Sepsis is a life-threatening organ dysfunction syndrome secondary to a dysregulated host response to infection, and the primary cause of death from infection, especially if not recognized and treated promptly [1]. A hallmark of sepsis is the heterogeneity of its presentation and its prognosis, due to the variability in pathogen and immune host response interactions.

In 2016, a consensus conference provided an updated definition of sepsis, with septic shock representing a subset of sepsis in which particularly profound circulatory, cellular, and metabolic abnormalities lead to substantially increased mortality [1].

The consensus definition emphasized the importance of timely recognition and prompt management of sepsis [2]. Available therapies and management for patients with sepsis remain limited to source control, administration of antibiotics, and supportive care [3]. Accumulated evidence suggests that the intrinsic heterogeneity of sepsis and variable stage at presentation posed challenges not only to clinical care but also to the conduct of clinical trials assessing interventions for sepsis. Therefore, identifying its sub-phenotypes is crucial for informing prognostic assessment and developing and evaluating effective treatment plans.

A prior study identified sepsis phenotypes at the time of patient presentation to the emergency department, using only routinely available Electronic Health Record (EHR) data in the clustering models [4]. The phenotypes were derived from a large observational cohort to ensure generalizability. This important study, however, did not account for the temporal registration and the rapidly evolving changes in patient physiological and laboratory values. Information acquired in the early course of sepsis can substantially enrich the clinical phenotypes, enable the identification of sub-phenotypes, and increase prognostic accuracy. Other studies have captured the dynamic nature of the clinical course in patients with sepsis using the change in the Sequential Organ Failure Assessment (SOFA) score that assesses the severity of organ dysfunction in ICU patients [5]. However, these scores have been used primarily as outcome measures to evaluate the overall course of organ dysfunction and to predict mortality.

To further advance the classification of sepsis, and identify potential subgroups, we incorporated medical context and temporal biomarker characteristics into the sepsis classification algorithms, early after sepsis onset.

Researchers have been studying disease phenotyping with the help of machine learning techniques and Electronic Health Records (EHRs) [6], [7], [8], [9], [10], which contain large amounts of patient-level information, including demographics, vital signals, lab tests, medications, and diagnosis. However, in recent review papers, Yang et al. and He et al. [11], [12] pointed out that most existing literature used purely data-driven approaches and seldom considered real-world medical use cases and corresponding medical interpretations. Limited work considers temporal information in the EHR longitudinal data. In addition, few existing studies perform non-overlapping clustering, i.e., each patient is commonly assigned to only one group (sub-phenotype).

Sepsis may initially be associated with dysfunction of one organ system and progress to involve multiple organ systems. Because of the involvement of multiple systems, a patient may exhibit more than one sub-phenotype. We thus develop a soft clustering method that allows each patient to be assigned to more than one sub-phenotype. At the same time, we take biomarker temporal information into account and incorporate clinical information into the soft clustering algorithm. By applying transformations to the soft clustering results, we obtain six novel sepsis hybrid sub-phenotypes. We evaluate the plausibility of the results by providing a biological explanation. Additionally, built upon the soft clustering results, we train and validate a sepsis early-warning model to predict the novel sepsis hybrid sub-phenotypes. The results suggest the newly identified hybrid sub-phenotypes provide characterizations of different sepsis progressions.