Written by: Maaike C SwetsZsuzsa BakkAnnette C WestgeestFelicia RuffinRachel KornJean-Francois JabbourPriscilla LaLuke TysallSimon DewarRebecca K SutherlandMark G J de BoerGeert H GroeneveldDavid H DockrellThomas W van der VaartVance G FowlerClark D RussellIDISA study team
Patients were included from 3 observational studies: a cohort from an ongoing UK (Scotland) retrospective study (Edinburgh cohort 2; n = 463) [3, 6]; the Dutch IDISA (Improved Diagnostic Strategies in Staphylococcus aureus bacteremia study) multicenter prospective study (IDISA cohort; n = 490) [12], and the US SABG-PCS (Staphylococcus aureus Bacteremia Group Prospective Cohort Study) prospective study (SABG-PCS cohort; n = 755) [10]. The Edinburgh 2 cohort included consecutive unique adults (aged ≥18 years) with monomicrobial SAB diagnosed between 28 August 2022 and 18 December 2024, approved by the South East Scotland Research Ethics Committee. No patients included in our previous analysis were included in this second cohort of patients. The IDISA cohort enrolled adults (aged ≥18 years) with SAB between 1 July 2017 to 30 September 2019, from 7 hospitals in the Randstad metropolitan region of the Netherlands, approved by the Medical Ethics Committee of the Academic Medical Centre Amsterdam. The cohort from SABG-PCS analyzed in this study included adults (aged ≥18 years) with monomicrobial SAB enrolled 1 January 2016 to 1 January 2020, at Duke University Medical Center in North Carolina, approved by the Duke Institutional Review Board.The Charlson Comorbidity Index (CCI) was used to define comorbid conditions, with adjustment for age. Vascular disease was defined as a history of peripheral vascular disease, myocardial infarction, or stroke. The route of acquisition of infection was classified based on the criteria outlined by Friedman et al [13]. The portal of entry was defined as the most probable entry point of S. aureus into the bloodstream. Metastatic infection was defined as the clinical, radiologic, or microbiologic identification of foci of infection remote from the portal of entry, thought to have arisen through hematogenous dissemination, and it included infective endocarditis.In the SABG-PCS cohort, heart rate, temperature, and creatinine level were collected as ordinal variables (Supplementary Table 1). For the Edinburgh 2 and IDISA cohorts, they were continuous variables. In the IDISA cohort, recorded vital signs and laboratory measurements were the most extreme values within 24 hours of the index blood culture. In the Edinburgh 2 and SABG-PCS cohorts, vital signs and laboratory measurements were recorded closest to the time of the index blood culture. In the Edinburgh 2 cohort, spa typing and Panton-Valentine leukocidin (pvl) gene quantitative reverse-transcription polymerase chain reaction were performed by the Scottish Microbiology Reference Laboratory. In the SABG-PCS cohort, typing and confirmation of USA300 genotype were performed as described elsewhere [10].
spa
Persistent bacteremia was defined as another positive blood culture >48 hours after the index blood culture in the Edinburgh 2 cohort [14] and >96 hours after the index culture in the IDISA cohort. In the SABG-PCS cohort, persistent bacteremia was defined as persistently positive blood cultures for ≥5 days after initiation of appropriate antimicrobial therapy. Recurrent bacteremia was defined as another positive blood culture within 90 days of stopping treatment in the Edinburgh 2 and SABG-PCS cohorts and within 90 days of the index blood culture in the IDISA cohort. In the Edinburgh 2 cohort, this was restricted to bacteremia caused by an isolate from the same spa-inferred clonal complex as the index bacteremia.
LCA was used to identify subgroups within the different cohorts. A detailed description was given elsewhere [6]. Indicator variables used to identify subgroups were the same as in our previous analysis except that endocarditis and other metastatic complications were included as a single variable rather than separately, due to data availability. Classes were formed without consideration of outcomes. Nonnormally distributed variables with right skew were log-transformed. Missing values were handled with full information maximum likelihood. A combination of the bayesian information criteria (BIC), number of classes, size of the smallest class, and clinical interpretation were used for model selection. We first identified models using the BIC (the most conservative index where a decrease indicates a more informative model), then checked smallest class sizes (to avoid uninformative and underpowered classes), and then applied clinical interpretation (eg, to ensure that all classes were clinically distinct).
