Development and Validation of 72-Hour Fluid Balance Trajectory Subphenotypes and Prognosis in Sepsis: A Multicenter Cohort Study

Development and Validation of 72-Hour Fluid Balance Trajectory Subphenotypes and Prognosis in Sepsis: A Multicenter Cohort Study | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Development and Validation of 72-Hour Fluid Balance Trajectory Subphenotypes and Prognosis in Sepsis: A Multicenter Cohort Study kangxing wang, Huaiyu Xiong, Yukun Zhu, Yongfang Zhou, Yan Kang This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8247018/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 26 Mar, 2026 Read the published version in Scientific Reports → Version 1 posted 14 You are reading this latest preprint version Abstract Background Fluid management in sepsis is controversial, and while persistent positive fluid balance is linked to adverse outcomes, the prognostic value of the temporal patterns (trajectories) of cumulative fluid balance (CFB) during the first 72 hours after sepsis diagnosis remains unclear. This study aimed to identify distinct subphenotypes based on 72-hour CFB trajectories in adult sepsis patients and evaluate their independent association with 28-day mortality. Methods This multicenter retrospective cohort study utilized three large-scale critical care databases (MIMIC-IV, MIMIC-III-CareVue, and eICU). We included adult patients meeting Sepsis-3 criteria with an ICU stay of over 72 hours. Group-Based Trajectory Modeling (GBTM) identified five distinct 72-hour CFB (mL/kg) subphenotypes in the MIMIC-IV development cohort (n = 16,069). Trajectories were validated in the MIMIC-III (n = 2,162) and eICU (n = 13,805) cohorts. We used Kaplan-Meier analysis and multivariable Cox proportional hazards models, including Inverse Probability of Treatment Weighting (IPTW), to assess the association with 28-day mortality, adjusting for baseline confounders. Results GBTM identified five reproducible CFB subphenotypes, notably a "Persistent Negative Balance" (Class 1), a "High, Rapid Decline" (Class 4), and a "Persistently High Balance" (Class 5). Class 5 patients exhibited the highest illness severity (e.g., highest SOFA scores). Kaplan-Meier analysis showed significant differences in 28-day survival (P < 0.001), with Class 1 and Class 4 having the best survival and Class 5 the worst. In the MIMIC-IV cohort, compared to the highest-risk Class 5 (reference) after full multivariable adjustment, all other trajectories were associated with significantly lower 28-day mortality: Class 1 (HR: 0.50, 95% CI: 0.44–0.56) and Class 4 (HR: 0.51, 95% CI: 0.46–0.57). These protective findings were consistent across all validation cohorts. Conclusion Sepsis patients exhibit five distinct, reproducible CFB trajectories strongly associated with 28-day mortality, independent of baseline severity. The finding that the "High, Rapid Decline" (Class 4) trajectory shares an excellent prognosis with the "Persistent Negative" (Class 1) trajectory challenges static fluid assessments. The ability to predict the high-risk Class 5 phenotype early using machine learning (AUC ≈ 0.81) provides a tangible pathway toward individualized fluid stewardship. Health sciences/Biomarkers Health sciences/Diseases Health sciences/Health care Health sciences/Medical research Health sciences/Risk factors Sepsis Fluid Therapy Group-Based Trajectory Modeling (GBTM) Subphenotypes Figures Figure 1 Figure 2 Figure 3 Figure 4 Background Sepsis, defined as life-threatening organ dysfunction caused by a dysregulated host response to infection, represents a major global health challenge [1,2]. Before the COVID-19 pandemic, sepsis affected nearly 50 million people worldwide annually and continues to be associated with a high risk of death, killing between one in three and one in six of those affected [3,4]. In the United States alone, sepsis is implicated in over one-third of in-hospital deaths, making it the most common cause of in-hospital mortality and the most expensive cause of hospitalization [5]. Despite advances in medical care, mortality from sepsis remains high, highlighting significant knowledge gaps and the persistent need for research to improve patient outcomes [2,6]. Early fluid resuscitation is considered a cornerstone of sepsis management [6–9]. Initial strategies, such as Early Goal-Directed Therapy (EGDT), showed early promise in improving outcomes [10]. However, subsequent large-scale randomized controlled trials (RCTs), namely the ProCESS, ARISE, and ProMISe studies, failed to demonstrate the superiority of EGDT over usual care in terms of survival [11–13]. These findings challenged the notion of "more is better" or the strict adherence to specific resuscitation protocols. Consequently, despite the critical importance of fluid resuscitation, the optimal fluid management strategy—encompassing the initial volume, the timing and methods for subsequent fluid administration, and appropriate resuscitation targets—remains a central controversy and a key unresolved issue in sepsis research and clinical practice [2,6]. Indeed, "how to individualize fluid resuscitation initially and beyond" has been identified as a top clinical research priority in the field of sepsis [6]. Fluid therapy in sepsis presents a "double-edged sword": while essential, excessive administration leading to fluid overload or persistent positive fluid balance is linked to increased mortality, acute kidney injury, and prolonged ventilation [14,15]. This underscores the need for precise fluid stewardship beyond initial resuscitation. Given the substantial heterogeneity among sepsis patients, individualized strategies that move beyond fixed protocols are increasingly sought. Therapeutic approaches ideally should be tailored to the patient's physiological state and the specific phase of resuscitation, as conceptualized by models like the Resuscitation-Optimization-Stabilization-Evacuation (ROSE) framework [16]. However, the limitations inherent in static fluid assessments and snapshot phenotyping approaches highlight a critical gap: the prognostic significance of temporal patterns in fluid balance during the initial 72 hours of sepsis remains largely unexplored [16,17]. It is unclear whether classifying patients based on these early fluid trajectories provides additional prognostic information beyond traditional risk stratification tools, such baseline severity scores or static cumulative fluid balance assessments at fixed time points. We hypothesized that identifying distinct subgroups of sepsis patients based on their 72-hour cumulative fluid balance trajectories, using Group-Based Trajectory Modeling (GBTM) [18], would reveal clinically meaningful phenotypes associated with significantly different risks for adverse outcomes, particularly 28-day mortality. Therefore, this study aimed to delineate these early fluid trajectories in large, multicenter cohorts of sepsis patients and evaluate their independent association with prognosis. Methods Data sources This multi-center retrospective cohort study utilized data from three large-scale, publicly available critical care databases. The development cohort, used for identifying fluid trajectory subphenotypes, was derived from the Medical Information Mart for Intensive Care IV (MIMIC-IV) database (version 3.1), representing patients admitted to a single academic medical center (Beth Israel Deaconess Medical Center, BIDMC) between 2008 and 2022 [19]. To assess temporal robustness, the MIMIC-III CareVue subset (v1.4), containing data from the same institution during an