A Wireless EEG System for Neurofeedback Training

Abstract

This paper presents a mobile, easy-to-maintain wireless electroencephalograph (EEG) system designed for work with children in a school environment. This EEG data acquisition platform is a small-sized, battery-powered system with a high sampling rate that is scalable to different channel numbers. The system was validated in a study of live z-score neurofeedback training for quantitative EEG (zNF-qEEG) for typical-reading children and those with developmental dyslexia (DD). This system reads and controls real-time neurofeedback (zNF) signals, synchronizing visual stimuli (low spatial frequency (LSF) illusions) with the alpha/theta (z-α/θ) score neural oscillations. The NF sessions were applied during discrimination of LSF illusions with different contrasts. Visual feedback was provided with color cues to remodulate neural activity in children with DD and their cognitive abilities. The combined zNF-qEEG and training with different visual magnocellular and parvocellular tasks (VTs) compensated for the deficits in the temporal areas affecting the occipitotemporal pathway more in the left-hemispheric ventral brain areas of the post-training children with dyslexia in the low-contrast LSF illusion and dorsal dysfunction in the high-contrast LSF illusion. The better α/θ scores for postD in the temporoparietal and middle occipital regions can be associated with an improvement in special frequency processing, while the better scores in the precentral and parietal cortices were due to an advancement in the temporal processing of the illusion. The improvements in the reading speeds were twice as high after 4 months of qEEG z-NF-VT training, with three times fewer omitted words and errors.

1. Introduction

The popularity of neurofeedback (NF) has grown significantly in recent years, making neuroscience more accessible to scientists and engineers as well as for various fields of physiological activity and tasks in pediatric studies. In computational neuroscience in particular, the functional neural network has been used to determine the relation between NF training and EEG connectivity in groups of children [1,2]. Currently, NFs are actively used for medical purposes [1,3] in combination with multi-modal stimuli for children with dyslexia [4,5,6,7,8]. An overview of the possibilities of EEG NF training also described quite favorably the prospects in this area [1,3,4,5,9,10].

One of the limiting factors for the widespread use of NF is the difficulty in reading EEG signals, as they are susceptible to various types of noise (e.g., electromagnetic), motion artifacts [11], and variability in EEG signal features [12]. Due to the non-stationary neuronal activity, the requirements for professional EEG mobile devices are high quality and a high resolution for the applied analog front-ends in them. The guidelines of the American Society of Clinical Neurophysiology recommend that the required features of effective EEG systems be a minimum number of channels to obtain a good spatial resolution (of 32 channels) and application of effective algorithms to detect events, improving the efficiency level of the system [13]. Commonly used EEG acquisition systems are either expensive or difficult to deploy on a large scale, as the system is fixed and difficult to expand, or the performance and usage scenarios are limited [1].

Wireless systems have significant improvements compared with traditional wired EEG systems, such as protection against electromagnetic interference and the flexibility of monitoring the out-of-hospital environment to transmit data in any situation [1,14]. Being lightweight and small enough, these systems are wearable, consume little power for long-term monitoring, and have the computing power to process data for online decision making [1,14]. The limitations that slow the widespread implementation of wireless EEG systems are a short battery life, moderate signal resolution, small number of channels, and wireless bandwidth [1]. Such quality can be better achieved by low-noise programmable analog-to-digital convertors (ADCs) for EEG and ECG signals, such as ADS1298/9 [9,15,16,17,18,19,20,21]. The professional mobile EEG apparatuses on the market are not for outdoor use or extensive NF research [1]. There are no low-cost and compact data acquisition systems available that allow good reconfiguration of scaling, hardware, and up to 32 or more channels spatially for outdoor widespread applications [1]. The EEG systems support streaming and event detection modes that have different power consumption. High sampling rates and a large number of channels are rarely achieved in a single device on modern wireless EEG acquisition platforms [1,9,15,16,17,18,19,20,21,22].

In applications intended for children in outdoor environments, the need for wireless systems capable of autonomous survey and record NF information processing becomes apparent. In outpatient settings of pediatric studies, EEG systems should be small and portable with low power consumption for long-term operation and have a sufficient computing ability for online data transferring and processing during various tasks [14,16].

Low-cost modular board amplifiers were used for EEG acquisition in the current study, maintaining significant cost savings. The proposed design would allow an easily scalable system with adaptive spatial and frequency characteristics for different NF-EEG tasks. Section 2 presents an overview of the real-time EEG acquisition system solutions and state-of-the-art designs. Different views of the architecture and technical decisions are shown in Section 3. The zNF-EEG procedure and results are described in Section 4. An overview of the results is discussed in Section 5. The system limitations and conclusions are presented in Section 6 and Section 7, respectively.

2. Related Work

2.1. Real-Time EEG Acquisition Platforms

There are several commercial electronic components that are used for noninvasive monitoring of brain waves (EEG devices) [20]. The creation of a new high-precision analog front-end design (AFE) for discrediting the analog EEG signal is still under experimentation [18]. Prototype systems for long-term continuous EEG monitoring of epileptic patients are based on wearable modules with battery-powered AFEs (ADS1299) with a sampling rate Fs of 250 ksps (up to 1 kbps), a 1 GHz microprocessor (DM3730) (ARM-cortex-A8), 32 active dry electrodes, and wireless data transmission (802.11 b/g) to a host computer [1,16] (

Table 1. Specifications of in-house EEG platforms, mainly built on low-noise, low power, and high-CMRR ADS1298/9 AFE chip. The dash (-) symbol indicates that the information is not available.

