Research Highlights
VA study applies 'natural language' technology to electronic medical records to find post-surgery complications
In a study at six VA medical centers, researchers used Google-like technology called "natural language processing" (NLP) to interpret free text in Veterans' electronic medical records—such as doctors' notes—and identify post-surgery complications. The new method, which outperformed an existing one that relies on administrative codes, could emerge as a powerful tool for quality improvement in VA's nationwide health system.
The findings appeared in the Aug. 24/31 issue of the Journal of the American Medical Association (JAMA).
The study used data on nearly 3,000 VA patients who underwent surgery between 1999 and 2006. Compared with a standard automated method that scans administrative data, NLP was better at picking up adverse post-surgery events such as lung, kidney, or heart problems. To provide a benchmark for both approaches, trained nurses manually reviewed the patient records and carefully looked for any clinical notes indicating complications.
In an accompanying JAMA editorial, Dr. Ashish Jha said the researchers, led by Dr. Harvey Murff at the Nashville VA Medical Center and Vanderbilt University, "push beyond the traditional uses of the [electronic health record] by demonstrating that natural language processing, when applied to electronic data, can help clinicians track adverse events after surgery." Jha, a health informatics expert with VA and Boston University, adds that while the study might seem esoteric to some, "its significance should not be underestimated. Instead, these findings suggest that [electronic health records] can transform health care delivery."
Usually, VA and other health systems rely on a set of patient safety indicators—developed by the Agency for Health Research and Quality in the early 2000s—to screen for surgery complications and other adverse events on a hospital—or system-wide basis. The indicators automatically scan billing data for diagnostic codes. They are used heavily in surveillance and quality-improvement efforts. The Centers for Medicare and Medicaid Services also use them to rank hospitals.
The indicators have drawbacks, though. At least two studies have questioned, for example, whether the indicators sometimes show conditions that existed before a patient entered the hospital, in which case they would not be valid markers of the quality of care.
The new study compared the indicator method against NLP, a burgeoning branch of computer science that teaches machines to make sense of human language. The science is already at work in everyday products such as Internet search engines or translation programs.
Health care researchers have been looking at the technology with a keen eye, especially in VA. The agency's pioneering electronic medical records system stores a wealth of patient data—including free text—that can be mined, with privacy safeguards, to improve patient safety and outcomes.
In the study, NLP was more "sensitive" than patient safety indicators in the six areas that the researchers chose to study: acute renal failure, pulmonary embolism, deep vein thrombosis, sepsis (a dangerous spread of bacteria following an infection), pneumonia, and heart attack. In other words, NLP was able to detect complications even in some cases where they were missed by the indicators. The NLP software read progress notes, imaging reports, discharge summaries, and other text entered in patients' records by clinicians.
On the downside, NLP was less "specific" than the indicator method-meaning, it produced some false alarms, perceiving a complication where none existed—but the difference was small. Moreover, the researchers suggest that NLP has advantages over the indicator method because search queries can be easily customized and refined to do an even more accurate job of finding problems. Another plus, say the researchers, is that NLP can potentially be used while a patient is still in the hospital to help doctors catch adverse events—something that would be less practical with the automated method that uses billing data.
This article originally appeared in the September 2011 issue of VA Research Currents.