A Comparative Analysis of Deep Learning and Traditional Machine Learning for Classifying Cognitive Workload from Raw EEG Signals

A Comparative Analysis of Deep Learning and Traditional Machine Learning for Classifying Cognitive Workload from Raw EEG Signals | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article A Comparative Analysis of Deep Learning and Traditional Machine Learning for Classifying Cognitive Workload from Raw EEG Signals Senushi Dinara This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7767198/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract The objective assessment of cognitive workload is critical for enhancing performance and safety in high-stakes environments such as aviation and process control. This study presents a comparative analysis of two machine learning paradigms for classifying cognitive workload into three distinct levels (Low, Moderate, High) using electroencephalography (EEG). We developed and evaluated a deep learning model based on a 1D Convolutional Neural Network (CNN) that processes raw time-series EEG data, and compared it against a traditional machine learning baseline, a Random Forest (RF) classifier, trained on hand-engineered statistical features. The CNN model achieved a superior test accuracy of 94.2%, significantly outperforming the Random Forest model, which achieved an accuracy of 62.0%. This 32.2% performance gap strongly indicates that the raw temporal structure of EEG signals contains discriminative features for workload classification that are not captured by standard statistical summaries. The results validate the efficacy of deep learning for automated feature extraction in neurophysiological data and provide a robust, deployable model for real-time cognitive workload monitoring systems. Biological sciences/Neuroscience Biological sciences/Neuroscience/Computational neuroscience Cognitive Workload EEG Deep Learning 1D Convolutional Neural Network Random Forest Classification BCI 1. Introduction Cognitive workload, the mental effort required to perform a task, is a crucial factor in human performance, particularly in complex, dynamic environments like aircraft cockpits, surgical suites, and industrial control rooms [ 1 , 2 ]. Overload can lead to errors, while underload may result in vigilance decrements. Therefore, an objective, real-time measure of cognitive workload is highly desirable for adaptive systems and operator monitoring. Electroencephalography (EEG) provides a non-invasive, high-temporal-resolution window into brain activity and has been widely established as a reliable indicator of cognitive workload [ 3 ]. Traditionally, machine learning approaches for EEG classification have relied on extracting hand-crafted features from the signal (e.g., band power, statistical moments) which are then fed into classifiers like Support Vector Machines or Random Forests [ 4 ]. While effective for some applications, this approach is limited by the expertise required for feature engineering and the potential to discard salient information embedded in the raw signal's temporal dynamics. Deep learning, particularly Convolutional Neural Networks (CNNs), offers a powerful alternative by learning hierarchical features directly from raw data [ 5 ]. In EEG analysis, 1D CNNs can scan across time and channels to automatically discover complex, task-relevant patterns without manual intervention. This research investigates the hypothesis that a 1D CNN model, by leveraging the full temporal structure of raw EEG, will significantly outperform a traditional Random Forest classifier based on engineered statistical features for a three-level cognitive workload classification task. We present a rigorous comparison of both approaches on the same dataset, demonstrating not only the superior performance of the deep learning model but also providing insights into the physiological manifestations of different workload states. 2. Methodology 2.1. Data Acquisition and Preprocessing The study utilized a multi-participant EEG dataset (e.g., from a BCI Hackathon or MATB-II variant), comprising a total of 13,410 one-second epochs across 15 participants. Each epoch consisted of 61 EEG channels sampled at 500 Hz, resulting in a data matrix of 61 channels × 500 time steps per epoch. The epochs were labeled according to three levels of induced cognitive workload: Class 0 (Low), Class 1 (Moderate), and Class 2 (High). The dataset was partitioned in a stratified manner into an 80% training set (10,728 epochs) and a 20% held-out test set (2,682 epochs) to ensure proportional class representation. A StandardScaler was applied to normalize the data channel-wise