To avoid a local maximum, 16 random starting values were used, with 50 iterations performed for each starting value [15]. The results were reviewed to confirm that the same maximum likelihood solution was found. Following class identification, we calculated the posterior probability of class membership for each individual across all identified classes, and we assigned each individual to the class with the highest posterior class membership probability. We corrected for misclassification error using bias-adjusted 3-step LCA [16]. LCA was done using the Latent GOLD 6.1 statistical software package [17].
To compare continuous variables, the Mann-Whitney U test was used for comparison between 2 cohorts and the Kruskal-Wallis test for comparison between all 3 cohorts. Categorical variables were compared by contingency table analysis, using χ2 test if cell counts were ≥5 and Fisher exact test if they were <5. P values <.05 was deemed to indicate statistical significance, and z scores were used to visualize the comparison of variables and outcomes between subphenotypes, calculated using the following formula: z = (value for subphenotype − mean for variable)/SD for variable. GraphPad Prism software ((version 10.5.0 for macOS) was used for data visualization and analysis.
RESULTS
Characteristics of the Edinburgh 2, IDISA, and SABG-PCS cohorts are compared in Table 1. Notable differences between these cohorts included a lower prevalence of chronic kidney disease (CKD) in the Edinburgh 2 cohort and a lower prevalence of people who inject drugs in the IDISA cohort. In the SABG-PCS cohort there was substantially greater multimorbidity (as determined with the CCI), despite a slightly lower median age. Infective endocarditis was diagnosed less frequently in the Edinburgh 2 cohort. MRSA bacteremia was uncommon in the Edinburgh 2 and IDISA cohorts (rates, 2.8% and 2.0%, respectively) compared with the high rate in the SABG-PCS cohort (40.2%). Persistent bacteremia was also more common in the SABG-PCS cohort despite a more stringent definition (requiring positive blood cultures for ≥5 days after starting treatment). Across the cohorts, most episodes of SAB were acquired in the community, and an unknown portal of entry was the most frequent individual category, followed by skin or soft-tissue infection and intravenous catheters.
Comparison of Cohort Characteristics
Abbreviations: CCI, Charlson Comorbidity Index; CKD, chronic kidney disease; CRP, C-reactive protein; IDISA, Improved Diagnostic Strategies in Staphylococcus aureus bacteremia study; IDU, injection drug use; IQR, interquartile range; MRSA, methicillin-resistant Staphylococcus aureus; NA, not available; NHS, National Health Service; SABG-PCS, Staphylococcus aureus Bacteremia Group Prospective Cohort Study; SSTI, skin or soft-tissue infection.
aData represent no. (%) of patients unless otherwise specified.
b P values based on Mann-Whitney U test for continuous values (displayed as median [IQR]) compared between 2 cohorts, Kruskal-Wallis test for continuous values compared between 3 cohorts, and χ2 test for categorical variables (displayed as no. [%]).
cClass-defining variable included in latent class analysis.
dDiabetes mellitus of any severity (data available for 754 patients in the SABG-PCS dataset).
ePeripheral vascular disease, myocardial infarction, or stroke.
fImplanted cardiac devices, including pacemakers and implantable automatic cardioverter-defibrillator and left ventricular assist devices but not including prosthetic heart valves.
gVital signs and laboratory measurements were recorded at the time of the index blood culture (see further details in Methods).
