By Michele McDonald
By analyzing huge swaths of patient data, computers are learning not only how to help doctors choose the best medical treatment for each patient, but also how to efficiently prepare medical bills and predict patients’ disabilities.
Dubbed “machine learning,” complex computer algorithms delve into data to boost individualized medicine, says Mason computational scientist Janusz Wojtusiak , director of the Machine Learning and Inference Laboratory  and the Center for Discovery Science and Health Informatics  at Mason’s College of Health and Human Services .
“Traditional research and traditional clinical trials focus on an average patient,” Wojtusiak says. “But if you’re a patient, you don’t care about the average patient. You want one specific treatment that will benefit you, have minimum risk, and will have the highest chance of a good outcome for you.”
Machine learning can help find the answer, Wojtusiak says. “It can be used to build individualized models. By observing patterns in past patients, machine learning can find out what are the best options for you, not an average.”
Complex algorithms help computers learn by doing, much like a child does with language or tasks, Wojtusiak says. The approach is similar to Amazon and Netflix when books and movies are recommended based on past choices.
The goal isn’t to replace the expertise of physicians but to give them the most complete information so they can make the best decision, Wojtusiak says. Inches-thick patient files are transformed into compact data. Lab tests, doctor’s notes, demographic information, and other details build an elaborate picture of that patient. When those data are combined with information about other patients, the computer can find new information that’s more than just numbers.
“With many machine learning methods, people analyze huge amounts of relatively simple data,” Wojtusiak says. “In health care, these are not just numbers, these are specific lab tests, specific diagnosis, specific treatments, specific notes that describe real patients. The true challenge is to create smart algorithms that understand what the data actually mean and put some meaning behind the data, and then be able to learn something better and easier from it.
Computers do more than create electronic records. Computers also can change the way patients live. One of the ways Wojtusiak’s team is doing this is through a pilot program with the Department of Veterans Affairs (VA).
The VA recently launched a two-year pilot program in Bay Pines, Florida, to help 45 veterans, who typically would live in nursing homes, stay in regular homes. These “foster” homes frequently are with family members, and the VA provides the medical care at these residences and requires they be extensively evaluated before they are approved as foster homes.
“[The pilot] started as an initiative to allow patients to stay where they like at a home instead of at an institution,” says Farrokh Alemi, an affiliate researcher at Mason’s Center for Discovery Science and Health Informatics and chief of performance improvement for the VA.
To assess how foster home patients are doing compared with those in nursing homes, the VA will ask questions to determine whether, for instance, patients fall more or less often or whether they are more or less depressed.
Alemi says they are using that information to evaluate foster home care as an alternative to nursing home care. “We’re looking at the outcomes of this care.”
Wojtusiak’s machines are crunching data that could be used to open up the program to 800 patients, Alemi says. “If the data support this, there will be rapid expansion of the program.”
Researchers at the Machine Learning and Inference Lab are also studying prostate cancer patients, comparing five different treatment options.
“The idea here is to find out which patients are similar, and then for similar patients, to find out what are the best treatment options,” Wojtusiak says. “Then when a new patient comes in, we can find out whether the patient is similar to those we analyzed before and whether specific results of comparison are helpful.”
In September 2012, Wojtusiak began work on another project with the VA that examines 2,000 heart failure patients. He projected that millions of lab tests and physician notes will be examined over the course of the study.
“We can identify what works and when it doesn’t work,” Alemi says. But machine learning can take the research further than that.
“We can predict who will live and who will die with great accuracy,” Alemi says. “It makes my hair stand on end. It’s an amazing time to be in health care.”
Despite the impact, people aren’t going to do what a computer advises unless they understand the science behind it. And it’s not enough for the algorithms to be accurate. The data need to be explained if they are to be accepted by physicians, according to Wojtusiak.
“It’s not a black box,” Wojtusiak says, comparing machine learning to the aviation device.
You can imagine that if you went to a surgeon and told him the “black box” suggests he operate on a patient, he or she is not likely to do it, says Wojtusiak. But if the computer model is able to show exact evidence behind the suggestion—based on specific patient factors and specific patterns based on previous patients—that the best course of action is to perform a specific type of a surgery, then the doctor is more likely to act on it.
In the end, all the number crunching is worth it. “Health care is the perfect area for machine learning methods because it’s extremely complex on one side and it’s hard to deal with on the other side,” Wojtusiak says. “You can really help people.”