Helping Doctors Decide

When Mike Hughes was studying machine learning for his computer science PhD, it seemed like many of his peers were focused on cats. “So much of what people were working on was things like object recognition in images—is there a cat in this cellphone picture or not?” Hughes recalled. He wanted more. He wanted to really help people.

Now, a little less than a year into his assistant professorship at Tufts, Hughes is building tools to help doctors make better decisions in the intensive care unit. By applying machine learning and statistical modeling to data already being collected during treatment—like blood pressure and heart rate—he can start to answer complex questions like, Can this patient be taken off intubation in the next few hours? “Because there’s a huge cost to taking somebody off and then realizing, Oh no, they’re struggling,” Hughes explained. He is also doing similar work with antidepressant recommendations for patients with major depression, a disease notoriously difficult to medicate.

Deployment of either of these systems is still a few years off, Hughes warned. But one big benefit of using readily available information from ICUs and psychiatrists is that once the tools are ready, they might be able to quickly make a difference. “There’s no additional input that doctors need to provide,” he said. “Literally in any hospital in the US that has an electronic records system we could, in short order, obtain this kind of input for any subject they were interested in.”

Building Better Tools

Professor Lenor Cowen with students

Computer Science Professor Lenore Cowen with students. (Photo: Anna Miller/Tufts University)

Big data is so big these days that researchers who want to use it don’t always have all the computational tools they need to make the best sense of it. That’s where people like Professor Lenore Cowen step in. Using graph theory and machine learning, she’s trying to find ways for scientists to better study how networks of genes might be connected in various disease states. In Facebook terms, it’s a little like looking at a chart of your different social connections. “If I have a list of all the people that you’re friends with, I might be able to infer some very interesting things about you, right?” Cowen said. “If all these other genes are also involved in a disease, then even though I didn’t know you were involved in this disease, maybe I can predict you are.”

Gene maps like these would give medical researchers much deeper insights into complicated, multi-gene diseases, which actually comprise the bulk of diseases out there—very few conditions come from only one easily identifiable gene, like Huntington’s disease. Cowen doesn’t investigate the medical research herself, but she makes her programs available to other scientists who do. “I am more building a toolbox,” she said, “and then people can take those tools and use them for whatever disease they’re interested in.”


Portrait of Professor Donna Slonlm

Donna Slonim, a professor of computer science. (Photo: Kelvin Ma/Tufts University)

Mitigating Down Syndrome

About a decade ago, Donna Slonim, a professor of computer science with a secondary appointment at the School of Medicine, began collaborating with a colleague, Diana Bianchi, now director of the National Institute of Child Health and Human Development at the NIH. The two of them were trying to find unique gene expressions of Down syndrome in the amniotic fluid of pregnant women. The work was meant to provide evidence that such an approach could yield useful information—Down syndrome, after all, was easy to validate with existing tests—but it revealed more than they expected. The researchers found signs that fetuses with Down syndrome were experiencing unusually high levels of oxidative stress, and suspected that could contribute to some of the neurological issues that characterize the condition.

The project snowballed. From there, Slonim and Bianchi were able to tap into big data resources like the Connectivity Map, an MIT-developed tool that links genetic information with existing drug effects, to start identifying certain drugs that might mitigate Down’s severity. It’s nothing miraculous, Slonim cautioned, mostly things like antioxidants, which are already found in food. “If somebody has a third copy of a chromosome, they’re still going to have a third copy of a chromosome,” she said. But there’s an opportunity to “address what’s going on during development and help nudge things in a good direction so maybe they don’t have so many hard problems.”

Visualizing Flu Season

Every year, the Centers for Disease Control and Prevention hosts a flu season forecast challenge, inviting academic labs to send their best-guess models for how the year’s outbreak will play out. But the resulting tangle of data can be hard to make sense of, especially for policy makers who could use better models to make key decisions such as when to release a flu vaccine.

“You end up with just a bunch of stuff that is plotted on top of itself, and it makes it very hard to see what’s going on or compare individual forecasts,” explained Andrea Brennen, a part-time student at Tufts with a particular interest in Remco Chang’s data visualization lab, as well as director of design and data visualization at In-Q-Tel Labs, which (as part of parent company In-Q-Tel) explores emerging technology areas of interest to national security.

Last year, Brennen teamed up with a public health group within In-Q-Tel to find a way to better communicate this mass of flu forecasts. The result, Viziflu, is an elegant, deceptively simple color-coded display mapping the probabilities of each model’s anticipated influenza forecast, stacked up against one another. It means users can easily see where the majority of models place the expected flu “peak week.”

It might seem simple, but it’s an answer to a problem Brennen sees a lot. “We are willing to invest so much energy and resources in building these really powerful analytic capabilities,” she said. “But then sometimes we don’t put quite enough emphasis on how to communicate the output of that work in a way that makes it really usable.” The team plans to keep working on the Viziflu project, and possibly expand it into other public health and epidemiological applications.

Shannon Fischer is a freelance writer and frequent contributor to Tufts Magazine. Send comments to