Identification and validation of clinical phenotypes in Staphylococcus aureus bloodstream infection and their association with mortality (FEN-AUREUS study)

Identification and validation of clinical phenotypes in Staphylococcus aureus bloodstream infection and their association with mortality (FEN-AUREUS study) Belén Gutiérrez-Gutiérrez, a , b , ∗ Belén Gallego-Mesa, a , b Achim J. Kaasch, c Matthias Riediger, c Siegbert Rieg, d Marta Trigo, e Sonsoles Salto-Alejandre, f Francisco Anguita-Santos, g Ángela Cano, h , b Andrea Prolo-Acosta, i Salvador López-Cárdenas, j María Teresa Pérez-Rodríguez, k Francisco Javier Martínez-Marcos, l Esperanza Merino-Lucas, m Blanca Anaya-Baz, n Ana Arizcorreta, o Antonio Plata-Ciézar, p Marco Piscaglia, q Alexandra Aceituno, r Lidia Romero-Calderón, a , b Daniel Hornuss, d Aurora Alemán-Rodríguez, a , b Adelina Gimeno-Gascón, f Esther Recacha, a , b Julián Torre-Cisneros, h Nicolás Merchante, e Álvaro Pascual, a , b Luis Eduardo López-Cortés, a , b and Jesús Rodríguez-Baño a , b a Unidad de Enfermedades Infecciosas y Microbiología, Hospital Universitario Virgen Macarena and Departamentos de Medicina y Microbiología, Universidad de Sevilla/Instituto de Biomedicina de Sevilla/CSIC, Seville, Spain b CIBER de Enfermedades Infecciosas (CIBERINFEC). Instituto de Salud Carlos III, Madrid, Spain c Institute of Medical Microbiology and Hospital Hygiene, Medical Faculty, Otto von Guericke University Magdeburg, Magdeburg, Germany d Division of Infectious Diseases, Department of Medicine II, Medical Center, Faculty of Medicine, University of Freiburg, Freiburg, Germany e Unidad de Enfermedades Infecciosas y Microbiología. Hospital Universitario de Valme, Instituto de Biomedicina de Sevilla (IBiS), Universidad de Sevilla, Sevilla, Spain f Clinical Unit of Infectious Diseases, Microbiology and Parasitology/Instituto de Biomedicina (IBIS), Virgen del Rocío University Hospital/ CSIC/University of Seville, Seville, Spain g Servicio de Enfermedades Infecciosas, Hospital Universitario San Cecilio, Granada, Spain h Unidad de Enfermedades Infecciosas, Hospital Universitario Reina Sofía, Universidad de Córdoba/Instituto Maimónides de Investigación Biomédica de Córdoba (IMIBIC), Córdoba, Spain i Servicio de Enfermedades Infecciosas, Hospital Virgen de la Victoria, Málaga, Spain j Unidad de Enfermedades Infecciosas y Microbiología Clínica, Hospital Universitario de Jerez/Departamento de Medicina y Cirugía, Universidad de Cádiz/Instituto de Investigación e Innovación Biomédica de Cádiz, Cádiz, Spain k Servicio de Medicina Interna, Hospital Álvaro Cunqueiro, Vigo, Pontevedra, Spain l Servicio de Enfermedades Infecciosas, Hospital Juan Ramón Jiménez, Huelva, Spain m Infectious Diseases Unit, Dr. Balmis General University Hospital, Alicante Institute for Health and Biomedical Research (ISABIAL), Clinical Medicine Department, University Miguel Hernández of Elche, Alicante, Spain n Unit of Infectious Diseases, Hospital Universitario Puerto Real/Institute of Biomedical Research and Innovation of Cádiz (INIBICA), Cádiz, Spain o Unidad de Enfermedades Infecciosas del Hospital Universitario Puerta del Mar, Cádiz, Spain p Servicio de Enfermedades Infecciosas, Hospital Regional de Málaga, Spain q Department of Infectious Diseases, ASST FBF Sacco, Milán, Italy r Hospital Universitario Torrecárdenas, Almería, Spain Summary Background Staphylococcus aureus bacteraemia (SAB) is heterogeneous in patients and infection-related features. The aim of the study was to identify clinical phenotypes among patients with SAB, to evaluate their association with mortality, and to derive and validate a simplified probabilistic model for phenotypes assignment. Methods Phenotypes were derived using two-stage cluster analysis of 2128 patients from the ISAC cohort (recruited between 2013 and 2015), analysing 62 variables. Cox regression assessed phenotype–mortality associations. Logistic regression was employed to develop a simplified probabilistic model for sub-phenotype allocation, validated in two external international cohorts: INSTINCT (1217 patients, recruited between 2006 and 2011) and FEN-AUREUS (1185 patients, recruited between January 2021 and October 2024). The association between sub-phenotypes and 30-day mortality in the validation cohorts was also assessed. *Corresponding author. Unidad Clínica de Enfermedades Infecciosas y Microbiología, Hospital Universitario Virgen Macarena, Avda Dr Fedriani 3, 41009, Seville, Spain. E-mail address: bgutierr[email protected] (B. Gutiérrez-Gutiérrez). eClinicalMedicine 2025;83: 103240 Published Online 7 May 2025 https://doi.org/10. 1016/j.eclinm.2025. 