Evidence Based Practice and Quality Assurance
Background and Goal of Study: To compare the usefulness of recursive partitioning (RP) to logistic regression (LR) analysis in studying the occurrence of unplanned hospital admission after outpatient surgery.
Materials and Methods: A case-control study had been conducted at a US university hospital and analyzed with LR. The data set was subjected to RP: 396 unplanned admissions (cases) and 396 patients not admitted (controls); data included age, gender, ASA physical status, anesthetic method, duration of anesthesia, surgical severity (5-point clinical-invasiveness scale), and whether anesthesia ended after 1600 h. RP algorithm CART (Classification & Regression Trees) was used to repetitively (recursively) divide (partition) the study population into subgroups in relation to the outcome; Fisher's Exact Test was used for hypothesis testing, with P < 0.05 considered to indicate statistical significance.
Results and Discussion: CART correctly classified outcomes of 80% of subjects, identifying a set of predictors for unplanned admission similar to those identified with LR: Anesthetic method (general or spinal/epidural anesthesia vs. alternate methods) and duration were the strongest predictors, with smaller contributions from age, gender, and surgical severity. General or spinal/epidural anesthesia was associated with a 5-fold higher rate of admission compared to alternate methods (P < 0.0001), which was influenced by duration of anesthesia. Even among shorter cases, admission was more likely for older men having higher severity surgery.
Conclusions: RP provided a more intuitive and patient-centric view of risk of unplanned hospital admission after outpatient surgery. Its tree-like output identified sources of unplanned admission as the method exploited lack of homogeneity among the population and interactions among risk factors. Indeed, different subgroups had somewhat different risk factors for the same outcome. While LR presents relative importance of risk factors succinctly (e.g., odds ratios), such a presentation inherently averages influence of risk factors over the entire population and ignores heterogeneity of the population and possibility of interactions among risk factors. RP and LR complement each other in providing a more comprehensive, informative perspective that is likely to be more helpful in quality improvement efforts.