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

ISCB2014_abstract_book

126 ISCB 2014 Vienna, Austria • Abstracts - Poster PresentationsWednesday, 27th August 2014 • 11:00-11:30 Monday25thAugustTuesday26thAugustThursday28thAugustAuthorIndexPostersWednesday27thAugustSunday24thAugust prognostic variables as related to outcome. This seems less expensive, less time consuming, and less specialized measurements and may prove use- ful in developing new therapeutic strategies and approaches. P3.5.145 A clinical diagnostic model using biomarkers and clinical characteristics for the identification of MODY patients BM Shields1 , T Mcdonald2 , K Owen3 , M Malecki4 , R Besser1 , A Jones1 , S Ellard2 , W Henley1 , A Hattersley1 1 University of Exeter Medical School, Exeter, United Kingdom, 2 Royal Devon and Exeter NHS Foundation Trust, Exeter, United Kingdom, 3 University of Oxford, Oxford, United Kingdom, 4 Jagiellonian University Medical College, Krakow, Poland   Aims/Objectives: Maturity-onset diabetes of the young (MODY) is a ge- netic form of diabetes, often misdiagnosed as Type 1 or Type 2 diabetes, resulting in inappropriate treatment. A diagnostic model to determine a patient’s probability of MODY based on clinical features (www.diabetes- genes.org) has been useful in determining which patients should receive expensive genetic testing. Biomarkers can also help identify MODY. We aimed to determine whether combining clinical features with biomarkers provides the best diagnostic model. Methods: Type 1 biomarkers (C-peptide and islet autoantibodies) and Type 2 biomarkers (hsCRP and HDL-cholesterol) were measured in 144 Type 1, 209 Type 2, and 172 MODY patients. Logistic regression mod- els were developed using the clinical features from the original model and biomarker results, with adjustment for the prevalence of MODY. Discriminative ability and classification improvement was assessed using ROC curves and net reclassification improvement (NRI). Results: In patients insulin treated from diagnosis, the addition of Type 1 biomarkers improved discrimination over clinical features alone (c-statis- tic=0.977 v 0.914, p=0.0004); NRI=1.57 (95% CI 1.29,1.85), p<0.0001). In pa- tients not insulin treated from diagnosis, Type 2 biomarkers made a small improvement in discrimination (c=0.986 v 0.981, p=0.06; NRI=0.80 (95% CI 0.56,1.04), p<0.0001). Both models had low prediction errors on jackknife crossvalidation (<6%) and good model fit. Conclusion/Summary: Excellent discrimination between MODY andType 1/Type 2 diabetes is achieved when biomarkers are used in combination with clinical characteristics. The addition of these markers to the online model will improve referrals sent for genetic testing for MODY. P3.5.173 Comparing different methods to develop prediction models for polytomous outcomes K Van Hoorde1,2 , K De Raedt1 , D Timmerman3,4 , S Van Huffel1,2 , B Van Calster3 1 KU Leuven Dept of Electrical Engineering / ESAT-STADIUS, Leuven, Belgium, 2 iMinds Medical Information Technologies, Leuven, Belgium, 3 KU Leuven Dept of Development & Regeneration, Leuven, Belgium, 4 University Hospitals Leuven Dept of Obstetrics & Gynecology, Leuven, Belgium   When diagnostic problems have a polytomous endpoint (i.e. more than two outcome categories) polytomous risk estimation can be useful al- though not common in clinical practice. For example, available models for ovarian tumor diagnosis simply predict the risk of malignancy, yet border- line malignancies often require less invasive fertility-sparing procedures than invasive cancers. The standard approach for polytomous outcomes is the multinomial lo- gistic regression (MLR) model. We compared MLR to four approaches that involve a combination of dichotomous logistic regression (LR) models to predict the risk that an ovarian tumor is benign, borderline, or invasive: (1) sequential dichotomous models, (2) combining models for each pair of categories, (3) individualized LR, and (4) combining models for each out- come category vs the rest. From a dataset including 5912 women and eleven predictor variables we drew a development and a validation set. The size of the development set was varied to have 5, 10, or 20 events per variable. Models were validated in terms of discrimination using the Polytomous Discrimination Index. Results for 300 development/validation set draws were averaged. Combining one-vs-rest models yielded the lowest discrimination between categories. The reason could be that the pooling of categories in a rest group obscures differences between individual categories. For this rea- son, the performance of sequential dichotomous models depended on the chosen sequence: borderline vs rest followed by benign vs invasive was the worst sequence. As expected apparent performance was overop- timistic when EPV=5. With EPV=20 optimism was notably smaller yet not negligible.   P3.5.174 Assessing the influence of case-mix heterogeneity on the discriminative ability of a risk model: the model-based concordance-index D van Klaveren1 , M Gönen2 , EW Steyerberg1 , Y Vergouwe1 1 Department of Public Health, Erasmus MC, Rotterdam, The Netherlands, 2 Epidemiology & Biostatistics, Memorial Sloan Kettering, New York, United States   The discriminative ability of a risk model in a new population depends on the validity of the regression coefficients, but also on the heterogeneity of the case-mix. We propose a model-based concordance-index (mbc- index) that assumes correct regression coefficients, to assess the influence of case-mix heterogeneity. We aimed to study the behaviour of the mbc- index in external validation data. We compared three concordance measures in a simulation study: Harrell’s c-index; the mbc-index, i.e. an extension to binary outcomes of Gönen and Heller’s censoring-robust concordance measure for time-to-event outcomes; and a previously proposed case-mix corrected c-index based on resampled outcomes. We first generated hypothetical samples of pa- tients (n=400 or n=1,600) for fitting regression models. To mimic different external validation settings, we simulated new patient data with different case-mix heterogeneity and different true regression coefficients. We per- formed the simulations (10.000 replications per setting) for binary and for time-to-event outcomes. As expected, the mbc-index was identical to the case-mix corrected c-index for binary outcomes. The mbc-index remained stable at the true value with increasing proportions of censoring of time- to-event outcomes, while the case-mix corrected c-index increased unfa- vourably. The mbc-index was further illustrated in two clinical examples. Comparison of the mbc-index in a new population with concordance measures in the model development population straightforwardly quan- tifies the loss (or gain) in a model’s discriminative ability due to less (or more) case-mix heterogeneity. The mbc-index is robust to censoring of time-to-event outcomes, in contrast to the previously proposed case-mix corrected c-index.

Pages Overview