L-FAME: Longitudinal Focused Attention Meditation EEG Dataset and Benchmark

arXiv:2605.22893 · eess.SP, cs.LG · Submitted 2026-05-21 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.

Rosa: We're looking at the paper: "L-FAME: Longitudinal Focused Attention Meditation EEG Dataset and Benchmark".

Dev: Here's the gist of the paper. L-FAME introduces a novel Longitudinal Focused Attention Meditation Electroencephalography (L-FAME) dataset and an accompanying benchmark to foster research into neural effects of various meditation practices over a six-week training period The complete dataset, preprocessing pipelines, and benchmark evaluation code will be publicly released, offering a valuable resource and a standardized framework for the development and comparison of new analytical methods in computational meditation research and EEG-based machine learning.

Rosa: Dataset Description The L-FAME dataset contains EEG recordings and psychological assessments from 74 healthy college participants, collected at two distinct time points: pre-intervention and postintervention Participants were randomly assigned to one of three distinct meditation groups: two mantra-based techniques (SA-TA-NA-MA and Hare Krishna) and one Breath Focus practice The dataset comprises high-resolution 64-channel EEG recordings from an initial cohort of 74 participants It includes a subset of 44 individuals who completed a six-week pre- and post-intervention protocol The data is organized and formatted in accordance with the Brain Imaging Data Structure (BIDS) standard, specifically the BIDS-EEG extension Experimental Design The study employed a structured longitudinal design involving three time points: pre-intervention, mid-intervention check-in, and post-intervention EEG sessions conducted before and after the 6-week intervention included a sequence of resting and meditation tasks During both the baseline and post-intervention EEG sessions, participants completed a standardized sequence of five tasks The protocol comprised a two-minute resting state with eyes open (restOE), followed by a four-minute resting state with eyes closed (restCE01) during which participants were instructed to allow their mind to wander Subsequently, participants engaged in an eight-minute active meditation (Medita) Following a second four-minute resting state (restCE02) utilizing identical instructions, the session concluded with an eight-minute silent meditation (slMedita) Benchmark Suite The authors propose a benchmark suite comprising three distinct classification tasks: (1) cognitive state decoding to distinguish between resting and meditation states, (2) fine-grained classification of the specific meditation techniques, and (3) cross-session adaptation to evaluate model generalization across the longitudinal time gap Task 3 evaluates model robustness across a six-week longitudinal gap, addressing signal nonstationarity and temporal concept drift Key Findings on Model Performance Across all tasks, a consistent pattern emerges where current models achieve high intra-subject performance but degrade significantly when generalizing across subjects or time Task 1 intra-subject AUC exceeds 97%, while cross-subject performance plateaus between 66% and 70% Deep learning architectures demonstrate exceptional proficiency in the intra-subject evaluation paradigm, with EEGNet achieving a peak AUC of 99.2% ± 1.2%, closely followed by EEGConformer at 97.2% ± 1.4% Task 3 indicates that longitudinal drift is substantial but can be corrected using as few as 30 calibration samples through full fine-tuning Neurophysiological Insights Representational geometry analyses demonstrate that six weeks of training initiates the differentiation of technique-specific neural representations Specifically, the similarity margin between within-class and between-class representations expands from 0.30 to 1.10 (Appendix D.2) An informative asymmetry emerges where Breath Focus, which relies on somatic attention, generates more separable cross-subject representations than the two mantra-based techniques This observation implies that the underlying cognitive substrate (somatic versus phonological) is potentially more determinative of cross-subject decoding performance than the specific content of the mantra Open Problems and Future Directions The L-FAME benchmark highlights three concrete open problems for the research community First, the persistent performance gap between intra-subject and cross-subject evaluations motivates domain generalization and transfer learning methods Second, the near-chance performance in the cross-subject classification of techniques indicates that general-purpose architectures lack inductive biases required to capture technique signatures that remain invariant across individuals Third, Task 3 demonstrates that models can transfer over time despite the expected longitudinal EEG drift The L-FAME benchmark serves as a standardized resource for the investigation of training-induced neuroplasticity and the personalized monitoring of meditation The authors introduced L-FAME, a novel 6-week longitudinal EEG dataset and benchmark designed to evaluate neural decoding across three focused attention meditation practices Our extensive baselines reveal a critical insight: while deep learning models achieve exceptional intra-subject accuracy, they struggle significantly with cross-subject generalization due to inherent inter-individual variability Furthermore, we demonstrated that longitudinal temporal drift is substantial but can be effectively mitigated via minimal few-shot calibration Limitations Reliance on self-reported practice adherence, sparse temporal sampling (two discrete time points), the lack of a control group, and a restricted college-aged demographic limit the broader clinical generalizability of the findings We plan to continue data collection in future work to supplement the dataset with appropriate control cohorts Acknowledgments We would like to acknowledge Devin O’Rourke and Sidharth Chhabra from The Harmony Collective, Ypsilanti, Michigan for their expert guidance in meditation training We also extend our gratitude to Michigan State University students Ab Basit Rafi Syed, Pratham Pradhan, Annie Wozniak, Vu Song Thuy Nguyen, Genevieve Orlewicz, and Alisia Coipel for their valuable assistance with data collection Ethics Statement All experimental procedures were approved by the Institutional Review Board (IRB) of Michigan State University All participants provided written informed consent prior to the commencement of the study The privacy rights of all human subjects have been observed throughout