Dr. Ananya Malhotra speaks to News-Medical about her latest research into pancreatic cancer, and how its prognosis could be improved by using artificial intelligence.
What led you to begin this research?
The prognosis of several cancers, including pancreatic tumors, has hardly improved in the last decades, contrasting with a general dramatic increase in survival for most cancers. We felt that a new approach was required.
Image Credit: Kateryna Kon/Shutterstock.com
Why is it problematic to screen everyone for pancreatic cancer?
Pancreatic cancer is a very rare disease (8-12 new cases diagnosed every year in a population of 100,000). Due to such a low incidence of this disease, screening the whole population is neither practical nor appropriate. Current diagnostic markers are very expensive per cancer and are challenged by cost and morbidity associated with invasive testing.
Why is screening important?
Cancer screening means testing for early signs of cancer in people who do not have any symptoms. The aim is to help to pick up cancers early, which means treatment may be more successful in terms of a higher chance of survival.
What are the requirements for screening and how does this relate to pancreatic cancer?
According to WHO, the criteria of an effective screening program can be found at - https://www.who.int/cancer/prevention/diagnosis-screening/screening/en/
To advocate a screening model for pancreatic cancer, given the very low prevalence of this disease, it will be most effective if we implement a targeted screening on a high-risk group of people.
Several biomarkers have been looked at, potentially detectable in urine.
We are not involved in any of this research and are not saying anything about which test or biomarker should be used, simply that if there were an accurate non-invasive test developed, the algorithm we are developing could be paired with it to facilitate targeted screening, i.e. screening among people who present increased likelihood to develop pancreatic cancer.
Image Credit: Carl Dupont/Shutterstock.com
What symptoms have been linked to early signs of pancreatic cancer? How could AI analyze these to determine high-risk individuals?
We included all symptoms that had previously been reported as having an association with pancreatic cancer in the literature. We also included others that occurred with a high frequency within health data of patients, some of them being diagnosed later with pancreatic cancer.
So, we included most of the things you would expect to see like abdominal pain, back pain, anemia, weight loss, diabetes, and jaundice as well as a few others which may not be so obvious, such as insomnia, fatigue, depression. The model predicts risk status based on all variables, so it is inappropriate to highlight just one or two which are particularly important in determining individuals at high-risk.
How did you test out the AI?
Pancreatic cancer patients are usually diagnosed when it is too late, but most of them have consulted their GP for non-specific reasons up to a couple of years before the cancer diagnosis. However, these individual symptoms are not associated with any increased risk of pancreatic cancer. We assume that it is possible to find a combination of such (non-specific) symptoms which will be associated with a higher risk of pancreatic cancer.
In order to identify this combination, we have used AI which allows the machine to study the data, look for patterns in data and make better decisions in the future based on the examples that we provide. These algorithms apply what has been learned in the past to new data to predict future events.
Compared to conventional programming techniques, machine learning algorithms automatically formulate the rules from the data, which is a lot more powerful. Since we are dealing with large volumes of data, the human eye cannot identify data trends and code rules.
Image Credit: everything possible/Shutterstock.com
What were your findings?
The pilot study found that in people under 60 years of age, the model could predict who was at higher risk of pancreatic cancer up to 20 months before diagnosis. Our model has estimated that around 1,500 tests need to be performed to save one life from pancreatic cancer. This is unlikely to be small enough to make screening viable just yet.
However, preliminary findings show that AI holds the potential to scale 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 from the general population, which is what we plan to do next (in the current study, the controls had other types of cancer).
How could this technology be used in the future to combat pancreatic cancer?
AI is an incredible tool for automated pattern recognition and regularities in the data. Using this information, the machine learning algorithms act and predict outcomes such as classifying data into different categories.
For example, in our project, we aim to develop an algorithm that produces a risk score for the incidence of pancreatic cancer in a patient. This risk score is based on all the symptoms experienced by patients who developed pancreatic cancer compared to those individuals who did not develop this disease.
What are the next steps for your research?
Our findings are based on patients diagnosed with pancreatic cancer between 2005-2010. We would like to extend our preliminary findings to a longer, more up-to-date data. We would also like to look at pre-diagnostic information before 2 years from diagnosis, which means we may be able to detect high-risk patients even earlier.
Our model has compromised on low specificity. The most probable cause of this could be the use of cancer patients as controls who share similar warning signs. Hence, we want to apply our methods to the general population.
Our results indicated the relative importance of variables like smoking and diabetes and considering that we would want to conduct a stratified analysis, by additionally matching controls on their smoking or diabetes status. At last, we would like to assess the relative cost-effectiveness of our methods to other screening programs.
Where can readers find more information?
We are in the process of submitting a research paper based on this study. However, more information can be found on our website-
About Dr. Ananya Malhotra
Dr. Ananya Malhotra is a research fellow at the London School of Hygiene & Tropical Medicine, working with the Inequalities in Cancer Outcomes Network.
Her current project is based on developing a machine learning algorithm that can facilitate early detection of pancreatic cancer. In July 2020, the preliminary findings of this study were presented at the ESMO World Congress on Gastrointestinal Cancer 2020 where it was well acclaimed.
She received a Young Investigator Award for the best abstract, as well as interviewed by ESMO Press & Media Affairs Committee for inclusion in their press program.
The press release was broadcasted on their News Forum and the same was published in EurekAlert!. Additionally, an ‘In Focus’ blog post was published by The American Journal of Managed Care.