Simultaneous EEG-fNIRS Data on Learning Capability via Implicit Learning Induced by Cognitive Tasks

Abstract

The development of real-time learning assessment tools is hindered by an incomplete understanding of the underlying neural mechanisms. To address this gap, this study aimed to identify the specific neural correlates of implicit learning, a foundational process crucial for skill acquisition. We collected simultaneous electroencephalography and functional near-infrared spectroscopy data from thirty healthy adults (ages 21–29) performing a serial reaction time task designed to induce implicit learning. By capturing both electrophysiological and hemodynamic responses concurrently at shared locations, this dataset offers a unique opportunity to investigate neurovascular coupling during implicit learning and gain deeper insights into the neural mechanisms of learning. The dataset is categorized into two groups: participants who demonstrated implicit learning (based on post-experiment interviews) and those who did not. This dataset enables the identification of prominent brain regions, features, and temporal patterns associated with successful implicit learning. This identification will form the basis for future real-time learning assessment tools.

Dataset: https://doi.org/10.17632/tsfs9fhn5y

Dataset License: CC-BY 4.0

1. Summary

Modern education faces the challenge of overwhelming information volume, making efficient learning strategies crucial and highlighting the importance of implicit learning skills. The ability to learn without explicit awareness, known as implicit learning, is crucial for mastering complex skills and making intuitive judgments, complementing the more structured nature of explicit instruction. This learning process typically unfolds incrementally across multiple exposures. Research has suggested that proficiency in implicit learning may predict a greater facility for acquiring similar skills, indicating its fundamental role in overall learning capability [1].

Despite its importance, understanding and assessing implicit learning in real-time remains a significant challenge. Traditional educational assessments, such as examinations and quizzes, are administered post-learning, offering delayed feedback that prevents immediate intervention. This temporal gap between the learning process and its evaluation constitutes a critical barrier to optimizing educational strategies. Our research confronts this challenge by leveraging multimodal neuroimaging. We aimed to construct a processing pipeline to systematically identify the neurophysiological markers of implicit learning. The dataset presented here comprises simultaneous EEG and fNIRS recordings acquired from participants performing cognitive tasks, providing a rich source of data to investigate the neural correlates of learning as it happens.

The potential for this line of inquiry is underscored by findings in related fields. For instance, numerous studies have demonstrated that neurofeedback training can help children with ADHD improve symptom regulation [2,3], illustrating the brain’s capacity for targeted training and enhancement. This inherent neuroplasticity suggests that a deeper understanding of the neural activities related to learning could unlock novel training paradigms. The dataset described herein was designed to provide such an understanding by capturing implicit learning events through simultaneous neural and hemodynamic measurements at identical cortical locations.

The combination of electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS) offers significant advantages over single-modality approaches. Both methods are portable and relatively inexpensive, facilitating replication and validation [4]. Critically, their characteristics are complementary. EEG provides exceptional temporal resolution, allowing for the detection of rapidly changing brain processes on the order of milliseconds, though its spatial accuracy can be limited by noise and signal propagation [5,6,7,8]. Conversely, fNIRS offers superior spatial resolution and is immune to the electromagnetic interference that affects EEG. While its measurement of the hemodynamic response is slower, taking several seconds to evolve, it is more effective at localizing the source of event-related neural activity [9,10,11,12].

By recording EEG and fNIRS data concurrently from the same locations, this dataset provides a more complete view of the underlying physiological processes, capturing both the fast electrophysiological activity and the slower, localized hemodynamic changes. This allows for a comprehensive investigation of neurovascular coupling and offers a richer understanding of brain function during learning. For biomedical engineers, this dataset serves as a valuable resource for developing sophisticated models capable of detecting implicit learning in real-time. Such a detection module could form the basis of a biofeedback system designed to modulate and enhance learning. For educators, these advancements could lead to highly personalized educational strategies, creating a more optimized learning process and ultimately fostering more effective learners.

2. Data Description

This dataset contains simultaneous EEG-fNIRS recordings from 30 healthy volunteers (23 male, 7 female, aged 21–29) with no prior learning disabilities. The data were acquired while participants performed a cognitive serial reaction time task designed to induce implicit learning. All participants provided written informed consent for the experiment, which was approved by the relevant institutional ethics committee.