earlier period (2001–2008), served as the temporal validation cohort [20]. Generalizability was assessed using the eICU Collaborative Research Database (eICU-CRD) (version 2.0) as the multi-center validation cohort, which includes data from over 200 hospitals across the United States (2014–2015) [21]. Data access was granted and data were extracted by an authorized researcher (Kangxing Wang, certification number: 13024213) after obtaining necessary institutional approvals and completing required ethical training. Study population We included adult patients (age > 18 years) admitted to the ICU who met the Sepsis-3 criteria, defined as suspected infection (indicated by concurrent administration of antibiotics and sampling of body fluids) combined with an acute increase in the Sequential Organ Failure Assessment (SOFA) score of ≥ 2 points [1]. Only the first ICU admission was included for patients with multiple admissions during a single hospitalization. The exclusion criteria were: (1) ICU length of stay < 72 hours; (2) multiple ICU admissions (only the first was retained, as mentioned); and (3) patients with fewer than two recorded fluid balance data points during the first 72 hours. The detailed patient selection process for all three databases is illustrated in Fig. 1 . Outcome Outcome The primary outcome was all-cause mortality within 28 days following ICU admission. The key secondary outcomes were the duration of mechanical ventilation (in hours) and the utilization of renal replacement therapy (RRT) during the ICU stay. Feature extraction Data extraction was performed using structured query language (SQL) in PostgreSQL (version 8.2). For each patient, a comprehensive set of baseline variables was obtained, primarily reflecting their clinical status within the first 24 hours after ICU admission unless otherwise specified. The extracted variables included demographics (age, sex, race); ICU type; severity of illness scores (Sequential Organ Failure Assessment [SOFA], Oxford Acute Severity of Illness Score [OASIS], Glasgow Coma Scale [GCS], and Charlson Comorbidity Index [CCI]); comorbidities identified by ICD-9/10 codes (hypertension, diabetes mellitus, chronic obstructive pulmonary disease [COPD], heart failure [HF], and stroke); infection site (lung, gastrointestinal, genitourinary, or other) [22]; the most abnormal vital signs within 24 hours (minimum mean arterial pressure, maximum heart rate, maximum respiratory rate, and maximum temperature); initial laboratory values (first recorded hemoglobin, white blood cell count, platelets, creatinine, lactate, PaO₂, PaCO₂, pH, potassium, sodium, chloride, and calcium); and interventions within 24 hours (mechanical ventilation, renal replacement therapy [RRT], and vasopressor use). Fluid balance trajectory construction The primary fluid balance trajectory variable was derived as follows: hourly net fluid balance data were extracted for the first 72 hours following the diagnosis of sepsis and aggregated into 24 consecutive 3-hour intervals. For each interval, the net fluid balance was normalized by the patient’s admission body weight (kg) to yield a value in mL/kg. These 3-hourly, weight-normalized net fluid balance values were winsorized at the 1st and 99th percentiles—based on the development cohort (MIMIC-IV)—to reduce the influence of extreme outliers. Statistical analysis Baseline characteristics were summarized as mean ± standard deviation (SD) or median [interquartile range (IQR)] for continuous variables, and as counts (percentages) for categorical variables. Comparisons were made using the Student’s t-test, Mann-Whitney U test, or Chi-square test, as appropriate. We assessed variables for missingness, and those with < 30% missing data were imputed using multiple imputation by chained equations (MICE). We utilized Group-Based Trajectory Modeling (GBTM) to identify distinct patient subphenotypes based on their cumulative fluid balance trajectories during the first 72 hours after sepsis diagnosis. To determine the optimal number of trajectory classes, we fitted models with a varying number of classes (e.g., 2–7) in the development cohort. The final model selection was guided by a combination of statistical criteria and clinical interpretability, including the Bayesian Information Criterion (BIC), average posterior probability (APP > 0.90), relative class size (> 10%), and the clinical plausibility of the trajectories. For external validation, the trained model was applied to the validation cohorts. Each patient was assigned to the trajectory class that yielded the lowest Mean Squared Error (MSE) when fitting their individual data to the class-specific polynomial functions [23]. Kaplan-Meier curves with the log-rank test were used for unadjusted survival comparisons. We employed multivariable Cox proportional hazards regression to evaluate the independent association between fluid trajectories and outcomes. Three sequential models were constructed: Model 1 (Crude Model), which included only the trajectory class as the predictor; Model 2 (Demographics-Adjusted Model), which adjusted for trajectory class, age, gender, and race; and Model 3 (Fully-Adjusted Model), which adjusted for all variables in Model 2 plus ICU type, severity of illness scores (SOFA, OASIS, GCS, Charlson Comorbidity Index), infection site (lung, GI, GU, other), baseline comorbidities (hypertension, diabetes, COPD, heart failure, stroke), 24-hour vital signs (minimum MAP, maximum heart rate, maximum respiratory rate, maximum temperature), initial laboratory values (hemoglobin, WBC, platelet, creatinine, lactate, PO2, PCO2, pH, potassium, sodium, chloride, calcium), and 24-hour interventions (mechanical ventilation, RRT, and vasopressor use). Several sensitivity analyses were performed to test the robustness of our findings. First, to further control for confounding, we applied an Inverse Probability of Treatment Weighting (IPTW) analysis. Second, to test the robustness of our missing data imputation method, we repeated the analysis utilizing random forest imputation for missing data instead of MICE. All analyses were conducted using R, version 4.2.2. Results Study Cohort and Trajectory Identification Based on the Group-Based Trajectory Modeling (GBTM) analysis, a five-class model was determined to be the optimal fit for the 72-hour cumulative fluid balance data. This model was selected based on the lowest Bayesian Information Criterion (BIC = 2317109), excellent average posterior probabilities (APP > 0.90 for all classes), and all class proportions being greater than 12%. The detailed model fit statistics are available in Supplementary Table S1-S2. The five distinct fluid trajectories are visualized in Fig. 2 A and were clinically characterized as follows: Class 1 (Blue line): "Persistent Negative Balance" (n = 3,772, 23.5%), which started with a low-positive balance and consistently trended downwards, ending near or below a 0 mL/kg cumulative balance by 72 hours; Class 3 (Yellow line): "Low-Stable Positive Balance" (n = 2,538, 15.8%), which maintained a stable, low-level positive fluid balance throughout the 72-hour period; Class 2 (Orange line): "Medium-Stable Positive Balance" (n = 3,372, 21.0%), which maintained a stable, medium-level positive fluid balance throughout the 72-hour period; Class 4 (Green line): "High, Rapid Decline" (n = 4,417, 27.5%), the largest group, characterized by a high initial cumulative balance that rapidly and steadily declined over 72 hours; and Class 5 (Red line): "Persistently High Balance" (n = 1,970, 12.3%), which started