Machine learning models like “random forest” may improve risk adjustment in sepsis patients by identifying important variables better than traditional models.
Elizabeth Munroe, MD
A study by Elizabeth Munroe, MD, and colleagues sought to assess risk-adjustment mortality models in patients with sepsis. “Traditional models do not fully incorporate acute physiology,” she outlined at the 2023 American Thoracic Society International Conference, held from May 19-24, in Washington, DC. “Perhaps new machine learning models can do this better.” The researchers used data from the Michigan Hospital Medicine Safety Consortium sepsis registry, including 5,303 cases of sepsis hospitalization in 31 hospitals across Michigan. The 90-day mortality rate of the sample was 27.0%. Several machine learning models were compared with a more traditional stepwise logistic regression model for the primary outcome of 90-day mortality. The models included patient characteristics, comorbidities, and parameters of acute physiology.
The stepwise logistic regression model had an area under the curve (AUC) of 0.77, whereas the best-performing machine learning model, called “random forest,” had an AUC of 0.90. Dr. Munroe explained that a disadvantage of the traditional model is its limited ability to assess interactions. In contrast, the more complex random forest accounts for interactions. Furthermore, random forest identified additional variables of importance for mortality in sepsis, such as creatinine and bilirubin levels, functional limitations, and dementia.
“Risk adjustment is important in sepsis quality improvement,” according to Dr. Munroe. “This study showed that machine learning models may help to improve risk adjustment in this population. A next step may be to use the variables that were identified by machine learning models to improve our traditional model.”
Copyright ©2023 Medicom Medical Publishers
