(400) Enhancing the APACHE II Score: Addition of Pharmacophenotypes to Predict Fluid Overload

Introduction: Fluid overload (FO) is challenging to predict in the intensive care unit (ICU) due to disease and treatment complexity and heterogeneity. Intravenous (IV) medications (including administrated volume) are a primary cause of FO but are challenging to evaluate as a predictor given the high frequency and time-dependency of their use. Here, we employed unsupervised machine learning methods to explore IV medication administration patterns (or pharmacophenotypes) to improve FO prediction in the ICU.

Methods: This retrospective cohort study included adults admitted to an ICU ≥72 hours with available fluid balance data. FO was defined as a positive fluid balance ≥7% of admission body weight within 72 hours of ICU admission. Unsupervised machine learning, including principal component analysis (PCA) and Restricted Boltzmann Machine (RBM), were used to evaluate medication administration records of patients with and without FO to identify pharmacophenotypes associated with FO. This pharmacophenotype was then added to a regression model that included the Acute Physiology and Chronic Health Evaluation (APACHE II) and receipt of diuretics to determine the effect on model predictiveness for identification of fluid overload.

Results: In a cohort of 927 patients that included 47,803 IV medication administrations, ten unique pharmacophenotypes were identified. Among the ten pharmacophenotypes, the mean number of Pharmacophenotype 7 medications administered was significantly greater in the FO (vs. no FO) cohort (25.6 vs.10.9 p< 0.0001). When Pharmacophenotype 7 was included in a FO prediction model including the APACHE II score and receipt of diuretics, model performance improved from an AUROC of 0.719 to an AUROC of 0.741 (p=0.027).

Conclusions: Using ML approaches, a unique pharmacophenotype was strongly associated with FO. Incorporation of this medication data improved the ability to predict FO occurrence in ICU patients compared to traditional prediction models. Application of this approach into real-time clinical applications may help improve early detection of FO to facilitate timely intervention.