Contiguous temporal withholding for hemispheric analysis in frontal EEG during cycling

Abstract

We present a methodological framework for evaluating temporal sensitivity in frontal EEG recordings during cycling. The approach departs from randomized train–test splits and instead implements contiguous temporal withholding to preserve physiological ordering. Class-wise recall is proposed as a time-indexed indicator of hemispheric recognizability under progressive temporal displacement.

The method was evaluated on proof-of-concept data obtained from four healthy volunteers performing sustained lower-limb exercise while frontal EEG was recorded. Signals were segmented into contraction-aligned epochs (i.e., time-locked signal segments) and organized as within-subject temporal sequences. A contiguous 20% segment of data was withheld across 20 equally spaced temporal positions, and model training was repeated five times under identical hyperparameters.

To validate temporal sensitivity, the method is evaluated on its ability to capture evolving hemispheric recognizability across the pedaling sequence, rather than separability. Interhemispheric differences are subsequently derived from these recall-based measures. We assess whether block-to-block recall modulation exceeds repetition-related variability. Results show consistent block-dependent modulation with repetition dispersion, indicating the method detects structured temporal shifts rather than stochastic training effects.

The framework provides a structured way to evaluate hemispheric differentiation during gradual cortical reorganization, without relying on maximal classification accuracy.

• Contiguous-block temporal withholding

• Repeated-training recall estimation as a temporal stability index

• Block-wise interhemispheric analysis

Graphical abstract

Specifications table

Subject area	Neuroscience
More specific subject area	Computational EEG analysis; Temporal evaluation of sequential neural time-series; Within-subject hemispheric classification using recurrent neural networks
Name of your method	Contiguous-Block Temporal Withholding for Block-wise Recall Evaluation in Ordered EEG Time-Series
Name and reference of original method	NA
Resource availability	The method is implemented in Python using TensorFlow/Keras. The repository will include: •Preprocessed contraction-aligned EEG datasets (.mat format) •Scripts for contiguous-block partitioning •Model training and evaluation scripts •CSV output templates •Environment specifications (Python and package versions) Repository link:The database of the presented cases is available at: http://u.pc.cd/vPK Code for BiLSTM analyses is available at: https://github.com/drcrane369/bilstm-contiguous-test

Background

Exercise-induced fatigue has been described as an active process of central regulation rather than a purely peripheral limitation [1]. Evidence indicates that fatigue unfolds as a temporally structured process rather than a binary pre/post condition. Prolonged effort is accompanied by progressive cortical reorganization, with large-scale network dynamics shifting under exhaustive exercise, including alterations in internal–external processing states and attentional networks [2]. Beta-band efficiency changes consistent with compensatory reorganization have also been reported during general fatigue [3]. Longitudinal intervention studies further suggest that exercise reshapes brain-level organization across time, as reflected in resistance training effects on brain aging metrics [4], reinforcing the view that neural responses to exercise are dynamically modulated rather than static.

Interhemispheric coordination constitutes a relevant dimension of this reorganization. Paradigms designed to study neural modulation induced by exercise demonstrate reductions in interhemispheric coherence under fatigue [5], and cycling resistance manipulations induce changes in functional connectivity and beta-band dynamics during later stages of exercise [6]. Cortical responses also depend on contraction type, load, and task parameters [7]. Most studies rely on group-level contrasts or predefined time windows, limiting resolution of within-subject temporal evolution during task execution.

Machine-learning methods have been applied to detect and classify exercise-related EEG changes. Yang and Ren [8] extracted multivariate features during exercise-induced fatigue and used a support-vector machine classifier to distinguish fatigue states. Resting-state EEG connectivity metrics have been associated with individual differences in fatigue tolerance; Li et al [9] showed that pre-exercise network features predicted physical fatigue resilience and performance in high-intensity tasks. These findings confirm that EEG contains structured information relevant to fatigue-related variability.