103240 www.thelancet.com Vol 83 May, 2025 1 Articles Findings Cluster analysis identified three clinical phenotypes based on the probable portal of entry: A (skin and soft tissues), 458 cases; B (vascular device-associated), 573 cases; and C (other portals of entry or unknown), 1097 cases. Their 30-day mortality was significantly different (13⋅1%, 18⋅2% and 25⋅3%, respectively, p < 0⋅001). Each phenotype contained two sub-phenotypes with differing characteristics and mortality risks. Also, three phenotypes were found in the INSTINCT cohort, which clustered on the same portals of entry, with two sub-phenotypes in each. When the simplified probabilistic model was applied, the sub-phenotypes showed significant associations with 30-day mortality in both validation cohorts. In INSTINCT, the aHRs were 1⋅93 (A2 vs A1), 3⋅40 (B2 vs B1), and 3⋅04 (C2 vs C1). In FEN-AUREUS, the aHRs were 2⋅02 (A2 vs A1), 2⋅11 (B2 vs B1), and 2⋅44 (C2 vs C1). Interpretation Patients with SAB can be classified into phenotypes and sub-phenotypes, each exhibiting considerable variations in mortality rates. To facilitate clinical application, a validated open-access algorithm and calculator for phenotype and sub-phenotype assignment have been developed, enabling their use at the time of SAB confirmation. This tool aims to support timely and personalised patient care. Funding Instituto de Salud Carlos III, Spanish Ministry of Science, Innovation and Universities (PI21/01801). Copyright © 2025 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/). Keywords: Staphylococcus aureus; Phenotypes; Mortality; Bacteraemia; Clinical profiles Introduction Staphylococcus aureus is one of the most frequent microorganisms causing disease in humans and a leading cause of bloodstream infections globally. 1,2 S. aureus bacteraemia (SAB) results in substantial morbidity and healthcare costs, with complications being common and mortality rates ranging between 20% and 40%. 1–3 Bacteraemic infections due to S. aureus are typically very heterogeneous in terms of patients’underlying conditions, pathogen characteristics, portals of entry, type of acquisition, site of infection and acute severity. 1,4,5 Moreover, there may be intrinsic clinical heterogeneity in SAB that extends beyond these variables, 6 making the establishment of management standards of care very complex. 7,8 This is reflected in the large practice variations across the world. 9 It also poses specific difficulties for both designing and interpreting the results of randomised trials. 6,10 We hypothesise that patients who develop SAB can be classified into clinical phenotypes associated with Research in context Evidence before this study We searched PubMed, Scopus, and medRxiv up to October 2024 using the terms [“Staphylococcus aureus”OR “S. aureus”] AND [“phenotypes”OR “clinical features”] with no language restrictions to identify any study characterising clinical phenotypes in bloodstream infections caused by Staphylococcus aureus (SAB). To our knowledge, only one prior study, based on the analysis of 1430 cases and published online in June 2024, attempted to classify SAB patients into phenotypes. No further studies explicitly characterising SAB phenotypes or investigating their association with outcomes were identified. Added value of this study Based on a cohort of 2128 patients, we identified, through comprehensive cluster analysis of clinical and laboratory data collected within the first 24 h from blood culture extraction, three distinct clinical phenotypes in SAB patients. Additionally, two sub-phenotypes were identified within each phenotype. Both the phenotypes and sub-phenotypes obtained differed in acquisition type, comorbidities, bacteraemia source, laboratory data, and antibiotic susceptibility, and showed significantly different 30-day mortality risks. Based on these results, we developed an algorithm using simplified probabilistic models that easily assign patients to phenotypes/sub-phenotypes, which was validated in two external cohorts of 1217 and 1185 cases, respectively. This study provides a novel framework to stratify SAB patients into phenotypes and sub-phenotypes with different mortality risks, enabling clinicians worldwide to identify high-risk patients requiring closer monitoring. To this end, we have made these models publicly and freely available through an online tool at http://fen-aureus.com. Implications of all the available evidence The identification and validation of phenotypes and subphenotypes in SAB represents a step forward in understanding this heterogeneous type of infection. These phenotypes and sub-phenotypes could be used, in addition to helping clinicians identify high-risk patients, to explore differences in the underlying pathophysiological mechanisms of phenotypes and sub-phenotypes, refine patient selection for clinical trials, or optimise therapeutic strategies. Articles 2 www.thelancet.com Vol 83 May, 2025 prognosis. The phenotypes may be the consequence of subjacent subtle different pathophysiological mechanisms and may also be associated with outcomes, which therefore may have important clinical and management implications. As examples in infectious diseases, clinical phenotypes have been identified in patients with sepsis 11 or COVID-19. 