Unsupervised analysis of fNIRS in the everyday world

By Meryem Yücel.

As fNIRS progresses into more natural and life like environments, new challenges in the acquisition and analysis of its signals emerge. A new blind source separation framework, recently published in NeuroImage, aims to provide some remedies.

It is an exciting time to be in fNIRS research: As instruments become smaller and wearable, they allow us to study the human brain under more naturalistic environments and during interaction with others. At the same time, data-driven tools from the domain of machine learning are entering the field and help to tackle the increasingly complex analysis of the signals acquired under those challenging conditions.

One such project with the aim to move “neurotechnology out of the lab” (von Lühmann, PhD Thesis, 2018) was pursued by Alexander von Lühmann and Klaus-Robert Müller at the Machine Learning Department at TU Berlin, Germany:  Using their newly designed wearable hybrid EEG-fNIRS instrument, the researchers performed an n-back based cognitive workload study, simultaneously measuring EEG, fNIRS and accelerometer signals in freely moving participants who had to take steps, bend down and stretch to press buttons according to the task. In collaboration with Tülay Adalı and Zois Boukouvalas at the Machine Learning for Signal Processing Department at UMBC Baltimore, USA, the group then developed a multimodal blind source separation framework (von Lühmann et al., NI, 2019) for the analysis of the heavily contaminated fNIRS signals to robustly reduce the effects of physiological and motion-induced artifacts. The framework combines the strengths of “ERBM-ICA”, a new kind of Independent Component Analysis that exploits higher order statistics and sample dependence, Canonical Correlation Analysis (CCA) with temporal embedding, and multimodality (accelerometer signals). This way, challenging characteristics in the fNIRS signals are addressed, such as non-instantaneous and non-constant coupling, correlated noise, and statistical dependency between fNIRS components. Applying this unsupervised approach, called “BLISSA2RD”, to their data, the group reported a strong reduction in systemic physiological effects of movement on the signals at a performance beyond that of conventional movement artifact correction. The framework further enables the single trial Brain-Computer Interface based classification of mental workload in freely moving operators. Dr. von Lühmann and colleagues point out that the framework is not limited to motion artifact rejection, however, and that it can be used for various other analysis approaches, particularly of interest for the commonly applied supervised General Linear Model for fNIRS.

To know more:

  • von Lühmann, A., Boukouvalas, Z., Müller, K.R. and Adalı, T., “A new blind source separation framework for signal analysis and artifact rejection in functional Near-Infrared Spectroscopy”, 2019. NeuroImage, vol. 200, p 72-88, DOI: 10.1016/j.neuroimage.2019.06.021
  • von Lühmann, A., “Multimodal instrumentation and methods for neurotechnology out of the lab”, 2018, PhD Thesis, Technical University Berlin, DOI: 10.14279/depositonce-7445