Consistent with our previous work, BIC values, class sizes, and clinical interpretability together favored the 5 class models (Supplementary Table 2). Overall, when class-defining variables were compared between the classes within each cohort, the 5 classes replicated the previously described subphenotypes A–E (Figure 1 and Supplementary Tables 3–5). Subphenotype A was associated with older age, male sex, and multimorbidity (dementia and vascular disease). Subphenotype B was associated with nosocomial SAB, intravenous catheter as the portal of entry, less metastatic infection, and overall less comorbidity. Subphenotype C was associated with community-acquired SAB from an unknown portal of entry ad with metastatic infection and higher C-reactive protein levels. Subphenotype D was associated with CKD, healthcare contact (either community-acquired healthcare-associated or nosocomial acquisition), and vascular disease. Hemodialysis was not included as a class-defining variable but was common in subphenotype D (Edinburgh 2, 16 of 34 cases; IDISA, 22 of 92; SABG-PCS, 76 of 124). Subphenotype E was associated with SAB in younger people without CKD, dementia, or vascular disease. This subphenotype was associated with acquisition through injection drug use (IDU), metastatic infection, and coexistent liver disease.
Reproducible clinical subphenotypes of Staphylococcus aureus bacteremia. Comparison of class sizes and class-defining variables between subphenotypes for the UK Edinburgh 2 cohort (A), the Dutch IDISA (Improved Diagnostic Strategies in Staphylococcus aureus bacteremia study) cohort (B), and the US SABG-PCS (Staphylococcus aureus Bacteremia Group Prospective Cohort Study) cohort (C). Letters A–E above columns represent the subphenotypes described in the text. Vertical bar graphs show the proportions of patients assigned to each subphenotype within each cohort. Heat map cells are shaded according to row z score (ie, comparing subphenotypes), except for “Acquisition” and “Portal of Entry,” where shading is by column z score (ie, comparing within each subphenotype). Intensity of red shading reflects a more positive z score (ie, above the mean), and intensity of blue shading reflects a more negative z score (ie, below the mean). Abbreviations: IDU, injection drug use; MRSA, methicillin-resistant S. aureus; SSTI, skin or soft-tissue infection.
Certain characteristics were less consistently associated with clinical subphenotype. Of particular clinical relevance, the presence of cardiac prosthetic material was variably associated with subphenotypes A, C, and D across the 3 cohorts. The association of subphenotype A with type of acquisition was also inconsistent (acquisition predominantly community onset and non–healthcare associated in the Edinburgh 2 cohort, community onset and healthcare associated in IDISA, and nosocomial in SABG-PCS).
The contribution of each class-defining variable to the clustering was quantified using the R2 value (Supplementary Figure 1), indicating that age, acquisition, CKD, IDU, metastatic infection, and portal of entry made the greatest individual contributions to class assignment. MRSA bacteremia was consistently associated with subphenotypes A (old age and multimorbidity), B (nosocomial) and E (IDU) but overall did not make a large contribution to the clustering.
The extent of multimorbidity, determined using the CCI, differed between the 3 cohorts, with patients in the SABG-PCS cohort having the highest CCI and those in the Edinburgh 2 cohort having the lowest (Table 1 and Supplementary Figure 2). CCIs were compared across subphenotypes, identifying the greatest extent of multimorbidity in subphenotypes A and D, and the lowest in B and E, consistent with our previous findings (Figure 2A). Despite the greater baseline multimorbidity, this pattern of differences in CCIs remained in the SABG-PCS cohort. Diabetes mellitus was not included as a class-defining variable but is associated with the risk of acquiring SAB [18]. A diagnosis of diabetes mellitus (of any severity) was consistently associated with subphenotypes D (CKD SAB) and A (older age, multimorbidity SAB) (Figure 2B).
Multimorbidity, measured with the Charlson Comorbidity Index (CCI) (A), and diabetes mellitus of any severity (B), both stratified by clinical subphenotype. A, Box-and-whisker plot drawn using the Tukey method. Boxes shows interquartile range; horizontal lines, median. CCIs were compared between subphenotypes within each cohort, using the Kruskal-Wallis test. C, Bars represent z scores comparing subphenotypes within each cohort. The raw proportion of patients with diabetes mellitus was compared between subphenotypes within each cohort, using the Fisher exact test. Abbreviations: IDISA, Improved Diagnostic Strategies in Staphylococcus aureus bacteremia study; SABG-PCS, Staphylococcus aureus Bacteremia Group Prospective Cohort Study.