The experimental task required participants to respond to a sequence of visual cues (pairs of colored boxes) by pressing a corresponding colored button on a keypad. The relationship between the cues and the correct response was governed by a set of rules, one of which was designed to be learned implicitly. Immediately following the experiment, a post-session interview was conducted to assess each participant’s explicit awareness of the underlying rules. This verbal report was used to classify participants into two distinct groups, which form the primary organizational basis for the dataset:

• Yes Implicit Learning (n = 9): Participants who could articulate all three task rules, including the hidden one.
• No Implicit Learning (n = 21): Participants who could articulate the first two explicit rules but showed no awareness of the third, hidden rule.

2.1. Dataset and Folder Structure

The dataset’s root folder, depicted in Figure 1, is organized to reflect the experimental design and data modalities. It contains folders for location files, documentation, and the primary participant data.

• Participant Data: The data are first divided into two folders corresponding to the classification described above: “Yes Implicit Learning” and “No Implicit Learning”. Within these folders, each participant is assigned a unique identifier (S00xx, where “xx” represents the participant’s assigned numerical ID code when they signed up for the experiment). Each participant’s folder is then further subdivided by modality (EEG and fNIRS).
• Location Files: This folder contains the “LocationCap32.ced” file for use with EEGLab and a diagram illustrating the experimental setup of EEG electrodes and fNIRS optodes.
• Documents and ReadMe: This folder provides documentation explaining data usage, a file named “Subject data.csv” listing each participant’s ID, sex, age, and their assigned implicit learning group, and a reference file “fNIRS to EEG Electrode channels reference.csv” for mapping fNIRS channels to the standard 10–20 EEG system.

2.2. EEG Data Files

The subfolder for each participant’s EEG data is organized as depicted in Figure 2. It contains the following three data types:

• Raw EEG Data (.edf): Unaltered data in European Data Format (EDF), converted directly from the Nihon Kohden Neurofax EEG-1100 machine with personal identifiers removed. This file includes all voltage measurements and event markers from the entire recording session and is ideal for researchers who wish to design custom preprocessing pipelines.
• Truncated EEG Data (.edf): Pre-processed data truncated to contain only the experimental task period, ready for import into MATLAB (R2020a) with the EEGLab toolbox (v2021.1).
• Event File (.txt): A two-column text file formatted for EEGLab. It contains the time markers for the beginning of each question and its corresponding answer, allowing for event-related analysis of the truncated data.

It should be noted that some participants’ data files are split into two parts (e.g., “part 1-1” and “part 1-2”). This occurred when participants requested a brief rest during the session, which was permitted as an ethical consideration for their comfort. Because this research focuses on identifying learning events that can occur at any time, all data segments—including those before and after a break—are considered essential and should be included in the analysis. These files can be safely concatenated chronologically to form a continuous record.

2.3. fNIRS Data Files

The organizational structure for the fNIRS data subfolder is shown in Figure 3. For each participant, the fNIRS folder contains data in a comma-separated value (.csv) format, structured as follows:

• fNIRS Recordings (.csv): For each participant, fNIRS data are provided in six separate files, corresponding to measurements from the left hemisphere (Probe1) and right hemisphere (Probe2). For each probe, there are three files detailing changes in oxygenated hemoglobin (HbO, denoted as “Oxy”), deoxygenated hemoglobin (HbR, denoted as “Deoxy”), and their difference (denoted as “Total”).
• File Content: Each CSV file contains 24 columns representing the fNIRS channels. A column labeled “Time” provides the time-stamp for each measurement, and a column labeled “Mark” contains event markers (one for question onset, two for answer). The channel mapping reference file in the “Documents” folder clarifies the correspondence of each column to the 10–20 system. Any column with a (-) symbol next to the probe name indicates that its polarity should be reversed during analysis. As with the EEG data, some participant recordings are split into two parts due to breaks and can be concatenated chronologically.

3. Methods

3.1. Participants

An a priori power analysis was conducted to determine the required sample size, following the guidelines of Lemeshow et al. [13]. The analysis was based on data from a study on working memory and serial reaction time tasks by Rose et al. [14], which showed mean response times of 640 ms and 580 ms for implicit and non-implicit learning conditions, respectively, with a standard deviation of 50 ms. With a confidence level of 95% (α = 0.05) and a statistical power of 80% (β = 0.2), the estimated minimum sample size was calculated to be 9 participants per group.

