PaperlayerView exact revision on arXiv

Source-linked AI summary

Decoding Imagined Speech: A Strictly Subject-Independent Approach Using EEG

Frederik Møllskov Trier, Xiaopeng Mao, Sadasivan Puthusserypady

arXiv:2609.29820v1cs.AI

TL;DR

Prior evaluations of the Kumar EEG dataset often leave their split protocols unclear, making cross-subject generalization uncertain. This study compares statistical time-domain and spectral pipelines under subject-wise cross-validation and trial-level voting; the Spectral Pipeline achieves higher coarse-level accuracy.

  • Problem

    Kumar-dataset evaluations often do not describe data splits sufficiently to establish whether reported performance reflects subject-independent generalization, which matters for practical BCI applications.

    Source: §I · p. 1

  • Method

    The study compares statistical time-domain and spectral feature pipelines on Kumar’s EEG dataset using subject-wise cross-validation, trial-level majority voting, and a random forest classifier.

    Source: §I · p. 1, C. Common Preprocessing and Segmentation · p. 2

  • Results

    49.03 ± 4.18% vs. 37.97 ± 3.79%: the Spectral Pipeline significantly exceeded the Kumar Pipeline in mean trial-wise coarse-level accuracy (p = 0.018).

    Source: §III · p. 4

  • Takeaways & Limitations

    Forward selection indicated that alpha and delta bands captured most of the discriminative structure, suggesting that improved generalization does not require high-dimensional feature spaces.

    Source: §V · p. 4

    The pipelines used minimal artifact removal, which the authors identify as a limitation because artifact removal improves data quality and representation of relevant brain activity.

    Source: §IV · p. 4

Abstract

from arXiv · show

Imagined speech decoding from electroencephalography (EEG) has gained increasing attention as a potential communication pathway for individuals with severe motor impairments, yet reported performance often relies on evaluation protocols that do not clearly reflect cross-subject generalization. This study presents a transparent baseline investigation of a multi-class imagined speech EEG dataset under a strictly subject-independent evaluation framework. Two preprocessing and feature extraction pipelines were compared: a time-domain statistical feature approach and a frequency-domain spectral bandpower approach, evaluated using subject-wise cross-validation and trial-level majority voting with a random forest classifier. The spectral pipeline achieved a significantly higher mean trial-wise accuracy than the statistical pipeline (49.03 $\pm$ 4.18% vs. 37.97 $\pm$ 3.79%) for coarse-level classification across subjects. Forward feature selection further indicated that a limited subset of frequency bands captured most of the discriminative information. Overall, this work provides a strong basis for future brain-computer interface studies targeting improved cross-subject generalization in EEG-based imagined speech decoding.

I. INTRODUCTION

Imagined-speech decoding is proposed as a communication route for people with severe motor impairment, but reliable assessment of cross-subject generalization remains unclear. This study compares two feature pipelines for the Kumar EEG dataset under a transparent subject-independent framework.

I. INTRODUCTION

  • Imagined-speech decoding is proposed as a potential communication route that translates brain activity into text or symbols, but EEG signals have low signal-to-noise ratios.The introduction connects feature quality with machine-learning decoding.Source: §I · p. 1
  • The dataset has three coarse classes—letters, digits, and objects—each divided into ten fine-grained subclasses.The original study used statistical time-domain features and a random forest, reporting 85.20% coarse-level and 67.03% fine-grained accuracy.Source: §I · p. 1
  • Prior Kumar-dataset studies used protocols that were not described sufficiently to determine whether splits were by window or trial and subject or non-subject, leaving generalizability uncertain.Strict subject-independent evaluation is particularly important for practical BCI applications.Source: §I · p. 1
  • The study compares the Kumar statistical time-domain pipeline with a frequency-domain spectral pipeline using subject-wise cross-validation and trial-level majority voting.The accompanying overview depicts the two preprocessing and feature-extraction paths followed by shared cross-validation and statistical performance comparison.Source: §I · p. 1, Figure 1 · p. 2

II. METHODOLOGY

The methodology compares two preprocessing and feature-extraction pipelines, then describes the shared dataset and evaluation framework. The dataset contains recordings from 23 people imagining targets from three coarse categories.

