Fuzzy classification of sepsis subtypes and implications for trajectory and treatment

Summary

Background

Sepsis is common and deadly, and subtypes are proposed to guide precision treatment. However, little is known about the uncertainty in subtype classification, and its implications for trajectory and treatment response.

Methods

In multiple electronic health record and trial data of adults with sepsis, we assigned patients clinical sepsis subtypes (α, β, γ, or δ-type), and measured uncertainty by defining core (≥90%) and margin (<90%) strata for each subtype according to model-derived membership probabilities. In multivariable logistic regression models, we determined the association between subtype, core/margin strata, and two outcomes, i.) change in subtype over 48 h and ii.) 365-day mortality in the ProCESS randomised trial.

Findings

We included 35,691 adult patients (mean age 68 [SD 16] years; 51% male, 85% White, 5.7% in-hospital mortality) with community-acquired sepsis according to Sepsis-3. Most patients changed clinical sepsis subtype during the 48 h after presentation (82%) regardless of initial subtype. The majority of patients were in the margin stratum of the subtype (α-type: 70%, β-type: 66%, γ-type: 64%), except for those in δ-type (18% margin strata). The odds of subtype change over 48 h was increased in the margin strata (interaction p = 0.023), where, for example, patients with the margin delta subtype had significantly higher odds than patients with alpha core (ref) subtype (odds ratio, 7.13; 95% confidence interval [CI], 5.16–9.85). For risk-adjusted 365-day mortality in the ProCESS trial, the effect of randomised treatment was modified by the subtype margin strata (interaction p = 0.026).

Interpretation

In patients with community sepsis, clinical subtypes are dynamic. Patients on the subtype margin are more likely to change groups, and uncertainty of subtype classification modified treatment effects.

Funding

National Institutes of Health, National Institute of General Medical Sciences (R35GM119519).

Research in context.

Evidence before this study

Clinical sepsis subtyping is proposed to facilitate personalisation of clinical care. The trajectory of assigned clinical subtypes following presentation is poorly understood. We searched PubMed for relevant original research articles assessing clinical sepsis subtypes and their trajectory over time. Using combinations of the terms “sepsis,” “subtype,” “phenotype,” “subphenotype”, “uncertainty”, “trajectory,” we filtered search results by the article type Cohort study and adult age and limited results to English-language articles published from database inception to August 1, 2024. We identified several studies of how ARDS and sepsis subtypes behave after presentation. Prior work generally concludes that subtypes may not be stable over daily assessment, and individual trajectory impacts outcomes (Slim et al., 2024 & Rambout et al., 2024).^1,2 Little is known about how uncertainty of initial classification impacts overall trajectory, how classification of patients changes by hour, and how these impacts treatment effects in trials.

Added value of this study

This work attempts to elucidate understanding of clinical sepsis subtype trajectory and determine how the uncertainty in initial classification effects whether or not subtypes change, and if this uncertainty modified the extent of treatment effect heterogeneity. Analysing subtype membership at 6-h intervals, we noted that most patients changed clinical subtype by 48 h after presentation (82%). Uncertainty of initial classification was common (63% were on the subtype margin, <90% probability of the assigned subtype). Patients on the margins were more likely to change subtype and a significant interaction between initial subtype, margin strata and odds of changing subtype in 48 h was noted within regression models (p-value <0.05). In a secondary analysis of the ProCESS randomised trial of early, goal directed therapy in septic shock patients,³ the observed effect of the intervention was heterogenous amongst core and marginal strata within subtypes.

Implications of all the available evidence

Our findings indicate that initial sepsis subtypes are fuzzy and change frequently within the initial 48 h from clinical presentation. Patients within the margin stratum were more dynamic than core patients and underlying uncertainty influenced treatment effects. Taken together, these results highlight challenges of leveraging clinical subtypes to develop precision therapies. Rigorous attention to subtype development is required to minimise uncertainty and incorporate their use in clinical trials.

Introduction

Sepsis is a complex and heterogenous syndrome, now proposed to have multiple prognostic and predictive subtypes.⁴ Whether derived using clinical data, biomarkers, or molecular methods, sepsis subtypes are typically discrete and bounded groups—with patients sharply classified as one subtype or not.⁵ This approach lends itself to greater applicability at the bedside and the delivery of subtype-specific treatments.