with the highest cumulative balance and maintained this persistently high positive balance with little to no decline over the 72-hour period. Critically, these five distinct trajectory patterns were consistently identified in both the MIMIC-III-CareVue (temporal validation) and eICU (external validation) cohorts, as shown in Fig. 2 B and 2 C. Clinical Characterization of Subphenotypes The baseline characteristics of the five subphenotypes in the MIMIC-IV cohort are presented in Table 1. A critical comparison emerges between Class 5 ("Persistently High Balance") and Class 4 ("High, Rapid Decline"), which represented the two most acutely ill profiles at baseline. Patients in Class 5 demonstrated the highest severity, with the highest median SOFA score (9.00), OASIS score (40.00), initial lactate (2.40 mmol/L), and the highest rates of 24-hour vasopressor use (70.91%) and mechanical ventilation (72.34%). In contrast, while Class 4 patients also presented with high initial illness severity (SOFA 7.00, Vaso 64.89%), they were distinguished by near-normal renal function (Creatinine 1.00 mg/dL) and a minimal 24-hour RRT rate (1.47%). This was in stark opposition to Class 5, which had significantly higher creatinine (1.40 mg/dL) and a dramatically higher RRT rate of 17.01%. The other three trajectories, Class 1 ("Persistent Negative"), Class 2 ("Medium-Stable"), and Class 3 ("Low-Stable"), generally represented patients with lower baseline severity scores. These static baseline differences were mirrored in the 72-hour dynamic physiological trends. As shown in Supplementary Figure S1, patients in Class 5 exhibited persistent shock, maintaining the lowest mean arterial pressure (MAP) and the highest heart rate over the 72-hour period. Furthermore, the 72-hour SOFA score trajectories (Supplementary Figure S1) revealed that Class 5 patients not only started with the highest organ dysfunction but also experienced a progressive worsening or failure to improve, particularly in the cardiovascular and renal sub-scores. This aligns perfectly with their high RRT rate and persistently high fluid balance observed in Table 1. Association with Primary Outcome Unadjusted analyses demonstrated a strong association between fluid trajectories and 28-day mortality. Class 5 ("Persistently High Balance") had the highest mortality rate (38.58%), while Class 1 ("Persistent Negative") and Class 4 ("High, Rapid Decline") had the lowest (16.54% and 17.70%, respectively) (Table 1). Kaplan-Meier survival analysis confirmed this finding, showing a significant separation in survival probabilities across all five groups (Log-rank p < 0.001) (Fig. 2 D). This clear prognostic stratification was robustly replicated in both the MIMIC-III-CareVue (Fig. 2 E) and eICU external validation cohorts (Fig. 2 F). In the multivariable Cox proportional hazards analysis (Table 2), using the highest-risk Class 5 as the reference, all other trajectories were associated with significantly lower mortality in the fully-adjusted Model 3. Specifically, in the MIMIC-IV cohort, the adjusted hazard ratios (aHR) were 0.50 (95% CI 0.44–0.56) for Class 1, 0.51 (95% CI 0.46–0.57) for Class 4, 0.73 (95% CI 0.64–0.83) for Class 3, and 0.87 (95% CI 0.78–0.96) for Class 2. These findings were confirmed in the Inverse Probability of Treatment Weighting (IPTW) model, which controlled for baseline confounding. This independent association remained highly consistent across the external validation cohorts, particularly for the protective associations seen in Class 1 and Class 4. Association with Secondary Outcomes The prognostic association of the fluid trajectories extended to secondary ICU outcomes (Supplementary Figure S2). The association with mechanical ventilation was highly robust across all cohorts: in the MIMIC-IV development cohort, Class 5 ("High-Persist") had a significantly higher ventilation burden (both hours and proportion) compared to all other groups (Kruskal-Wallis p < 0.001; all post-hoc p.adj < 0.001). This finding was robustly validated in both the eICU and MIMIC-III cohorts, which showed a nearly identical pattern of significant differences. In contrast, the association with CRRT utilization was strong in the development cohort (Class 5 vs. all others, p.adj < 0.001), partially validated in the eICU cohort (where Class 1 showed significantly lower use than other groups), and not observed in the MIMIC-III temporal validation cohort (Kruskal-Wallis p = 0.78). Subgroup Analyses To assess the consistency of the primary outcome, we performed subgroup analyses on the association between fluid trajectories and 28-day mortality, using Class 5 as the reference (Fig. 3 ). The protective effect of the other trajectories (especially Class 1 and Class 4) was consistent across most subgroups, including age (=65 years, P for interaction = 0.531) and baseline SOFA score ( 5, P for interaction = 0.755). However, we observed significant interactions for patients with baseline heart failure (HF, P for interaction = 0.001), stroke (P for interaction = 0.019), and vasopressor use (Vaso, P for interaction = 0.001). Specifically, in patients with pre-existing stroke, the protective effects of Class 1, Class 2, and Class 3 were attenuated and lost statistical significance (p = 0.06, 0.32, and 0.32, respectively). Furthermore, the protective associations of Class 1 and Class 4 appeared to be even stronger in patients with HF or those receiving vasopressors, while the effects of Class 2 and 3 were attenuated in these same high-risk groups. Early Prediction of the High-Risk (Class 5) Trajectory Finally, to explore the feasibility of early identification, we developed and validated several machine learning models using baseline data to predict patient membership in the high-risk Class 5 ("Persistently High Balance") trajectory. Among the six models tested, the XGBoost model demonstrated the best and most robust performance (Fig. 4 A). It achieved a high Area Under the Receiver Operating Characteristic (AUC) of 0.811 in the MIMIC-IV development (internal validation) cohort. Critically, this high predictive performance was well-sustained during external validation, achieving an AUC of 0.793 in the multi-center eICU cohort and 0.809 in the MIMIC-III-CareVue temporal validation cohort. Model interpretability using SHAP analysis (Fig. 4 B) identified baseline severity scores (OASIS, SOFA), Ca, Infection, and Lactate as the most important features for predicting this high-risk fluid trajectory. As shown in the SHAP plot, high initial values of OASIS, SOFA, and Lactate, and a lower level of Ca, were all strongly associated with an increased likelihood of being classified into Class 5. Sensitivity Analyses Two key sensitivity analyses were conducted to confirm the robustness of our primary findings. First, as shown in the primary analysis (Table 2), the associations remained consistent after applying an Inverse Probability of Treatment Weighting (IPTW) model to account for baseline confounding. Second, to test the robustness of our missing data imputation method, we repeated the entire analysis using Random Forest imputation instead of MICE. The results were nearly identical to our primary findings (Supplementary Table S3); for example, in the MIMIC-IV Model 3, the aHR for Class 1 vs. Class 5 was 0.49 (95% CI 0.43–0.56), compared to 0.50 (95% CI 0.44–0.56) in the main analysis. Discussion In this large, multicenter retrospective cohort study utilizing three distinct critical care databases (MIMIC-IV, eICU, and MIMIC-III-CareVue), we successfully identified and