Julia Schiffner

Relationship between anti-Müllerian hormone and antral follicle count across the menstrual cycle using the Beckman Coulter Access assay in comparison with Gen II manual assay.

Abstract Background: The study aim was to validate Beckman Coulter’s fully automated Access Immunoassay System (BC Access assay) for anti-Müllerian hormone (AMH) and compare it with Beckman Coulter’s Modified Manual Generation II assay (BC Mod Gen II), with regard to cycle AMH fluctuations and antral follicle counts. Methods: During one complete menstrual cycle, transvaginal ultrasound was performed on regularly menstruating women (n=39; 18–40years) every 2 days until the dominant ovarian follicle reached 16mm, then daily until observed ovulation; blood samples were collected throughout the cycle. Number and size of antral follicles was determined and AMH levels measured using both assays. Results: AMH levels measured by the BC Access assay vary over ovulatory menstrual cycles, with a statistically significant pre-ovulatory decrease from –5 to +2 days around objective ovulation. Mean luteal AMH levels were significantly lower (–7.99%) than mean follicular levels but increased again towards the end of the luteal phase. Antral follicle count can be estimated from AMH (ng/mL, BC Access assay) concentrations on any follicular phase day. BC Access assay-obtained AMH values are considerably lower compared with the BC Mod Gen II assay (–19% on average); conversion equation: AMH BC Access (ng/mL)=0.85 [AMH BC Mod Gen II (ng/mL)]0.95. Conclusions: AMH levels vary throughout the cycle, independently of assay utilised. A formula can be used to convert BC Access assay-obtained AMH levels to BC Mod Gen II values. The number of antral follicles can be consistently estimated from pre-ovulatory AMH levels using either assay.

D-optimal plans for variable selection in data bases

This paper is based on an article of Pumplun et al. (2005a) that investigates the use of Design of Experiments in data bases in order to select variables that are relevant for classification in situations where a sufficient number of measurements of the explanatory variables is available, but measuring the class label is hard, e. g. expensive or time-consuming. Pumplun et al. searched for D-optimal designs in existing data sets by means of a genetic algorithm and assessed variable importance based on the found plans. If the design matrix is standardized these D-optimal plans are almost orthogonal and the explanatory variables are nearly uncorrelated. Thus Pumplun et al. expected that their importance for discrimination can be judged independently of each other. In a simulation study Pumplun et al. applied this approach in combination with five classification methods to eight data sets and the obtained error rates were compared with those resulting from variable selection on the basis of the complete data sets. Based on the D-optimal plans in some cases considerably lower error rates were achieved. Although Pumplun et al. (2005a) obtained some promising results, it was not clear for different reasons if D-optimality actually is beneficial for variable selection. For example, D-efficiency and orthogonality of the resulting plans were not investigated and a comparison with variable selection based on random samples of observations of the same size as the D-optimal plans was missing. In this paper we extend the simulation study of Pumplun et al. (2005a) in order to verify their results and as basis for further research in this field. Moreover, in Pumplun et al. D-optimal plans are only used for data preprocessing, that is variable selection. The classification models are estimated on the whole data set in order to assess the effects of D-optimality on variable selection separately. Since the number of measurements of the class label in fact is limited one would normally employ the same observations that were used for variable selection for learning, too. For this reason in our simulation study the appropriateness of D-optimal plans for training classification methods is additionally investigated. It turned out that in general in terms of the error rate there is no difference between variable selection on the basis of D-optimal plans and variable selection on random samples. However, for training of linear classification methods D-optimal plans seem to be beneficial.

Localized Logistic Regression for Categorical Influential Factors

In localized logistic regression (cp. Loader, Local regression and likelihood, Springer, New York, 1999; Tutz and Binder, Statistics and Computing 15:155–166, 2005) at each target point where a prediction is required a logistic regression model is fitted locally. This is achieved by weighting the training observations in the log-likelihood based on their distances to the target observation. For interval-scaled influential factors these weights usually depend on Euclidean distances. This paper aims to combine localized logistic regression with dissimilarity measures more suitable for categorical data.