System	AFE	Sampling Rate (Hz)	No. of Ch/ Electrode	Accuracy (dB SNR)	MCU/MPU	I/O	CMRR dB	Disadvantages
Vo et al., 2017 [9]	ADS1299	250 (up to 1 kHz)	8 wet gel	-	STM32F4 (ARM Cortex-M4)	Bluetooth	110	Shorter connection range, changing the baud rate of Bluetooth with increased number of channels
Myung, Yoo, 2013 [15]	IC chip	512	16 wet gel	-	STM32F103	Wi-Fi 802.11 WizFi210 with UART communication	-	high power consumption, limited CPU speed
Pinho et al., 2016 [16]	ADS1299	250 (up to 1 kHz)	32 active dry	-	DM3730 (ARM Cortex-A8)	Wi-Fi 802.11 b/g/n	110	High power, long cables, non-optimal system size, expensive Wi-Fi communication mode
Senevirathna et al. [17]	ADS1299	250	8 wet gel	8	SAM G55	Bluetooth Mate Silver	110	230.4 kBaud rate, wet electrodes
Consul-Pacareu et al., 2017 [18]	Proposed AFE design	250	2 wet gel	55	PSoC3 (CY8C3866-030LTI)	Bluetooth (RN-42)	110	Packet loss or delay in transmission
Uktveris, Jusas, 2018 [19]	ADS1298	250 (up to 1 KHz)	16–64 dry passive	15	16 MHz Atmega2560 Max speed	Bluetooth mini HM-11 BLE 4.0 (0.2 M baud), ESP8266 Wi-Fi module	97	Changing the baud rate of Bluetooth with increased of channels, limited 0.2 M baud
Vargas et al., 2021 [21]	ADS1299	250	8 dry	-	ESP32	ESP8266-12E/ SPI interfaces with a baud rate (0.9 M baud)	110	TI-ADS1015 drives, piezo sensors combine neural and physical activities
Feng et al., 2016 [22]	ADS1299	250 (up to 1 KHz)	16 dry passive	25	BeagleBone Black(AM3358 (ARM Cortex-A8; 1 GHz)	USB (Wi-Fi dongle)	110	Real-time response, high power, size and cost
Toresano et al., 2017 [23]	ADS1299	250	16 wet gel	91 %	84 MHz ATSAM3X8E (ARM Cortex-M3)	SPI communication protocol	110	Wire equipment
Proposed platform	ADS1298	250	40 dry passive	15	PIC18F2420	nRF24L01 (up to 2 Mbps, up to 256 channels)	97	3 modules of the system

). The problems of this prototype were the maximum battery life and long cables due to the non-optimal system size [1]. Another portable EEG system [22] used for speller application on locked-in syndrome patients also integrated ADS1299 AFEs to support 16 independent channels (Fs, 250 Hz). This EEGu2 system had a more powerful processing element: BeagleBone Black with a 1 GHz processor (AM3358) (ARM Cortex-A8). The sufficient processing capability of data with high accuracy (25 dB SNR) and long life operation of the battery (up to 12 h) enable rapid embedded deployment of the speller application. This prototype needs its real-time response, total required power, size, and cost improved (Table 1).

In order to protect the system from various AC and other noise sources, the EEG is transmitted to the host computer via a wireless connection. Wireless data transfer replaces long cables to electrodes and limits cable wobble, which introduces signal noise and artifacts [1]. Other systems based on a single ADS1299 have different microcontrollers such as SAM G55 [17] for continuous patient monitoring with schizophrenia, and STM32F4 [9] for different brain–computer applications, sending data over Bluetooth. The Bluetooth transfer of the EEG data due to the low power device consumption has bottlenecks, such as a shorter connection range and low data rates at higher EEG channel numbers or higher sample rates. In the EEG system [15], the bottleneck is changed by higher-bandwidth Wi-Fi 802.11 communication technology [15], but the limiting factor is the speed of the microcontroller (MCU). Other decisions include the use of two communication components: (1) a Bluetooth mini module HM-11 BLE 4.0 and (2) a Wi-Fi module ESP8266-12E through universal asynchronous receiver-transmitter and serial peripheral interfaces (SPIs) [19,21]. A wireless communication protocol based on SPI was also implemented in 16-channel EEG systems with ADS1298 AFEs and a 16 MHz Atmega2560 MCU [19], as well as ADS1299 and an 84 MHz ATSAM3X8E MCU (based on the ARM Cortex-M3 processor) [23]. In these EEG systems, the rejection ratios of the common mode (CMRR) for signals up to 70 Hz were −97 dB for ADS1298 [19] and −110 dB for ADS1299 [23]. Since the high channel number usually implies a low sampling rate under the available transmission bandwidth, these systems can rarely work simultaneously with high channel numbers and high sampling frequencies (Table 1).

2.2. Electrode Sensors and EEG Cap

Both EEG sensors (dry and wet) could be appropriate to EEG data acquisition for systems with 24 bit analog-to-digital convertors (ADCs) because of their high input impedance (1 MΩ) and CMRR (–97/110 dB), with achievable sampling rates up to 1 KHz [1,24]. The advantages of dry sensors were greater than those of wet electrodes in recent studies [24,25]. The application of traditional wet-sensor systems is largely limited by their difficult portability and long preparation time to outdoor settings.

3. System Architecture of the EEG Device

3.1. General Appearance

The functional block of the developed EEG device comprised the transmitting and receiving modules (Figure 1 and Figure A1). The transmission module consists of an analog front-end (AFE), a microcontroller, and a transmitter. Communication between the AFE and the software platform in the microprocessor part occurs with a transceiver via an SPI for program control.

The receiving part of one module included a receiver, a microcontroller, and a USB communication module (Figure 1 and Figure A1). The development of the real-time EEG data acquisition system was based on wearable modules with battery-powered, analog-to-digital, 24 bit resolution preamplifiers with conversion chips (ADS129x) that had a sampling frequency of up to 1 ksps per channel and low-power microcontrollers. They send data via wireless transmission to a host high-performance central processing unit (CPU) from 40 dry electrodes during tasks. The system was validated in a 12-channel EEG neurofeedback study in a school environment through research of typical-reading children and those with DD.

3.2. Transmission Module

3.2.1. Analog-to-Digital Convertor Module

The main focus was on the choice of the analog front-end (AFE), which supports a maximum number of channels with a high resolution for measurements of EEG and electrocardiogram (ECG) signals and good noise reduction capabilities. The AFE design enables the initial acquisition of analog biosignals by electrodes and their digital conversion (Figure 1 and Figure A1).

A conventional board for EEG data acquisition was founded on high-precision modules (ADS1298/9) to convert the data from analog to digital format [26]. The ADS129x enabled the elaboration of size-reduced medical systems with scalable characteristics and low power and total cost. The ADS129x AFEs were used in a multi-channel, battery-powered EEG system with significantly reduced board space compared with alternative EEG AFEs that provided fewer channels [18] and required additional external components such as anti-aliasing RC filters [26] and a high-Q active notch filter [18] for noise reduction. The choice of ADS129x in the EEG scheme implemented in the current study was mainly guided by the requirement of minimum power consumption when collecting weak electrophysiological signals (Figure 1 in [26]). The ADS129x AFEs were ADC chips programmable by configuring their registers. An on-chip internal reference circuit (right-leg-drive (RLD) amplifier) provided reference voltages to the device. The power supply module could be configured to 3.3 V, 0 V, or 5 V. The low-consumption device could be powered by 3.3 V, which was suitable for the battery power. Including more than one Wilson central terminal amplifiers reduced the total noise because of the noise averaging from the passive summing network at the output of these amplifiers. The capabilities of the right-leg-driving (RLD) circuit better reduced the noise and counteracted general interference, such as from power lines, and electromagnetic radiation from various radio transmitters and cells in mobile phones and routers. All in-phase LCD circuits were significantly reduced.

