1,721,071 research outputs found
Prognostic Molecular Classification of Breast Cancer Based on Features Extracted from a Scale Space
No full textBreast cancer is one of the most prevalent cancers affecting females in the world. In recent years, many cancer researchers have been trying to determine molecular prognosis tools that predict cancer patient treatment response and/or chance of survival. In particular, the determination of gene expression signatures obtained by feature selection methods applied to large microarray datasets has shown potential. The main purpose of this study is to extend these gene signatures and molecular prognostic classifiers by investigating features constructed from a scale-space representation of the microarray data. Here, we construct a scale space by first mapping all genes to a one-dimensional functional space using protein family information. Next, we applied successive smoothing to the expression values resulting in one scale-space representation of the gene expression data from one sample. At the lowest scale, the scale space contains the original gene expression values, whereas at higher scales meta-features are formed, which are weighted sums of groups of genes. To test whether a scale-space representation is useful we performed feature selection and classification on a publicly available breast cancer expression dataset. We found that, instead of signatures consisting of single genes, meta-genes (i.e. groups of genes) that exist at higher scales were preferentially selected. We furthermore determined cross-validation errors using seven distinct classifiers (NMC, LDC, QDC, FISHERC, PARZENC, 3NNC, and LOGLC) and found that better performance is obtained using the scale-space representation than with the traditional representation of the gene expression data. As a result, we conclude that the scale-space analysis constitutes a potent way of selecting molecular signatures and is useful for prognostic classification.Pattern Recognition and BioinformaticsIntelligent SystemsElectrical Engineering, Mathematics and Computer Scienc
Depth extraction, refinement and confidence estimation from image data
No full textDepth extraction is one of the important steps of D computer vision (CV). Although, it has been researched for many decades and there are variety of methods already that addresses depth extraction, there is no perfect solution that satisfies the needs of all CV algorithms. Stereo vision is implemented in CV as a matching algorithm where an image region in one image is matched to another region in the other image and the disparity between the matches indicates the depth of the region. One of the fundamental issues in stereo matching is the repetition of pattern problem. When there are patterns that are repeated along the search path, stereo algorithms can wrongly match the searched pattern with its repetition in the other image. We showed that a solution to this problem is to use enhanced features that will distinguish the correct pattern from its repetitions. Therefore, the search space is limited to the pattern itself rather than its repetitions and correct disparities can be found. Although stereo algorithms require many computations, these computations are independent from each other which make effective parallelization of these algorithms on a GPU possible. However, their parallelization efficiency relies highly on their architecture. In order to optimize the performance of stereo algorithms, it is important to consider both their accuracy and their parallelization performance. We showed that certain architectures of stereo matching provide better parallelization capability while providing similar accuracies with other architectures. Another way of measuring the depth of a scene is to use depth sensors. After the release of the Microsoft Kinect, depth sensors have been increasingly used in CV applications. Kinect provides dense and real-time depth measurements of indoor scenes which has sufficient quality for many CV applications. However, its quality is not enough for accurate D reconstructions especially on the boundaries of objects. Since there is a mismatch between the RGB and depth map of the Kinect, depth refinement algorithms that consider all of their input depth information as correct, fail to refine depth maps accurately. We showed that, to accurately refine regions around boundaries, refinement algorithms should mark outliers and do the refinement based on the trustworthy part of the depth map. Another fundamental problem of depth sensors including the Kinect is transparent surfaces. On transparent regions, Kinect fail to estimate any depth measurements. Since depth refinement algorithms require sparse depth estimations on a surface in order to estimate the unknown depth, they fail to refine the depth on the transparent surfaces correctly. To fully recover transparent objects on the depth map, we propose to use stereo matching between IR and RGB views of the Kinect in a fully connected energy minimization framework. Our refinement strategy can fully recover transparent objects and it can correct the errors from Kinect measurements and stereo matching estimations. Stereo matching requires distinctive similarity measures to match pixels between two images. Different similarity measures perform differently depending on the noise and texture of the regions. It is important to combine their advantages to increase the accuracy of the matching. To measure which similarity measure performs better than others on a local region, we used stereo confidences. According to the confidence of each measure, multiple measures are adaptively fused. The result of fusion provides more robust and accurate matching compared to any of the fused similarities and any fusion of them with static weights. Finally, we proposed a novel confidence measure for medical image registration based on similar measures from stereo matching confidences. The proposed confidence measure is shown to be correlated with error from expert control points. Besides, our confidence measure can indicate the error as a continuous score, on any region of the image.Pattern Recognition and BioinformaticsElectrical Engineering, Mathematics and Computer Scienc
