Artefact detection and removal algorithms for EEG diagnostic systems

The electroencephalogram (EEG) is a medical technology that is used in the monitoring of the brain and in the diagnosis of many neurological illnesses. Although coarse in its precision, the EEG is a non-invasive tool that requires minimal set-up times, and is suitably unobtrusive and mobile to allow continuous monitoring of the patient, either in clinical or domestic environments. Consequently, the EEG is the current tool-of-choice with which to continuously monitor the brain where temporal resolution, ease-of- use and mobility are important. Traditionally, EEG data are examined by a trained clinician who identifies neurological events of interest. However, recent advances in signal processing and machine learning techniques have allowed the automated detection of neurological events for many medical applications. In doing so, the burden of work on the clinician has been significantly reduced, improving the response time to illness, and allowing the relevant medical treatment to be administered within minutes rather than hours. However, as typical EEG signals are of the order of microvolts (μV ), contamination by signals arising from sources other than the brain is frequent. These extra-cerebral sources, known as artefacts, can significantly distort the EEG signal, making its interpretation difficult, and can dramatically disimprove automatic neurological event detection classification performance. This thesis therefore, contributes to the further improvement of auto- mated neurological event detection systems, by identifying some of the major obstacles in deploying these EEG systems in ambulatory and clinical environments so that the EEG technologies can emerge from the laboratory towards real-world settings, where they can have a real-impact on the lives of patients. In this context, the thesis tackles three major problems in EEG monitoring, namely: (i) the problem of head-movement artefacts in ambulatory EEG, (ii) the high numbers of false detections in state-of-the-art, automated, epileptiform activity detection systems and (iii) false detections in state-of-the-art, automated neonatal seizure detection systems. To accomplish this, the thesis employs a wide range of statistical, signal processing and machine learning techniques drawn from mathematics, engineering and computer science. The first body of work outlined in this thesis proposes a system to automatically detect head-movement artefacts in ambulatory EEG and utilises supervised machine learning classifiers to do so. The resulting head-movement artefact detection system is the first of its kind and offers accurate detection of head-movement artefacts in ambulatory EEG. Subsequently, addtional physiological signals, in the form of gyroscopes, are used to detect head-movements and in doing so, bring additional information to the head- movement artefact detection task. A framework for combining EEG and gyroscope signals is then developed, offering improved head-movement arte- fact detection. The artefact detection methods developed for ambulatory EEG are subsequently adapted for use in an automated epileptiform activity detection system. Information from support vector machines classifiers used to detect epileptiform activity is fused with information from artefact-specific detection classifiers in order to significantly reduce the number of false detections in the epileptiform activity detection system. By this means, epileptiform activity detection which compares favourably with other state-of-the-art systems is achieved. Finally, the problem of false detections in automated neonatal seizure detection is approached in an alternative manner; blind source separation techniques, complimented with information from additional physiological signals are used to remove respiration artefact from the EEG. In utilising these methods, some encouraging advances have been made in detecting and removing respiration artefacts from the neonatal EEG, and in doing so, the performance of the underlying diagnostic technology is improved, bringing its deployment in the real-world, clinical domain one step closer.

Profiling side-channel attacks on cryptographic algorithms

Traditionally, attacks on cryptographic algorithms looked for mathematical weaknesses in the underlying structure of a cipher. Side-channel attacks, however, look to extract secret key information based on the leakage from the device on which the cipher is implemented, be it smart-card, microprocessor, dedicated hardware or personal computer. Attacks based on the power consumption, electromagnetic emanations and execution time have all been practically demonstrated on a range of devices to reveal partial secret-key information from which the full key can be reconstructed. The focus of this thesis is power analysis, more specifically a class of attacks known as profiling attacks. These attacks assume a potential attacker has access to, or can control, an identical device to that which is under attack, which allows him to profile the power consumption of operations or data flow during encryption. This assumes a stronger adversary than traditional non-profiling attacks such as differential or correlation power analysis, however the ability to model a device allows templates to be used post-profiling to extract key information from many different target devices using the power consumption of very few encryptions. This allows an adversary to overcome protocols intended to prevent secret key recovery by restricting the number of available traces. In this thesis a detailed investigation of template attacks is conducted, along with how the selection of various attack parameters practically affect the efficiency of the secret key recovery, as well as examining the underlying assumption of profiling attacks in that the power consumption of one device can be used to extract secret keys from another. Trace only attacks, where the corresponding plaintext or ciphertext data is unavailable, are then investigated against both symmetric and asymmetric algorithms with the goal of key recovery from a single trace. This allows an adversary to bypass many of the currently proposed countermeasures, particularly in the asymmetric domain. An investigation into machine-learning methods for side-channel analysis as an alternative to template or stochastic methods is also conducted, with support vector machines, logistic regression and neural networks investigated from a side-channel viewpoint. Both binary and multi-class classification attack scenarios are examined in order to explore the relative strengths of each algorithm. Finally these machine-learning based alternatives are empirically compared with template attacks, with their respective merits examined with regards to attack efficiency.

