AI Predicts Future Pancreatic Cancer Risk

According to new research led by investigators at Harvard Medical School and the University of Copenhagen in collaboration with VA Boston Healthcare System, Dana-Farber Cancer Institute, and the Harvard T.H. Chan School of Public Health, an artificial intelligence tool successfully identified people at the highest risk for pancreatic cancer up to three years before diagnosis using only the patients’ medical records.

The findings, published on May 8 in Nature Medicine, suggest that AI-based population screening could be useful in identifying those at high risk for the disease and could speed up the diagnosis of a condition that is all too often discovered at advanced stages when treatment is less effective and outcomes are poor, according to the researchers. Pancreatic cancer is one of the deadliest malignancies in the world, and its death toll is expected to rise.

There are currently no population-based screening techniques for pancreatic cancer. Those with a family history of pancreatic cancer and certain genetic variants that predispose them to the disease are examined in a targeted manner. However, such targeted screens may miss cases that fall outside of those categories, according to the researchers.

“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,” said study co-senior investigator Chris Sander, faculty member in the Department of Systems Biology in the Blavatnik Institute at HMS. “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.”

Sander said that if used on a large scale, such a method could lead to early detection of pancreatic cancer, better outcomes, and longer patient lives.

“Many types of cancer, especially those hard to identify and treat early, exert a disproportionate toll on patients, families and the healthcare system as a whole,” said study co-senior investigator Søren Brunak, professor of disease systems biology and director of research at the Novo Nordisk Foundation Center for Protein Research at the University of Copenhagen.

“AI-based screening is an opportunity to alter the trajectory of pancreatic cancer, an aggressive disease that is notoriously hard to diagnose early and treat promptly when the chances for success are highest,” Brunak said.

The AI algorithm was trained on two independent data sets comprising 9 million patient records from Denmark and the United States in the new study. Based on the data in the records, the researchers “asked” the AI model to hunt for warning indications.
The program was able to forecast which patients are likely to acquire pancreatic cancer in the future based on combinations of illness codes and the timing of their emergence. Notably, many of the symptoms and disease codes had nothing to do with or were caused by the pancreas.

The researchers evaluated the AI models’ capacity to predict persons at high risk of illness development throughout time scales of six months, one year, two years, and three years.

Overall, each version of the AI system was significantly more accurate than current community-wide estimates of disease incidence — defined as how frequently a condition arises in a population over a certain time period — in predicting who would develop pancreatic cancer. The researchers believe the model is at least as accurate as current genetic sequencing procedures, which are often available only for a limited percentage of patients in data sets.

The so-called “angry organ”

Certain common malignancies, such as those of the breast, cervix, and prostate gland, can be detected using very basic and highly effective procedures, such as a mammography, Pap screening, and blood test.

These screening technologies have improved illness outcomes by allowing for early detection and intervention during the most curable stages.

Pancreatic cancer, on the other hand, is more difficult and costly to detect and test for. Physicians focus primarily on family history and the presence of genetic abnormalities, which, while crucial markers of future risk, frequently overlook many patients.

One notable feature of the AI tool is that it may be used on any patient with available health data and medical history, not simply those with a known family history or genetic susceptibility to the disease. This is especially crucial, according to the researchers, because many high-risk individuals may be unaware of their genetic susceptibility or family history.

In the absence of symptoms and without a clear indication that a patient is at high risk for pancreatic cancer, doctors may be understandably hesitant to offer more sophisticated and costly testing, such as CT scans, MRIs, or endoscopic ultrasounds.

When these tests are used and suspicious lesions are detected, the patient must go through a biopsy procedure. The organ is located deep within the abdomen and is difficult to stimulate and inflame. Because of its irritation, it has gained the nickname “the angry organ.”

According to the researchers, an AI tool that identifies those at highest risk for pancreatic cancer will ensure that clinicians test the proper population while sparing others from unnecessary testing and treatments.

Only around 44 percent of persons identified with pancreatic cancer in the early stages survive five years following diagnosis, but only about 12 percent of cases are diagnosed that early. According to the researchers, patients whose malignancies have spread beyond their original place of origin have a 2 to 9 percent chance of survival.

“That low survival rate is despite marked advances in surgical techniques, chemotherapy, and immunotherapy,” Sander said. “So, in addition to sophisticated treatments, there is a clear need for better screening, more targeted testing, and earlier diagnosis, and this where the AI-based approach comes in as the first critical step in this continuum.”

Previous Diagnosis Foreshadow Future Danger
The researchers created various versions of the AI model for the current study and trained them on the health information of 6.2 million patients from Denmark’s national health system over a 41-year period. Over time, 23,985 of those patients developed pancreatic cancer.

During training, the algorithm identified patterns suggestive of future pancreatic cancer risk based on disease trajectories, or whether the patient had particular symptoms that occurred in a specific order throughout time.

Diagnoses such as gallstones, anemia, type 2 diabetes, and other GI issues, for example, predicted a higher risk of pancreatic cancer within three years of examination.

Less surprisingly, pancreatic inflammation was substantially predictive of future pancreatic cancer within a two-year time frame.

The researchers emphasize that none of these diagnoses should be regarded as predictive or cause of future pancreatic cancer. However, the pattern and sequence in which they occur over time provide clues for an AI-based surveillance model and may alert physicians to actively monitor or test patients at high risk.

The best performing algorithm was then tested on an entirely new set of patient records – a U.S. Veterans Health Administration data set of approximately 3 million records spanning 21 years and includes 3,864 individuals diagnosed with pancreatic cancer.

On the US data set, the tool’s predicted accuracy was slightly lower. This was most likely due to the fact that the American dataset was obtained over a shorter period of time and featured slightly different patient demographic profiles — the total Danish population in the Danish data set vs current and former military people in the Veterans’ Affairs data set.

The algorithm’s predicted accuracy improved once it was retrained from scratch on the US dataset. According to the experts, this highlights two critical points: To begin, ensure that AI models are trained on high-quality, abundant data. Second, access to large representative datasets of clinical information aggregated nationally and internationally is required.

In the absence of globally valid models, AI models should be trained on local health data to guarantee that their training represents the unique characteristics of local populations.

Source Link