Development and Validation of a Multivariable Prediction Model for All-Cause Cancer Incidence Based on Health Behaviours in the Population Setting

Abstract

Background: We examined if it was possible to use routinely available, self-reported data on health behaviours to predict incident cancer cases in the Ontario population.

Methods: This retrospective cohort study involved 43 696 female and 36 630 male respondents from Ontario, who were >20 years old and without a prior history of cancer, to the Canadian Community Health Survey (CCHS) cycles 2.1-4.1. The outcome of interest was malignant cancer from any site, termed all-cause cancer, determined from the Ontario Cancer Registry. Predictor variables in the risk algorithm were health behaviours including smoking status, pack-years of smoking, alcohol consumption, fruit and vegetable consumption and physical activity level. A competing-risk Cox proportional hazard model was utilized to determine hazard of incident cancer. The developed risk prediction tool was validated in the CCHS cycle 1.1 on 14 426 female and 11 970 male survey respondents.

Results: Incident cancer was predicted with a high degree of calibration (differences between observed and predicted values for females 2.97%, for males 4.23%) and discrimination (C-statistic: females 0.76, males 0.83). Similar results were obtained in the validation cohort.

Conclusions: Routinely collected self-reported information on health behaviours can be used to predict incident cancer in the Ontario population. This type of risk prediction tool is valuable for public health purposes of estimating population risk of incident cancer, as well as projection of future risk in the population over time.