In total, 30 healthy volunteers participated in the study (23 males, 7 females; age range: 21–29 years). All participants reported having no prior history of learning disabilities and had normal color vision. The experiment was carried out in accordance with the Declaration of Helsinki. All participants volunteered and signed written informed consent forms. The study protocol was approved by the Experimental Ethics Committee of the Faculty of Information Science and Electrical Engineering, Kyushu University (ISEE H26-3, 23 June 2014).

3.2. Experimental Design and Procedure

The study employed a cognitive serial reaction time task (SRTT) specifically designed to induce and measure implicit learning. This paradigm was selected due to its established efficacy in studying the acquisition of knowledge without explicit awareness. The SRTT format requires participants to respond to a series of stimuli that appear random but are governed by underlying, unstated rules. As participants are repeatedly exposed to these patterns, they often exhibit improved performance (e.g., faster reaction times) without being able to articulate the rules, thus providing a robust behavioral marker of implicit learning. Participants were seated in front of a monitor and provided with a three-button keypad (marked red, green, and blue).

The task consisted of a sequence of up to 180 trials. Each trial began with the presentation of a “question,” which was a pair of colored boxes on the monitor. Participants were instructed to respond by pressing the single “correct” button on the keypad within a 3 s window. This response window, creating a 12 s block for each four-question trial, was deliberately chosen as a compromise to accommodate the different temporal characteristics of the two neuroimaging modalities. While EEG captures fast electrophysiological activity with millisecond precision, the hemodynamic response measured by fNIRS is inherently slower, evolving over several seconds. The 12 s trial duration was therefore designed to be long to capture a meaningful hemodynamic response for fNIRS analysis while still allowing for event-related analysis of the faster EEG signals and maintaining participant engagement.

The participants were informed that their goal was to respond correctly and to try to discover the underlying rules that determined the correct answer. Immediately after the 3 s response window, the correct color was displayed on the screen as feedback, and the next question appeared. Each trial was composed of four such questions, with a random delay of 0 to 1 s between trials. If a participant appeared to have deduced the rules before completing all 180 trials, the experiment continued for an additional 10 trials at the observer’s discretion before concluding. The experimental setup is illustrated in Figure 4.

The relationship between the colored boxes and the correct answer was governed by three rules:

• Rule 1 (Explicit): If the two boxes in the pair are different colors, the correct answer is the third color not shown (e.g., if red and green are shown, the answer is blue).
• Rule 2 (Explicit): If the two boxes are the same color, the correct answer is that same color (e.g., if two red boxes are shown, the answer is red). An incorrect response to these first two rules was met with an audible beep as negative feedback.
• Rule 3 (Implicit): The correct answer to the fourth question of any given trial is always identical to the correct answer for the first question of that same trial. No explicit cues or feedback were provided to signal this underlying pattern.

This design is intended for the third rule to be learned implicitly. The primary behavioral measure for detecting this learning was a reduction in response time for the fourth question relative to others in the trial.

3.3. EEG-fNIRS Data Acquisition

Simultaneous EEG and fNIRS data were recorded continuously throughout the experimental task. Electrophysiological signals were acquired using a Nihon Kohden Neurofax EEG-1100 system at a sampling frequency of 1500 Hz. Concurrently, hemodynamic activity was measured using a Hitachi ETG-7100 optical topography system with a sampling frequency of 10 Hz.

A single integrated cap was used, which held EEG electrodes alongside fNIRS transmitters and receivers. In accordance with the established principles of fNIRS signal localization, each EEG electrode was physically positioned over the midpoint of a corresponding transmitter–receiver pair. This arrangement aimed to align the recording of electrophysiological activity (EEG) with the source of the hemodynamic response (fNIRS) from the same cortical region. The specific layout of the measurement channels is shown in Figure 5.

3.4. Behavioral Assessment and Participant Classification

To obtain objective behavioral evidence of learning, we first analyzed participant response times (RTs). The primary behavioral marker for implicit knowledge acquisition was a reduction in RT for the fourth question of each trial, which was governed by the hidden rule. We compared the average RT for the fourth question against the average RT for the first three questions (which followed only explicit rules). A significantly faster RT on the fourth question would indicate that a participant had learned the underlying pattern, even if they were not consciously aware of it.