II. METHODOLOGY

  • The methodology section covers the comparison framework, Kumar’s EEG dataset, and descriptions of both pipelines.Source: §II · p. 2

A. Proposed Method

  • The proposed comparison contrasts the Kumar Pipeline with the Spectral Pipeline while keeping the cross-validation framework identical.Fig. 1 presents the comparison as a left-to-right flow from preprocessing and feature extraction to statistical performance comparison.Source: A. Proposed Method · p. 2

B. Dataset

  • The dataset comprises 690 ten-second recordings from 23 subjects aged 15–40, who imagined 30 targets spanning characters, digits, and everyday images.EEG was collected with 14 Emotiv EPOC+ channels, captured at 2048 Hz and down-sampled to 128 Hz; coarse-level classification distinguishes the three target groups.Source: B. Dataset · p. 2

C. Common Preprocessing and Segmentation

Shared processing retains the central ten seconds of each trial, segments the data into overlapping windows, and applies subject-wise feature normalization. The section also documents dataset indexing corrections and the trial-level evaluation workflow.

C. Common Preprocessing and Segmentation

  • Because trials contain 12 seconds without onset markers, the central 10 seconds were retained, and digit-condition folders were manually re-indexed without changing trials.The authors attributed index discrepancies to errors because the dataset documentation consistently reports 23 subjects.Source: C. Common Preprocessing and Segmentation · p. 2
  • Each 10-second trial was divided into 1-second windows with 75% overlap across all 14 channels, balancing temporal and frequency resolution for spectral features.Subject-wise z-score normalization was applied before classification; the feature value x is normalized using its subject-wise mean µ and standard deviation σ.Source: C. Common Preprocessing and Segmentation · p. 2
  • Fig. 2 shows representative stimuli from the character, digit, and image classes, arranged from left to right and adapted from [4].Source: Figure 2 · p. 2
  • Fig. 3 depicts subject-wise folds in which held-out-subject window predictions are combined by majority voting to obtain trial-wise accuracy averaged across folds.Source: Figure 3 · p. 3

D. Kumar Pipeline

The Kumar Pipeline follows the original study’s statistical-feature approach, with overlapping windows added to align its evaluation with the Spectral Pipeline. It computes four statistics per channel and transforms skewed standard-deviation features before normalization.

D. Kumar Pipeline

  • The Kumar Pipeline smooths 128 Hz EEG with a 5-point moving-average filter and uses the shared 1-second overlapping windows for comparability with the Spectral Pipeline.Source: D. Kumar Pipeline · p. 3
  • Four statistics—standard deviation, root mean square, sum, and signal energy—produce a 56-dimensional feature vector from 14 channels for each window.The resulting feature matrix has size 27600 × 56.Source: D. Kumar Pipeline · p. 3
  • Because standard-deviation features were positively skewed, they were logarithmically transformed before all features were normalized.Source: D. Kumar Pipeline · p. 3

E. Spectral Pipeline

The Spectral Pipeline filters EEG, estimates power spectral density, and converts frequency-band power into normalized features for classification.

E. Spectral Pipeline

  • A 50 Hz zero-phase notch filter removes powerline interference, followed by a zero-phase 4th-order Butterworth bandpass filter from 0.5–60 Hz.The bandpass filtering reduces slow drift while preserving the temporal structure needed for FFT-based spectral estimation.Source: E. Spectral Pipeline · p. 3
  • Power spectral density is estimated per window with a periodogram, and absolute band power is calculated by trapezoidal integration.The band-power equation uses the PSD estimate and frequency resolution to compute power within each frequency band.Source: E. Spectral Pipeline · p. 3
  • Five bands span delta [0.5, 4), theta [4, 8), alpha [8, 14), beta [14, 30), and gamma [30, 64] Hz; log transformation and z-score normalization yield 70 features per window.The 70 features comprise five band-power values across 14 channels.Source: E. Spectral Pipeline · p. 3

F. Evaluation and Statistical Analysis

The evaluation uses subject-wise cross-validation, aggregates window predictions to trials, and compares pipeline performance across subjects.

