Practice variation in sepsis management in the eICU database

Abstract

Sepsis is a life-threatening syndrome triggered by an infection. Despite international guidelines, sepsis management varies between sites. This unwanted practice variation may affect negatively the quality of care but enables theoretically retrospective studies for finding optimal treatment strategies. The goal of the present master thesis was to find a relevant way to model practice variation in the management of sepsis-induced circulatory failure. Sepsis patients were retrieved from the eICU critical care database. Nine treatments and nine relevant covariates were selected from domain knowledge. Practice variation was successively investigated in four intensive care units and four hospitals respectively. Missing values were imputed using forward filling, linear interpolation and the Multiple Imputation by Chained Equations algorithm. For each analysis, two logistic regression models were successively trained and calibrated. The first model yielded propensity estimates for being treated in a particular site given covariates. The second model was trained on the subset of patients having a reasonable probability of being treated in all the sites and yielded propensity estimates for being treated with a particular treatment. Practice variation was first defined as the expected difference in propensity for treatment between two sites and then characterized for a given patient with given covariates as the distance between the likelihood to get a certain treatment in a certain site s and the expectation of the likelihood to get the same treatment over all the sites with the assumption that this patient had the same likelihood of being treated in s as he had in the actual data. A pairwise comparisons of propensity for treatments between sites revealed variations up to 12.5%. At a patient level, practice variation distributions showed a similar positively skewed distribution for both analyses and revealed variations up to 5.8%. We demonstrated the feasibility of modeling practice variation among distinct sites in the management of sepsis-induced circulatory failure using retrospective data. The significance of this variation should be further evaluated by investigating which treatment policies are associated with a better outcome.