DDALAB: Identifying Latent States from Neural Recordings with Nonlinear Causal Analysis

University Of California, San Diego, La Jolla CA

Investigators

Abstract

Summary The goal of this proposal is to develop DDALAB, a software platform that will make it possible for researchers to identify latent cortical states and analyze the flow of information in large populations of neurons using Delay Differential Analysis (DDA). Although DDA can be used to analyze any time series data, we will initially focus on EEG recordings from the scalp and iEEG data recordings directly from the brain. In addition to developing software making it easy for an investigator to analyze their own recordings, we will also develop interfaces with recordings stored in OpenNeuro archive, a data repository funded the The BRAIN Initiative. These data can be analyzed and visualized with the DDALAB running on local computers or imported directly from OpenNeuro into the NEMAR resource and processed via the Neuroscience Gateway (NSG) at the San Diego Supercomputer Center (SDSC) for High Performance Computing (HPC). We propose to integrate DDALAB into the existing ecosystem supported by the BRAIN Initiative. Delay Differential Analysis (DDA) is a nonlinear, time-domain technique that fits time series waveforms, which complements commonly used frequency domain techniques based on linear Fourier analysis. DDA has a number of advantages for analyzing brain recordings: • DDA is able to extract nonlinear features in recordings that are invisible to linear techniques. • Neural recordings and other time series can be accurately fit with a few low-order time-delayed polynomial terms, typically having only 3 parameters. This reduces overfilling and makes DDA insensitive to most artifacts, allowing DDA to be used for online analysis of raw recordings without preprocessing. • The output of DDA is a highly compressed version of the time series because noise and artifacts are ignored. DDA extracts and distills brain signals from raw data for later analysis. • Much less data are required to specify a model compared with machine learning. • The same set of DDA models fits recordings across subjects, suggesting that DDA is capturing fundamental properties of cortical dynamics. • Fewer time points are needed in a moving window compared with spectral windows, improving the time resolution. DDALAB will provide data analysis for identifying latent changes in cortical states and visualization tools that can be used to extract estimates for the directed flow of information between brain areas. These methods can be applied by the research community at large to analyze a wide range of brain recordings and to develop better treatments for patients with brain diseases. The software developed in this proposal will be openly available through GitHub with an Open Source Software license. Users will not have to buy commercial software or depend on proprietary data formats.