Predicting LFP from EEG time series

[danpolus]

I have a dataset of simultaneous EEG (64 channels, 250Hz) and LFP (6 channels, 250Hz) recordings. I want to use the Autoformer to predict the LFP time series from EEG. I am new to transformers, and have had very limited success with this task so far. Would you be able to provide guidance for this task?
Thank you!

[Jasmine-12138]

Did you create this dataset yourself, or is it from a public dataset? Are these spontaneous EEG signals or task-related EEG signals? Can you share some information about your data?

[danpolus]

The data were collected by our group. They consist of both visual paradigm trials and a resting state continuous recording.

[Jasmine-12138]

I am exploring the deep learning method to decode the LFP from EEG signals too. But the network I used is based on CNN not autoformer and the median of Pearson coefficient achieved 0.4 across subjects and channels. I have no idea whether the performance is good enough and the model may be overfitted because of the limited data size (5 subjects, 1231 trials of working memory task with duration of 8s, 8 channels EEG and 64 channels LFP) I used to train. Therefore I am curious about how many data you can provide to train. In addition, why are your LFP only 6 channels? It’s strange because intracranial recordings usually have far more channels than six (a single electrode has had eight channels already).

[danpolus]

Very interesting! Who are your subjects? We acquire LFP from DBS sensing electrodes from PD patients, with 3 contacts in each STN. We have 40 patients, each has 6 minutes of resting state and 400 trials (2sec each) of a visual response task. Unfortunately, I haven't reached convergence in my training process yet, so I can't really tell how much data is enough. I can only provide a basic recommendation to use transfer learning.

…

On Wed, 7 May 2025 at 14:39, Jasmine-12138 ***@***.***> wrote: *Jasmine-12138* left a comment (thuml/Autoformer#232) <#232 (comment)> I am exploring the deep learning method to decode the LFP from EEG signals too. But the network I used is based on CNN not autoformer and the median of Pearson coefficient achieved 0.4 across subjects and channels. I have no idea whether the performance is good enough and the model may be overfitted because of the limited data size (5 subjects, 1231 trials of working memory task with duration of 8s, 8 channels EEG and 64 channels LFP) I used to train. Therefore I am curious about how many data you can provide to train. In addition, why are your LFP only 6 channels? It’s strange because intracranial recordings usually have far more channels than six (a single electrode has had eight channels already). — Reply to this email directly, view it on GitHub <#232 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AGYK2NVFSYDY25ZNJ7RRFGT25HWFZAVCNFSM6AAAAAB3ZBWEA6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDQNJYGI2TCNBUHA> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[Jasmine-12138]

That‘s really something! Simutaneous recordings of EEG and LFP in human are hard-won data. The public dataset was recorded from patients with epilepsy and intracranial EEG was recorded using microeletrodes. I roughly learned about Aotutransform and found that the prediction of sequence is made from the history of the sequnece itself, but not the other simutaneous recordings. Is that not so consistent to your task?

[danpolus]

I'm not familiar with Aotutransform. Did you mean Autoformer? If so, it seems that it can be adapted for my purpose. However, I don't have the know-how for it