Fakultät für Mathematik, Informatik und Statistik - Digitale Hochschulschriften der LMU - Teil 01/02

This thesis is concerned with the development of flexible continuous-time survival models based on the accelerated failure time (AFT) model for the survival time and the Cox relative risk (CRR) model for the hazard rate. The flexibility concerns on the one hand the extension of the predictor to take into account simultaneously for a variety of different forms of covariate effects. On the other hand, the often too restrictive parametric assumptions about the survival distribution are replaced by semiparametric approaches that allow very flexible shapes of survival distribution. We use the Bayesian methodology for inference. The arising problems, like e. g. the penalization of high-dimensional linear covariate effects, the smoothing of nonlinear effects as well as the smoothing of the baseline survival distribution, are solved with the application of regularization priors tailored for the respective demand.

The considered expansion of the two survival model classes enables to deal with various challenges arising in practical analysis of survival data. For example the models can deal with high-dimensional feature spaces (e. g. gene expression data), they facilitate feature selection from the whole set or a subset of the available covariates and enable the simultaneous modeling of any type of nonlinear covariate effects for covariates that should always be included in the model. The option of the nonlinear modeling of covariate effects as well as the semiparametric modeling of the survival time distribution enables furthermore also a visual inspection of the linearity assumptions about the covariate effects or accordingly parametric assumptions about the survival time distribution.

In this thesis it is shown, how the p>n paradigm, feature relevance, semiparametric inference for functional effect forms and the semiparametric inference for the survival distribution can be treated within a unified Bayesian framework. Due the option to control the amount of regularization of the considered priors for the linear regression coefficients, there is no need to distinguish conceptionally between the cases p<=n and p>n. To accomplish the desired regularization, the regression coefficients are associated with shrinkage, selection or smoothing priors. Since the utilized regularization priors all facilitate a hierarchical representation, the resulting modular prior structure, in combination with adequate independence assumptions for the prior parameters, enables to establish a unified framework and the possibility to construct efficient MCMC sampling schemes for joint shrinkage, selection and smoothing in flexible classes of survival models. The Bayesian formulation enables therefore the simultaneous estimation of all parameters involved in the models as well as prediction and uncertainty statements about model specification.

The presented methods are inspired from the flexible and general approach for structured additive regression (STAR) for responses from an exponential family and CRR-type survival models. Such systematic and flexible extensions are in general not available for AFT models. An aim of this work is to extend the class of AFT models in order to provide such a rich class of models as resulting from the STAR approach, where the main focus relies on the shrinkage of linear effects, the selection of covariates with linear effects together with the smoothing of nonlinear effects of continuous covariates as representative of a nonlinear modeling. Combined are in particular the Bayesian lasso, the Bayesian ridge and the Bayesian NMIG (a kind of spike-and-slab prior) approach to regularize the linear effects and the P-spline approach to regularize the smoothness of the nonlinear effects and the baseline survival time distribution. To model a flexible error distribution for the AFT model, the parametric assumption for the baseline error distribution is replaced by the assumption of a finite Gaussian mixture distribution. For the speci

Mobilität in Form des Transports von Waren und Personen ist ein wesentlicher Bestandteil unserer heutigen Gesellschaft, da diese einen enormen Einfluss auf die Wirtschaftlichkeit und das soziale Leben hat. Nichts verkörpert die Begriffe Individualität, Flexibilität und Freiheit mehr als das eigene Auto und ist - in der Masse - gleichzeitig deren größte Bedrohung. Insbesondere in Megastädten konzentrieren sich die mit dem Verkehr verbundenen Probleme, die neben Staus auch zu einer überlasteten Infrastruktur führen und erhebliche Konsequenzen für die Umwelt nach sich ziehen.

