PaperlayerView exact revision on arXiv

Source-linked AI summary

An Evaluation of the EEG alpha-to-theta and theta-to-alpha band Ratios as Indexes of Mental Workload

Bujar Raufi, Luca Longo

arXiv:2202.12937v2eess.SPcs.AIcs.HCcs.LG

TL;DR

Evidence is limited on whether alpha-to-theta and theta-to-alpha EEG ratios help discriminate self-reported mental workload. The study extracts high-level features from these ratios and trains classifiers using EEG from rest and multitasking conditions. Alpha-to-theta indexes yielded models with higher performance than other indexes, while the electrode-count comparison remained inconclusive.

  • Problem

    Although EEG bands and some ratios have been studied, evidence is limited on how alpha-to-theta and theta-to-alpha ratios discriminate levels of self-reported mental workload.

    Source: §2 · p. 4

  • Method

    The study forms alpha-to-theta and theta-to-alpha indexes from EEG recordings and uses high-level features to train models against users’ self-reported workload ratings.

    Source: §1 · p. 3, Experiment Design and Dataset… · p. 5

  • Results

    Alpha-to-theta indexes enabled models with minimum 18.4%–30.2% higher performance, measured by f1-score, accuracy, precision, and recall, than other indexes.

    Source: §5 · p. 19

  • Takeaways & Limitations

    The authors conclude that alpha-to-theta and theta-to-alpha workload indexes can significantly discriminate self-reported workload and be used to design detection models.

    Source: §5 · p. 19

    The results do not conclusively show that using more electrodes, especially in the parietal region, improves discrimination of self-reported workload.

    Source: §5 · p. 19

Abstract

from arXiv · show

Many research works indicate that EEG bands, specifically the alpha and theta bands, have been potentially helpful cognitive load indicators. However, minimal research exists to validate this claim. This study aims to assess and analyze the impact of the alpha-to-theta and the theta-to-alpha band ratios on supporting the creation of models capable of discriminating self-reported perceptions of mental workload. A dataset of raw EEG data was utilized in which 48 subjects performed a resting activity and an induced task demanding exercise in the form of a multitasking SIMKAP test. Band ratios were devised from frontal and parietal electrode clusters. Building and model testing was done with high-level independent features from the frequency and temporal domains extracted from the computed ratios over time. Target features for model training were extracted from the subjective ratings collected after resting and task demand activities. Models were built by employing Logistic Regression, Support Vector Machines and Decision Trees and were evaluated with performance measures including accuracy, recall, precision and f1-score. The results indicate high classification accuracy of those models trained with the high-level features extracted from the alpha-to-theta ratios and theta-to-alpha ratios. Preliminary results also show that models trained with logistic regression and support vector machines can accurately classify self-reported perceptions of mental workload. This research contributes to the body of knowledge by demonstrating the richness of the information in the temporal, spectral and statistical domains extracted from the alpha-to-theta and theta-to-alpha EEG band ratios for the discrimination of self-reported perceptions of mental workload.

1 INTRODUCTION

Mental workload is a complex construct used to investigate human performance, and EEG bands have been proposed as indicators of workload. This study addresses limited evidence on whether alpha-to-theta and theta-to-alpha ratios can help discriminate users’ self-reported workload.

1 INTRODUCTION

  • Mental workload is described as the cognitive cost of performing a task within a finite timeframe, but its definition remains ambiguous.The paper presents workload as a construct used to investigate human performance.Source: §1 · p. 1
  • EEG offers high time resolution and decomposes signals into frequency bands, including theta and alpha, which prior work has related to cognitive processes and workload.The paper distinguishes subjective, task-performance, and physiological approaches to workload measurement.Source: §1 · p. 2
  • The study investigates whether high-level features from alpha-to-theta and theta-to-alpha ratios can classify levels of self-reported mental workload.The proposed experiment generates ratio time series and extracts features for classification models.Source: §1 · p. 3

2 RELATED WORK

  • Research had not established how alpha-to-theta and theta-to-alpha ratios contribute to discriminating users’ self-reported workload, despite prior work on EEG bands and other ratios.The article frames this gap as a question about high-level features extracted from the ratios and their combinations.Source: §1 · p. 2, §2 · p. 4

3 DESIGN AND METHODOLOGY

The study uses a comparative empirical design to test whether workload indexes based on alpha-to-theta and theta-to-alpha ratios discriminate self-reported workload. Its pipeline processes EEG recordings, forms frontal and parietal electrode clusters, and uses the resulting indexes for classification.