the research Appendix Documentation The appendix provides supplementary material organized into four sections, including comprehensive dataset documentation adhering to the Datasheets for Datasets framework This section presents detailed participant demographics (Section A.1), attrition analysis with MCAR validation (Section A.2), and the complete BIDS-compliant dataset structure alongside its processing derivatives (Section A.3) The appendix details the extended experimental paradigms, including criteria for participant inclusion and exclusion (Section B.1), three meditation techniques and the corresponding training protocols (Section B.2), procedures for EEG sessions (Section B.3), and outcomes of the psychometric questionnaires (Section B.4) The third section outlines the methodology and the data processing pipeline in detail, encompassing parameters for EEG preprocessing (Section C.1), configurations for model training (Section C.2), protocols for benchmark tasks (Section C.3), and details regarding hardware and software implementations The fourth section reports additional experimental results for all three benchmark tasks, featuring ablations of the evaluation protocols (Section D.1), longitudinal representational analyses (Section D.2), and learning curves for few-shot adaptation (Section D.3) Data Structure Details The raw EEG data are provided in BrainVision format, with all supplementary information, such as electrode locations and event markers, complying with the BIDS guidelines Furthermore, the continuous data are pre-segmented into five tasks (and saved in the EEGLAB format To ensure signal quality for machine learning, we implemented a preprocessing pipeline focusing on automated artifact suppression and signal retention Preprocessing Pipeline For data in cleaned EEG in Section 3.3, the High-pass conditioning utilized a 1 Hz zero-phase Butterworth filter Zapline-plus removed 60 Hz interference via spectral integration without distorting neural oscillations ASR was configured with a 25-standard-deviation burst detection threshold and 0.2 maximum bad-channel tolerance Channels flagged by ASR were interpolated using spherical splines to maintain the 64-channel manifold Model Training Details For intra-subject evaluations, the data are partitioned into 20-second blocks, alternating between training and testing sets in an 80%/20% ratio Unlike standard chronological splitting, this block-wise strategy is utilized exclusively in the intra-subject setting to account for non-stationary temporal dynamics throughout the session To prevent data leakage between adjacent blocks, 4-second safety gaps are inserted (Figure 10) Benchmark Task Methodology For Task 2, a stratified intersubject cross-validation approach is the only viable evaluation strategy For Task 3, we evaluate the model performance under zero-shot, 10-shot (consisting of 8.5 seconds of data for both resting and meditation states), and 30-shot (consisting of 18.5 seconds for each state) configurations The full post-session data from the intra-blockwise training of Task 1 is utilized as the performance ceiling Full fine-tuning consistently outperforms linear probing across all models and shot counts, and both strategies substantially exceed the zero-shot baseline even at 10 shots Results on Technique Separability The results reveal a striking representational shift where the BF group becomes clearly isolated into a highly distinct cluster in the post-intervention latent space The similarity between the HK and SA centroids decreases to 0.250, reflecting a measurable divergence despite their shared cognitive foundation The BF centroid becomes structurally independent, exhibiting near-orthogonal or inverse relationships with HK at −0.055 and SA at −0.482 This stark quantitative shift from high positive correlation to orthogonal or negative similarity rigorously confirms that the somatic-directed BF technique cultivates a profoundly distinct neural signature compared to the inner-speech modalities following structured practice Final Assessment The L-FAME benchmark highlights three concrete open problems for the research community First, the persistent performance gap between intra-subject and cross-subject evaluations motivates the development of domain generalization and transfer learning methods Second, the near-chance performance in the cross-subject classification of techniques indicates that general-purpose architectures lack inductive biases required to capture technique signatures that remain invariant across individuals Third, Task 3 demonstrates that models can transfer over time despite the expected longitudinal EEG drift The L-FAME benchmark serves as a standardized resource for the investigation of training-induced neuroplasticity and the personalized monitoring of meditation We introduced L-FAME, a novel 6-week longitudinal EEG dataset and benchmark designed to evaluate neural decoding across three focused attention meditation practices Our extensive baselines reveal a critical insight: while deep learning models achieve exceptional intra-subject accuracy, they struggle significantly with cross-subject generalization due to inherent inter-individual variability Furthermore, we demonstrated that longitudinal temporal drift is substantial but can be effectively mitigated via minimal few-shot calibration Dataset Documentation A.1 Detailed Participant Demographics shows that the three groups were completely balanced at the pre-test phase Consistent Attrition and Preserved Post-Test Balance indicates that this attrition was non-selective and did not introduce demographic bias A.2 Attrition Cohorts Analysis and MCAR Validation demonstrates robust baseline parity, confirming that neither biological factors nor the specific nature of the assigned meditation protocol (HK, SA, or BF) served as a primary catalyst for dropout Data Structure Details The raw EEG recordings are provided in BrainVision format, with all supplementary information, such as electrode locations and event markers, complying with the BIDS guidelines Furthermore, the continuous data are pre-segmented into five tasks (and saved in the EEGLAB format To ensure signal quality for machine learning, we implemented a preprocessing pipeline focusing on automated artifact suppression and signal retention Preprocessing Pipeline For data in cleaned EEG in Section 3.3, the High-pass conditioning utilized a 1 Hz zero-phase Butterworth filter Zapline-plus removed 60 Hz interference via spectral integration without distorting neural oscillations ASR was configured with a 25-standard-deviation burst detection threshold and 0.