validated five distinct and reproducible subphenotypes based on 72-hour cumulative fluid balance trajectories following sepsis diagnosis. Our primary finding is that these dynamic trajectories are a robust and independent predictor of 28-day mortality, providing prognostic information beyond traditional static baseline severity scores. As hypothesized, the Class 5 ("Persistently High Balance") trajectory was associated with the worst outcomes, including the highest mortality, longest duration of mechanical ventilation, and greatest CRRT utilization. However, the central and most significant finding of this study is that the Class 4 ("High, Rapid Decline") trajectory—a pattern perfectly aligning with the successful "Evacuation" (de-resuscitation) phase of the ROSE framework—was associated with a favorable prognosis nearly identical to that of the Class 1 ("Persistent Negative Balance") trajectory. This observation, which remained robust after rigorous adjustment for baseline confounding using an IPTW model, strongly suggests that the dynamic process of successfully reversing a high positive fluid balance may be as prognostically protective as a primary state of persistent negative balance. A central finding of this study is the stark contrast in outcomes between Class 5 and Class 4. Admittedly, the Class 5 subphenotype was characterized by more severe baseline renal impairment, as evidenced by higher creatinine and a 17.01% 24-hour RRT rate in Table 1. However, a critical finding is that even after rigorously adjusting for these baseline confounders (including renal function and SOFA score) using an IPTW model, the Class 5 trajectory remained independently associated with significantly higher mortality compared to Class 4. This strongly implies that the trajectory itself—the process of persistent high fluid balance—carries a prognostic significance independent of the baseline state. The 72-hour dynamic data in Supplementary Figure S3 provides a clear physiological explanation for this persistent risk. The vital sign trajectories demonstrate that Class 5 patients remained in a state of persistent shock (lowest MAP, highest HR) that failed to reverse. Moreover, their SOFA sub-score trajectories (Supplementary Figure S4) revealed progressive or non-improving organ dysfunction, particularly in the Cardiovascular (Panel B) and Renal (Panel E) scores. Therefore, the Class 5 trajectory likely represents a phenotype of "refractory shock and failed de-resuscitation," where persistent fluid administration is both a consequence of, and a contributor to, a vicious cycle of non-reversing organ failure, such as AKI [24–26]. In contrast, Class 4, whose dynamic SOFA scores stabilized or improved, represents a "successful resuscitation" cohort with the capacity to be de-resuscitated—a capability increasingly recognized as a critical determinant of survival in sepsis [27] . A key limitation of trajectory analysis is that it is a descriptive, post-hoc classification. To bridge this gap for clinical practice, we demonstrated that the high-risk Class 5 trajectory can be accurately predicted within the first 24 hours. While many machine learning models have been developed to predict sepsis mortality [28], our focus on predicting a dynamic trajectory aligns with the strategic move toward identifying actionable subphenotypes [29]. The SHAP analysis revealed that this prediction was driven by intuitive clinical features. The importance of features like OASIS, SOFA, and Lactate is well-documented in prognostication, as exemplified by the Sepsis-3 definitions [1], but the inclusion of Ca (Calcium) is notable, as hypocalcemia is increasingly recognized as a marker of severity and poor prognosis in sepsis [30]. This model provides a proof-of-concept for a crucial clinical tool: an early warning system, a type of intervention that has shown promise for improving sepsis outcomes by facilitating earlier intervention [31]. Our subgroup analyses confirmed the robustness of the primary findings across most key subgroups, including age and baseline SOFA score (P for interaction > 0.05). However, we observed several clinically significant interactions (Fig. 4 ). Most notably, in patients with baseline heart failure (HF, P-interaction = 0.001), the protective effects of all other trajectories (Classes 1–4) relative to Class 5 were significantly magnified compared to non-HF patients. This suggests that for patients with compromised cardiac function, the survival benefit of avoiding the "Persistently High" (Class 5) trajectory is even more pronounced, adding critical nuance to the controversy surrounding the 30mL/kg bolus in this population [6,32,33]. Furthermore, in patients with stroke (P-interaction = 0.019), only the Class 4 ("High, Rapid Decline") trajectory remained significantly protective (p < 0.05), while the benefits of other trajectories were attenuated. This may reflect the clinical "tightrope walk" of neurocritical care, where an "aggressive-followed-by-rapid-removal" strategy (Class 4) may represent the optimal balance between cerebral perfusion and avoiding cerebral edema [34]. An important and honest finding of our study was the heterogeneity in secondary outcomes. While the association with mechanical ventilation was remarkably consistent across all three databases, the association with CRRT was strong in MIMIC-IV, partially validated in eICU, and absent in the MIMIC-III cohort (p = 0.78). We believe this does not invalidate the finding, but rather reflects a critical real-world variance in clinical practice over time. The MIMIC-III cohort (2001–2008) represents an older era of critical care. During that period, the optimal timing of RRT initiation was a major debate, with many observational studies and meta-analyses suggesting a potential survival benefit for "early" RRT [35]. However, several landmark randomized controlled trials (RCTs) published after this period, most notably the AKIKI trial (2016) [36] and the STARRT-AKI trial (2020) [37], failed to show a survival benefit for an "accelerated" or "early" strategy compared to a "standard" or "delayed" approach. These trials have fundamentally shifted global practice toward a more conservative threshold for RRT initiation [38]. This evolution of practice likely explains the null finding in the older MIMIC-III cohort, while the clear association we observed in the more contemporary MIMIC-IV and eICU cohorts more accurately reflects current practice. Previous literature on sepsis fluid management has often been contentious, largely focusing on static, 24-hour cumulative fluid balance or snapshot assessments [39]. Numerous meta-analyses have now robustly confirmed that a high static positive fluid balance at 24, 48, or 72 hours is strongly linked to increased mortality [40]. Our work significantly extends this knowledge. By using a 72-hour trajectory-based approach, we move beyond a simple "positive vs. negative" dichotomy. We demonstrate that a dynamic pattern of "High, Rapid Decline" (Class 4) is highly protective—a critical nuance entirely missed by static analysis, which might have simply labeled both Class 4 and Class 5 as "high CFB." This methodology is similar to how trajectory analysis of biomarkers like lactate, C-reactive protein, or even SOFA scores has provided new prognostic insights in sepsis [41,42]. Our finding that Class 4 and Class 1 share a similar, favorable prognosis strongly supports a paradigm shift away from the rigid "liberal vs. restrictive" debate and toward the more modern, personalized concepts of "fluid stewardship" and