Operating characteristics for the design and optimisation of classification systems
No full textIn statistical pattern recognition, problems involve distinguishing of various concepts or classes, based on the development of classifiers/discriminators. These exploit discriminatory information existing in measurements originating from objects. A trained classifier results in a partitioning in measurement space, providing some separation between the various classes. In the (typical) case of class overlap, this partitioning inherently results in a trade-off between the various possible classification errors that may occur. This partitioning can be modified to adjust these trade-offs. Given class abundances, a classifier can be evaluated at a given partitioning. However, variations in the abundances leads to an altered classifier performance. These fundamental aspects behind classifier design and evaluation can be studied within the framework of classifier operating characteristics, which is the topic of this dissertation. The contents consist of a number of published/accepted journal and conference papers, contextualised into a number of chapters representing various aspects of operating characteristic analysis. First the well-known two-class operating characteristic is considered, with two new analyses that are useful in certain circumstances. Next, the extension to the elusive multiclass case is considered, showing how standard 2-class operating characteristics analyses can be extended theoretically to the multiclass case. The challenge behind the multiclass extension is shown to be of a computational nature, with the calculation size increasing exponentially with the number of classes. The primary thesis contribution is then presented, consisting of a number of approaches and philosophies that can be used to overcome the computational challenges. Of primary importance is the finding that most practical problems are such that not all dimensions of the operating characteristic interact together significantly. Next it is shown how the operating characteristic approach can be used to design classifiers in ill-defined environments. In these problems some classes may be poorly represented, and the goal of the classifier design is to protect against these unforeseen conditions. Finally, it is shown that operating characteristics can be applied to a multi-stage classifier setup, allowing for a holistic design incorporating interactions between classes, and the classifier stages.Electrical Engineering, Mathematics and Computer Scienc
Data integration strategies for bioinformatics with applications in biomarker and network discovery
No full textNowadays, with large amounts of data becoming available, solving biological quests is becoming more and more a data-driven activity. To support this, there is a need for tools that enable the integration of the many sources of data. This thesis presents several avenues that can be taken, showing how integration can support research in the life sciences. Biological data can be integrated at several levels. We present a categorization of these different strategies within the context of the Data-Information-Knowledge-Wisdom paradigm. We argue that bioinformatics research should not only concentrate on the individual levels, but also on the transitions between these domains. Throughout the thesis we present possible solutions for a number of these different transformations. One of these transformations bridges the gap that exists between the two lowest levels of data integration: data representations (e.g. databases) and data analysis (e.g. pattern recognition, statistical analysis). The wealth of different data sources, each describing a different aspect of the molecular system (such as gene expressions, locations on genome, physical binding partners etc.), have driven data representation approaches towards flat and flexible formats. In contrast, data analysis prefers structured multi-dimensional array-based data formats. We argue that this gap cannot be closed, but rather needs to be bridged through the use of novel query systems. As a solution, we introduce the tool IBIDAS, which allows one to easily handle not only tables, but also more complexly structured data, making it a flexible tool for the exploration and analysis of data. Another question concerns at what point different data sources need to be integrated. One strategy (`late' integration) is to analyse each data source separately, after which the results are integrated. This strategy however prevents the discovery of connections that transcend the individual data sources. The alternative `early' integration strategy, in which the data is first concatenated (i.e. as feature vector) before analysis, is however not always feasible when complex data types, such as DNA sequences, need to be taken into account. We advocate an `intermediate' data integration approach, in which each data source is first transformed into a suitable “kernel space”. In this space, data can be integrated in a straightforward manner. This can even be done in a non-linear fashion, after which the data can be analyzed together. We show the strength of such “kernel method” when combining data sources to predict interacting proteins. When combining similar data, we emphasize that one should consider this as a data integration problem too, instead of just concatenating the data. As an example, batch-effects can seriously affect the data distributions of gene expression experiments. These effects need to be resolved when analyzing these experiments jointly. One way to solve this is to normalize data before it is joined. We show that it is necessary to take into account as much information as possible about the way in which the data is created into a normalization scheme. By modeling the effects that deteriorate your data, seemingly uninformative data sets can become again a rich source of information. As an example we applied