Decoding Analysis

This is an example of a decoding analysis performed on functional near-infrared spectroscopy (fNIRS) data using MNE-Python, scikit-learn, and MNE-NIRS.

Detailed information about decoding of neural signals can be found in the MNE-Python documentation. For example see Decoding (MVPA), Linear classifier on sensor data, Decoding source space data. This example will use the techniques covered in the MNE-Python tutorials, but applied specifically to fNIRS data.

This script is an example of analysis performed in the manuscript Luke et. al. (2021) [1].

Note

This tutorial uses data stored using the BIDS format [2].

MNE-Python allows you to process fNIRS data that is not in BIDS format. Simply modify the read_raw_ function to match your data type. See data importing tutorial to learn how to use your data with MNE-Python.

# Authors: Robert Luke <mail@robertluke.net>
#          Alexandre Gramfort <alexandre.gramfort@inria.fr>
# License: BSD (3-clause)

# Import common libraries
import contextlib
import os

import numpy as np
from mne import Epochs, events_from_annotations
from mne.decoding import Scaler, Vectorizer, cross_val_multiscore

# Import MNE-Python processing
from mne.preprocessing.nirs import beer_lambert_law, optical_density

# Import MNE-BIDS processing
from mne_bids import BIDSPath, get_entity_vals, read_raw_bids
from sklearn.linear_model import LogisticRegression

# Import sklearn processing
from sklearn.pipeline import make_pipeline

# Import MNE-NIRS processing
from mne_nirs.datasets.fnirs_motor_group import data_path

Set up directories

Requires MNE-BIDS fNIRS branch

This section of code requires the MNE-BIDS fNIRS branch. See instructions at the top of the page on how to install. Alternatively, if your data is not in BIDS format, skip to the next section.

First we will define where the raw data is stored. We will analyse a BIDS dataset. This ensures we have all the metadata we require without manually specifying the trigger names etc. We first define where the root directory of our dataset is. In this example we use the example dataset audio_or_visual_speech.

root = data_path()
dataset = BIDSPath(
    root=root, suffix="nirs", extension=".snirf", task="tapping", datatype="nirs"
)
subjects = get_entity_vals(root, "subject")

Define individual analysis

More details on the epoching analysis can be found at Waveform individual analysis. A minimal processing pipeline is demonstrated here, as the focus of this tutorial is to demonstrate the decoding pipeline. In this example only the epochs for the two conditions we wish to decode between are retained.

def epoch_preprocessing(bids_path):
    with open(os.devnull, "w") as f, contextlib.redirect_stdout(f):
        raw_intensity = read_raw_bids(bids_path=bids_path).load_data()

    raw_od = optical_density(raw_intensity)
    raw_od.resample(1.5)
    raw_haemo = beer_lambert_law(raw_od, ppf=6)
    raw_haemo = raw_haemo.filter(
        None, 0.6, h_trans_bandwidth=0.05, l_trans_bandwidth=0.01, verbose=False
    )

    events, event_dict = events_from_annotations(raw_haemo, verbose=False)
    epochs = Epochs(
        raw_haemo,
        events,
        event_id=event_dict,
        tmin=-5,
        tmax=30,
        reject=dict(hbo=100e-6),
        reject_by_annotation=True,
        proj=True,
        baseline=(None, 0),
        detrend=1,
        preload=True,
        verbose=False,
    )

    epochs = epochs[["Tapping/Right", "Tapping/Left"]]
    return raw_haemo, epochs

Run analysis on all participants

Next we loop through each measurement and decode between the control and audio condition. Here we compute a single spatio-temporal metric approach that simultaneously uses all channels and time points to estimate the experimental condition. The data is scaled for each channel by the mean and standard deviation from all time points and epochs, after which they were vectorized to comply with the scikit-learn data structure, and a logistic regression classifier was applied using the liblinear solver. This approach classifies the data within, rather than across, subjects.

for chroma in ["hbo", "hbr"]:
    st_scores = []
    for sub in subjects:
        bids_path = dataset.update(subject=sub)
        raw_haemo, epochs = epoch_preprocessing(bids_path)

        epochs.pick(chroma)

        X = epochs.get_data()
        y = epochs.events[:, 2]

        clf = make_pipeline(
            Scaler(epochs.info), Vectorizer(), LogisticRegression(solver="liblinear")
        )

        scores = 100 * cross_val_multiscore(
            clf, X, y, cv=5, n_jobs=1, scoring="roc_auc"
        )

        st_scores.append(np.mean(scores, axis=0))

    print(
        f"Average spatio-temporal ROC-AUC performance ({chroma}) = "
        f"{np.round(np.mean(st_scores))} % ({np.round(np.std(st_scores))})"
    )

Average spatio-temporal ROC-AUC performance (hbo) = 89.0 % (10.0)
Average spatio-temporal ROC-AUC performance (hbr) = 85.0 % (14.0)

Conclusion

Data were epoched then decoding was performed on the hbo signal and the hbr signal. The HbO signal decodes the conditions with 6% greater accuracy than the HbR signal. For further discussion about the efficacy of fNIRS signals in decoding experimental condition see Luke et. al. (2021) [1].

Bibliography

[1] (1,2)

Robert Luke, Maureen Shader, Eric Larson, Alexandre Gramfort, Adrian KC Lee, and David McAlpine. Oxygenated hemoglobin signal provides greater predictive performance of experimental condition than de-oxygenated. bioRxiv, 2021. doi:10.1101/2021.11.19.469225.

[2]

Robert Luke, Robert Oostenveld, Helena Cocks, Guiomar Niso, Maureen Shader, Felipe Orihuela-Espina, Hamish Innes-Brown, Stephen Tucker, David Boas, Remi Gau, Taylor Salo, Stefan Appelhoff, Christopher Markiewicz, David McAlpine, and Luca Pollonini. Fnirs-bids, the brain imaging data structure extended to functional near-infrared spectroscopy. OSF Preprints, 2023. doi:10.31219/osf.io/7nmcp.

Estimated memory usage: 462 MB