JMIR Publications

ABSTRACT

Background:

China has the largest burden of esophageal cancer (EC). Prediction models can be used to identify high-risk individuals for intensive lifestyle intervention and endoscopy screening. However, current models are limited by a small sample size and a lack of external validation, and none of them can be embedded into the booming electronic health records (EHRs) in China.

Objective:

This study aims to develop and validate an absolute risk prediction model for esophageal cancer (EC) in the Chinese population. We especially assessed whether models that contain only electronic health record (EHR)-available predictors performed well.

Methods:

A prospective cohort recruiting 510,145 participants free of cancer from both high-risk and low-risk EC areas in China was used to develop EC models. Another prospective cohort of 18,441 participants was used for validation. A flexible parametric model was used to develop a 10-year absolute risk model taking account of competing risk (full model). The full model was then abbreviated by keeping only EHR-available predictors. We internally and externally validated models using the area under the receiver-operating characteristic curve (AUC) and calibration plot and compared them based on classification measures.

Results:

During a median of 11.1 years of follow-up, we observed 2,550 EC incident cases. The models included age, sex, regional risk level (high-risk areas: two study regions; low-risk areas: other eight regions), education, family history of cancer (above predictors: simple model), smoking, alcohol drinking, BMI (intermediate model), physical activity, hot tea consumption, and fresh fruit consumption (full model). The performance was only slightly compromised after the abbreviation. The simple and intermediate models showed good calibration and excellent discriminating ability with an AUCs (95% confidence intervals) of 0.822 (0.783-0.861) and 0.830 (0.792-0.867) in the external validation, and 0.871 (0.858-0.884) and 0.879 (0.867-0.892) in the internal validation.

Conclusions:

Three nested models were developed and validated. Even the simple model with only five predictors available from the EHR had excellent discrimination and good calibration. The simple and intermediate models have the potential to be widely used for both primary and secondary prevention.