September 1, 2022 – It’s hard to work out what the road ahead will look like for a cancer patient. Much evidence is considered, such as the patient’s health and family history, grade and stage of the tumor, and characteristics of the cancer cells. But in the end, the outlook comes down to health professionals analyzing the facts.
That can lead to “large-scale variability,” says Faisal Mahmood, PhD, an assistant professor in the Division of Computational Pathology at Brigham and Women’s Hospital. Patients with similar cancers can have very different prognosis, with some being more (or less) accurate than others, he says.
So he and his team have developed an artificial intelligence (AI) program that can form a more objective — and potentially more accurate — assessment. The aim of the study was to determine whether the AI was a workable idea, and the team’s results were published in cancer cell.
And because prognosis is key in determining treatments, greater accuracy may mean greater treatment success, Mahmood says.
“[This technology] has the potential to generate more objective risk assessments and subsequently more objective treatment decisions,” he says.
Building the AI
The researchers developed the AI using data from The Cancer Genome Atlas, a public catalog of profiles of various cancers.
Their algorithm predicts cancer outcomes based on: histology (a description of the tumor and how fast the cancer cells are likely to grow) and genomics (using DNA sequencing to tumor at the molecular level). Histology has been the diagnostic standard for more than 100 years, as genomics is increasingly used, notes Mahmood.
“Both are now commonly used for diagnosis in major cancer centers,” he says.
To test the algorithm, the researchers chose the 14 cancer types with the most data available. When histology and genomics were combined, the algorithm gave more accurate predictions than with either information source alone.
Not only that, but the AI used other markers — such as the patient’s immune response to treatment — without being told, the researchers found. This could mean that the AI can discover new markers that we don’t even know about yet, Mahmood says.
While more research is needed, including large-scale trials and clinical trials, Mahmood is confident that this technology will be used for real patients someday, probably in the next 10 years.
“In the future, we will see large-scale AI models that can incorporate data from multiple modalities,” he says, such as radiology, pathology, genomics, medical records and family history.
The more information the AI can process, the more accurate the assessment will be, Mahmood says.
“Then we can continuously assess patient risk in a computational, objective way.”