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

ISCB2014_abstract_book

ISCB 2014 Vienna, Austria • Abstracts - Oral Presentations 23Monday, 25th August 2014 • 14:00-15:30 Monday25thAugustTuesday26thAugustThursday28thAugustAuthorIndexPostersWednesday27thAugustSunday24thAugust C07.5 Power and sample size of trials with a partially nesting design for binary outcomes C Roberts1 , E Batistatou1 , S Roberts1 1 University of Manchester, Manchester, United Kingdom Introduction: Partial Nesting describes the situation where some subjects are in clusters and others are not. This design may arise in individually ran- domized trials of non-pharmacological interventions where patients are clustered in one treatment arm due to treatment but not in the compara- tor. We consider sample size and power for binary outcomes of trials of this design. Methods: Formulae for determining power and sample size on the scale of proportions, log-odds and using an arc-sine transformation are pre- sented. These are compared with empirical power estimated using an ad- justed test of proportions, a summary measures test, a logistic model with random intercept and a logistic GEE model. Empirical power is determined using simulation with 40,000 replications for each scenario and a range of study sizes and proportions appropriate for estimation of small to medium treatment effects. Results: There is some under- and over-estimation of the formula deter- mined power relative to empirical power up to a maximum of 6%. For methods on the scale of proportions increased variance of the test statis- tics inflates the type II error thereby reducing empirical power relative to the formula. Power determined using the arc-sine formula tends to per- form better than the proportions method. For logistic models, the empiri- cal power appears to be increased relative to the proposed formulae by small sample bias. Conclusions: All methods of calculating power gave approximations to empirical power adequate for many practical situations. There were nev- ertheless differences between methods of sample size and power deter- mination and between determined and empirical power.   C08 Prediction models: case studies C08.1 Multidimensional assessment of the predictive ability of a Trichotomous-outcome model in the NICHD Collaborative Pediatric Critical Care Research Network (CPCCRN) R Holubkov1 , MM Pollack2 , AE Clark1 , T Funai1 , JM Dean1 1 University of Utah School of Medicine, Salt Lake City, United States, 2 University of Arizona College of Medicine, Phoenix, United States The CPCCRN Trichotomous Outcome Prediction in Critical Care (TOPICC) study is applying generalized logistic regression to a cohort of 10,000 critically ill children, predicting three outcome states (dead, alive with new morbidity, or good outcome at hospital discharge). We have implemented various graphical and analytic approaches to describe the model´s diag- nostic performance over the three outcome levels. Three-dimensional plots of the ROC hypersurface are informative when rotated to show individual pairwise ROC curves on the surface borders; the hypersurface can also be “sliced” to show diagnostic ability for two outcomes conditional on a particular correct prediction rate for the third. The Volume under the Surface (VUS) statistic may be intuitively presented as a brute-force integration of the ROC hypersurface, or as proportion of all possible one-patient-per-outcome triplets correctly classified (Mossman, MedDecisMaking 1999). However, clinicians do not have applied experi- ence with the VUS and its properties, such as its 1/6 rather than 1/2 value under a model without any discriminative ability. Example hypothetical ROC hypersurfaces under “pathological”, poor/perfect prediction settings were helpful when presenting our model´s performance. Extensions of the c-index summarizing performance over pairs of out- comes are more interpretable to clinicians than VUS as numerical sum- maries of model performance. For example, we have found the ordinal c-index of van Calster et al (BiometricalJ 2012) to be a useful integration of VUS and pairwise approaches. We will describe how the above approaches, and setting-specific modifi- cations, were used in TOPICC to summarize diagnostic ability of our three- level outcome model. C08.2 Developing dynamic prediction models for acute diseases KL Phung1 , M Wolbers1,2 1 Oxford University Clinical Research Unit, Ho Chi Minh City, Viet Nam, 2 Nuffield Department of Medicine, University of Oxford, Oxford, United Kingdom Dynamic prediction models which incorporate longitudinal covariate data and allow for temporal updating of predictions are increasingly popular. Unlike models based only on baseline information, such models use data efficiently and are able to capture the dynamic evolution of the disease in each individual. Several methods for developing such models have been proposed including extensions of classical regression models (Cox regres- sion with time-dependent covariates, landmarking) and joint models. These approaches have been developed and applied mainly in chronic diseases, such as cancer or cardiovascular diseases, where the disease progresses slowly over a long time period. However, dynamic and updat- able prediction models are also very attractive to physicians working in the acute setting, where diseases evolve rapidly over a short period of a few hours or days. We will summarize differences between chronic and acute diseases with respect to the types of outcomes of primary interest, expected associa- tions between longitudinal predictors and outcomes, and the amount of longitudinal data available. We then discuss implications of these dif- ferences for the development of dynamic prediction models. Our recom- mendations for dynamic modeling in the setting of acute diseases will be illustrated with a dataset of 2614 hospitalized children with dengue infec- tion, an acute disease commonly seen in tropical regions, where the out- come of interest is the occurrence of shock. We will compare several pos- sible modeling approaches and discuss how time-dependent accuracy measures and reclassification methods can be used to compare dynamic models and models based on baseline information only. C08.3 A patient-specific predicting tool for functional recovery after stroke A Douiri1 , JJ Grace1 , C McKevitt1 , AG Rudd1 , CDA Wolfe1 1 King’s College London, London, United Kingdom Predicting recovery over time at various stages of rehabilitation after stroke could potentially allow for sequential monitoring of patients and early identification and management of patients with poor recovery. This study aims to develop and validate a patient-specific tool for pre- dicting recovery trajectories post-stroke. Data from 495 patients after first-ever stroke between 2002-2004 were determined from the ongoing population-based South London Stroke Register for the model develop- ment. Functional recovery was assessed using Barthel Index (BI) at 1, 2, 3, 4, 6, 8, 12, 26 and 52 weeks post-stroke. The selected model incorporates reliable prognostic factors which are prevalent in stroke and which have good clinical accessibility. Penalized iteratively reweighted nonlinear least squares for generalized linear mixed models were adapted to develop re- covery curve models. The internal prediction error of the model was as- sessed using leave-one-out cross-validation. Temporal validations were conducted using two different cohorts between the period of 2005-2011 to evaluate the overall accuracy of the recovery curves, and to assess the

Pages Overview