Learning to Predict Cortical Potentials using Simultaneous Transcranial Recordings

Abstract Characterizing the transform from cortical potentials to scalp electrodereadings (the forward problem of brain tomography) and from scalp elec-trode readings to cortical potentials (the inverse problem) remain chal-lenging research problems. A major reason for this is the lack of groundtruth data on both sides of the mapping: to our knowledge, no publishedstudies to date have analyzed channel characteristics derived from simul-taneous high-density cortical and scalp recordings. Usingsimultaneousrecordings from a high-density array of 258 EEG electrodes and 42 sub-dural ECoG electrodes implanted in a human subject, we present the firstsuch characterization of the cortex-scalp transfer function using knownground truth data. We show that a linear method suffices to sol ve theforward problem. We apply nonlinear statistical machine learning meth-ods based on non-parametric inference for the inverse problem. Specif-ically, we show that a Gaussian process method employing minimal as-sumptions about the transfer function can recover important temporaland spectral features of the underlying brain signal. Applying a particlefilter to the cortical estimate improves the model’s accurac y in both thetemporal and spectral domains.