Im Rahmen dieser Arbeit werden einige Ansätze vorgestellt und deren technische Umsetzung erläutert. Aus Sicht der Benutzer werden Anwendungen zur Förderung des kollektiven und gemeinschaftlichen Transports sowie ein Ansatz zur gemeinschaftlichen Parkraumverwaltung präsentiert. Im Anschluss wird aus der Sicht der Mobilitätsanbieter ein kooperativer Ansatz für einen flexiblen und bedarfsorientierten Tür-zu-Tür Transportdienst beschrieben. Abschließend wird auf ein System zur gemeinschaftlichen Schadstoffüberwachung eingegangen, welches einerseits eine detaillierte Grundlage für Infrastrukturbetreiber und Stadtplaner bietet und andererseits als Basis für umweltsensitive Anwendungen genutzt werden kann.

Mit der Unterstützung von Informations- und Kommunikationstechnologien in Kombination mit mobilen Endgeräten sowie auf der Basis des gemeinschaftlichen Zusammenwirkens, leisten die entwickelten Anwendungen und Systeme damit einen Beitrag zur Förderung einer effizienten und nachhaltigen Mobilität in Megastädten.

Methods for constructing high-risk zones, which can be used in situations where a spatial point pattern has been observed incompletely, are introduced and evaluated with regard to unexploded bombs in federal properties in Germany.

Unexploded bombs from the Second World War represent a serious problem in Germany. It is desirable to search high-risk zones for unexploded bombs, but this causes high costs, so the search is usually restricted to carefully selected areas. If suitable aerial pictures of the area in question exist, statistical methods can be used to determine such zones by considering patterns of exploded bombs as realisations of spatial point processes. The patterns analysed in this thesis were provided by Oberfinanzdirektion Niedersachsen, which supports the removal of unexploded ordnance in federal properties in Germany. They were derived from aerial pictures taken by the Allies during and after World War II.

The main task consists of finding as small regions as possible containing as many unexploded bombs as possible. In this thesis, an approach based on the intensity function of the process is introduced: The high-risk zones consist of those parts of the observation window where the estimated intensity is largest, i.e. the estimated intensity function exceeds a cut-off value c. The cut-off value can be derived from the risk associated with the high-risk zone. This risk is defined as the probability that there are unexploded bombs outside the zone.

A competing approach for determining high-risk zones consists in using the union of discs around all exploded bombs as high-risk zone. The radius is chosen as a high quantile of the nearest-neighbour distance of the point pattern. In an evaluation procedure, both methods yield comparably good results, but the theoretical properties of the intensity-based high-risk zones are considerably better.

A further goal is to perform a risk assessment of the investigated area by estimating the probability that there are unexploded bombs outside the high-risk zone. This is especially important as the estimation of the intensity function is a crucial issue for the intensity-based method, so the risk cannot be determined exactly in advance. A procedure to calculate the risk is introduced. By using a bootstrap correction, it is possible to decide on acceptable risks and find the optimal, i.e. smallest, high-risk zone for a fixed probability that not all unexploded bombs are located inside the high-risk zone.

The consequences of clustering are investigated in a sensitivity analysis by exploiting the procedure for calculating the risk. Furthermore, different types of models which account for clustering are fitted to the data, classical cluster models as well as a mixture of bivariate normal distributions.

The goal of bioinformatics is to develop innovative and practical methods and algorithms for bio-

logical questions. In many cases, these questions are driven by new biotechnological techniques,

especially by genome and cell wide high throughput experiment studies.

In principle there are two approaches:

1. Reduction and abstraction of the question to a clearly defined optimization problem, which

can be solved with appropriate and efficient algorithms.

2. Development of context based methods, incorporating as much contextual knowledge as

possible in the algorithms, and derivation of practical solutions for relevant biological ques-

tions on the high-throughput data. These methods can be often supported by appropriate

software tools and visualizations, allowing for interactive evaluation of the results by ex-

perts.