to a mean of zero and a standard deviation of one. To manage computational memory constraints, a Keras Data Generator was implemented to apply scaling and reshaping on the fly during model training, loading only one batch of data at a time. 2.2. Deep Learning Model: 1D Convolutional Neural Network (CNN) The CNN was designed to process the raw EEG data with an input shape of (500, 61), representing 500 time steps and 61 channels. · Architecture: The core feature extraction module consisted of three sequential blocks, each containing a 1D Convolutional layer (with 32, 64, and 128 filters, respectively), a ReLU activation function, and a 1D Max-Pooling layer. This design progressively extracts localized temporal features while reducing dimensionality. The output was flattened and passed to a fully connected Dense layer with 128 units and a 50% Dropout layer for regularization. The final output layer used a softmax activation function to produce a probability distribution over the three classes. · Training: The model was compiled with the Adam optimizer and categorical cross-entropy loss function, and trained for 20 epochs. Table 1: 1D CNN Architecture Layer Type Configuration Output Shape Rationale Input - (500, 61) Raw EEG epoch. Conv1D filters=32, kernel_size=3, activation='ReLU' (498, 32) Extract local temporal patterns. MaxPooling1D pool_size=2 (249, 32) Dimensionality reduction. Conv1D filters=64, kernel_size=3, activation='ReLU' (247, 64) Learn more complex features. MaxPooling1D pool_size=2 (123, 64) Further downsampling. Conv1D filters=128, kernel_size=3, activation='ReLU' (121, 128) High-level feature extraction. MaxPooling1D pool_size=2 (60, 128) Final pooling. Flatten - (7680) Prepare for dense layers. Dense units=128, activation='ReLU' (128) High-level aggregation. Dropout rate=0.5 (128) Prevent overfitting. Dense (Output) units=3, activation='softmax' (3) Class probability distribution. 2.3. Traditional Machine Learning Model: Random Forest (RF) To establish a performance baseline, a Random Forest classifier was implemented using a feature-engineered approach. · Feature Engineering: Seven statistical features were calculated for each of the 61 EEG channels over the 500-sample epoch: Mean, Median, Standard Deviation, Minimum, Maximum, Interquartile Range (IQR), and Skewness. This resulted in a feature vector of 427 dimensions per epoch (61 channels × 7 features). · Model and Training: A Random Forest classifier with 100 decision trees and a maximum depth of 15 was trained on the resulting feature matrix. The feature extraction and model fitting were performed in a memory-efficient, participant-by-participant loop to prevent RAM exhaustion. 3. Results 3.1. Deep Learning Model Performance The 1D CNN model demonstrated exceptional performance on the unseen test set, achieving a final test accuracy of 94.2% with a corresponding loss of 0.16. The close tracking of training and validation accuracy curves indicated successful generalization with minimal overfitting. 3.2. Traditional Machine Learning Model Performance The Random Forest classifier achieved a significantly lower overall test accuracy of 62.0%. A detailed classification report and confusion matrix provide further insight into its performance. Table 2: Random Forest Classification Report Class Precision Recall F1-Score Support 0 (Low) 0.68 0.78 0.72 894 1 (Moderate) 0.56 0.43 0.49 894 2 (High) 0.60 0.65 0.63 894 Accuracy 0.62 Macro Avg 0.61 0.62 0.61 2682 Table 3: Random Forest Confusion Matrix True \ Predicted Class 0 Class 1 Class 2 Class 0 (Low) 695 115 84 Class 1 (Moderate) 205 383 306 Class 2 (High) 126 184 584 3.3. Comparative Performance Summary The central finding of this study is the stark performance difference between the two modeling paradigms, as summarized below. Table 4: Model Comparison Model Type Input Data Final Test Accuracy Deep Learning (1D CNN) Raw Time-Series (500, 61) 94.2% Traditional ML (Random Forest) 427 Engineered Features 62.0% 4. Discussion The primary objective of this research was to determine the most effective machine learning approach for classifying cognitive workload from EEG. The results provide a clear and definitive answer. 4.1. Superiority of Deep Learning for Temporal Feature Extraction The most significant finding is the 32.2% absolute difference in accuracy between the CNN and Random Forest models. This substantial gap underscores a critical point: the raw, millisecond-level temporal dynamics of the EEG signal contain a wealth of discriminative information that is lost when the signal is reduced to summary statistics like mean or standard deviation. The 1D CNN's convolutional filters successfully learned to identify complex, localized patterns across time and channels that are highly predictive of workload state, a capability beyond the scope of the manually engineered features. This result strongly validates the use of end-to-end deep learning models for complex EEG classification tasks, as they circumvent the limitations and biases of manual feature engineering. 