this to data sets that study the relationship between the transcriptome of stem cells and the effectivity of these cells in bone regeneration. Instead of integrating data for a single problem, one can also integrate data for a class of problems. We show that the machine learning concept can elegantly solve such integration problems. By making use of the similarities between the problem domains, learning parameters can be restricted. This approach was applied in the analysis of 'materiomics' data for the new TopoChip platform. Measurements that characterized the reactions of cells to individual material surfaces were noisy, making it difficult to adequately compare these surface effects. However, by taking into account the similarities between surfaces, and by integrating data across these similar surfaces, results were improved significantly. As an encompassing example of data integration we finally show how a combination of integration methods can be put together to link two other integration levels: pattern recognition and causal model inference. In this example, numerous data sources are being used to predict cause-effect relationships between genes in perturbation experiments. The used data sources describe various aspects of proteins, protein-protein interactions and protein-DNA interactions. These descriptions of the physical components of a cell are related to cause-effect interactions between the genes, in such a way that data from perturbation experiments is explained. We combine kernel-based integration methods with a method that constructs a causal model, showing that cause-effect relationships can be accurately predicted. Taken together, this thesis explores several data integration levels and approaches. Given the complexity of biology, we believe that data integration will become more and more essential in bioinformatics and that this dissertation only has set the first steps on this road.Intelligent SystemsElectrical Engineering, Mathematics and Computer Scienc
Multi-Scale Pattern Recognition for Image Classification and Segmentation
No full textScale is an important parameter of images. Different objects or image structures (e.g. edges and corners) can appear at different scales and each is meaningful only over a limited range of scales. Multi-scale analysis has been widely used in image processing and computer vision, serving as the basis for many high-level image analysis systems. One such high-level system is based on supervised learning as studied in pattern recognition and machine learning, which might take the results from multi-scale analysis as its input. Supervised learning defines a classifier to assign objects into different categories, and learns the classifier with some example objects whose category labels are known. A common characteristic of the current multi-scale analysis methods, however, is that they are designed without specific assumptions about the high-level image analysis systems. The problem is that, different tasks need images to be analysed at different scales, that is, they need different multi-scale analysis. For example, for the same image containing a person, small scales are needed if the problem is to segment the eyes, while large scales are needed when one wants to segment the person. In many applications, the task is defined only with some given example images and it is not known a priori the right scale to conduct analysis. This asks for multi-scale analysis frameworks which can adapt to the different tasks. The aim of this thesis is to study such adaptive multi-scale frameworks based on supervised learning. It focuses on three important aspects in multi-scale analysis: scale selection, scale invariance, and scale combining. Scale selection addresses the problem of choosing a right scale to detect an object or to analyse an image. Scale invariance is the ability to deal with objects appearing at arbitrary sizes. Scale combining concerns the combination of information from all scales. General learning frameworks are proposed for these three aspects. Examples are shown for image segmentation and classification problems. A learning-based scale selection method is proposed for supervised image segmentation. Supervised segmentation trains a classifier based on some given segmented images, which assigns the pixels of an image into different classes or segments. The input of the classifier is features extracted from a neighbourhood at each pixel, and the scale of this neighbourhood is a crucial parameter of the features. Scale is usually selected as the size of a certain image structure, which is, however, not necessarily the best for the segmentation task. Keeping this in mind, the selected scale for supervised segmentation is redefined as the one at which pixels from different classes are best separable. A general scale selection scheme is proposed, which relies on the classifier for segmentation to measure the class separability. Experiments are presented, which show that this scheme can indeed choose scales that are best for the segmentation problem and thus leads to significantly improved performance. Based on the proposed scale selection scheme, a scale-invariant classification framework is proposed for supervised image segmentation. This classifier can deal with images from arbitrary scales. Consequently, the same segmentation result will be obtained when an image is resized. The classifier is trained with image features from all scales, and thus able to handle images from any scales. To make the classifier not biased on particular scales, the right proportion of features from different scales is needed. Scale invariance of the classification is achieved with the proposed scale selection scheme in the testing phase, which finds the right scales for image structures of different sizes. A learning model closely related to the proposed scale-invariant classification is multiple-instance learning (MIL). MIL is a generalised supervised-learning framework that represents an object as a bag consisting of many feature vectors called instances. Only some of the instances in the