Context based methods are often much more complex and require more involved algorithmic

techniques to get practical relevant and efficient solutions for real world problems, as in many

cases already the simplified abstraction of problems result in NP-hard problem instances. In

many cases, to solve these complex problems, one needs to employ efficient data structures and

heuristic search methods to solve clearly defined sub-problems using efficient (polynomial) op-

timization (such as dynamic programming, greedy, path- or tree-algorithms).

In this thesis, we present new methods and analyses addressing open questions of bioinformatics

from different contexts by incorporating the corresponding contextual knowledge.

The two main contexts in this thesis are the protein structure similarity context (Part I) and net-

work based interpretation of high-throughput data (Part II).

For the protein structure similarity context Part I we analyze the consistency of gold standard

structure classification systems and derive a consistent benchmark set usable for different ap-

plications. We introduce two methods (Vorolign, PPM) for the protein structure similarity recog-

nition problem, based on different features of the structures.

Derived from the idea and results of Vorolign, we introduce the concept of contact neighbor-

hood potential, aiming to improve the results of protein fold recognition and threading.

For the re-scoring problem of predicted structure models we introduce the method Vorescore,

clearly improving the fold-recognition performance, and enabling the evaluation of the contact

neighborhood potential for structure prediction methods in general.

We introduce a contact consistent Vorolign variant ccVorolign further improving the structure

based fold recognition performance, and enabling direct optimization of the neighborhood po-

tential in the future. Due to the enforcement of contact-consistence, the ccVorolign method has

much higher computational complexity than the polynomial Vorolign method - the cost of com-

puting interpretable and consistent alignments.

Finally, we introduce a novel structural alignment method (PPM) enabling the explicit modeling

and handling of phenotypic plasticity in protein structures. We employ PPM for the analysis of

effects of alternative splicing on protein structures. With the help of PPM we test the hypothesis,

whether splice isoforms of the same protein can lead to protein structures with different folds

(fold transitions).

In Part II of the thesis we present methods generating and using context information for the

interpretation of high-throughput experiments.

For the generation of context information of molecular regulations we introduce novel textmin-

ing approaches extracting relations automatically from scientific publications.

In addition to the fast NER (named entity recognition) method (syngrep) we also present a novel,

fully ontology-based context-sensitive method (SynTree) allowing for the context-specific dis-

ambiguation of ambiguous synonyms and resulting in much better identification performance.

This context information is important f

Dynamic imaging series acquired in medical and biological research are often analyzed with the help of compartment models. Compartment models provide a parametric, nonlinear function of interpretable, kinetic parameters describing how some concentration of interest evolves over time. Aiming to estimate the kinetic parameters, this leads to a nonlinear regression problem. In many applications, the number of compartments needed in the model is not known from biological considerations but should be inferred from the data along with the kinetic parameters. As data from medical and biological experiments are often available in the form of images, the spatial data structure of the images has to be taken into account.

This thesis addresses the problem of parameter estimation and model selection in compartment models. Besides a penalized maximum likelihood based approach, several Bayesian approaches-including a hierarchical model with Gaussian Markov random field priors and a model state approach with flexible model dimension-are proposed and evaluated to accomplish this task. Existing methods are extended for parameter estimation and model selection in more complex compartment models. However, in nonlinear regression and, in particular, for more complex compartment models, redundancy issues may arise. This thesis analyzes difficulties arising due to redundancy issues and proposes several approaches to alleviate those redundancy issues by regularizing the parameter space.

The potential of the proposed estimation and model selection approaches is evaluated in simulation studies as well as for two in vivo imaging applications: a dynamic contrast enhanced magnetic resonance imaging (DCE-MRI) study on breast cancer and a study on the binding behavior of molecules in living cell nuclei observed in a fluorescence recovery after photobleaching (FRAP) experiment.

Direct touch input on interactive surfaces has become a predominating standard for the manipulation of digital information in our everyday lives. However, compared to our rich interchange with the physical world, the interaction with touch-based systems is limited in terms of flexibility of input and expressiveness of output. Particularly, the lack of tactile feedback greatly reduces the general usability of a touch-based system and hinders from a productive entanglement of the virtual information with the physical world.

