The End Of Knowledge - Vault 3 - G.TEC BCI & Neurotechnology Spring School 2024 - Day 3

graph LR classDef niediek fill:#f9d4d4, font-weight:bold, font-size:14px; classDef artifacts fill:#d4f9d4, font-weight:bold, font-size:14px; classDef oscar fill:#d4d4f9, font-weight:bold, font-size:14px; classDef evaluation fill:#f9f9d4, font-weight:bold, font-size:14px; classDef misc fill:#f9d4f9, font-weight:bold, font-size:14px; A[Johannes Gruenwald] --> B[OSCAR removes EEG artifacts. 1] A --> C[EEG/ECoG artifacts: physiological, electrical,
transient, permanent. 2] C --> D[Pathological activity difficult to
distinguish from artifacts. 3] A --> E[OSCAR goals: eliminate artifacts, preserve
activity, recover BCI, feedback. 4] E --> F[OSCAR decomposes, whitens, removes
artifacts, reconstructs signals. 5] E --> G[OSCAR: unsupervised, adaptive, automated,
three versions. 6] G --> H[OSCAR Live: real-time, low-latency,
standard EEG. 7] G --> I[OSCAR Pro: offline, no
channel/bandwidth limits. 8] G --> J[OSCAR BCI: real-time, optimized
for ERP-based BCIs. 9] H --> K[OSCAR Live integrates with
g.tec amplifiers, recorders. 10] B --> L[Demo: OSCAR Live removes
artifacts in real-time. 11] L --> M[Demo: OSCAR Live preserves
visual evoked potentials. 12] I --> N[OSCAR Pro focuses on
offline source reconstruction. 13] G --> O[Synthetic datasets assess OSCAR
Pro's source reconstruction. 14] O --> P[OSCAR Pro visually matches
reference, preserves oscillations. 15] O --> Q[OSCAR Pro quantitatively improves
cleaned-reference similarity. 16] O --> R[OSCAR Pro evaluated on
1,500 synthetic datasets. 17] O --> S[OSCAR Pro substantially improves
R^2 in contaminated segments. 18] J --> T[OSCAR Pro/BCI evaluated on
P300 BCI datasets. 19] G --> U[BCI performance compared: contaminated,
OSCAR Pro, OSCAR BCI. 20] U --> V[OSCAR Pro/BCI improves auditory
P300 BCI accuracy. 21] U --> W[OSCAR Pro/BCI improves visual
P300 BCI accuracy. 22] J --> X[Demo: OSCAR BCI in
real-time P300 BCI game. 23] E --> Y[OSCAR addresses most ideal
artifact removal goals. 24] C --> Z[OSCAR can remove pathological
activity like epileptic spikes. 25] F --> AA[OSCAR outperforms ASR, recovers
online signals better. 26] F --> AB[OSCAR uses statistical properties,
not deep learning. 27] F --> AC[Specific artifacts can't be
excluded, but isolated. 28] H --> AD[OSCAR integrated into g.tec
hardware, not open source. 29] F --> AE[OSCAR robust to channel
counts, 16-channel 'sweet spot'. 30] class A,B niediek; class C,D,Z artifacts; class E,F,G,H,I,J,K,L,M,N,AA,AB,AC,AD,AE oscar; class O,P,Q,R,S,T,U,V,W,X evaluation; class Y misc;

Resume:

1.- Johannes Gruenwald demonstrates artifact removal in EEG data using the OSCAR method to remove movement, EMG, eye movement artifacts and more.

2.- Artifacts in EEG/ECoG can be physiological/electrical and transient/permanent, with common types being EMG, EOG, ECG, cable movement, and stimulation artifacts.

3.- Pathological activity like epileptic spikes is difficult to distinguish from artifacts based on amplitude and spectral characteristics and may get removed.

4.- OSCAR aims to eliminate artifacts, preserve brain activity, recover BCI performance, provide signal quality feedback, produce realistic EEG, and work automatically in real-time.

5.- OSCAR decomposes signals into frequency bands, buffers them, performs spatiotemporal whitening, identifies and removes artifact components, inverts whitening, and reconstructs signals.

6.- OSCAR is unsupervised, adaptive, automated, and targets specific use cases with three versions: OSCAR Live, OSCAR Pro, and OSCAR BCI.

7.- OSCAR Live provides real-time, low-latency artifact removal for standard EEG up to 64 channels, optimizing source signal reconstruction.

8.- OSCAR Pro does offline EEG/ECoG artifact removal without channel/bandwidth limits, optimizing source signal reconstruction using larger processing windows.

9.- OSCAR BCI provides real-time artifact removal optimized for ERP-based BCIs like P300 spellers, operating in the spatiotemporally whitened space.

10.- OSCAR Live is being integrated into g.tec's g.Nautilus, g.HIamp, g.USBamp amplifiers and recorders. OSCAR Pro will be in g.BSanalyze.

11.- Johannes demonstrates real-time artifact removal with OSCAR Live using a 16-channel g.Nautilus while moving and talking, showing effective corrections.

12.- An oddball visual evoked potential paradigm is demonstrated with OSCAR Live, showing consistent responses despite movement artifacts.

13.- OSCAR Pro focuses on profound offline artifact removal, aiming to reconstruct the source signal underneath artifacts.

14.- Synthetic datasets with known reference signals were created by applying artifact templates to clean EEG to assess OSCAR Pro's source reconstruction.

15.- Visually, OSCAR Pro's cleaned signals closely match the artifact-free reference signals, even preserving underlying oscillations like alpha rhythms.

16.- Quantitatively, OSCAR Pro greatly improves similarity (R^2) between cleaned and reference signals across frequency bands, especially in artifact-contaminated segments.

17.- OSCAR Pro was evaluated on 1,500 synthetic datasets with combinations of natural/provoked artifacts, contamination levels, subjects, and repeated realizations.

18.- OSCAR Pro substantially improves R^2 between cleaned and reference signals in contaminated segments across delta, theta, alpha, beta bands and broadband EEG.

19.- The impact of OSCAR Pro on BCI performance was evaluated using datasets with real artifacts from auditory and visual P300 BCIs.

20.- BCI performance is compared between contaminated, OSCAR Pro cleaned offline, and OSCAR BCI cleaned real-time signals using cross-validated classification.

21.- For auditory EPs, OSCAR Pro improved accuracy from 47% to 60%, and OSCAR BCI to 65% across 15 subjects.

22.- For visual P300, OSCAR Pro improved accuracy from 66% to 71%, and OSCAR BCI to 77% across 10 subjects.

23.- A demo video shows OSCAR BCI used in a real-time P300 BCI game called BrainBuddy developed by Martin Spüler.

24.- In summary, OSCAR versions address most of the initially defined goals for an ideal EEG artifact removal tool.

25.- OSCAR can remove pathological activity like epileptic spikes that deviate from normal EEG characteristics.

26.- OSCAR outperforms methods like ASR (artifact subspace reconstruction), recovering signals better in online scenarios.

27.- OSCAR does not use deep learning - it relies on statistical signal properties.

28.- Specific artifact types can't be excluded, but clean and original signals can be subtracted to isolate artifacts like eye blinks.

29.- The OSCAR algorithm is integrated into g.tec's hardware systems for efficiency and is not open source.

30.- OSCAR's performance is robust to different channel counts, with a 16-channel "sweet spot". Up/downsampling may introduce issues.

Knowledge Vault built byDavid Vivancos 2024