Brian C. Dean

Standardized database development for EEG epileptiform transient detection: EEGnet scoring system and machine learning analysis

The routine scalp electroencephalogram (rsEEG) is the most common clinical neurophysiology procedure. The most important role of rsEEG is to detect evidence of epilepsy, in the form of epileptiform transients (ETs), also known as spike or sharp wave discharges. Due to the wide variety of morphologies of ETs and their similarity to artifacts and waves that are part of the normal background activity, the task of ET detection is difficult and mistakes are frequently made. The development of reliable computerized detection of ETs in the EEG could assist physicians in interpreting rsEEGs. We report progress in developing a standardized database for testing and training ET detection algorithms. We describe a new version of our EEGnet software system for collecting expert opinion on EEG datasets, a completely web-browser based system. We report results of EEG scoring from a group of 11 board-certified academic clinical neurophysiologists who annotated 30-s excepts from rsEEG recordings from 100 different patients. The scorers had moderate inter-scorer reliability and low to moderate intra-scorer reliability. In order to measure the optimal size of this standardized rsEEG database, we used machine learning models to classify paroxysmal EEG activity in our database into ET and non-ET classes. Based on our results, it appears that our database will need to be larger than its current size. Also, our non-parametric classifier, an artificial neural network, performed better than our parametric Bayesian classifier. Of our feature sets, the wavelet feature set proved most useful for classification.

Inter-rater agreement on identification of electrographic seizures and periodic discharges in ICU EEG recordings

OBJECTIVE This study investigated inter-rater agreement (IRA) among EEG experts for the identification of electrographic seizures and periodic discharges (PDs) in continuous ICU EEG recordings. METHODS Eight board-certified EEG experts independently identified seizures and PDs in thirty 1-h EEG segments which were selected from ICU EEG recordings collected from three medical centers. IRA was compared between seizure and PD identifications, as well as among rater groups that have passed an ICU EEG Certification Test, developed by the Critical Care EEG Monitoring Research Consortium (CCEMRC). RESULTS Both kappa and event-based IRA statistics showed higher mean values in identification of seizures compared to PDs (k=0.58 vs. 0.38; p<0.001). The group of rater pairs who had both passed the ICU EEG Certification Test had a significantly higher mean IRA in comparison to rater pairs in which neither had passed the test. CONCLUSIONS IRA among experts is significantly higher for identification of electrographic seizures compared to PDs. Additional instruction, such as the training module and certification test developed by the CCEMRC, could enhance this IRA. SIGNIFICANCE This study demonstrates more disagreement in the labeling of PDs in comparison to seizures. This may be improved by education about standard EEG nomenclature.

No sensor left behind: enriching computing education with mobile devices

The use of mobile app development in pre-college computing education is rapidly gaining momentum due to the increasingly widespread use of mobile devices. To fully realize the learning potential of this technology in the classroom, however, one may need to re-examine traditional curricular approaches originating from desktop computing environments. In this work, we describe our experience with a new high-school computing camp designed from the ground up to engage students by taking full advantage of the specific benefits of mobile devices, such as built-in cameras, GPS, networking, and sensors measuring touch, sound, acceleration, and orientation. We describe the design of our camp including materials and examples used. We assess the effectiveness of this instructional approach by demonstrating a statistically significant increase in interest in future computing endeavors. We also comment on the use of MIT App Inventor to ease the transition, particularly for novice programmers, to more sophisticated Java-based apps.

Faster Algorithms for Stable Allocation Problems

We consider a high-multiplicity generalization of the classical stable matching problem known as the stable allocation problem, introduced by Baïou and Balinski in 2002. By leveraging new structural properties and sophisticated data structures, we show how to solve this problem in O(mlog n) time on a bipartite instance with n vertices and m edges, improving the best known running time of O(mn). Building on this algorithm, we provide an algorithm for the non-bipartite stable allocation problem running in O(mlog n) time with high probability. Finally, we give a polynomial-time algorithm for solving the “optimal” variant of the bipartite stable allocation problem, as well as a 2-approximation algorithm for the NP-hard “optimal” variant of the non-bipartite stable allocation problem.

Web-based collection of expert opinion on routine scalp EEG: software development and interrater reliability.