12 The objective of this study was to identify SAB phenotypes, investigate their association with mortality, and develop simplified probabilistic models for assigning patients to these phenotypes, with the aim of facilitating their application in clinical practice. Methods Study design and population This study is part of the FEN-AUREUS project (clinicaltrials.gov identifier: NCT06574399). Clinical phenotypes in patients with SAB were derived using data from the International Staphylococcus aureus Collaboration (ISAC) prospective cohort (2128 patients recruited from 2013 to 2015 from 11 hospitals in 5 countries); the correlation of phenotypes with 30-day mortality was assessed, and a simplified probabilistic model for phenotype allocation was derived. The phenotypes allocation model and the association of phenotypes with mortality were externally validated in the INvasive S. aureus INfections CohorT (INSTINCT), a prospective cohort involving 1217 patients from 2 hospitals, recruited from 2006 to 2011. The methodological aspects of both cohorts were previously published. 13,14 Finally, the FEN-AUREUS prospective cohort, performed in 19 centres from two countries from January 2021 to October 2024 and including 1185 patients, was also used for contemporary external validation. In all cohorts, patients were eligible for the study if they were at least 18 years of age and had at least one blood culture positive for S. aureus with accompanying clinical symptoms and signs of infection. Written informed consent was obtained from all participants prior to inclusion in each of the three prospective cohorts (ISAC, INSTINCT, and FENAUREUS). An overview of the analyses performed in the derivation and validation cohorts is presented in Fig. 1. The FEN-AUREUS project was approved by the Ethics Committee of Biomedical Research of Andalusia (1428-N-22). The study is reported according to the STROBE recommendations. Variables and definitions The variables considered to derive the clinical phenotypes were collected within the first 24 h from blood culture extraction and included age, gender, comorbidities, blood chemistry and blood cell counts, likely S. aureus-portal of entry, source of bacteraemia, severity of infection and methicillin-susceptibility. The full list of variables is shown in Table 1. Additionally, variables related to early clinical management, such as antibiotic treatment, source control or ID consultation, and centre, were included. The main endpoint was 30-day all-cause mortality; patients discharged before day 30 were ISAC cohort (Derivaon cohort) (n=2,128) INSTINCT cohort (1st External validaon cohort) (n=1,217) Databases Derivaon cohort Phenotypes Likely portal of entry + Predicve models Comparison External validaon cohorts Mortality B Procedures AC Mortality Phenotypes + Sub-phenotypes Sub-phenotype A1 Sub-phenotype A2 Sub-phenotype B1 Sub-phenotype B2 Sub-phenotype C1 Sub-phenotype C2 Cluster analysis FEN-AUREUS cohort (2nd External validaon cohort) (n=1,185) INSTINCT FEN-AUREUS Fig. 1: Databases used and procedures performed. Articles www.thelancet.com Vol 83 May, 2025 3 contacted by phone at the end of the follow-up period to assess their status. The probable portal of entry for SAB refers to the anatomical site through which the pathogen is believed to have entered the bloodstream. This was determined based on specific criteria (Supplementary Table S1), with only permitted one portal per case. They were classified as vascular catheter-related, skin and soft tissues, respiratory tract, and genito-urinary tract. If no probable portal could be determined, the entry was classified as unknown. The site of infection refers to the primary organ site where the infection was located at the time the first positive blood culture was taken. The site of infection might or might not coincide with the portal of entry (e.g., a vascular catheter may be the portal of entry in a patient who later develops knee arthritis as the site of infection). Eventually, more than one source could be identified in a patient. Sources were determined based on