Published on
Updated
Reading 4 mins.
A very recent study is formal: well-trained, artificial intelligence algorithms would be able to detect pancreatic cancer and predict whether patients will develop the disease, up to three years before a doctor can make the same diagnosis.
Pancreatic cancer is known to be one of the most dreaded: because it is often detected too late, its five-year survival rate is on average 12%. But in the face of this pessimistic outlook, Danish and American academics believe that artificial intelligence (AI) could help by detecting cancer at earlier stages and reliably predicting which patients are most at risk of developing the disease. Their work was published on Monday May 8 in the journal Nature.
More accurate than human algorithms in all tests performed
The researchers trained AI algorithms on 9 million medical records obtained from the Danish National Patient Registry and data from American veterans. The models were trained to correlate diagnostic codes (medical labels describing different conditions) with pancreatic cancer. Certain diagnostic codes for jaundice, abdominal and pelvic pain, weight loss, for example, are most closely related to the disease when found in patients about six months before diagnosis. Others like type 2 diabetes, anemia or inflammation of the pancreas is usually found earlier.
The researchers “asked” the AI model to look for telltale signs based on the data in the recordings. Based on the combinations of disease codes and when they occurred, the model was able to predict which patients are likely to develop pancreatic cancer in the future.
The researchers tested different versions of the AI models for their ability to detect people at high risk of developing disease on different time scales: 6 months, one year, two years and three years.
Overall, each version of the AI algorithm was significantly more accurate in predicting who would develop pancreatic cancer than current estimates without AI. The researchers said they believe the model was at least as accurate in predicting the onset of the disease as current genetic sequencing tests which are generally only available for a small subset of patients in data sets. .
AI predicts, but needs to be perfected
According to the documented tests, the AI would therefore be able to learn from the signs of the human body that can be linked to gradual changes, signs of pancreatic cancer. However, the AI cannot determine the cause.
“This model is still in its infancy, and while the AI system can make reasonably accurate predictions, it cannot, or currently cannot, identify causal mechanisms or events. As often in science, correlation is useful for prediction, but causality is much more difficult to establish”said Chris Sander, co-lead researcher of the study.
To be even more accurate and reliable, future AI-based screening tools for pancreatic cancer will need to be trained on local population-specific data, which presents a new challenge. A model trained on data from Danish patients, for example, was not as accurate when applied to American patients. “Given the experience of Denmark and one or two US healthcare systems, this means that in each country with different conditions and systems, it is best to retrain the model locally. AI needs a lot of data to train in. Access in different locations is not easy, as medical records are and should be confidential, so local approval and data security are essential”Sander said.
With artificial intelligence, another way to approach cancer?
The study demonstrates, however, that AI used on actual clinical records has the potential to “to shift the focus from late-stage to early-stage cancer treatment, improve patients’ quality of life and increase the benefit/cost ratio of cancer care, we read in the press release published with the study. Unfortunately, the software used in the study cannot yet be used to run screening programs. Improvements are needed, but it promises to be a concrete and applicable hope in the not so distant future:
“Once a monitoring program is in place, the actual computational costs to apply the software are moderate. Training is what consumes considerable computing resources. Actual clinical testing to see early signs of cancer or to detecting a cancer when it is still very small is expensive”, added Sander.
Still, the team thinks that as technology improves and operating costs come down, AI could become a valuable screening tool in the future.
“Many types of cancer, especially those that are difficult to identify and treat early, exert a disproportionate impact on patients, families and the healthcare system as a whole. AI-based screening is an opportunity to alter the trajectory of pancreatic cancer, an aggressive disease notoriously difficult to diagnose early and treat quickly when the chances of success are highest.”concluded Søren Brunak, co-lead researcher of the study.