Title and authors: Rosa: So we're looking at this paper called L-FAME, which is the Longitudinal Focused Attention Meditation EEG Dataset and Benchmark. It’s designed to test how neural activity changes over six weeks when people are training in different meditation practices.

Dev: Yeah, the authors are Li, Syed, Alzweri, Liu, Cohen and Ravishankar from Michigan State University. They set up a whole dataset and a benchmark for this kind of research.

Rosa: What does this actually mean for us? It means they’ve created a way to look at how meditation affects the brain over time in seventy-four college students <ref:2605.22893#pg1>.

Dev: It gives researchers a standardized framework to compare different ways they analyze meditation data using machine learning and EEG.

The paper's summary: Rosa: So, what’s the core idea behind L-FAME? Basically, they collected EEG recordings and psychological assessments from seventy-four healthy college participants at two different times: before and after a six-week meditation training period <ref:2605.22893#pg1,EEG recordings and psychological assessments from 74 healthy college participants>.

Dev: They assigned these participants to three groups: two mantra-based techniques like SA-TA-NA-MA and Hare Krishna, and one Breath Focus practice. The goal is to see the neural effects of these different practices over that time span.

Rosa: The dataset includes high resolution sixty-four channel EEG recordings, organized in a BIDS format, which makes it pretty standard for other research groups to use.

