Development and external validation of prediction algorithms to improve early diagnosis of cancer

Abstract

Abstract Cancer prediction algorithms are used in the UK to identify individuals at high probability of having a current, as yet undiagnosed cancer with the intention of improving early diagnosis and treatment. Here we develop and externally validate two diagnostic prediction algorithms to estimate the probability of having cancer for 15 cancer types. The first incorporates multiple predictors including age, sex, deprivation, smoking, alcohol, family history, medical diagnoses and symptoms (both general and cancer-specific symptoms). The second additionally includes commonly used blood tests (full blood count and liver function tests). We use multinomial logistic regression to develop separate equations in men and women to predict the absolute probability of 15 cancer types using a population of 7.46 million adults aged 18 to 84 years in England. We evaluate performance in two separate validation cohorts (total 2.64 million patients in England and 2.74 million from Scotland, Wales and Northern Ireland). The models have improved performance compared with existing models with improved discrimination, calibration, sensitivity and net benefit. These algorithms provide superior prediction estimates in the UK compared with existing scores and could lead to better clinical decision-making and potentially earlier diagnosis of cancer.