However, sepsis is not discrete, but rather fuzzy and difficult to diagnose.⁶ This is the case both when clinicians determine who is septic (or not) as well as who may be a member of what sepsis subtype (or not). Further uncertainty is encountered when subtypes are studied over time, influenced by treatment or natural disease progression.¹ The classification of patients with sepsis into subtypes may be a key step towards precision treatment, but the clinical implications of variability within and between groups is unknown.

To address these challenges, we sought to determine the epidemiology and outcomes of clinical sepsis subtypes relative to the certainty with which patients were grouped. We used observational data from an integrated health system to study the trajectory of patients’ subtype relative to group membership, and ProCESS randomised trial data to determine the implications of fuzzy assignment of sepsis subtype on treatment effects.

Methods

The project was approved by the University of Pittsburgh Human Research Protection Office (STUDY 19030218) and conducted under data use agreement (PRO12020657). The cohort data were obtained under a waiver of informed consent and with authorisation under the Health Insurance Portability and Accountability Act. Data were abstracted and reported in accordance with the STROBE statement.⁷ Written informed consent was obtained for clinical trial data per published trial procedures (ProCESS ClinicalTrials.gov number, NCT00510835).

Overview

The study approach involved two datasets and multiple analyses. For the first goal (subtype trajectory) we identified clinical sepsis subtypes using electronic health record (EHR) data available at presentation in a large database of hospital encounters using previously reported methods.⁴ We then assessed subtype trajectory and membership over the ensuing 48 h of hospitalisation in 6-h time epochs. For the second goal (fuzzy classification), we used the probability of subtype assignment to understand which patients were at the core or the margin (strata) of the subtype. Then, in randomised trial data, we evaluated whether subtype trajectories are independent of treatment. For the third goal (impact of fuzzy classification on treatment effects), we examined ProCESS trial data to determine how treatment with fluid resuscitation (early, goal directed therapy vs. usual care) was independently associated with subtypes in the core vs. margin strata.

Data sources and definitions

We used data from inpatient EHR (CERNER Co., Kansas City, MO) at 12 community and academic hospitals in the UPMC healthcare system. This included demographics such as age, patient-reported sex and race, comorbid conditions, procedural, organ support, and clinical outcomes. Race was abstracted from the registration system data and categorised in accordance with the Centres for Medicare and Medicaid Services EHR meaningful use data set. Comorbid conditions were classified by hospitalisation diagnosis International Classification of Diseases, Clinical Modification, of the 9th and 10th revision codes. We abstracted vital signs, laboratory results, evidence of organ dysfunction as measured by the sequential organ failure assessment (SOFA) score within 6 h of sepsis onset,⁸ and time from sepsis onset to antibiotic administration. Date of death in the healthcare system was captured using the monthly updated Social Security Death Index linked outpatient records. Randomised clinical trial data in ProCESS compared early goal-directed therapy (EGDT), a multicomponent resuscitation strategy, to alternative resuscitation approaches in patients with septic shock at 31 sites in the United States from 2008 to 2013 and reported no benefit for mortality.⁶

Patients

In the observational data, we included hospitalised adults (age ≥18 years) with community-acquired sepsis. We defined community-acquired sepsis as evidence of International Consensus Criteria for Sepsis-3 in the EHR,⁹ including i) evidence of suspected infection within 24 h of hospital presentation and ii) presence of organ dysfunction within 6 h of suspected infection. A suspected infection was defined by the administration of antibiotics as well as acquisition of a culture specimen. Acute organ dysfunction was defined as ≥2 SOFA points. Patients were included in the observational data from 1 January 2013 to 31 December 2014 with follow up through 31 December 2019 (the “Validation Cohort” from previously reported work).⁴ In the ProCESS trial data, all patients met a variation of the Sepsis-2 criteria.¹⁰ To understand the trajectory in sepsis subtypes, we restricted analyses to subjects who survived and remained hospitalised for 48 h in the observational data, or who were hospitalised at randomisation and day 3 in the ProCESS trial (the “ProCESS Cohort” from previously reported work).⁴