bag are informative about the label of the object, while others share the same probability distribution for objects from different classes. In the training phase, only the labels of bags (not instances) are known, and a classifier is trained to separate bags into different classes. These characteristics make MIL fit well for multi-scale image analysis, as an object can be represented with a set of features from all scales and only features from some scales are informative. Features from other scales are uninformative as the object becomes too blurred or too small to be distinguished from other objects. Observing that MIL algorithms usually make effective use of only one, not all, informative instance in a bag, we propose a new MIL model to. A simple MIL classifier is obtained, which performs very well for numerous data sets in the experiments. Combining information from multiple scales is studied based on the dissimilarity representation. It has been recognised that information from more than one scale can be useful for image analysis and should be exploited for better performance. For learning-based image analysis, multi-scale information is usually combined by concatenating features from all scales, which typically creates an enormously high-dimensional feature vector and thus makes learning difficult. We use the dissimilarity representation as it enables to combine multi-scale information without increasing the dimensionality of the representation space. It represents an image with dissimilarities by comparing it with a set of reference images. Multi-scale information is exploited by computing dissimilarities at each scale and then combining these dissimilarities. Various rules are proposed and tested with real-world image classification problems. The results show that simple combining rules can already improve significantly upon the best result from the individual scales, and more adaptive rules, which exploit certain structures along the scale, can lead to even better results.Intelligent SystemsElectrical Engineering, Mathematics and Computer Scienc
Supervised classification and spatial dependency analysis in human cancer using high throughput data
No full textElectrical Engineering, Mathematics and Computer Scienc
Finding cancer genes in copy number data and insertional mutagenesis data
No full textCancer is a genetic disease. Step-wise alteration of genes that have a normal function in the cell can lead to the transformation of a healthy cell into a malignant cancer cell. Cancer genes provide several traits to the cell that allow it to become malignant. These traits have been researched for many years, and currently one knows quite well what has to change in a normal cell before a tumor can be formed. For example, cells must divide continuously, escape the immune system and cause the growth of new blood vessels among others. There are many genes that can cause these processes when deregulated, and each individual tumor alters a different combination of genes to acquire its tumorigenic traits. Knowing which combination of mutations was sustained by a tumor is important as this might make the tumor susceptible or resistant to certain treatments. There are many ways in which cancer genes get mutated. This thesis studies two ways in which cancer genes are mutated. The first way of mutation comes from gains and losses of gene DNA called DNA copy number alterations (CNAs). These alterations occur due to the fact that tumors generally lose their ability to correctly repair damage to their DNA. CNAs can alter the expression of cancer genes and thereby cause cancer. Not only cancer-related genes will be affected by CNAs, also non-related DNA can be damaged. The challenge is to separate the truly oncogenic CNAs from the non-oncogenic passenger CNAs, as these oncogenic CNAs point to novel cancer genes that can be new drug targets. This thesis introduces two methods of finding cancer genes by examining DNA copy number alterations. Multiple comparable tumor samples are used to detect regions in the DNA that are altered significantly more often than other regions, indicating that they are more important for tumor development and therefore probably causative. Analogously, a novel method is introduced to find pairwise regions in the DNA that are preferentially lost or gained together (co-occurring) or preferentially not together (mutually exclusive). It is shown that co-occurring CNAs primarily target genes that are highly similar in function. A detailed analysis of a group of three mutually exclusive CNAs made it possible to associate a novel function to a known cancer gene. The second source of mutations concerns insertions of viral or transposon DNA. These agents insert their DNA in the host genome, which can cause activation or inactivation of host genes. Occasionally they can perturb genes that allow the cell to acquire cancer-related traits. In the end these insertions will cause a tumor. By carefully examining the tumor DNA one is able to reconstruct which genes caused the cancer. Of course, not all insertions were instrumental in the development of the cancer, so also in this case the passenger events and truly causal events have to be separated. This thesis used a novel approach called Shear-Splink that allows determination of the relative number of integrations in a single tumor. Each tumor will present a variety of insertions, each with its own abundance. By examining this abundance it is possible to distinguish between insertions that happened early in tumor development (and are therefore highly abundant) and insertions that are simple passengers or only essential for a small number of cells in the tumor (who will be lowly abundant). In this thesis this has been applied to a study of mouse mammary tumor virus (MMTV), a retrovirus that causes breast cancer in mice through insertion of its DNA in the mouse genome. Results show that by examining the insertion abundance a model through which the tumor has developed can be recovered. Overall this thesis contributes to the analysis of tumor-causing events and especially to the determination of which combination of events is necessary to cause a tumor.Pattern Recognition and BioinformaticsElectrical Engineering, Mathematics and Computer Scienc
