A Proposed Method of Automating Data Processing for Analysing Data Produced from Eye Tracking and Galvanic Skin Response

Abstract

The use of eye tracking technology, together with other physiological measurements such as psychogalvanic skin response (GSR) and electroencephalographic (EEG) recordings, provides researchers with information about users’ physiological behavioural responses during their learning process in different types of tasks. These devices produce a large volume of data. However, in order to analyse these records, researchers have to process and analyse them using complex statistical and/or machine learning techniques (supervised or unsupervised) that are usually not incorporated into the devices. The objectives of this study were (1) to propose a procedure for processing the extracted data; (2) to address the potential technical challenges and difficulties in processing logs in integrated multichannel technology; and (3) to offer solutions for automating data processing and analysis. A Notebook in Jupyter is proposed with the steps for importing and processing data, as well as for using supervised and unsupervised machine learning algorithms.

1. Introduction

The use of eye tracking technology, together with other physiological biomarker measurements such as psychogalvanic skin response (GSR) and electroencephalographic recordings (EEG), provides researchers with information about users’ physiological behavioural responses during their learning processes when performing various types of tasks. These records are obtained through sensors and provide non-subjective information that is put together with other data from the user (learning outcomes, perceived satisfaction with the learning process, perception of the use of metacognitive strategies, etc.). This technology has been applied in a range of fields, each with different objectives. For example, in marketing it has been used to understand client behaviour in order to achieve customer loyalty and increase sales [1,2,3,4,5,6,7,8,9,10,11]. In healthcare sciences, this technology can be applied for diagnosis (e.g., mental illness [12], autism spectrum disorder [13,14,15], examining medical imaging [16], mental fatigue [17], application in various clinical needs [18], Alzheimer’s [19], dysfunctional cerebral networks [20], various disorders [21], Rett’s Syndrome [22], and activity with prosthetic limbs [23]). Eye tracking technology has also been used in therapeutic intervention (for example, in Autism Spectrum Disorder [15], cerebral palsy [24], in patients with depression [25,26], in Parkinson’s [27], with various addictions [28], and for anorexia nervosa [29]). It has also been used in education in order to understand how learners process information when doing various tasks, with the ultimate aim being to improve students’ learning processes. Eye tracking technology can be used in educational settings to analyse different learning styles [30], for educational rehabilitation from problems related to dyslexia [31], and to examine the effectiveness of educational methodologies such as evidenced-based learning [32].

More specifically in educational settings, this technology is commonly used in higher education, especially with health sciences students [33,34,35,36,37,38,39]. One of the specific functions the technology offers is that it can give the teacher or researcher data that, once processed and analysed, will help in the design of explanatory models of how learning occurs in different types of learners [30].

Eye tracking technology is also being used in driving to understand the characteristics of different drivers in pursuit of better driving and fewer accidents [40,41]. It is also being used to test different effects on driving of things such as advertising [42].

Recent studies on human perception indicate that higher levels of attention are related to variables such as gender, age, and educational level. Studies on vigilance indicate that higher levels of self-attention [35] are related to shorter total fixation durations and fewer fixation counts while the user looks at picture stimuli [40]. Also, the use of animated images seems to improve learning outcomes [41]. Similarly, recent studies call for increased research into the effects of self-regulated learning strategy (SRL) training on student motivation, engagement and performance [43]. Specifically, using think-aloud protocols involving self-regulated strategies appears to improve learning outcomes [44]. Similarly, feedback on user performance appears to enhance deep learning in learners [45]. Along these lines, using materials that include self-regulated simulation activities has been shown to be very effective in increasing information acquisition and mitigating the effects of prior knowledge about the task on learning outcomes [46].

In other studies we have addressed the meanings of metrics from applying eye tracking technology integrated with GSR devices and their relationship to the use of cognitive and metacognitive strategies as well as learning profiles [37,47,48]. One reference that may help researchers to prepare metrics is Holmqvist et al. [49], a summary of which can be found in

Table 1. Some representative metrics and interpretation in integrated multichannel eye tracking.

