MEG data comes from a neuroimaging technique that allows to scan the brain’s magnetic field. Multiple sensors (eg magnetometers) are placed on the human scalp and their recordings can be of major importance in neuroscience research. One can for instance infer from brain data the state of a patient that has mental disorders.
You can download the data in the following link (password 123): https://surfdrive.surf.nl/files/index.php/s/3bDWFzLx3smTNTn
Once downloaded and uncompressed, you should end up with 2 folders : “Intra” and “Cross”. The folder “Intra” contains 2 folders : train and test. The folder “Cross” contains 4 folders: train, test1, test2, and test3.
The files contained in each of those folders have the “h5” extension. In order to read them, you need to use the h5py library ( that you can install using “pip install h5py” if you don’t have it already ). This type of files can contain datasets identified by a name. For simplicity, each file contains only 1 dataset. The following code snippet can read the file ”Intra/train/rest 105923 1.h5”:
import h5py
def get_dataset_name(file_name_with_dir):
filename_without_dir = file_name_with_dir.split('/')[-1]
temp = filename_without_dir.split('_')[:-1]
dataset_name = "_".join(temp)
return dataset_name
filename_path="Intra/train/rest_105923_1.h5"
with h5py.File(filename_path, 'r') as f:
dataset_name = get_dataset_name(filename_path)
matrix = f.get(dataset_name)[()]
print(type(matrix))
print(matrix.shape)The files have the following format: “taskType subjectIdentifier number.h5” where taskType can be rest, task motor, task story math, and task working memory. In practice, these tasks correspond to the activities performed by the subjects:
The subject identifier is made of 6 numbers, and the number at the end corresponds to a chunk part. This number has no particular meaning (splitted files are easier to handle in terms of memory management). The folder “Intra” contains the files of 1 subject only. In the folder “Cross”, 2 subjects are contained in the train folder while the 3 test folders contain different subjects from the ones contained in the train folder. As seen in the section above, each file is represented by a matrix of shape 248 x 35624. The number of rows, 248, corresponds to the number of magnetometer sensors placed on the human scalp. The number of columns, 35624, corresponds to the time steps of a recording.
In brain decoding, 2 types of classifications are performed. The first one is intra-subject classification, where deep learning is used to train and test models using the same subject(s). The second type, called cross-subject classification, happens when we train a model with a set of subjects, but test the model on new, unseen subjects. In this assignment, you are asked to perform both intrasubject and cross-subject classification. The goal will be to accurately classify whether the subject is in one of the following states: rest, math, memory, motor. Tasks:
As you have seen in figure 1, the order of magnitude of this data is 10e-15, which might not be adapted for deep learning tasks. A common approach to tackle this problem is to do min-max scaling, making all the data scale to values between 0 and 1. Another common approach is Z-score normalization. More specifically, a time wise scaling/normalization is more suitable.
The machine that made the recording of this data used a sample rate of 2034 Hz, meaning that every second corresponds to 2034 samples, or data points. Therefore every file corresponds to a duration of approximately 17.5 seconds. A common approach in neuroscience research is to consider that not every samples are significant, and to perform downsampling. A major advantage of this technique is that it makes deep learning training faster, while not necessarily having a negative impact on the accuracy.
Since the train folder of the ”Cross” directory contains 64 files, it might be difficult to load everything in the memory for the training. A simple workaround is to use a loop. For instance, the first iteration would load a small subpart of all the files (eg: 8 files), to fit the model to this data. The second iteration would load the next subpart (the next 8 files), to fit it etc ...