Dev: They even have a subset of forty-four people who finished the full six-week pre and post intervention protocol. It’s structured longitudinally from the start.

The paper's improvements: Rosa: One thing they focus on is improving how models handle those subtle, purely internally driven temporal shifts that happen during training. They want to see if a model can track those changes over time.

Dev: They propose a benchmark with three classification tasks. First, decoding the cognitive state—resting versus meditation. Second, classifying the specific technique they practiced. And third, testing how well a model adapts across those six weeks of time gap between sessions.

Rosa: That third task is really important because it tests if a model will generalize well over that long period and handle things like signal nonstationarity and temporal concept drift.

Dev: They found that current models do pretty well when they are only looking at one person, but their performance drops off significantly when they try to apply those models to different people or across time points.

Conclusion: Rosa: So, to wrap up the L-FAME paper, it shows that while deep learning models can be very accurate for a single subject in decoding meditation states, they struggle a lot with generalizing across different subjects or over time.

Dev: They also found that the longitudinal drift is substantial—the signal changes a lot over six weeks—but they showed you can correct that drift using as few as thirty calibration samples through full fine-tuning.

Rosa: That’s a pretty practical point for building systems that might be used to monitor meditation or health in the future, because you don't need retraining every time.

Dev: I agree, and this whole L-FAME study sets up a standard way for researchers to look at training-induced neuroplasticity and how we can personalize monitoring.

Taro: From my side, it’s interesting that they showed Breath Focus generated more separable neural representations between subjects than the mantra techniques, suggesting somatic attention is a key driver here.

Angqi Li, Ab Basit Rafi Syed, Hamzeh Alzweri, Taosheng Liu, Barry H. Cohen, Saiprasad Ravishankar

Michigan State University · New York University

eess.SP, cs.LG

Submitted: 2026-05-21

Updated: 2026-11-07

Comments: Accepted at NeurIPS 2026 (Evaluations & Datasets Track). Code and dataset available at: https://huggingface.co/datasets/L-FAME-Dataset-Benchmark/L-FAME

License: http://creativecommons.org/licenses/by-nc-sa/4.0/

Importance score: 82/100

The gist: L-FAME introduces a novel Longitudinal Focused Attention Meditation Electroencephalography (L-FAME) dataset and an accompanying benchmark to foster research into neural effects of various meditation…

Key concepts

Longitudinal Focused Attention Meditation EEG (L-FAME)
This is a novel dataset combining high-resolution 64-channel EEG recordings with psychological assessments from college students over six weeks. It tracks neural activity before and after training in three different meditation styles to study how brain activity changes during practice.
Intra-subject vs. Cross-subject Performance Gap
This refers to the difference in model accuracy when testing on data from the same person versus data from different people. Models perform very well within one individual but drop significantly when tested on new subjects, highlighting a major challenge in applying AI to personalized brain research.
Longitudinal Temporal Drift
This describes how brain signals change over time, even within the same person's training period. The study found this drift is substantial but can be corrected by fine-tuning models with just 30 calibration samples, showing that models can adapt to these changes.
Technique-Specific Neural Representations
This concept shows that different meditation techniques create distinct patterns in brain activity. Breath Focus, for example, developed a unique neural signature compared to mantra techniques, suggesting the type of attention used (somatic vs. phonological) is important.

Terminology

Summary

L-FAME introduces a novel Longitudinal Focused Attention Meditation Electroencephalography (L-FAME) dataset and an accompanying benchmark to foster research into neural effects of various meditation practices over a six-week training period The complete dataset, preprocessing pipelines, and benchmark evaluation code will be publicly released, offering a valuable resource and a standardized framework for the development and comparison of new analytical methods in computational meditation research and EEG-based machine learning. p. 0

Where the paper says it

We provide comprehensive baseline results for these tasks utilizing a range of classical machine learning algorithms and deep learning architectures. The complete dataset, preprocessing pipelines, and benchmark evaluation code will be publicly released, offering a valuable resource and a standardized framework for the development and comparison of new analytical methods in computational meditation research and EEG-based machine learning.

