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

ISCB2014_abstract_book

24 ISCB 2014 Vienna, Austria • Abstracts - Oral PresentationsMonday, 25th August 2014 • 14:00-15:30 Monday25thAugustTuesday26thAugustThursday28thAugustAuthorIndexPostersWednesday27thAugustSunday24thAugust prognostic accuracy and utility for the classification of patients with poor recovery at 3 and 12 months. Predictive accuracy of the recovery curves was acceptable, with a root mean square deviation of 3.32 BI points. The prognostic accuracies to predict poor recovery at both 3 and 12 months were also satisfactory (94%, 95% CI [90.9-96.9] and 89%, 95% CI [84.4%- 93.2] respectively). This presentation will describe and discuss the differ- ent statistical and computational methodologies adopted in the develop- ment and validation of the final recovery curve model.   C08.4 A tree based model for thyroid cancer prognostication M Banerjee1 , D Muenz1 , M Haymart1 1 University of Michigan, Ann Arbor, United States Thyroid cancer is becoming an increasingly common cancer and yet little is known about the prognostic factors associated with survival. Controversy also exists over appropriate treatment for thyroid cancer. Prognostic models are needed to determine correlates of overall survival and identify subgroups of patients with poor prognosis who may benefit from earlier therapeutic intervention. In this talk we present a tree-based model for thyroid cancer prognostication using data from the US National Cancer Database. Trees are conceptually simple yet powerful, and are be- ing increasingly used in biomedical studies for analyzing censored sur- vival data where the primary goal is prognostication of patients. To gain accuracy in prediction and address instability in a single tree, an ensemble of trees is typically grown and the predictions are averaged across the trees in the ensemble. In this talk, we present a methodology for identi- fying the most representative tree from the ensemble based on several tree distance metrics. Out of bag error based on the cumulative hazard estimate is computed for the representative tree. For the thyroid cancer data, the representative tree from the ensemble was able to identify four distinct prognostic groups, defined by age, gender, lymph node involve- ment, tumor size, and metastasis status. Five year survival rates in these groups ranged between 64% and 99%.The prognostic groups derived can provide guidance for patient management, clinical trial design, and future treatment policy. The representative tree itself can be used as a decision making tool in the clinical setting.   C08.5 Stratified weighted regression for subgroup signatures from prognostic models with molecular data V Weyer1 , H Binder1 1 University Medical Center Mainz, IMBEI, Mainz, Germany In the analysis of high dimensional molecular data with time-to-event endpoints, developing subgroup signatures is one way for taking indi- vidual heterogeneity into account. We propose an alternative method to subgroup analysis based on weighted regression. As an application we consider RNA-Seq data from acute myeloid leukemia (AML) patients with different cytogenetic risk profiles, where a survival gene expression sig- nature for cytogenetic low risk patients is to be developed while taking information from high risk patients into account. For signature develop- ment we use automated variable selection by componentwise boosting with a weighted Cox regression partial log-likelihood, allowing for differ- ent baseline hazards in the different subgroups by strata. Thus, the partial likelihood takes the form of a weighted product of terms, one for each stratum. Further we propose two approaches for automatically choos- ing the weights which are based on resampling methods. For evaluation, we consider model stability and prediction performance. In a simulation study and in the real data set of AML patients the stratified approach is compared to an unstratified variant. The stratified approach is seen to per- form well in terms of identifying of important factors as well as with re- spect to prediction performance. Automated selection of weights is seen to adequately identify situations where information from the respective other subgroup is useful. Thus stratified weighted regression with auto- mated weight selection seems promising when subgroups need to be taken into account for signature development in a time-to-event setting.   C09 Latent variable methods C09.1 Treatment effect estimation in latent variable models with structural misspecification A Kifley1,2 1 University of Sydney, Sydney, Australia, 2 Macquarie University, Sydney, Australia Latent variable models are sensitive to misspecifications of the nature of the relationships between observed variables and unobserved underly- ing latent variables that may be of primary interest. However misspecifica- tions of this type are likely to occur in practice. In this study, we evaluate the performance of reflective latent variable models in estimating treat- ment or exposure effects when presented with observed item measures that include a mixture of formative and reflective item types. Reflective models assume that observed items serve merely as indicators of the sta- tus of the underlying latent variables, while formative items in fact affect the latent variables directly. We explore the sensitivity of global treatment or exposure effect estimates to levels of direct, indirect and mediated ef- fects of treatment, under a range of different conditions. We find a ten- dency toward overestimation of treatment effects by the reflective model if, in truth, the treatment affects formative items present in the assessment with little or no direct treatment effect on the latent variable of interest. We find a weaker tendency toward underestimation of treatment effects by the reflective model if, in truth, the treatment directly affects the latent variable but does not affect potentially formative items that are included. Problems in estimation are substantially greater if the assessment is pre- dominantly formative and the formative items share strong similarities with each other. Our simulation studies were motivated by issues arising in analysis of health-related quality of life data, but are relevant to many other applications of latent variable modelling. C09.2 Joint modeling of longitudinal binary and continuous responses E Kurum1 , R Li2 , S Shiffman3 , W Yao4 1 Istanbul Medeniyet University, Istanbul, Turkey, 2 The Pennsylvania State University, University Park, United States, 3 University of Pittsburgh, Pittsburgh, United States, 4 Kansas State University, Manhattan, United States Motivated by an empirical analysis of ecological momentary assessment data (EMA) collected in a smoking cessation study, we propose a joint modeling technique for estimating the time-varying association between two intensively measured longitudinal responses: a continuous one and a binary one. A major challenge in joint modeling these responses is the lack of a mul- tivariate distribution. We suggest introducing a normal latent variable underlying the binary response and factorizing the model into two com- ponents: a marginal model for the continuous response, and a conditional model for the binary response given the continuous response.We develop a two-stage estimation procedure and establish the asymptotic normality of the resulting estimators. We also derived the standard error formulas

Pages Overview