Optimized approach to decision fusion of heterogeneous data for breast cancer diagnosis.

As more diagnostic testing options become available to physicians, it becomes more difficult to combine various types of medical information together in order to optimize the overall diagnosis. To improve diagnostic performance, here we introduce an approach to optimize a decision-fusion technique to combine heterogeneous information, such as from different modalities, feature categories, or institutions. For classifier comparison we used two performance metrics: The receiving operator characteristic (ROC) area under the curve [area under the ROC curve (AUC)] and the normalized partial area under the curve (pAUC). This study used four classifiers: Linear discriminant analysis (LDA), artificial neural network (ANN), and two variants of our decision-fusion technique, AUC-optimized (DF-A) and pAUC-optimized (DF-P) decision fusion. We applied each of these classifiers with 100-fold cross-validation to two heterogeneous breast cancer data sets: One of mass lesion features and a much more challenging one of microcalcification lesion features. For the calcification data set, DF-A outperformed the other classifiers in terms of AUC (p < 0.02) and achieved AUC=0.85 +/- 0.01. The DF-P surpassed the other classifiers in terms of pAUC (p < 0.01) and reached pAUC=0.38 +/- 0.02. For the mass data set, DF-A outperformed both the ANN and the LDA (p < 0.04) and achieved AUC=0.94 +/- 0.01. Although for this data set there were no statistically significant differences among the classifiers' pAUC values (pAUC=0.57 +/- 0.07 to 0.67 +/- 0.05, p > 0.10), the DF-P did significantly improve specificity versus the LDA at both 98% and 100% sensitivity (p < 0.04). In conclusion, decision fusion directly optimized clinically significant performance measures, such as AUC and pAUC, and sometimes outperformed two well-known machine-learning techniques when applied to two different breast cancer data sets.

Lognormal and gamma mixed negative binomial regression

In regression analysis of counts, a lack of simple and efficient algorithms for posterior computation has made Bayesian approaches appear unattractive and thus underdeveloped. We propose a lognormal and gamma mixed negative binomial (NB) regression model for counts, and present efficient closed-form Bayesian inference; unlike conventional Poisson models, the proposed approach has two free parameters to include two different kinds of random effects, and allows the incorporation of prior information, such as sparsity in the regression coefficients. By placing a gamma distribution prior on the NB dispersion parameter r, and connecting a log-normal distribution prior with the logit of the NB probability parameter p, efficient Gibbs sampling and variational Bayes inference are both developed. The closed-form updates are obtained by exploiting conditional conjugacy via both a compound Poisson representation and a Polya-Gamma distribution based data augmentation approach. The proposed Bayesian inference can be implemented routinely, while being easily generalizable to more complex settings involving multivariate dependence structures. The algorithms are illustrated using real examples. Copyright 2012 by the author(s)/owner(s).

Communications inspired linear discriminant analysis

We study the problem of supervised linear dimensionality reduction, taking an information-theoretic viewpoint. The linear projection matrix is designed by maximizing the mutual information between the projected signal and the class label. By harnessing a recent theoretical result on the gradient of mutual information, the above optimization problem can be solved directly using gradient descent, without requiring simplification of the objective function. Theoretical analysis and empirical comparison are made between the proposed method and two closely related methods, and comparisons are also made with a method in which Rényi entropy is used to define the mutual information (in this case the gradient may be computed simply, under a special parameter setting). Relative to these alternative approaches, the proposed method achieves promising results on real datasets. Copyright 2012 by the author(s)/owner(s).

Gene selection using iterative feature elimination random forests for survival outcomes.