F. Evaluation and Statistical Analysis

  • The random forest uses bootstrap aggregation, majority voting across trees, and the Gini impurity criterion, with 200 trees and other settings left at scikit-learn defaults.This classifier choice is consistent with Kumar et al..Source: F. Evaluation and Statistical Analysis · p. 3
  • Five-fold cross-validation uses subject-wise splits and class stratification; window predictions within each test fold are majority-voted into trial-level predictions.Final performance averages trial-wise accuracies across folds.Source: F. Evaluation and Statistical Analysis · p. 3
  • Pipeline comparison averages trial accuracies within each subject, then applies a paired Wilcoxon signed-rank test to the subject-wise differences.The test statistic uses ranks of absolute nonzero differences, with each rank signed according to its difference.Source: F. Evaluation and Statistical Analysis · p. 3

G. Forward Feature Selection

Forward feature selection tests how individual frequency bands and their combinations contribute to spectral-pipeline classification accuracy.

G. Forward Feature Selection

III. RESULTS

For coarse-level decoding, the Spectral Pipeline outperformed the Kumar Pipeline; class-wise performance and frequency-band combinations were also examined.

III. RESULTS

  • 49.03 ± 4.18% vs. 37.97 ± 3.79% mean trial-wise accuracy: the Spectral Pipeline significantly outperformed the Kumar Pipeline for coarse-level prediction.The paired Wilcoxon test gave p = 0.018, with rank-biserial correlation = 0.57.Source: Table I · p. 4, §III · p. 4
  • Both pipelines achieved their highest class-wise accuracy for characters, with lower performance for digits and images.Source: Table II · p. 4, §III · p. 4
  • Alpha had the highest single-band mean accuracy, adding delta substantially improved performance, and α+δ+γ+θ achieved the highest mean accuracy; no band combination differed significantly from full-band performance.Table III marks mean accuracies exceeding the full-band model.Source: §III · p. 4, Table III · p. 4

IV. DISCUSSION

The spectral pipeline outperformed the time-domain pipeline under subject-independent evaluation, while class-dependent performance and minimal artifact removal qualify the findings.

IV. DISCUSSION

  • The Spectral Pipeline outperformed the Kumar Pipeline, indicating better cross-subject generalization for frequency-domain features in this setting.Absolute accuracies were lower than those reported by Kumar et al. under the strictly subject-independent protocol.Source: §IV · p. 4
  • Characters were classified more accurately than digits and images, while alpha-band power was most discriminative and alpha-plus-delta performance was not significantly different from the full-band model.The findings suggest that much of the coarse-level classification information was captured by a limited set of frequency bands.Source: §IV · p. 4
  • Minimal artifact removal constrained the pipelines; channel selection and additional spectral features are potential directions for improving discrimination.The study did not perform channel selection and used only absolute band power as spectral features.Source: §IV · p. 4

V. CONCLUSION

The study establishes a subject-independent comparison of statistical and spectral EEG features, finding better coarse-level performance with spectral features and identifying a useful limited band subset.

V. CONCLUSION

  • Under strict subject-wise cross-validation, the spectral band-power pipeline outperformed the time-domain statistical pipeline for coarse-level imagined-speech classification.The study presents this as a realistic assessment of cross-subject generalization on the Kumar EEG dataset.Source: §V · p. 4
  • Forward feature selection found that a limited subset, particularly alpha and delta bands, captured most of the discriminative structure, while performance varied across imagined-speech categories.The conclusion states that improved generalization did not require high-dimensional feature spaces.Source: §V · p. 4
  • The work offers a reproducible baseline for subject-independent imagined-speech research and a foundation for more generalizable EEG-based BCI systems.Its findings underscore feature representation when evaluation protocols reflect real-world constraints.Source: §V · pp. 4–5
Loading 2609.29820v1…