clinical data, imaging studies, and isolation of the microorganism in specific source-related samples. Sources considered are listed in Supplementary Table S1. Source control was defined as follows: if the bacteraemia was associated with a removable intravascular device, the source was considered controlled if the device was removed within 3 days. Alternatively, if the bacteraemia was related to a removable non-device source, such as an abscess, the source was considered controlled if the non-device related source was addressed (e.g., drained or surgically removed) within the same timeframe. Missing data and phenotypes derivation The proportion of missing data per variable in the ISAC cohort is shown in the appendix (Supplementary Table S2); after discarding the hypothesis that data was missing completely at random by Little’s MCAR test, missing data were completed by multiple imputation using the Markov chain Monte Carlo method. To identify the phenotypes in the derivation cohort (ISAC cohort), we first assessed the distribution of values and correlations among variables using the chisquare test for categorical variables and Pearson’s correlation coefficient for continuous variables, after verifying assumptions of normality and linearity. When these assumptions were not fulfilled, a Spearman’srankcorrelationmatrixwasusedasanalternative to evaluate associations and detect potential collinearity between continuous variables. When two variables were highly correlated, one of them was excluded to eliminate collinearity A total of 62 variables were included to derive the clinical phenotypes. As the aim was to explore the potential existence of phenotypes, no preselection of variables was performed. An unsupervised two-step cluster analysis was then conducted using both continuous and Factor A a (n = 458) B b (n = 573) C c (n = 1097) p-value Demographics Female sex 155 (33⋅8) 212 (37⋅0) 410 (37⋅4) 0⋅44 Median age in years (IQR) 66 (53–77) 62 (50–73) 67 (54–78) <0⋅001 Comorbidities Myocardial infarction 77 (16⋅8) 96 (16⋅8) 147 (13⋅4) 0⋅09 Peripheral vascular disease 73 (15⋅9) 55 (9⋅6) 89 (8⋅1) <0⋅001 Dementia 43 (9⋅4) 19 (3⋅3) 84 (7⋅7) <0⋅001 Chronic lung disease 54 (11⋅8) 91 (15⋅9) 150 (13⋅7) 0⋅14 Leukaemia 13 (2⋅8) 22 (3⋅8) 26 (2⋅4) 0⋅23 Lymphoma 13 (2⋅8) 35 (6⋅1) 31 (2⋅8) 0⋅002 Solid cancer without metastasis 36 (7⋅9) 54 (9⋅4) 99 (9⋅0) 0⋅68 Metastatic solid cancer 16 (3⋅5) 71 (12⋅4) 96 (8⋅8) <0⋅001 Chronic renal disease 100 (21⋅8) 201 (35⋅1) 228 (20⋅8) <0⋅001 Cerebrovascular disease No cerebrovascular disease 387 (84⋅5) 492 (85⋅9) 942 (85⋅9) 0⋅91 Mild or no residual neurological defect 49 (10⋅7) 58 (10⋅1) 105 (9⋅6) Hemiplegia 22 (4⋅8) 23 (4⋅0) 50 (4⋅6) Chronic liver disease No liver disease 397 (86⋅7) 516 (90⋅1) 916 (83⋅5) 0⋅003 Mild 33 (7⋅2) 29 (5⋅1) 80 (7⋅3) Moderate or severe 28 (6⋅1) 28 (4⋅9) 101 (9⋅2) Diabetes mellitus Not diabetic 278 (60⋅7) 397 (69⋅3) 800 (72⋅9) <0⋅001 Without end-organ damage 95 (20⋅7) 92 (16⋅1) 191 (17⋅4) With end-organ damage 85 (18⋅6) 84 (14⋅7) 106 (9⋅7) HIV infection 9 (2⋅0) 6 (1⋅0) 15 (1⋅4) 0⋅45 Invasive procedures or treatment Hemodialysis 37 (8⋅1) 134 (23⋅4) 67 (6⋅1) <0⋅001 Systemic corticosteroid for more than 2 weeks 26 (5⋅7) 54 (9⋅4) 63 (5⋅7) 0⋅011 Neutropenia 2 (0⋅4) 25 (4⋅4) 35 (3⋅2) 0⋅001 Current chemotherapy 5 (1⋅1) 84 (14⋅7) 78 (7⋅1) <0⋅001 Immunesuppressive therapy 27 (5⋅9) 38 (6⋅6) 58 (5⋅3) 0⋅53 Organ or bone marrow transplantation 14 (3⋅1) 24 (4⋅2) 43 (3⋅9) 0⋅62 Previous surgery (last month) 4(0⋅9) 1 (0⋅2) 3 (0⋅3) 0⋅14 Vascular catheter at onset 77 (16⋅8) 457 (79⋅8) 298 (27⋅2) <0⋅001 Acquisition of infection Community-acquired 181 (39⋅5) 30 (5⋅2) 471 (42⋅9) <0⋅001 Healthcare-associated 152 (33⋅2) 200 (34⋅9) 274 (25⋅0) <0⋅001 Nosocomial 126 (27⋅5) 343 (59⋅9) 352 (32⋅1) <0⋅001 Main site of infection d Skin and soft tissues excluding surgical wound and deep tissue 192 (41⋅9) 17 (3⋅0) 39 (3⋅6) <0⋅001 Surgical wound 88 (19⋅2) 7 (1⋅2) 0 (0⋅0) <0⋅001 (Table 1 continues on next page) Articles 4 www.thelancet.com Vol 83 May, 2025 categorical variables, allowing the optimal number of clusters to emerge naturally without imposing any predefined assumptions. This approach included a pre-clustering step based on a log-likelihood distance measure, followed by hierarchical clustering. The optimal number of clusters was automatically determined using Schwarz’s Bayesian Information Criterion (BIC). To assess the quality and robustness of the clusters, silhouette analysis was applied. A sensitivity analysis excluding variables with more than 30% missing data was performed to validate the reliability of the identified phenotypes. The characteristics of patients across the identified