The classification of patients into clinical sepsis subtypes followed previously published methods (Table S1).⁴ We used 29 variables, including demographic variables (e.g., age, sex, Elixhauser comorbidities), vital signs (e.g., heart rate, respiratory rate, Glasgow coma scale score, systolic blood pressure, temperature, and oxygen saturation (SaO₂)), markers of inflammation (e.g., white blood cell count, premature neutrophil count (‘bands’), erythrocyte sedimentation rate, and c-reactive protein), markers of organ dysfunction or injury (e.g., alanine aminotransferase (ALT), aspartate aminotransferase (AST), total bilirubin, blood urea nitrogen (BUN), creatinine, international normalised ratio (INR), partial pressure of oxygen (PaO₂), platelets, and troponin) and serum glucose, sodium, haemoglobin, chloride, bicarbonate, lactate, and albumin. Patients were classified every 6 h for the 48 h of observational data and at admission and day 3 for the ProCESS trial. We determined the subtype membership of each patient in each 6-h time epoch using the standardised, normalised Euclidean distance from the measured variable to the published SENECA derivation subtype centroid. The minimal distance then identified the patient as α, β, γ, or δ sepsis subtype.

Outcomes

The primary clinical outcome in observational data was a change in subtype compared to the initial time epoch (yes/no) at any time epoch during the 48 h. For the ProCESS trial, a change in subtype was defined by re-assessment using case report form data from 0 h to 72 h after enrolment. For treatment effects in trial data, we used the primary outcome of survival at 365 days, and secondary outcomes were 60-day inpatient mortality, intensive care unit length of stay (ICU LOS), total days of mechanical ventilation, and total days of vasopressor support.

Statistical analyses

To determine the epidemiology of sepsis subtypes, we first assessed candidate variable distributions, and missingness. Multiple imputation using chained random forests with predictive mean matching was used to account for missing data.¹¹ In observational data, imputation was conducted within each of the eight 6-h time intervals during the 48-h observation period (n = 1000 regression trees and 20% sample fraction for each). In the ProCESS trial, imputation was performed within the baseline and the day 3 data (n = 1000 regression trees and 100% sample fraction for each). Log transformation was used for nonnormal data (Table S2).

To determine the fuzzy classification of sepsis subtypes, we used percentile of membership derived from the Euclidean distance used to assign subtype membership. We determined membership percentile by calculating the distance to the assigned subtype divided by the sum of distances to all subtype centres. This value was then ranked for all patients in the cohort and a percentile of membership was assigned. We defined two strata per subtype, “marginal” and “core”. A core strata member was a patient in whom the percentile of membership was ≥90% in the first time epoch, and a marginal strata member was all others within that subtype. We report descriptive patient characteristics before and after imputation, across subtypes, and by core vs. marginal strata within subtype. We explored alternative cut points for defining core and marginal strata, alternatively defining core strata as patients in whom the percentile of membership was ≥85% and ≥95% in the first time epoch, and a marginal strata member was all others within that subtype and visualised patterns of transition with bar graphs.

Core and marginal membership were visualised with t-distributed stochastic neighbour embedding plots (which show multidimensional data in 2 dimensions). The trajectory of subtype membership was also visualised using heatmaps across time epochs, stratified by core vs. marginal members. Finally, multivariable logistic regression was used to determine the association between core vs. margin strata and clinical subtype change (ever during the 48 h). This model is reported after testing for an a priori interaction between core/marginal strata and initial subtype, with a significance threshold of <0.1.

To determine the impact of fuzzy classification in the ProCESS trial, we used the same methods for data standardisation, handling of missing data, subtype assignment, and descriptive comparisons as in the observational data. First, we re-analysed a multivariable logistic regression model to evaluate the association of core vs. marginal membership with subtype change, independent of randomised treatment arm. This model seeks to limit variability in clinician treatment which may confound the pathway between subtype strata and its association with subtype change.