4.2. The Challenge of Classifying Moderate Workload The analysis of the Random Forest's performance offers a secondary, important insight. The model struggled most severely with Class 1 (Moderate Workload), as evidenced by its lowest recall (0.43) and F1-score (0.49). The confusion matrix (Table 3) reveals that 205 Moderate epochs were misclassified as Low, and 306 were misclassified as High. This suggests that the "Moderate" cognitive state is a physiologically ambiguous transition zone, sharing characteristics with both the relaxed (Low) and overloaded (High) states. The CNN's ability to learn nuanced, non-linear boundaries in the high-dimensional raw data was essential for accurately dissecting this ambiguous class. 4.3. Limitations and Future Work A limitation of this study is the use of a specific, task-induced workload dataset. Future work should validate the generalizability of the trained CNN model on data from different experimental paradigms and participant populations. Furthermore, techniques like Grad-CAM could be applied to the trained CNN to identify which time periods and channels most strongly influence the classification decision, enhancing the interpretability of the model. The high-performing CNN model has been saved and is readily deployable. The immediate next step is to integrate this model into an edge computing system for low-latency, real-time cognitive workload assessment in operational settings, such as providing feedback during pilot training or triggering alerts for system operators. 5. Conclusion This study successfully demonstrated that a 1D Convolutional Neural Network can classify cognitive workload from raw EEG data with high fidelity (94.2% accuracy), dramatically outperforming a traditional Random Forest classifier based on statistical features (62.0% accuracy). The findings confirm that deep learning is superior for automatically extracting the complex temporal patterns in EEG that are indicative of mental state, and that these patterns are more informative than hand-crafted features. The developed model provides a robust, accurate, and readily deployable backend for real-time cognitive workload monitoring systems, with significant potential for improving safety and performance in high-demand professions. References Wickens CD (2002) Multiple resources and performance prediction. Theoretical Issues Ergon Sci 3(2):159–177 Gevins A, Smith ME (2003) Neurophysiological measures of cognitive workload during human-computer interaction. Theoretical Issues Ergon Sci 4(1–2):113–131 Antonenko P, Paas F, Grabner R, van Gog T (2010) Using electroencephalography to measure cognitive load. Educational Psychol Rev 22(4):425–438 Lotte F et al (2018) A review of classification algorithms for EEG-based brain–computer interfaces: a 10-year update. J Neural Eng 15(3):031005 Lawhern VJ EEGNet: a compact convolutional neural network for EEG-based brain–computer interfaces. Journal of Neural Engineering, 15(5), 056013. Data sets 1.Center for Data Science and Neuroergonomics. (2021). Passive BCI Hackathon – Neuroergonomics 2021 (Version 1.0) [Data set]. Zenodo., Santiago-Espada Y, Myer RR, Latorella KA, Comstock JR Jr. et al (2018) (2011). The Multi-Attribute Task Battery II (MATB-II) Software for Human Performance and Workload Research: A User’s Guide [Data set / Technical report]. NASA Technical Reports Server. https://matb.larc.nasa.gov/ 4.Zenodo. (n.d.). Data repository [Data set]. https://zenodo.org References Wickens, C. D. (2002). Multiple resources and performance prediction. Theoretical Issues in Ergonomics Science, 3(2), 159-177. Gevins, A., & Smith, M. E. (2003). Neurophysiological measures of cognitive workload during human-computer interaction. Theoretical Issues in Ergonomics Science, 4(1-2), 113-131. Antonenko, P., Paas, F., Grabner, R., & van Gog, T. (2010). Using electroencephalography to measure cognitive load. Educational Psychology Review, 22(4), 425-438. Lotte, F., et al. (2018). A review of classification algorithms for EEG-based brain–computer interfaces: a 10-year update. Journal of Neural Engineering, 15(3), 031005. Lawhern, V. J., et al. (2018). EEGNet: a compact convolutional neural network for EEG-based brain–computer interfaces. Journal of Neural Engineering, 15(5), 