In most cases, machine learning is used to classify predefined states or predict outcomes, with emphasis on discriminative accuracy. EEG signals are characterized by high dimensionality and substantial noise, which can limit the interpretability of purely performance-driven approaches. Recent work has highlighted the need for methods capable of capturing latent structure in EEG data beyond predefined labels [10].

The present framework does not aim to optimize classification performance. Instead, it proposes recall as an operational index of hemispheric recognizability across temporally ordered data segments. In this context, “recognizability” refers to the consistency with which class-specific patterns can be identified by the model, and should not be interpreted as a direct physiological measurement. Because fatigue unfolds sequentially, evaluation based on contiguous temporal withholding preserves acquisition order and enables examination of how hemispheric recognizability evolves across the exercise bout.

In related conceptual work (Medical Hypotheses), hemispheric recognizability during sustained cycling is described using class-wise recall as a temporally indexed metric. That publication introduces the evaluation logic but did not provide procedural specifications. The motivation for the present article is to formalize that evaluation strategy as a reproducible methodological framework. The method withholds continuous segments of data while preserving their original temporal order. A fixed proportion of the ordered sequence is reserved as a continuous block and shifted across predefined positions to estimate performance under controlled temporal displacement.

The framework is designed for within-subject EEG datasets organized as sequential epochs, such as contraction-aligned recordings obtained during sustained cycling. It specifies procedures for:

• 1.Constructing temporally ordered frontal EEG datasets
• 2.Defining contiguous-block test partitions
• 3.Implementing repeated model training under identical hyperparameters
• 4.Computing and aggregating class-wise recall across block positions

Repeated training quantifies dispersion associated with stochastic optimization and separates block-related variation from repetition-related variability. The methodology complements the conceptual publication by providing implementation details and reproducibility guidelines for application to other sequential EEG paradigms.

Positioning with respect to existing temporal validation methods

The proposed contiguous-block temporal withholding strategy differs from standard temporal validation approaches such as blocked cross-validation, sliding window validation, walk-forward validation, and chronological splits. These methods are primarily designed to estimate predictive performance under temporal constraints. In contrast, the present framework is not intended to optimize or estimate generalization accuracy. Instead, it treats class-wise recall as a time-indexed observable, enabling the characterization of structured variation across contiguous segments of an ordered dataset.

In this sense, the method does not aim to identify a globally optimal predictive model. Rather, it is designed to detect how hemispheric recognizability patterns evolve across the temporal sequence, specifically in terms of differential hemispheric expression. This temporal profiling cannot be directly obtained from validation strategies, where performance is typically collapsed into a single estimate or evaluated under progressively expanding training sets. (This objective aligns with the conceptual framework proposed in the related Medical Hypotheses article.)

The proposed method systematically displaces a fixed-length contiguous test block across predefined positions while maintaining a constant test proportion, enabling direct comparison of recall values across temporal locations. In addition, repeated training under identical hyperparameters is incorporated to quantify stochastic (i.e., random) variability, allowing separation of block-dependent effects from training-induced dispersion. Under this formulation, performance is treated as a time-varying quantity rather than a single summary metric.

Method details

Experimental design and EEG acquisition

The experimental protocol follows previously reported schemes for the simultaneous acquisition of brain, muscular, and cardiac signals during exercise-induced fatigue [11]. Unlike that design, the current study emphasizes frontal cortical dynamics by implementing a higher spatial-density EEG montage focused on this region.

Four healthy adult volunteers (3 male, 1 female; mean age 25.25 ± 0.96 years) participated in a controlled cycling protocol using a Lode Corival ergometer. All participants provided written informed consent prior to data collection. The experimental procedure consisted of two sessions conducted on separate days (approximately one week apart).

Determination of individualized fatigue threshold (FT)

In the first session, each participant performed 3-minute cycling trials with incremental load increases of 30 W per trial. Pedaling cadence was maintained at 60 RPM.

