Can we screen for pancreatic cancer? Identifying a sub-population of patients at high risk of subsequent diagnosis using machine learning techniques applied to primary care data

Abstract

Background: Pancreatic cancer (PC) represents a substantial public health burden. Pancreatic cancer patients have very low survival due to the difficulty of identifying cancers early when the tumour is localised to the site of origin and treatable. Recent progress has been made in identifying biomarkers for PC in the blood and urine, but these cannot be used for population-based screening as this would be prohibitively expensive and potentially harmful.

Methods: We conducted a case-control study using prospectively-collected electronic health records from primary care individually-linked to cancer registrations. Our cases were comprised of 1,139 patients, aged 15-99 years, diagnosed with pancreatic cancer between January 1, 2005 and June 30, 2009. Each case was age-, sex- and diagnosis time-matched to four non-pancreatic (cancer patient) controls. Disease and prescription codes for the 24 months prior to diagnosis were used to identify 57 individual symptoms. Using a machine learning approach, we trained a logistic regression model on 75% of the data to predict patients who later developed PC and tested the model's performance on the remaining 25%.

Results: We were able to identify 41.3% of patients < = 60 years at 'high risk' of developing pancreatic cancer up to 20 months prior to diagnosis with 72.5% sensitivity, 59% specificity and, 66% AUC. 43.2% of patients >60 years were similarly identified at 17 months, with 65% sensitivity, 57% specificity and, 61% AUC. We estimate that combining our algorithm with currently available biomarker tests could result in 30 older and 400 younger patients per cancer being identified as 'potential patients', and the earlier diagnosis of around 60% of tumours.

Conclusion: After further work this approach could be applied in the primary care setting and has the potential to be used alongside a non-invasive biomarker test to increase earlier diagnosis. This would result in a greater number of patients surviving this devastating disease.

Fig 1. Schematic of multi-stage screening process starting with machine-learning derived algorithm.

Fig 2. Model parameters obtained in each time window (0–24,1–24,…,20–24 months) prior to diagnosis.

a) AUC (%) b) Sensitivity (%) c) Specificity (%).

Fig 3. AUC, sensitivity and specificity of the models with % of the population recommended for biomarker (Stage 2) testing.