Metric	Unit of Measurement	Meaning	Interpretation in the Context of the Human Learning
Fixation count	Count	Number of fixations on a stimulus or part of a stimulus.	A higher number of fixations may be related to a difficulty in processing that information because it is novel or difficult for the learner to process.
Fixation duration	Milliseconds	Duration of fixation.	Refers to the reaction times of the learner. A longer duration may be related to a higher cognitive load in the processing of the stimulus. It may also be related to the use of metacognitive orientation strategies, i.e., searching for information or relating it to previous knowledge.
Saccade count	Count	Refers to the shift of gaze from one part of the stimulus to another.	A higher number of saccades implies the use of metacognitive orientation strategies, i.e., searching for information or relating it to prior knowledge. Likewise, the greater the amplitude of the saccade, the lower the cognitive effort, although this may also be related to information processing problems. Younger learners apply shorter saccades.
Pupil diameter	Millimetres	The mean pupil diameter is collected for all fixations within an AOI during a time interval.	It refers to the interest that a stimulus or part of it has for the learner. A larger pupil diameter may be related to increased cognitive load and/or difficulty in processing a task.
Number of visits		Number of bindings within a stimulus or part of a stimulus (of an area of interest or AOI).	Refers to attention to a stimulus or part of a stimulus.
Scan Path Length or Gaze Point	X and Y position coordinates	Refers to the chain of fixations in order of succession.	It involves a pattern of visual tracking on a stimulus. It gives information about how each learner processes information.
SCR count	Count	The number of skin conductance responses (SCRs), for each Interval in Time of Interest.	It provides information about the emotional state of a learner. The SCR count can be used to identify which specific moments of a dynamic stimulus and specific information within a stimulus elicit an emotional response in a learner. A higher SCR count indicates a higher level of emotional arousal.
ER SCR Amplitude	Microsiemens	The amplitude of each event related skin conductance response (ER-SCR), for each Interval in Time of Interest. Time of Interest intervals that do not have an ER-SCR are calculated using filtered GSR data.	When an SCR occurs between 1 and 5 s after an event (ER-SCR), it is considered whether the stimulus elicits an emotional reaction in the learner. This measure can provide information about different emotional states such as: anxiety, stress, frustration, or relaxation.
GSR average	Microsiemens	The mean of the average galvanic skin response (GSR) signal, after filtering for each time of interest.	When environmental conditions are held constant, slow fluctuations in the GSR signal (between seconds and minutes) reflect changes in the participant’s emotional arousal. The researcher can use the average GSR metric in different sections of the session to determine whether a learner might be stressed, frustrated, or relaxed during the course of an experiment.

.

How to Extract Records from Eye Tracking Devices and How to Process and Analyse Records

The devices that apply eye tracking technology include records of static metrics (fixations, saccades, pupil diameter, blinking, etc.). They also apply various measuring parameters to those metrics such as velocity, duration, etc. These metrics are found by applying statistical resources such as frequencies, means, standard deviations, etc. (examples of these measures according to parameter may be found in Table 1).

Eye tracking devices also include measurement of dynamic metrics (these involve the points where the participant is looking on the screen’s cartesian co-ordinates), called gaze point or scan path. An example is shown in Figure 1.

Nonetheless, more thorough analysis of dynamic metrics requires the application of machine learning techniques. These may be supervised techniques, which include techniques such as regression (about predicting numerical values), including linear regression, Support Vector Machine (SVM), Decision Tree, and Neural Networks. The other supervised techniques are classification (about predicting categories). These include SVM, discriminant analysis, nearest neighbour, and naïve Bayes. Machine learning techniques may also be unsupervised (which are applied without using labels in order to discover hidden patterns). Unsupervised techniques include clustering algorithms (which help to group data based on similarity, such as: k-means, k-means++, etc.), and are applied to reduce dimensionality (such as in Principal Component Analysis—PCA) [50,51].

That said, it is important to remember that the data that can be extracted from eye tracking devices are unprocessed. This means that, in order to be able to apply statistical techniques or machine learning algorithms (supervised or unsupervised), they must first be processed. According to García et al. [52], pp. 10–13, the steps in the procedure are:

• Produce the database.
• Define the problem.
• Understand the problem.
• Process the data, with the following sub-steps:

  4.1.

  Prepare the data: clean the data, which includes operations to correct erroneous data, filter incorrect data from the dataset, and reduce unnecessary data. Other data cleaning tasks involve detecting discrepancies and dirty data (fragments of original data that make no sense)—these tasks are more closely related to understanding the original data and usually need human involvement in the process—and transformation of the data (in this step, the data are converted or consolidated so that the result of the data-mining process can be applied or be more effective). Sub-tasks within transformation include smoothing, construction of characteristics, and data aggregation or data summary; these require human supervision.

