“Superusers” of medical services are the small fraction of patients that account for huge consumption of medical services. An article published August 14, 2019 in JAMA Surgery (online) reports on the application of machine learning methods to Medicare data on 1,049,160 Medicare patients who underwent surgery, and were then tracked over the next year to assess their need for additional hospitalization. The factors most contributing to additional hospitalization were:
- hemiplegia/paraplegia
- weight loss
- congestive heart failure with chronic kidney disease stages
While individual factors driving superuser hospitalization had been previously identified, this was the first long-term longitudinal study, and also the first one to use machine learning data mining methods to identify interactions (like congestive heart failure with kidney disease). The method used was “logic forest,” an ensemble of logic regression models. Logic regression takes binary variables as predictors, and uses boolean combinations of them (AND, OR) in its prediction model. It produces trees similar to those in Classification and Regression Trees (CART) methods; for details see this article.