Bryan Lee, MD
Dr.
Cooper University Hospital
Camden, NJ, United States
Disclosure information not submitted.
Nitin Puri, MD
Director, Center for Critical Care Services
Cooper University Health Care
Camden, NJ
Disclosure information not submitted.
Sharad Patel, MD
Attending Physician
Cooper University Hospital, United States
Disclosure information not submitted.
Title: Net Fluid Balance Effect on Mortality in Septic ICU Patients With Diabetes
Introduction: Fluids have been a crucial tenet of sepsis management(1) but the ideal amount of resuscitation remains unknown. Previous studies have shown that positive fluid balance in septic patients requiring the ICU have been associated with increased mortality(2), though the mechanism remains unclear or if mortality risk is heterogeneous amongst different subgroups. Using machine learning analysis, net positive fluid balance in the critically ill septic patients with diabetes shows a strong signal toward mortality.
Methods: We extracted data from the eICU Collaborative Research Database. Data filtered for an admission diagnosis of Sepsis leading to a total of 17097 patients within the dataset. We further seperated diabetic and non-diabetic patients, with 4362 and 12735 patients respectively. There were 34 features extracted. Mortality was the dependent variable in our prediction models. We split the datasets with a 70/30 distribution. Missing data imputed with Sklearn Iterative Imputer library. We performed binning of the net total fluid balance using Sklearn KBinsDiscretizer with a final bin count of ten. We used a gradient boosting decision tree algorithm variant called CatBoost. Feature Importance was determined with SHapley Additive exPlanations (SHAP) library. To understand the marginal contribution of increasing fluid balances on mortality, we generated a SHAP partial dependence plot.
Results: Partial dependence plots for septic ICU patients with and without diabetes for 20 patient characteristics' contribution to mortality. Net positive fluid balance was shown to be the third most contributing characteristic in diabetic patients but 19th in non diabetic patients. When the patient’s net fluid balance was plotted against mortality and superimposed with APACHE scores, increasing mortality was seen in patients with higher net positive amongst the highest APACHE scores whereas this correlation was not seen in non-diabetics.
Conclusions: Machine learning is a useful tool at identifying patient characteristics’ contribution to mortality. There was a strong signal for net positive fluid balance’s association with mortality amongst the septic patients with diabetes in the ICU that is not seen in the non-diabetic population.Our study is hypothesis generating and further prospective evaluation is warranted.