Kenneth A. Colwell

Bayesian Approach to Dynamically Controlling Data Collection in P300 Spellers

P300 spellers provide a noninvasive method of communication for people who may not be able to use other communication aids due to severe neuromuscular disabilities. However, P300 spellers rely on event-related potentials (ERPs) which often have low signal-to-noise ratios (SNRs). In order to improve detection of the ERPs, P300 spellers typically collect multiple measurements of the electroencephalography (EEG) response for each character. The amount of collected data can affect both the accuracy and the communication rate of the speller system. The goal of the present study was to develop an algorithm that would automatically determine the necessary amount of data to collect during operation. Dynamic data collection was controlled by a threshold on the probabilities that each possible character was the target character, and these probabilities were continually updated with each additional measurement. This Bayesian technique differs from other dynamic data collection techniques by relying on a participant-independent, probability-based metric as the stopping criterion. The accuracy and communication rate for dynamic and static data collection in P300 spellers were compared for 26 users. Dynamic data collection resulted in a significant increase in accuracy and communication rate.

Increasing BCI communication rates with dynamic stopping towards more practical use: an ALS study

OBJECTIVE The P300 speller is a brain-computer interface (BCI) that can possibly restore communication abilities to individuals with severe neuromuscular disabilities, such as amyotrophic lateral sclerosis (ALS), by exploiting elicited brain signals in electroencephalography (EEG) data. However, accurate spelling with BCIs is slow due to the need to average data over multiple trials to increase the signal-to-noise ratio (SNR) of the elicited brain signals. Probabilistic approaches to dynamically control data collection have shown improved performance in non-disabled populations; however, validation of these approaches in a target BCI user population has not occurred. APPROACH We have developed a data-driven algorithm for the P300 speller based on Bayesian inference that improves spelling time by adaptively selecting the number of trials based on the acute SNR of a users EEG data. We further enhanced the algorithm by incorporating information about the users language. In this current study, we test and validate the algorithms online in a target BCI user population, by comparing the performance of the dynamic stopping (DS) (or early stopping) algorithms against the current state-of-the-art method, static data collection, where the amount of data collected is fixed prior to online operation. MAIN RESULTS Results from online testing of the DS algorithms in participants with ALS demonstrate a significant increase in communication rate as measured in bits/min (100-300%), and theoretical bit rate (100-550%), while maintaining selection accuracy. Participants also overwhelmingly preferred the DS algorithms. SIGNIFICANCE We have developed a viable BCI algorithm that has been tested in a target BCI population which has the potential for translation to improve BCI speller performance towards more practical use for communication.

Utilizing a Language Model to Improve Online Dynamic Data Collection in P300 Spellers

P300 spellers provide a means of communication for individuals with severe physical limitations, especially those with locked-in syndrome, such as amyotrophic lateral sclerosis. However, P300 speller use is still limited by relatively low communication rates due to the multiple data measurements that are required to improve the signal-to-noise ratio of event-related potentials for increased accuracy. Therefore, the amount of data collection has competing effects on accuracy and spelling speed. Adaptively varying the amount of data collection prior to character selection has been shown to improve spelling accuracy and speed. The goal of this study was to optimize a previously developed dynamic stopping algorithm that uses a Bayesian approach to control data collection by incorporating a priori knowledge via a language model. Participants (n = 17) completed online spelling tasks using the dynamic stopping algorithm, with and without a language model. The addition of the language model resulted in improved participant performance from a mean theoretical bit rate of 46.12 bits/min at 88.89% accuracy to 54.42 bits/min (p <; 0.0065) at 90.36% accuracy.

Projected Accuracy Metric for the P300 Speller

The P300 Speller brain-computer interface (BCI) is a virtual keyboard that allows users to type without requiring neuromuscular control. P300 Speller research commonly aims to improve the system accuracy, which is typically estimated by spelling a small number of characters and calculating the percent spelled correctly. In this paper we introduce a new method for estimating the long-term (“projected”) accuracy, which utilizes all available flash data and a probabilistic model of the Speller system to produce an estimate with lower variance and lower granularity than the standard measure. We apply the new method to 110 previously-collected P300 Speller runs to confirm its consistency, and simulate spelling runs from real subject data to demonstrate lower variance on the accuracy estimate for any given amount of data.

Sparse model inversion and processing of spatial frequency-domain electromagnetic induction sensor array data for improved landmine discrimination

Frequency-domain electromagnetic induction (EMI) sensors have been shown to provide target signatures which enable discrimination of landmines from harmless clutter. In particular, frequency-domain EMI sensors are well-suited for target characterization by inverting a physics-based signal model. In many model-based signal processing paradigms, the target signatures can be decomposed into a weighted sum of parameterized basis functions, where the basis functions are intrinsic to the target under consideration and the associated weights are a function of the target sensor orientation. When sensor array data is available, the spatial diversity of the measured signals may provide more information for estimating the basis function parameters. After model inversion, the basis function parameters can form the foundation of model-based classification of the target as landmine or clutter. In this work, sparse model inversion of spatial frequency-domain EMI sensor array data followed by target classification using a statistical model is investigated. Results for data measured with a prototype frequency-domain EMI sensor at a standardized test site are presented. Preliminary results indicate that extracting physics-based features from spatial frequency-domain EMI sensor array data followed by statistical classification provides an effective approach for classifying targets as landmine or clutter.