de-escalation, as encapsulated by the ROSE framework [43]. This study has several significant strengths. First, to our knowledge, it is the largest and most rigorously validated study of its kind, using a robust "development + temporal validation + multicenter external validation" design across three massive databases. This ensures the high generalizability and robustness of our identified trajectories. Second, we used advanced statistical methods, including GBTM for novel phenotype discovery, IPTW-adjusted Cox models to rigorously mitigate confounding by indication, and machine learning to provide a clear pathway for clinical translation. Finally, our focus on dynamic trajectories, rather than static snapshots of fluid balance, provides a more profound insight into the heterogeneity of fluid management in sepsis. We must also acknowledge several limitations. First, this is a retrospective observational study; therefore, we can only report associations, not causation. Although we used IPTW, unmeasured confounders may still exist. Second, our fluid balance data is derived from EHRs and can be prone to inaccuracies, a well-documented limitation of all large database research. Third, GBTM statistically assigns patients to their "most likely" trajectory, which may not perfectly capture every individual's clinical course. Finally, our ML model, while robust, requires further prospective validation before it can be implemented in clinical practice. Conclusion In conclusion, we identified and validated five distinct 72-hour fluid balance trajectories in sepsis patients following their diagnosis. These trajectory subphenotypes are strongly and independently associated with 28-day mortality, with the "Persistently High Balance" (Class 5) trajectory representing a unique high-risk phenotype uniquely associated with profound organ failure and death. The "High, Rapid Decline" (Class 4) trajectory, however, is associated with a highly favorable prognosis, equivalent to that of a persistent negative balance. These findings challenge static assessments of fluid balance and provide a new dynamic framework for prognosticating sepsis. Furthermore, the ability to predict the high-risk Class 5 trajectory early using machine learning offers a tangible pathway toward individualized fluid stewardship in the ICU. Abbreviations AKI Acute kidney injury APP Average Posterior Probability AUC Area Under the Receiver Operating Characteristic BIC Bayesian Information Criterion CCI Charlson Comorbidity Index CFB Cumulative Fluid Balance CI Confidence Interval CNS Central Nervous System COPD Chronic Obstructive Pulmonary Disease CRRT Continuous Renal Replacement Therapy DBP Diastolic Blood Pressure eICU eICU Collaborative Research Database GBTM Group-Based Trajectory Modeling GCS Glasgow Coma Scale GI Gastrointestinal GU Genitourinary Hb Hemoglobin HF Heart failure HR Hazard Ratio ICU Intensive Care Unit IPTW Inverse Probability of Treatment Weighting MAP Mean Arterial Pressure MICE Multiple Imputation by Chained Equations MIMIC-III Medical Information Mart for Intensive Care-III CareVue subset MIMIC-IV Medical Information Mart for Intensive Care IV MSE Mean Squared Error OASIS Oxford Acute Severity of Illness Score RRT Renal Replacement Therapy SBP Systolic Blood Pressure SHAP SHapley Additive exPlanations SOFA Sequential Organ Failure Assessment WBC White blood cell count XGBoost eXtreme Gradient Boosting Declarations Ethics approval and consent to participate: The study was conducted in accordance with the Declaration of Helsinki. Ethical approval and individual patient consent were waived for all three databases (MIMIC-IV, MIMIC-III-CareVue, and eICU-CRD) because they contain de-identified health information that is publicly available for research purposes. The use of the MIMIC-IV and MIMIC-III databases for this study was specifically approved by the Massachusetts Institute of Technology Institutional Review Board. The authorized researcher (certification number 13024213) completed the required data user training, which granted access approval for all three publicly available cohorts, including eICU-CRD. Consent for publication: Not applicable. Availability of data and materials The datasets utilized in this study were obtained from three large-scale, publicly available critical care databases: the Medical Information Mart for Intensive Care IV (MIMIC-IV, version 3.1), the MIMIC-III CareVue subset (v1.4), and the eICU Collaborative Research Database (eICU-CRD, version 2.0). Data access was granted and data were extracted by an authorized researcher (Kangxing Wang, certification number: 13024213) after obtaining necessary approvals and completing ethical training. The datasets are publicly accessible for research purposes via PhysioNet. Competing Interests: The authors declare no competing financial interests. Fundings: The study was approved by National Key R&D Program of China（2022YFC2504500）and 1.3.5 project for disciplines of excellence, West China Hospital, Sichuan University（ZYGD23012） Author Contributions WKX, XHY, and YYK contributed equally to this work as co-first authors. WKX was the primary contributor, responsible for the study conception, essential data acquisition (coding and extraction), performing the main statistical analysis, and drafting the manuscript. XHY refined the methodology, provided technical support for the analysis, and assisted with data management. YYK assisted with the statistical modeling, interpretation of results, and critical revision of the manuscript. KY and ZYF served as co-corresponding authors, secured funding, and supervised the study. Both KY and ZYF critically revised the manuscript for intellectual content. All authors read and approved the final manuscript. Acknowledgments: Not applicable References Singer M, Deutschman CS, Seymour CW, Shankar-Hari M, Annane D, Bauer M, et al. The third international consensus definitions for sepsis and septic shock (sepsis-3). JAMA. 2016;315:801. https://doi.org/10.1001/jama.2016.0287 Meyer NJ, Prescott HC. Sepsis and septic shock. Hardin CC, editor. N Engl J Med. 2024;391:2133–46. https://doi.org/10.1056/NEJMra2403213 Rudd KE, Johnson SC, Agesa KM, Shackelford KA, Tsoi D, Kievlan DR, et al. Global, regional, and national sepsis incidence and mortality, 1990–2017: analysis for the global burden of disease study. Lancet. 2020;395:200–11. https://doi.org/10.1016/S0140-6736(19)32989-7 Fleischmann-Struzek C, Mellhammar L, Rose N, Cassini A, Rudd KE, Schlattmann P, et al. Incidence and mortality of hospital- and ICU-treated sepsis: results from an updated and expanded systematic review and meta-analysis. Intensive Care Med. 2020;46:1552–62. https://doi.org/10.1007/s00134-020-06151-x Rhee C, Jones TM, Hamad Y, Pande A, Varon J, O’Brien C, et al. Prevalence, underlying causes, and preventability of sepsis-associated mortality in US acute care hospitals. JAMA Netw Open. 2019;2:e187571. https://doi.org/10.1001/jamanetworkopen.2018.7571 De Backer D, Deutschman CS, Hellman J, Myatra SN, Ostermann M, Prescott HC, et al. Surviving sepsis campaign research priorities 2023. Crit Care Med. 2024;52:268–96. https://doi.org/10.1097/CCM.0000000000006135 Nishie H. Guidelines for management of severe sepsis and septic shock. Okayama Igakkai Zasshi (j Okayama Med Assoc). 