Abstract 880: Development and validation of an electronic health record-based absolute risk prediction model for esophageal cancer in the primary care setting in China

Abstract

Abstract Background This study aims to develop a set of electronic health record (EHR)-based absolute risk prediction models for esophageal cancer (EC), which are nationally applicable and easily integrated into the EHR system in China. Methods The study used data from China Kadoorie Biobank (CKB), a large nationwide prospective cohort recruited during 2004-2008 in ten regions. In a random two-thirds of the 510,145 participants with no prior cancer diagnoses, we developed four nested models, according to the availability of predictors in EHRs. A flexible parametric survival model was fitted to calculate a 10-year absolute risk of EC accounting for the competing risk of mortality from other diseases. Besides discrimination and calibration, we estimated a range of performance indices (e.g. sensitivity and specificity) and compared the model performance with current screening criteria in the remaining one-third. Results During a median of 11.1 years of follow-up, we observed 2,550 EC incident cases. The most sophisticated model included EHR variable (age, sex, region risk level [of ten regions: Hui county in Henan province and Pengzhou in Sichuan province], education level, family history of cancer, smoking, alcohol drinking, body mass index), and physical activity, hot tea consumption, and fresh fruit consumption. In the validation dataset, as more predictors included, the area under the receiver-operating characteristic curve statistically increased from 0.748 (0.732-0.763) to 0.883 (0.867-0.895). We observed a great agreement between predicted and observed risk, except for a slight underestimation in the top deciles. Taking 0.40% as the 10-year risk cut-off, models had sensitivities and specificities of around 80% and largely outperformed the current screening criteria. Conclusions Based only on non-invasive predictors, even the model including only five predictors in the EHRs exhibited excellent calibration and discrimination. Utilizing limited predictors in EHRs, our models are useful to develop nationwide risk-stratified screening strategies for esophageal cancer. Citation Format: Yuting Han, Canqing Yu, Yu Guo, Pei Pei, Ling Yang, Yiping Chen, Huaidong Du, Junshi Chen, Zhengming Chen, Dezheng Huo, Jun Lv, Liming Li. Development and validation of an electronic health record-based absolute risk prediction model for esophageal cancer in the primary care setting in China [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2021; 2021 Apr 10-15 and May 17-21. Philadelphia (PA): AACR; Cancer Res 2021;81(13_Suppl):Abstract nr 880.