Robust, Automated Methods for Filtering and Processing Neural Signals

This dissertation presents novel tools for robust filtering and processing of neural signals. These tools improve upon existing methods and were shown to be effective under a variety of conditions. They are also simple to use, allowing researchers and clinicians to focus more time on the analysis of neural data and making many tasks accessible to non-expert personnel. The main contributions of this research were the creation of a generalized software framework for neural signal processing, the development of novel algorithms to filter common sources of noise, and an implementation of a brain-computer interface (BCI) decoder as an example application.

The framework has a modular structure and provides simple methods to incorporate neural signal processing tasks and applications. The software was found to maintain precise timing and reliable communication between components. A simple user interface allowed real-time control of all system parameters, and data was efficiently streamed to disk to allow for offline analysis.

One common source of contamination in neural signals is line noise. A method was developed for filtering this noise with a variable bandwidth filter capable of tracking a sinusoid’s frequency. The method is based on the adaptive noise canceling (ANC) technique and is referred to here as the adaptive sinusoid canceler (ASC). This filter effectively eliminates sinusoidal contamination by tracking its frequency and achieving a narrow bandwidth. The ASC was found to outperform comparative methods including standard notch filters and an adaptive line enhancer (ALE).

Ocular artifacts (OAs) caused by eye movement can also present a large problem in neural recordings. Here, a wavelet-based technique was developed for efficiently removing these artifacts. The technique uses a discrete wavelet transform with an automatically selected decomposition level to localize artifacts in both time and frequency before removing them with thresholding. This method was shown to produce superior reduction of OAs when compared to regression, principal component analysis (PCA), and independent component analysis (ICA).

Finally, the removal of spatially correlated broadband noise such as electromyographic (EMG) artifacts was addressed. A method termed the adaptive common average reference (ACAR) was developed as an effective method for removing this noise. The ACAR is based on a combination of the common average reference (CAR) and an ANC filter. In a convergent process, a weighted CAR provides a reference to an ANC filter, which in turn provides feedback to enhance the reference. This method outperformed the standard CAR and ICA under most circumstances.

As an example application for the methods developed in this dissertation, a BCI decoder was implemented using linear regression with an elastic net penalty. This decoder provides automatic feature selection and a robust feature set. The software framework was found to provide reliable data for the decoder, and the filtering algorithms increased the availability of neural features that were usable for decoding.