Clinical Phenotyping of Adults With Suspected Sepsis in Uganda Reinforces Severe Tuberculosis as a Key Driver of Physiologic Severity and Mortality Risk

Abstract Rationale: Global sepsis incidence is highest in sub-Saharan Africa (SSA), a high HIV burden setting where inciting pathogens are unique. Despite this, clinical heterogeneity inherent to sepsis in SSA, including relationships with host-pathogen profiles, is poorly understood. We hypothesized that an unsupervised learning approach, applied to clinical data from patients hospitalized with suspected sepsis in Uganda, would identify patient subgroups associated with distinctive host-pathogen profiles and clinical outcomes. Methods: We analyzed clinical and microbiologic data from two prospective cohorts of adults hospitalized with severe, undifferentiated infection (i.e., suspected sepsis) at public hospitals in Uganda (discovery cohort [Entebbe, urban], N=301; validation cohort [Tororo, rural], N=298). To identify patient subgroups in the discovery cohort, we applied hierarchical clustering on principal components to readily available clinical data, followed by a k-means procedure to consolidate cluster membership. Input variables included body temperature, heart rate, respiratory rate, systolic blood pressure, oxygen saturation, and ordinal assessment of consciousness. In the validation cohort, we assigned patients to subgroups based on their proximity (in Euclidean distance) to the centroid of subgroups identified in the discovery cohort. Results: In the discovery cohort, two clinical subgroups were identified (subgroup 1, N=198, subgroup 2, N=103). Compared with subgroup 1, subgroup 2 was associated with more deranged vital signs, higher physiological severity scores, and significantly higher risk of 30-day mortality (19% vs 31%, p=0.004). Subgroup 2 was associated with a significantly higher prevalence of HIV co-infection (43% vs 69%, p<0.001) and microbiologically-diagnosed tuberculosis (TB) (11% vs 29%, p<0.001) (Figure 1). Patient assignment to cluster centroids in the validation cohort reproduced subgroups with similar distributions (subgroup 1, N=208, subgroup 2, N=90), profiles of physiologic severity, and mortality risks (15% vs 42%, p=<0.001). While subgroup 2 remained significantly associated with a high prevalence of severe TB in the validation cohort (17% vs 28%, p=0.030), differences in HIV co-infection were not (49% vs 50%, p=0.8). Although not statistically significant, prevalence of malaria was consistently higher in subgroup 1 (23% vs. 15%, p=0.088 [discovery], 33% vs. 17%, p=0.004 [validation]). Discussion: Using unsupervised learning methods, we consistently identified two prognostically-enriched subgroups among adults hospitalized with suspected sepsis in Uganda. High-risk subgroups were consistently associated with higher prevalence of severe TB. Continued efforts are needed to determine the optimal approach to diagnosis and treatment of TB sepsis in high-burden settings, including host and pathogen directed therapeutics.