Gesture Recognition by Computer Vision: An Integral Approach
No full textThe fundamental objective of this Ph.D. thesis is to gain more insight into what is involved in the practical application of a computer vision system, when the conditions of use cannot be controlled completely. The basic assumption is that research on isolated aspects of computer vision often leads to `too' general solutions. That these solutions lack the robustness and accuracy, which could only be achieved by an integral approach of a specific application. Furthermore, an integral approach, and actually trying out a computer vision system in practice, can lead to new insights that can determine the direction of future research in computer vision. The application for the research in this thesis is automatic sign recognition for feedback in active learning with an electronic learning environment for sign language. The goal of this learning environment is to enlarge the vocabulary of deaf and hard of hearing children, between the age of 3 and 5, in order to facilitate in decreasing a delay in language development. The research has been focussed on a number of aspects that were assumed to have the most influence on the robustness of sign recognition. These were: tracking of movements, the extraction of relevant structure information from an image, skin color detection, including the third dimension of hand locations, dealing with variations of time as well as shape of a sign and reducing the required effort to teach the system to recognize a new sign. `Particle filtering' is a popular method to track hand movement. However, tests with the CONDENSATION algorithm show contradictions in dealing with different situations. When the motion is unpredictable (as is the case with tracking of human hands) a particle filter has difficulty to keep track of the object. It turns out that, under different conditions, different strategies are required to deal with this in the best possible way. Isophote properties can be used as local abstractions of an image. One advantage of isophote properties is that they are independent of image contrast. In experiments with face detection using isophote properties, the results are superior to using pixels, gradients, or the popular Haar features. Because face detection requires significant computational cost, and the methods involved are less suitable for detection of hands, it is appealing to detect these body parts on the basis of their color alone. Unfortunately, color behaves less predictable in practical situations, than can be described by a single light-reflection model. Deviations from physical models for reflection are caused by properties and settings of the camera that is used, but also by the combination of different light sources and reflections. By combining these uncertainties in a more general model, robustness can be obtained in unknown circumstances. Unfortunately, this generalization comes at the price of accuracy in more friendly conditions. To combine robustness with accuracy, we have proposed an adaptive chromatic model, which can use a small set of measurements to model variation of the color of a face, using a bi-modal piecewise linear model in the red/green/blue space. Sign language takes place in a three-dimensional space, while images only allow measurements in two dimensions. Therefore, we have used stereometry to convert the measured hand locations in the images from two cameras into three-dimensional positions of the hands in space. The experiments show that this richer information does indeed lead to an improvement in sign recognition. Alternatively, the perspective of a single wide-angle camera at a short distance turned out to achieve a comparable improvement. However, the disadvantage of the latter solution is a decreased robustness, because perspective depends highly on the location of a person relative to the camera. Using dynamic recognition methods, like ``Hidden Markov Models'' (HMM) or Statistical ``Dynamic Time Warping'' (SDTW) a sequence of measured features of a person can be recognized as a specific sign. These models are able to deal with differences in tempo, contrary to conventional methods of pattern recognition, which can only deal with a fixed set of features. However, one of the disadvantages of HMM and SDTW is that they assume that what is important for estimating time warping, is equally important to class recognition. Furthermore, they are based on the factorization of probabilities for different time points, preventing the modeling of dependencies between measurements at different time steps. For these reasons, we have proposed to separate time warping and classification into subsequent processing steps. Experiments show a significant improvement over HMM or SDTW alone. In practice, it is difficult to obtain many examples of signs from different persons, in order to train a recognition system. To make the system robust to small training sets, we let the system make use of sign classes that were already trained with many examples. Here, we assumed that, when a part of the new sign is very similar to a part of a learned sign, its variation can be modeled in the same way. With a single example as the training material, this generalizing system performed comparable to when five examples are used in the regular training method. From this thesis, it can be concluded that robustness is not only relevant for practical applications of computer vision, but also deserves a place in fundamental research. Combining vantage points from different disciplines, such as physics, machine learning, neuropsychology and human computer interaction, makes sure that all aspects of a computer vision process can be integrally taken into account. With this, more robust solutions can be obtained than with each of the disciplines separately.MediamaticsElectrical Engineering, Mathematics and Computer Scienc
