Identifying and analyzing sepsis states: A retrospective study on patients with sepsis in ICUs

Abstract

Sepsis accounts for more than 50% of hospital deaths, and the associated cost ranks the highest among hospital admissions in the US. Improved understanding of disease states, progression, severity, and clinical markers has the potential to significantly improve patient outcomes and reduce cost. We develop a computational framework that identifies disease states in sepsis and models disease progression using clinical variables and samples in the MIMIC-III database. We identify six distinct patient states in sepsis, each associated with different manifestations of organ dysfunction. We find that patients in different sepsis states are statistically significantly composed of distinct populations with disparate demographic and comorbidity profiles. Our progression model accurately characterizes the severity level of each pathological trajectory and identifies significant changes in clinical variables and treatment actions during sepsis state transitions. Collectively, our framework provides a holistic view of sepsis, and our findings provide the basis for future development of clinical trials, prevention, and therapeutic strategies for sepsis.

Fig 1. Illustration of proposed framework: The data preparation phase extracts 42 variables (demographic profiles, vital signs, laboratory tests, mechanical ventilation status of the patients, and the comorbidity profiles) from 16,546 distinct sepsis patients admitted to Beth Israel Deaconess Medical Center from the MIMIC-III database.

In the data analysis phase, we use archetypal analysis to find distinct states in sepsis. We validate that each state corresponds to patient clusters that are statistically distinct from the distribution of the cohort as a whole, and the SOFA score, SIRS score, and mortality rate are calculated to characterize each sepsis state. In primary function analysis, selected features from archetypes are used to identify the primary functions (namely, nervous system, inflammation and infection, liver function, kidney function, coagulation, respiratory function, and, cardiovascular function) of each sepsis state. In etiological analysis, we find correlation between pre-existing comorbidity profiles (30 types) and sepsis states. Finally, in progression analysis, we use higher-order Markov chains to model the dynamics of pathological processes of sepsis. We then use archetypal analysis to identify distinct types of sepsis state transitions and use z-score analysis to find representative clinical markers of each state transition.

Fig 2. Visualization (using low-dimensional UMAP embedding) of the six derived sepsis states.

Colors represent different sepsis states. The average SOFA score, average SIRS score, and mortality rate are used to characterize sepsis states. Based on these scores, we characterize states A1 (blue), A2 (orange), and A3 (green) as ‘moderate condition’, ‘inflammation’, and ‘mild condition’, respectively, and we characterize states A4 (red), A5 (brown), and A6 (black) as ‘Multiple Organ Dysfunction Syndrome (MODS)’.

Fig 3. Visualization of the selected features (21 features in total) by Q_j(P_K), Qj′(PK), and Variation test.

Q_j(P_K) calculates the discriminative power of feature i for a given clustering as the ratio of inter-cluster inertia to the total inertia computed using feature i. $Q_j^{\prime }\left(P_K\right)$ calculates the discriminative power of feature i as the ratio of inter-cluster inertia computed using feature i to total inter-cluster inertia computed using all features. Variation test selects features that have the lowest probability of overlap across clusters (please see Methods section Feature selection methods [56] for more details). Note that there is a significant overlap between features chosen by these selection criteria. However, each criterion yields a distinct set of features significantly associated with different sepsis states.

Fig 4. Spider-plot of primary functions affected in each sepsis state (represented by corresponding colors).

There are seven different dimensions of primary function for each sepsis sate. The measured dimensions are nervous system, inflammation and infection, liver function, kidney function, coagulation, respiratory function, and cardiovascular function, respectively. The scale of each dimension ranges from 0 to 10, with higher values indicating higher affect on the primary function.

Fig 5

(A) The population distribution for each sepsis type stratified by age. (B) The population distribution for each sepsis type stratified by weight. (C) The population distribution for each sepsis type stratified by the number of comorbidities before infection. (D) Z-score analysis of the comorbidity profiles (row) of each sepsis type (column). Entries approaching red in intensity indicate that the comorbidity profiles are expressed in the corresponding sepsis states, and entries closer to blue indicate that the comorbidity profiles are suppressed in corresponding sepsis states. Entries with p-value >0.05. are set to 0.

Fig 6

(A) Third-order transition graph: Edges approaching red in color indicate higher transition probabilities, and edges approaching black indicate lower transition probabilities. (B) First-order transition graph. Averaged fluids, maximum dosage of vasopressor, the probability of using mechanical ventilator, and the mortality rate are shown on the transitions.

Fig 7

(A) Visualization (using low-dimensional UMAP embedding) of the six derived gradient groups. (B) Z-score analysis of the gradients of clinical measurements (row) of each gradient group (column). Entries approaching red indicate that the gradient of the clinical readings are expressed in the corresponding gradient groups, and entries closer to blue indicate that the gradient of the clinical readings are suppressed in corresponding gradient groups. (C) Probability histogram of state transitions of each gradient group.