3 DESIGN AND METHODOLOGY

  • The empirical comparison tests whether ratio-based workload indexes discriminate self-reported workload better than single-band indexes, and whether using more adjacent electrodes improves discrimination.These are the study’s H1 and H2 hypotheses.Source: §3 · p. 4
  • The process selects frontal theta and parietal alpha electrodes, computes alpha-to-theta and theta-to-alpha indexes, extracts high-level features, and trains and evaluates classifiers.The figure presents this sequence as the study’s classification pipeline.Source: Figure 1 · p. 5

Experiment Design and Dataset Description

  • The STEW dataset contains EEG recordings from 48 subjects during rest and a multitasking SIMKAP test, with a 1–9 self-reported workload rating after each condition.Recordings used 14 channels, a 128 Hz sampling frequency, and 2.5 minutes per condition.Source: Experiment Design and Dataset… · p. 5

A EEG Denoising Pipeline

  • The preprocessing pipeline applies average re-referencing, 1 Hz high-pass filtering, and ICA-based source separation and artefact removal.Fourteen components are generated; outliers beyond z-score ±3 are treated as artefacts before inverse ICA reconstructs the signal.Source: A EEG Denoising Pipeline · p. 6

B Forming Cluster Combinations

  • The available headset electrodes were organized into three frontal clusters and one parietal cluster for analysis of theta and alpha activity.Table 1 presents the selected clusters and electrode combinations.Source: B Forming Cluster Combinations · p. 6, Table 1 · p. 6

C Formation of the Mental Workload indexes from clusters of EEG alpha and theta bands

The section defines EEG workload indexes from alpha- and theta-band cluster ratios, extracts and selects time-series features, and trains classifiers against self-reported workload labels.

C Formation of the Mental Workload indexes from clusters of EEG alpha and theta bands

  • Three alpha-to-theta and theta-to-alpha ratios pair average alpha- and theta-band PSD values from frontal and parietal electrode clusters, yielding indexes per non-overlapping 1-second window.The configurations combine individual clusters and their ratios.Source: C Formation of the Mental Workload… · p. 7

D Feature Extraction from indexes and Selection

  • TSFEL extracts statistical, frequency-domain, and temporal-domain features from workload-index time series to capture properties not available from raw indexes alone.Examples include variability, slope, peak-to-peak, kurtosis, mean, and median.Source: D Feature Extraction from indexes and… · p. 7
  • SelectKBest ranks features by ANOVA F-value; the feature count is halved until performance drops, then Pearson correlation helps reduce multicollinearity.The authors also describe a ±0.5 correlation threshold as optimal in cited prior work.Source: D Feature Extraction from indexes and… · pp. 7–8

E Models Training

  • Models target self-reported workload rather than Rest or Simkap condition because task conditions do not determine each user’s subjective workload.The paper notes that experience and other factors can make workload vary within the same condition.Source: E Models Training · p. 8
  • Synthetic data are added because the 48-subject dataset yields few training instances; quality assessment combines field-correlation, deep-structure, and field-distribution stability.Lower average Jensen-Shannon distance indicates higher field-distribution stability; the overall synthetic-data quality score averages the three stability measures.Source: E Models Training · p. 9
  • Logistic regression, SVM, and decision trees are trained using randomized 70% subject splits and tested on the remaining 30% across 100 iterations.The training passage also describes z-score normalization using the training-sample mean and standard deviation.Source: E Models Training · pp. 8–10

F Models Evaluation

  • Accuracy, precision, recall, and F1-score evaluate models on unseen data, with F1-score included to account for the precision–recall trade-off when testing hypotheses H1 and H2.The passages define accuracy, precision, recall, and F1-score using binary-classification outcomes.Source: F Models Evaluation · p. 10

4 RESULTS

The results section follows the same order as the steps outlined in the design section.