Algorithms for sequence-based reverse metabolic engineering
No full textMicroorganisms are employed on a large scale in a variety of industrial production processes, such as production of food and ingredients, beverages, pharmaceutical compo- nents and fine- and bulk chemicals. Improving the behaviour of these microorganisms is of crucial importance in the design of efficient production processes. In this thesis we focus use directed evolution, which is a method to improve the phe- notype of a microorganism without knowing what to change in the DNA to obtain the desired effect. Improved strains are obtained through the natural evolution processes of variation and selection by enforcing a strong selective pressure in the laboratory. The mutations that occur during such short evolution can be determined by sequencing the DNA of the improved strains and comparing it to the starting strain. Finally, relating the relevant mutations to the enforced phenotype is called reverse engineering. We describe the reverse engineering of mutations in ADY2 enabling lactate transport across the cell wall and mutations in ACE2, which in combination with aneuploidy, could be related to the origin of multi-cellular growth in yeast. Next-generation sequencing was employed to reconstruct genomes and infer mutations. This technology has made tremendous progress since its introduction in 2008 and has become affordable for individual laboratories. Sequencing poses many challenges to the field of bioinformatics, since genomes are long and repetitive (e.g. yeasts have millions of base pairs and humans billions), but current sequencing technology can only read up to 100 to 1000 base pairs in sequence. The computational methods developed in this thesis enable the use of sequencing technology in industrial microbiology research. A starting strain in this field is the yeast Saccharomyces cerevisiae CEN.PK113-7D. To determine its genome sequence we devel- oped an algorithm that uses a Tabu-search that exploits previously sequenced genomes and heterogeneous sequencing data types. Additionally, two methods were developed to infer mutations between genomes in absence of a high-quality closely related reference genome. First, we describe an algorithm that estimates the number of times a DNA subsequence exists in the full genome using a Poisson mixture model. We used this method to infer copy number differences between aneuploid lager brewing yeasts. Second, we describe an algorithm that uses graph decomposition to find variation in metagenomes, which are a combination of many bacterial genomes within one sample and play an important role in industrial fermentation, such as in production of cheese and yoghurt.Intelligent SystemsElectrical Engineering, Mathematics and Computer Scienc
A Critical Perspective On Microarray Breast Cancer Gene Expression Profiling
No full textMicroarrays offer biologists an exciting tool that allows the simultaneous assessment of gene expression levels for thousands of genes at once. At the time of their inception, microarrays were hailed as the new dawn in cancer biology and oncology practice with the hope that within a decade diseases like breast cancer would be solved. Various high-profile publications showed the immense potential of this technique in breast cancer event prediction and breast cancer subtyping. From these studies it became clear that breast cancer at the molecular level is not a single disease, but comprises a heterogeneous set of subtypes associated with clear differences in gene expression patterns and clinical outcomes. However, as microarrays became more popular, it became apparent that the accurate analysis and interpretation of microarray data provided a plethora of unique challenges. From a biological, as well as a technical perspective microarray data is complex, while the high feature-to-sample ratio associated with microarray studies rendered many classic statistical procedures useless. To make matters worse, various publications emerged that showed severe stability problems in the model fits of early pilot studies and showed that these studies were often overly optimistic. As a result the reliability of microarray based experiments in general was openly questioned. Given the multitude of different factors which may or may not influence results it is clear that a proper evaluation of microarray breast cancer profiling is both crucial and challenging. This dissertation provides a number of carefully devised protocols, by which the influence of important sources of variation can be isolated, controlled and/or explicitly quantified, even in the absence of a gold standard. Instead of applying these protocols to data from small spike-in or dilution studies, they were applied to a large collection of real life breast cancer datasets of considerable size. Furthermore, we extensively studied breast cancer subtyping and the evaluation of subtype-specific predictors constructed on these, from both a practical and a theoretical perspective. This work shows that the evaluation of subtype-specific event prediction, based on divide and conquer schemes brings various new statistical challenges. For a variety of frequently encountered performance measures from machine learning several decompositions of the overall performance into subtype-specific performances are provided which show that the relation between subtype-specific and overall performance can be highly complex and counterintuitive. Furthermore, the experiments in this dissertation show that with modern processing techniques and a standardized approach it is possible to construct extremely stable subtyping schemes. However, the selected approach has a strong impact on the obtained results, suggesting that a stringent standardization of the methodologies used for subtyping is not sufficient for the consistent assignment of subtypes to individual patient samples. From these findings we conclude that the molecular subtypes of breast cancer are not sufficiently well understood and need further refinements.Intelligent SystemsElectrical Engineering, Mathematics and Computer Scienc
- …