For our development, we chose implementation in the Thin Quad Flat Packages (TQFP) build. The connection to multiple devices was applied as shown in the technical overview of the ADS129x (Figure 45 in [26]). A “daisy chain” configuration was applied to connect the ADS1298 devices (Figure A2). Communication with the microcontroller was conducted via an SPI port (Figure A1). Most functions of ADS129x were implemented through five pins, such as configuration and control through the registers, the transfer of digital data, and leading out to a header for a convenient wireless connection (Figure A1). When multiple devices were used, all devices in the chain operated in the same register setting, and DIN was shared, reducing the SPI communication signals to four regardless of the number of devices (Figure A2). This architecture allowed a faster SCLK rate speed.

3.2.2. Transmitter

In the current study, a 2.4 GHz nRF24L01 transceiver module for low-power applications (1.9–3.6 V) accomplished wireless communication with a laptop (Figure 1 in [27]). This module was configured and controlled via a serial peripheral interface (SPI). Various operating modes of autonomous protocols are supported via a built-in baseband protocol engine, reducing system costs by managing all high-speed link layer operations. Data flow is provided by internal FIFOs between the MCU and the radio front-end with Gaussian frequency-shift keying (GFSK) modulation (Figure 1 and Figure A1). The nRF24L01 supported a configurable air data rate (up to 2 Mbps), frequency channel (2.4–2.4835 GHz), and output power with a programmable range (0, −6, −12, and −18 dBm) and a wide power supply range (1.9–3.6 V; Figure 1 in [27]). Power-saving modes, combined with high air data rates in the nRF24L01 low-power design, were appropriate for the current study. The packet-based data characteristics (enhanced shock burst) of nRF24L01 enabled high-performance communication with cheap host microcontrollers and achieving very low power [27]. The nRF24L01 stayed in TX active mode during transmitting packets. In an urban environment, the power provided a transmitter-receiver distance of several tens of meters. It was possible to operate several transceiver modules simultaneously without them interfering with each other.

3.2.3. Microcontroller

The choice of a high-performance central processing unit (CPU) was inappropriate for the present battery-powered EEG device. The selected CPU should reach an analog front-end sampling speed of 1 ksps to communicate with the device’s wireless modules. A microcontroller PIC18F2420 [28] was chosen for a CPU at 16 MHz. It can operate at a supply voltage of 3.3 V, with the low power consumption in operating mode being up to 40 MHz. The microcontroller can switch to sleep mode (1 MHz). Its hardware SPI port made programming the system very easy. The supply current was passed through the pin of the bypass capacitor and then to the supply pin for a more effective bypass. Potential crosstalk between multiple ADCs on the same printed circuit board (PCB) was minimized by broad power-supply traces. Minimizing the power-supply and reference-return trace couplings involved routing both traces separately on the AVSS pin as a star connection.

Overall, the fine performance of the ADS1298/9 comprised high input impedance and satisfactory temporal accuracy of simultaneous acquisition of multi-channel EEG data.

3.2.4. Receiving Block

The receiving unit was implemented based on a commercial module USB to nRF24L01 [27]. During the active RX mode, the nRF24L01+ radio, as a receiver, presents the demodulated data to the baseband protocol engine. It constantly searches for a valid packet, which can be cut by the protocol engine, when an address and a valid cyclic redundancy check match. The packet payload is presented in a free slot at the RX FIFOs (Figure A1). A received power detector (RPD) signal is available when RF signals >−64 dBm are inside the receiving frequency channel. The state of the RPD is latched upon receiving a valid packet, which then indicates the signal strength from our transmitter.

The module was a complete device that plugged directly into a computer’s USB port. All settings were determined using AT commands to a serial interface for cellular modules (so-called modem instructions). The communication set-up software package called USB nRF24L01 + Config v0.9 through a USB adapter is a free web-based tool. The USB communication was based on the CH340, a widely available chip with an available free driver running on all versions of Windows.

3.2.5. Software

Three ADC boards were plugged together for 40-channel EEG acquisition. The wirelessly transmitted data from the hardware front-end to the host computer made the system portable. This allowed the connection of up to 3 modules of 16 channels to the PC via different receivers connected to the USB ports on the PC. The instructions for the PIC18F2420 were supported by the software tools for the family of devices (MPLAB environment, MPLAB C18 C-compiler, and MPASM assembly language [28]). The main program takes a sequence of ADS129x task-related converted data, standby mode, input common mode, and reading and writing registers, controlling the ADS129x operations including acquisition, translation, and formatting of data. Data transmission to the software platform was performed by the standalone program.

The core functions of the software were data storage and real-time signal visualization (Figure A3). The software package provided reading, visualization, and recording of up to 40 channels. Algorithms for online analysis of the signal were implemented to provide real-time, stimulus-generated feedback, and functioning during neurophysiological tasks to create visual and auditory stimuli. Through temporal synchronization of stimuli events and data recording, the EEG signals and the task conditions can be compared with each other for further analysis. The real-time trace visualization of electrophysiological signals enabled the user to control the program through receive (”Start” and ”Stop”) and record buttons (”Stimulus” and ”NF”) (Figure A3). The generated stimuli, relevant to the external input, elicited EEG responses to certain experimental events synchronized to the EEG time stamps. The system implemented completed hardware and software EEG signal acquisition in various tasks. The stimuli were generated in C/C++ software on an NVIDIA GeForce GTX 1080 graphics card (Dell Alienware). At the same time, the acquisition front-ends for the EEG data recorded the 40 channels and the corresponding timestamps on a laptop (Acer Predator). Each dataset was constructed using EEG time stamps in the signal aligned to the stimulus events. A low-cost expansion of the EEG system for neurofeedback application was adapted for broad public use. The signal processing by the software was an important part, being the hardware component of the NF-EEG system where data collection takes place for high-speed performance in C/C+ and real-time signal processing in MATLAB [1,22].

For validation of the EEG system, a comparison of the data collection of the proposed system with the controlled data of the Nikon Kohden EEG system was performed (Figure 2). Both datasets correlated well, showing good agreement for the data in the time and frequency domains.

The experiments showed an average of a 25 dB SNR over 10 trials (20 μV EEG signal) which contained 250 × 10 samples with a gain of 12 and sampling frequency of 250 Hz (Figure 1B; gains 1–12). The internal noise in the AFE analog circuit was quantified by analyzing the input-referred noise with short positive and negative analog inputs. The experiment showed the peak-to-peak and root mean square input-referred noise (Vpp = 2.29 μV; Vrms = 0.34 μV) under the data reported in the manual of the ADS129x AFE (2.4 μV; 0.4 μV) at a sampling frequency of 250 Hz and bandwidth of 0.01–70 Hz of the signal.

