An artificial intelligence (AI) tool has been developed that can predict the risk of pancreatic cancer using disease trajectories. The AI algorithm, published in Nature Medicine, could improve the detection of at-risk patients, increasing life expectancy and quality of life.
Pancreatic cancer prognosis
Pancreatic cancer is a leading cause of cancer-related mortality and incidence is increasing worldwide. At present, the disease is usually detected at a late stage, with around 80% of patients diagnosed with locally advanced or distant metastatic disease. At this late stage, long-term survival rates are low (5-year survival rate is 2-9% of patients).
Fortunately, prognosis is better in early-stage disease as patients can be cured using a combination of surgery, chemotherapy and radiotherapy. Improving the early detection of pancreatic cancer therefore has the potential to increase patient survival and reduce overall mortality.
Realistic risk prediction
The current study was a collaboration between researchers in Denmark and in the USA. The researchers set out to address the challenge of early pancreatic cancer detection in the general population. The development of a realistic risk prediction model such as this requires the selection of appropriate machine learning (ML) methods, that can work with large and noisy sequential datasets.
The team of researchers built on previous ML and AI methods based on clinical data and disease trajectories, which use information encoded in the time sequence of clinical events. An investigation was first performed using the Danish National Patient Registry (DNPR), which contains data for 8.6 million patients, and then on a smaller subset of patients from the United States Veterans Affairs (US-VA). A diverse set of ML methods were tested to optimise the extraction of information from these records, in the prediction of pancreatic cancer.
The results from the ML testing indicated that an AI-based screening method could be a valuable tool in identifying those with an increased risk of developing pancreatic cancer. Using the DNPR cohort, a high area under the receiver operating characteristic (AUROC) curve value of 0.88 was shown when looking at cancer occurrence within 36 months. The performance of the model dropped slightly when also considering the US-VA data, but still showed a respectable AUROC of 0.71.
Implementing real-world clinical prediction
An advantage of the AI and ML prediction presented here is that screening of large populations is inexpensive, which could then narrow down the number of individuals needed for more costly clinical screening. The performance of the models shown could be sufficient for initial screening, leading to real-world clinical prediction and surveillance programs. Chris Sander, co-senior investigator on the study, stated: “One of the most important decisions clinicians face day to day is who is at high risk for a disease, and who would benefit from further testing, which can also mean more invasive and more expensive procedures that carry their own risks. An AI tool that can zero in on those at highest risk for pancreatic cancer who stand to benefit most from further tests could go a long way toward improving clinical decision-making.”