phenotypes were compared using the Chi-square test for categorical variables and either ANOVA or the Kruskal–Wallis test for continuous variables, depending on whether the assumption of normality was met. Because the phenotypes were defined by the portal of entry, we thenperformedanewtwo-stepclusteranalysisfor each of the phenotypes obtained, using the same methodology described above, to identify subphenotypes within each phenotype. The characteristics of patients in each of these sub-phenotypes were also analysed. Derivation and external validation of a parsimonious predictive models for phenotypes Although assignment to phenotypes was straightforward, the high number of variables considered and the complexity in their distribution in the different subphenotypes make assigning patients to them inapplicable in clinical practice. To solve this, we developed probabilistic predictive models for sub-phenotypes assignment by performing one logistic regression analysis in the ISAC cohort per phenotype, with subphenotype as the outcome variable. Variables included in the multivariable models were those with a p-value <0⋅20 in the corresponding univariable logistic regression analyses. We used the variance inflation factor value to detect potential collinearity, and interactions were tested. The variables were selected using a manual backward selection process. The predictive ability of the final models for observed phenotypes assignment was checked by calculating their area under the receiver operating characteristic curves (AUROC) with 95% confidence intervals (CI). These sub-phenotype-assignment predictive models were applied to patients in the INSTINCT and FEN-AUREUS cohort for their external validation; because the model-derived formula used for the assignment probabilities calculation corresponds with a binary logistic regression, each patient within a phenotype was assigned to sub-phenotype 1 or 2 according to whichever had a higher probability (>50%). We then assessed the distribution of clinical variables across the assigned phenotypes and subphenotypes. Factor A a (n = 458) B b (n = 573) C c (n = 1097) p-value (Continued from previous page) Central venous catheter (including PICC) 7(1⋅5) 252 (44⋅0) 0 (0⋅0) <0⋅001 Peripheral venous catheter 11 (2⋅4) 182 (31⋅8) 0 (0⋅0) <0⋅001 Infected intravascular thrombus 10 (2⋅2) 14 (2⋅4) 11 (1⋅0) 0⋅05 Implanted vascular device 3(0⋅7) 95 (16⋅6) 44 (4⋅0) <0⋅001 Native heart valve 21 (4⋅6) 19 (3⋅3) 82 (7⋅5) 0⋅003 Prosthetic heart valve 9 (2⋅0) 4 (0⋅7) 27 (2⋅5) 0⋅04 Epidural or intraspinal empyema 5(1⋅1) 4 (0⋅7) 37 (3⋅4) <0⋅001 Vertebral bone/disc 18 (3⋅9) 7 (1⋅2) 87 (7⋅9) <0⋅001 Native joint 33 (7⋅2) 11 (1⋅9) 76 (6⋅9) <0⋅001 Prosthetic joint 13 (2⋅9) 3 (0⋅5) 28 (2⋅6) 0⋅01 Other bone-related source 42 (9⋅2) 1 (0⋅2) 20 (1⋅8) <0⋅001 Deep tissue infection or abscess 49 (10⋅7) 6 (1⋅0) 100 (9⋅1) <0⋅001 Pneumonia 28 (6⋅1) 9 (1⋅6) 180 (16⋅4) <0⋅001 Source not established 25 (5⋅5) 17 (3⋅0) 362 (33⋅0) <0⋅001 Other sources 20 (4⋅4) 27 (4⋅7) 44 (4⋅0) 0⋅73 Other infection features Previous nonbacteraemic S. aureus infection in the last 12 weeks 18 (3⋅9) 21 (3⋅7) 58 (5⋅3) 0⋅25 More than 2 days with symptoms 144 (31⋅4) 74 (12⋅9) 255 (23⋅2) <0⋅001 Median days from admission until confirmation e of S. aureus bacteremia (IQR) 0(0–5) 5 (0–13) 1 (0–6) <0⋅001 Antibiotic susceptibility Methicillin-resistant 105 (22⋅9) 94 (16⋅4) 175 (16⋅0) 0⋅003 Physical examination data Hypotension (arterial blood pressure<70 mmHG) 96 (21⋅1) 111 (19⋅4) 236 (21⋅5) 0⋅59 Inotropes use 179 (16⋅3) 59 (12⋅9) 56 (9⋅8) 0⋅001 Lowest Glasgow score, median (IQR) 15 (14–15) 15 (14–15) 15 (13–15) <0⋅001 Highest temperature (◦C), median (IQR) 38⋅3 (37⋅4–38⋅8) 38⋅5 (38⋅1–39⋅0) 38⋅2 (37⋅5–38⋅7) <0⋅001 Maximum heart rate (bpm), median (IQR) 98 (89–111) 98 (91–113) 98 (90–115) 0⋅6 Laboratory analysis Lowest O2 saturation in %, median (IQR) 94 (93–96) 95 (93–96) 94 (92–96) 0⋅001 Highest FiO2%, median (IQR) 24 (21–40) 26 (21–35) 28 (21–40) 0⋅02 Serum bilirubin (mg/dL), median (IQR) 0⋅58 (0⋅35–0⋅90) 0⋅75 (0⋅47–1⋅12) 0⋅70 (0⋅44–1⋅08) 0⋅53 Serum creatinine (mg/dL), median (IQR) 1⋅24 (0⋅85–1⋅90) 1⋅19 (0⋅80–3⋅50) 1⋅16 (0⋅79–2⋅00) 0⋅019 (Table 1 continues on next page) Articles www.thelancet.com Vol 83 May, 2025 5 Prognostic assessment of the phenotypes and subphenotypes Mortality until day 30 in patients belonging to the different phenotypes and sub-phenotypes in the derivation cohort (ISAC) was plotted using Kaplan–Meier curves and compared by the log-rank test. The association between sub-phenotypes and 30-day mortality was adjusted using Cox regression models that included management variables and centre only if they showed an association with 30-day mortality (p < 0⋅20) in the univariable analyses. These analyses were repeated in the external validation cohorts INSTINCT and FENAUREUS for the assigned phenotypes and sub-phenotypes. All analyses were performed with IBM SPSS Statistics 29, SPM 8⋅3 and R software version 4.4.1. Role of the funding source The funders of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. All authors had access to all the data. The corresponding author had final responsibility for the decision to submit for publication. Results Derivation of phenotypes The features of the 2128 patients in the ISAC cohort, used for derivation of phenotypes, was previously reported in detail. 