(1434) Machine Learning-Driven Temporal Subphenotypes of Early Sepsis

Introduction: Sepsis is a life threating illness and is associated with high morbidity and mortality. Recent studies have delineated distinct subphenotypes of sepsis but these do not account for temporal evolution of disease. We aimed to identify temporal subphenotypes of sepsis within 24h of intensive care unit (ICU) admission while accounting for progressive evolution of disease process.

Methods: Using MIMIC-IV database we identified adult patients with sepsis admitted to ICU. We included only first ICU admission for each patient and excluded those who died within first 24h of ICU admission. We used Long Short Term Memory (LSTM) Autoencoder to represent timeseries nature of dynamic features followed by KMeans clustering on available dynamic features (hourly vital signs, SOFA scores, vasopressor use, mechanical ventilator use and labs) to identify temporal subphenotypes of these patients over first 24 hours. We used ANOVA and chi squared tests for comparisons within subphenotypes (SP).

Results: Among 18,728 patients we identified 4 distinct temporal subphenotypes (SP 1 with n=7190, SP2 with n=737, SP3 with n=1123 and SP4 with n=9678). SP 1 had median age 67yr with admission SOFA score 2 & no vasopressor support on admission. Their max 24h SOFA score was 4 & total 24h vasopressor dose was 2.8mcg/kg norepinephrine equivalent dose (NEeD). In comparison, SP 2 had median age 65y, admission SOFA score 4 and admission vasopressor dose 0.74mcg/kg NEeD. Their max 24h SOFA score was 11 and total 24h vasopressor dose was 114.1mcg/kg NEeD. SP 3 and 4 in comparison had median age 64 & 69y, admission SOFA score 2 each but max 24h SOFA score was higher for subphenotype 3 (8 vs 5) as was total vasopressor dose (24.1 vs 7.3 mcg/kg NEeD). SP 2 had highest mortality (40%), followed by SP 3(28%), SP 4(12%) and SP 1(8%). All comparisons were significant at p< 0.001.

Conclusions: We identified 4 distinct temporal subphenotypes of patients with early sepsis with differences in clinical characteristics, disease evolution and outcomes. These clusters provide valuable insights into heterogeneity in progression of early sepsis and provide a foundation of personalization of treatments.