Understanding the Limits of Large Datasets

North Carolina medical journal, 2014

European Journal of Cancer and Clinical Oncology

The quality of the recorded diagnosis is a major limit to the usefulness of Cancer Registry statistics that is easily overlooked by users of the data. With data from a largepopulationbased cancer registry as an example, we demonstrate how Registry statistics could be improved by wider use of three simple indices, name& (1) the proportion histologically verified (adjusted for age), (2) the proportion of verified cases with an uninformative diagnosis, and (3) the proportion of cases that are staged. We believe that greater awareness of the deficiencies of Cancer Registry statistics will lead to a more critical interpretation of them, and help stimulate efforts to rectif matters.

Cancer Medicine, 2021

BackgroundMissing patient reported outcomes data threaten the validity of PRO‐specific findings and conclusions from randomized controlled trials by introducing bias due to data missing not at random. Clinical Research Associates are a largely unexplored source for informing understanding of potential causes of missing PRO data. The purpose of this qualitative research was to describe factors that influence missing PRO data, as revealed through the lived experience of CRAs.MethodsMaximum variation sampling was used to select CRAs having a range of experiences with missing PRO data from academic or nonacademic centers in different geographic locations of Canada. Semistructured interviews were audio‐recorded, transcribed verbatim, and analyzed according to descriptive phenomenology.ResultsEleven CRAs were interviewed. Analysis revealed several factors that influence missing PRO data that were organized within themes. PROs for routine clinical care compete with PROs for RCTs. Both the ...

Methods in Molecular Biology, 2009

Cancer registries provide systematically collected information on cancer incidence, prevalence, mortality, and survival of different cancers. Aggregated and de-identified patient-level information on cancer is available for analysis from individual cancer registries, nationally from the Surveillance, Epidemiology, and End Results program, the Centers for Diseases Control and Prevention, the North American Association of Central Cancer Registries; and internationally from the International Association of Cancer Registries. Over the past few decades, the type and extent of cancer-related information captured by different cancer registries have been greatly expanded by linkage with other population-based information sources, such as the census data and the Centers for Medicare and Medicaid Services claims data. In addition, sophisticated statistical analytical techniques have been developed that have greatly expanded the traditional purview of cancer registries focused on descriptive epidemiology and disease quantification to a much broader analytical horizon ranging from study of cancer etiology; rare cancers in specific demographic groups; interaction of environmental and genetic factors in causation of cancer; impact of co-morbidities, race, geographic, socioeconomic, and provider-related factors on access, diagnosis, and treatment; outcomes and end results of cancer treatment; and cancer control initiatives to diverse areas of cancer care disparity, public health policy, public health education, and importantly, cost-effectiveness of cancer care. Thus, it is not surprising that cancer registries have increasingly become indispensable parts of local, national, and international cancer control programs, and it is certain that cancer registries will continue to be extraordinary resources of information for clinicians, researchers, scientists, policy makers, and the public in our fight against cancer.

British journal of cancer, 1997

There is a need to evaluate cancer services and provide a baseline on current treatment success and organization. This study shows that this process may be severely hindered by case note destruction or inaccessibility and incomplete information. This is an ongoing problem that needs to be addressed now.

Journal of oncology practice / American Society of Clinical Oncology, 2015

JCO precision oncology, 2020

PURPOSE Our goal was to identify the opportunities and challenges in analyzing data from the American Association of Cancer Research Project Genomics Evidence Neoplasia Information Exchange (GENIE), a multiinstitutional database derived from clinically driven genomic testing, at both the inter-and the intra-institutional level. Inter-institutionally, we identified genotypic differences between primary and metastatic tumors across the 3 most represented cancers in GENIE. Intra-institutionally, we analyzed the clinical characteristics of the Vanderbilt-Ingram Cancer Center (VICC) subset of GENIE to inform the interpretation of GENIE as a whole. METHODS We performed overall cohort matching on the basis of age, ethnicity, and sex of 13,208 patients stratified by cancer type (breast, colon, or lung) and sample site (primary or metastatic). We then determined whether detected variants, at the gene level, were associated with primary or metastatic tumors. We extracted clinical data for the VICC subset from VICC's clinical data warehouse. Treatment exposures were mapped to a 13-class schema derived from the HemOnc ontology. RESULTS Across 756 genes, there were significant differences in all cancer types. In breast cancer, ESR1 variants were over-represented in metastatic samples (odds ratio, 5.91; q , 10 −6). TP53 mutations were overrepresented in metastatic samples across all cancers. VICC had a significantly different cancer type distribution than that of GENIE but patients were well matched with respect to age, sex, and sample type. Treatment data from VICC was used for a bipartite network analysis, demonstrating clusters with a mix of histologies and others being more histology specific. CONCLUSION This article demonstrates the feasibility of deriving meaningful insights from GENIE at the inter-and intra-institutional level and illuminates the opportunities and challenges of the data GENIE contains. The results should help guide future development of GENIE, with the goal of fully realizing its potential for accelerating precision medicine.

Pathology & Oncology Research, 2002

Cancer Causes & Control, 2011

Objective-Cancer incidence and mortality statistics provide limited insight regarding the cancer survivor population and its needs. Cancer prevalence statistics enumerate cancer survivors-those currently living with cancer. Commonly used limited-duration prevalence (LDP) methods yield biased estimates of the number of survivors. National estimates may not allow sufficient granularity to inform local survivorship programs. In this study, complete prevalence (CP) methods are applied to actual North Carolina Central Cancer Registry (NCCCR) data to generate better, more informative prevalence estimates than previous methods. Methods-Data included all incident cases for 1995-2007 from the NCCCR and US Census population data. SEER*Stat software was used to calculate 13-year LDP. ComPrev software was used to estimate CP for each cancer site, gender, and race combination. Results-CP methods estimated 362,810 survivors in North Carolina on January 1, 2008, 40% more than LDP estimates of 258,556, with substantial racial, regional, and gender differences in prevalence rankings of several cancers. Conclusion-CP estimates are substantially higher than previous prevalence estimates. This study found previously unrecognized racial, regional, and gender differences. State and local programs may apply these methods using their own data to develop better, more detailed estimates to improve planning for their specific survivor populations' needs.

arXiv (Cornell University), 2022

Missing data is a common concern in health datasets, and its impact on good decision-making processes is well documented. Our study's contribution is a methodology for tackling missing data problems using a combination of synthetic dataset generation, missing data imputation and deep learning methods to resolve missing data challenges. Specifically, we conducted a series of experiments with these objectives; a) generating a realistic synthetic dataset, b) simulating data missingness, c) recovering the missing data, and d) analyzing imputation performance. Our methodology used a gaussian mixture model whose parameters were learned from a cleaned subset of a real demographic and health dataset to generate the synthetic data. We simulated various missingness degrees ranging from 10%, 20%, 30%, and 40% under the missing completely at random scheme MCAR. We used an integrated performance analysis framework involving clustering, classification and direct imputation analysis. Our results show that models trained on synthetic and imputed datasets could make predictions with an accuracy of 83% and 80% on a) an unseen real dataset and b) an unseen reserved synthetic test dataset, respectively. Moreover, the models that used the DAE method for imputation yielded the lowest log loss an indication of good performance, even though the accuracy measures were slightly lower. In conclusion, our work demonstrates that using our methodology, one can reverse engineer a solution to resolve missingness on an unseen dataset with missingness. Moreover, though we used a health dataset, our methodology can be utilized in other contexts.