  4.2.

  Integrate the data (covers the fusion of data from various data stores). This process is applied to avoid redundancies and inconsistencies in the resulting dataset. Typical operations in data integration are identification and unification of variables and domains, analysis of the correlation of attributes, duplication of tuples, and detection of conflicts in data values from different sources.

  4.3.

  Normalise the data (this refers to the unity of the measure used, as this may affect the data analysis). This means that all of the attributes must be expressed in the same units of measurement and must use a common scale or range. This step is particularly useful in statistical learning methods.

  4.4.

  Impute missing data (this is a form of cleaning data, the aim of which is to fill in missing values using a reasonable estimation via different methods, such as applying a mean value).

  4.5.

  Identify noise (this refers to the act of smoothing in data transformation, the main objective of which is to detect random errors or variance in a measured variable).

• Apply machine learning techniques.
• Evaluation.
• Exploit the results.

In summary, using integrated multichannel eye tracking technology can be of great help to professionals in various fields (marketing, health, education, driving, etc.). However, the processed information these devices provide is very basic. Therefore, we believe that ordinary users (researchers, teachers, professionals, etc.) should be able to get information about the results simply and easily [53]. Current research indicates that machine learning techniques can help make data analysis more effective, as they help people to interpret the data [54,55,56,57]. Many concepts in machine learning are related to ideas that were previously addressed in psychology and neuroscience [58].

Therefore, our research field could benefit from proposals for automating the preprocessing and processing [52] of data recorded by devices applying integrated eye tracking techniques (i.e., data from eye tracking and other data such from integrated GSR devices or electroencephalographic response (EEG)). There is a gap between the output of such software and the data analysis that researchers would perform. The present paper seeks to bridge that gap, aiming to explore this issue in greater depth in order to help researchers in this field.

In view of the above, the objective of the current study was to suggest a procedure for processing data extracted from these types of devices by applying integrated multichannel, eye tracking technology.

2. Materials and Methods

2.1. Participants

A convenience sample was used to illustrate the proposed procedure to readers by example. More specifically, the study used a sample of 17 students (nine women and eight men) in the fourth year of a biomedical engineering degree at the University of Burgos. They were between 21 and 22 years old, with a standard deviation for both genders between 0.7 and 0.9 years. Participation was voluntary and without any financial reward. The students were told about the study objectives and signed their informed consent before taking part. In addition, the study was approved beforehand by the University of Burgos Bioethics Committee (report no. IO 03/2022).

2.2. Instruments

(a)

Tobii Pro Lab v. 1.241.54542 [59]. This is a hardware and software platform that is used by researchers to analyse human behaviour. It has a high degree of flexibility and can be used to perform advanced studies on attention and provide deep analysis of cognitive processes. This software was chosen due to it currently being one of the most widely used in research [60], as it makes it easy to integrate other measuring instruments, such as GSR.

(b)

Shimmer 3 GSR+ [61]. The Shimmer3 GSR+ (Galvanic Skin Response) unit provides connections and pre-amplification for a Galvanic Skin Response (Electrodermal Resistance Measurement—EDR/Electrodermal Activity (EDA) data acquisition channel). The GSR+ unit is suitable for measuring the electrical characteristics or conductance of the skin. The device is compatible for use in conjunction with Tobii Pro Lab v. 1.241.54542 and the metrics are integrated with those obtained with the Tobii Pro Lab v. 1.241.54542 device.

(c)

Virtual Laboratory 8—Eye tracking technology applied to Early Intervention II. A virtual classroom (also called a lab) in the eEarlyCare-T Research Project. This classroom is free and open access after logging into the platform https://www2.ubu.es/eearlycare_t/en/project (accessed on 8 August 2024) [62]. Images from the virtual classroom are provided in Figure A1.

(d)

Python libraries: Pandas, sklearn, matplotlib, seaborn.

2.3. Procedure

The objective of this study was to suggest a procedure for processing data extracted from an eye tracking device that incorporated GSR recording. More specifically, the study used a Tobii Pro Lab v. 1.241.54542 [59] with Shimmer 3 GSR+ [61]. It addressed the technical difficulties and record processing challenges that users may have to deal with when using this integrated multichannel technology. Lastly, the study also aimed to propose potential solutions for automating data analysis via the design of a Jupyter Notebook in which machine learning algorithms were applied.