056013. Data sets Center for Data Science and Neuroergonomics. (2021). Passive BCI Hackathon – Neuroergonomics 2021 (Version 1.0) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.4728939 National Aeronautics and Space Administration. (n.d.). Life Sciences Data Archive (LSDA) [Data set]. Lyndon B. Johnson Space Center. https://lsda.jsc.nasa.gov Santiago-Espada, Y., Myer, R. R., Latorella, K. A., & Comstock, J. R., Jr. (2011). The Multi-Attribute Task Battery II (MATB-II) Software for Human Performance and Workload Research: A User’s Guide [Data set / Technical report]. NASA Technical Reports Server. https://matb.larc.nasa.gov/ Zenodo. (n.d.). Data repository [Data set]. https://zenodo.org Additional Declarations There is NO Competing Interest. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7767198","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":525333222,"identity":"e20e0dfa-82a1-4ff5-a8cb-c61ad9540941","order_by":0,"name":"Senushi Dinara","email":"data:image/png;base64,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","orcid":"https://orcid.org/0009-0008-7205-205X","institution":"NA","correspondingAuthor":true,"prefix":"","firstName":"Senushi","middleName":"","lastName":"Dinara","suffix":""}],"badges":[],"createdAt":"2025-10-02 14:20:33","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7767198/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7767198/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":93106371,"identity":"f6328a9e-233e-4b1b-b57e-03214c280ab6","added_by":"auto","created_at":"2025-10-09 06:48:49","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":13480,"visible":true,"origin":"","legend":"","description":"","filename":"Nasaresearch.docx","url":"https://assets-eu.researchsquare.com/files/rs-7767198/v1/36ff064f9fc0276cf533540a.docx"},{"id":93106373,"identity":"e7a7eaf4-106f-4933-aff9-bf17ae811b4f","added_by":"auto","created_at":"2025-10-09 06:48:49","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":342,"visible":true,"origin":"","legend":"","description":"","filename":"rs7767198.json","url":"https://assets-eu.researchsquare.com/files/rs-7767198/v1/438149bcc89a63217b830e40.json"},{"id":93106370,"identity":"f6abaaf6-bc4e-4bca-b0ab-ddc134e26022","added_by":"auto","created_at":"2025-10-09 06:48:49","extension":"xml","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":24187,"visible":true,"origin":"","legend":"","description":"","filename":"rs77671982enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7767198/v1/d4194557f1bc165a9ea78fdf.xml"},{"id":93106374,"identity":"7b7eee63-67c6-4cc7-87fc-f074061e8b26","added_by":"auto","created_at":"2025-10-09 06:48:49","extension":"xml","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":22260,"visible":true,"origin":"","legend":"","description":"","filename":"rs77671982structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7767198/v1/77f0290bbb104dbdfd2d2dab.xml"},{"id":93106372,"identity":"f587ba98-6ccf-4f9b-a59a-0a28f5db7254","added_by":"auto","created_at":"2025-10-09 06:48:49","extension":"html","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":27146,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7767198/v1/8401dc74909156f7d3106db7.html"},{"id":93106375,"identity":"9de0500d-ac49-46a6-8129-5144e007dc52","added_by":"auto","created_at":"2025-10-09 06:48:53","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":434259,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7767198/v1/41f6ce31-c7ef-4f36-9587-19b9d7ccb3e2.pdf"}],"financialInterests":"There is \u003cb\u003eNO\u003c/b\u003e Competing Interest.","formattedTitle":"A Comparative Analysis of Deep Learning and Traditional Machine Learning for Classifying Cognitive Workload from Raw EEG Signals","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eCognitive workload, the mental effort required to perform a task, is a crucial factor in human performance, particularly in complex, dynamic environments like aircraft cockpits, surgical suites, and industrial control rooms [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Overload can lead to errors, while underload may result in vigilance decrements. Therefore, an objective, real-time measure of cognitive workload is highly desirable for adaptive systems and operator monitoring.\u003c/p\u003e\u003cp\u003eElectroencephalography (EEG) provides a non-invasive, high-temporal-resolution window into brain activity and has been widely established as a reliable indicator of cognitive workload [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. Traditionally, machine learning approaches for EEG classification have relied on extracting hand-crafted features from the signal (e.g., band power, statistical moments) which are then fed into classifiers like Support Vector Machines or Random Forests [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. While effective for some applications, this approach is limited by the expertise required for feature engineering and the potential to discard salient information embedded in the raw signal's temporal dynamics.