Learning from label proportions for a visual matrix speller (ERP)

Learning from label proportions for a visual matrix speller (ERP) dataset from Hübner et al 2017 [1]_.

Dataset Overview

• Code: Huebner2017
• Paradigm: p300
• DOI: 10.1371/journal.pone.0175856
• Subjects: 13
• Sessions per subject: 3
• Events: Target=10002, NonTarget=10001
• Trial interval: [-0.2, 0.7] s
• Runs per session: 9
• Session IDs: session_1
• File format: BrainVision

Acquisition

• Sampling rate: 1000.0 Hz
• Number of channels: 31
• Channel types: eeg=31, misc=6
• Channel names: C3, C4, CP1, CP2, CP5, CP6, Cz, EOGvu, F10, F3, F4, F7, F8, F9, FC1, FC2, FC5, FC6, Fp1, Fp2, Fz, O1, O2, P10, P3, P4, P7, P8, P9, Pz, T7, T8, x_EMGl, x_GSR, x_Optic, x_Pulse, x_Respi
• Montage: standard_1020
• Hardware: BrainAmp DC
• Reference: nose
• Ground: FCz
• Sensor type: passive Ag/AgCl
• Line frequency: 50.0 Hz
• Impedance threshold: 20.0 kOhm
• Cap manufacturer: EasyCap
• Auxiliary channels: EOG (1 ch, vertical), pulse, respiration

Participants

• Number of subjects: 13
• Health status: healthy
• Age: mean=26.0, std=1.5
• Gender distribution: female=5, male=8
• BCI experience: mostly naive
• Species: human

Experimental Protocol

• Paradigm: p300
• Number of classes: 2
• Class labels: Target, NonTarget
• Trial duration: 25.0 s
• Study design: Visual ERP speller copy-spelling task using a 6x7 grid with learning from label proportions (LLP) classifier. Two sequences with different target/non-target ratios: sequence 1 (3 targets/8 stimuli), sequence 2 (2 targets/18 stimuli). Unsupervised calibrationless approach.
• Feedback type: visual
• Stimulus type: character matrix
• Stimulus modalities: visual
• Primary modality: visual
• Synchronicity: synchronous
• Mode: online
• Training/test split: False
• Instructions: Copy-spelling task: subjects spelled the sentence 'FRANZY JAGT IM KOMPLETT VERWAHRLOSTEN TAXI QUER DURCH FREIBURG' three times
• Stimulus presentation: soa_ms=250, stimulus_duration_ms=100, grid_size=6x7, highlighting_method=salient (brightness enhancement, rotation, enlargement, trichromatic grid overlay), viewing_distance_cm=80, screen_size_inches=24

HED Event Annotations

Schema: HED 8.4.0 | Browse: https://www.hedtags.org/hed-schema-browser

  Target
    ├─ Sensory-event
    ├─ Experimental-stimulus
    ├─ Visual-presentation
    └─ Target

  NonTarget
    ├─ Sensory-event
    ├─ Experimental-stimulus
    ├─ Visual-presentation
    └─ Non-target

Paradigm-Specific Parameters

• Detected paradigm: p300
• Number of targets: 42
• Stimulus onset asynchrony: 250.0 ms

Data Structure

• Trials: 12852
• Trials context: 68 highlighting events per character, 63 characters per sentence, 3 sentences = 68633 = 12852 EEG epochs per subject. Each epoch is a Target (10002) or NonTarget (10001) event.

Preprocessing

• Data state: raw
• Preprocessing applied: False

Signal Processing

• Classifiers: LLP (Learning from Label Proportions), shrinkage-LDA, EM-algorithm
• Feature extraction: mean amplitude per time interval
• Frequency bands: analyzed=[0.5, 8.0] Hz

Cross-Validation

• Method: 5-fold chronological cross-validation
• Folds: 5
• Evaluation type: within_subject

Performance (Original Study)

• Accuracy: 84.5%
• Auc: 0.975
• Online Spelling Accuracy: 84.5
• Post Hoc Spelling Accuracy: 95.0
• Accuracy After Rampup: 90.2
• Supervised Auc: 0.975
• Max Spelling Speed Chars Per Min: 2.4

BCI Application

• Applications: speller, communication
• Environment: laboratory
• Online feedback: True