Surface electromyography (sEMG) from the vastus lateralis muscle and ECG lead II were recorded using a Biopac MP36 system at 500 Hz. sEMG was filtered online (0.5–150 Hz, IIR, Q = 0.707). ECG was filtered (5–150 Hz, IIR, Q = 0.707) with an additional 60 Hz notch filter.

The fatigue threshold (FT) was defined as the highest sustainable load prior to a reduction in sEMG power. If the participant could not continue or heart rate reached 180 BPM without a clear power drop, FT was defined as the intermediate load between the final two trials.

Data from this phase were used exclusively to determine individualized load.

Fatigue protocol at individualized load

In the second session, participants performed repeated 5-minute cycling trials at their individualized FT. Rest intervals between trials were fixed at 2 min. The session ended when the participant could no longer complete a trial or heart rate reached 180 BPM.

During this phase, EEG and sEMG signals were recorded simultaneously.

EEG was acquired using a wireless TMSi MOBITA system at 500 Hz. Twenty frontal electrodes were positioned according to the 10–10 system at the following locations:

Fp1, Fp2, AFz, F7, F5, F3, F1, Fz, F2, F4, F6, F8, FT7, FC5, FC3, FC1, FC2, FC4, FC6, FT8.

Signals were filtered online (5–100 Hz, IIR, Q = 0.707) with a 60 Hz notch filter. AFz was used as reference during acquisition.

An external pulse generator synchronized EEG, sEMG, and ECG systems. Pulses marked the start and end of each pedaling trial.

Signal segmentation and dataset construction

Surface EMG signals recorded during the fatigue protocol were used to define contraction-aligned epochs. Muscle burst onset and offset were detected using a nonlinear burst detection procedure applied to the sEMG signal.

Each detected contraction defined one epoch. At a cadence of 60 RPM, approximately 300 contractions were obtained per 5-minute trial, although the final number of usable epochs varied per participant after preprocessing and artifact exclusion. Fatigue during the trials was verified through the detection of sEMG power progressive reduction.

For each contraction epoch, frontal EEG data were extracted synchronously from all available channels. For hemispheric analyses, AFz and Fz were excluded, resulting in 18 frontal channels (9 left hemisphere, 9 right hemisphere).

For each participant, EEG data were organized as within-subject spatiotemporal samples. Each epoch was represented as a multichannel temporal segment of size:

$$\left(\text{number}\text{of}\text{channels}=18\right)\times \left(\text{number}\text{of}\text{time}\text{samples}\text{per}\text{epoch}\right).$$

Epoch duration varied across participants depending on contraction length. No temporal rescaling or warping was applied. Within a participant, all retained epochs were zero-padded or cropped consistently so that sequence length remained fixed for that participant. Sequence length ranged between 164 and 212 time samples (sampling rate = 500 Hz).

Samples were preserved in acquisition order, such that contiguous indices corresponded to temporally adjacent contractions along the fatigue progression.

All analyses were conducted independently for each participant. No cross-subject pooling was performed. The number of pedaling sessions and total contraction epochs per participant are summarized in Table 1. Note that the signals used to construct the dataset with the total 14,500 examples correspond to EEG after segmentation and preprocessing. No special EEG features were extracted or used in this study.

Table 1.

Overview of the dataset.

Subject	Pedaling session	Total epochs	Total EEG (epochs × 18)
V1	2	100	1800
V2	3	100	5400
V3	3	62	3348
V4	3	40	2160

Note: One epoch corresponds to one muscle contraction detected from the sEMG signal during the fatigue protocol.

Contiguous-block temporal evaluation and recall estimation

Input representation and dataset structure

For a given participant and channel, contraction-aligned EEG epochs were represented as fixed-length univariate temporal sequences. Let:

• •$N$denotes the total number of retained epochs for that participant,
• •$T$denotes the fixed sequence length after padding/cropping,

Sequence length $T$was participant-specific and ranged from 164 to 212 time samples. Within a participant, all epochs shared identical length. Input tensors were defined as:

$$X\in {\mathbb{R}}^{N\times T\times 1}$$

where the last dimension corresponds to the univariate temporal signal.