Second, we illustrate 365-day survival using Kaplan Meier curves for subtype, randomised treatment, and core vs. marginal strata. We used multivariable logistic regression to unpack whether previously reported heterogeneity of treatment effect for early-goal directed therapy is consistent for patients in both core and margin. In this model, we adjusted for potential patient characteristics such as patient age, race, sex, SOFA score at randomisation, and Charlson comorbidity index to reduce variance.

Data are presented as mean (SD) or median [IQR], as appropriate based on histograms and QQ plots. For comparisons, we used ANOVA and Kruskal Wallis tests, for continuous data, and chi square tests for categorical data. There was no adjustment for type I error due to multiple comparisons, and therefore the findings from these analyses should be considered exploratory. Analyses were performed with Stata 18.0 (StataCorp, College Station, Texas), and R 4.4.1 (R Core Team, Vienna, Austria).

Role of the funding source

The funding sources had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication. This material is based upon work supported by the National Institutes of Health. The contents of this paper are solely the responsibility of the authors and do not necessarily represent those of the National Institutes of Health, or any enrolling clinical site. The senior and first author had full access to all the data in the study and had final responsibility for the decision to submit for publication.

Results

Patients

Among 1,137,588 adult admissions to 12 hospitals in a large integrated health system, we included 35,691 adult admissions (mean age 68 [SD 16] years; 51% male, 85% White) with community-acquired sepsis according to Sepsis-3 (Table 1). In the ProCESS trial, we studied 821 subjects with septic shock (mean age 60 (SD 16) years, 56% male, 68% White). Data were similar before and after multiple imputation (see Table S3). The four sepsis subtypes (α, β, γ, or δ) were assigned in both datasets (see Tables S4 and S5), with the α-type most common (n = 12,045, 34%), followed by β-type (n = 10,224, 29%), γ-type (n = 9303, 26%), and δ-type (n = 4119, 12%). When comparing demographics, coexisting conditions, and sepsis characteristics, the δ-type was consistently more ill, with greatest serum lactate (median 3.0 [IQR, 1.7–5.4] mmol/L), ICU admission (64%), and inpatient mortality (14%). The frequency distribution was skewed towards patients in the δ-type in the ProCESS trial (25%), where 60-day inpatient mortality was high (24%).

Table 1.

Patient characteristics.

Variablea	Observational data n = 35,691	Randomised trial data n = 821
Age (years), mean (SD)	68 (16)	60 (16)
Male sex, no. (%)	18,368 (51%)	459 (56%)
Race, no. (%)b
White	30,410 (85%)	558 (68%)
Black	4192 (12%)	200 (24%)
Other	1089 (3%)	63 (8%)
Elixhauser/Charlson Comorbidity Index, mean (SD)c	1.3 (1.2)	2.6 (2.5)
SOFA score, maximum, mean (SD)d	3.7 (2.0)	6.8 (3.4)
Vital signs
Heart rate, beats per minute, mean (SD)	96 (21)	103 (21)
Respiratory rate, breaths per min, mean (SD)	23 (7)	23 (7)
Systolic blood pressure, mmHg, mean (SD)	111 (25)	100 (24)
Serum lactate, mmol/L, median [IQR]	1.8 [1.2–2.8]	4.3 [2.5–5.7]
Total bilirubin, mg/dL, median [IQR}	0.7 [0.4–1.2]	0.9 [0.6–1.5]
Serum creatinine, mg/dL, median [IQR]	1.3 [0.9–2.1]	1.6 [1.1–2.6]
Hospital organ support and outcomes
Mechanical ventilation, no. (%)e	6094 (17%)	272 (33%)
Vasopressors, no. (%)e,f	4355 (12%)	125 (15%)
Intensive care, no. (%)e	13,317 (37%)	731 (89%)
In-hospital mortality, no. (%)	2025 (6%)	104 (13%)

Abbreviations: ICU, intensive care unit; IQR, interquartile range; SD, standard deviation; SIRS, systemic inflammatory response syndrome criteria; SOFA, sequential organ failure assessment score.

^aCorresponds to minimum or maximum value, as appropriate, within 6 h of presentation or prior to randomisation.

^bOther race corresponds to Chinese, Filipino, Hawaiian, American Indian/Alaskan, Asian, Hawaiian/Other Pacific Islander, Middle Eastern, Native American, Not specified, or Pacific Islander.

