Predicting Patient Outcomes

By designing algorithms to sift through reams of data in clinical records, the team is helping clinicians find better ways to predict which patients will experience complications and then intervene to keep them healthy.

In a recent project with Duke Health colleagues, the researchers analyzed data from five years’ worth of electronic health records gathered by the Southeastern Diabetes Initiative (taking care to protect individual privacy). With more than 16,000 records and thousands of potential parameters to consider, they enlisted help from a student Data+ team to reduce the data to a count of the number of times patients visited the hospital, took a medication or had a procedure.

“It turns out we can do pretty well, even with such a simplified dataset,” said Henao. “With those metrics, our machine learning algorithm could predict any one of 13 potential comorbidities over a six-month period better than existing tools.”

Now, with colleagues Katherine Heller in Statistics and Erich Huang in Biostatistics & Bioinformatics, they’re taking a similar approach to help Duke Health improve overall care quality and reduce medical costs. Using a massive database of insurance claims, the team built a model that combines data on diagnostics, medications and procedures with demographic information to predict which patients are at risk of admission and readmission within six months. The early results proved promising, encouraging Henao to begin refining the model.

“If we can identify the patients most at risk and prioritize their follow-ups, we can raise the level of care while decreasing hospital costs for everyone,” said Henao.