Samples were stored in acquisition order. Index $i$corresponds to the $i$-th contraction in temporal progression. No reordering was performed prior to contiguous-block selection. Binary class labels $Y$were encoded to represent either rigjt (1) or left (0) hemispheres, using one-hot representation for hemispheric class membership. Label encoding was consistent across all repetitions and block configurations.

Contiguous-block temporal test strategy

Temporal evaluation was implemented by reserving a contiguous block of samples as test data.

Let:

• •$N$= total number of samples,
• •$p=0.20$ (test proportion),
• •$B=20$ (number of block positions).

The test block length was defined as: $\text{block}\_\text{len}=\lfloor p·N\rfloor$

Block start indices were computed as: $s_b=\text{round}\left(\frac{b\left(N-\text{block}\_\text{len}\right)}{B-1}\right),b=0,1,\dots ,B-1$

For each block $b$:

• •Test set: $X_{\text{test}}=X\left[s_b:s_b+\text{block}\_\text{len}\right]$
• •Development set: all remaining samples outside the contiguous interval.

Blocks may partially overlap depending on $N$. The block length remains constant across all positions. Test samples were not shuffled, to preserve the temporal structure required by the method.

Training, validation, and model specification

For each block configuration:

• 1.The development set was randomly permuted using NumPy’s random generator.
• 2.90 % of samples were assigned to training.
• 3.10 % were assigned to validation.

The test set was excluded from all training, validation, and model selection procedures.

A Bidirectional LSTM (Long Short-Term Memory) classifier was implemented using tf.keras, although the evaluation strategy is not restricted to this architecture. Alternative models such as multilayer perceptrons [12], 1D convolutional neural networks [13,14], or other recurrent networks could be employed. BiLSTM was selected because it captures bidirectional temporal dependencies within sequential EEG data [15], aligning with the objective of modeling progressive cortical reorganization during sustained effort.

The architecture was:

• •Bidirectional(LSTM(32, return_sequences=False))
• •Dropout (0.5)
• •Dense layer (64 units, ReLU activation)
• •Dropout (0.4)
• •Dense output layer (2 units, softmax activation)

Additional implementation details:

• •LSTM activation: tanh
• •Recurrent activation: sigmoid
• •Kernel initialization: Glorot uniform (TensorFlow default)
• •Bias initialization: zeros (default)

Training parameters:

• •Optimizer: Adam (learning rate = 0.001, clipnorm = 1.0)
• •Loss: categorical cross-entropy
• •Batch size: 32
• •Maximum epochs: 40
• •Early stopping: monitor = validation loss, patience = 8, restore best weights
• •ReduceLROnPlateau: monitor = validation loss, patience = 4, factor = 0.5

All hyperparameters described in this section (architecture, optimizer settings, training parameters, and regularization components) were kept fixed across all repetitions and block configurations.

Repeated training protocol

For each block position, training and evaluation were repeated $R=5$ times.

For each repetition:

• 1.TensorFlow backend session was cleared.
• 2.The model was re-instantiated and recompiled.
• 3.Optimizer state was reset.
• 4.Development data were reshuffled.
• 5.Training and validation split was re-applied.

No fixed random seed was enforced across repetitions. Variability across runs reflects stochastic initialization and batch-level ordering. Predictions were obtained on the contiguous test set after training convergence.

It is important to note that the proposed evaluation framework is not restricted to the Bi-LSTM architecture. Any classifier capable of operating on the given input representation can be used. The methodological contribution lies in the temporal evaluation strategy rather than in the model design, and the observed behavior is therefore expected to reflect the evaluation procedure rather than a specific architectural choice.

Recall computation

For each repetition and block:

A confusion matrix was computed.