\u003c/p\u003e\u003cp\u003eDeep learning, particularly Convolutional Neural Networks (CNNs), offers a powerful alternative by learning hierarchical features directly from raw data [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. In EEG analysis, 1D CNNs can scan across time and channels to automatically discover complex, task-relevant patterns without manual intervention.\u003c/p\u003e\u003cp\u003eThis research investigates the hypothesis that a 1D CNN model, by leveraging the full temporal structure of raw EEG, will significantly outperform a traditional Random Forest classifier based on engineered statistical features for a three-level cognitive workload classification task. We present a rigorous comparison of both approaches on the same dataset, demonstrating not only the superior performance of the deep learning model but also providing insights into the physiological manifestations of different workload states.\u003c/p\u003e"},{"header":"2. Methodology","content":"\u003ch2\u003e2.1. Data Acquisition and Preprocessing\u0026nbsp;\u003c/h2\u003e\n\u003cp\u003eThe study utilized a multi-participant EEG dataset (e.g., from a BCI Hackathon or MATB-II variant), comprising a total of 13,410 one-second epochs across 15 participants. Each epoch consisted of 61 EEG channels sampled at 500 Hz, resulting in a data matrix of 61 channels \u0026times; 500 time steps per epoch. The epochs were labeled according to three levels of induced cognitive workload: Class 0 (Low), Class 1 (Moderate), and Class 2 (High).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe dataset was partitioned in a stratified manner into an 80% training set (10,728 epochs) and a 20% held-out test set (2,682 epochs) to ensure proportional class representation. A StandardScaler was applied to normalize the data channel-wise to a mean of zero and a standard deviation of one. To manage computational memory constraints, a Keras Data Generator was implemented to apply scaling and reshaping on the fly during model training, loading only one batch of data at a time.\u0026nbsp;\u003c/p\u003e\n\u003ch2\u003e2.2. Deep Learning Model: 1D Convolutional Neural Network (CNN)\u0026nbsp;\u003c/h2\u003e\n\u003cp\u003eThe CNN was designed to process the raw EEG data with an input shape of (500, 61), representing 500 time steps and 61 channels.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u0026middot; Architecture: The core feature extraction module consisted of three sequential blocks, each containing a 1D Convolutional layer (with 32, 64, and 128 filters, respectively), a ReLU activation function, and a 1D Max-Pooling layer. This design progressively extracts localized temporal features while reducing dimensionality. The output was flattened and passed to a fully connected Dense layer with 128 units and a 50% Dropout layer for regularization. The final output layer used a softmax activation function to produce a probability distribution over the three classes.\u003c/p\u003e\n\u003cp\u003e\u0026middot; Training: The model was compiled with the Adam optimizer and categorical cross-entropy loss function, and trained for 20 epochs.\u003c/p\u003e\n\u003ctable style=\"width: 100%;\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 100.0000%;\"\u003eTable 1: 1D CNN Architecture\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 100.0000%;\"\u003e\n \u003cp\u003eLayer Type Configuration Output Shape Rationale\u003c/p\u003e\n \u003cp\u003eInput - (500, 61) Raw EEG epoch.\u003c/p\u003e\n \u003cp\u003eConv1D filters=32, kernel_size=3, activation=\u0026apos;ReLU\u0026apos; (498, 32) Extract local temporal patterns.\u003c/p\u003e\n \u003cp\u003eMaxPooling1D pool_size=2 (249, 32) Dimensionality reduction.\u003c/p\u003e\n \u003cp\u003eConv1D filters=64, kernel_size=3, activation=\u0026apos;ReLU\u0026apos; (247, 64) Learn more complex features.\u003c/p\u003e\n \u003cp\u003eMaxPooling1D pool_size=2 (123, 64) Further downsampling.\u003c/p\u003e\n \u003cp\u003eConv1D filters=128, kernel_size=3, activation=\u0026apos;ReLU\u0026apos; (121, 128) High-level feature extraction.\u003c/p\u003e\n \u003cp\u003eMaxPooling1D pool_size=2 (60, 128) Final pooling.\u003c/p\u003e\n \u003cp\u003eFlatten - (7680) Prepare for dense layers.