LFP-Analyze

Because of the data security, we just upload our code.

Introduction

This project is mainly about Local Field Potential (LFP). Thoulgh we can't publish our data, we can present its format.

The data is 3-channel time series. The sampling rate is 1kHz and the length of series is 1024 points (-500ms~523ms). However, it is only the signal 0ms~523ms that matters.

The data is devided into two part, "Target" and "Result". We have different classification tasks for different parts.

Target:

• Classify ChosenDirection (1-4)
• Classify ChosenOption (1-7)

Result:

• Classify ChosenOption (1-7)
• Classify Reward [1 3 5 9]

The data was saved as ".mat" file. The structure of data files is followed:

Igo08282012-01/

• Event.mat (include "TrialInfo.ChosenOption", "TrialInfo.ChosenDirection" and "TrialInfo.Reward")
• LFP.mat (include "LFPResultAll" and "LFPTargAll")
• Spike.mat

Igo08302012-01/

• Event.mat (include "TrialInfo.ChosenOption", "TrialInfo.ChosenDirection" and "TrialInfo.Reward")
• LFP.mat (include "LFPResultAll" and "LFPTargAll")
• Spike.mat

Requirements

Please make sure that you have already installed these toolkits.

• libsvm-weights: https://www.csie.ntu.edu.tw/~cjlin/libsvmtools/#weights_for_data_instances
• Matching Pursuit: https://github.com/supratimray/MP

Please add the folder "functions/" to your path in advance.

Prepare Data

You just need to run this part once to build the workspace.

At first, you can run "extract_and_merge_data.m" to merge the data from "Igo08282012-01/" and "Igo08302012-01/" (of course you can set another name). Then you will get a folder named "data/" which contains 3 subfolders --"0828/", "0830/" and "merge/". These are workspaces of the following operations.

Then you can run "prepare_data.m" to do Gabor Decomposition (We use Matching Pursuit to do this) on data and get spectrograms of data. Please note that when building X_tf, a.k.a spectrogram, the size of X_tf will change. Therefore DON'T ignore the scale (?Hz/pixel or ?ms/pixel). You should run it twice to get X_tfs of 'targ' and 'result'.

If you want to see whether you do previous steps correctly, you can run "retrieve_display.m" to plot the original data, the reconstructed data and the spectrogram of a trial (3 channels).

It may look like this:

Plot Average Spectrogram

After you have built the data workspace, you can plot the average spectrogram of data by running "plot_average_spectrogram.m". If you want to plot average spectrograms of different labels, you can run "plot_average_spectrogram_of_label.m".

If you run "plot_average_spectrogram.m", it may look like this:

• "Result"

• "targ"

Analyze Accuracy Matrix

After you have built the data workspace, you can analyze the accuracy matrix of data by running "analyze_acc_matrix.m" and you can also plot the accuracy matrix by running "plot_acc_matrix.m".

In "analyze_acc_matrix.m", we use libsvm-weights to classify the data (more details can be found in "analyze_acc_matrix.m"). In order to speed up the program, we use parallel computing on channels. However, it still cost a lot of time. In this step, you have two choices to balance the labels. One is giving different weights to different labels and the other one is downsampling. You can just change two boolean variables in "analyze_acc_matrix.m" to decide which method to use (more details can be found in "analyze_acc_matrix.m"). We recommend the latter.

The accuracy matrix may look like this:

• val_split = 0.5; tag = "result"; y_name = "Reward"; downsample.