(1411) Machine Learning Analysis of Sociodemographic Factors on Continuous Renal Replacement Therapy Dosing

Introduction: Acute Kidney Injury (AKI) is subject to social determinants of health (SDOH) that may also influence Continuous Renal Replacement Therapy (CRRT) dosing and outcomes in the intensive care unit (ICU). However, the number and strength of these SDOH factors remains unknown. This study sought to identify the number and strength of SDOH factors associated with CRRT dosing and ICU outcomes using machine learning methods.

Methods: This single-center retrospective study included ICU adults (aged 18+) admitted to Beth Israel Deaconess Medical Center (BIDMC) between January 2008 to December 2019. SDOH features included age, race, insurance type, and marital status. Clinical features included Sequential Organ Failure Assessment (SOFA) score, Simplified Acute Physiology Score (SAPS), age, mechanical ventilation status, and ICU length of stay. CRRT patients that received guideline-based dosing ( < 25 ml/kg/hr) vs. non-guideline-based dosing (>25 ml/kg/hr) were analyzed using unsupervised machine learning methods. Uniform Manifold Approximation and Projection (UMAP) was used for phenotype analysis and derivation of dosing clusters. A parsimonious social determinants-based ICU mortality prediction machine learning model was developed using logistic regression, random forest, eXtreme Gradient Boosting (XGBoost), and CatBoost. Feature importance analysis was conducted using the SHapley Additive exPlanations (SHAP) plot.

Results: There were 1156 patients included for phenotype dosing analysis, and 2454 for mortality prediction. Average age and SOFA score were 63 years and 4.6, respectively. The UMAP analysis resulted in multiple clusters where cluster 1’s significant features included age (p-0.01), SOFA (p-0.01), and mechanical ventilation status (p-0.05); cluster 2’s were race (p -0.04) and cluster 3’s was ICU length of stay (p-0.01). For ICU mortality, the XG Boost model had the highest accuracy (0.84). The important features identified by SHAP plot for mortality prediction included race, age, marital status, SAPS, and ICU length of stay.

Conclusions: Combined with clinical characteristics, various SDOH factors, such as age and race, may influence CRRT dosing and mortality. Further studies are needed to determine the extent of SDOH influence in larger, more diverse cohorts.