Researchers do not normally work with very large samples of participants because the experimental process for these types of studies is laborious and needs to be tailored. This means spending a lot of time on each user in each experimental phase. Despite the small numbers of participants, each one produces a large amount of diverse records, meaning a huge volume of data. Integrated multichannel eye tracking devices facilitate the processing and analysis of data through frequency analysis—for example, through data extracted using the heat map model, which visualises the most common gaze points recorded for the object of study, be that an image, text, video, or, web page. Figure 2 shows an example.

In addition, integrated multi-channel eye tracking technology provides information about the gaze positioning chain in hierarchical order of appearance. This is a dynamic measure called scan path or gaze point, an example of which is shown in Figure 3.

Furthermore, a variety of metrics can be extracted for each participant once Areas of Interest (AOIs) have been established. AOIs can be divided into relevant AOIs and non-relevant AOIs. Relevant AOIs refer to the essential information in the learning process. The users should pay more attention to them. Non-relevant AOIs refer to information that the user should not pay attention to.

Table 2. Most representative metrics, meaning, and unit of measurement.

Metric	Meaning	Unit of Measurement
Fixation point	The normalized horizontal and vertical coordinate of the fixation point.	Normalized coordinates (DACS)
Average pupil diameter	The average diameter of the pupil of the fixation. Calculated using the resulting pupil diameter after applying pupil diameter.	Millimeters
Saccade direction	The angle of the straight line between the preceding fixation and succeeding fixation. This can only to applied to whole saccades.	Degrees
Average velocity	The average velocity across all samples belonging to the saccade, even outside the interval.	Degrees/second
Peak velocity	The maximum velocity across all samples belonging to the saccade, even outside the interval.	Degrees/second
Saccade amplitude	The amplitude for whole saccades.	Degrees
Mouse position	The position of the mouse.	Pixels (DACS)
Galvanic skin response (GSR)	The raw galvanic skin response signal of the participant.	Microsiemens
Average GSR	The average galvanic skin response (GSR) signal, after filtering, for an interval.	Microsiemens
Number of SCR	The number of skin conductance responses (SCRs) for an interval.	Count
Amplitude of event related SCR	The amplitude of each event-related skin conductance response (ER-SCR), for an interval. ER-SCRs are calculated using filtered GSR data.	Microsiemens

gives an example of some of these metrics, the meaning of the record, and the unit of measurement. These metrics refer specifically to the Tobii Pro Lab technology v. 1.241.54542 [59]. In addition, researchers can extract metrics for saccades and fixations about parameters of duration, start, end, mapping, direction, etc. These records can be produces in various formats, including .xlsx, .tsv, .plof, etc.

However, in order to be able to analyse the recorded data, researchers must process and analyse most of these records. Furthermore, once processed, statistical or machine learning techniques (supervised or unsupervised) must be applied in order to analyse and subsequently interpret the data [52]. Figure 4 shows an outline of the process that should be followed. It applies a more complex log analysis procedure than the default analysis offered by the devices.

2.4. Data Analysis

The objective of this study was not to test any research hypotheses, but rather to suggest how to create a Notebook in Jupyter as a way of providing an example of how one might process data recorded via integrated multichannel eye tracking technology (in this case incorporating a GSR device). More specifically, we applied supervised machine learning algorithms (regression—linear regression in this case) and unsupervised machine learning algorithms (clustering: using k-means) using the Python libraries: pandas, sklearn, matplotlib, and seaborn.

3. Results

A Notebook was made in Jupyter with the steps for importing and processing the database extracted from Tobii Pro Lab v. 1.241.54542 [59], after a pre-processing phase. In this phase, all detected noise was cleaned. The steps to be followed are described below.

• Step 1. Clean and filter the files (one file per participant was exported), in this case, they were extracted in .xlsx format.
• Step 2. Import libraries to read the extracted files.

Import

import pandas as pd

• Step 3. Create a function to concatenate the files

Input:

• data_folder: path to the folder containing data files
• extension: file extension (e.g., .xlsx)

Output:

• concatenated_data: A DataFrame formed by concatenating all the files.

Process:

• Initialize files as the list of files in data_folder.
• Initialize an empty list list_dataframes.
• For each file in files:

  ○

  Set full_path as the complete path of the file.