Tone onset detection using an auditory model

Onset detection is an important step for music transcription and other tasks frequently encountered in music processing. Although several approaches have been developed for this task, neither of them works well under all circumstances. In Bauer et al. (Einfluss der Musikinstrumente auf die Gute der Einsatzzeiterkennung, 2012) we investigated the influence of several factors like instrumentation on the accuracy of onset detection. In this work, this investigation is extended by a computational model of the human auditory periphery. Instead of the original signal the output of the simulated auditory nerve fibers is used. The main challenge here is combining the outputs of all auditory nerve fibers to one feature for onset detection. Different approaches are presented and compared. Our investigation shows that using the auditory model output leads to essential improvements of the onset detection rate for some instruments compared to previous results.

Benchmarking Classification Algorithms on High-Performance Computing Clusters

Comparing and benchmarking classification algorithms is an important topic in applied data analysis. Extensive and thorough studies of such a kind will produce a considerable computational burden and are therefore best delegated to high-performance computing clusters. We build upon our recently developed R packages BatchJobs (Map, Reduce and Filter operations from functional programming for clusters) and BatchExperiments (Parallelization and management of statistical experiments). Using these two packages, such experiments can now effectively and reproducibly be performed with minimal effort for the researcher. We present benchmarking results for standard classification algorithms and study the influence of pre-processing steps on their performance.

Comparison of classical and sequential design of experiments in note onset detection

Design of experiments is an established approach to parameter optimization of industrial processes. In many computer applications however it is usual to optimize the parameters via genetic algorithms. The main idea of this work is to apply design of experiment’s techniques to the optimization of computer processes. The major problem here is finding a compromise between model validity and costs, which increase with the number of experiments. The second relevant problem is choosing an appropriate model, which describes the relationship between parameters and target values. One of the recent approaches here is model combination. In this paper a musical note onset detection algorithm will be optimized using design of experiments. The optimal algorithm parameter setting is sought in order to get the best onset detection accuracy. We try different design strategies including classical and sequential designs and compare several model combination strategies.

Bias-Variance Analysis of Local Classification Methods

In recent years an increasing amount of so called local classification methods has been developed. Local approaches to classification are not new. Well-known examples are the k nearest neighbors method and classification trees (e.g. CART). However, the term ‘local’ is usually used without further explanation of its particular meaning, we neither know which properties local methods have nor for which types of classification problems they may be beneficial. In order to address these problems we conduct a benchmark study. Based on 26 artificial and real-world data sets selected local and global classification methods are analyzed in terms of the bias-variance decomposition of the misclassification rate. The results support our intuition that local methods exhibit lower bias compared to global counterparts. This reduction comes at the price of an only slightly increased variance such that the error rate in total may be improved.

Comparison of parameter optimization techniques for a music tone onset detection algorithm

Design of experiments is an established approach to parameter optimization for industrial processes. In many computer applications, however, it is usual to optimize the parameters via genetic algorithms or, recently, via sequential parameter optimization techniques. The main idea of this work is to analyse and compare parameter optimization approaches which are usually applied in industry with those applied for computer optimization tasks using the example of a tone onset detection algorithm. The optimal algorithm parameter setting is sought in order to get the best onset detection accuracy. We vary in our work essential options of the parameter optimization strategies like size and constitution of the initial designs in order to assess their in uence on the evaluation results. Furthermore we test how the instrumentation and the tempo of music pieces a ect the optimal parameter setting of the onset detection algorithm.

Local Analysis of SNP Data

SNP association studies investigate the relationship between complex diseases and one’s genetic predisposition through Single Nucleotide Poly- morphisms. The studies provide the analyst with a wealth of data and lots of challenges as the moderate to small risk changes are hard to detect and, moreover, the interest focusses not on the identification of single influential SNPs, but of (high-order) SNP interactions. Thus, the studies usually contain more variables than observations. An additional problem arises as there might be alternative ways of developing a disease. To face the challenges of high dimension, interaction effects and local differences, we use associative classification and localised logistic regression to classify the observations into cases and controls. These methods contain great potential for the local analysis of SNP data as applications to both simulated and real-world whole-genome data show.

Comparison of Local Classification Methods

In this paper four local classification methods are described and their statistical properties in the case of local data generating processes (LDGPs) are compared. In order to systematically compare the local methods and LDA as global standard technique, they are applied to a variety of situations which are simulated by experimental design. This way, it is possible to identify characteristics of the data that influence the classification performances of individual methods. For the simulated data sets the local methods on the average yield lower error rates than LDA. Additionally, based on the estimated effects of the influencing factors, groups of similar methods are found and the differences between these groups are revealed. Furthermore, it is possible to recommend certain methods for special data structures.