Survival Analysis 1
Presenter: Rachel Altman
When: Monday, July 11, 2016 Time: 11:00 AM - 12:30 PM
Room: Salon B Carson Hall (Level 2)
Session Synopsis:
Assessing Prediction Error in Survival Times: Application to Ovarian Cancer
Random survival forests (RSFs) are a tool for predicting survival times. Forests are grown using a set of (possibly censored) survival times with associated predictor variables. One challenge surrounding the use of RSFs is the assessment of prediction error; mean squared error (MSE), the commonly used measure of prediction error in the traditional random forest setting, is not an option if some observations are censored. An alternative measure of prediction error that has been suggested in the literature is the C-index, which summarizes the concordance between the observed and predicted values and can incorporate some censored cases. However, in this talk, we show that this measure of prediction error can behave in undesirable ways (and quite differently than MSE).
Survival Analysis 1
Presenter: IRANTZU BARRIO
When: Monday, July 11, 2016 Time: 11:00 AM - 12:30 PM
Room: Salon B Carson Hall (Level 2)
Session Synopsis:
Polychotomization of continuous variables in Cox proportional hazards regression models: application to patients with stable COPD.
The Cox proportional hazards model is the most common survival prediction model for analyzing time to event data. In medical research, a common strategy when developing survival models is to categorize continuous covariates. However, agreement on criteria to select cut points does not always exist. Previous work on the estimation of optimal cut points with censored data has been done but with the aim of seeking for a unique cut point (Sima and Gönen, 2013). In this work, we propose a methodology for the polychotomization of continuous predictor variables in a Cox proportional hazards regression model considering the maximal discrimination attained. To measure the discrimination ability of the model, we considered the concordance probability index, and two different estimators were studied: the c-index (Harrell, 1982) and the concordance probability estimator (CPE, Gönen and Heller ,2005). The algorithms used to select the optimal cut points, called Addfor and Genetic have been presented elsewhere (Barrio et al, 2015). We conducted a simulation study to evaluate the empirical performance of both the c-index and CPE estimators when it comes to select the optimal cut points for the categorization of continuous variables. Simulations were performed for different sample sizes (500 and 1000), number of cut points (1,2 and 3) and censoring rates (20%,50% and 70%). The method performs successfully when it comes to search two or three cut points. However, when the aim is to search a unique cut point, the methods performance depends largely on the location of the theoretical optimal cut point.This methodology has been applied to a cohort of 543 patients with stable chronic obstructive pulmonary disease (COPD, Esteban et al., 2014). An important predictor for COPD mortality is the forced expiratory volume in one second in percentile (FEV1%).The most commonly used prediction models for the evolution of patients with COPD use a categorized version of FEV1%, but not all of them use the same cut points. We considered categorizing the predictor variable FEV1% in a multiple Cox proportional hazards regression model taking into account the effect of age and dyspnea. We got that the optimal cut points were two, 30 and 50 with an estimated c-index of 0.734 which outperformed previous categorization proposals.
Survival Analysis 1
Presenter: Daniel Conn
When: Monday, July 11, 2016 Time: 11:00 AM - 12:30 PM
Room: Salon B Carson Hall (Level 2)
Session Synopsis:
Metric Learning for Right Censored Survival Outcomes
In this paper we adapt the metric learning methodology to censored outcomes. Metric learning is an extension of kernel regression designed to overcome the flaws of kernel regression in moderate or high dimensions. In metric learning, the kernel function is learned from the data via various optimization routines. This data adaptive kernel function is effectively able to remove unimportant covariates from the analysis and use lower dimensional structure to increase the estimation efficiency. In this way, metric learning achieves satisfactory results when the number of features is moderate in comparison to the sample size. Our extension of metric learning to survival data takes advantage of Leurgans synthetic data approach. We demonstrate our novel method by predicting survival times for sample of small cell lung cancer cases from the SEER (Surveilance Epidemiology and End Result) database.
Survival Analysis 1
Presenter: Pierre Joly
When: Monday, July 11, 2016 Time: 11:00 AM - 12:30 PM
Room: Salon B Carson Hall (Level 2)
Session Synopsis:
Projections of health indicators for chronic disease under Semi-Markov assumption
Chronic diseases are a growing public health problem. Thus, their economic, social and demographic burdens are increasing in years to come. The Prevalence is the number of cases of a disease in a population at a specific time and this can be a very important information to plan future actions of the health system. We propose an estimation of the age-specific prevalence and other health indicators (overall life expectancy, life expectancy without the disease, life-long probability of the disease ) taking into account the non-homogeneity of the age-specific mortality of subjects over time. Up to now, the method used to make projections assumed non-homogeneous Markov assumption in an illness-death model. The age and the calendar time were taken into account in all parameters estimations, nevertheless the time spent with the disease was not considered. This work aims to develop the method with semi-Markov assumption to model the mortality among diseased, considering the time spent with the disease. The method is applied for estimating several health indicators for dementia in France in 2030. The French National Institute of Statistics provides French demographic projections detailed by year, age and gender. Incidence of dementia and mortality of demented subjects can be estimate thanks to data from large cohort studies.