Although many feature selection methods for classification have been developed, there is a need to identify genes in high-dimensional data with censored survival outcomes. Traditional methods for gene selection in classification problems have several drawbacks. First, the majority of the gene selection approaches for classification are single-gene based. Second, many of the gene selection procedures are not embedded within the algorithm itself. The technique of random forests has been found to perform well in high-dimensional data settings with survival outcomes. It also has an embedded feature to identify variables of importance. Therefore, it is an ideal candidate for gene selection in high-dimensional data with survival outcomes. In this paper, we develop a novel method based on the random forests to identify a set of prognostic genes. We compare our method with several machine learning methods and various node split criteria using several real data sets. Our method performed well in both simulations and real data analysis.Additionally, we have shown the advantages of our approach over single-gene-based approaches. Our method incorporates multivariate correlations in microarray data for survival outcomes. The described method allows us to better utilize the information available from microarray data with survival outcomes.

The Representation of Emotion in Autonomic and Central Nervous System Activity

Phenomenologically, humans effectively label and report feeling distinct emotions, yet the extent to which emotions are represented categorically in nervous system activity is controversial. Theoretical accounts differ in this regard, some positing distinct emotional experiences emerge from a dimensional representation (e.g., along axes of valence and arousal) whereas others propose emotions are natural categories, with dedicated neural bases and associated response profiles. This dissertation aims to empirically assess these theoretical accounts by examining how emotions are represented (either as disjoint categories or as points along continuous dimensions) in autonomic and central nervous system activity by integrating psychophysiological recording and functional neuroimaging with machine-learning based analytical methods. Results demonstrate that experientially, emotional events are well-characterized both along dimensional and categorical frameworks. Measures of central and peripheral responding discriminate among emotion categories, but are largely independent of valence and arousal. These findings suggest dimensional and categorical aspects of emotional experience are driven by separable neural substrates and demonstrate that emotional states can be objectively quantified on the basis of nervous system activity.

Predictive Modeling of Loss-of-Heterozygosity Events in Yeast

During mitotic cell cycles, DNA experiences many types of endogenous and exogenous damaging agents that could potentially cause double strand breaks (DSB). In S. cerevisiae, DSBs are primarily repaired by mitotic recombination and as a result, could lead to loss-of-heterozygosity (LOH). Genetic recombination can happen in both meiosis and mitosis. While genome-wide distribution of meiotic recombination events has been intensively studied, mitotic recombination events have not been mapped unbiasedly throughout the genome until recently. Methods for selecting mitotic crossovers and mapping the positions of crossovers have recently been developed in our lab. Our current approach uses a diploid yeast strain that is heterozygous for about 55,000 SNPs, and employs SNP-Microarrays to map LOH events throughout the genome. These methods allow us to examine selected crossovers and unselected mitotic recombination events (crossover, noncrossover and BIR) at about 1 kb resolution across the genome. Using this method, we generated maps of spontaneous and UV-induced LOH events. In this study, we explore machine learning and variable selection techniques to build a predictive model for where the LOH events occur in the genome.

Randomly from the yeast genome, we simulated control tracts resembling the LOH tracts in terms of tract lengths and locations with respect to single-nucleotide-polymorphism positions. We then extracted roughly 1,100 features such as base compositions, histone modifications, presence of tandem repeats etc. and train classifiers to distinguish control tracts and LOH tracts. We found interesting features of good predictive values. We also found that with the current repertoire of features, the prediction is generally better for spontaneous LOH events than UV-induced LOH events.

Statistical analysis of crystallization database links protein physico-chemical features with crystallization mechanisms.

X-ray crystallography is the predominant method for obtaining atomic-scale information about biological macromolecules. Despite the success of the technique, obtaining well diffracting crystals still critically limits going from protein to structure. In practice, the crystallization process proceeds through knowledge-informed empiricism. Better physico-chemical understanding remains elusive because of the large number of variables involved, hence little guidance is available to systematically identify solution conditions that promote crystallization. To help determine relationships between macromolecular properties and their crystallization propensity, we have trained statistical models on samples for 182 proteins supplied by the Northeast Structural Genomics consortium. Gaussian processes, which capture trends beyond the reach of linear statistical models, distinguish between two main physico-chemical mechanisms driving crystallization. One is characterized by low levels of side chain entropy and has been extensively reported in the literature. The other identifies specific electrostatic interactions not previously described in the crystallization context. Because evidence for two distinct mechanisms can be gleaned both from crystal contacts and from solution conditions leading to successful crystallization, the model offers future avenues for optimizing crystallization screens based on partial structural information. The availability of crystallization data coupled with structural outcomes analyzed through state-of-the-art statistical models may thus guide macromolecular crystallization toward a more rational basis.

