Revolutionary AI tool forecasts pancreatic cancer risk up to three years in advance

Ground-breaking research spearheaded by Harvard Medical School, in collaboration with the University of Copenhagen, VA Boston Healthcare System, Dana-Farber Cancer Institute, and the Harvard T.H. Chan School of Public Health, has developed an artificial intelligence (AI) instrument capable of identifying individuals at the greatest risk of developing pancreatic cancer up to three years before diagnosis, using solely their medical records.

The study, published in Nature Medicine on May 8, indicates that implementing AI-driven population screening could be a key strategy in detecting those at a high risk of pancreatic cancer earlier. This could, in turn, hasten the diagnosis of a condition often detected at advanced stages when treatment options are less effective, resulting in poorer outcomes. Pancreatic cancer, one of the world’s deadliest malignancies, is anticipated to increase its mortality toll.

At present, there is an absence of population-wide screening tools for pancreatic cancer. Targeted screenings are performed for individuals with certain genetic mutations or a family history that increases their risk of developing the disease. However, these screenings may overlook other cases not fitting these criteria, the researchers highlighted.

The study’s co-senior investigator, Chris Sander, a faculty member in the Department of Systems Biology at the Blavatnik Institute at HMS, underscored the significance of the AI tool. “Deciding who is at a high risk for a disease and would benefit from additional testing is one of the most challenging determinations clinicians have to make. The tests can be more invasive, more costly, and carry their own risks. An AI tool that accurately identifies those at the highest risk for pancreatic cancer and who would gain the most from additional tests could greatly enhance clinical decision-making.”

If implemented widely, this AI-driven method could expedite the detection of pancreatic cancer, lead to earlier treatment, and improve patient outcomes, possibly extending their life spans.

“AI-driven screening provides the opportunity to change the course of pancreatic cancer, a formidable disease that is exceptionally challenging to diagnose early and treat promptly,” said study co-senior investigator Søren Brunak, a professor of disease systems biology and research director at the Novo Nordisk Foundation Center for Protein Research at the University of Copenhagen.

In this novel study, the researchers trained the AI algorithm on two separate data sets, containing a total of 9 million patient records from Denmark and the United States. They instructed the AI model to identify potential signs of pancreatic cancer risk based on the data in the records.

The model could predict patients likely to develop pancreatic cancer in the future by identifying combinations of disease codes and the timing of their occurrence. Interestingly, many of the symptoms and disease codes were not directly related to or derived from the pancreas.

The researchers evaluated different versions of the AI models for their capacity to identify individuals at a heightened risk of disease development over different timescales – 6 months, one year, two years, and three years.

Overall, each iteration of the AI algorithm proved considerably more precise in predicting who would develop pancreatic cancer than current estimates of disease incidence in the general population. The researchers proposed that the model is likely as accurate in predicting disease onset as the existing genetic sequencing tests, which are generally only accessible to a small subset of patients in data sets.

Screening techniques for certain prevalent cancers, such as breast, cervix, and prostate cancer, rely on relatively straightforward and highly effective techniques, such as a mammogram, a Pap smear, and a blood test. These methods have significantly improved the outcomes for these diseases by ensuring early detection and intervention.

In contrast, pancreatic cancer poses greater challenges and costs in terms of screening and testing. Doctors predominantly focus on family history and the presence of genetic mutations. While these are crucial indicators of future risk, they often overlook many patients.

The AI tool presents a significant advantage in its potential applicability to any patient for whom health records and medical history are available, not solely those with a known family history or genetic predisposition for the disease. This is particularly important, the researchers noted, because many patients at a high risk may not be aware of their genetic predisposition or family history.

In the absence of clear indications that a person is at high risk for pancreatic cancer and without symptoms, clinicians may understandably hesitate to recommend more sophisticated and costlier testing methods such as CT scans, MRI, or endoscopic ultrasound.

When these tests are performed and suspicious lesions are detected, the patient must undergo a procedure to obtain a biopsy. Given its deep placement in the abdomen, the pancreas is difficult to reach and easy to inflame, leading to its nickname as “the angry organ.”

The researchers advocate for an AI tool that singles out those at the greatest risk for pancreatic cancer. This would ensure clinicians are testing the correct population, while also preventing others from undergoing unnecessary testing and additional procedures.

The survival rate for those diagnosed with pancreatic cancer in its early stages is about 44 percent, five years post-diagnosis. However, only 12 percent of cases are diagnosed at this stage. The survival rate decreases dramatically to 2 to 9 percent for those with tumours that have spread beyond their origin, the researchers estimated.

Chris Sander emphasised, “Despite significant advancements in surgical techniques, chemotherapy, and immunotherapy, the survival rate remains low. Therefore, besides advanced treatments, there’s a pressing need for better screening, more focused testing, and earlier diagnosis. This is where the AI-based approach serves as the initial critical step in this process.”

For the current study, the researchers created multiple versions of the AI model and trained them on the health records of 6.2 million patients from Denmark’s national health system over a 41-year span. Of these patients, 23,985 developed pancreatic cancer over time.

During the training, the algorithm identified patterns suggesting future pancreatic cancer risk based on disease trajectories. For instance, diagnoses such as gallstones, anaemia, type 2 diabetes, and other gastrointestinal-related issues pointed to a higher risk for pancreatic cancer within three years of evaluation.

Inflammation of the pancreas was a strong predictor of future pancreatic cancer within an even shorter time span of two years.

The researchers caution that none of these diagnoses on their own should be deemed indicative or causative of future pancreatic cancer. However, the pattern and sequence in which they occur over time provide clues for an AI-based surveillance model and could prompt physicians to closely monitor or test those at elevated risk.

Next, the researchers tested the best-performing algorithm on an entirely new set of patient records it had not previously seen — a U.S. Veterans Health Administration data set comprising nearly 3 million records over 21 years, including 3,864 individuals diagnosed with pancreatic cancer.

The tool’s predictive accuracy was somewhat lower on the US data set. The researchers attributed this to the shorter collection period and the different patient population profiles in the U.S. dataset compared to the Danish dataset.

When the algorithm was retrained from scratch on the U.S. dataset, its predictive accuracy improved. This, the researchers said, underscores the importance of training AI models on high quality, rich data and the necessity of access to large representative datasets of clinical records aggregated nationally and internationally.

In the absence of globally valid models, AI models should be trained on local health data to ensure their training reflects the specific characteristics of local populations.