13 The cluster analysis identified 3 clinical phenotypes based on the portal of entry: phenotype A (portal of entry: skin and soft tissues), 458 cases (21⋅5%); phenotype B (portal of entry: vascular catheter), 573 cases (26⋅9%); and phenotype C (others or unknown portal of entry), 1097 cases (51⋅6%). The silhouette score was 0⋅61, indicating good quality of clustering. The features of the patients, according to the phenotypes, are shown in Table 1. Overall, compared to patients in the other two phenotypes, those in phenotype A had fewer days of hospitalisation before developing bacteraemia, higher platelet counts, serum creatinine levels, CRP values and neutrophils count by the time of blood culture extraction, and a higher proportion of methicillinresistant isolates. In contrast, patients in phenotype B were younger and exhibited a higher frequency of cancer, neutropenia, chronic renal disease, and the presence of a vascular catheter at the onset of infection, with SAB more frequently being nosocomial. The duration of symptoms prior to blood culture was shorter than in the other phenotypes. They also had a longer history of prior hospitalisation. Patients classified in phenotype C showed a higher percentage of community-acquired SAB, liver disease, and pneumonia as a source, and lower oxygen saturation. In each of these phenotypes, the two-step cluster analysis was repeated, which allowed identifying two sub-phenotypes with each of them: sub-phenotypes A1 and A2; B1 and B2; C1 and C2. Sub-phenotypes differed in acquisition type, comorbidities, bacteraemia source, laboratory data, and methicillin-resistance (Supplementary Tables S3–S5). In phenotype A, subphenotype A1 showed a moderately severe profile, with fewer comorbidities but higher CRP levels. In contrast, A2 included higher rates of renal disease, diabetes mellitus with end-organ damage, methicillinresistance, the highest platelet and creatinine levels, and the lowest lymphocyte count. Within phenotype B, sub-phenotype B1 exhibited moderate complexity, characterized by a lower prevalence of nosocomial infections, the highest oxygen saturation, and lowest white blood cell and neutrophil counts. Meanwhile, B2 was more severe, with a high rate of nosocomial infection, infections associated with central venous catheters, and higher methicillin resistance. Finally, among patients in phenotype C, sub-phenotype C1 was characterised as mild, with few chronic conditions, whereas C2 included older patients with multiple age-related comorbidities and the highest percentage of unknown foci (Supplementary Figure S1). 30-Day mortality in phenotypes and subphenotypes of the derivation cohort (ISAC) The three phenotypes were associated with different 30day mortality rates: 18⋅1% (83/458), 13⋅1% (75/573) and 25⋅3% (277/1097) for phenotypes A, B and C, respectively (log-rank test p < 0⋅001) (Fig. 2A, Table 2). The sub-phenotypes also exhibited significant differences in 30-day mortality within each phenotype: A1, 15⋅7% (58/ 370) vs A2, 28⋅4% (25/88), p = 0⋅003; B1, 11⋅0% (55/498) vs B2, 26⋅7% (20/75), p = 0⋅001; and C1, 15⋅1% (61/403) vs C2, 31⋅1 (216/694), p < 0⋅001 (Fig. 2B, C, and D, Table 2). Factor A a (n = 458) B b (n = 573) C c (n = 1097) p-value (Continued from previous page) Serum CRP (mg/L), median (IQR) 156⋅0 (54⋅5–231⋅3) 103⋅0 (36⋅0–163⋅0) 118⋅0 (37⋅0–191⋅0) <0⋅001 Blood white cell count (×10 3 /μL), median (IQR) 13⋅0(9⋅0–16⋅9) 10⋅9(7⋅3–14⋅5) 12⋅8(8⋅6–16⋅8) <0⋅001 Neutrophil count (×10 3 /μL), median (IQR) 10⋅8(7⋅5–14⋅8) 9⋅3(6⋅0–12⋅6) 10⋅7(7⋅1–14⋅5) <0⋅001 Lymphocyte count (×10 3 /μL), median (IQR) 1⋅2(0⋅7–1⋅7) 0⋅85 (0⋅45–1⋅25) 0⋅95 (0⋅55–1⋅35) <0⋅001 Platelet count (×10 3 / μL), median (IQR) 220 (142–298) 185 (123–252) 188 (114–256) <0⋅001 Data are expressed as No. (%) except where specified. a Portals of entry: skin or soft tissue infection unrelated to surgical wounds: 324 cases (70⋅7%); surgical wounds: 134 cases (29⋅3%). b Portals of entry: central vascular catheter (including PICC): 271 cases (47⋅3%); peripheral vascular catheter (including arterial line): 206 cases (35⋅9%); other implanted vascular device (e.g., pacemaker, stent, graft) with early-onset infections following device implantation: 96 cases (16⋅8%). c Portals of entry: respiratory tract: 198 cases (18⋅8%); Genito-urinary tract: 123 cases (11⋅2%); Portal not known: 776 cases (70⋅7%). d More than one allowed. e Confirmation refers to the moment when blood culture results become available. Table 1: Features and outcomes of patients in the derivation cohort and in the phenotypes obtained. Articles 6 www.thelancet.com Vol 83 May, 2025 Fig. 2: Kaplan–Meier survival analysis: 30-day mortality by phenotype and sub-phenotype in the derivation cohort. A) Kaplan–Meier survival analysis: 30-day mortality by phenotype based on the likely portal of entry in the derivation cohort. B) Kaplan–Meier Survival Analysis: 30-Day mortality by sub-phenotype in patients belonging to the A phenotype in the derivation cohort. C) Kaplan–Meier Survival Analysis: 30-Day mortality by sub-phenotype in patients belonging to the B phenotype in the derivation cohort. D) Kaplan–Meier Survival Analysis: 30-Day mortality by sub-phenotype in patients belonging to the C phenotype in the derivation cohort. Articles www.thelancet.com Vol 83 May, 2025 7 The association of A2, B2 and C2 with higher 30-day mortality was confirmed in multivariable Cox regression analyses conducted after adjusting for clinical management variables such as source control, infectious diseases consultation and antimicrobial treatment, as well as for centre effect. The adjusted hazard ratios (aHR) for mortality were 1⋅86 (95% CI 1⋅16–2⋅99, p = 0⋅01) for A2 respect to A1; 2⋅50 (95% CI 1⋅47–4⋅26, p = 0⋅0007) for B2 respect to B1; and 1⋅97 (95% CI 1⋅47–2⋅62, p<0⋅0001) for C2 respect to C1 (Table 3). Additionally, consultation by an infectious diseases specialist and source control within day 3 were also protective for mortality across all three phenotypes. Predictive models for sub-phenotypes assignment To facilitate the clinical application of phenotype and sub-phenotype identification, we performed a multivariable logistic regression analysis within each phenotype using the sub-phenotype as dependent variable. The final multivariable logistic regression models are detailed in Table 3 and included 7 variables for classifying patients in sub-phenotypes A1 or A2, 5 for B1 or B2, and 14 for C1 or C2. These models exhibited very good predictive ability, with AUROCs (95% CI) of 0⋅86 (0⋅82–0⋅91), 0⋅88 (0⋅83–0⋅92), and 0⋅89 (0⋅86–0⋅91), respectively (Supplementary Table S6). An online, freely available application to assign SAB patients to sub-phenotypes based on the developed probabilistic models is accessible at https://fen-aureus.com/. Validation in external cohorts We replicated the two-phase cluster analysis in the INSTINCT validation cohort, confirming the presence of three phenotypes, each subdivided into two subphenotypes. The phenotypes also clustered based mostly on the portal of entry: the first included 355 cases, with 80⋅5% having skin and soft tissues as the portal of entry; the second comprised 432 cases, 99⋅8% of which had a vascular catheter as the portal of entry; and the third phenotype included 430 cases, with 95⋅1% associated with an unknown or other portal of entry (Supplementary Table S7). Phenotype Mortality (n/N; %) Sub-phenotype Mortality (n/N; %) p (log-rank test) ISAC cohort A 83/458 (18⋅1) A1 58/370 (15⋅7) 0⋅003 A2 25/88 (28⋅4) B 75/573 (13⋅1) B1 55/498 (11⋅0) 0⋅001 B2 20/75 (26⋅7) C 277/1097 (25⋅3) C1 61/403 (15⋅1) <0⋅0001 C2 216/694 (31⋅1) INSTINCT cohort A 37/243 (15⋅2) A1 20/166 (12⋅0) 0⋅04 A2 17/77 (22⋅1) B 61/458 (13⋅3) B1 38/383 (9⋅9) <0⋅0001 B2 23/75 (30⋅7) C 139/516 (26⋅9) C1 21/165 (12⋅7) <0⋅0001 C2 118/351 (33⋅6) FEN-AUREUS cohort A 69/283 (24⋅4) A1 31/181 (17⋅1) 0⋅0002 A2 38/102 (37⋅3) B 37/573 (16⋅9) B1 76/505 (15⋅0) 0⋅0007 B2 21/68 (30⋅9) C 94/329 (28⋅6) C1 19/123 (15⋅4) <0⋅0001 C2 75/206 (36⋅4) Table 2: 30-Day mortality rates by phenotypes and sub-phenotypes in the ISAC derivation cohort and in the INSTINCT and FEN-AUREUS external validation cohorts. Variable A phenotype B phenotype C phenotype aHR (95% CI) p aHR (95% CI) p aHR (95% CI) p ID consultation 0⋅51 (0⋅31–0⋅84) 0⋅01 0⋅50 (0⋅30–0⋅84) 0⋅009 0⋅35 (0⋅27–0⋅44) <0⋅0001 Source control within 3 days 0⋅39 (0⋅18–0⋅84) 0⋅02 0⋅45 (0⋅28–0⋅71) 0⋅001 0⋅34 (0⋅20–0⋅58) <0⋅0001 Sub-phenotype 2 a 1⋅86 (1⋅16–2⋅99) 0⋅01 2⋅50 (1⋅47–4⋅26) 0⋅0007 1⋅97 (1⋅47–2⋅62) <0⋅0001 High risk centre b 1⋅90 (1⋅10–3⋅28) 0⋅02 1⋅34 (1⋅00–1⋅80) 0⋅05 a A2 sub-phenotype with respect to A1 in A phenotype; B2 sub-phenotype