\u003c/p\u003e\n \u003cp\u003eDense units=128, activation=\u0026apos;ReLU\u0026apos; (128) High-level aggregation.\u003c/p\u003e\n \u003cp\u003eDropout rate=0.5 (128) Prevent overfitting.\u003c/p\u003e\n \u003cp\u003eDense (Output) units=3, activation=\u0026apos;softmax\u0026apos; (3) Class probability distribution.\u003c/p\u003e\u003cbr\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003ch2\u003e2.3. Traditional Machine Learning Model: Random Forest (RF)\u003c/h2\u003e\n\u003cp\u003eTo establish a performance baseline, a Random Forest classifier was implemented using a feature-engineered approach.\u003c/p\u003e\n\u003cp\u003e\u0026middot; Feature Engineering: Seven statistical features were calculated for each of the 61 EEG channels over the 500-sample epoch: Mean, Median, Standard Deviation, Minimum, Maximum, Interquartile Range (IQR), and Skewness. This resulted in a feature vector of 427 dimensions per epoch (61 channels \u0026times; 7 features).\u003c/p\u003e\n\u003cp\u003e\u0026middot; Model and Training: A Random Forest classifier with 100 decision trees and a maximum depth of 15 was trained on the resulting feature matrix. The feature extraction and model fitting were performed in a memory-efficient, participant-by-participant loop to prevent RAM exhaustion.\u003c/p\u003e"},{"header":"3. Results","content":"\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e\n \u003ch2\u003e3.1. Deep Learning Model Performance\u003c/h2\u003e\n \u003cp\u003eThe 1D CNN model demonstrated exceptional performance on the unseen test set, achieving a final test accuracy of 94.2% with a corresponding loss of 0.16. The close tracking of training and validation accuracy curves indicated successful generalization with minimal overfitting.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\n \u003ch2\u003e3.2. Traditional Machine Learning Model Performance\u003c/h2\u003e\n \u003cp\u003eThe Random Forest classifier achieved a significantly lower overall test accuracy of 62.0%. A detailed classification report and confusion matrix provide further insight into its performance.\u003c/p\u003e\n \u003ctable style=\"width: 100%;\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 100.0000%;\"\u003eTable 2: Random Forest Classification Report\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 100.0000%;\"\u003e\n \u003cp\u003eClass Precision Recall F1-Score Support\u003c/p\u003e\n \u003cp\u003e0 (Low) 0.68 0.78 0.72 894\u003c/p\u003e\n \u003cp\u003e1 (Moderate) 0.56 0.43 0.49 894\u003c/p\u003e\n \u003cp\u003e2 (High) 0.60 0.65 0.63 894\u003c/p\u003e\n \u003cp\u003eAccuracy 0.62\u003c/p\u003e\n \u003cp\u003eMacro Avg 0.61 0.62 0.61 2682\u003c/p\u003e\u003cbr\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003cp\u003e\u003cbr\u003e\u003c/p\u003e\n \u003ctable style=\"width: 100%;\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 100.0000%;\"\u003eTable 3: Random Forest Confusion Matrix\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 100.0000%;\"\u003e\n \u003cp\u003eTrue \\ Predicted Class 0 Class 1 Class 2\u003c/p\u003e\n \u003cp\u003eClass 0 (Low) 695 115 84\u003c/p\u003e\n \u003cp\u003eClass 1 (Moderate) 205 383 306\u003c/p\u003e\n \u003cp\u003eClass 2 (High) 126 184 584\u003c/p\u003e\u003cbr\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e\n \u003ch2\u003e3.3. Comparative Performance Summary\u003c/h2\u003e\n \u003cp\u003eThe central finding of this study is the stark performance difference between the two modeling paradigms, as summarized below.\u003c/p\u003e\n \u003ctable style=\"width: 100%;\" class=\"fr-table-selection-hover\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 100.0000%;\"\u003eTable 4: Model Comparison\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 100.0000%;\"\u003e\n \u003cp\u003eModel Type Input Data Final Test Accuracy\u003c/p\u003e\n \u003cp\u003eDeep Learning (1D CNN) Raw Time-Series (500, 61) 94.2%\u003c/p\u003e\n \u003cp\u003eTraditional ML (Random Forest) 427 Engineered Features 62.0%\u003c/p\u003e\u003cbr\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e"},{"header":"4. Discussion","content":"\u003cp\u003eThe primary objective of this research was to determine the most effective machine learning approach for classifying cognitive workload from EEG. The results provide a clear and definitive answer.\u003c/p\u003e\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\u003ch2\u003e4.1. Superiority of Deep Learning for Temporal Feature Extraction\u003c/h2\u003e\u003cp\u003eThe most significant finding is the 32.2% absolute difference in accuracy between the CNN and Random Forest models. This substantial gap underscores a critical point: the raw, millisecond-level temporal dynamics of the EEG signal contain a wealth of discriminative information that is lost when the signal is reduced to summary statistics like mean or standard deviation. The 1D CNN's convolutional filters successfully learned to identify complex, localized patterns across time and channels that are highly predictive of workload state, a capability beyond the scope of the manually engineered features.