  ○

  If full_path is a valid file and its extension matches extension:

  ▪

  If the extension is .xlsx:

  ▪

  Display message: “Reading file i + 1/len(files): file”.

  ▪

  Read the file into a DataFrame df.

  ▪

  Append df to list_dataframes.

  ▪

  Else: Continue to the next file.

  ○

  If an error occurs while reading the file:

  ▪

  Display message: “Error reading file file: error_message”.

• If list_dataframes is not empty:

  ○

  Concatenate all DataFrames in list_dataframes into concatenated_data.

  ○

  Display message: “All files have been processed!”.

  ○

  Return concatenated_data.

• Else:

  ○

  Display message: “No files were found for processing”.

  ○

  Return None.

• Step 4. Data Preprocessing.

Input:

• data: DataFrame containing the data
• delete_rows: dictionary where keys are column names and values are rows to be deleted
• filter_rows: dictionary where keys are column names and values are rows of interest
• select_columns: list containing the columns to be kept

Output:

• processed_data: A DataFrame with the processed data.

Process:

• For each column, values in delete_rows:

  ○

  Remove rows from data where the column contains any of the values.

• For each column, values in filter_rows:

  ○

  Filter data to keep only rows where the column contains any of the values.

• Select columns of interest in data using select_columns.
• Return processed_data, which is now the processed DataFrame.

• Step 5. Data Processing with Filters and Selection.

Process:

• Create a dictionary delete_rows:

  ○

  Keys: ‘Participant name’

  ○

  Values: [‘Test XX 2’, ‘Participant_X3’, ‘NP_01’, ‘NP_01_2’, ‘NP_02’]

  ○

  Purpose: To remove rows where the ‘Participant name’ matches any value in the list.

• Create a dictionary filter_rows:

  ○

  Keys: ‘Sensor’

  ○

  Values: [‘Mouse’, ‘GSR’]

  ○

  Purpose: To filter rows where the ‘Sensor’ column contains only ‘Mouse’ and ‘GSR’.

• Create a list select_columns:

  ○

  [‘Participant name’, ‘Gender’, ‘Audio’, ‘Recording duration’, ‘Eye movement type’, ‘Galvanic skin response (GSR)’]

  ○

  Purpose: To select the columns of interest from the DataFrame.

• Applythe processing:

  ○

  Call processing_Data function with the parameters: data, delete_rows, filter_rows, and select_columns.

• Return the processed data as processed_data.

• Step 6. Data grouping and aggregation

Input:

• data: DataFrame containing participant data

Output:

• grouped: DataFrame with data grouped by ‘Participant name’ and aggregated metrics

Process:

• Group the data by ‘Participant name’, applying the following aggregation functions:

  ○

  Gender: Select the first value (assuming gender is consistent across all rows).

  ○

  Audio: Select the first value (assuming the audio characteristic is consistent across all rows).

  ○

  Recording duration: Select the first value (assuming test duration is the same across all rows).

  ○

  Eye movement type:

  ▪

  If both ‘Fixation’ and ‘Saccade’ exist in the column:

  ▪

  Compute the ratio of ‘Fixation’ events to ‘Saccade’ events.

  ▪

  Else: Return None.

  ○

  Galvanic skin response (GSR): Compute the mean value for GSR.

• Reset the index of the grouped DataFrame.
• Rename the ‘Eye movement type’ column to ‘Fixation to Saccade Ratio’ for a more descriptive name.
• Return the grouped DataFrame.

• Step 7. Load data

Input:

• location: File path of the Excel file

Output:

• data: Loaded DataFrame from the Excel file

Process:

• Load the Excel file from location into a DataFrame data
• Return the data

• Step 8. Combine data

Input:

• data1: First DataFrame
• data2: Second DataFrame
• column: Column name on which to merge both DataFrames

Output:

• data: Merged DataFrame based on the common column

Process:

• Merge data1 and data2 on the column
• Return the merged data

An example is presented in Figure 5.

An example of the final integration of the data is presented in Figure 6.

• Step 9. Clustering preparation

Input:

• data: DataFrame containing the original data

Output:

• data_normalized: DataFrame with encoded, imputed, and normalized data ready for clustering

ProcAess:

• Detect Boolean variables:

  ○

  Identify Boolean columns.