Dataset Description

The L-FAME dataset contains EEG recordings and psychological assessments from 74 healthy college participants, collected at two distinct time points: pre-intervention and postintervention Participants were randomly assigned to one of three distinct meditation groups: two mantra-based techniques (SA-TA-NA-MA and Hare Krishna) and one Breath Focus practice The dataset comprises high-resolution 64-channel EEG recordings from an initial cohort of 74 participants It includes a subset of 44 individuals who completed a six-week pre- and post-intervention protocol The data is organized and formatted in accordance with the Brain Imaging Data Structure (BIDS) standard, specifically the BIDS-EEG extension

Experimental Design

The study employed a structured longitudinal design involving three time points: pre-intervention, mid-intervention check-in, and post-intervention EEG sessions conducted before and after the 6-week intervention included a sequence of resting and meditation tasks During both the baseline and post-intervention EEG sessions, participants completed a standardized sequence of five tasks The protocol comprised a two-minute resting state with eyes open (restOE), followed by a four-minute resting state with eyes closed (restCE01) during which participants were instructed to allow their mind to wander Subsequently, participants engaged in an eight-minute active meditation (Medita) Following a second four-minute resting state (restCE02) utilizing identical instructions, the session concluded with an eight-minute silent meditation (slMedita)

Benchmark Suite

The authors propose a benchmark suite comprising three distinct classification tasks: (1) cognitive state decoding to distinguish between resting and meditation states, (2) fine-grained classification of the specific meditation techniques, and (3) cross-session adaptation to evaluate model generalization across the longitudinal time gap Task 3 evaluates model robustness across a six-week longitudinal gap, addressing signal nonstationarity and temporal concept drift

Key Findings on Model Performance

Across all tasks, a consistent pattern emerges where current models achieve high intra-subject performance but degrade significantly when generalizing across subjects or time Task 1 intra-subject AUC exceeds 97%, while cross-subject performance plateaus between 66% and 70% Deep learning architectures demonstrate exceptional proficiency in the intra-subject evaluation paradigm, with EEGNet achieving a peak AUC of 99.2% ± 1.2%, closely followed by EEGConformer at 97. p. 7

Where the paper says it

Deep learning architectures demonstrate exceptional proficiency in the intra-subject evaluation paradigm, with EEGNet achieving a peak AUC of 99. 2%, closely followed by EEGConformer at 97.

2% ± 1.4% Task 3 indicates that longitudinal drift is substantial but can be corrected using as few as 30 calibration samples through full fine-tuning

Neurophysiological Insights

Representational geometry analyses demonstrate that six weeks of training initiates the differentiation of technique-specific neural representations Specifically, the similarity margin between within-class and between-class representations expands from 0.30 to 1.10 (Appendix D. p. 2

Where the paper says it

Collected demographic data included age, sex, and handedness. A comprehensive summary of the demographics of the participants is provided in Table 2 and the detailed participant demographics is in Appendix A.

2) An informative asymmetry emerges where Breath Focus, which relies on somatic attention, generates more separable cross-subject representations than the two mantra-based techniques This observation implies that the underlying cognitive substrate (somatic versus phonological) is potentially more determinative of cross-subject decoding performance than the specific content of the mantra

Open Problems and Future Directions

The L-FAME benchmark highlights three concrete open problems for the research community First, the persistent performance gap between intra-subject and cross-subject evaluations motivates domain generalization and transfer learning methods Second, the near-chance performance in the cross-subject classification of techniques indicates that general-purpose architectures lack inductive biases required to capture technique signatures that remain invariant across individuals Third, Task 3 demonstrates that models can transfer over time despite the expected longitudinal EEG drift The L-FAME benchmark serves as a standardized resource for the investigation of training-induced neuroplasticity and the personalized monitoring of meditation

Conclusion

The authors introduced L-FAME, a novel 6-week longitudinal EEG dataset and benchmark designed to evaluate neural decoding across three focused attention meditation practices Our extensive baselines reveal a critical insight: while deep learning models achieve exceptional intra-subject accuracy, they struggle significantly with cross-subject generalization due to inherent inter-individual variability Furthermore, we demonstrated that longitudinal temporal drift is substantial but can be effectively mitigated via minimal few-shot calibration