The distribution of spa-inferred clonal complexes of isolates from the Edinburgh 2 and SABG-PCS cohorts differed substantially (Supplementary Table 6), with only CC8 and CC15 identified in both cohorts. The distribution of clonal complexes did not differ substantially between subphenotypes within either cohort (Figure 3). Panton-Valentine leukocidin toxin gene carriage was uncommon among isolates from the Edinburgh 2 cohort (7 of 463 isolates, [1.5%]) and did not differ substantially across the subphenotypes (Supplementary Figure 3A). The USA300 clone accounted for 14.7% (111 of 755) of isolates in the SABG-PCS cohort. Although numerically more common in subphenotype E, this difference was not statistically significant (P = .2) (Supplementary Figure 3B).
Staphylococcus aureus spa-inferred clonal complex (CC) in bloodstream isolates from the Edinburgh 2 cohort (A) and the SABG-PCS (Staphylococcus aureus Bacteremia Group Prospective Cohort Study) cohort (B), stratified by subphenotype. Proportions represent the proportion of isolates within each subphenotype. Abbreviation: NA, not available.
Subphenotype A was consistently associated with highest 90-day all-cause mortality rate (47.9% for Edinburgh 2, 78.3% for IDISA, and 77.9% for SABG-PCS), whereas B (21.1%, 39.4%, and 20.2%, respectively) and especially E (4.0%, 11.2%, and 21.2%) were associated with the lowest (Figure 4A, Supplementary Figure 4, and Supplementary Tables 3–5). Persistent bacteremia was most common in subphenotypes C (4.1% for Edinburgh 2, 16.1% for IDISA, and 36.0% for SABG-PCS) and E (10.2%, 14.5%, and 53.7%, respectively) (Figure 4B and Supplementary Tables 3–5). Subphenotype D was consistently associated with recurrent bacteremia, although this difference was not statistically significant (Figure 4C and Supplementary Tables 3–5).
Survival and microbiologic outcomes differ between clinical subphenotypes, as shown by comparison of 90-day all-cause mortality rate (A), persistent bacteremia (B), and recurrent bacteremia (C) by subphenotypes. Bars represent z scores comparing subphenotypes within each cohort. The raw proportion of each outcome was compared between subphenotypes within each cohort using Fisher exact test. Abbreviations: IDISA, Improved Diagnostic Strategies in Staphylococcus aureus bacteremia study; SABG-PCS, Staphylococcus aureus Bacteremia Group Prospective Cohort Study.
DISCUSSION
Approaches to achieve patient stratification remain an unmet need in SAB, to facilitate personalized clinical management and to address the heterogeneity in responses to treatments investigated in clinical trials. In the current study we have analyzed 3 cohorts of adults with SAB, including a total of 1708 patients, replicating the 5 clinical subphenotypes we described elsewhere [6]. Importantly, we included geographically diverse observational studies reflecting “real-world” patients with SAB, including varied S. aureus genetic lineages, and a high proportion of MRSA and USA300 bacteremia. Consistent with our original report, subphenotypes represent 5 clinically distinct archetypes of SAB [7], and they differ in survival and microbiologic outcomes. The reproducible identification of clinically relevant subphenotypes supports the contention that investigation of SAB as a single syndrome is overly simplistic and may be impeding efforts to personalize care and improve outcomes.
