A novel prediction model to identify patients with early-stage pancreatic cancer.

Abstract

Listen

Translate article

e16801 Background: Pancreatic cancer has an insidious presentation with four-in-five patients presenting with disease not amenable to potentially curative surgery. Efforts to screen patients for pancreatic cancer using population-wide strategies have proven ineffective. We applied a machine learning approach to create an early prediction model drawing on the content of patients’ electronic health records (EHRs). Methods: We used patient data from OptumLabs which included de-identified data extracted from patient EHRs collected between 2009 and 2017. We identified patients diagnosed with pancreatic cancer at age 40 or later, which we categorized into early-stage pancreatic cancer (ESPC; n = 3,322) and late-stage pancreatic cancer (LSPC; n = 25,908) groups. ESPC cases were matched to non-pancreatic cancer controls in a ratio of 1:16 based on diagnosis year and geographic division, and the cohort was divided into training (70%) and test (30%) sets. The prediction model was built using an eXtreme Gradient Boosting machine learning algorithm of ESPC patients’ EHRs in the year preceding diagnosis, with features including patient demographics, procedure and clinical diagnosis codes, clinical notes and medications. Model discrimination was assessed with sensitivity, specificity, positive predictive value (PPV) and area under the curve (AUC) with a score of 1.0 indicating perfect prediction. Results: The final AUC in the test set was 0.841, and the model included 583 features, of which 248 (42.5%) were physician note elements, 146 (25.0%) were procedure codes, 91 (15.6%) were diagnosis codes, 89 (15.3%) were medications and 9 (1.54%) were demographic features. The most important features were history of pancreatic disorders (not diabetes or cancer), age, income, biliary tract disease, education level, obstructive jaundice and abdominal pain. We evaluated model performance at varying classification thresholds. When applied to patients over 40 choosing a threshold with a sensitivity of 20% produced a specificity of 99.9% and a PPV of 2.5%. The model PPV increased with age; for patients over 80, PPV was 8.0%. LSPC patients identified by the model would have been detected a median of 4 months before their actual diagnosis, with a quarter of these patients identified at least 14 months earlier. Conclusions: Using EHR data to identify early-stage pancreatic cancer patients shows promise. While widespread use of this approach on an unselected population would produce high rates of false positives, this technique could be employed among high risk patients, or paired with other screening tools.