Limitations

Reliance on self-reported practice adherence, sparse temporal sampling (two discrete time points), the lack of a control group, and a restricted college-aged demographic limit the broader clinical generalizability of the findings We plan to continue data collection in future work to supplement the dataset with appropriate control cohorts

Acknowledgments

We would like to acknowledge Devin O’Rourke and Sidharth Chhabra from The Harmony Collective, Ypsilanti, Michigan for their expert guidance in meditation training We also extend our gratitude to Michigan State University students Ab Basit Rafi Syed, Pratham Pradhan, Annie Wozniak, Vu Song Thuy Nguyen, Genevieve Orlewicz, and Alisia Coipel for their valuable assistance with data collection

Ethics Statement

All experimental procedures were approved by the Institutional Review Board (IRB) of Michigan State University All participants provided written informed consent prior to the commencement of the study The privacy rights of all human subjects have been observed throughout the research

Appendix Documentation

The appendix provides supplementary material organized into four sections, including comprehensive dataset documentation adhering to the Datasheets for Datasets framework This section presents detailed participant demographics (Section A.1), attrition analysis with MCAR validation (Section A. p. 9

Where the paper says it

Furthermore, we demonstrated that longitudinal temporal drift is substantial but can be effectively mitigated via minimal few-shot calibration. Limitations & Future Work: Although attrition reduced the longitudinal cohort to 44 participants, statistical evaluations confirm that dropouts occurred completely at random (MCAR), thereby preserving the overall integrity of the sample (see Appendix A.

2), and the complete BIDS-compliant dataset structure alongside its processing derivatives (Section A. p. 4
Where the paper says it

Following the BIDS guidelines, each data file is accompanied by a corresponding metadata file in JSON format. Table 3: Overview of dataset structure and processing derivatives.

3) The appendix details the extended experimental paradigms, including criteria for participant inclusion and exclusion (Section B. p. 3
Where the paper says it

1 Longitudinal Design Screening. Potential adult novice participants were recruited primarily from Michigan State University (MSU) campus and underwent an initial screening process based on specific inclusion and exclusion criteria (see Section B.

1), three meditation techniques and the corresponding training protocols (Section B. p. 2
Where the paper says it

[19] MBSR 11 25-45 6-week intervention L-FAME (Ours) Focused Attention† 74 18–32 6-week intervention ∗ Only a subset of 30 participants underwent the meditation training. † Three different FAM techniques.

2), procedures for EEG sessions (Section B. p. 4
Where the paper says it

During this session, they performed the same tasks as in the baseline, including the EEG data recording detailed in Section 3. 2 EEG Session Tasks During both the baseline and post-intervention EEG sessions, participants completed a standardized sequence of five tasks.

3), and outcomes of the psychometric questionnaires (Section B. p. 3
Where the paper says it

Finally, the participants completed all baseline data collection, including an EEG recording session detailed in Section 3. 2 and questionnaires (see Appendix B.

4) The third section outlines the methodology and the data processing pipeline in detail, encompassing parameters for EEG preprocessing (Section C. p. 4
Where the paper says it

Meditation recordings were partitioned into five segments defined in Section 3. To ensure signal quality for machine learning, we implemented a preprocessing pipeline focusing on automated artifact suppression and signal retention.

1), configurations for model training (Section C. p. 4
Where the paper says it

Within one week following the completion of the 6-week training period, participants returned for a post-intervention session. During this session, they performed the same tasks as in the baseline, including the EEG data recording detailed in Section 3.

2), protocols for benchmark tasks (Section C. p. 5
Where the paper says it

Training Data. This subset establishes benchmark tasks’ training and testing data for ML/DL following established preprocessing protocols [33, 34].

3), and details regarding hardware and software implementations The fourth section reports additional experimental results for all three benchmark tasks, featuring ablations of the evaluation protocols (Section D.1), longitudinal representational analyses (Section D. p. 8
Where the paper says it

Appendices D. 3 provide detailed subgroup analyses, representational geometry experiments, and adaptation ablations for each task.

