MIT experts develop AI models that can detect pancreatic cancer early

Two machine learning models dubbed the “PRISM” neural network were created using five million patient health records.

Getty Images

Researchers at MIT’s CSAIL division, which focuses on computer engineering and AI development, built two machine learning algorithms that can detect pancreatic cancer at a higher threshold than current diagnostic standards. The two models together formed to create the “PRISM” neural network. It is designed to specifically detect pancreatic ductal adenocarcinoma (PDAC), the most prevalent form of pancreatic cancer.

The current standard PDAC screening criteria catches about 10 percent of cases in patients examined by professionals. In comparison, MIT’s PRISM was able to identify PDAC cases 35 percent of the time.

While using AI in the field of diagnostics is not an entirely new feat, MIT’s PRISM stands out because of how it was developed. The neural network was programmed based on access to diverse sets of real electronic health records from health institutions across the US. It was fed the data of over 5 million patient’s electronic health records, which researchers from the team said “surpassed the scale” of information fed to an AI model in this particular area of research. “The model uses routine clinical and lab data to make its predictions, and the diversity of the US population is a significant advancement over other PDAC models, which are usually confined to specific geographic regions like a few healthcare centers in the US,” Kai Jia, MIT CSAIL PhD senior author of the paper said.

MIT’s PRISM project started over six years ago. The motivation behind developing an algorithm that can detect PDAC early has a lot to do with the fact that most patients get diagnosed in the later stages of the cancer’s development — specifically about eighty percent are diagnosed far too late.

The AI works by analyzing patient demographics, previous diagnoses, current and previous medications in care plans and lab results. Collectively, the model works to predict the probability of cancer by analyzing electronic health record data in tandem with things like a patient’s age and certain risk factors evident in their lifestyle. Still, PRISM is still only able to help diagnose as many patients at the rate the AI can reach the masses. At the moment, the technology is bound to MIT labs and select patients in the US. The logistical challenge of scaling the AI will involve feeding the algorithm more diverse data sets and perhaps even global health profiles to increase accessibility.

Nonetheless, this isn't MIT’s first stab at developing an AI model that can predict cancer risk. It notably developed a way to train models how to predict the risk of breast cancer among women using mammogram records. In that line of research, MIT experts confirmed, the more diverse the data sets, the better the AI gets at diagnosing cancers across diverse races and populations. The continued development of AI models that can predict cancer probability will not only improve outcomes for patients if malignancy is identified earlier, it will also lessen the workload of overworked medical professionals. The market for AI in diagnostics is so ripe for change that it is piquing the interest of big tech commercial companies like IBM, which attempted to create an AI program that can detect breast cancer a year in advance.