Differentiating Ductal Adenocarcinoma of the Pancreas from Benign Conditions Using Routine Health Records: A Prospective Case-Control Study

SIMPLE SUMMARY: Pancreatic ductal adenocarcinoma (PDAC) constitutes a devastating disease with late diagnosis and poor overall survival, complicated by clinical presentations similar to benign pancreatic diseases. We aimed to analyse clinical parameters for differentiating suspected PDAC from benign conditions. The study holistically explored the presenting symptoms and routine laboratory test results of pancreatic disease patients during their consultation at secondary/tertiary care, including their demographic, lifestyle and comorbidity characteristics. Developed on a prospectively recruited cohort, this is the first machine learning-based prediction model that differentiates PDAC patients from those with benign conditions with a high degree of recall (sensitivity) and precision (positive predictive value). The model could serve to support clinicians’ decisions when assessing patients with pancreatic pathology and separating potential malignant candidates from benign ones for urgent referral to a tertiary centre. This could improve upon the current UK guidelines enabling early detection of PDAC by developing a digital referral tool. ABSTRACT: The study aimed to develop a prediction model for differentiating suspected PDAC from benign conditions. We used a prospective cohort of patients with pancreatic disease (n = 762) enrolled at the Barts Pancreas Tissue Bank (2008-2021) and performed a case-control study examining the association of PDAC (n = 340) with predictor variables including demographics, comorbidities, lifestyle factors, presenting symptoms and commonly performed blood tests. Age (over 55), weight loss in hypertensive patients, recent symptoms of jaundice, high serum bilirubin, low serum creatinine, high serum alkaline phosphatase, low red blood cell count and low serum sodium were identified as the most important features. These predictors were then used for training several machine-learning-based risk-prediction models on 75% of the cohort. Models were assessed on the remaining 25%. A logistic regression-based model had the best overall performance in the validation cohort (area-under-the-curve = 0.90; Spiegelhalter’s z = −1·82, p = 0.07). Setting a probability threshold of 0.15 guided by the maximum F2-score of 0.855, 96.8% sensitivity was reached in the full cohort, which could lead to earlier detection of 84.7% of the PDAC patients. The prediction model has the potential to be applied in primary, secondary and emergency care settings for the early distinction of suspected PDAC patients and expedited referral to specialist hepato-pancreatico-biliary services.