Engineering better medical outcomes

AVDr. Ali Vahdati, assistant professor of engineering, is working to improve precision medicine using computers and machine learning. His research using semantic data integration, standardization and dimensionality reduction to predict outcomes in bariatric surgery was recently published in the journal Computers in Biology and Medicine.

Along with collaborators at UNC Wilmington, UNC Greensboro and Harvard, Vahdati has developed software designed to help doctors and surgeons make more informed clinical decisions for specific patients. With electronic health records, Vahdati said, there is a lot of data available to physicians — but it comes from many different sources in different formats, with differing levels of quality.

“So if the physician wants to perform a surgery on the subject and wants to look at previous results to see what’s the best course of treatment for this patient, how can you utilize all of that data with a limited amount of time?” he said. “We were trying to help with that problem and alleviate some of the obstacles and challenges that physicians have.”

The first step the software undertakes is to integrate the data from different sources, formats and file types. The second step is standardization using the Unified Medical Language System, a dictionary of medical terminology.

“It’s a large amount of data, so we need to reduce the amount of data somehow to the most important attributes,” Vahdati said.

The software uses machine learning algorithms to find the most relevant attributes in the data to the specific problem so that finally, the data can be analyzed to provide the physician with the needed information.

For the published paper, the process was applied to bariatric surgery outcomes.

“There is a public database from the National Surgical Quality Improvement Program,” Vahdati said. “We tested it on this and … the algorithm was able to reduce the size of the dataset that we have from 136 megabytes to 5 megabytes. And then we reduced the number of attributes that we needed to search for from 250 to 20. On the same computer it took about 20 minutes to run before; when we reduced the dataset using machine learning it took less than a minute.”

The team is continuing to work to make the process easier for users without a background in computer science. The process could be useful in studying diabetes and heart disease, he said.

“The goal is to develop a better understanding of how the body operates and how we can prevent disease, how we can better diagnose it, and how we can have more successful surgery outcomes,” Vahdati said.

The full paper can be found at