\u003c/p\u003e\u003cp\u003eThis result strongly validates the use of end-to-end deep learning models for complex EEG classification tasks, as they circumvent the limitations and biases of manual feature engineering.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e\u003ch2\u003e4.2. The Challenge of Classifying Moderate Workload\u003c/h2\u003e\u003cp\u003eThe analysis of the Random Forest's performance offers a secondary, important insight. The model struggled most severely with Class 1 (Moderate Workload), as evidenced by its lowest recall (0.43) and F1-score (0.49). The confusion matrix (Table\u0026nbsp;3) reveals that 205 Moderate epochs were misclassified as Low, and 306 were misclassified as High. This suggests that the \"Moderate\" cognitive state is a physiologically ambiguous transition zone, sharing characteristics with both the relaxed (Low) and overloaded (High) states. The CNN's ability to learn nuanced, non-linear boundaries in the high-dimensional raw data was essential for accurately dissecting this ambiguous class.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e\u003ch2\u003e4.3. Limitations and Future Work\u003c/h2\u003e\u003cp\u003eA limitation of this study is the use of a specific, task-induced workload dataset. Future work should validate the generalizability of the trained CNN model on data from different experimental paradigms and participant populations. Furthermore, techniques like Grad-CAM could be applied to the trained CNN to identify which time periods and channels most strongly influence the classification decision, enhancing the interpretability of the model.\u003c/p\u003e\u003cp\u003eThe high-performing CNN model has been saved and is readily deployable. The immediate next step is to integrate this model into an edge computing system for low-latency, real-time cognitive workload assessment in operational settings, such as providing feedback during pilot training or triggering alerts for system operators.\u003c/p\u003e\u003c/div\u003e"},{"header":"5. Conclusion","content":"\u003cp\u003eThis study successfully demonstrated that a 1D Convolutional Neural Network can classify cognitive workload from raw EEG data with high fidelity (94.2% accuracy), dramatically outperforming a traditional Random Forest classifier based on statistical features (62.0% accuracy). The findings confirm that deep learning is superior for automatically extracting the complex temporal patterns in EEG that are indicative of mental state, and that these patterns are more informative than hand-crafted features. The developed model provides a robust, accurate, and readily deployable backend for real-time cognitive workload monitoring systems, with significant potential for improving safety and performance in high-demand professions.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eWickens CD (2002) Multiple resources and performance prediction. Theoretical Issues Ergon Sci 3(2):159\u0026ndash;177\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGevins A, Smith ME (2003) Neurophysiological measures of cognitive workload during human-computer interaction. Theoretical Issues Ergon Sci 4(1\u0026ndash;2):113\u0026ndash;131\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAntonenko P, Paas F, Grabner R, van Gog T (2010) Using electroencephalography to measure cognitive load. Educational Psychol Rev 22(4):425\u0026ndash;438\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLotte F et al (2018) A review of classification algorithms for EEG-based brain\u0026ndash;computer interfaces: a 10-year update. J Neural Eng 15(3):031005\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLawhern VJ EEGNet: a compact convolutional neural network for EEG-based brain\u0026ndash;computer interfaces. Journal of Neural Engineering, 15(5), 056013. Data sets 1.Center for Data Science and Neuroergonomics. (2021). Passive BCI Hackathon \u0026ndash; Neuroergonomics 2021 (Version 1.0) [Data set]. Zenodo., Santiago-Espada Y, Myer RR, Latorella KA, Comstock JR Jr. et al (2018) (2011). The Multi-Attribute Task Battery II (MATB-II) Software for Human Performance and Workload Research: A User\u0026rsquo;s Guide [Data set / Technical report]. NASA Technical Reports Server. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://matb.larc.nasa.gov/\u003c/span\u003e\u003cspan address=\"https://matb.larc.nasa.gov/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e 4.Zenodo. (n.d.). Data repository [Data set]. https://zenodo.org\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eWickens, C. D. (2002). Multiple resources and performance prediction. Theoretical Issues in Ergonomics Science, 3(2), 159-177.\u003c/li\u003e\n \u003cli\u003eGevins, A., \u0026amp; Smith, M. E. (2003). Neurophysiological measures of cognitive workload during human-computer interaction. Theoretical Issues in Ergonomics Science, 4(1-2), 113-131.\u003c/li\u003e\n \u003cli\u003eAntonenko, P., Paas, F., Grabner, R., \u0026amp; van Gog, T. (2010). Using electroencephalography to measure cognitive load. Educational Psychology Review, 22(4), 425-438.\u003c/li\u003e\n \u003cli\u003eLotte, F., et al. (2018). A review of classification algorithms for EEG-based brain\u0026ndash;computer interfaces: a 10-year update. Journal of Neural Engineering, 15(3), 031005.\u003c/li\u003e\n \u003cli\u003eLawhern, V. J., et al. (2018). EEGNet: a compact convolutional neural network for EEG-based brain\u0026ndash;computer interfaces. Journal of Neural Engineering, 15(5), 056013.\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003e\u003cstrong\u003eData sets\u003c/strong\u003e\u003c/p\u003e\n\u003col start=\"1\" type=\"1\"\u003e\n \u003cli\u003eCenter for Data Science and Neuroergonomics. (2021). Passive BCI Hackathon \u0026ndash; Neuroergonomics 2021 (Version 1.0) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.4728939\u003c/li\u003e\n \u003cli\u003eNational Aeronautics and Space Administration. (n.d.). Life Sciences Data Archive (LSDA) [Data set]. Lyndon B. Johnson Space Center. https://lsda.jsc.nasa.gov\u003c/li\u003e\n \u003cli\u003eSantiago-Espada, Y., Myer, R. R., Latorella, K. A., \u0026amp; Comstock, J. R., Jr. (2011). The Multi-Attribute Task Battery II (MATB-II) Software for Human Performance and Workload Research: A User\u0026rsquo;s Guide [Data set / Technical report]. NASA Technical Reports Server. https://matb.larc.nasa.gov/\u003c/li\u003e\n \u003cli\u003eZenodo. (n.d.). Data repository [Data set]. https://zenodo.org\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Cognitive Workload, EEG, Deep Learning, 1D Convolutional Neural Network, Random Forest, Classification, BCI","lastPublishedDoi":"10.21203/rs.3.rs-7767198/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7767198/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cdiv language=\"En\" class=\"ArticleSubTitle\"\u003eThe objective assessment of cognitive workload is critical for enhancing performance and safety in high-stakes environments such as aviation and process control. This study presents a comparative analysis of two machine learning paradigms for classifying cognitive workload into three distinct levels (Low, Moderate, High) using electroencephalography (EEG). We developed and evaluated a deep learning model based on a 1D Convolutional Neural Network (CNN) that processes raw time-series EEG data, and compared it against a traditional machine learning baseline, a Random Forest (RF) classifier, trained on hand-engineered statistical features. The CNN model achieved a superior test accuracy of 94.2%, significantly outperforming the Random Forest model, which achieved an accuracy of 62.0%. This 32.2% performance gap strongly indicates that the raw temporal structure of EEG signals contains discriminative features for workload classification that are not captured by standard statistical summaries. The results validate the efficacy of deep learning for automated feature extraction in neurophysiological data and provide a robust, deployable model for real-time cognitive workload monitoring systems.\u003c/div\u003e","manuscriptTitle":"A Comparative Analysis of Deep Learning and Traditional Machine Learning for Classifying Cognitive Workload from Raw EEG Signals","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-10-09 06:48:44","doi":"10.21203/rs.3.rs-7767198/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"48784343-1808-4a30-a8c6-a270c0883a06","owner":[],"postedDate":"October 9th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":55827713,"name":"Biological sciences/Neuroscience"},{"id":55827714,"name":"Biological sciences/Neuroscience/Computational neuroscience"}],"tags":[],"updatedAt":"2025-10-09T06:48:44+00:00","versionOfRecord":[],"versionCreatedAt":"2025-10-09 06:48:44","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7767198","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7767198","identity":"rs-7767198","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}