4 RESULTS

  • The results section follows the same order as the steps outlined in the design section.This passage describes the section’s organization.Source: §4 · p. 10

I EEG artefact removal

The study removed ICA components from EEG recordings in both conditions, reporting average removal counts and describing the resulting signal reconstruction.

I EEG artefact removal

  • 1.61 components were removed on average per subject in Rest, compared with 1.46 in Simkap.Figure 2 depicts component-removal occurrence across 48 subjects and 14 components in the two conditions.Source: I EEG artefact removal · p. 10, Figure 2 · p. 11
  • At most one component per subject was identified as significantly different from the others by the ±3-standard-deviation criterion and removed by zeroing it.The EEG data were then reconstructed using inverse ICA.Source: I EEG artefact removal · p. 10
  • Because at least one bad component was identified and removed for most subjects, the authors say some artefact was reasonably removed, facilitating later alpha- and theta-band computation.Source: I EEG artefact removal · p. 11

II Evaluation of Feature Extraction and Selection

Feature selection reduced 210 candidate statistics to about seven features, while models trained on the original data underfit and ratio indexes performed better than standalone alpha or theta indexes. The small spectral-entropy variance motivated synthetic-data training.

II Evaluation of Feature Extraction and Selection

  • 210 temporal- and frequency-domain statistical features were assessed, and iterative selection identified about seven top features as the stopping point.The selected features were retained for subsequent training; Pearson correlations were also examined to reduce multicollinearity.Source: II Evaluation of Feature Extraction… · p. 11, Figure 3 · p. 12
  • Most selected features for the at-1 alpha-to-theta index fell between −0.5 and +0.5 correlation in both Rest and Simkap conditions.The authors report that other indexes were mostly consistent with this example.Source: II Evaluation of Feature Extraction… · p. 11, Figure 4 · p. 12

III Evaluation of the training set across indexes

  • The original-data training set comprised 48 subjects, and its limited variance was linked to bias that impaired generalization.The authors applied synthetic data to train more robust models.Source: III Evaluation of the training set… · p. 13
  • On original data, alpha-to-theta and theta-to-alpha indexes achieved 70% accuracy, versus more than 60% for standalone alpha and theta indexes; learning curves indicated underfitting and inability to generalize from test data.Spectral-entropy analysis revealed small variance, which the authors say increases bias; they applied synthetic data generation to train robust models.Source: III Evaluation of the training set… · p. 13

IV Synthetic data evaluation

Synthetic datasets were generated separately for Rest and Simkap, expanding the combined data to 228 subjects while retaining the original dataset’s intrinsic properties.

IV Synthetic data evaluation

  • Synthetic quality scores exceeded 87% across the selected workload indexes, placing them in the authors’ “Excellent” range.The paper defines scores from 80% to 100% as Excellent.Source: IV Synthetic data evaluation · p. 13
  • The synthesis input comprised 48 subjects with 150 one-second EEG segments per workload index, representing 2.5 minutes per participant.Separate synthetic datasets were created for the Rest and Simkap conditions.Source: IV Synthetic data evaluation · p. 13
  • The study added 180 synthetic subjects, producing a combined dataset of 228 subjects with 150 segments per index.Each segment represented one second of EEG activity.Source: IV Synthetic data evaluation · p. 13

V Validation of models for discriminating self-reported perceptions of mental workload

Support Vector Machines and Logistic Regression had the best observed classification accuracy, but their performance was not significantly different from Decision Trees. The first two alpha-to-theta and theta-to-alpha indexes generally outperformed their component indexes.

