Risk prediction of pancreatic cancer using clinical risk factors and biomarkers

Abstract

Pancreatic cancer has a poor prognosis with 5-year survival rates ranging from 5% to 15% and is often diagnosed at an advanced stage due to vague and non-specific symptoms. Routine screening for the general population is not recommended due to the disease's low incidence (4.9 per 100,000 people), but screening may be appropriate for high-risk populations if individuals at sufficient risk can be identified. Screening is typically advised for those with a strong family history or a known genetic predisposition, although these factors only account for 5-10% of cases. Therefore, the aim of the PhD thesis was to develop a pancreatic cancer risk prediction model based on established clinical risk factors, serum biomarkers, and genetic factors, to identify those people at high-risk of developing the disease within the general population, designed to inform screening decisions in a primary care setting.

Firstly, a population based pancreatic cancer audit using the Northern Ireland Cancer Registry was used to profile the characteristics of patients presenting with pancreatic cancer (n=539). Descriptive statistics were used to quantify the characteristics that patients with pancreatic cancer present with or are noted in the timeframe shortly before diagnosis. Results showed that the majority of patients were diagnosed at advanced stages (53% stage IV), indicating the need for improved early detection. Smoking and diabetes were prevalent, with smoking being more common than population estimates and diabetes being more common in early-stage patients. Symptoms were associated with late-stage diagnosis, with jaundice being more common in early stages.

Secondly, a systematic review was conducted to evaluate the existing literature on risk prediction models based on clinical risk factors for identifying high-risk individuals for pancreatic cancer. The review found 35 studies covering 40 models that developed, validated, or updated risk prediction models for pancreatic cancer based on clinical risk factors. The studies described risk prediction models for pancreatic cancer in the general population, patients with diabetes, and individuals with gastrointestinal symptoms. However, only eight models were externally validated and 13 were only internally validated. The discrimination of models varied from 0.61 to 0.98, which may be due to differences in the target population and predictors included in the models. The majority of studies were rated to have a high risk of bias due to poor reporting and methodological limitations based on Prediction Model Risk of Bias Assessment Tool. Six studies were judged to be at low risk of bias indicating an appropriate study design and selection of suitable established predictors.

Thirdly, by using the UK Biobank cohort data, which included 357,047 (357 pancreatic cancer cases) individuals aged 50-73 years, this thesis investigated the model performance of a risk prediction model based on established clinical risk factors to predict 5-year risk of pancreatic cancer. The clinical risk factors included in the model were age, body mass index, sex, smoking status, new-onset diabetes, and pancreatitis. The results demonstrated that the developed model had good model calibration and a modest discriminative ability, with an AUROC of 0.66 after correcting for optimism. A nomogram was created to facilitate clinical decision making, and selecting an example score of 12 or more to give a high specificity (97.7%), would result in low sensitivity (7.28%), leading to 293 additional screenings for each additional case of pancreatic cancer diagnosed. The absolute 5-year risk of pancreatic cancer was estimated to be 3.4 per 1,000 people.

In a subsequent chapter, to improve the model performance of the risk prediction model based on clinical risk factors only, the addition of the serum biomarkers, glucose and HbA1c, that had been previously shown by other studies to be strongly positively correlated to pancreatic cancer risk, was evaluated. The results showed that adding glucose and HbA1c to the clinical risk factor model led to a marginal improvement in model performance, with an optimism corrected AUROC of 0.67 compared to the model based solely on clinical risk factors (AUROC: 0.66). However, when least absolute shrinkage and selection operator (LASSO) regression was used to choose the biomarkers to be included in the model, the optimism corrected AUROC remained at 0.67 even after adding more biomarkers (glucose, HbA1c, cystatin C, IGF1, and urate), indicating the practicality of the simpler model.

Finally, a study incorporating genetic factors as a polygenic risk score (PRS) in addition to clinical risk factors and biomarkers improved model performance compared to a model with no PRS (optimism corrected Harrell's C-statistic: 0.70 versus 0.68). However, absolute risk remained low even when selecting the top 1% highest risk (absolute 5-year risk of 6.0 per 1000 people). An alternative two-stage approach was used to identify high-risk individuals, where moderate-risk individuals were pre-selected using a clinical risk factor model for blood and genetic testing. The second model incorporating clinical risk factors, biomarkers, and PRS was then used to stratify those pre-selected based on their clinical risk for further clinical investigation. The results showed a modest increase in absolute 5-year risk of 34.3 per 1000 people when using the top 1% as the cut-off for each of the stage. However, based on the net-benefit analysis at an absolute 5-year risk threshold of 34.3 per 1000 people, it appears that the approach is more beneficial than a strategy of either treating everyone or treating no one.

To conclude, the risk prediction model for pancreatic cancer using clinical risk factors, biomarkers, and genetic factors has been shown, in this thesis, to identify a cohort that has a higher prevalence of the disease than the general population. However, it is unsure whether the low absolute risk would be clinically useful in practice. The two-stage process showed potential in identifying high-risk individuals, but external validation is needed. Future research should focus on improving model performance and exploring the feasibility of the two-stage process. However, limitations in cost-effectiveness, impact on healthcare systems, and patient acceptance must be considered.

Thesis is embargoed until 31 July 2026.

Date of Award	Jul 2024
Original language	English
Awarding Institution	Queen's University Belfast
Sponsors	Northern Ireland Department for the Economy
Supervisor	Andrew Kunzmann (Supervisor), Helen Coleman (Supervisor) & Victoria Child (Supervisor)