249: PRINCIPAL COMPONENT ANALYSIS IDENTIFIES DISCRETE PATTERNS OF HOSPITAL CAPABILITIES RELATED TO SEPSIS

Abstract

Introduction: Sepsis patients treated at hospitals with higher sepsis case volumes have better outcomes. Case volume is a proxy for hospital capability. We explored patterns of hospital characteristics that may better reflect hospital capabilities among sepsis patients. Methods: We used the New York State Inpatient Database to analyze 90,051 adult sepsis encounters in 2018 at 157 non-federal hospitals. We used patient-level data to construct 14 a priori identified, hospital-level resource use characteristics and classified them into distinct capability constructs: clinical volume (4 measures: bed size, annual volumes of sepsis, emergency department [ED] visits, ICU admissions); diagnostic (3 measures: annual volumes of magnetic resonance imaging [MRI], CT scans, ultrasound); therapeutic capacity (7 measures: renal replacement therapy, mechanical ventilation [MV], ECMO, central venous catheterization, percutaneous nephrostomy [PN] and cholecystostomy, and major surgical procedures). Using Principal Component Analysis with varimax rotation, we analyzed the dimensionality of, and reduced the 14 characteristics to 3 interpretable, linear data combinations (principal components (PC). We calculated PC scores for each hospital as a sum of standardized values for each component multiplied by the respective PC loading. Results: Among 157 hospitals, 50% were teaching, 80% were urban, with median (IQR) bed size: 195 (97-381), sepsis volume: 416 (144-737), ED volume: 29860 (16251-50102), ICU volume: 775 (292-1665), MRI: 1 (0-16), CT scan: 2 (0-72), and diagnostic ultrasound: 122 (20-697). The 3 selected PCs together explained 81% of the data variation (PC1, 63%; PC2, 12%, PC3; 6%). PC1 was highly correlated on clinical volume (range of loadings: 0.66, ED volume – 0.92, sepsis volume) and therapeutic capacity (0.72, PN – 0.94, MV). PC2 was correlated on diagnostic constructs (0.63, ultrasound – 0.94, MRI), while PC3 correlated specifically with ECMO capability (0.94) Conclusions: Discrete, interpretable patterns of hospital characteristics correlate with pre-specified hospital capability constructs. Further studies should explore correlation of PCs with sepsis mortality, and performance of PCs compared to sepsis volume.