From mixed signals to neural code: compensating volume conduction for noninvasive generic neural interfaces.

Abstract

Introduction: Noninvasive neural interfaces promise scalable access to neural information without the risks of implanted sensors, but their fundamental limitation is the transformation imposed by the volume conductor between neural sources and sensors. Biological tissues spatially and temporally filter, mix, and disperse neural activity, such that recorded signals (e.g. EEG, ENG, surface EMG) primarily reflect convolution with tissue-dependent impulse responses rather than the underlying neural information.

Area covered: We frame this challenge using a generic convolutive model in which neural sources are observed through volume-conductor filters and noise. Two complementary strategies are discussed: direct compensation, which seeks to separate and recover subsets of neural sources through deconvolution methods, and indirect compensation, which learns representations that are invariant to volume-conductor variability from large, diverse datasets.

Expert opinion: We argue that progress in noninvasive interfacing will depend on explicit recognition and compensation of the volume conductor effect, either directly or indirectly. Together, these strategies point toward noninvasive neural interfaces that can scale beyond subject-specific calibration by isolating neural information from tissue-dependent distortions.