Knowledge Vault 3/22 - G.TEC BCI & Neurotechnology Spring School 2024 - Day 2
Investigation of high frequency oscillations of iEEG with computational
intelligence for the rapid localization of seizure onset zone in epilepsy
Nuri Firat Ince, Mayo Clinic (USA)
<Resume Image >

Concept Graph & Resume using Claude 3 Opus | Chat GPT4 | Llama 3:

graph LR classDef moved fill:#f9d4d4, font-weight:bold, font-size:14px; classDef modalities fill:#d4f9d4, font-weight:bold, font-size:14px; classDef domains fill:#d4d4f9, font-weight:bold, font-size:14px; classDef parkinsons fill:#f9f9d4, font-weight:bold, font-size:14px; classDef hfos fill:#f9d4f9, font-weight:bold, font-size:14px; classDef epilepsy fill:#d4f9f9, font-weight:bold, font-size:14px; classDef detection fill:#f9d4d4, font-weight:bold, font-size:14px; classDef processing fill:#d4f9d4, font-weight:bold, font-size:14px; classDef gui fill:#d4d4f9, font-weight:bold, font-size:14px; classDef collaborative fill:#f9f9d4, font-weight:bold, font-size:14px; A[Nuri Firat Ince,] --> B[Ince: UH to Mayo
for neuro research 1] A --> C[Signal modalities vary in
spatial, temporal resolution 2] A --> D[Ince's domains: Parkinson's,
epilepsy, brain tumors 3] D --> E[Parkinson's: beta LFP
activity, optimal DBS 4] D --> F[Parkinson's HFOs: subtypes,
subthalamic nucleus 5] D --> G[Epilepsy: intracranial EEG,
seizure onset zone 6] G --> H[HFOs: seizure areas,
visually challenging 7] H --> I[Automated HFO detectors needed,
artifact challenges 8] I --> J[Unsupervised pipeline clusters
HFOs vs artifacts 9] I --> M[Pseudo-HFOs from noise
misidentified as HFOs 12] H --> N[Sparse processing reconstructs
HFOs vs pseudo-HFOs 13] N --> O[Random forests improve
HFO, seizure zone accuracy 14] I --> P[GUI for HFO annotation,
AI-assisted interpretation 15] I --> Q[Adaptive local HFO
representation developed 16] G --> V[Responsive neurostimulation,
HFO-triggered 21] V --> W[Simulink: wireless EEG,
closed-loop stimulation 22] A --> Z[Multi-institution collaborations,
Neural Engineering group 25] A --> AA[Open source HFO algorithms,
animal model adaptable 26] A --> AB[Deep learning interest,
segmented signal input 27] A --> AC[SEEG main data source,
some tumor ECoG 28] A --> AD[Postdoc, PhD positions: neural
interfaces, processing, ML 29] A --> AE[Collaborative ECoG maps
sulcus, evoked potentials 30] class B moved; class C modalities; class D,E,F domains; class G,H,I,J,M,N,O,P,Q,V,W epilepsy; class Z,AA,AB,AC,AD,AE collaborative;


1.-Nuri Ince recently moved his lab from University of Houston to Mayo Clinic to research neural biomarkers in neurological disorders.

2.-Various signal modalities are available with different spatial and temporal resolutions to extract information and identify neural biomarkers.

3.-Nuri's lab operates in domains including Parkinson's disease, epilepsy and brain tumors, with a focus on seizure onset zone localization.

4.-In Parkinson's, beta band activity in local field potentials correlated with optimal deep brain stimulation contacts.

5.-High-frequency oscillations (HFOs) between 200-400 Hz distinguished Parkinson's subtypes and helped localize the subthalamic nucleus target.

6.-In epilepsy, intracranial EEG is used to identify the seizure onset zone over days of monitoring, which has risks.

7.-HFOs recognized as potential biomarkers occur frequently in seizure generating areas even before seizures, but are difficult to visually identify.

8.-Automated HFO detectors are needed but face challenges distinguishing real HFOs from artifact waveforms that appear similar when filtered.

9.-In 2016, an unsupervised HFO detection pipeline using time-frequency analysis was proposed to cluster and distinguish HFOs from artifacts.

10.-HFO clusters correlated with seizure onset zones in most patients, suggesting HFOs could predict seizure generating areas without seizures themselves.

11.-Motor cortex HFOs had arbitrary waveforms while seizure onset zone HFOs exhibited stereotyped morphology, a potential biomarker to distinguish pathological events.

12.-Realistic recordings pose challenges with locally corrupted signals generating pseudo-HFOs that can be misidentified as real HFOs.

13.-Sparse signal processing was used to reconstruct candidate HFO events with predefined waveform dictionaries to distinguish real from pseudo-HFOs.

14.-Random forest classifiers using sparse approximation features significantly improved accuracy of identifying real HFOs and the seizure onset zone.

15.-A graphical user interface was developed to annotate HFOs and incorporate the AI tool to assist neurologists' interpretations.

16.-Approximating the entire HFO signal has limitations, so an adaptive sparse local representation approach was developed.

17.-Learned waveform dictionaries were used to represent HFOs in overlapping signal segments, improving robustness to noise in EMU and intraoperative settings.

18.-Intraoperative HFO rates showed no significant difference from EMU rates, suggesting feasibility of earlier seizure onset zone localization.

19.-Denoising eliminated many fast ripple artifacts misidentified as HFOs and improved seizure onset zone localization, especially in noisy intraoperative recordings.

20.-Ranking contacts by denoised HFO rates achieved 95% accuracy localizing the seizure onset zone with the top 1-2 contacts intraoperatively.

21.-Responsive neurostimulation aims to control seizures by stimulating the seizure onset zone when HFO biomarkers are detected by an implant.

22.-A Simulink interface was established to collect intracranial EEG with a wireless brain implant and clinical amplifier for closed-loop stimulation.

23.-The implant had lower signal quality than the clinical amplifier but captured spikes well, and strategies improved wireless data transmission.

24.-The implant is being used to study neuromodulation in and out of seizures with a MATLAB/Simulink platform under development.

25.-The research involved collaborations across institutions including Baylor, Texas Children's, and Mayo Clinic within a new Neural Engineering and Bioelectronics group.

26.-Open source algorithms exist for HFO detection, including from Nuri's lab, that can be adapted for animal models as well.

27.-Deep learning approaches feeding the network segmented signals are of interest and will be explored by an incoming graduate student.

28.-Most data presented was from stereo EEG, with some ECoG collected from brain tumor patients during awake brain mapping.

29.-Postdoc and PhD positions are available in Nuri's labs focusing on neural interfacing, signal processing and machine learning for biomarker discovery.

30.-A collaborative study with Nuri's team used high-density ECoG to map the central sulcus and detect evoked potentials.

Knowledge Vault built byDavid Vivancos 2024