Class-wise recall for class $c$was defined as: $\text{Recal}{\mathrm{l}}_c=\frac{T{P}_c}{T{P}_c+F{N}_c}$where:

• •$T{P}_c$= true positives for class $c$,
• •$F{N}_c$= false negatives for class $c$.

Recall was retained as the primary block-wise index because the objective was not to maximize overall classification accuracy, but to examine the extent to which each class contained learnable temporal patterns. A high recall for a given class indicates that a large proportion of its instances were correctly identified, suggesting that the model captured consistent patterns associated with either or both hemispheres. In this sense, recall serves as an indicator of pattern learnability, as it reflects whether features present in the training blocks resonate with those appearing in the corresponding test blocks.

Aggregation and output structure

For each block position, recall was computed separately for every repetition. All individual recall values were preserved. In total, five recall values per block were obtained for each full pass of the 20 contiguous block positions.

For each participant and channel, the following values were stored:

• •block_id
• •start_index
• •end_index
• •recall_class_1 (all repetitions)
• •recall_class_2 (all repetitions)

Outputs were exported in CSV format for downstream analysis. The resulting structure preserves the complete repetition-wise recall values while providing a temporally ordered recall profile across 20 contiguous block positions for each participant and channel.

Algorithm 1. contiguous-block temporal evaluation with repeated training

Input:

$X\in {\mathbb{R}}^{N\times T\times 1}$(ordered samples)

$Y$(labels)

Test proportion $p=0.20$

Number of blocks $B=20$

Repetitions $R=5$

1. Compute: $\text{block}\_\text{len}=\lfloor pN\rfloor$

2. Compute start indices: $s_b=\text{round}\left(\frac{b\left(N-\text{block}\_\text{len}\right)}{B-1}\right),b=0,\dots ,B-1$

3. For each block $b$:

  3.1 Define contiguous test set

  $X_{test}=X\left[s_b:s_b+\text{block}\_\text{len}\right]$

  3.2 Define development set as remaining samples

  3.3 For $r=1,\dots ,R$:

    (a) Clear backend session

    (b) Rebuild and compile model

    (c) Shuffle development set

    (d) Split into training and validation sets

    (e) Train with early stopping

    (f) Predict on $X_{test}$

    (g) Compute class-wise recall

  3.4 Aggregate recall across repetitions

4. Return block indices, recall values

Computational environment

All analyses were conducted in Python (version 3.8.20) using the following primary libraries: TensorFlow (version 2.11.0), NumPy (version 1.24.3), SciPy (version 1.10.1), scikit-learn (version 1.3.0), and pandas (version 2.0.3).

Model training and evaluation were executed in the Spyder IDE within a Windows 64-bit operating system environment. Computation was performed using the CPU version of TensorFlow (no GPU detected in the execution environment).

Computations were conducted on an Intel-based 64-bit architecture (Intel64 Family 6 Model 151) with 32 GB RAM.

Exact package versions and environment specifications are provided in the accompanying code repository to ensure full reproducibility.

Data and code availability

The preprocessed datasets used in this study, including contraction-aligned EEG sequences and corresponding hemispheric labels, will be made publicly available upon publication.

The repository includes:

• •Participant-specific .mat files (Subject_1.mat to Subject_4.mat) containing electrode-wise temporal segments (X) and hemisphere labels (Y_text, where 0 = Left and 1 = Right).
• •Metadata describing the number of pedaling sessions, contraction epochs per session, and total electrode-wise segments per participant.
• •The main evaluation script (Test_block.py) implementing the contiguous temporal block test strategy.
• •Full model training and evaluation code, including computation of confusion matrices and class-wise performance metrics.
• •Requirements.txt file and environment documentation specifying exact software versions for reproducibility.

Exact package versions and computational environment specifications are provided in the repository to facilitate full methodological transparency and reproducibility.

Method validation

Validation framework

Method validation was performed using the cycling dataset described above. The objective was to verify that the contiguous-block procedure produces measurable temporal variation in recall that exceeds repetition-level variability.