Evaluating Brain-Computer Interface Performance Using Color in the P300 Checkerboard Speller

OBJECTIVE Current Brain-Computer Interface (BCI) systems typically flash an array of items from grey to white (GW). The objective of this study was to evaluate BCI performance using uniquely colored stimuli. METHODS In addition to the GW stimuli, the current study tested two types of color stimuli (grey to color [GC] and color intensification [CI]). The main hypotheses were that in a checkboard paradigm, unique color stimuli will: (1) increase BCI performance over the standard GW paradigm; (2) elicit larger event-related potentials (ERPs); and, (3) improve offline performance with an electrode selection algorithm (i.e., Jumpwise). RESULTS Online results (n=36) showed that GC provides higher accuracy and information transfer rate than the CI and GW conditions. Waveform analysis showed that GC produced higher amplitude ERPs than CI and GW. Information transfer rate was improved by the Jumpwise-selected channel locations in all conditions. CONCLUSIONS Unique color stimuli (GC) improved BCI performance and enhanced ERPs. Jumpwise-selected electrode locations improved offline performance. SIGNIFICANCE These results show that in a checkerboard paradigm, unique color stimuli increase BCI performance, are preferred by participants, and are important to the design of end-user applications; thus, could lead to an increase in end-user performance and acceptance of BCI technology.

Attribute-driven transfer learning for detecting novel buried threats with ground-penetrating radar

Ground-penetrating radar (GPR) technology is an effective method of detecting buried explosive threats. The system uses a binary classifier to distinguish “targets”, or buried threats, from “nontargets” arising from system prescreener false alarms; this classifier is trained on a dataset of previously-observed buried threat types. However, the threat environment is not static, and new threat types that appear must be effectively detected even if they are not highly similar to every previously-observed type. Gathering a new dataset that includes a new threat type is expensive and time-consuming; minimizing the amount of new data required to effectively detect the new type is therefore valuable. This research aims to reduce the number of training examples needed to effectively detect new types using transfer learning, which leverages previous learning tasks to accelerate and improve new ones. Further, new types have attribute data, such as composition, components, construction, and size, which can be observed without GPR and typically are not explicitly included in the learning process. Since attribute tags for buried threats determine many aspects of their GPR representation, a new threat type’s attributes can be highly relevant to the transfer-learning process. In this work, attribute data is used to drive transfer learning, both by using attributes to select relevant dataset examples for classifier fusion, and by extending a relevance vector machine (RVM) model to perform intelligent attribute clustering and selection. Classification performance results for both the attribute-only case and the low-data case are presented, using a dataset containing a variety of threat types.

ALS Population Assessment of a Dynamic Stopping Algorithm Implementation for P300 Spellers

P300-based BCIs have been proposed as potential communication alternatives for individuals whose severe neuro-muscular limitations, e.g. due to amyotrophic lateral sclerosis (ALS), preclude their use of most commercially available assistive technologies. However, BCIs are currently limited by their relatively slower selection rates due to the significant amount of data collection required to improve the signal-to- noise ratio (SNR) of the elicited brain responses to achieve desired system accuracy levels. The conventional strategy is to average over a fixed amount of data prior to BCI decision making, an approach that might be inefficient given the inherent variation in a user’s responses during BCI use. There is need for the development of improved algorithms that can demonstrate increased performance in online testing, especially in target end-user populations. We have developed an algorithm that uses a Bayesian approach to collect only the amount of data necessary to reach a specified confidence level in the BCI’s decision based on continuous evaluation of the quality of a user’s responses. We further optimized the algorithm by incorporating statistical information about the user’s language. Results from online testing in participants with ALS demonstrate that using the Bayesian dynamic stopping algorithm resulted in a significant reduction in character selection time with minimal effect on accuracy, compared to conventional static data collection. In post-use surveys, the participants overwhelmingly preferred the dynamic stopping algorithms.

Evaluating Brain-Computer Interface Performance in an ALS Population: Checkerboard and Color Paradigms

The objective of this study was to investigate the performance of 3 brain-computer interface (BCI) paradigms in an amyotrophic lateral sclerosis (ALS) population (n = 11). Using a repeated-measures design, participants completed 3 BCI conditions: row/column (RCW), checkerboard (CBW), and gray-to-color (CBC). Based on previous studies, it is hypothesized that the CBC and CBW conditions will result in higher accuracy, information transfer rate, waveform amplitude, and user preference over the RCW condition. An offline dynamic stopping simulation will also increase information transfer rate. Higher mean accuracy was observed in the CBC condition (89.7%), followed by the CBW (84.3%) condition, and lowest in the RCW condition (78.7%); however, these differences did not reach statistical significance (P = .062). Eight of the eleven participants preferred the CBC and the remaining three preferred the CBW conditions. The offline dynamic stopping simulation significantly increased information transfer rate (P = .005) and decreased accuracy (P < .000). The findings of this study suggest that color stimuli provide a modest improvement in performance and that participants prefer color stimuli over monochromatic stimuli. Given these findings, BCI paradigms that use color stimuli should be considered for individuals who have ALS.

Improving buried threat detection in ground-penetrating radar with transfer learning and metadata analysis

Ground-penetrating radar (GPR) technology has proven capable of detecting buried threats. The system relies on a binary classifier that is trained to distinguish between two classes: a target class, encompassing many types of buried threats and their components; and a nontarget class, which includes false alarms from the system prescreener. Typically, the training process involves a simple partition of the data into these two classes, which allows for straightforward application of standard classifiers. However, since training data is generally collected in fully controlled environments, it includes auxiliary information about each example, such as the specific type of threat, its purpose, its components, and its depth. Examples from the same specific or general type may be expected to exhibit similarities in their GPR data, whereas examples from different types may differ greatly. This research aims to leverage this additional information to improve overall classification performance by fusing classifier concepts for multiple groups, and to investigate whether structure in this information can be further utilized for transfer learning, such that the amount of expensive training data necessary to learn a new, previously-unseen target type may be reduced. Methods for accomplishing these goals are presented with results from a dataset containing a variety of target types.