Researchers at MIT were able to categorize 40% of cancers, which helped them “predict survival and the best treatments.”
For a small number of cancer patients, doctors can’t figure out where the disease started in the body.
Researchers at the Massachusetts Institute of Technology (MIT) have made an artificial intelligence model that looks at the patient’s genetic information and guesses where the tumor first showed. This will help doctors determine where cancers of unknown primary (CUP) originated.
A study in Nature Medicine found that when researchers used the new AI model on 900 patients with cancers of unknown cause, they correctly classified at least 40% of tumors.
Researchers say this knowledge could help doctors determine the best treatments for these cancer patients.
Related Posts
“Our study showed that the AI model we made, OncoNPC, can use routinely collected genomic data to help doctors decide how to treat patients with cancer of unknown primary (CUP) tumors, which are hard to diagnose and have few treatment options,” Intae Moon, the lead author of the study and an MIT graduate student in electrical engineering and computer science, told the media.
We demonstrated that CUP tumors share the same genetic and prognostic characteristics as their predicted cancer types and may benefit from treatments already available if they follow OnoNPC predictions.
The experts also found that 15% of the patients could have gotten better care if they had known where their cancer came from.
Researchers say there were problems with the study.
Even though the researchers used data from different places to train their AI model, Moon pointed out that the clinical data they used for their thorough study came from only one school.
This could make it harder to use the results in other places.
“Another limitation is that most of the training data patients were White (83.2%),” Moon said. “This could mean that the tool is more accurate for White patients.”
“Even though it still worked pretty well for other ethnicities, a deeper look is needed to make sure that the model is helping a wide range of patients.”
Also, only 22 of the most common kinds of cancer were used to identify tumors. This means that if a tumor is of a type not on the list, forecasts could be less accurate.
Moon said, “We hope to figure this out gradually as we get more and more information.”
“Finally, even though our results show that patients with CUPs who are grouped by our algorithm respond better to “matching” treatments, this is still a retrospective study,” he said.
“A prospective, randomized study would be needed to prove that the link is cause and effect.”
Moon said that the experts see OncoNPC as a tool to use with traditional cancer medicines, not as an alternative.
“It’s important to check the results of our study at different institutions,” he said. “This is one of the most important next steps.”
“In the long run, we hope that this will lead to more research on how different CUPs are and more treatment options.”
In the future, the researchers want to add unorganized data like pathological pictures and clinical notes to the AI to learn more about cancer.
Moon said, “This could make it better at doing many things, like directly predicting survival and the best treatments.”
“A crucial step”
The MIT study did not include Dr. Tinglong Dai, who teaches operations management and business analytics at the Johns Hopkins Carey Business School in Baltimore, Maryland. But he said this study is an “important step toward figuring out the best ways to treat people with cancer whose cause is unknown.”
He told Fox News Digital, “The results are encouraging, and they tell us important things about how we might handle such complicated cases.”
“However, it should be noted that the study was done in the past, so more field studies are needed to see how well it works in the real world.”
Dai said that how doctors combine and use these models in their daily work is crucial in using this method in the real world.
“Their acceptance and effective use of the model will be key to turning evidence into strategies that can be put into action,” he said.
Dai said again that adding random data, like pictures from pathology, could add more levels of information and improve the accuracy of predictions.
He also said, “Integrating multiple data sources will almost certainly result in a more robust approach.”