Validation focused on three measurable properties:

• 1.Block-wise recall variation in hemispheric recognizability across temporally shifted test segments.
• 2.Variation of interhemispheric differences derived from recall differences across blocks.

Temporal withholding sensitivity

Block-wise recall values (20 temporal positions, averaged across five repetitions per block) were used to quantify temporal modulation for each participant and hemisphere.

The following quantities were computed:

• •Minimum recall across block positions
• •Maximum recall across block positions
• •Dynamic range R = max − min
• •Mean standard deviation across repetitions (Mean SD)
• •Maximum standard deviation (Max SD)
• •Ratio R / Mean SD

Table 2 summarizes these metrics.

Table 2.

Block-wise recall dynamic range and repetition variability across participants and hemispheres.

Min and Max of average recalls were computed across the 20 contiguous temporal blocks. R = Max − Min. Mean SD and Max SD correspond to standard deviation across five repetitions per block.

Volunteer	Hemisphere	Min	Max	R	Mean SD	Max SD	R /Mean SD
V1	Left	0.48	0.73	0.25	0.042	0.073	5.95
V1	Right	0.44	0.65	0.21	0.036	0.081	5.83
V2	Left	0.58	0.94	0.36	0.035	0.086	10.28
V2	Right	0.25	0.86	0.61	0.026	0.053	23.46
V3	Left	0.39	0.70	0.31	0.055	0.096	5.64
V3	Right	0.49	0.77	0.28	0.052	0.099	5.38
V4	Left	0.59	0.78	0.19	0.028	0.057	6.78
V4	Right	0.53	0.81	0.28	0.030	0.070	9.33

Across participants and hemispheres:

• •Dynamic range R ranged from 0.19 to 0.61.
• •Mean repetition SD ranged from 0.024 to 0.066.
• •R / Mean SD ratios ranged from 3.8 to 25.4.

In all cases, dynamic range across block positions exceeded repetition-related variability. That R consistently surpasses Mean SD indicates that changes in recall are predominantly associated with the temporal position of the withheld segment rather than with instability across training runs. Therefore, recall profiles reflect sensitivity to contiguous temporal displacement, i.e., the method differentiates between block-dependent variation and repetition-level dispersion, which supports its capacity to detect structured temporal modulation rather than random fluctuations introduced by model training.

To provide a formal statistical assessment of this modulation, a Friedman test was applied across the 20 contiguous temporal blocks for each participant and hemisphere. In all cases, a significant effect of block on recall was observed (all p < 0.001), indicating that recall varies systematically across temporal segments. This result supports the presence of block-dependent modulation beyond repetition-related variability.

Interhemispheric difference

For each contiguous temporal block $b\in \left\{1,\dots ,20\right\}$and each repetition, interhemispheric difference was computed from class-wise recall as:

$${\Delta }_{\text{recall}}\left(b\right)=\text{Recal}{\mathrm{l}}_{\text{left}}\left(b\right)-\text{Recal}{\mathrm{l}}_{\text{right}}\left(b\right)$$

For each participant, the following quantities were calculated from the 20 block-wise mean differences:

• •Mean Δ across blocks
• •Minimum Δ
• •Maximum Δ
• •Δ range (Max − Min)
• •Mean standard deviation of Δ across repetitions

Δ range quantifies the amplitude of interhemispheric variation across the exercise bout, whereas Mean SD(Δ) estimates repetition-level dispersion. A sign-change indicator was also computed to determine whether Δ crossed zero at least once across block positions, reflecting shifts in hemispheric dominance over time.

Results can be found in Table 2. Across participants, Δ range exceeded Mean SDΔ in all cases, indicating that interhemispheric variation across contiguous temporal segments was larger than repetition-related variability. These results support interhemispheric differences derived from recall exhibit structured temporal modulation rather than noise-driven fluctuation.