2013;125:153–7. https://doi.org/10.4044/joma.125.153 Rhodes A, Evans LE, Alhazzani W, Levy MM, Antonelli M, Ferrer R, et al. Surviving sepsis campaign: international guidelines for management of sepsis and septic shock: 2016. Intensive Care Med. 2017;43:304–77. https://doi.org/10.1007/s00134-017-4683-6 Evans L, Rhodes A, Alhazzani W, Antonelli M, Coopersmith CM, French C, et al. Surviving sepsis campaign: international guidelines for management of sepsis and septic shock 2021. Intensive Care Med. 2021;47:1181–247. https://doi.org/10.1007/s00134-021-06506-y Rivers E, Muzzin A. Early goal-directed therapy in the treatment of severe sepsis and septic shock. N Engl J Med. 2001; Mouncey PR, Osborn TM, Power GS, Harrison DA, Sadique MZ, Grieve RD, et al. Trial of early, goal-directed resuscitation for septic shock. N Engl J Med. 2015;372:1301–11. https://doi.org/10.1056/NEJMoa1500896 The ARISE Investigators and the ANZICS Clinical Trials Group. Goal-directed resuscitation for patients with early septic shock. N Engl J Med. 2014;371:1496–506. https://doi.org/10.1056/NEJMoa1404380 The ProCESS Investigators. A randomized trial of protocol-based care for early septic shock. N Engl J Med. 2014;370:1683–93. https://doi.org/10.1056/NEJMoa1401602 Liu VX, Morehouse JW, Marelich GP, Soule J, Russell T, Skeath M, et al. Multicenter implementation of a treatment bundle for patients with sepsis and intermediate lactate values. Am J Respir Crit Care Med. 2016;193:1264–70. https://doi.org/10.1164/rccm.201507-1489OC Lat I, Coopersmith CM, De Backer D. The surviving sepsis campaign: fluid resuscitation and vasopressor therapy research priorities in adult patients. Crit Care Med. 2021;49:623–35. https://doi.org/10.1097/CCM.0000000000004864 Monnet X, Lai C, Teboul J-L. How I personalize fluid therapy in septic shock? Crit Care. 2023;27:123. https://doi.org/10.1186/s13054-023-04363-3 Ruste M, Schweizer R, Fellahi J-L, Jacquet-Lagrèze M. Fluid removal tolerance during the de-escalation phase: is preload unresponsiveness the best guiding candidate? Crit Care. 2023;27:154. https://doi.org/10.1186/s13054-023-04444-3 Nagin DS, Odgers CL. Group-based trajectory modeling in clinical research. Annu Rev Clin Psychol. 2010;6:109–38. https://doi.org/10.1146/annurev.clinpsy.121208.131413 Johnson AEW, Bulgarelli L, Shen L, Gayles A, Shammout A, Horng S, et al. MIMIC-IV, a freely accessible electronic health record dataset. Sci Data. 2023;10:1. https://doi.org/10.1038/s41597-022-01899-x Johnson A, Pollard T, Mark R. MIMIC-III clinical database CareVue subset [Internet]. PhysioNet; [cited 2025 Oct 20]. https://doi.org/10.13026/8A4Q-W170 Moukheiber D, Temps W, Molgi B, Li Y, Lu A, Nannapaneni P, et al. Northwestern ICU (NWICU) database [Internet]. PhysioNet; [cited 2025 Sep 20]. https://doi.org/10.13026/S84W-1829 Hu W, Chen H, Ma C, Sun Q, Yang M, Wang H, et al. Identification of indications for albumin administration in septic patients with liver cirrhosis. Crit Care. 2023;27:300. https://doi.org/10.1186/s13054-023-04587-3 Wang Z, Wang W, Xu J, He Q, Sun C, Xie S, et al. Development and validation of dynamic clinical subphenotypes in acute pancreatitis patients using vital sign trajectories in intensive care units: a multinational cohort study. Signal Transduction Targeted Ther. 2025;10:180. https://doi.org/10.1038/s41392-025-02261-4 Malbrain MLNG, Van Regenmortel N, Saugel B, De Tavernier B, Van Gaal P-J, Joannes-Boyau O, et al. Principles of fluid management and stewardship in septic shock: it is time to consider the four D’s and the four phases of fluid therapy. Ann Intensive Care. 2018;8:66. https://doi.org/10.1186/s13613-018-0402-x Nandhabalan P, Ioannou N, Meadows C, Wyncoll D. Refractory septic shock: our pragmatic approach. Crit Care. 2018;22:215. https://doi.org/10.1186/s13054-018-2144-4 Ostermann M, Straaten HMO, Forni LG. Fluid overload and acute kidney injury: cause or consequence? Crit Care. 2015;19:443, s13054-15-1163–7. https://doi.org/10.1186/s13054-015-1163-7 Silversides JA, Major E, Ferguson AJ, Mann EE, McAuley DF, Marshall JC, et al. Conservative fluid management or deresuscitation for patients with sepsis or acute respiratory distress syndrome following the resuscitation phase of critical illness: a systematic review and meta-analysis. Intensive Care Med. 2017;43:155–70. https://doi.org/10.1007/s00134-016-4573-3 Fleuren LM, Klausch TLT, Zwager CL, Schoonmade LJ, Guo T, Roggeveen LF, et al. Machine learning for the prediction of sepsis: a systematic review and meta-analysis of diagnostic test accuracy. Intensive Care Med. 2020;46:383–400. https://doi.org/10.1007/s00134-019-05872-y Seymour CW, Kennedy JN, Wang S, Chang C-CH, Elliott CF, Xu Z, et al. Derivation, validation, and potential treatment implications of novel clinical phenotypes for sepsis. JAMA. 2019;321:2003. https://doi.org/10.1001/jama.2019.5791 Niu D, Bai H, Zong Y. The association between ionized calcium level and 28-day mortality in patients with sepsis: a cohort study. Sci Rep. 2025;15:22761. https://doi.org/10.1038/s41598-025-05090-1 Al-Juhani A, Desoky R, Iskander Z, Alshehri KT, Alshehri AA, Almuhaimid A, et al. Advances in data-driven early warning systems for sepsis recognition and intervention in emergency care: a systematic review of diagnostic performance and clinical outcomes. Cureus [Internet]. 2025 [cited 2025 Nov 11]; https://doi.org/10.7759/cureus.89882 Wiss AL, Doepker BA, Hoyte B, Olson LM, Disney KA, McLaughlin EM, et al. Impact of initial fluid resuscitation volume on clinical outcomes in patients with heart failure and septic shock. J Intensive Med. 2023;3:254–60. https://doi.org/10.1016/j.jointm.2023.05.001 Jones TW, Smith SE, Tuyl JSV, Newsome AS. Sepsis with preexisting heart failure: management of confounding clinical features. 2022; Malbrain MLNG, Wong A, Nasa P, Ghosh S, editors. Rational use of intravenous fluids in critically ill patients [Internet]. Cham: Springer International Publishing; 2024 [cited 2025 Nov 11]. https://doi.org/10.1007/978-3-031-42205-8 Luo K, Fu S, Fang W, Xu G. The optimal time of initiation of renal replacement therapy in acute kidney injury: A meta-analysis. Oncotarget. 2017;8:68795–808. https://doi.org/10.18632/oncotarget.17946 Gaudry S, Hajage D, Schortgen F, Martin-Lefevre L, Pons B, Boulet E, et al. Initiation strategies for renal-replacement therapy in the intensive care unit. N Engl J Med. 2016;375:122–33. https://doi.org/10.1056/NEJMoa1603017 The STARRT-AKI Investigators. Timing of initiation of renal-replacement therapy in acute kidney injury. N Engl J Med. 2020;383:240–51. https://doi.org/10.1056/NEJMoa2000741 Barbar SD, Clere-Jehl R, Bourredjem A, Hernu R, Montini F, Bruyère R, et al. Timing of renal-replacement therapy in patients with acute kidney injury and sepsis. N Engl J Med. 2018;379:1431–42. https://doi.org/10.1056/NEJMoa1803213 Pfortmueller CA, Dabrowski W, Wise R, Van Regenmortel N, Malbrain MLNG. Fluid accumulation syndrome in sepsis and septic shock: pathophysiology, relevance and treatment—a comprehensive review. Ann Intensive Care. 2024;14:115. https://doi.org/10.1186/s13613-024-01336-9 Messmer AS, Zingg C, Müller M, Gerber JL, Schefold JC, Pfortmueller CA. Fluid overload and mortality in adult critical care patients—a systematic review and meta-analysis of observational studies*. Crit Care Med. 2020;48:1862–70. https://doi.org/10.1097/CCM.0000000000004617 Yang J, Ma B, Tong H. Lymphocyte count trajectories are associated with the prognosis of sepsis patients. Crit Care. 2024;28:399. https://doi.org/10.1186/s13054-024-05186-6 Xu Z, Mao C, Su C, Zhang H, Siempos I, Torres LK, et al. Sepsis subphenotyping based on organ dysfunction trajectory. Crit Care. 2022;26:197. https://doi.org/10.1186/s13054-022-04071-4 Zampieri FG, Bagshaw SM, Semler MW. Fluid therapy for critically ill adults with sepsis: a review. JAMA. 