  ○

  Store these column names in booleans.

• Detect non-Boolean categorical variables:

  ○

  Identify categorical columns.

  ○

  Store these column names in categorical.

• Encode categorical variables:

  ○

  Convert categorical variables into Boolean (binary) variables.

  ○

  Store the result in data_encoded.

• Convert Boolean columns to binary values:

  ○

  For each column in booleans:

  ▪

  Convert the column values to binary (0 or 1).

• Handle missing values:

  ○

  Apply a mean imputation strategy to fill missing values:

  ▪

  Impute missing values using mean in data_encoded.

  ▪

  Store the imputed data in data_filled.

• Normalize the data:

  ○

  Apply StandardScaler to normalize the data:

  ▪

  Scale the imputed data so that all features have the same scale.

  ▪

  Store the normalized data in data_normalized.

• Return the data_normalized.

• Final Step: Apply Clustering Preparation to Combined Data

• Call clustering_preparation(data_combined) to normalize and prepare the data for clustering.
• Return data_normalized.

• Algorithm: K-Means Clustering (Elbow Method)

Input:

• data: DataFrame containing the preprocessed data for clustering
• k_range: List of values representing the range of potential cluster numbers

Output:

• inertia: List of inertia values for different cluster sizes

Process:

• Initialize an empty list inertia to store the inertia values for each K.
• For each value of k in k_range:

  ○

  Create a KMeans object with the following parameters:

  ▪

  n_clusters = k: Specifies the number of clusters.

  ▪

  random_state = 42: Ensures reproducibility.

  ▪

  n_init = 10: Number of times the KMeans algorithm will be run with different initializations.

  ○

  Fit the KMeans model to data using the fit method.

  ○

  Append the inertia value of the current KMeans model to the inertia list.

• Algorithm: K-Means Clustering Visualization

Input:

• data: Original DataFrame with all data (including the cluster assignment)
• normalised_data: Data that has been preprocessed and normalized for clustering
• delete_columns: List of columns to be removed (if necessary) before visualization
• clusters: The number of clusters from the KMeans model

Output:

• A plot showing relationships between variables involved in clustering.

Process:

• Initialize a figure for the plot with size (12, 12).
• Create a pairplot using Seaborn (sns.pairplot):

  ○

  Set data as the source DataFrame.

  ○

  Specify the variables to plot using numerical_cols (the numerical columns involved in the clustering).

  ○

  Set hue to ‘Cluster’ to color the points by their assigned cluster.

  ○

  Use the palette=‘tab10’ to specify a color palette for the clusters.

  ○

  Set plot_kws with additional plot options:

  ▪

  alpha = 0.6: Transparency of the points.

  ▪

  s = 100: Size of the scatter plot points.

  ▪

  edgecolor = ‘w’: White edge around the points.

• Add a title to the plot:

  ○

  Title format: “Relationships between variables involved in clustering with KMeans({clusters} clusters)”.

  ○

  Adjust the title position with y=1.02.

• Save the plot as a PDF file:

  ○

  File name: Variables_KMeans_{clusters}Cluster.pdf.

• Show the plot using plt.show().

In this case the graph points to the application of k = 2 (see Figure 7).

An example of the graph obtained is shown in Figure 8.

• Step 10. Linear Regression and Feature Impact Visualization

Input:

• data: DataFrame containing the feature variables and target variable
• target_column: The column name of the target variable to be predicted

Output:

• df_coefficients: DataFrame containing the coefficients and p-values of each feature

Process:

• Split the DataFrame into features (X) and target variable (y):

  ○

  Set y as the target variable column.

  ○

  Set X as the remaining features.

• Scale the features using StandardScaler:

  ○

  Apply StandardScaler to scale the feature values to the same scale.

  ○

  Store the scaled data in X_scaled.

• Initialize the Linear Regression model:

  ○

  Create a LinearRegression object.

• Create empty lists to store coefficients and p-values:

  ○

  coefficients: Stores the regression coefficients (weights).

  ○

  p_values: Stores the p-values for each coefficient, indicating their statistical significance.

• Fit the linear regression model for each feature:

  ○

  For each feature column in X:

  ▪

  Extract the feature column from X_scaled: X_temporal = X_scaled[:, [i]].

  ▪

  Fit the model.

  ▪

  Append the coefficient (model.coef_[0]) to coefficients.

