynamic changes of alpha powers within a NFT [26].
Statistical analysis
Demographic data (age, education, gender) in the two groups were analyzed by independent t test or Chi square test, respectively. The normality and equal variance of the data were assessed for a parametric statistic. Mean alpha power ratio and total alpha duration throughout 12 training sessions in the two groups were analyzed by two-way analysis of variance (ANOVA) with one-factor repetition, if appropriate, followed by t test with Bonferroni correction. Accuracies of the backward digital span task and word-pair task were assessed by two-way ANOVA with one-factor repetition. The temporal relationship in activities of different channels was calculated by Pearson correlation coefficient r. Furthermore, independent t test was used to compare r values between two channels. All statistical analyses were performed by SigmaPlot. Data were expressed as the mean ± standard error of the mean. A two-tailed significance level was set at p < 0.05.