Photo by Robert Hood / Fred Hutch News Service
In its April issue, the American Journal of Epidemiology recognized a paper by Fred Hutchinson Cancer Research Center faculty member Dr. Elizabeth Krakow as one of its 10 best articles published in 2017. Each spring, the editors of AJE select the top 10 articles published by the journal in the previous year that represent “the best in the field,” according to the journal’s announcement.
Besides Krakow’s article, fellow top-10s include research on hysterectomy and subsequent diabetes, exposures to electromagnetic fields from mobile phones, and alcohol use in the military.
In their winning paper, Krakow and colleagues show how a machine learning-based method for mining historical data could inform therapeutic choices, such as for preventing and treating a common complication of bone marrow transplant in an individual patient.
“Our work was meant as a demonstration project to show how disease registries could be leveraged to yield the same sorts of insights about personalizing the longitudinal sequence of treatments as one might get from a prospective, multistage randomized trial,” Krakow said in emailed comments about the work, which was based on the transplant complication known as graft-vs.-host disease.
When a cancer patient has a complex medical history, a little-studied disease and many potential treatment options, published guidelines and studies rarely indicate one clear treatment choice. So in everyday clinical practice, physicians recommend treatments to their patients based on an amalgam of published and anecdotal evidence, and patient-specific characteristics and preferences. Then, they adjust treatment as needed as time goes on.
In what Krakow calls “algorithm-informed treatment,” a computer would generate a treatment recommendation for a specific time point in an individual patient’s therapeutic course. The recommendation would be the product of a complicated mathematical formula the computer developed from studying the outcomes of treatment in a data set comprising many past, real-life patients.
“Our intent is to make the relationship between a combination of time-varying, patient-specific characteristics and the treatment decision explicit, to give clinicians a deeper evidence base to inform their everyday decisions,” said Krakow, who led this research as a master’s student under the guidance of Dr. Erica Moodie of McGill University.
In their paper, the researchers demonstrated how their algorithm could, for example, tell a physician that a particular regimen for GVHD prevention would give a patient a 16 percent higher chance of surviving two years with no signs of cancer as compared to other preventive treatments, taking into account aspects of that patient’s demographics, disease biology and treatment history. In contrast, another patient with the same diagnosis but other differences would probably do better with another preventive therapy, the algorithm showed.
The algorithm in the team’s paper was derived using certain medical data from more than 9,500 transplant patients in a research database managed by the Center for International Blood and Marrow Transplant Research. While the incomplete nature of these data means that it cannot be applied to clinical practice, Krakow said their approach can be used with other data sets to develop algorithms that could themselves be clinically validated and used in patient care.
Krakow is now developing algorithm-informed treatment methodologies for the care of patients with relapsed acute myeloid leukemia.
Susan Keown, a staff writer at Fred Hutchinson Cancer Research Center, has written about health and research topics for a variety of research institutions, including the National Institutes of Health and the Centers for Disease Control and Prevention. Reach her at email@example.com or on Twitter @sejkeown.
Are you interested in reprinting or republishing this story? Be our guest! We want to help connect people with the information they need. We just ask that you link back to the original article, preserve the author’s byline and refrain from making edits that alter the original context. Questions? Email us at firstname.lastname@example.org