^cElixhauser (used in observational data) is a method of categorising comorbidities of patients based on the International Classification of Diseases (ICD) diagnosis codes found in administrative data, ranging from 0 to 31; Charlson (used in ProCESS trial) is a method of categorising comorbidities of patients based on the International Classification of Diseases (ICD) diagnosis codes found in administrative data, ranging from 0 to 24.

^dSOFA score corresponds to the severity of organ dysfunction, reflecting six organ systems each with a score range of 0–4 points (cardiovascular, hepatic, haematologic, respiratory, neurological, renal), with a total score range of 0–24 points.

^eAt any time during hospitalisation.

^fAny administration of norepinephrine, epinephrine, vasopressin, phenylephrine, dopamine, dobutamine, or milrinone during hospitalisation.

Sepsis subtypes over time

Most patients with sepsis changed subtype within 48 h of presentation (n = 29,205 of 35,691, 82%, Fig. 1). Subtypes changed most often for patients who started in the γ-type (n = 8564 of 9303, 92%), compared to the α-type (n = 9551 of 12,422, 79%), β-type (n = 7588 of 10,224, 74%), or δ-type (n = 3502 of 4119, 85%). On average, the time to subtype change was shortest in the γ-type (mean (SD) hours to change, 14 (13) hours, see Table S10). Patterns of subtype change were similar between patients presenting with and without septic shock (see Fig. S1).

Fig. 1.

Trajectory of clinical sepsis subtypes. Left column are patients with highest subtype membership probabilities (“core patients”). The right lighter shaded column are patients with lower membership probabilities (“margin patients”). Upper panels A and B show the proportion of subjects in original subtype in 6-h intervals from presentation to 48 h after arrival. Lower panels C and D illustrate a heat map of transition patterns. Green is α-type, purple is β-type, red is γ-type, and blue is δ-type. Core patients have >90% membership probability for their assigned clinical sepsis subtype, and margin patients with <90% membership probability.

Fuzzy classification of sepsis subtypes

The majority of patients were in the margin of the assigned subtypes (overall, 63%; α-type, 70%; β-type, 66%; γ-type, 64%; median probability of membership overall, 0.86; α-type, 0.83; β-type, 0.85; γ-type, 0.85; δ-type, 0.94; see Fig. S2, Table S7). Clinical characteristics of patients in the core and margin were largely similar (see Table S6 and S8). However, patients in the margin were a minority in the δ-type (18%) and were distinctly less ill than patients in the core δ-type. For example, despite similar demographics, patients in the margin δ-type had, on average, increased blood pressure, lower transaminases, higher platelet counts, and reduced serum lactate than patients in the core δ-type (see Table S9). Such differences were also seen in the ProCESS trial data (see Table S10).

Patients on the margins were more likely to change subtype (margin, 85% changed; core, 77% changed subtype, see Table S11). As visualised in t-SNE plots (Fig. 2), the patients who changed subtype commonly overlapped with those in the marginal strata. Patients in the marginal strata were more likely to change subtype than core strata, irrespective of threshold used to define core and margin (see Fig. S3). In regression models, there was a significant interaction between initial subtype, margin strata, and the odds of changing subtype in 48 h (interaction p value < 0.05). For example, compared to the core α-type (referent group), we observed the core β-type was associated with less subtype change (odds ratio (OR), 0.71, 95% CI: 0.64, 0.79) and a greater change in patients in the margin β-type (OR, 1.60, 95% CI: 1.45, 1.76). Patients in the margin were consistently associated with greater subtype change in γ- and δ-types (Fig. 3).

Fig. 2.

t-SNE plots of subtype patients and proportion who change subtype. Panels A–D illustrate t-distributed stochastic neighbour embedding plots of clinical sepsis subtypes where lighter shaded dots are margin strata and darker shaded dots are core strata. Panels E–H overlap black dots for subtype members who transition to another subtype. Panel I shows the cumulative proportion of patients who transition subtype over 48 h. Green is α-type, purple is β-type, red is γ-type, and blue is δ-type. Light grey are not members of that subtype. Core patients have >90% membership probability for their assigned clinical sepsis subtype, and margin patients have <90% membership probability.