V Validation of models for discriminating self-reported perceptions of mental workload

  • Table 2 covers synthetic scores across workload indexes and Rest and Simkap conditions, while Table 3 compares learning techniques across four evaluation metrics.The metrics are f1-score, accuracy, recall, and precision.Source: Table 2 · p. 14, Table 3 · p. 14
  • Support Vector Machines and Logistic Regression showed the best classification accuracy, but no significant differences from Decision Trees were found across the reported metrics.The two-tailed tests used α = 0.05; Figure 6 presents classification results across learning techniques.Source: V Validation of models for… · p. 14, Table 3 · p. 14, Figure 6 · p. 15
  • The first two alpha-to-theta indexes outperformed their component alpha and theta indexes across learning techniques.The first two theta-to-alpha indexes also outperformed their component indexes.Source: V Validation of models for… · p. 14
  • Comparisons among ratio indexes with differing electrode counts showed statistically significant performance differences after Bonferroni correction at α = 0.005.The analysis compared at-1 and at-3 with at-2, and ta-1 and ta-3 with ta-2; the supplied passage does not state which performed better.Source: V Validation of models for… · p. 15

5 DISCUSSION

The discussion reports that alpha-to-theta and theta-to-alpha indexes discriminate self-reported workload better than individual band indexes, while the effect of electrode count remains unresolved. It also describes synthetic-data augmentation as a way to address small EEG datasets.

5 DISCUSSION

  • The authors present synthetic-data generation as a way to expand limited EEG datasets and describe constructing indexes from alpha and theta bands individually and in combination.Statistical features from these indexes are used to discriminate self-reported workload perceptions.Source: §5 · p. 16
  • From an initial 60% accuracy for individual alpha and theta indexes on the original dataset, classifier performance increased by 8% to 20% with synthetic data augmentation.Table 5 is identified as comparing model performance on the original dataset and the dataset combined with synthetic data.Source: §5 · p. 17, Table 5 · p. 18
  • 18.4%–30.2% higher performance was reported for workload ratio indexes, especially alpha-to-theta indexes, than for the other indexes across f1-score, accuracy, precision, and recall.The discussion also reports that at-1, at-2, ta-1, and ta-2 better discriminate self-reported perceptions than their individual band counterparts.Source: §5 · p. 19
  • The electrode-count hypothesis remains inconclusive: at-2 and ta-2 perform better despite using indexes with fewer electrodes, and the proposed explanation involving the sequential Rest and SIMKAP activities requires further validation.The authors hypothesize that the task transition may affect neuromodulation in parietal regions and neuronal suppression in frontal regions.Source: §5 · p. 19
  • The results support alpha-to-theta and theta-to-alpha ratios as discriminators of self-reported workload, but do not establish that using more electrodes improves discrimination.The authors report that classification metrics indicate good performance across almost all ratio indexes.Source: §5 · p. 19

6 CONCLUSION

The conclusion reports that alpha-to-theta and theta-to-alpha EEG ratios can serve as indexes for discriminating self-reported workload, while identifying validation and scope extensions for future work.

6 CONCLUSION

  • The alpha-to-theta and theta-to-alpha ratios can be treated as mental workload indexes, with extracted higher-level features supporting models that distinguish two levels of self-reported workload.The levels are “suboptimal MWL” and “super optimal MWL.”Source: §6 · p. 19
  • The authors propose replication with additional publicly available datasets to further validate the contribution.Source: §6 · p. 20
  • Future work would evaluate tasks beyond those employed, including examples from automobile, human-computer interaction, and education settings.Source: §6 · p. 20
  • The authors propose using multi-channel EEG from a larger electrode pool and evaluating additional indexes built from different alpha- and theta-band electrode clusters.Source: §6 · p. 20
  • Future work would add task-load conditions of incremental complexity and objective task-performance measures as dependent features.Source: §6 · p. 20

FUNDING

The research was funded by an MCSA Post-doc CareerFIT fellowship.

FUNDING

  • The research was funded by the MCSA Post-doc CareerFIT fellowship, supported by Enterprise Ireland and the European Commission.Fellowship reference: MF2020 0144.Source: FUNDING · p. 20

http://arxiv.org/ps/2202.12937v2

This section lists figure-file references, including files named for the EEG pipeline and classifier and feature accuracies.

http://arxiv.org/ps/2202.12937v2

Related papers

Loading 2202.12937v2…