3.2.6. Electrode Sensors and Cap

The wireless system with dry sensors was more applicable in outdoor environments (Figure 3 and

Table 2. Specifications of the dry sensors in the wireless EEG acquisition system.

Characteristics	Spring-Loaded Sensors	Foam Sensors
Impedance	200–500 KΩ	200–500 KΩ
Size (mm)	15 × 15 × 14	15 × 15 × 14
Weight	1.8 g	0.8 g
Position	Hairy area	No hair area

). The time to prepare such a system was less than 5 min without using conductive gel and cleaning the skin. Each dry EEG sensor was configured as a matrix with a stellated geometry of 16 long-length golden and spring-loaded pins that better contacted an irregular scalp [29]. Other foam-based sensors made of a polymer and covered in electrically conductive fabric were used as ground and reference electrodes. Both types of sensors achieved low impedance of the skin–electrode interface (less than 5 kΩ). The dry sensors were built permanently into soft elastic 64-channel configured EEG caps [30] with three sizes (56, 58, and 60 cm), chosen in accordance with the child’s head. The adjustability of the EEG caps is a limitation of most devices, requiring different headsets to collect data for people with scalp size discrepancies that increase the cost of the system.

4. Materials and Methods

4.1. Participants

A controlled comparison of an experimental group with controls included typical readers (n = 36; 26 boys and 10 girls; mean age: 8–9 years; SD = 0.58) and children with developmental dyslexia (DD, n = 72; 52 boys and 20 girls). Children with disabilities unrelated to reading did not take part in the study (Table S1). The controls were from the same school grade and sociodemographic background without dyslexia and without learning or language difficulties (Table S1). The criteria to include 8–9-year-old children in the experimental group were the following: (1) DD diagnosed by a school logopedist and psychologist under the DD classification (Table S1); (2) several brain regions with abnormal waves; (4) completing 12 qEEG z-NF sessions and visual training; and (5) an intelligence quotient >85 according to the Raven intelligence scale (Table S1). These pretest sessions were conducted to identify children with core (magnocellular) visual deficits and the multi-sensory integration deficits underlying dyslexia. Intervention to be effective must be tailored to defective skills. Children with DD were assigned to a pre-training experimental group (preD) with six qEEG z-NF intervention sessions over a one-week period (for each hemisphere). The control group also had six neurofeedback sessions. Over a period of 4 months, a visual-based training procedure (VT, Supplementary Materials) and six other qEEG z-NF sessions were administered to a treatment group with DD (post-training group (postD)). The behavioral measures of reading included word reading, accuracy, and speed measured at the beginning and end of the four-month period (intervention). The study used a pretest-training-post-test design to assess the training effect of combined programs (12 completed qEEG z-NF sessions and visual training) based on visual tasks, with a specific focus on gains in skill-related brain functioning for reading.

4.2. NF-EEG Brain Locations

The 40-channel EEG placement was according to the standard 10-20 and 10-10 systems [31,32] and separated sensors for the ground, sensors placed on the forehead, and reference sensors in both mastoid processes (

Table 3. Forty sensor locations covering adjacent brain areas. Twelve NF sensors (eight areas) are marked in bold.

		10-20	System
Gyrus	Abbreviation	Electrode (L-R Site)	Closest Brodmann Areas	Functional Name
middle frontal gyrus	MFG	Fz, F3-F4	BA8	intermediate frontal cortex (includes frontal eye fields)
inferior frontal gyrus	IFG	F7-F8	BA45/47	Broca’s area, orbital frontal cortex
postcentral gyrus	PSTCG	C3-C4	BA123/6	primary somatosensory and motor cortices
posterior part of superior temporal gyrus	STG	T7-T8	BA21/22	MT and MST
inferior parietal lobe	IPL	P3	BA39/7/19	angular, precuneus, associative visual V3 area
inferior parietal lobe	IPL	P4	BA39/40/7	angular, supramarginal, gyri, precuneus
middle occipital gyrus	MOG	O1-O2	BA19	associative visual V3 area
precentral gyrus	PRECG	Cz	BA4/6	primary motor and premotor cortices
superior parietal gyrus	SPL	Pz	BA7	precuneus, 7P
inferior temporal gyrus	ITG	P7	BA37/19	lateral occipitotemporal gyrus, adjacent to posterior fusiform, lingual cortex V5 and V3
inferior temporal gyrus	ITG	P8	BA37	occipitotemporal gyrus
		10-10	system
superior frontal cortex	SFC	AF3-AF4	BA9	dorsolateral prefrontal cortex (DLFC)
anterior part of the inferior temporal gyrus	ATG	FT9-FT10	BA20/BA38	temporal pole, areas TE/AIT
middle frontal gyrus	MFG	FC3-FC4	BA6	premotor and supplementary motor cortices: pre-SMA and SMA
inferior frontal gyrus	IFG	FC5-FC6	BA44/45	opercular and triangular parts of Broca’s area
precentral gyrus	PRECG	C1-C2	BA4/6/123	primary motor, premotor, and somatosensory cortices
postcentral gyrus	PSTCG	C5-C6	BA123/40/43	primary somatosensory cortex and supramarginal gyrus with extension into the Sylvian fissure to PFo
superior parietal gyrus	SPL	CP1-CP2	BA5/7	areas PGa, 7A, and 7PC, LIP
inferior parietal lobe	IPL	CP3-CP4	BA40/123	supramarginal gyrus (subareas PFt and PFm; ventral intraparietal sulcus VIP or IPSmot)
middle temporal gyri	MTG	TP7	BA21/37/22	MT+, V5, and MST
middle temporal gyri	MTG	TP8	BA21/22/37/ 20	medial superior and middle temporal areas, lateral occipitotemporal or posterior inferior temporal gyri, adjacent to posterior fusiform or lingual gyrus; MST, MT+, V5, and V4
superior occipital gyrus	SOG	PO3	BA19/18/39/7	dorsal visual cortex or parieto-occipital sulcus, angular gyrus, and precuneus; pIPS, V3A, and POs
superior occipital gyrus	SOG	PO4	BA19/18/39	dorsomedial parietooccipital visual V6 and V6A, ventral portion of posterior intraparietal sulcus with dorsal portion of retinotopical V3A and V7
middle occipital gyrus	MOG	PO7-PO8	BA18/19	ventral visual cortices V3v and V2

and Figure 4). The stretch cap held the 40 embedded sensors that were prepositioned in the international montages, which minimized electrode placement errors through measurements (10% and 20% as well as 10% and 5%) of the distances between the skull landmarks along the midline from the nasion (Nz) to the inion (Iz) and from the left to the right pre-auricular points (A1 and A2, respectively) correctly aligned in the horizontal and vertical planes with placement Cz. The EEG sensors (10-20: Fz, Cz, Pz, O1-O2, and Oz; 10-10: FT9-FT10, TP7-TP8, PO7-PO8) for the NF experiment are underlined in Table 3. Their choice was according to the functional connectivity studies [6,7] of the typically developed children. We hypothesized that qEEG z-NF may lead to a functional reorganization in DD children that consists of a reduced interhemispheric low-frequency theta band and an increased alpha band in the regions involved in the reading network in the temporal lobe, occipital cortex, and those which intervene in the articulation and gesticulation of words as frontal and precentral cortices.