Fig. 3.

Odds of clinical sepsis subtype change in 48 h after presentation. Panel A shows odds ratio (95% confidence intervals) for logistic regression model in observational data (n = 36,591), with interaction between initial subtype and core vs. margin strata (interaction p < 0.01). Panel B reports a similar model and interaction in ProCESS trial data (n = 821), adjusting for randomised treatment arm. Dark circles to the left of 1.0 correspond to a reduced adjusted odds of changing subtype, and to the right an increased adjusted odds of changing subtype. Patients in core have >90% membership probability for their assigned clinical sepsis subtype, and patients in margin have <90% membership probability.

Subtype change independent of treatment

When subtypes mapped to the ProCESS trial, 63% of subjects were in the margin of the assigned subtype. In regression models adjusted for randomised treatment arm (EGDT vs. usual care), we observed a significant interaction between subtype, margin strata, and subtype change at 72 h (p = 0.06, see Fig. 3). For example, compared to patients in the core α-type (referent group), we observed an increased odds of subtype change for patients in the margin δ-type (margin δ-type aOR = 3.25, 95% CI: 1.78, 5.95). Treatment arm was not independently associated with subtype change (p > 0.05).

Heterogeneity of treatment effects for core and marginal subtype members

The observed effect of early, goal directed therapy vs. usual care was differential across core and marginal strata in subtypes (Fig. 4). For example, the α-type had no EGDT treatment benefit overall, in core, or margin strata for survival at 365-days (see Figs. S4, S5 and S6). However, in patients in the δ-type, the signal for harm from the EGDT treatment effect on 365-day mortality was observed in patients in the code but not in the margin (Fig. 4).

Fig. 4.

365-day survival for δ-type subjects in the ProCESS trial. Kaplan Meier plots for 365-day survival, stratified by treatment arm (dotted, early goal direct therapy; solid, usual care) for all patients in δ-type in Panel A (n = 205), patients in core δ-type in Panel B (n = 76), and patients in margin δ-type in Panel C (n = 129). At risk subject counts shown below graphs. Models are adjusted for patient age, race, sex, SOFA at randomisation, and Charlson comorbidity index. Abbreviation: EGDT, early, goal directed therapy.

Discussion

In this retrospective analysis of multiple datasets from patients with sepsis, most patients changed clinical subtypes after presentation. The classification of subtypes was fuzzy, with more than of half of patients in the margin of their assigned type. Patients on the margin of the initial subtype were more likely to change subtypes in 48 h than those in the core, independent of treatment. The conclusions about the treatment benefit or harm of fluid resuscitation were sensitive to whether patients began in the subtype core or margin, especially in patients in the δ-type.

Clinical sepsis subtypes are now reproduced in thousands of patients in multicenter, international datasets.^1,2,5 Building on this generalisability, these data show that subtypes are dynamic, with most patients changing types within 48 h of presentation. Multiple mechanisms likely explain subtype change, including clinician variability in care, timing of source control, natural progression of disease, or even chronic disease burden. This contrasts with earlier reports from subtypes in acute respiratory distress syndrome (ARDS), where subtype change over multiple days was less common.⁸ Such findings could be explained by differences in data source (trial vs. cohort), data collection (electronic health record vs. case report form), and cadence of subtype assignment (hourly vs. daily).

Subtypes are not discrete. Like the well-known diagnostic uncertainty in sepsis, there is subtype classification uncertainty, as demonstrated by the broad distribution of membership probabilities. Patients with lower membership probabilities, on the margins, were more likely to change groups independent of illness acuity and even treatment. A suggestion of subtype uncertainty was earlier reported in paediatric patients with septic shock using transcriptomic data, and recent work has sought to create a metrics to quantify uncertainty in assignment between subtype models.¹² Fundamentally, these data suggest that all clinical subtype members are not the same, and that exploration for a single underlying biologic mechanisms or treatable trait inside clinical subtypes may be premature.

