Mortality risk model combines EHR data, language processing to account for frailty

Applying natural language learning and deep neural networks to mortality risk models could help predict cardiovascular outcomes with more accuracy than modern support vector machines, researchers said at the 67th annual American College of Cardiology conference in Orlando.

In an effort to represent heart patients more wholly and predict postoperative outcomes after major cardiovascular procedures, Yijan Shao, PhD, and colleagues designed a risk model that used electronic health record (EHR) data to estimate frailty—a factor crucial to the care of elderly patients but one that’s frequently unaccounted for.

Death after major cardiovascular procedures is common among older patients, Shao said, but frailty isn’t taken into consideration in clinical prediction models.

“The way you build models, it’s prespecified usually,” she said. “This is a more unbiased approach.”

The researchers built a frailty ontology database by extracting more than 2.5 million text snippets from clinical notes entered into EHR systems, later narrowing the field to 1,000 of the most popular terms and phrases associated with frailty. Though the former step was carried out by a computer, Shao said human experts were the ones who annotated the remaining thousand terms, and a natural language processing classifier was trained on the annotated snippets. When applied, the classifier identified 855,000 positive mentions.

Using EHR data to create a system for quantifying frailty means clinicians are able to account for frailty without additional, time-consuming tests and can account for time-varying patient characteristics. In the future, Shao said, it could be implemented at the point of care.

Shao and colleagues applied the mechanism to two machine learning methods—a deep neural network and support vector machine—in a study population of 21,355 American veterans who underwent their first major cardiac procedure in 2014. Participants were studied for a period of one year, during which 6.8 percent of the group died.

Deep neural networks were the most successful in predicting mortality in these patients, Shao said, and saw higher levels of accuracy than support vector machine models. Factors other than frailty were taken into consideration, including all-cause hospitalizations and medication orders, but frailty was beneficial for clinical decision-making.

“A deep neural network approach amy help achieve better performance than traditional machine learning methods on predicting postoperative outcomes,” Shao said. “Longitudinal data, which naturally fit to the convolutional neural network, can be effectively used to improve prediction performance.”