Multimorbidity, and diabetes mellitus specifically, was more prevalent in the subphenotypes associated with highest mortality rates (A and D), highlighting the need to understand how these processes affect host defense and response to treatment. Immune dysfunction associated with CKD could also be a specific contributor to the increased mortality risk and potentially to recurrence in subphenotype D. This would be consistent with the overall increased risk of infection-related death in people with CKD, including in comparison with other comorbid conditions [19–21]. Identification of metastatic foci was more prevalent in subphenotypes C (community acquired metastatic SAB) and E (IDU-associated SAB), which were consistently associated with persistent bacteremia but not with death. The predisposition to metastatic infection in subphenotype C remains unclear, not related to the host or pathogen characteristics investigated in this study. Investigation of host responses and the impact of pathogen genetic variation at the whole-genome level in this subphenotype could be particularly informative to define host and/or pathogen determinants of metastatic infection during SAB.
Our study has several strengths, providing confidence in our conclusions. The current analysis includes a large cohort from the United States with a high prevalence of MRSA bacteremia (and the USA300 clone), which was missing from our original study. Compared with other variables, MRSA infection itself did not have a large impact on class assignment, providing reassurance that the subphenotypes are generalizable to settings with a higher prevalence of MRSA. Related to this, the distribution of S. aureus genetic lineages was almost entirely different in the Edinburgh 2 versus SABG-PCS cohorts, providing further reassurance regarding generalizability. We studied 3 well-characterized patient cohorts which used the same definitions of the class-defining variables used in the analysis. There was a very high degree of overlap in availability of the variables, and all patients had follow-up extending ≥90 days after the date of the index blood culture. Two of the 3 cohort studies were prospective. The use of 3 observational cohorts provides confidence that these subphenotypes are applicable to patients encountered during “real-life” clinical practice, which can differ from patients included in clinical trials [5, 22].
Subjectivity in model selection represents an important limitation to our approach. For the Edinburgh 2 cohort, the 5-class model had the lowest BIC value and best clinical interpretability. In both the IDISA and SABG-PCS cohorts, although the BIC value marginally favored the 4-class models, clinical interpretability favored the 5-class models, with an acceptable number of patients in the smallest class and no significant differences in entropy (as a measure of class separation). In addition, the study time periods were not overlapping between cohorts, data collection was retrospective in the Edinburgh 2 cohort, and definitions of persistent SAB differed. Despite the differences in the definition of persistent SAB, consistent associations with subphenotypes were observed, supporting the robustness of these associations.
The competing risk of death was not considered when examining associations between subphenotype membership and persistent or recurrent SAB. The low number of recurrences in each cohort prevented statistical confirmation of a difference between the subphenotypes. Granular definitions of sepsis (such as the Sequential Organ Failure Assessment score) were not recorded for all cohorts so could not be included in the model nor compared between subphenotypes. Future work will define a prediction model for use in clinical practice and research studies for subphenotype assignment of individual patients. This will then allow prospective evaluation of the impact of subphenotype assignment on personalizing diagnostic and therapeutic strategies in clinical trials.
An alternative approach has recently been reported by Gutiérrez-Gutiérrez and colleagues [23], with a focus on early clinical risk stratification using variables available within the first 24 hours after the index blood culture. This approach identified 3 clusters based on portal of entry (skin or soft tissue; intravascular, mainly catheter related; and other/unknown) with subsequent identification of clusters within each with higher or lower 30-day mortality risk. We agree that this is a useful and carefully validated tool for clinicians seeking to stratify the risk of death early in infection. This differs from and complements our approach, which has defined holistic subphenotypes of disease which differ in mortality risk and also microbiologic outcomes and response to adjunctive rifampicin (in the ARREST trial).
In summary, we report 5 reproducible clinical subphenotypes of SAB that provide a framework to rationalize the substantial heterogeneity intrinsic to this high morbidity and mortality infection. The current study extends the total number of patients included in our analyses to 3138, spanning 5 separate studies from the United Kingdom, Spain, the Netherlands, and the United States. We have also established the relevance of these subphenotypes in the setting of MRSA bacteremia, the USA300 clone, and patients with a spectrum of multimorbidity. We think that these subphenotypes are robust and can be used to stratify recruitment to and analysis of clinical trials in SAB, in addition to investigation of disease pathogenesis.
VF
Dr. Vance Fowler
Infectious Disease · General Infectious DiseaseDuke Medical Centervance
