Knowledge Vault 3/33 - G.TEC BCI & Neurotechnology Spring School 2024 - Day 3
Towards finding an objective and data-driven approach
for defining real TMS-EEG from sham responses
Ahmadreza Keihani, University of Pittsburgh (USA)
<Resume Image >

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

graph LR classDef real fill:#f9d4d4, font-weight:bold, font-size:14px; classDef sham fill:#d4f9d4, font-weight:bold, font-size:14px; classDef model fill:#d4d4f9, font-weight:bold, font-size:14px; classDef data fill:#f9f9d4, font-weight:bold, font-size:14px; classDef future fill:#f9d4f9, font-weight:bold, font-size:14px; A[Ahmadreza Keihani] --> B[Defining real vs sham
TMS-EEG is challenging. 1] A --> C[Objective, data-driven approach to
differentiate real from sham. 2] A --> D[TMS-EEG signals reflect post-synaptic
potentials summation. 3] A --> E[TMS-EEG setup: coil, channels,
amplifiers, neuronavigation, masking. 4] B --> F[Conde 2018: real TMS-EEG
resembled sham, sparked controversy. 5] F --> G[Experts critiqued Conde's methods,
showed real TMS-EEG lateralization. 6] F --> H[Conde criticized: small coil,
common reference, inadequate masking. 7] B --> I[Recent studies dissociate real
TMS-EEG from sham. 8] I --> J[UCL, Milan labs use
advanced methods. 9] C --> K[Speaker proposes machine learning
to label real vs sham. 10] K --> L[Key comparisons: real vs
sham, within and between sessions. 11] L --> M[Within-session: real accuracy should
increase, sham stays flat. 12] L --> N[Between-session: real post-response differs
from sham, baselines similar. 13] K --> O[Tested on datasets from
speaker's lab, Milan, UCL. 14] O --> P[Milan: real, realistic sham,
auditory with/without masking, electrical. 15] O --> Q[UCL: real with/without masking,
testing model on imperfect data. 16] K --> R[Within-session results met expectations. 17] K --> S[Between-session: real post-stimulation differed
from all sham. 18] K --> T[Single-subject results confirmed group
findings, identified outliers. 19] K --> U[Model differentiated imperfectly masked
real from fully masked TMS. 20] K --> V[Limitations: motor cortex data,
prefrontal more challenging. 21] K --> W[Replicate with data from
more labs beyond Milan, UCL. 22] K --> X[More sophisticated models could
improve performance. 23] C --> Y[Four key characteristics define
real TMS-EEG response. 24] C --> Z[Objective tool to verify
TMS-EEG data validity. 25] O --> AA[Milan, UCL differences: devices,
thresholding, sample sizes. 26] E --> AB[TMS intensity: motor evoked
potentials or online EEG. 27] O --> AC[Some Milan data openly
available on bioRxiv soon. 28] B --> AD[Real TMS-EEG: lateralized, propagating
response, specific peaks, amplitudes. 29] K --> AE[Model uses single-trial, averaged
accuracy to identify real TMS-EEG. 30] class A,C,K,Z model; class B,F,G,H,I,Y,AD real; class D,E,J,O,P,Q,AA,AB,AC,AE data; class L,M,N,R,S,T,U,V,W,X future;


1.-Defining real vs sham TMS-EEG responses is challenging and has sparked debate in the field.

2.-The talk aims to find an objective, data-driven approach to differentiate real TMS-EEG signals from sham.

3.-TMS-EEG signals reflect summation of excitatory and inhibitory post-synaptic potentials from pyramidal neurons and interneurons.

4.-Key TMS-EEG setup components include the coil, EEG channels, amplifiers, neuronavigation, and noise masking.

5.-A 2018 study by Conde et al. found real TMS-EEG responses resembled sham, sparking controversy.

6.-Experts critiqued Conde's study, showing real TMS-EEG has lateralized responses under the coil that propagate.

7.-Criticisms of Conde's study included using a small coil, common average reference, and inadequate noise masking.

8.-Recent studies aim to dissociate real TMS-EEG from sham effects using carefully designed protocols.

9.-Studies from UCL and Milan labs use advanced methods to define real vs sham TMS-EEG.

10.-The speaker proposes an objective solution to label real vs sham TMS-EEG using machine learning.

11.-The approach uses key comparisons between real TMS and sham conditions within and between sessions.

12.-Within-session, real TMS accuracy should start moderate and increase, while sham stays flat.

13.-Between-session, real TMS post-response should differ from sham, while pre-stimulation baselines are similar.

14.-The method was tested on rich datasets from the speaker's lab, Milan, and UCL.

15.-Milan data had real TMS, realistic sham, auditory with/without masking, and electrical stimulation conditions.

16.-UCL data had real TMS with/without noise masking, testing the model on imperfect data.

17.-Within-session results showed real TMS accuracy increased while sham was flat, meeting expectations.

18.-Between-session results showed real TMS post-stimulation differed from all sham conditions.

19.-Single-subject results confirmed the group average findings, with the model identifying outlier subjects.

20.-The model could differentiate imperfectly masked real TMS from fully masked based on early response similarity.

21.-Current limitations are the model is based on motor cortex data, with prefrontal areas being more challenging.

22.-The approach should be replicated with data from more labs beyond Milan and UCL.

23.-More sophisticated machine learning models beyond RNNs and LSTMs could further improve performance.

24.-Four key characteristics define if TMS-EEG data represents a real response based on the model.

25.-The approach provides an objective tool to verify the validity of TMS-EEG data.

26.-Differences between Milan and UCL datasets included TMS devices, thresholding procedures, and sample sizes.

27.-Determining TMS intensity relies on measuring motor evoked potentials or online EEG responses.

28.-Some of the Milan data will be made openly available on bioRxiv within 1-2 months.

29.-Real TMS-EEG has a lateralized and propagating response with specific peaks and amplitudes.

30.-The machine learning model uses moderate single-trial accuracy and high averaged accuracy to identify real TMS-EEG.

Knowledge Vault built byDavid Vivancos 2024