HDAC inhibitors cause site-specific chromatin remodeling at PU.1-bound enhancers in K562 cells

BACKGROUND: Small molecule inhibitors of histone deacetylases (HDACi) hold promise as anticancer agents for particular malignancies. However, clinical use is often confounded by toxicity, perhaps due to indiscriminate hyperacetylation of cellular proteins. Therefore, elucidating the mechanisms by which HDACi trigger differentiation, cell cycle arrest, or apoptosis of cancer cells could inform development of more targeted therapies. We used the myelogenous leukemia line K562 as a model of HDACi-induced differentiation to investigate chromatin accessibility (DNase-seq) and expression (RNA-seq) changes associated with this process. RESULTS: We identified several thousand specific regulatory elements [~10 % of total DNase I-hypersensitive (DHS) sites] that become significantly more or less accessible with sodium butyrate or suberanilohydroxamic acid treatment. Most of the differential DHS sites display hallmarks of enhancers, including being enriched for non-promoter regions, associating with nearby gene expression changes, and increasing luciferase reporter expression in K562 cells. Differential DHS sites were enriched for key hematopoietic lineage transcription factor motifs, including SPI1 (PU.1), a known pioneer factor. We found PU.1 increases binding at opened DHS sites with HDACi treatment by ChIP-seq, but PU.1 knockdown by shRNA fails to block the chromatin accessibility and expression changes. A machine-learning approach indicates H3K27me3 initially marks PU.1-bound sites that open with HDACi treatment, suggesting these sites are epigenetically poised. CONCLUSIONS: We find HDACi treatment of K562 cells results in site-specific chromatin remodeling at epigenetically poised regulatory elements. PU.1 shows evidence of a pioneer role in this process by marking poised enhancers but is not required for transcriptional activation.

Model selection approaches for non-linear system identification: a review

The identification of non-linear systems using only observed finite datasets has become a mature research area over the last two decades. A class of linear-in-the-parameter models with universal approximation capabilities have been intensively studied and widely used due to the availability of many linear-learning algorithms and their inherent convergence conditions. This article presents a systematic overview of basic research on model selection approaches for linear-in-the-parameter models. One of the fundamental problems in non-linear system identification is to find the minimal model with the best model generalisation performance from observational data only. The important concepts in achieving good model generalisation used in various non-linear system-identification algorithms are first reviewed, including Bayesian parameter regularisation and models selective criteria based on the cross validation and experimental design. A significant advance in machine learning has been the development of the support vector machine as a means for identifying kernel models based on the structural risk minimisation principle. The developments on the convex optimisation-based model construction algorithms including the support vector regression algorithms are outlined. Input selection algorithms and on-line system identification algorithms are also included in this review. Finally, some industrial applications of non-linear models are discussed.

Classifying Network Protocols: A ‘2 Way’ Flow Approach

The identification and classification of network traffic and protocols is a vital step in many quality of service and security systems. Traffic classification strategies must evolve, alongside the protocols utilising the Internet, to overcome the use of ephemeral or masquerading port numbers and transport layer encryption. This research expands the concept of using machine learning on the initial statistics of flow of packets to determine its underlying protocol. Recognising the need for efficient training/retraining of a classifier and the requirement for fast classification, the authors investigate a new application of k-means clustering referred to as 'two-way' classification. The 'two-way' classification uniquely analyses a bidirectional flow as two unidirectional flows and is shown, through experiments on real network traffic, to improve classification accuracy by as much as 18% when measured against similar proposals. It achieves this accuracy while generating fewer clusters, that is, fewer comparisons are needed to classify a flow. A 'two-way' classification offers a new way to improve accuracy and efficiency of machine learning statistical classifiers while still maintaining the fast training times associated with the k-means.

Topological mappings between graphs, trees and generalized trees

We present novel topological mappings between graphs, trees and generalized trees that means between structured objects with different properties. The two major contributions of this paper are, first, to clarify the relation between graphs, trees and generalized trees, a graph class recently introduced. Second, these transformations provide a unique opportunity to transform structured objects into a representation that might be beneficial for a processing, e.g., by machine learning techniques for graph classification. (c) 2006 Elsevier Inc. All rights reserved.