OR WAIT null SECS
No 2 hospitals are alike, and their predictive analytics tools should probably reflect that.
No 2 hospitals are alike, and the predictive analytics tools that they use must acknowledge that. In a new report, researchers claim that they have developed a generalizable model for predicting patients at risk of one of the most common—and serious—hospital-based infections, based on that hospital’s unique population.
The team analyzed electronic health records (EHR) data from nearly 200,000 admissions to University of Michigan Hospitals (UM) and more than 60,000 admissions to Massachusetts General Hospital (MGH), pulling out invariant (think demographic) and variant (like medical condition) metrics. They were looking to predict stratify patient risk for Clostridium difficile (C. difficile) infection, which impacts about half a million Americans each year.
And they believe the system that they built does just that. They pulled about 4,800 unique features from the UM dataset and roughly 1,800 from MGH. Applying an L2 regularized logistical regression to the models, they applied them to additional data from the 2 health systems.
The institutions had noticeably different populations and even EHR systems, but the methodology produced models that showed respectable area under the curve—0.82 for UM, 0.75 for MGH. In cases where C. difficile was correctly identified, the models detected it 5 days earlier, on average, than a doctor had diagnosed it. That gap in detection time isn’t just dramatic: It can reduce care costs and potentially save lives. It can also help hospital staff contain a potential infection, reducing the risk of it spreading to other patients.
“I see tools and approaches such as these as potentially very effective at reducing the incidence of C. difficile,” one of the authors, Erica Shenoy, MD, PhD, told our sister publication MD Magazine. “If we are able to diagnose patients earlier, we can implement effective infection prevention tools to prevent transmission to other patients.”
Although the 2 different models did share some predictive factors, the authors noted that many of the most important variables actually differed between the environments. “In contrast to traditional approaches that focus on developing models that apply across hospitals, our generalizable approach yields risk-stratification models tailored to an institution,” they wrote, which allows for better targeting of infection prevention strategies.