4.3. NF-EEG Experiment

The reinforcement technique (neurofeedback) for the magnocellular pathways in DD children consisted of visual stimuli with different contrasts vertically flicking in an external noise field (Figure 5 and Figure A3C).

The discrimination of different contrast Gabor gratings with illusionary doubling motion (LSF illusion: flicker with a 15 Hz counter phase and low spatial frequency 2 cycles per visual angle degree) consisted of the following conditions: a low-contrast level 6% and high-contrast level 12% of the threshold level of 50%, defined in previous psychophysics experiments [33]. The stimuli were presented in the center on a monitor and were arranged in a 40-trial pseudo-randomized series with 1.5–3.5 s intervals between stimuli (20 trials/condition). A fixation white cross appeared for 100 ms at the start of each trial. After 200 ms, the LSF illusion appeared for 200 ms at a viewing distance of 210 cm from the monitor (Figure 5). A colored cross was presented on the monitor for 100 ms in red when the EEG prints were below the set threshold and in a green sign when the EEG stamps were above it. The child determined the illusion with low contrast by pressing a button with the left hand and the illusion with high contrast by pressing a button with the right hand. The trial’s end was marked with a brief beep. A response time period comprised the period of 100–1500 ms from the stimulus onset. The acquisition of the EEG data was synchronized with the stimuli onset. Trial-by-trial neurofeedback was perceived as a different-colored cross.

4.4. NF-EEG Procedure

The developed system was validated via a 12-channel EEG neurofeedback study in a school environment through research on typical-reading children and children with DD (Figure A3). Neurofeedback via evoked EEG oscillations can improve the performance of a child with DD, compensating for their magnocellular deficit during the visual LSF illusion task [2].

The real-time z-score neurofeedback device (qEEG z-NF) provided a 24-bit conversion with a 256 Hz sampling frequency and second-order notch filtering for removing the powerline noise at 50 Hz, as well as an EEG bandwidth of 0.1–70 Hz with a type II Chebyshev filter. The time-locked trials during the presentation of the flicking Gabor gratings (~250 ms) were frequency-decoded via a Morlet wavelet (MATLAB) in δ: 3–4, θ: 4.5–8, α: 8.5–12, β: 13.5–20, 20.5–30, and γ: 30.5–48, 52–70 Hz. The qEEG z-NF α/θ scores were determined from the α and θ frequency peaks assessed in each hemisphere for eight brain regions (Table 3). A mean baseline with a duration of 2.5 min at the first experimental feedback-free session with the eyes opened for each child was applied as a measure to define the z-score’s threshold in the software norm data [34]. The absolute qEEG standardized index (|prior z-score| ≥ 1.0) was used as the target z-score, defined by logarithmic transformation of the qEEG data features to obtain the optimal mean and standard deviation. It minimized the influence of variables affecting the EEG characteristics by tending toward a standard normal distribution with a mean of zero and variance of one. Several brain areas (Table 3) and their abnormal α/θ scores were enrolled as candidates for feedback in studies of dyslexia [6,7]. During the sessions, the z-scores for the seven bands for the selected channels were computed moment by moment during the stimulus duration. Their absolute pre-z-scores ≥1.0 were highlighted as being the targeted z-scores by site and frequency.

The z-α/θ scores were either up- or down-trained to approximate to the normal score. This allowed specific underlying areas of the brain to be targeted for training and could lead to a reduced number of sessions. The NF latency, or the delivery time of the EEG trial to the observed qEEG-z-NF sign on the monitor, determines the schedule of reinforcement so that it significantly affects the outcome of the operant conditioning [35]. The child received positive feedback as a visual cue when the z-scores fell within a “target window” (±1 standard deviation of the normative mean) and the z-score that was outside of the target window moved toward the target window. The visual feedback was marked with a green sign when the z-score was above the threshold, and the mark was a red sign when the z-score was under the threshold. The qEEG z-NF protocol was performed for 10 min (for each brain hemisphere) in 6 sessions twice for the DD group before a visual perceptional training procedure [33] and 4 months after that, and one procedure was carried out for the controls. The mean pre- and post-assessment qEEG z-NF sessions was 12. After preprocessing, the average artifact-free trial number was 772 epochs of data for each group or condition.

The protocol included (1) the left-hemispheric θ decrease and α-wave augmentation at regions adjacent to the reading functional network (left inferior frontal, posterior superior temporal, ventral occipitotemporal, and middle occipital cortex) and (2) the compensated left-hemispheric α/θ score increase and right-hemispheric α/θ decrease in brain laterality for the dyslexic group from reducing the slow θ waves to reach a new learning information level from the brain.

4.5. Statistics

Nonparametric bootstrap tests compared the threshold-selected z-scores for pairs of brain locations of the groups (controls vs. pre-training group with DD, controls vs. post-training group with DD, and pre-training vs. post-training group with dyslexia) and the hemispheres (left vs right side), with subsets of 1000 permutations for each comparison. The waves with |z-scores| higher than the group standard deviation were categorized as outliers, which were excluded from the comparisons. The Bonferroni correction for the multiple comparisons was applied to the significance level in the permutation tests. For behavioral analysis of the task (correct responses, response times, and reading achievements), non-parametric Kruskal–Wallis tests were performed in pre-post conditions.

4.6. Results

4.6.1. Behavioral Results

In both conditions of the LSF illusion task, the pre-training dyslexics had slower responses and fewer correct responses than the other groups (Figure 6). Before training, the children with dyslexia (preD) had more difficulty discriminating the low-contrast illusion than the other groups (Con and postD; Figure 5). The postD group showed higher percentages of correct responses and faster response times in the low-contrast condition after qEEG z-NF-VT training in the within-group comparison (preD vs. postD, p = 8.7 × 10⁻⁷, χ² = 24.79; Figure 5). The improved discrimination accuracy and reaction time were observed in the post-D group compared with the other groups (p < 0.015, χ² > 5.94).

