The End Of Knowledge - Vault 3 - G.TEC BCI & Neurotechnology Spring School 2024 - Day 10

graph LR classDef main fill:#f9d4d4, font-weight:bold, font-size:14px; classDef eeg fill:#d4f9d4, font-weight:bold, font-size:14px; classDef hgm fill:#d4d4f9, font-weight:bold, font-size:14px; classDef bci fill:#f9f9d4, font-weight:bold, font-size:14px; classDef hfo fill:#f9d4f9, font-weight:bold, font-size:14px; A[Michael Jordan
Mostafa Mohammadpour] --> B[Jordan presents 3D mapping of
EEG and HFO detection. 1] A --> C[Electrode montages define signal
locations for visualizations. 2] C --> D[2D montages use schematics, 3D
use brain models. 3] C --> E[IAG Montage Creator makes
exportable 2D/3D montages. 4] E --> F[Imports FreeSurfer 3D brain
models from MRI. 5] E --> G[Imports surfaces, resections, DTI,
enables manual placement. 6] E --> H[Post-op CT for automatic
grid/depth electrode localization. 7] E --> I[Imports coordinates to semi-automatically
place electrodes. 8] C --> J[Montages used for real-time 3D
and offline analysis. 9] A --> K[Mohammadpour discusses epilepsy biomarkers:
seizure onset, spikes, HFOs. 10] K --> L[Seizure patterns: initiation, increase,
termination. Onset guides surgery. 11] K --> M[HFOs >80Hz on spikes
localize onset better. 12] K --> N[HFO detection challenges: artifacts.
True HFOs are islands. 13] K --> O[Time-domain HFO detectors filter,
threshold energy, duration, peaks. 14] K --> P[Authors detect seizure onset
zones automatically with LDA. 15] P --> Q[Good overall performance, weaker
for uncommon patterns. 16] K --> R[Detected spike propagation sequences
across electrodes in space-time. 17] R --> S[Clustered spikes by shape,
kept textbook morphologies. 18] K --> T[Spikes, HFOs localized to
onset zones off medication. 19] K --> U[Spike sequences, pathological HFOs
closest to onset zones. 20] K --> V[Real-time spike detection with
adaptive thresholds in Simulink. 21] V --> W[Comparable to offline methods,
visualized on 3D brain. 22] A --> X[Guger discusses ECoG-based BCIs
for clinical and research use. 23] class A main; class B,C,D,E,F,G,H,I,J eeg; class K,L,M,N,O,P,Q,R,S,T,U,V,W hfo; class X bci;

Resume:

1.- Michael Jordan, a skilled programmer, presents on 3D mapping of grid and stereo EEG activity and high frequency oscillation detection.

2.- Defining electrode montages representing signal locations is important for topographical visualizations, anatomical mapping, and comparisons to other modalities like fMRI.

3.- Simple 2D montages use schematic images or photos, while 3D montages use template brains or subject-specific MRI-based brain models.

4.- The IAG Montage Creator is a standalone application that creates 2D and 3D montages exportable as a single file.

5.- It imports data from FreeSurfer, a powerful free tool that reconstructs 3D brain models from MRI scans in 5-24 hours.

6.- It also imports cortex surfaces from other tools, resection areas, DTI tracks, and allows manual electrode placement aided by parcellation maps.

7.- Post-operative CT scans enable automatic localization of grid and depth electrodes, compensating for brain shift between pre- and post-operative scans.

8.- Coordinates from other systems can be imported to semi-automatically place electrodes on the 3D model.

9.- The montages are used for real-time 3D visualization in MATLAB/Simulink and offline plotting for analysis and publication.

10.- Mostafa Mohammadpour discusses epilepsy biomarkers - the seizure onset zone, irritative zone of interictal spikes, and high frequency oscillations (HFOs).

11.- Seizures show initiation, increase in frequency/amplitude, and termination patterns. The seizure onset zone is a "gold standard" for surgical resection.

12.- Spikes last 20-200ms. Pathological HFOs >80Hz riding on spikes can localize the seizure onset zone better than spikes alone.

13.- Challenges in HFO detection include artifacts. True HFOs appear as "islands" in time-frequency maps while artifacts elongate across frequencies.

14.- Existing time-domain HFO detectors bandpass filter, calculate energy, threshold, and apply duration and peak number criteria.

15.- The authors aimed to automatically detect seizure onset zones, which are usually marked subjectively by doctors after reviewing multiple seizures.

16.- They collected a dataset of several seizure onset patterns, extracted time and time-frequency features, and trained an LDA classifier.

17.- Classification performance was good overall but weaker for uncommon onset patterns like delta rhythms due to imbalanced data.

18.- Next, they detected sequences of propagating spikes across adjacent electrodes within 15-50ms and 2-5cm using an existing algorithm.

19.- To identify true spikes, they clustered detected spikes by shape using DTW distance and kept clusters resembling textbook spike morphologies.

20.- 35% of true spikes and 62% of non-spikes localized to seizure onset zones. True spikes clustered closer to onset zones.

21.- In 3 patients, they compared rates of spikes, ripples (80-250Hz), fast ripples (250-500Hz), and pathological HFOs between medicated and unmedicated states.

22.- Event rates were higher off medication as expected. Optimizing rate thresholds yielded 95% specificity for localizing seizure onset zones.

23.- Spike sequences and pathological HFOs localized closest to onset zones, outperforming isolated spikes, ripples and fast ripples.

24.- Lastly, they developed a real-time Simulink model to detect spikes using adaptive thresholds that stabilize after ~3 minutes of data.

25.- The real-time detector performed comparably to offline methods. Spikes were visualized on a 3D brain and as per-minute spike rates.

Knowledge Vault built byDavid Vivancos 2024