Back to Search
Start Over
Abstract 5300: Risk models for cancer screening cohorts assembled from electronic health records: Application to calculating risks that underlie current cervical cancer screening guidelines
- Source :
- Cancer Research. 77:5300-5300
- Publication Year :
- 2017
- Publisher :
- American Association for Cancer Research (AACR), 2017.
-
Abstract
- Introduction: We developed risk models for cancer screening cohorts assembled from electronic health records. We applied them to a cohort of 1.4 million women undergoing cervical cancer screening with human papillomavirus (HPV) and Pap “cotesting” at Kaiser Permanente Northern California (KPNC) to develop risk estimates to inform screening guidelines. Methods: Cohorts assembled using electronic health records present 3 challenges that make it inappropriate to use Kaplan-Meier or Cox models. First, the time of disease onset for an individual is unobserved, and falls between screens (interval-censoring). Second, there is also prevalent disease (left-censoring). Third, prevalent disease is not always immediately diagnosed (e.g. in those with negative screening results), and thus some incident disease is actually missed prevalent disease. To address the challenges, we propose a “logistic-Weibull” model that is a logistic regression for prevalent disease and a Weibull survival regression for interval-censored incident disease. We also propose a non-parametric method (no covariates) to check the assumptions of the logistic-Weibull model. We calculate risks of cervical intraepithelial neoplasia grade 3 and cancer (CIN3+) for each possible abnormal cotesting result. As an illustrative example of the biases that can occur in real data, we present the Kaplan-Meier, logistic-Weibull, and non-parametric cumulative risk curves for 34,261 women at KPNC who test Pap negative and HPV positive at enrollment. Results: The non-parametric method estimates 1.99% prevalent risk of CIN3+ and 5.68% 7-year cumulative risk of CIN3+ among women who are Pap-negative/HPV-positive at enrollment. In contrast, the Kaplan-Meier method estimates merely 0.18% prevalent risk of CIN3+, primarily because of prevalent disease not diagnosed at baseline. Furthermore, the Kaplan-Meier method overestimates 7-year cumulative risk as 7.37% because it equates the time of onset with the time of diagnosis, and thus overestimates the hazards of disease onset at later times. In contrast, the logistic-Weibull estimates of 1.87% prevalent risk of CIN3+ and 5.84% 7-year cumulative risk of CIN3+ are close to the risk estimates from the non-parametric method. Discussion: The Kaplan-Meier method provided poor risk estimates while logistic-Weibull model-based risks were close to the risk estimates from the non-parametric method. Our findings support use of the logistic-Weibull models over Kaplan-Meier methods for developing the risk estimates that underlie current U.S. cervical cancer screening guidelines. Citation Format: Li C. Cheung, Qing Pan, Noorie Hyun, Mark Schiffman, Barbara Fetterman, Philip E. Castle, Thomas Lorey, Hormuzd A. Katki. Risk models for cancer screening cohorts assembled from electronic health records: Application to calculating risks that underlie current cervical cancer screening guidelines [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2017; 2017 Apr 1-5; Washington, DC. Philadelphia (PA): AACR; Cancer Res 2017;77(13 Suppl):Abstract nr 5300. doi:10.1158/1538-7445.AM2017-5300
Details
- ISSN :
- 15387445 and 00085472
- Volume :
- 77
- Database :
- OpenAIRE
- Journal :
- Cancer Research
- Accession number :
- edsair.doi...........c96fd319577c4219758039a96b4f5211