4.6.2. Reading Results

The improvements in the reading speeds were twice as high after four months of qEEG z-NF-VT training with three times fewer missed words and errors (p = 0.001; Table S2). The word reading span also improved after training (p = 0.001). Before training, the dyslexics showed slower voice response times (1497.96 ± 30.5) during the word reading compared with post training (1266.16 ± 27.9 ms). After training, there was a significant improvement in voice response compared with before training (p < 0.0001, χ² = 30.4; Table S2). Their performance significantly improved in the visual words and pseudoword discrimination task, reducing reading time and increasing words read correctly (p < 0.03; Table S2). The dyslexic group showed higher success rates and slower reaction times compared with both words and pseudowords after training (p < 0.03, χ² = 4.51 for words; p = 0.02, χ² = 5.29 for pseudowords). Their reaction times did not change significantly (Table S2).

4.6.3. z-NF Training Results

The qEEG z-NF α/θ scores for all sessions are presented in Figure 7 for all groups (

Table 4. Statistics of α/θ scores between sensors of the groups for qEEG z-NF trainings (p < 0.05).

Scores at Sensors	Controls	Pre- Training Dyslexics	Post- Training Dyslexics	Controls vs. Pre-Training Dyslexics		Controls vs. Post-Training Dyslexics		Pre- and Post- Training Dyslexics
Low-contrast		mean ± s.e.		p	χ2	p	χ2	p	χ2
Fz	2.06 ± 0.04	1.65 ± 0.02	2.47 ± 0.05	1.9 × 10-10	40.8	2.4 × 10−11	44.6	2.7 × 10−40	176.6
Cz	2.10 ± 0.04	1.73 ± 0.01	2.66 ± 0.08	0.0010	10.6	0.0131	6.2	5.2 × 10−4	12.1
Pz	2.27 ± 0.05	2.21 ± 0.04	3.44 ± 0.08	0.0039	8.34	3.3 × 10−40	176.2	3.7 × 10−32	139.4
Oz	2.48 ± 0.06	1.97 ± 0.04	2.00 ± 0.03	4.0 × 10−17	70.8	7.2 × 10−6	20.2	2.0 × 10−11	44.9
FT9	1.68 ± 0.02	2.12 ± 0.03	2.13 ± 0.03	1.3 × 10−22	95.7	4.9 × 10−33	143.4	0.303	1.06
TP7	2.26 ± 0.05	1.82 ± 0.02	2.62 ± 0.07	1.9 × 10−5	18.3	7.2 × 10−10	37.9	2.2 × 10−24	103.8
PO7	2.12 ± 0.05	1.64 ± 0.02	2.82 ± 0.08	2.7 × 10−9	35.4	2.4 × 10−15	62.7	3.7 × 10−38	166.8
O1	2.13 ± 0.04	2.01± 0.03	2.27 ± 0.05	0.4743	0.51	0.0339	4.5	0.193	1.71
FT10	2.08 ± 0.04	2.57 ± 0.06	2.77 ± 0.07	1.9 × 10−14	58.6	4.7 × 10−11	43.3	0.79	0.07
TP8	2.50 ± 0.06	2.06 ± 0.03	2.43 ± 0.04	5.1 × 10−4	12.1	8.8 × 10−4	11.1	1.2 × 10−9	36.9
PO8	2.73 ± 0.07	2.21 ± 0.03	2.28 ± 0.03	6.1 × 10−7	24.9	0.202	1.63	0.0015	10.1
O2	2.10 ± 0.03	2.20 ± 0.03	2.50 ± 0.04	0.1045	2.64	3.4 × 10−18	75.3	1.9 × 10−11	45.1
all sensors	2.27 ± 0.04	1.98 ± 0.03	2.52 ± 0.05	7.2 × 10−6	20.1	7.8 × 10−4	11.3	1.3 × 10−12	50.3
midline	2.10 ± 0.04	1.95 ± 0.03	2.28 ± 0.05	0.705	0.14	1.6 × 10−22	95.4	1.5 × 10−12	50.1
left side	1.99 ± 0.04	1.89 ± 0.03	2.64 ± 0.08	0.0357	4.41	3.4 × 10−7	25.96	1.5 × 10−12	50.1
right side	2.33 ± 0.05	2.22 ± 0.04	2.49 ± 0.05	0.187	1.74	1.5 × 10−5	18.73	0.0018	9.7
left vs right p; χ2	1.2e-12; 50.5	2.3e-06; 31.2	0.192; 1.7
High-contrast
Fz	1.97 ± 0.03	2.34 ± 0.06	2.02 ± 0.03	2.4 × 10−4	13.5	6.2 × 10−4	11.7	2.5 × 10−9	35.5
Cz	3.24 ± 0.09	1.70 ± 0.02	2.81 ± 0.07	5.6 × 10−49	216.4	0.002	9.5	3.5 × 10−38	166.9
Pz	1.92 ± 0.02	2.44 ± 0.06	2.84 ± 0.04	2.0 × 10−6	22.6	9.9 × 10−62	274.9	1.8 × 10−28	122.5
Oz	2.26 ± 0.05	2.10 ± 0.03	2.21 ± 0.03	0.215	1.54	0.002	9.77	0.008	7.02
FT9	1.98 ± 0.04	1.51 ± 0.02	2.40 ± 0.06	1.8 × 10−18	76.9	8.0 × 10−8	28.8	4.4 × 10−45	198.5
TP7	2.04 ± 0.03	1.66 ± 0.02	2.58 ± 0.06	1.4 × 10−28	122.9	6.7 × 10−4	11.6	1.9 × 10−42	186.4
PO7	1.93 ± 0.03	1.76 ± 0.02	2.98 ± 0.08	1.4 × 10−7	27.7	2.0 × 10−22	94.9	8.6 × 10−40	174.3
O1	2.45 ± 0.06	2.49 + 0.07	2.15 ± 0.04	0.062	3.49	0.040	4.21	1.2 × 10−8	32.4
FT10	2.27 ± 0.04	2.30 ± 0.04	2.70 ± 0.06	0.032	4.61	5.3 × 10−4	12.0	3.2 × 10−9	35.1
TP8	2.50 ± 0.06	2.23 ± 0.03	2.27 ± 0.04	0.007	7.30	0.048	4.61	0.613	0.26
PO8	2.33 ± 0.04	2.59 ± 0.05	2.27 ± 0.06	8.8 × 10−5	15.4	8.6 × 10−6	19.8	1.2 × 10−19	82.2
O2	2.47 ± 0.06	2.16 ± 0.04	2.12 ± 0.03	4.6 × 10−5	16.6	9.4 × 10−20	82.7	0.12	2.5
all sensors	2.30 ± 0.05	2.11 ± 0.05	2.44 ± 0.06	2.2 × 10−6	22.4	0.003	8.86	1.1 × 10−4	14.9
midline	2.13 ± 0.05	1.82 ± 0.03	2.34 ± 0.05	1.3 × 10−5	18.9	0.017	5.66	1.1 × 10−11	46.2
left side	2.02 ± 0.04	1.63 ± 0.03	2.51 ± 0.06	9.3 × 10−20	82.8	3.9 × 10−8	30.2	4.1 × 10−43	189.5
right side	2.35 ± 0.04	2.30 ± 0.04	2.40 ± 0.05	0.18	1.89	0.15	2.1	0.94	0.01
left vs right p; χ2	9.7e-10; 37.4	1.3e-61; 274.3	0.131; 2.29

