Please activate JavaScript!
Please install Adobe Flash Player, click here for download

ISCB2014_abstract_book

ISCB 2014 Vienna, Austria • Abstracts - Oral Presentations 27Monday, 25th August 2014 • 14:00-15:30 Monday25thAugustTuesday26thAugustThursday28thAugustAuthorIndexPostersWednesday27thAugustSunday24thAugust C11 Functional data analysis C11.1 Functional data analysis of temporal glucose curves compared with gold standard measurements of insulin sensitivity and beta-cell function KF Frøslie1,2 , K Færch3 , J Røislien1 1 University of Oslo, Oslo, Norway, 2 Oslo University Hospital, Oslo, Norway, 3 Steno Diabetes Center A/S, Gentofte, Denmark A recent study of blood glucose curves in pregnant women, applying functional principal component analysis (FPCA) of oral glucose tolerance test (OGTT) data, identified the general glucose level and the timing of the glucose peak as the two main modes of temporal variation between indi- viduals. The latter was a significant predictor of gestational diabetes later in pregnancy. Glucose curve trajectories may also reflect distinct physi- ological processes. The aim of the present study was to extract glucose curve characteristics from OGTTs in healthy, non-pregnant individuals, and compare these to gold standard measurements of physiological features. OGTT glucose data from 20 participants in the Inter99 study were ana- lysed by FPCA. Glucose curve characteristics were compared with mea- surements of insulin sensitivity and beta-cell function obtained from eu- glycaemic hyperinsulinemic clamps and intravenous glucose tolerance tests. The first two functional principal components (FPCs) explained 65% and 19%, respectively, of the variance in glucose curves. The first FPC (FPC1) represented the general postprandial glucose levels during the OGTT and the timing of the postprandial glucose peak. High FPC1 scores (high levels and late peak) were associated with low insulin sensitivity (r=-0.43) and low first-phase insulin response (r=-0.41). FPC2 represented the “flatness” of the curve, with high scores (flat curve, higher than average postprandial glucose values in the later half of the OGTT) being associated with high first-phase insulin response (r=0.43), but low insulin sensitivity (r=-0.22). The curve characteristics derived by FPCA adds to the understanding of the various physiological mechanisms that are mirrored in glucose curves.   C11.2 Unsupervised classification of functional data based on covariance structures F Ieva1 , AM Paganoni2 , N Tarabelloni2 1 Università degli studi di Milano, Milano, Italy, 2 Politecnico di Milano, Milano, Italy We address the problem of performing an unsupervised classification of samples from two populations that differ in terms of variability, rather than location. This is of interest in biomedical context, where the dichot- omy between physiological and pathological features often shows an in- teresting pattern in change of variability. To this aim, we recur to a proper notion of distance between covariance structures, which is the basis of our new classification method. We formulate it in an abstract setting, suitable for both standard finite-dimensional, and functional data analysis. In order to detect the two groups of samples, we search, inside the set of possible recombinations of samples, the labelling that maximizes the distance be- tween the induced covariance structures, under the hypothesis that the true subdivision realizes this maximum. We identify the members of the two estimated populations as those who fulfil the two maximally-distant covariance structures. Special care is taken of traslating this procedure into a heuristic algorithm, able to restrain the explosive complexity of a naive, greedy search. We also propose an exploratory technique to early assess the successful application of our method. This is based on informa- tion drawn from an approximate permutation test performed on data, in which our target is the distribution of the covariances´ distance under dif- ferent recombinations of samples. We also point out that, when analysing functional samples, best results are obtained by improving the estimation of covariances with a shrinkage approach.We test our method first on syn- thetic data, and then on real data.   C11.3 Functional data analysis in radiobiology and radiation epidemiology MA Benadjaoud1,2,3 , H Cardot4 , F de Vathaire1,2,3 1 Radiation Epidemiology Group CESP - INSERM 1018, Villejuif, France, 2 Institut Gustave Roussy, Villejuif, France, 3 Université Paris Sud, Le Kremlin-Bicêtre, France, 4 Institut de Mathématiques de Bourgogne, Dijon, France In many fields, each observation consists of discrete measurements col- lected over a continuum. These data points can be thought as discrete sampling from an underlying smooth process.The functional data analysis (FDA) aims to analyze data providing by curves, (hyper)surfaces…etc as opposed to a point or a finite-dimensional vector and extract additional information contained in the functions and their derivatives, not normally available through traditional methods. We propose to illustrate the performance of the FDA technics through two applications from the ionizing radiations field. The first application concerns the normal tissue complication probability (NTCP) modeling in external radiotherapy. The most widely NTCP model is the Lyman-Kutcher-Burman (LKB) model based on the reducing of the dose distribution to an equivalent uniform dose (EUD) using a power law. In this example, we propose a non-parametric NTCP model where the weights dose values in EUD are estimated flexibly using logistic model based on the scores of functional principal component analysis conduct- ed on the estimated dose probability density functions. The estimated parameter function leads to a better understanding of the dose-volume effect. The second application focus on the analysis of the temporal response of H2AX, an important marker of a dangerous radio-induced DNA lesion: the double strand breaks (DSBs). In these data, each subject is a cluster wherein three H2AX temporal curves are measured. We used a multilevel functional principal component analysis to quantify intra-and-inter subject variability and investigate the association between the H2AX temporal dynamics and the risk of radiation-induced second malignancies. C11.4 Use of finite mixture models for dietary patterns analysis N Sauvageot1 , A Alkerwi1 , A Albert2 , M Guillaume2 1 CRP Santé, Luxembourg, Luxembourg, 2 University of Liege, Liege, Belgium Free-living individuals have multifaceted diets and consume foods in nu- merous combinations. The effect of the overall diet beyond that of single foods can be studied with dietary pattern analysis. Furthermore, the di- etary pattern approach reduces data-dimensionality and alleviates prob- lems of model over-fitting and residual confounding that occur with the statistical analysis of many food items. Most recent dietary pattern analyses have used factor and cluster analysis. We describe a finite mixture modelling (FMM) approach for dietary pat- tern analysis and show its advantages over previous ones. First, FMM allows estimating pattern prevalence directly from the model parameters as opposed to the subjective joint classification of the factors. Moreover, in contrast to ‚hard‘ assignment of clustering methods, FMM also produces posterior cluster membership probabilities for each subject providing measures of uncertainty of the associated classification. Second, it allows problems in determining the number of clusters and

Pages Overview