EHR Data May Predict Sepsis Mortality, Study Shows

Ken Terry

March 28, 2014

Vital signs and laboratory data in electronic health records (EHRs) can be used to predict whether hospitalized patients will develop the high blood lactate levels that are often associated with sepsis, according to a study performed at the University of California, Davis (UC Davis).

Using lactate levels along with arterial pressure and respiratory rate data, Eren Gultepe, a PhD candidate from the Department of Biomedical Engineering, UC Davis, and colleagues were able to predict the mortality of patients with sepsis in the majority of cases. Their study was published in the March issue of the Journal of the American Medical Informatics Association.

Patients with sepsis have a lower mortality risk if diagnosed earlier, but it is difficult to test all adults with suspected sepsis. The study was designed to find out whether data mining and machine learning could be used to identify patients at risk for hyperlactatemia. The researchers also hoped to raise compliance with treatment guidelines for severe sepsis, and thus improve patient outcomes.

The researchers used an EHR database containing 1492 patients who had symptoms of sepsis. They excluded patients whose records were incomplete with respect to 7 variables of interest (including vital signs, lactate level, and white blood count), leaving them with a sample of 741 patients.

Ilias Tagkopoulos, PhD, from the Department of Computer Science and the UC Davis Genome Center, and a study coauthor, acknowledged to Medscape Medical News that the quality of data in EHRs, in which information is often missing from structured fields or is incorrect, poses a serious obstacle to the development of this kind of clinical decision support system.

Whereas, in this retrospective study, the researchers could focus just on the cases that had adequate data, in a real-world setting, the predictive power of the approach would be considerably diluted if all the required data were not available on every patient, noted Dean Sittig, PhD, a professor at the University of Texas Health Sciences Center in Houston and an expert on clinical decision support.

Nevertheless, the study demonstrated the feasibility of creating a decision support tool for sepsis, Dr. Tagkopoulos said. "We used multiple machine learning algorithms to identify the important features from the data and use them to create predictive classifiers. Then, at the end, we performed statistical tests to evaluate how well the classifier was able to distinguish between patients who had high risk or low risk of mortality."

In an analysis including patients with and without sepsis, the researchers found that they could accurately predict lactate levels based on vital signs and white blood cell count (accuracy 0.99; discrimination, 1.00 area under the receiver operating characteristic curve [AUC]). When they considered only those patients with sepsis, the authors were able to predict mortality with an accuracy of 0.73 and discrimination of 0.73 AUC, using just 3 variables: median lactate level, mean arterial pressure, and median absolute deviation of the respiratory rate.

The research team plans to replicate the findings in larger studies, Dr. Tagkopoulos said. They will also expand their model to include additional information, such as the presence of organ dysfunction. Moreover, they will try to predict the effects of different actions within the sepsis protocol to find out when to apply antibiotics, when to place patients on intravenous fluids, and so forth.

Dr. Sittig told Medscape Medical News that he regards the UC Davis study as a "preliminary step" in the research that must be done before the machine learning techniques can be applied in acute care settings. One big caveat, he noted, is that the study failed to show that EHR data could be used to predict mortality from sepsis in time to do something about it.

In addition, although the authors suggest the model could reduce the number of false alerts because of its accuracy, the fact that half of the patients had to be eliminated from the study cohort because of poor data belies that assertion, Dr. Sittig said

"One of the reasons we have such bad alert fatigue now is that we don't have good data in the system," he noted.

Many researchers and companies are trying to use big data techniques, in which they apply supercomputers to large datasets, to analyze EHR data and extract patterns that can improve patient care, Dr. Sittig pointed out. Although most of these efforts have been in vain, he said, it makes sense to use computers to track variables associated with particular conditions because computers are better than humans at making decisions when multiple variables are involved.

Finally, Dr. Sittig noted that the current study is not the first attempt to use EHR data in sepsis care. Cerner, a leading EHR vendor, has already developed a decision support tool for sepsis and has distributed it to its customers for free. Dr. Sittig said his own hospital is testing this tool now.

"It identifies patients that are sick; it doesn't always identify people who have sepsis," he said. "We're trying to figure out whether we've affected the mortality rate. We're certainly sending more people to the [intensive care unit] faster than before, but we haven't shown that it has reduced the number of heart attacks or deaths in the hospital."

This work was supported by the Center for Information Technology Research in the Interest of Society and by the National Center for Advancing Translational Sciences, National Institutes of Health. The authors and Dr. Sittig have disclosed no relevant financial relationships.

J Am Med Inform Assoc. 2014;21:315-325. Abstract


Comments on Medscape are moderated and should be professional in tone and on topic. You must declare any conflicts of interest related to your comments and responses. Please see our Commenting Guide for further information. We reserve the right to remove posts at our sole discretion.