Representation Learning and Spectral Clustering for the Development and External Validation of Dynamic Sepsis Phenotypes: Observational Cohort Study

Background Recent attempts at clinical phenotyping for sepsis have shown promise in identifying groups of patients with distinct treatment responses. Nonetheless, the replicability and actionability of these phenotypes remain an issue because the patient trajectory is a function of both the patient’s physiological state and the interventions they receive. Objective We aimed to develop a novel approach for deriving clinical phenotypes using unsupervised learning and transition modeling. Methods Forty commonly used clinical variables from the electronic health record were used as inputs to a feed-forward neural network trained to predict the onset of sepsis. Using spectral clustering on the representations from this network, we derived and validated consistent phenotypes across a diverse cohort of patients with sepsis. We modeled phenotype dynamics as a Markov decision process with transitions as a function of the patient’s current state and the interventions they received. Results Four consistent and distinct phenotypes were derived from over 11,500 adult patients who were admitted from the University of California, San Diego emergency department (ED) with sepsis between January 1, 2016, and January 31, 2020. Over 2000 adult patients admitted from the University of California, Irvine ED with sepsis between November 4, 2017, and August 4, 2022, were involved in the external validation. We demonstrate that sepsis phenotypes are not static and evolve in response to physiological factors and based on interventions. We show that roughly 45% of patients change phenotype membership within the first 6 hours of ED arrival. We observed consistent trends in patient dynamics as a function of interventions including early administration of antibiotics. Conclusions We derived and describe 4 sepsis phenotypes present within 6 hours of triage in the ED. We observe that the administration of a 30 mL/kg fluid bolus may be associated with worse outcomes in certain phenotypes, whereas prompt antimicrobial therapy is associated with improved outcomes.


Table of Contents
. STROBE Statement-checklist of items that should be included in reports of observational studies 9

Statistical Methods:
In our modeling of phenotype dynamics as a markov decision process, only interventions within the first 3 hours from ED triage were considered so that the effect of early interventions on a patient's trajectory could be examined. To ensure a clear separation between the variables related to transition and those related to the severity of illness, we excluded age, CCI, and SOFA from our transition models. However, we included the initial phenotype of the patient, which captures a set of features that includes age, comorbidities, and SOFA elements (among others), as a confounder to adjust for patient state. To evaluate the dynamics of our derived phenotypes against traditional patient categorizations, we also repeated this analysis using SOFA groups instead of phenotypes. Significant predictors of transition were determined from Wald tests on the odds ratios. Adjustment for multiple comparisons via the Bonferroni method was done by setting a conservative significance threshold of p < 0.01.   The presence of cluster membership overlap between consecutive runs of k-means demonstrates that this approach results in less consistent clustering than spectral clustering on representations (Figures S1-S2). Figure S4. Cumulative Density Function for k-means clustering Figure S4: Cumulative density function for consensus k-means clustering on normalized EHR data.
The absence of a flattened curve demonstrates inconsistent clustering results.

Figure S5. Phenotype Chord Diagram
The chord diagram shows the clinical profile of patients in the 4 phenotypes we identified. Size of each individual ribbon represents the portion of abnormal features within that body system. For example, phenotypes 3 and 4 are more likely to have abnormal hepatic and inflammatory biomarkers compared to phenotypes 1 and 2. State specific objectives, including any prespecified hypotheses 5

Study design 4
Present key elements of study design early in the paper 5-6 Setting 5 Describe the setting, locations, and relevant dates, including periods of recruitment, exposure, follow-up, and data collection 5-6 Participants 6 (a) Cohort study-Give the eligibility criteria, and the sources and methods of selection of participants. Describe methods of follow-up Case-control study-Give the eligibility criteria, and the sources and methods of case ascertainment and control selection. Give the rationale for the choice of cases and controls Cross-sectional study-Give the eligibility criteria, and the sources and methods of selection of participants Give the source of funding and the role of the funders for the present study and, if applicable, for the original study on which the present article is based 1 *Give information separately for cases and controls in case-control studies and, if applicable, for exposed and unexposed groups in cohort and cross-sectional studies.    Here we explore the degree to which missingness alone accounts for cluster membership. First, we train four logistic regression models to assign cluster membership at ED Triage + 3 hours based only on whether each measurement is missing or not. We then compare this against four separate models trained on the true clinical values. We report the AUCs of each of these models in the table above. As shown, severity of illness and the presence of data are indeed entangled. This is particularly apparent for M2 (the cluster 2 assignment model) which represents the healthiest phenotype. However, using information from the clinical variables significantly improves the performance of the assignment. Therefore, missingness does not fully explain cluster membership.   At a significance threshold of p < 0.01, early administration of fluids is the only intervention significantly associated with transition. Since the administration of vasopressors is one of the SOFA score criteria, this may be a result of the observed inverse correlation between the use of fluids and the administration of pressors. At a significance threshold of p < 0.01, no intervention is significantly associated with transition.