  ▪

  Add a constant column to X_temporal for calculating the intercept: X_temporal sm.add_constant(X_temporal).

  ▪

  Create and fit the OLS model.

  ▪

  Append the p-value (model_sm.pvalues [1]) to p_values.

• Create a DataFrame with the coefficients and p-values:

  ○

  Store the results in df_coefficients.

• Sort the DataFrame by the absolute value of the coefficient:

  ○

  Sort df_coefficients by coefficient magnitude.

• Visualize the regression coefficients:

  ○

  Create a figure for the bar plot.

  ○

  Use Seaborn to create a bar plot.

  ○

  Set plot titles and axis labels.

  ○

  Save the plot as a PDF.

  ○

  Show the plot.

• Return the DataFrame with the sorted coefficients and p-values (df_coefficients).

Specifically, the regression coefficients and their significance at α = 0.05 were found. As

Table 3. Lineal Regression for each of the variables.

Variables (Characteristics)	Standardized Beta Coefficients	Standard Errors of Coefficients	p-Value
Type of presentation	0.48	0.23	0.07
Recording duration	−0.45	0.23	0.07
Gender	−0.15	0.25	0.57
GSR	−0.16	0.24	0.73
Fixation saccade ratio	−0.19	0.24	0.52

shows, none of them were significant. There was only a pattern of difference in the variables “Type of presentation” α = 0.07 and “Recording duration”.

4. Discussion

The aim of this study was to highlight the need for integrated multichannel eye tracking devices to also offer—in addition to analysis of metrics based on using simple statistical tests (e.g., duration of an interval, number of events, time of each event, total duration of a fixation on an AOI, mean duration of fixations on an AOI, the number of fixations on an AOI, the duration of visits to a stimulus, the number of visits, the number of clicks on an AOI, gaze direction, eye position, mean duration of fixations, minimum duration of fixations, maximum duration of fixations, pupil diameter, etc.)—the possibility of processing records by applying other analytical techniques, such as more complex statistical or machine learning techniques (both supervised and unsupervised). Various studies have made effective use of supervised machine learning techniques (classification and regression) and unsupervised techniques (clustering) to interpret eye tracking data [50,51,54,55,56,57]. The objective is for the processing of all of this data to be done as smoothly as possible, automating the phases that researchers currently have to do by hand [52]. This paper provides an example of how data may be extracted and how machine learning algorithms may be automated, specifically in this case with the Python programming language—although the procedure may be used with other languages, such as R and Matlab.

5. Conclusions

Currently, devices applying eye tracking technology give researchers and users simple information for data analysis (frequencies, means, simple analysis of gaze paths for each participant). Extracting this data is relatively painless. However, the devices record much more information than that. To make the best analysis of all of this data, these records need to be pre-processed and processed. This requires an understanding of computation or IT. This is a major limitation on making use of the potential offered by multichannel integrated eye tracking technology. This is a pressing issue, as we believe that this technology should be habitually applied in advertising, health, education, and driving (among other areas).

5.1. Study Limitations

The limitations of this study are related to the small sample of user data. Because of that, future studies will increase the size of the samples in order to produce more versatile Jupyter Notebooks that can be used in different research fields. In addition, this study only applied two machine learning algorithms (supervised: linear regression; unsupervised: k-means clustering). Because of that, other studies will expand the use of supervised and unsupervised algorithms.

5.2. Future Lines of Research

The use of integrated multichannel eye tracking technology can offer professionals in marketing, education, and health (among others) real-time analysis of the processing of all of the information the devices record. This functionality will be extremely useful in various fields of application (for example, in education it will be useful for teachers to have prediction, classification, or clustering data to help them create tailored educational resources. Similarly, doctors and psychologists may find help in individualised differential diagnosis and treatment). In other words, it opens up a wide range of possibilities for practical application. However, the obstacle of data processing means that this technology is not being used to its fullest potential. In this article, we have presented an example of how the phases of data pre-processing and processing may be automated, including the use of supervised and unsupervised machine learning algorithms. We would encourage researchers to continue working along these lines, and those responsible for the technology to incorporate these analytical resources into their multichannel eye tracking devices. Current researchers who avail themselves of machine learning techniques are more effective at analysing eye tracking data. All of this will contribute to this technology being used more habitually, and to improving the conclusions studies can produce, helping to improve research in the fields in which it is applied.