This thesis proposes remote tactile feedback as a novel method to provide programmed tactile stimuli supporting direct touch interactions. The overall principle is to spatially decouple the location of touch input (e.g. fingertip or hand) and the location of the tactile sensation on the user's body (e.g. forearm or back). Remote tactile feedback is an alternative concept which avoids particular challenges of existing approaches. Moreover, the principle provides inherent characteristics which can accommodate for the requirements of current and future touch interfaces.

To define the design space, the thesis provides a structured overview of current forms of touch surfaces and identifies trends towards non-planar and non-rigid forms with more versatile input mechanisms. Furthermore, a classification highlights limitations of the current methods to generate tactile feedback on touch-based systems. The proposed notion of tactile sensory relocation is a form of sensory substitution. Underlying neurological and psychological principles corroborate the approach. Thus, characteristics of the human sense of touch and principles from sensory substitution help to create a technical and conceptual framework for remote tactile feedback.

Three consecutive user studies measure and compare the effects of both direct and remote tactile feedback on the performance and the subjective ratings of the user. Furthermore, the experiments investigate different body locations for the application of tactile stimuli. The results show high subjective preferences for tactile feedback, regardless of its type of application. Additionally, the data reveals no significant differences between the effects of direct and remote stimuli. The results back the feasibility of the approach and provide parameters for the design of stimuli and the effective use of the concept.

The main part of the thesis describes the systematical exploration and analysis of the inherent characteristics of remote tactile feedback. Four specific features of the principle are identified: (1) the simplification of the integration of cutaneous stimuli, (2) the transmission of proactive, reactive and detached feedback, (3) the increased expressiveness of tactile sensations and (4) the provision of tactile feedback during multi-touch. In each class, several prototypical remote tactile interfaces are used in evaluations to analyze the concept. For example, the PhantomStation utilizes psychophysical phenomena to reduce the number of single tactile actuators. An evaluation with the prototype compares standard actuator technologies with each other in order to enable simple and scalable implementations. The ThermalTouch prototype creates remote thermal stimuli to reproduce material characteristics on standard touchscreens. The results show a stable rate of virtual object discrimination based on remotely applied temperature profiles. The AutmotiveRTF system is implemented in a vehicle and supports the driver's input on the in-vehicle-infotainment system. A field study with the system focuses on evaluating the effects of proactive and reactive feedback on the user's performance.

The main contributions of the dissertation are: First, the thesis introduces the principle of remote tactile feedback and defines a design space for this approach as an alternative method to provide non-visual cues on interactive surfaces. Second, the thesis describes technical examples to rapidl

Aktuelle Methoden zur dynamischen Modellierung von biologischen Systemen sind für Benutzer ohne

mathematische Ausbildung oft wenig verständlich. Des Weiteren fehlen sehr oft genaue Daten und detailliertes Wissen über Konzentrationen, Reaktionskinetiken oder regulatorische Effekte. Daher erfordert eine computergestützte Modellierung eines biologischen Systems, mit Unsicherheiten und grober Information umzugehen, die durch qualitatives Wissen und natürlichsprachliche Beschreibungen zur Verfügung gestellt wird.

Der Autor schlägt einen neuen Ansatz vor, mit dem solche Beschränkungen überwunden werden können. Dazu wird eine Petri-Netz-basierte graphische Darstellung von Systemen mit einer leistungsstarken und dennoch intuitiven Fuzzy-Logik-basierten Modellierung verknüpft. Der Petri Netz und Fuzzy Logik (PNFL) Ansatz erlaubt eine natürlichsprachlich-basierte Beschreibung von biologischen Entitäten sowie eine Wenn-Dann-Regel-basierte Definition von Reaktionen. Beides kann einfach und direkt aus qualitativem Wissen abgeleitet werden. PNFL verbindet damit qualitatives Wissen und quantitative Modellierung.