2), and learning curves for few-shot adaptation (Section D. p. 7
Where the paper says it

Table 7: Task 3: Cross-Session Generalization and Longitudinal Adaptation. Note that the k-shot calibration performance reported in this table corresponds to the full fine-tuning strategy (updating all layers).

3)

Data Structure Details

The raw EEG data are provided in BrainVision format, with all supplementary information, such as electrode locations and event markers, complying with the BIDS guidelines Furthermore, the continuous data are pre-segmented into five tasks (and saved in the EEGLAB format To ensure signal quality for machine learning, we implemented a preprocessing pipeline focusing on automated artifact suppression and signal retention

Preprocessing Pipeline

For data in cleaned EEG in Section 3.3, the High-pass conditioning utilized a 1 Hz zero-phase Butterworth filter Zapline-plus removed 60 Hz interference via spectral integration without distorting neural oscillations ASR was configured with a 25-standard-deviation burst detection threshold and 0. p. 4

Where the paper says it

This included zero-phase high-pass filtering [29], Zapline-plus line-noise removal [30], and Artifact Subspace Reconstruction (ASR) [31]. Neural activity was isolated via Independent Component Analysis (ICA) with ICLabel classification [32].

p. 4
Where the paper says it

This included zero-phase high-pass filtering [29], Zapline-plus line-noise removal [30], and Artifact Subspace Reconstruction (ASR) [31]. Neural activity was isolated via Independent Component Analysis (ICA) with ICLabel classification [32].

2 maximum bad-channel tolerance Channels flagged by ASR were interpolated using spherical splines to maintain the 64-channel manifold

Model Training Details

For intra-subject evaluations, the data are partitioned into 20-second blocks, alternating between training and testing sets in an 80%/20% ratio Unlike standard chronological splitting, this block-wise strategy is utilized exclusively in the intra-subject setting to account for non-stationary temporal dynamics throughout the session To prevent data leakage between adjacent blocks, 4-second safety gaps are inserted (Figure 10)

Benchmark Task Methodology

For Task 2, a stratified intersubject cross-validation approach is the only viable evaluation strategy For Task 3, we evaluate the model performance under zero-shot, 10-shot (consisting of 8.5 seconds of data for both resting and meditation states), and 30-shot (consisting of 18. p. 0

Where the paper says it

Participants were randomly assigned to one of three distinct meditation groups: two mantra-based techniques (SA-TA-NA-MA and Hare Krishna) and one Breath Focus practice. Leveraging this unique longitudinal and comparative dataset, we propose a benchmark suite comprising three distinct classification tasks: (1) cognitive state decoding to distinguish between resting and meditation states, (2) fine-grained classification of the specific meditation techniques, and (3) cross-session adaptation to evaluate model generalization across the longitudinal time gap.

5 seconds for each state) configurations The full post-session data from the intra-blockwise training of Task 1 is utilized as the performance ceiling Full fine-tuning consistently outperforms linear probing across all models and shot counts, and both strategies substantially exceed the zero-shot baseline even at 10 shots

Results on Technique Separability

The results reveal a striking representational shift where the BF group becomes clearly isolated into a highly distinct cluster in the post-intervention latent space The similarity between the HK and SA centroids decreases to 0.250, reflecting a measurable divergence despite their shared cognitive foundation The BF centroid becomes structurally independent, exhibiting near-orthogonal or inverse relationships with HK at −0.055 and SA at −0.482 This stark quantitative shift from high positive correlation to orthogonal or negative similarity rigorously confirms that the somatic-directed BF technique cultivates a profoundly distinct neural signature compared to the inner-speech modalities following structured practice

Final Assessment

The L-FAME benchmark highlights three concrete open problems for the research community First, the persistent performance gap between intra-subject and cross-subject evaluations motivates the development of domain generalization and transfer learning methods Second, the near-chance performance in the cross-subject classification of techniques indicates that general-purpose architectures lack inductive biases required to capture technique signatures that remain invariant across individuals Third, Task 3 demonstrates that models can transfer over time despite the expected longitudinal EEG drift The L-FAME benchmark serves as a standardized resource for the investigation of training-induced neuroplasticity and the personalized monitoring of meditation