Note: A comparison of z-scores between (1) pairs of brain locations of controls vs. pre-training group with DD; (2) controls vs. post-training group with DD, (3) pre-training vs. post-training group with dyslexia, as well as a comparison between aforementioned group’s pairs of (4) all z-scores, (5) z-scores at the midline sensors, (6) left-hemispheric z-scores, and (7) right-hemispheric z-scores. Hemispheric comparison of sensors for each group (last row in each condition). The highest z-scores corresponding to Figure 7 are indicated in bold.

).

In both conditions, the α/θ z-scores of the controls (Con) were significantly higher than those for the children with dyslexia before training (preD) in more brain locations (Con vs. preD: p < 7.2 × 10⁻⁶, χ² > 20.1; average score of all sensors: mean ± se, Con: 2.27 ± 0.04; preD: 1.98 ± 0.03; Table 4). There were no differences between the Con and preD groups for the z-score of the midline, right-side sensors, (O1-2) at low contrast (p > 0.1045, χ² < 2.64), or for the z-score of the right-side sensors (Oz, O1) under the high-contrast condition (p > 0.062, χ² < 3.49). In both contrasts, significantly higher z-scores were found for the left-hemispheric sensors of Con compared with those of the preD group (p < 0.0357, χ² > 4.41; low contrast: Con = 1.99 ± 0.04; preD = 1.89 ± 0.03; high contrast: Con = 2.03 ± 0.04; preD = 1.63 ± 0.03).

Although the α/θ scores of the preD group were significantly lower than those of the other groups (Table 4), they had significantly higher scores than the Con group at FT9 and at the right-hemispheric FT10 and O2 (low contrast) and Pz, FT10, and PO8 (high contrast) while showing no differences at O1-2 (preD vs. Con, low) and O1-Oz (high contrast).

The α/θ scores of all post-training dyslexic sensors were significantly higher relative to the correspondent sensors of the preD group (postD vs. preD: p < 0.008, χ² > 7.02; Figure 7 and Table 4). There were no differences at the occipital and temporal cortices between the preD and postD z-scores at O1 and FT9-FT10 (low) and at TP8-O2 (high contrast).

For both conditions, the right-hemispheric α/θ scores of the controls and dyslexic children before training were significantly higher than the left-hemispheric scores (left vs right side: p < 2.3 × 10⁻⁶, χ² > 31.2; Table 4), whereas a non-significant hemispheric difference in the post-training dyslexic group was found (postD left vs right side: p > 0.131, χ² < 2.29; Table 4).

In a condition of low-contrast discrimination, the highest α/θ scores were found in the postD group at FT10, TP7, Pz, PO7, which covered the anterior and middle temporal, superior parietal, and middle occipital areas, in the Con group at TP8, PO8, and Oz, and the preD group at FT10, Pz, PO8, and O2 (Table 4) among the sensors in a group. In the high-contrast discrimination, the highest α/θ scores were found for the postD group at PO7, Pz, and Cz, in the Con group at Cz, TP8, and O1-O2, and in the preD group at the PO8 and O1 across the sensors in a group (Table 4).

Lower α/θ z-scores at areas adjacent to the precentral gyrus and bilateral cuneus (Cz and Oz under high contrast) were found in the preD group compared with the Con group, similar to the reading tasks [6]. The preD α/θ scores were elevated at the anterior temporal cortex, related to efforts by compensatory mechanisms for the deficits in visual and motor processing (Figure 7). The postD group improved the α/θ oscillations at the premotor, primary motor, right anterior temporal, right occipitotemporal (TP8), and left middle occipital (PO7) cortices due to their more active involvement in the NF sessions (Figure 7). The Con vs. postD group comparisons showed that the qEEG z-NF also produced better effects in the occipital, middle temporal, and precentral areas of the controls (low contrast: TP8 and Oz; high contrast: O1-2 and Cz; Table 4).

The sensor statistics for the postD group (vs. Con) yielded significantly higher α/θ scores at the SPL, left MOG, the MFG (low contrast: Pz and PO7; high contrast: Fz) and lower α/θ scores at the right middle occipital cortex (high contrast: PO8-O2). The middle temporal, lateral occipitotemporal, dorsal associative visual, and occipital areas in the post-training children with DD contributed to the compensatory α and θ mechanisms, reducing right-hemispheric dominance over the left. The intra-hemispheric α/θ scores significantly increased at the middle temporal and occipital cortices (both conditions: TP7 and PO7) and the middle frontal, precentral, and superior parietal cortices (low contrast: Fz, Cz, and Pz) for the post-training group (vs. Con and preD).

5. Discussion

The proposed EEG system design, which is easily scalable and adaptable to various EEG tasks, explored the α/θ z-NF effects during discrimination of different contrast LSF illusions on children with DD.

Mechanisms compensating parvo- or magnocellular deficits were induced during the qEEG z-NF training. By simultaneously reducing excess θ frequencies and increasing power at the α frequencies, these mechanisms normalized the frequency power spectrum associated with the specific experimental conditions. Specific frequency changes, such as decreased α/θ scores and α oscillations, were found in the posterior areas of the right hemisphere. Other changes under increased α/θ scores (e.g., θ activity decrease and α activity increase) were observed in the anterior areas. The postD group was characterized by emphasized α/θ (α power) increases at the posterior regions in the left brain hemisphere compared with the other groups.

