COMPUTATION FOR THE BEDSIDE: Optimizing Patient Care
How some tools are already impacting patients
Medical decision-making is often more art than science, requiring physicians to exercise judgment in the face of complex factual circumstances. But now a few tools offer the opportunity to computationally optimize patient care. Here, we present several recent projects that have begun making a difference in patients’ lives. They span a range of medical scenarios affecting those with chronic illnesses, including HIV and kidney disease, as well as those undergoing CT scans or facial bone surgery.
Optimizing AIDS Treatment Protocols
People living with HIV often develop drug-resistant forms of the virus after a period of treatment. To select effective new drug cocktails, their doctors must filter through a mass of information about the patients’ viral mutations and load, past drug regimens, immune counts, and symptoms. This problem cries out for a computational solution. One such tool, HIV-TrePS, launched as a free online service in October 2010 and is already being used in over 40 countries.
Physicians can log in, input their patients’ baseline parameters (about 80 variables), and they will receive a list of drug combinations and an assessment of the probability that they will reduce the virus to undetectable levels, explains Brendan Larder, PhD, scientific chair of the HIV Resistance Response Database Initiative (RDI), a nonprofit research group in the United Kingdom which, along with several collaborators, created HIV-TrePS.
Larder and his colleagues trained the HIV-TrePS system using data from 60,000 real patients and a computational learning method called random forests, which builds decision trees and detects patterns. During testing and validation, the tool correctly predicted whether a combination therapy will lead to response or failure approximately 80 percent of the time. And physicians changed their treatment plans about one-third of the time based on its recommendations, Larder’s team has found.
In a few months, they will launch a similar system that doesn’t require viral genotyping data, which is unavailable in most developing countries. “The accuracy drops to around 70 to 75 percent,” Larder says, “but that’s still pretty good.”
Optimizing CT Images to Reduce Radiation Exposure
X-ray computed tomography (CT) scans provide useful, clear 2-D and 3-D images, but require using low doses of radiation. The question is: Can computers help radiologists get sufficiently clear images at the lowest necessary radiation dose? Already, computers can reduce radiation 20 to 30 percent by automatically adjusting to a patient’s body size and the thickness of the body part being scanned. But recent research at the Mayo Clinic in Rochester, Minnesota goes further.
“We showed you can use almost half the radiation dose with a lower tube energy and get diagnostically acceptable images,” says Joel Fletcher, MD, a radiologist at the Mayo Clinic. Lowering the tube energy makes the iodine contrast (which is injected into the patient) brighter, but it also makes the image noisy. To deal with that, Mayo has developed a novel computational way to de-noise the data. The work was published in Academic Radiology in October 2010.
Mayo’s novel method—projection-space de-noising—reads how the x-ray tube current is modulating to adapt to the patient size and shape, predicts where the noise will be, and corrects for it before the data have been reconstructed into an image for interpretation. The resulting images compared favorably with routine images reconstructed using a higher radiation dose.
It may be possible to lower the dose further if radiologists can increase their tolerance for noise. “There is evidence that radiologists are just as good at looking at noisy images, even without fancy de-noising software,” Fletcher says. It’s a tougher case to prove that with higher noise, diagnoses are still optimal. But that may be the optimization of the future.
Optimizing Facial Bone Replacements
Patients with cancer or injuries to the face sometimes need bones replaced. Today, surgeons fashion solutions based on what has worked before in other patients, not knowing for sure if the result will be strong enough for important functions, such as chewing, which can exert up to 700 Newtons of force. In recent work, researchers at Ohio State University used topological optimization to improve the design of these bone grafts.
“For each specific function, you can optimize the structure,” says Alok Sutradhar, PhD, a post-doctoral researcher and lead author on the paper published in Proceedings of the National Academy of Sciences in July 2010.
Topological optimization does not try to mimic structure that was there before. Starting with a rectangular prism filled with material, as well as boundary conditions that fix the height, the algorithm iteratively designs the optimal shape to withstand the loads (the direction and size of the forces) exerted by, say, chewing and swallowing. Videos accompanying the research show the gradual evolution of the optimal shape. “The algorithm is such that it will take you to the correct shape in less than an hour,” Sutradhar says.
Sutradhar’s group is now testing the optimized shape in a skeletal model, confirming that it does actually withstand the required forces. In the future, scientists might be able use tissue engineering to grow bones on a topologically optimized scaffold, he adds.
Nationwide Optimization of Live Kidney Donor/Recipient Matches
People who need kidney transplants often have friends and relatives who are willing to donate. They will even donate to a different person as long as their loved one gets a kidney out of the transaction. This might happen either in a “cycle” of multiple donor/recipient pairs (pair A gives to pair B who gives to pair C who gives to pair A) or in a chain (where an altruistic donor sets off a chain reaction of donation from one pair to the next and the next and so on).
Organ centers are therefore faced with the constant problem of matching multiple potential donors with multiple potential recipients. This scenario created a challenging computational problem that caught the attention of Tuomas Sandholm, PhD, professor of computer science at Carnegie Mellon University. He created an optimization algorithm that, following a successful pilot program, began being implemented nationwide in October 2010.
Unlike predecessors who have tackled the problem, Sandholm’s algorithm solves the problem optimally and scales to larger populations without any simplifications. All the possible combinations of cycles would be “more than the number of atoms in the universe,” Sandoholm says. “The algorithm has to prove that there are combinations that aren’t worth trying. Otherwise you’re dead in the water.” His algorithm identifies combinations of cycles and chains that you shouldn’t even try. The problem has interesting computational bits and pieces, Sandholm says. For example, he had to constrain the algorithm such that no donor gives out more than one kidney; assign weights to maximize better quality over lower quality matches; deal with the fact that computer memory was the bottleneck rather than speed; and design the rules of the exchange.
The work described here suggests additional opportunities: optimizing drug protocols for hepatitis; optimizing hip replacement designs; and optimizing liver and bone marrow transplants. And there may be other low-hanging fruit out there just waiting for someone to pluck them and find amazing satisfaction. As Sandholm says, “It’s very rewarding and unusual for a computer scientist to be able to save lives like this.”