Developing Novel Machine Learning Approaches for Distinguishing Neural Time-Series Patterns

Abstract

Machine learning (ML) has made a significant impact on the field of neuroscience. Tools from ML allow researchers to gain insights into large and complex datasets that would be difficult to obtain otherwise. In this work, we introduce novel ML approaches for analyzing high-dimensional neural activity datasets. More specifically, distinguishing between complex nonlinear neural time-series patterns is a challenging problem in neuroscience. Accurately classifying different time-series patterns could be useful for a wide variety of clinical and research applications. Some examples include detecting seizures in the context of epilepsy, brain-machine interface applications, and real-time detection of events from clinical streams of data. Previously, a number of approaches have been put forward for neural time-series classification. On the one hand, there are simple distance metrics like Euclidean distance, which can be computed quickly, but do not yield accurate classifications. On the other hand, there are more complex approaches such as deep neural networks which offer high accuracy but are training data intensive. We introduce a novel approach based on reservoir computing, termed as TRAKR (short for state tracker), which provides the high accuracy of complex deep supervised methods while preserving the benefits of simple distance metrics in being applicable to fewer examples of training data. We validate TRAKR on synthetic time series datasets generated from chaotic data recurrent neural networks. We also apply TRAKR to a benchmark dataset in the domain of ML – permuted sequential MNIST (psMNIST) – and show that it achieves high classification accuracy. In addition, we use TRAKR for distinguishing neural time-series patterns from animal and human electrophysiological recordings. We apply TRAKR to electrocorticography (ECoG) activity from the macaque orbitofrontal cortex (OFC). We find that TRAKR can distinguish behaviorally relevant neural activity patterns from different macaque experimental trials accurately. Similarly, we apply TRAKR to local field potentials (LFP) collected from patients who undergo subcallosal cingulate deep brain stimulation (SCC DBS) surgery for treatment-resistant depression. We find that TRAKR can accurately distinguish neural signals collected before and after electrical stimulation of the SCC. In all these cases, TRAKR performs on par with deep networks and outperforms simple distance metrics. Thus, TRAKR is a viable ML tool for the classification of neural time-series activity, offering the potential to generate new insights into the information encoded in neural circuits from smaller amounts of data.