The story of understanding and battling pancreatic cancer, a deadly and elusive disease, goes back to the 18th century.
The key to beating this cancer lies in early detection, but that’s tricky because the pancreas is hidden deep inside the abdomen, making it hard to spot problems early on.
Scientists from MIT and elsewhere embarked on a mission to identify patients at high risk for the most common type of pancreatic cancer, pancreatic ductal adenocarcinoma (PDAC).
They developed two advanced machine-learning models to detect this cancer early.
To make sure their models were robust and applicable to a wide range of people, the team tapped into a vast database of health records from various U.S. institutions.
This collaboration ensured the models were reliable for different populations and locations.
The two models they created, the PRISM neural network, and a logistic regression model, proved to be more effective than current screening methods.
While standard screening catches about 10% of PDAC cases, PRISM can detect 35% of cases, using the same risk criteria.
Using AI to spot cancer risks isn’t new. It’s been done with mammograms, lung CT scans, Pap smears, and HPV tests.
However, PRISM stands out because it was developed and tested on an extensive database of over 5 million patients, a much larger scale than previous studies.
The PRISM models analyze regular clinical and lab data, offering predictions based on a diverse American population. This wide-ranging approach is a step forward compared to other models that focus on specific regions.
Kai Jia, an MIT Ph.D. student, highlighted how unique regularization techniques in training made the models more generalizable and easier to understand.
David Avigan, a Harvard Medical School professor, believes this method could lead to new ways to identify and screen high-risk patients for early intervention.
The idea for PRISM came from a need to improve diagnosis. Currently, about 80-85% of pancreatic cancer patients are diagnosed too late for a cure.
The team hoped that clues hidden in electronic health records (EHRs) could serve as early warning signs. They used these records to train their models, offering a scalable way to predict risks in healthcare.
PRISM’s two models analyze EHR data like patient age, medical history, and lab results. The neural network model, PrismNN, uses complex patterns in the data to give a risk score for PDAC, while PrismLR uses logistic regression for a simpler probability score.
Understanding how these models make their predictions, known as interpretability, is crucial for gaining doctors’ trust.
The team managed to narrow down thousands of potential predictive features in a patient’s EHR to about 85 key indicators, including age, diabetes diagnosis, and frequent doctor visits.
These indicators are automatically identified by the model and align with what doctors know about pancreatic cancer risks.
The PRISM models still have a way to go. They currently rely on U.S. data, so they need testing and tweaking for worldwide use.
The team’s future goals include broadening the model’s reach to international data and adding more biomarkers for a finer risk assessment.
Their ultimate aim is to integrate these models into everyday healthcare. They envision a system where the models work quietly in the background, analyzing patient data and alerting doctors to high-risk cases without extra work for the physicians.
This integration could lead to early interventions for high-risk patients, potentially saving lives before symptoms even appear.
This groundbreaking work, authored by Jia, Appelbaum, and MIT Professor Martin Rinard, paves the way for AI to play a vital role in detecting and managing one of the most challenging cancers, offering hope for earlier, life-saving treatments.
If you care about cancer, please read studies about a new method to treat cancer effectively, and this low-dose, four-drug combo may block cancer spread.
For more information about cancer prevention, please see recent studies about nutrient in fish that can be a poison for cancer, and results showing this daily vitamin is critical to cancer prevention.
The research findings can be found in eBioMedicine.
Copyright © 2024 Knowridge Science Report. All rights reserved.