Computerized detection of epileptiform transients (ETs), characterized by interictal spikes and sharp waves in the EEG, has been a research goal for the last 40 years. A reliable method for detecting ETs would assist physicians in interpretation and improve efficiency in reviewing long-term EEG recordings. Computer algorithms developed thus far for detecting ETs are not as reliable as human experts, primarily due to the large number of false-positive detections. Comparing the performance of different algorithms is difficult because each study uses individual EEG test datasets. In this article, we present EEGnet, a distributed web-based platform for the acquisition and analysis of large-scale training datasets for comparison of different EEG ET detection algorithms. This software allows EEG scorers to log in through the web, mark EEG segments of interest, and categorize segments of interest using a conventional clinical EEG user interface. This software platform was used by seven board-certified academic epileptologists to score 40 short 30-second EEG segments from 40 patients, half containing ETs and half containing artifacts and normal variants. The software performance was adequate. Interrater reliability for marking the location of paroxysmal activity was low. Interrater reliability of marking artifacts and ETs was high and moderate, respectively.

The Unsplittable Stable Marriage Problem

The Gale-Shapley “propose/reject” algorithm is a well-known procedure for solving the classical stable marriage problem. In this paper we study this algorithm in the context of the many-to-many stable marriage problem, also known as the stable allocation or ordinal transportation problem. We present an integral variant of the Gale-Shapley algorithm that provides a direct analog, in the context of “ordinal” assignment problems, of a well-known bicriteria approximation algorithm of Shmoys and Tardos for scheduling on unrelated parallel machines with costs. If we are assigning, say, jobs to machines, our algorithm finds an unsplit (non-preemptive) stable assignment where every job is assigned at least as well as it could be in any fractional stable assignment, and where each machine is congested by at most the processing time of the largest job.

A linear programming approach to reconstructing subcellular structures from confocal images for automated generation of representative 3D cellular models

This paper presents a novel computer vision algorithm to analyze 3D stacks of confocal images of fluorescently stained single cells. The goal of the algorithm is to create representative in silico model structures that can be imported into finite element analysis software for mechanical characterization. Segmentation of cell and nucleus boundaries is accomplished via standard thresholding methods. Using novel linear programming methods, a representative actin stress fiber network is generated by computing a linear superposition of fibers having minimum discrepancy compared with an experimental 3D confocal image. Qualitative validation is performed through analysis of seven 3D confocal image stacks of adherent vascular smooth muscle cells (VSMCs) grown in 2D culture. The presented method is able to automatically generate 3D geometries of the cells boundary, nucleus, and representative F-actin network based on standard cell microscopy data. These geometries can be used for direct importation and implementation in structural finite element models for analysis of the mechanics of a single cell to potentially speed discoveries in the fields of regenerative medicine, mechanobiology, and drug discovery.

Morphology-based wavelet features and multiple mother wavelet strategy for spike classification in EEG signals

New wavelet-derived features and strategies that can improve autonomous EEG classifier performance are presented. Various feature sets based on the morphological structure of wavelet subband coefficients are derived and evaluated. The performance of these new feature sets is superior to Gulers classic features in both sensitivity and specificity. In addition, the use of (scalp electrode) spatial information is also shown to improve EEG classification. Finally, a new strategy based upon concurrent use of several mother wavelets is shown to result in increased sensitivity and specificity. Various attempts at reducing feature vector dimension are shown. A non-parametric method, k-NNR, is implemented for classification and 10-fold cross-validation is used for assessment.

Finite termination of augmenting path algorithms in the presence of irrational problem data

This paper considers two similar graph algorithms that work by repeatedly increasing flow along augmenting paths: the Ford-Fulkerson algorithm for the maximum flow problem and the Gale-Shapley algorithm for the stable allocation problem (a many-to-many generalization of the stable matching problem). Both algorithms clearly terminate when given integral input data. For real-valued input data, it was previously known that the Ford-Fulkerson algorithm runs in polynomial time if augmenting paths are chosen via breadth-first search, but that the algorithm might fail to terminate if augmenting paths are chosen in an arbitrary fashion. However, the performance of the Gale-Shapley algorithm on real-valued data was unresolved. Our main result shows that, in contrast to the Ford-Fulkerson algorithm, the Gale-Shapley algorithm always terminates in finite time on real-valued data. Although the Gale-Shapley algorithm may take exponential time in the worst case, it is a popular algorithm in practice due to its simplicity and the fact that it often runs very quickly (even in sublinear time) for many inputs encountered in practice. We also study the Ford-Fulkerson algorithm when augmenting paths are chosen via depth-first search, a common implementation in practice. We prove that, like breadth-first search, depth-first search also leads to finite termination (although not necessarily in polynomial time).

A Linear Programming Approach for Automated Localization of Multiple Faults

In this paper, we address the problem of localizing faults by analyzing execution traces of successful and unsuccessful invocations of the application when run against a suite of tests. We present a new algorithm, based on a linear programming model, which is designed to be particularly effective for the case where multiple faults are present in the application under investigation. Through an extensive empirical study, we show that in the case of both single and multiple faults, our approach outperforms a host of prominent fault localization methods from the literature.