Medical Research
Presenter: Martin Bland
When: Thursday, July 14, 2016 Time: 2:30 PM - 4:00 PM
Room: Salon C Carson Hall (Level 2)
Session Synopsis:
Clinical trial by television
In 2013, I was asked to help design and analyse some very small controlled clinical trials of popular health remedies for a television series, Health Freaks. In 2015, a different television series, Trust Me, Im a Doctor, followed this up with a large, uncontrolled trial of one of the remedies: duct tape for the treatment of warts and verrucae. I was again asked to assist. I shall describe and compare the problems and outcomes of each of these trials. I shall go on to discuss their possible implications for improving the public understanding of both clinical trial research and statistics.
Medical Research
Presenter: Laurent Briollais
When: Thursday, July 14, 2016 Time: 2:30 PM - 4:00 PM
Room: Salon C Carson Hall (Level 2)
Session Synopsis:
Markov-Renewal Multistate Model for Colorectal Cancer Screening Evaluation in Lynch Syndrome families
The evaluation of screening procedures for cancer detection in high-risk families raises many statistical challenges due to the complexity of disease and screening processes. We discuss the application of multi-state (MS) models in the context of screening by colonoscopy and sigmoidoscopy in Lynch Syndrome families, characterized by very high-risk of colorectal cancer (CRC) and other cancers. The states are defined based on individuals' screening results, whose values are measured intermittently at each screening. We propose a novel Markov-renewal MS model to assess how screening intervals affect different transition intensities and the probability of transitioning to CRC by assuming that the transition intensity functions are related to the elapsed time since the last screening visit. This model allows the transition intensities to depend on an individuals baseline characteristics and time-dependent covariates such as types of polyps. We discuss here the estimation of transition intensities, covariate effects, and transition probabilities for the proposed Markov-renewal model. An ascertainment-corrected likelihood is proposed to deal with the non-random sampling of families. Simulation study results indicate good performances in terms of bias and efficiency for a practical family sample size of 100. The methodology is applied to a series of 18 large LS families from Newfoundland harbouring a founder MSH2 gene mutation. Our main result shows that the probability of transitioning to CRC is reduced by half for an individual with a polyp detected vs. no polyp detected, provided that individuals are screened every 2 years. Keywords: Multi-state model; Markov-renewal process; Intermittent observation; Cancer screening; Lynch syndrome family; Colonoscopy
Medical Research
Presenter: Bruce Craig
When: Thursday, July 14, 2016 Time: 2:30 PM - 4:00 PM
Room: Salon C Carson Hall (Level 2)
Session Synopsis:
Assessing Inter-rater Agreement of Immunohistochemistry Scores
Compositional data are non-negative vectors whose elements sum to one. This type of data occurs in many research areas where the relative magnitudes between the vectors elements are of primary interest. For immunohistochemistry (IHC) analysis, pathologists typically give vectors of scores representing the proportions of cells of a tissue sample in each of four different staining categories. These scores will vary across pathologists not only because of inherent within-rater variability but also because the staining-level cutpoints that define the categories are subjective. In this paper, we propose a novel Bayesian approach, enabled by Markov chain Monte Carlo, to investigate differences in the pattern of vector scores across pathologists, while accounting for intra-rater variabilities. We also propose the use of the Bhattacharyya coefficient as an overall measure of agreement. This approach is needed because existing agreement measures either involve converting the vector to a univariate value, thereby losing information, or they fail to account for the sum-to-one restriction. Numerous simulation studies are used to demonstrate the validity of our model and the advantages of our approach over the more traditional ones. To enhance the use of this methodology and help with the design of future agreement studies, an R Shiny package is also introduced.