We introduced L-FAME, a novel 6-week longitudinal EEG dataset and benchmark designed to evaluate neural decoding across three focused attention meditation practices Our extensive baselines reveal a critical insight: while deep learning models achieve exceptional intra-subject accuracy, they struggle significantly with cross-subject generalization due to inherent inter-individual variability Furthermore, we demonstrated that longitudinal temporal drift is substantial but can be effectively mitigated via minimal few-shot calibration

Dataset Documentation

A.1 Detailed Participant Demographics shows that the three groups were completely balanced at the pre-test phase Consistent Attrition and Preserved Post-Test Balance indicates that this attrition was non-selective and did not introduce demographic bias A.2 Attrition Cohorts Analysis and MCAR Validation demonstrates robust baseline parity, confirming that neither biological factors nor the specific nature of the assigned meditation protocol (HK, SA, or BF) served as a primary catalyst for dropout

The raw EEG recordings are provided in BrainVision format, with all supplementary information, such as electrode locations and event markers, complying with the BIDS guidelines Furthermore, the continuous data are pre-segmented into five tasks (and saved in the EEGLAB format To ensure signal quality for machine learning, we implemented a preprocessing pipeline focusing on automated artifact suppression and signal retention

For data in cleaned EEG in Section 3.3, the High-pass conditioning utilized a 1 Hz zero-phase Butterworth filter Zapline-plus removed 60 Hz interference via spectral integration without distorting neural oscillations ASR was configured with a 25-standard-deviation burst detection threshold and 0.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Develop domain generalization methods for continuous cognitive states that are specifically tailored to non-stationary neural drifts, addressing non-stationary temporal shifts by employing transfer learning strategies like zero-shot and few-shot calibration.

  2. Enhance model robustness against inter-individual variability by integrating task priors derived from the specific meditation modality; this can lead to the development of meditation/cognition-specific priors to overcome the performance gap between intra-subject and cross-subject evaluations.

  3. Implement architectural constraints, such as utilizing Shallow ConvNet, to implicitly regularize deep learning models against overfitting on subject-specific artifacts, thereby enhancing cross-population generalizability.

  4. Create a framework for personalized monitoring by enabling efficient deployment of longitudinal brain-computer interfaces without complete model retraining, leveraging the finding that full fine-tuning consistently and decisively outperforms linear probing in few-shot adaptation.

  5. Develop classification methods capable of robustly isolating distinct neural signatures, specifically identifying how the somatic-directed BF technique cultivates a profoundly distinct neural signature compared to the inner-speech modalities following structured practice.

Abstract

We introduce a novel Longitudinal Focused Attention Meditation Electroencephalography (L-FAME) dataset and an accompanying benchmark, designed to foster research into the neural effects of various meditation practices and the evolution of these effects over a six-week training period. The dataset includes baseline EEG recordings from 74 healthy college participants, of whom 44 completed the six-week intervention and returned for post-intervention recordings, along with psychological assessments. Participants were randomly assigned to one of three distinct meditation groups: two mantra-based techniques (SA-TA-NA-MA and Hare Krishna) and one Breath Focus practice. Leveraging this unique longitudinal and comparative dataset, we propose a benchmark suite comprising three distinct classification tasks: (1) cognitive state decoding to distinguish between resting and meditation states, (2) fine-grained classification of the specific meditation techniques, and (3) cross-session adaptation to evaluate model generalization across the longitudinal time gap. We provide comprehensive baseline results for these tasks utilizing a range of classical machine learning algorithms and deep learning architectures. The complete dataset, preprocessing pipelines, and benchmark evaluation code are being publicly released, offering a valuable resource and a standardized framework for the development and comparison of new analytical methods in computational meditation research and EEG-based machine learning. The dataset is available at https://huggingface.co/datasets/L-FAME-Dataset-Benchmark/L-FAME

Related papers