Many prior reports suggest that specific subtypes have a differential treatment benefit from interventions in critical illness, including oxygenation strategies, fluid administration, or immune modulation.¹³ These data are typically uncovered in secondary analyses of prior randomised trials, where subtype-specific treatment effects would be assumed to be discretely bounded and homogeneous across subtype members. However, the analysis of the ProCESS trial herein suggests the opposite, that treatment effects are distributed differently for subtype patients with different membership probabilities. For example, early, goal-directed therapy was associated with harm only in core delta-type members, not those on the margins with lower membership probabilities.

These findings have additional implications. First, they highlight the tension between pragmatic strategies to move subtype classification to clinical care which emphasise ease & discrete-ness -vs.- the reality of fuzzy boundaries between groups. Second, particularly for trialists seeking enrichment strategies for precision therapies, clinical subtypes may need both multiple, complimentary statistical models to confirm membership, as well as quantitative measures of certainty to reduce misclassification or methods that more precisely calculate conditional average treatment effects for individuals. It is unlikely that averaged, post hoc treatment effects observed in a clinical subtype will extend to all members in prospective evaluation. Third, subtypes derived from single trials or datasets using a single approach may be more susceptible to misclassification, perhaps, than those built from the layering of validated subtypes—so-called subtype ensembles.⁵ Fourth, recent work suggests that any grouping strategy may be inferior to estimating the predicted treatment effects for the individual.¹⁴

This study has several limitations. First, only clinical subtypes were analysed, and the inclusion of other subtype labels from muti-omic data could change conclusions about subtype trajectories and membership. Second, no trajectory models were developed to understand group membership changes over time. The study concept was to inform trial enrichment and design, where knowledge of a subtype probability would only be available at enrolment, not the longitudinal behaviour. Third, because missing data were common for some variables included in models, multiple imputation was used. Prior work demonstrates clinical subtypes are robust to missing-ness, feature selection, and case mix.⁴ Fourth, we evaluated subtype change over a short window, and longitudinal behaviour may be different later in the sepsis episode. Fifth, subtype behaviour and membership may be influenced by the data source, and these findings may not be generalisable to all health systems and are not generalisable to low, middle-income countries.

Conclusions

In this retrospective analysis of datasets from patients with sepsis, clinical subtypes were dynamic and not discretely bounded. Patients with fuzzy classification on the margins were more likely to change groups and experience different treatment effects than those with greater subtype membership probabilities. Further research is needed to determine the impact of these data on mechanistic discovery and randomised trial enrichment using clinical subtypes.

Contributors

JNK and CWS directly accessed and verified the underlying data reported in the manuscript. JNK, SI, PCN, KMR, JC, LT, DY, DCA, RA, LvV, and CWS met criteria for authorship and approve of the submission of this manuscript. No medical writers were used in the writing of this manuscript.

Design and conduct of the study: JNK, KMR, JC, LT, DY, DCA, CWS.

Data acquisition, analysis, and interpretation of the data: JNK, SI, PCN, KMR, JC, LT, DY, DCA, RA, LvV, and CWS.

Statistical Analysis: JNK, CWS, JC, LT.

Initial draft drafting of the manuscript: JNK, CWS.

Critical revision of the article for important intellectual content: JNK, SI, PCN, KMR, JC, LT, DY, DCA, RA, LvV, and CWS.

Final approval of the article: JNK, SI, PCN, KMR, JC, LT, DY, DCA, RA, LvV, and CWS.

Obtaining Funding: CWS, DY, DCA.

Data sharing statement

Deidentified data and the data dictionary will be shared with approval from the CRISMA Center and a signed data access agreement. All requests should be sent to seymourc@pitt.edu.

Declaration of interests

Outside of this submitted work, the following declaration of interests include: CWS, JC, and LT report grants from the National Institutes of Health. LT reports grants from the National Science Foundation. LvV reports grants from the Netherlands Organisation for Health Research and Development ZonMW. CWS reports consulting fees from Deepull, reports support for attending meetings and travel from ISICM, CCCF, ISF, Monash University, ESICM, and JAMA, and reports participation in the DSMB for the RENOVATE trial. CWS is also Associate Editor at JAMA, receiving compensation from the American Medical Association.

The remaining authors declare no other interests.