Medical Research
Presenter: Audrey Mauguen
When: Thursday, July 14, 2016 Time: 2:30 PM - 4:00 PM
Room: Salon C Carson Hall (Level 2)
Session Synopsis:
ESTIMATING THE PROBABILITY OF CLONAL RELATEDNESS IN CASES WITH TWO TUMORS
When two tumors arise in a patient, a key question is to determine whether they are two independent primary tumors, or whether one is a relapse from the other. This can be done by comparing the mutational profiles characterizing each tumor. If one of the tumors is a relapse, the two tumors have a common clonal history, and the mutations that occurred during the clonal period are shared by the two tumors. Conversely, if the two tumors are independent, shared mutations are observed only by chance. Ostrovnaya et al. (Ann Appl Stat 2015) recently developed a test of the hypothesis that the tumors are independent in this setting. We have built upon these ideas to develop diagnostic probabilities that a particular case is clonal versus independent. This involves a random effects model that can be applied to a learning dataset to estimate both the overall probability that a case is clonal (?) and the distribution of individual clonal signals {?j} that represent proportions of mutations in each tumor pair that are clonal mutations. This quantity is equal to 0 when the case is non clonal (ie, the two tumors are independent). Assuming that the distribution of ?j among clonal cases is Beta(?,?), we can estimate the parameters ?, ?, and ? using maximum likelihood estimation, and use these to estimate the diagnostic probabilities for each case. The accuracy of the proposed method is assessed through simulations, both in large and small sample sizes, and applied to a dataset involving breast cancer cases.
Medical Research
Presenter: Xu Zhao
When: Thursday, July 14, 2016 Time: 2:30 PM - 4:00 PM
Room: Salon C Carson Hall (Level 2)
Session Synopsis:
Affinity Analysis for chronic disease screening and prevention
Recommendations for chronic disease prevention, screening and management (CDPSM) rely on physicians having accurate and up-to-date patient information in their electronic medical records (EMR). Affinity analysis, also called association rule mining, examines the extent to which items occur together. It is commonly used in market basket analysis, e-commerce, online search analysis and bioinformatics to study interesting relations between variables in large databases. Our primary objective is to assess the applicability of association rule mining in large health care database from University of Toronto Practice Based Research Network (UTOPIAN PBRN) extracted October 2015. Identification of co-occurring CDPSM items in EMR may be important for service bundling and provision of more efficient health care items by physicians. We considered 9 CDPSM: BP, BMI, WC, LDL, FBG_A1c, Smoking, Alcohol, Diet and Exercise coded as binary data. To mine strong associations from the power set of the potential association rules, we used the Apriori algorithm implemented in arules package in R. The strength of rules was determined by support (prevalence), confidence (prediction of presence of another item when one item is chosen) and lift (the degree to which those two occurrences are dependent on one another). Support was highest for BP (82.01% of patients had a BP recorded) and lowest for Diet (3.6%). Confidence was highest if BMI was present, 98.46% had BP. Lifts were greater than 1 for all items studied indicating that they are non-independent. Some interesting association rules were also found, for example, BMI and BP, Alcohol and Smoking, LDL and FBG_A1c. Discovering and understanding associations between different items may help to plan for more effective and efficient recording of CDPSM.
Medical Research
Presenter: Mehdi Rostami
When: Thursday, July 14, 2016 Time: 2:30 PM - 4:00 PM
Room: Salon C Carson Hall (Level 2)
Session Synopsis:
Modeling spatially correlated survival data in hip fracture among residents from the long term care facilities in BC, Canada
In public health settings, epidemiologists and health professionals are interested in studying the spatial variation in survival data to reveal the underlying factors, which, in turn, assist them in identifying regions requiring spatialattention. Ignoring the correlations within clusters or between clusters may lead to incorrect standard errors of the estimates of parameters of interest. In the present study, we investigated the spatial random effect model for modeling geographically clustered survival data in hip fracture among residents from the long term care facilities in BC, Canada. Our main objective is to explain the pattern of hip fracture using important demographic, socioeconomic and clinical factors, while accounting for possible spatial correlation in the hazard among the geographic areas. We investigate the goodness-of-fit of our chosen model in bayesian frame work. We also investigate the impact of misspecifying the correlation structure of the random effect terms on estimating the effects of risk factors through simulation studies. Keywords: Bayesian hierarchical models, frailty models, Markov Chain Monte Carlo (MCMC), spatial association, survival modeling