Survival Analysis 1
Presenter: Patience Nyakato
When: Monday, July 11, 2016 Time: 11:00 AM - 12:30 PM
Room: Salon B Carson Hall (Level 2)
Session Synopsis:
Use of Sample-based methodology to obtain corrected estimates of retention in care of newly diagnosed HIV positive patients before the initiation of ART in Uganda.
Background: Access to antiretroviral therapy (ART) has expanded considerably in sub-Saharan Africa but AIDS-related mortality remains high, partly due to attrition from care. This attrition can lead to potential biases in estimates of retention and survival among HIV patients. In order to evaluate the effect on HIV care programs and appropriately direct scarce resources, it is essential to have accurate estimates of outcomes among HIV infected patients. We used sample-based methodology to improve the accuracy of estimates of retention among new pre-ART patients at the Infectious Diseases Institute (IDI) in Kampala, Uganda. Methods: This was a retrospective cohort study among newly diagnosed HIV positive patients, not yet eligible for ART and had registered at 3 Kampala clinics (Kisenyi, Kawala and Kitebi) served by IDI between January-May 2015. Out of all that were considered lost to follow up (LTFU) defined as missing a scheduled appointment for 3 months or more, a random sample of 125 was traced to determine their HIV care outcome status. Tracing was done through telephone calls and home visits. Corrected estimates of retention, were obtained using Frangakis & Rubin method of double sampling. Weighted survival anlaysis was then done so that each patient found to be in care after tracing is given a probability. Results: A total of 871 (24% male, 29 years, median cd4 cell count of 617) patients had registered for pre-ART care at the clinics in the period between January-May 2015. Of these, 213 (38% male, median age of 29 years, median cd4 cell count of 636) were lost to follow up and a random sample of 125 was intensively traced. Outcomes were successfully ascertained for 85 (68%) participants. 3 (3.5%) were dead, 30 (35.3%) of those alive were in care elsewhere (self transfers), 2 (2.4%) were still actively attending the clinics but under different names and 50 (58.8%) were out of care. After correcting for retention using a probability weight of 2.51 (213/3+32+50), there was a 4.5% increase in retention estimates. Conclusion: Estimates of retention in care and mortality for the pre-ART patients maybe under estimated if clinic records are the sole source of data because they include patients that are considered LTFU but are in care elsewhere or in the same clinic under different identity.
Survival Analysis 1
Presenter: Morten Valberg
When: Monday, July 11, 2016 Time: 11:00 AM - 12:30 PM
Room: Salon B Carson Hall (Level 2)
Session Synopsis:
Prostate-specific antigen testing for prostate cancer: Emptying a pool of susceptible individuals?
Background: After the introduction of the prostate specific antigen (PSA) test in the 1980s, a sharp increase in the incidence rate of prostate cancer was seen in the United States. The age-specific incidence patterns exhibited remarkable shifts to younger ages, and declining rates were observed at old ages. Similar trends were seen in Norway. We investigate whether these features could be explained by the existence of subgroups of the populations that are especially susceptible to prostate cancer. Methods: We analyzed incidence data from the United States Surveillance, Epidemiology, and End Results program for 1973-2010, comprising 511 027 prostate cancers in men ?40 years old, and national Norwegian incidence data for 1953-2011, comprising 113 837 prostate cancers in men ?50 years old. We developed a frailty model where only a proportion of the population can develop prostate cancer. The increased risk of being diagnosed with the cancer due to the massive use of PSA testing is taken into account. Results: The proportion of men that were susceptible was 39.9% (95% CI: 38.2%, 41.6%) in the United States and 30.4% (95% CI: 28.9%, 32.0%) in Norway. The frailty model describes the changing age-specific incidence patterns across birth cohorts well. Conclusion: The peaking cohort-specific age-incidence curves of prostate cancer may be explained by the underlying heterogeneity in prostate cancer risk. Furthermore, the introduction of the PSA test seems to have driven the peak in the incidence rate toward younger ages by inducing a larger depletion of a pool of individuals susceptible to this cancer.