Predicting risk of pancreatic cancer in individuals with new-onset type-2 diabetes in primary care: protocol for the development and validation of a clinical prediction model (QPancreasD)

Background Pancreatic cancer continues to have an extremely poor prognosis in part due to late diagnosis. 25% of pancreatic cancer patients have a prior diagnosis of diabetes, and hence identifying individuals at risk of pancreatic cancer in those with recently diagnosed type 2 diabetes may be a useful opportunity to identify candidates for screening and early detection. In this study, we will comparatively evaluate regression and machine learning-based clinical prediction models for estimating individual risk of developing pancreatic cancer two years after type 2 diabetes diagnosis. Methods In the development dataset, we will include adults aged 30-84 years with incident type-2 diabetes registered with QResearch primary care database. Patients will be followed up from type-2 diabetes diagnosis to first diagnosis of pancreatic cancer as recorded in any one of primary care records, hospital episode statistics, cancer registry data, or death records. Cox-proportional hazards models will be used to develop a risk prediction model for estimating individual risk of developing pancreatic cancer during up to 2 years of follow-up. We will perform variable selection using a combination of clinical and statistical significance approach i.e. HR <0.9 or >1.1 and p<0.01. Linear predictors and baseline survivor function at 2 years will be used to compute absolute risk predictions. Internal-external cross-validation (IECV) framework across geographical regions within England will be used to assess performance and pooled using random effects meta-analysis using: (i) model fit in terms of variation explained by the model Royston & Sauerbrei's R2D, (ii) calibration slope and calibration-in-the-large, and (iii) discrimination measured in terms of Harrell's C and Royston & Sauerbrei's D-statistic. Further, we will evaluate machine learning (ML) approaches for the clinical prediction model using neural networks (NN) and XGBoost. The model predictors and performance of these will be compared with the results of those derived from the regression-based strategy. Discussion The proposed study will develop and validate a novel risk prediction model to aid early diagnosis of pancreatic cancer in patients with new-onset diabetes in primary care. With an enhanced decision-risk tool for use at point-of care by general practitioners to assess pancreatic cancer risk, it may improve decision-making so that at-risk patients are rapidly prioritised to aid early diagnosis of pancreatic cancer in patients with newly diagnosed diabetes.