Code to reproduce results for the Biomag challenge on dementia screening, based on the Python library MNE for MEG/EEG data processing.
In order to use MNE, the given files need first to be converted to the Fieldtrip format.
This requires the usage of the matlab library SPM to load the data with spm_eeg_load.m, and the matlab library Fieldtrip to convert the data to Fieldtrip files with spm2fieldtrip.
No scripts could be provided here for this specific step.
MNE provides tools to convert data written in the Fieldtrip format to .fif files, which contain raw data designed for MNE.
To obtain such data, run the command:
python convert_fiedltrip_to_fif.py
BIDS stands for Brain Imaging Data Structure and is a simple way to organize neuroimaging data. The aim of BIDS is to be a standard format to make the data accessible to everyone and avoid confusion. The MNE-BIDS-pipeline is a processing pipeline for MEG and EEG data stored under the BIDS format. It includes several steps like preprocessing, epoching, evoked responses, etc.
To convert the .fif files to BIDS, run the command:
python convert_to_bids.py
The preprocessing requires the usage of MNE-BIDS-pipeline, along with a configuration file config_biomag.py.
We recommend downloading the MNE-BIDS-pipeline repository and placing it in the same folder this repository is downloaded, such that its relative position would be ../mne-bids-pipeline.
See installation instructions.
Run the following command to start the preprocessing:
python ../mne-bids-pipeline/run.py --config config_biomag.py --n_jobs 20 --steps=preprocessing
In our case, the data are resampled to 200 Hz, bandpassed between 0.1 Hz and 49 Hz, and changed into epochs of 10 seconds without overlap.
To compute the features based on the PSD (Power Spectral Density) or the covariance, run the command:
python compute_features.py
To train the model (logistic regression on PSD features and covariances) and get the prediction on the test set, run the command:
python final_model.py
It will generate a CSV file containing the subjects ID, their estimated diagnoses and the probability of the estimated diagnose computed by our model.