Validation criteria

The method was considered empirically validated when the following measurable conditions were satisfied:

• 1.Block-wise hemispheric recognizability dynamic range (R) exceeds repetition-level mean SD.
• 2.Mean SD across repetitions remains bounded relative to block-wise variation.
• 3.The range of interhemispheric differences (Δ) exceeds the mean standard deviation of Δ.
• 4.Δ values vary across block positions.

All participants met these criteria.

Validation therefore demonstrates that the contiguous-block procedure produces measurable temporal modulation of recall that is larger than variability introduced by repeated random training. (Table 3)

Table 3.

Interhemispheric differences derived from recall across contiguous temporal blocks. Mean_Δ is the average Δ (Recall_left − Recall_right) across 20 blocks. Δ_min and Δ_max are the minimum and maximum block-wise mean Δ values. R_Δ is the range of Δ, defined as Δ_max − Δ_min. Mean_SD(Δ) is the mean repetition-level SD of Δ across blocks. “Sign” indicates whether Δ crossed zero across the ordered blocks; the number in parentheses is the total count of sign changes.

Vol	Mean_Δ	Δ_min	Δ_max	R_Δ	Mean_SD(Δ)	Sign
V1	0.03	−0.16	0.17	0.33	0.08	Yes (6)
V2	0.15	−0.28	0.66	0.94	0.06	Yes (3)
V3	−0.04	−0.38	0.21	0.59	0.10	Yes (6)
V4	0.06	−0.12	0.20	0.32	0.053	Yes (2)

Limitations

The following methodological constraints should be considered.

• •The procedure is strictly implemented at the within-subject level. Models are trained and evaluated independently for each participant, and no cross-subject modeling, transfer learning, or pooled analysis is performed. The framework is not intended for inter-individual comparison, ranking, or generalization assessment. Its scope is limited to characterizing temporal modulation of recall within each subject’s ordered EEG sequence.
• •The validation is conducted on a limited number of participants, as the objective is to demonstrate the behavior of the proposed method under controlled within-subject temporal conditions rather than to establish population-level generalizability. Evaluation across larger and more diverse populations is required to assess the robustness of block-dependent modulation and constitutes a direction for future work.
• •The evaluation relies on contiguous temporal segmentation. Datasets with a small number of samples (low $N$) may produce very short block lengths after applying $\lfloor 0.20N\rfloor$, which can reduce stability of block-wise estimates.
• •The method assumes that samples are temporally ordered and represent sequential acquisition. It is not applicable to datasets lacking intrinsic temporal structure or to datasets where acquisition order is arbitrary.
• •The magnitude of block-wise variation depends on consistent preprocessing. Changes in filtering parameters, referencing scheme, artifact rejection criteria, or epoch definition may alter recall profiles.
• •Validation was performed using sustained cycling EEG data. Application to other experimental paradigms with different temporal dynamics may require adjustment of block proportion, number of blocks, or repetition count.

Ethics statements

All procedures involving human participants were conducted in accordance with the ethical standards of the institutional research committee and with the principles of the World Medical Association Declaration of Helsinki. The study was approved by the institutional ethics committee appointed by the Director of the Center (approval no MTY08122022). Written informed consent was obtained from all participants prior to inclusion in the study. No identifying personal information is reported in this manuscript.

Declaration of generative AI and AI-assisted technologies in the manuscript preparation process

During the preparation of this work the authors used ChatGPT (OpenAI) to support language editing. The tool was used to improve clarity and organization of the text. After using this tool, the authors critically reviewed and edited the manuscript and take full responsibility for the content of the published article.

CRediT author statement

RCHV: Conceptualization; Methodology; Software; Data curation; Formal analysis; Validation; Visualization; Writing – original draft. DG: Conceptualization; Resources; Supervision; Project administration; Funding acquisition; Writing – review & editing. MC: Conceptualization; Methodology; Formal analysis; Validation; Supervision; Writing – review & editing.

Declaration of interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Data availability

Repositories with data and allgorithms use in this work are shared in the manuscript.