Low pre-D α/θ scores were observed during the first NF sessions from the left MOG to the MFG (low-contrast LSF illusions) and from the left MTG to the ATG (high-contrast LSF illusions). Recent NF training sessions resulted in higher post-D α/θ scores at the left MOG across the SPL to the PRECG (low- and high-contrast conditions). The reading skills were also improved after NF training of children with disabilities in learning (Table S2). Studies of EEG-based NF have shown that the combination of qEEG-zNF and other treatment approaches leads to better outcomes than an NF alone [36]. Combining the training effect of NF on the primary visual deficit of DD with the effects of visual perceptional training on specific visual deficiencies for dyslexia [33,36] improved reading by stimulating the connectivity of weakly connected brain regions in the reading network. The design of the EEG system was convenient to apply in schools. Although the additional visual training for specific DD deficits extended the training period, it still fell short of age-related maturation. Therefore, the NF contribution of the protocol to improving the cognitive and reading skills of children with DD was combined solely with the contributions of the visual training and could not be attributed to age-related maturation or other school learning.

The precentral, left ATC, SPL (precuneus), and lateral occipital cortex showed significant changes as a result of training in the postD group. The changes in the middle occipital cortices and the precuneus were significantly affected by the task features. Under the discrimination conditions, the SMA and the postcentral cortex reduced their activation, as did the bilateral precuneus, revealing the sensorimotor and visual processing deficits in the preD group, respectively. During reading tasks, reduced activation was also found in the SMA and the right IPL in the DD group [6]. The increased activity in the IFG and ATC in the preD group could be explained by more executive efforts and the low automatization of the tasks (LSF illusion and reading). The reduced activity extending posteriorly from the right postcentral cortex to the SPL in children with DD (vs. the controls) could explain the reduced reading speed, occurring due to insufficient sensory feedback. The reduced activity in the right dorsomedial parietooccipital, bilateral preparietal, and superior parietal regions and the MOG of the preD group, compared with the controls, could be related to insufficient processing of visual-spatial information. The effective sensory stimulus processing and selective attention of the post-training dyslexics (for the high-contrast condition) and the controls (for the low-contrast discrimination) could be related to the decreased α frequency power and α/θ activity in the occipital areas [37,38], as well as in the frontoparietal network (high θ frequency activity). The high θ oscillations in the frontal cortex of the pre-training group related to the low-contrast condition could be associated with the requirements for more cognitive resources for the task’s performance. Additionally, the left MTG and left MOG of the preD group showed low α/θ scores (high θ power). The language processing was subserved in both hemispheres by the MST and its right side for visuospatial information storage [39]. Hence, the lower α/θ lateralization of MST may imply a compensatory mechanism in the post-training DD children [39]. During the stimulus, right-hemispheric high α/θ score (high α-band power) was observed in the preD group, while in the postD group, widespread α/θ growth was found in both hemispheres. In general, attention-relevant α activity was not restricted to visual stimulus processing [40,41], whereas pre-stimulus α power was modulated by accuracy and attention to stimulus processing. The low α/θ scores (increased θ-band activity) bilateral at the frontal, temporal, and parietal areas in the pre-training dyslexics could be related to mental fatigue, worse performance [42], and also aberrations in global brain organization of verbal and spatial cognitive skills [43].

The left-lateralized α/θ scores of the post-training dyslexics (α-band power [44]) correlated with focusing on processing local LSF elements [45], while the right-side high α/θ scores (high α activity) in the pre-training group correlated with the worse global attention bias. The selection of attention at the local or global level of information is governed by a cortical attentional mechanism established in the temporoparietal junction and depending on the visual task-related conditions [46]. Global form identification was associated with higher activation of the lingual gyrus in the right hemisphere and the local element identification with that of the inferior occipital gyrus in the left hemisphere [47,48]. The occipital cortex rapidly relays the coarse information about the LSF illusion through the parietal and frontal cortices (dorsal route) before spreading to the inferotemporal cortex, mediating LSF recognition (ventral route [49]).

Better α and θ activity of the postD group in the frontal, temporoparietal, and middle occipital regions was associated with detection from the visual areas of the LSF identification Gabor gratings during the neurofeedback. They recognize coarse LSF information more quickly through projections to the orbitofrontal cortex, which is strongly and reciprocally connected to the temporal cortex [50]. In the right hemisphere, the recognition of the fine LSF illusion elements preferentially occurred from the occipital cortex to the ITG [48], whereas the temporal and orbitofrontal cortices were involved during its coarse recognition, suggesting strong intra-hemispheric functional interactions. Through the top-down connections of the frontal, temporal, and parietal cortices to the visual cortex [47,51], the fine stimulus features were analyzed and selected, whereas the cerebral hemispheres were complementary in the postD group’s stimulus processing.

The better left-hemispheric α/θ scores of the postD group in the temporoparietal and middle occipital regions can be associated with an improvement in spatial frequency processing [52], while the better scores in the precentral and parietal cortices can be associated with an advancement of the temporal processing efficiency of the illusion through the magnocellular pathway in both contrast conditions [53].

6. Limitations

Due to the variable characteristics of the EEG signal, the development of a battery-powered mobile Wi-Fi NF-EEG system is challenging. Some professional EEG-NF devices are available from many vendor companies, but due to the high price, they are not intended for public use. Scaling and reconfiguring the hardware to allow for over 64 channels would reduce many of the technical characteristics of mobile systems. The only limitation of a stand-alone system with a high-performance processor operating at maximum load is the maximum battery life. Charging it complicates the collection of EEG data in these devices. Participant data security may be at risk on account of the sensitivity of wireless connections or due to any movement of cables. Therefore, wireless EEG systems could encrypt data prior to wireless transfer and use shorter sensor-ADC connections or a head mounting. Overall, the price of the system only increases due to the headset adjustability, requiring multiple headset caps for people with different head sizes.

7. Conclusions

The EEG acquisition system maintained significant savings due to its development of the presented low-cost modular board amplifiers. The proposed design allowed easy adaptation to various NF-EEG task applications, being scalable in regard to spatial and temporal resolutions. In this application designed for children in an outdoor environment, the need for an autonomous wireless system, research, and recording capability for NF information processing was evident. In the school-based setting of the children’s study, this system was small and portable with low power consumption for long-term operation. The EEG system had sufficient computing ability for online data transfer, processing, supported streaming, and event detection modes during various complex tasks.

The stimulated qEEG z-NF brain self-regulation of determinative brain regions has also been shown at the behavioral level to be more effective than NF regulation of a single brain area. An outcome of combined qEEG z-NF-VT training was functional brain reorganization and emerging neuroplasticity [2,36,53]. Through specific visual tasks (VTs), the qEEG z-NF alpha-theta network better affected other frequency networks related to the reading and transferred the learned effect to improvements in cognitive and reading abilities by strengthening weakly connected brain areas. Stimulating the dyslexia-unaffected remote brain areas in the ventral route by training can facilitate the affected regions in the dorsal route and hemispheric laterality.