2023;329:1967. https://doi.org/10.1001/jama.2023.7560 Tables Table 1 and 2 are available in the Supplementary Files section. Additional Declarations No competing interests reported. Supplementary Files Table1.xlsx Table2.xlsx SupplementaryTableS1.xlsx SupplementaryTableS2.xlsx SupplementaryTableS3.xlsx SupplementaryFigureS1.png SupplementaryFigureS2.png SupplementaryFigureS3.png SupplementaryFigureS4.png Cite Share Download PDF Status: Published Journal Publication published 26 Mar, 2026 Read the published version in Scientific Reports → Version 1 posted Editorial decision: Revision requested 09 Jan, 2026 Reviews received at journal 02 Jan, 2026 Reviews received at journal 02 Jan, 2026 Reviews received at journal 23 Dec, 2025 Reviews received at journal 23 Dec, 2025 Reviewers agreed at journal 19 Dec, 2025 Reviewers agreed at journal 18 Dec, 2025 Reviewers agreed at journal 18 Dec, 2025 Reviewers agreed at journal 18 Dec, 2025 Reviewers invited by journal 17 Dec, 2025 Editor invited by journal 08 Dec, 2025 Editor assigned by journal 02 Dec, 2025 Submission checks completed at journal 02 Dec, 2025 First submitted to journal 01 Dec, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8247018","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":562549494,"identity":"3be8db6b-88b6-4423-8852-37aba4ec7a7b","order_by":0,"name":"kangxing wang","email":"","orcid":"","institution":"West China Hospital of Sichuan University","correspondingAuthor":false,"prefix":"","firstName":"kangxing","middleName":"","lastName":"wang","suffix":""},{"id":562549495,"identity":"c8552b2f-c634-4e41-9c72-470697decd48","order_by":1,"name":"Huaiyu Xiong","email":"","orcid":"","institution":"West China Hospital of Sichuan University","correspondingAuthor":false,"prefix":"","firstName":"Huaiyu","middleName":"","lastName":"Xiong","suffix":""},{"id":562549496,"identity":"6e11013d-a154-4fc8-bdc8-ed49eef4024d","order_by":2,"name":"Yukun Zhu","email":"","orcid":"","institution":"West China Hospital of Sichuan University","correspondingAuthor":false,"prefix":"","firstName":"Yukun","middleName":"","lastName":"Zhu","suffix":""},{"id":562549497,"identity":"651c00be-3108-4544-82d2-146380a4fbb4","order_by":3,"name":"Yongfang Zhou","email":"","orcid":"","institution":"West China Hospital of Sichuan University","correspondingAuthor":false,"prefix":"","firstName":"Yongfang","middleName":"","lastName":"Zhou","suffix":""},{"id":562549498,"identity":"a59e58f8-995f-48b1-a12b-fb07df824265","order_by":4,"name":"Yan Kang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA0klEQVRIiWNgGAWjYBACCSgtB6HYSNBiTLqWxAaitUi2nz38mqfmTvr89jMGDB/KDjPwz27Ar0WaJy/NmufYs9wNZ3IMGGecO8wgcecAfi1yDDlmxjxsh3M3SPAYMPO2HWYwkEggoIX/DVDLv8Pp8jOAWv4So0VaIsf4MdDwBIYbQC2MxGiRnPHGjHFu32HDDWfSCg72nEvnkbhBQIvE+RzjD2++HZaXbz+88cGPMms5/hkEtAABmxQPlHUAiHnwqIQD5o8/iFE2CkbBKBgFIxcAAHf0QIqxyCiqAAAAAElFTkSuQmCC","orcid":"","institution":"West China Hospital of Sichuan University","correspondingAuthor":true,"prefix":"","firstName":"Yan","middleName":"","lastName":"Kang","suffix":""}],"badges":[],"createdAt":"2025-12-01 07:08:13","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8247018/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8247018/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1038/s41598-026-46063-2","type":"published","date":"2026-03-26T16:10:00+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":98780853,"identity":"0fbaaaeb-0311-4a8d-b622-d207dc826347","added_by":"auto","created_at":"2025-12-22 12:31:45","extension":"png","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":249544,"visible":true,"origin":"","legend":"","description":"","filename":"Figure1.png","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/a5093a180dc5bd8e9e92b5cd.png"},{"id":98780744,"identity":"b420248e-999b-4ebc-ad30-996117e6d53c","added_by":"auto","created_at":"2025-12-22 12:31:37","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":86165,"visible":true,"origin":"","legend":"","description":"","filename":"article.docx","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/24e7c3e26bbd5aab08da0531.docx"},{"id":98766858,"identity":"aad17ec6-6855-4f88-9e52-8d6f52dde48e","added_by":"auto","created_at":"2025-12-22 10:19:40","extension":"png","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":4702844,"visible":true,"origin":"","legend":"","description":"","filename":"Figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/5c106c26aaf9915de17144e9.png"},{"id":98766848,"identity":"19d3c016-9253-4cdf-bab0-322fcd11516d","added_by":"auto","created_at":"2025-12-22 10:19:40","extension":"xlsx","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":9766,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryTableS1.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/4ed0d4740b7a93447660bf8a.xlsx"},{"id":98766859,"identity":"e0a53ac2-4d9f-42f6-b4e1-7f9d7da67df6","added_by":"auto","created_at":"2025-12-22 10:19:40","extension":"jpg","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":4404973,"visible":true,"origin":"","legend":"","description":"","filename":"Figure3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/6857c53c70c7af8b7b48a241.jpg"},{"id":98779146,"identity":"8269a40f-df3c-411a-9ab7-df4272ce4bc6","added_by":"auto","created_at":"2025-12-22 12:30:00","extension":"xlsx","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":9814,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryTableS2.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/1af2ae38cfaa01e45e21dbc2.xlsx"},{"id":98766861,"identity":"3836b5dd-864c-4bcb-9afa-b411a21b963f","added_by":"auto","created_at":"2025-12-22 10:19:40","extension":"png","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":5244904,"visible":true,"origin":"","legend":"","description":"","filename":"Figure4.png","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/891849265e543cbf50a52ca9.png"},{"id":98779502,"identity":"d1736e05-d179-4a1c-92eb-4d3df035189a","added_by":"auto","created_at":"2025-12-22 12:30:25","extension":"xlsx","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":10142,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryTableS3.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/ad02ca7995174bb640c47479.xlsx"},{"id":98780446,"identity":"54f94ecb-d4a2-4760-bb1a-d2dfbf724a8d","added_by":"auto","created_at":"2025-12-22 12:31:21","extension":"xlsx","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":14078,"visible":true,"origin":"","legend":"","description":"","filename":"Table1.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/0c99a1273bd76a7dcdcedbfa.xlsx"},{"id":98766892,"identity":"70dc59a4-55f8-484f-8632-d80804fb0c08","added_by":"auto","created_at":"2025-12-22 10:19:42","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":89566095,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFigureS1.png","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/071aff0a644e5c2dd743d511.png"},{"id":98766879,"identity":"1233f941-91e6-4186-9aff-e12b2237e086","added_by":"auto","created_at":"2025-12-22 10:19:41","extension":"xlsx","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":10520,"visible":true,"origin":"","legend":"","description":"","filename":"Table2.