To definitively assess each participant’s level of explicit awareness, a structured post-experiment interview was conducted immediately after the data acquisition session. Participants were asked to describe and articulate any patterns or rules they had discovered. Based on their verbal reports, participants were categorized into one of two groups, which provided the labels for subsequent supervised machine learning analysis:

• No Implicit Learning (n = 21): Participants who successfully articulated the first two explicit rules but expressed no awareness of the third, hidden rule concerning the fourth question in each trial.
• Yes Implicit Learning (n = 9): Participants who were able to verbally describe all three rules, including the pattern associated with the fourth question.

For the participants ultimately categorized in the “Yes Implicit Learning” group, the point of learning acquisition was identified on an individual basis. The learning point was operationally defined as the first trial where the response time for the implicit-rule question (Question 4) was faster than the lower bound of the 95% confidence interval of the mean RT for the explicit-rule questions (Questions 1–3). The confidence interval was calculated from each participant’s performance in the previous trials. All trials for that participant from the learning point onward were classified as belonging to the post-learning phase.

The results of the response time analysis, which provided the quantitative basis for our grouping, are summarized in

Table 1. Comparison of response times (RT) before and after implicit learning acquisition.

Participant Group and Phase	Mean RT for Explicit Rules (Q1–3) (ms)	Mean RT for Implicit Rule (Q4) (ms)	Difference (ms) (Explicit-Implicit)	Wilcoxon Signed-Rank Test, p-Value
Yes Implicit Learning (n = 9)
Pre-learning Trials	1148.5 ± 176.2	1128.3 ± 131.4	20.2	0.4961
Post-learning Trials	1027.9 ± 99.8	666.7 ± 155.3	361.2	0.0039
No Implicit Learning (n = 21)
Full Experiment Average	973.7 ± 94.2	956.3 ± 76.5	17.4	0.1315

. The data reveals a clear learning effect for the participants ultimately categorized into the “Yes Implicit Learning” group. As shown in the table, this group demonstrated no significant difference in response time between the explicit and implicit trials during their pre-learning phase. However, in the post-learning phase, they exhibited a large and statistically significant reduction in RT for the implicit-rule question (Mean Q4 RT = 666.7 ms ± 155.3 ms) compared to the average of the first three questions (Mean Q1–3 RT = 1027.9 ms ± 99.8 ms, p < 0.005). In contrast, the “No Implicit Learning” group showed no significant RT difference between trial types when averaged across the entire experiment, confirming their status as a stable control group.

To further validate the quality of the neural recordings and demonstrate their utility, a preliminary analysis focusing solely on the EEG signals from this dataset has been previously published [15]. In that study, Multiscale Entropy (MSE) was computed from various EEG frequency bands to serve as features for machine learning classification. The results demonstrated that by using an Artificial Bee Colony (ABC) optimization algorithm for feature selection, a classifier could distinguish between the “Yes” and “No” implicit learning groups with 95% confidence. This finding confirms that the EEG data, when considered alone, contains robust, high-quality information sufficient to differentiate learning states, directly addressing the potential for single-modality analysis.

4. User Notes

While this dataset provides a valuable resource for investigating the neurophysiology of implicit learning, several limitations should be noted. First, the data were collected from a specific and homogeneous population: Asian participants aged 21 to 29. Consequently, the results may not be directly generalizable to individuals of other ethnicities or age ranges, where learning processes and their neural signatures could differ.

Second, all participants were university students, representing a cohort with a high and uniform level of education. This is a potential confound, as educational background can significantly influence cognitive strategies and brain function related to learning. To build upon this work, future research should prioritize the recruitment of participants from more diverse populations. Investigating these phenomena across a wider spectrum of ages, educational levels, and cultural contexts will be a critical next step in developing a universally applicable model of implicit learning. However, the collection of simultaneous EEG-fNIRS data is inherently resource-intensive. It requires specialized, co-compatible equipment, and the experimental setup for each participant is meticulous and time-consuming, involving the precise placement of both electrodes and optodes. The limited availability of such sophisticated instrumentation and the trained personnel required to operate it makes large-scale data collection challenging.