(Reuters Health) - “Smart” search programs can ease the process of systematically reviewing new medical research, a key step in getting the best practices from laboratories to doctors' offices, U.S. researchers say.
The Institute of Medicine (now the National Academy of Medicine) says clinical practice guidelines should be based on a systematic review of the evidence, lead author Dr. Paul Shekelle from RAND Corporation in Santa Monica, California, told Reuters Health by email.
“We know that clinical practice guidelines go out of date over time, as new evidence accumulates. An impediment to the regular updating of clinical practice guidelines is the time and resources needed to update the systematic review,” he said.
Typically, researchers and their assistants perform computer searches to identify anywhere from a few to thousands of new research studies, then they determine which ones are relevant and assemble the information into updated guidelines and recommendations.
Shekelle and colleagues thought machines could do more of the job and do it faster, so they compared machine-learning methods with the standard search methods for identifying new information.
They tested the idea on three health conditions: gout, low bone density and osteoarthritis of the knee. The smart search program “learned” which key terms to look for by analyzing words from studies that were included in prior reviews on each topic.
In all three cases, computers - provided only with the titles and summaries of articles included in previous reviews - reduced the number of articles researchers had to screen further by 67 to 83 percent, according to the results in Annals of Internal Medicine.
The machine-learning method missed only two articles that humans would have identified, for an overall accuracy of 96 percent. And neither of these articles would have changed the ultimate evidence reviews, Shekelle’s team concludes.
“Machine learning methods are very promising as a way to reduce the amount of time and effort for the literature search, which in turn should make it easier to update the systematic review, which in turn can facilitate keeping clinical practice guidelines up to date,” Shekelle said.
The approach would “shorten the time from completion of research studies to adoption of effective treatments in clinical practice,” said Dr. Alfonso Iorio from McMaster University in Hamilton, Ontario, Canada, who coauthored an editorial accompanying the report.
“Also, it will allow more efficient update of doctors about what works and what does not, saving lives and dollars,” he said by email.
Iorio thinks this method would have the greatest impact in the fields of cardiology, diabetes, respiratory disease and cancer. “But any field would benefit,” he said.
“In the near future, artificial intelligence will also be used to match to individual need with the best available health care intervention - one necessary step to get this is proper classification on existing and newly generated knowledge,” Iorio said.
“The critical step is training properly the computer systems - we need to ensure research dollars are provided to ensure this training is done by serious and independent researchers and controlled by public institutions.”
SOURCE: bit.ly/2rslYlR and bit.ly/2rhBLA0 Annals of Internal Medicine, online June 13, 2017.