Background: Risk models provide a means to quantify the benefit of cancer prevention measures. However, existing risk prediction models mainly focused on estimating risk for individual cancers.
Objectives: To facilitate risk-informed primary prevention measures and effective implementation of multicancer early detection test, we propose a comprehensive multicancer risk prediction model.
Methods: Cancer risk prediction models were developed using the UK Biobank (UKB) cohort and externally validated in the European Prospective Investigation into Cancer and Nutrition (EPIC) cohort. We fitted 23 cancer-specific risk models for males and 26 for females using Cox-proportional hazards model with a 10-year time horizon. Cancer-specific informative risk factors were selected based on existing literature and LASSO-penalized regression. Total cancer risk was estimated by adding up the sub-hazards from the cancer-specific models. Preliminary model performance for males was evaluated by 10-repeated 10-fold cross-validated area under the curve (AUC – i.e. model discrimination) and the ratio of expected to observed events (E/O – i.e. model calibration), respectively.
Results: A total of 218,720 males with complete smoking status were included, comprising 107,552 never-smokers and 111,168 ever-smokers. In the internal 10-fold cross validation, 23 cancer-specific models developed using LASSO-selected variables performed well in terms of AUC and E/O. Preliminary AUCs ranged from 0.572 for soft tissue cancer to 0.798 for lung cancer in participants with a history of smoking exposure. The model AUCs for several other cancers exceeded 0.700, including esophageal adenocarcinoma (0.719), esophageal squamous cell carcinoma (0.723), other respiratory cancers (0.726), bladder cancer (0.747), mesothelioma (0.759),  and liver cancer (0.767). Most models demonstrated good calibration, with E/Os ranging between 1.033 for oropharynx cancer and 1.170 for esophageal squamous cell carcinoma over 10-years of follow-up. The total cancer model yielded an AUC of 0.695, with a cross-validated E/O of 1.111.
Conclusions: Our preliminary risk models provided appropriate risk estimates and demonstrated the strength of the modelling framework. During the meeting we will present the performance of the final models.