Tags

• Pathology: Healthy
• Modality: Visual
• Type: Research

Documentation

• DOI: 10.1371/journal.pone.0175856
• License: CC-BY-4.0
• Investigators: David Hübner, Thibault Verhoeven, Konstantin Schmid, Klaus-Robert Müller, Michael Tangermann, Pieter-Jan Kindermans
• Senior author: Michael Tangermann
• Contact: david.huebner@blbt.uni-freiburg.de; michael.tangermann@blbt.uni-freiburg.de; p.kindermans@tu-berlin.de
• Institution: Albert-Ludwigs-University
• Department: Brain State Decoding Lab, Cluster of Excellence BrainLinks-BrainTools, Department of Computer Science
• Address: Freiburg, Germany
• Country: DE
• Repository: Zenodo
• Data URL: http://doi.org/10.5281/zenodo.192684
• Publication year: 2017
• Funding: BrainLinks-BrainTools Cluster of Excellence funded by the German Research Foundation (DFG), grant number EXC 1086; bwHPC initiative, grant INST 39/963-1 FUGG; European Union's Horizon 2020 research and innovation programme under the Marie Sklodowska-Curie grant agreement No 657679; Special Research Fund from Ghent University; BK21 program funded by Korean National Research Foundation grant No. 2012-005741
• Ethics approval: Ethics Committee of the University Medical Center Freiburg; Declaration of Helsinki
• Keywords: brain-computer interface, BCI, event-related potentials, ERP, P300, learning from label proportions, LLP, unsupervised learning, calibrationless, visual speller

Abstract

Using traditional approaches, a brain-computer interface (BCI) requires the collection of calibration data for new subjects prior to online use. This work introduces learning from label proportions (LLP) to the BCI community as a new unsupervised, and easy-to-implement classification approach for ERP-based BCIs. The LLP estimates the mean target and non-target responses based on known proportions of these two classes in different groups of the data. We present a visual ERP speller to meet the requirements of LLP. For evaluation, we ran simulations on artificially created data sets and conducted an online BCI study with 13 subjects performing a copy-spelling task. Theoretical considerations show that LLP is guaranteed to minimize the loss function similar to a corresponding supervised classifier. LLP performed well in simulations and in the online application, where 84.5% of characters were spelled correctly on average without prior calibration.

Methodology

The experiment used a modified visual ERP speller with a 6×7 grid. Two distinct stimulus sequences with different target/non-target ratios were used: sequence 1 had 3 targets in 8 stimuli, sequence 2 had 2 targets in 18 stimuli. Each trial consisted of 4 sequences of length 8 and 2 sequences of length 18, totaling 68 highlighting events per character. The LLP algorithm exploited these known proportions to reconstruct mean target and non-target ERP responses without requiring labeled data. The classifier was reset at the start of each sentence and retrained after each character. Subjects spelled a German pangram sentence three times. One subject (S2) had prior EEG experience; others were naive. Sessions lasted about 3 hours including setup. Participants were compensated 8 Euros per hour.

References

Hübner, D., Verhoeven, T., Schmid, K., Müller, K. R., Tangermann, M., & Kindermans, P. J. (2017) Learning from label proportions in brain-computer interfaces: Online unsupervised learning with guarantees. PLOS ONE 12(4): e0175856. https://doi.org/10.1371/journal.pone.0175856

.. versionadded:: 0.4.5 Appelhoff, S., Sanderson, M., Brooks, T., Vliet, M., Quentin, R., Holdgraf, C., Chaumon, M., Mikulan, E., Tavabi, K., Hochenberger, R., Welke, D., Brunner, C., Rockhill, A., Larson, E., Gramfort, A. and Jas, M. (2019). MNE-BIDS: Organizing electrophysiological data into the BIDS format and facilitating their analysis. Journal of Open Source Software 4: (1896). https://doi.org/10.21105/joss.01896

Pernet, C. R., Appelhoff, S., Gorgolewski, K. J., Flandin, G., Phillips, C., Delorme, A., Oostenveld, R. (2019). EEG-BIDS, an extension to the brain imaging data structure for electroencephalography. Scientific Data, 6, 103. https://doi.org/10.1038/s41597-019-0104-8

---

Generated by MOABB 1.5.0 (Mother of All BCI Benchmarks) https://github.com/NeuroTechX/moabb