CLINICAL AND ECONOMIC IMPACT OF A MACHINE LEARNING DRIVEN TWO-STAGE SCREENING STRATEGY FOR EARLY PANCREATIC CANCER DETECTION IN THE UK
Author(s)
Han-I Wang1, Ananya Malhotra, PhD2, Bernard Rachet, PhD2, Laura Woods Woods, PhD3.
1Student, University of York, York, United Kingdom, 2London School of Hygiene & Tropical Medicine, London, United Kingdom, 3Newcastle University, Newcastle Upon Tyne, United Kingdom.
1Student, University of York, York, United Kingdom, 2London School of Hygiene & Tropical Medicine, London, United Kingdom, 3Newcastle University, Newcastle Upon Tyne, United Kingdom.
OBJECTIVES: Early detection of pancreatic ductal adenocarcinoma (PDAC) remains a profound clinical challenge due to low baseline incidence, late symptoms, and rapid progression. This study evaluates the clinical and economic feasibility of an innovative two-stage screening process utilising our newly developed machine-learning algorithm paired with a subsequent non-invasive biomarker (the "ML-biomarker strategy") compared with current UK practice.
METHODS: Our stage 1 machine-learning algorithm was trained on the UK Clinical Practice Research Datalink (CPRD) using 63 pre-diagnostic symptoms and prescription features to identify high-risk individuals at least a year prior to standard diagnosis. Diagnostic accuracy metric was integrated with a secondary non-invasive biomarker. To evaluate downstream economic impact, a hybrid decision-tree and partitioned survival model (PSM) was developed from a UK NHS perspective over a lifetime horizon using monthly cycles. Survival and initial stage distributions were derived from the UK National Cancer Registry, while costs (£, 2025/26) and health utilities were sourced from Hospital Episode Statistics and published literature. Both costs and quality-adjusted life years (QALYs) were discounted at 3.5% annually.
RESULTS: The CPRD-trained XGBoost algorithm was the most optimal classifier for identifying high-risk individuals at an average of 18 months prior to standard diagnosis. Compared to standard care, the ML-biomarker strategy improved the population-average incremental health gain by 0.0148 QALYs, driven by early-detection stage-shifting. Incremental lifetime costs increased by £169.09, reflecting upfront diagnostic deployment and extended survival management. The Incremental Cost-Effectiveness Ratio (ICER) was £11,425 (95% CI: £10,585 to £12,318) per QALY gained. At the UK willingness-to-pay threshold of £20,000/QALY, the probability of ML-biomarker being cost-effectiveness was 100%.
CONCLUSIONS: The ML-biomarker strategy bridges advanced data science and clinical economics and is likely to be cost-effective if implemented in primary care as a gatekeeper to downstage disease and maximise patient QALYs.
METHODS: Our stage 1 machine-learning algorithm was trained on the UK Clinical Practice Research Datalink (CPRD) using 63 pre-diagnostic symptoms and prescription features to identify high-risk individuals at least a year prior to standard diagnosis. Diagnostic accuracy metric was integrated with a secondary non-invasive biomarker. To evaluate downstream economic impact, a hybrid decision-tree and partitioned survival model (PSM) was developed from a UK NHS perspective over a lifetime horizon using monthly cycles. Survival and initial stage distributions were derived from the UK National Cancer Registry, while costs (£, 2025/26) and health utilities were sourced from Hospital Episode Statistics and published literature. Both costs and quality-adjusted life years (QALYs) were discounted at 3.5% annually.
RESULTS: The CPRD-trained XGBoost algorithm was the most optimal classifier for identifying high-risk individuals at an average of 18 months prior to standard diagnosis. Compared to standard care, the ML-biomarker strategy improved the population-average incremental health gain by 0.0148 QALYs, driven by early-detection stage-shifting. Incremental lifetime costs increased by £169.09, reflecting upfront diagnostic deployment and extended survival management. The Incremental Cost-Effectiveness Ratio (ICER) was £11,425 (95% CI: £10,585 to £12,318) per QALY gained. At the UK willingness-to-pay threshold of £20,000/QALY, the probability of ML-biomarker being cost-effectiveness was 100%.
CONCLUSIONS: The ML-biomarker strategy bridges advanced data science and clinical economics and is likely to be cost-effective if implemented in primary care as a gatekeeper to downstage disease and maximise patient QALYs.
Conference/Value in Health Info
2026-11, ISPOR Europe 2026, Vienna, Austria
Value in Health, Volume 29, Issue 12S
Code
EE77
Topic
Economic Evaluation, Methodological & Statistical Research
Topic Subcategory
Cost/Cost of Illness/Resource Use Studies
Disease
Oncology