Can Artificial Intelligence Detect Early-Stage Pancreatic Cancer?

July 2, 2020

Using artificial intelligence to comb through electronic health records to look for the subtle signs that individuals often present with to their doctor in the years before a more advanced diagnosis of pancreatic cancer is made showed promise in a early trial to see if earlier detection is possible.

Could artificial intelligence (AI) be used to identify patients with early-stage pancreatic cancer close to 2 years before the disease becomes apparent, and thus harder to treat?

A presentation at the ESMO World Congress on Gastrointestinal Cancer Thursday suggested that using AI to comb through electronic health records (EHRs) to look for the subtle signs that individuals often present with to their doctor in the years leading up to diagnosis is a possibility but will need a larger trial.

Pancreatic cancer is typically diagnosed in more advanced stages, with few treatment options and low survival rates; fewer than 1 in 10 patients live for 5 years or more after diagnosis. Globally, there were 458,918 new cases reported in 2018; in the United States, it affects nearly 57,000 Americans annually.

Nonspecific symptoms, including gastrointestinal issues and back pain, may begin more frequently in years prior to diagnosis, said the authors; other early signs are stool changes, itchy skin, poor appetite, and weight loss. Considered separately, these symptoms are unlikely to trigger further investigations.

The preliminary study used EHRs from the UK’s Clinical Practice Research Datalink (CPRD), which collects deidentified patient data from a network of general practitioner practices; the primary care data are then linked to other health records, in this case, national cancer registrations.

The data set included information about 1378 patients aged 15 to 99 diagnosed with pancreatic cancer from 2005 to 2010. Each patient was matched by age and sex to 4 individuals who did not get pancreatic cancer. Disease, symptom, and prescription drug codes for the 24 months prior to diagnosis were used to define 58 individual symptoms.

The researchers said in a statement they used machine learning to create logistic regression and random forest models. These models were trained on 75% of the data and tested on the remaining 25%.

Early results showed that pancreatic cancer diagnosis could be accurately predicted for 60% of the patients younger tha age 60, with an area under the curve of 61%, sensitivity of 76%, and a specificity of 45%.

"Our model has estimated that around 1,500 tests need to be performed to save one life from pancreatic cancer," said study author Ananya Malhotra, PhD, a research fellow in statistics, London School of Hygiene & Tropical Medicine.

However, that sample size is not enough to make screening a reality just yet, she said.

In addition, there is not a reliable screening test using blood. There is the carbohydrate antigen 19-9 assay, which may detect an antigen released by pancreatic cancer cells, but it is also released via other conditions and it is not expressed by every patient with cancer of the pancreas.

A minority of patients (about 4% to 7%) have a germline BRCA mutation; overall, the number of patients with a particular molecular alteration are too low to carry out a clinical trial.

Still Malhorta said the study shows "potential to narrow down the number of people we need to screen. We should be able to reduce this quite a lot further by matching pancreatic cancer patients to controls

"Pairing this predictive model with a non-invasive screening test, followed by scans and biopsies, could lead to earlier diagnosis for a significant proportion of patients and a greater number of patients surviving this cancer," added Malhotra. A future study will compare patients with pancreatic cancer to controls from the general population, as the controls in this study had other types of cancer.

Reference

Malhotra A, Rachet B, Bonaventure A, Pereira S, Woods L. 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. Presented at: ESMO World Congress on Gastrointestinal Cancer, July 1-4, 2020. Abstract SO-13.