with respect to B1 in B phenotype; and C2 sub-phenotype with respect to C1 in C phenotype. b In Phenotype A, “high-risk centre”was not significantly associated with mortality in the final model. Table 3: Multivariable cox regression of 30-day mortality for each of the identified phenotypes in the derivation cohort. Articles 8 www.thelancet.com Vol 83 May, 2025 Phenotypes, sub-phenotypes, and the corresponding probabilistic models were exclusively derived from the derivation cohort (ISAC) and then directly applied to the external validation cohorts INSTINCT and FENAUREUS. For phenotype assignment, patients were classified based on their portal of entry. Subsequently, sub-phenotypes were assigned according to the proposed algorithm (Fig. 3). The resulting distribution in the INSTINCT cohort of phenotypes A, B and C was 243 (20⋅0%), 458 (37⋅6%), and 516 patients (42⋅4%), respectively; and in the FEN-AUREUS cohort: 283 (23⋅9%), 573 (48⋅3%) and 329 (27⋅8%), respectively (Supplementary Tables S8 and S9). 30-Day mortality in the phenotypes and subphenotypes of the external validation cohorts In the external validation cohorts, INSTINCT and FENAUREUS, the three phenotypes identified significant differences in 30-day mortality rates: 15⋅2% and 24⋅4% for phenotype A; 13⋅3% and 16⋅9% for phenotype B; and 26⋅9% and 28⋅6% for phenotype C, respectively (Table 2 and Supplementary Figures S2 and S3). Furthermore, significant differences in 30-day mortality were also noted among the sub-phenotypes within each of the three phenotypes in both cohorts (Table 2), confirming the results of the derivation cohort. The aHRs for mortality were: 1⋅93 (95% CI: 1⋅01–3⋅71, p=0⋅04) for sub-phenotype A2 vs A1; 3⋅40 (95% CI: 2⋅02–5⋅72, p < 0⋅0001) for B2 vs B1; and 3⋅04 (95% CI: 1⋅84–5⋅02, p < 0⋅0001) for C2 vs C1 in the INSTINCT cohort (Supplementary Table S10). In the FENAUREUS cohort, the corresponding aHRs were: 2⋅02 (95% CI: 1⋅25–3⋅29, p = 0⋅004); 2⋅11 (95% CI: 1⋅30–3⋅44, p=0⋅002) and 2⋅44 (95% CI: 1⋅47–4⋅06) (Supplementary Table S11). Discussion SAB is a condition characterised by significant heterogeneity, making the establishment of standardised management protocols highly complex. 7,8 In this study, we identified three phenotypes of SAB based on probable portals of entry and six sub-phenotypes (two within each phenotype), taking into account baseline patient and infection characteristics, including demographics, comorbidities, acquisition type, site of infection, antibiotic susceptibility, and clinical and laboratory data from an international cohort. Both phenotypes and subphenotypes were associated with 30-day mortality. These phenotypes could provide an explanation for this heterogeneity and support the identification and management of these patients. To our knowledge, only one previous study 15 has aimed to identify clinical subphenotypes in SAB patients, using latent class analysis in both observational and trial cohorts. That study confirmed the existence of distinct, prognostically relevant groups. However, key methodological differences limit comparability. Swets et al. selected variables based on expert consensus and modelled class structure using both statistical and clinical criteria, whereas we applied an unsupervised, fully data-driven approach without preselection of variables. Their analyses focused primarily on trial populations, while our study was grounded in large, real-world cohorts with standardised data collection. Additionally, while Swets et al. suggest the possibility of differential treatment responses, our focus was on early clinical risk stratification, externally validated and operationalised through an open-access tool for routine application. A recent study by Russell et al. 16 highlighted distinct clinical trajectories in SAB, including early death, metastatic infection, and endocarditis. Although not based Confirmaon of S. aureus bacteraemia (Days 1-2 aer blood culture) A (skin and so ssues) Probabilisc model to assign subphenotypes within phenotype A** Subphenotype A1 Subphenotype A2 B (vascular catheter) Probabilisc model to assign subphenotypes within phenotype B** Subphenotype B1 Subphenotype B2 C (other/unknown) Probabilisc model to assign subphenotypes within phenotype C** Subphenotype C1 Subphenotype C2 Portal of entry* Fig. 3: Proposed clinical management algorithm for the classification of patients with S. aureus bacteraemia based on identified phenotypes and sub-phenotypes. *See definitions to consider a probable portal of entry in Supplementary material. **Visit webpage: https:// fen-aureus.com/ for phenotypes and subphenotypes assignment. Articles www.thelancet.com Vol 83 May, 2025 9