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/113f10e53663f4ff528b9edc.xlsx"},{"id":98766862,"identity":"8b4ea0e9-7dc1-48d8-84e9-0885f1a948b2","added_by":"auto","created_at":"2025-12-22 10:19:40","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1931186,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFigureS2.png","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/a9b8e6209ca72436dc4579a1.png"},{"id":98766872,"identity":"4fe7f10c-ab7d-4ca4-864d-0c1bbae9121d","added_by":"auto","created_at":"2025-12-22 10:19:41","extension":"png","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":5100585,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFigureS3.png","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/9bba22c68a127c0ec5adc64f.png"},{"id":98780756,"identity":"3a6df06d-252c-4f8d-bed3-14fd79e4f54b","added_by":"auto","created_at":"2025-12-22 12:31:38","extension":"png","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3040185,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFigureS4.png","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/c5e0608f959a3fe351ebb48c.png"},{"id":98780401,"identity":"0aec9789-1d3f-41ab-bf08-7f1372f460a6","added_by":"auto","created_at":"2025-12-22 12:31:18","extension":"json","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":8395,"visible":true,"origin":"","legend":"","description":"","filename":"492ee1fa9713457b81ced6dbc7b70fcc.json","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/2db400651758dcfe019ff445.json"},{"id":98766865,"identity":"19907091-6b57-43f0-8a4e-c4ea2aecbb90","added_by":"auto","created_at":"2025-12-22 10:19:40","extension":"xml","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":87653,"visible":true,"origin":"","legend":"","description":"","filename":"492ee1fa9713457b81ced6dbc7b70fcc1enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/f282eaf2987c0434727cd0d6.xml"},{"id":98766869,"identity":"be12dd71-34f0-45dc-b126-cba679e9eac5","added_by":"auto","created_at":"2025-12-22 10:19:40","extension":"png","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":249544,"visible":true,"origin":"","legend":"","description":"","filename":"Figure1.png","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/ad76d4b0089a4bf40f4c389e.png"},{"id":98766867,"identity":"0c3ecff6-840f-4dcf-a1c9-a55b1cb84baa","added_by":"auto","created_at":"2025-12-22 10:19:40","extension":"png","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":4702844,"visible":true,"origin":"","legend":"","description":"","filename":"Figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/b9acec2253e70523b905c6d2.png"},{"id":98779157,"identity":"07ead038-8c43-4121-a76b-217cb1221383","added_by":"auto","created_at":"2025-12-22 12:30:00","extension":"jpg","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":4404973,"visible":true,"origin":"","legend":"","description":"","filename":"Figure3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/99f14c7d5e1b3d60049aabd0.jpg"},{"id":98766870,"identity":"9941b207-a801-41d3-ad2a-f8823a2b6c1d","added_by":"auto","created_at":"2025-12-22 10:19:40","extension":"png","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":5244904,"visible":true,"origin":"","legend":"","description":"","filename":"Figure4.png","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/aa51522d9cca2105c59baf25.png"},{"id":98766891,"identity":"f0ac9aa7-0d7d-41df-8473-60c9366479bb","added_by":"auto","created_at":"2025-12-22 10:19:41","extension":"png","order_by":20,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":89566095,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFigureS1.png","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/7b1248842a8f6d1099eb5387.png"},{"id":98780887,"identity":"5b390bd8-c27a-485b-8602-bc9d5907e259","added_by":"auto","created_at":"2025-12-22 12:31:48","extension":"png","order_by":21,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1931186,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFigureS2.png","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/e285b6ee6b60b811ebf02ec6.png"},{"id":98779495,"identity":"d1a7ce78-57b1-4ec2-8a66-ff76bc893293","added_by":"auto","created_at":"2025-12-22 12:30:25","extension":"png","order_by":22,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":5100585,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFigureS3.png","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/dc33e0c0c726f59d302c6a9c.png"},{"id":98778898,"identity":"1fdb1727-9e28-4e05-8245-546aca7c9464","added_by":"auto","created_at":"2025-12-22 12:29:47","extension":"png","order_by":23,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3040185,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFigureS4.png","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/6b8310983b89ac02551d9aad.png"},{"id":98779557,"identity":"bde041d2-7283-4ba3-8302-3fc991f369e0","added_by":"auto","created_at":"2025-12-22 12:30:27","extension":"png","order_by":24,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":157493,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure1.png","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/6191046a012af5e8ea08a908.png"},{"id":98766886,"identity":"a445a9be-6c5e-4fe0-9345-06f99c5eef68","added_by":"auto","created_at":"2025-12-22 10:19:41","extension":"png","order_by":25,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1336294,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure2.png","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/256298fe7dbb509be73624dc.png"},{"id":98780854,"identity":"fb016c79-0758-4f80-b0b5-43eabca36be8","added_by":"auto","created_at":"2025-12-22 12:31:45","extension":"png","order_by":26,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":655762,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure3.png","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/53687ec315edaabfd186cdad.png"},{"id":98778694,"identity":"cdb642ff-2b13-4143-8962-10df2cbef169","added_by":"auto","created_at":"2025-12-22 12:29:32","extension":"png","order_by":27,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1166829,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure4.png","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/1291cbd7ebce8830ed5fcc3c.png"},{"id":98766887,"identity":"422e5cca-d7f7-4a57-80f7-3f5127882e47","added_by":"auto","created_at":"2025-12-22 10:19:41","extension":"png","order_by":28,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":10895760,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineSupplementaryFigureS1.png","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/7261ab7a17560f8ac5551777.png"},{"id":98780262,"identity":"b72fdfdd-0f9a-472e-969b-2fedf52f0150","added_by":"auto","created_at":"2025-12-22 12:31:11","extension":"png","order_by":29,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1161813,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineSupplementaryFigureS2.png","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/a569dee2ca45b6e9142e558c.png"},{"id":98766888,"identity":"9069dd99-2d9c-4c68-a2fb-b351a3302d00","added_by":"auto","created_at":"2025-12-22 10:19:41","extension":"png","order_by":30,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1534543,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineSupplementaryFigureS3.png","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/e945f3e8bd2cc332e4a94535.png"},{"id":98766883,"identity":"5b910476-4619-4aaf-887b-8f2bd26ba9c2","added_by":"auto","created_at":"2025-12-22 10:19:41","extension":"png","order_by":31,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":945061,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineSupplementaryFigureS4.png","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/ed3083fcd810c986fc534274.png"},{"id":98766878,"identity":"8e21a00d-7c20-47fd-8cb7-bd46763c72d6","added_by":"auto","created_at":"2025-12-22 10:19:41","extension":"xml","order_by":32,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":87040,"visible":true,"origin":"","legend":"","description":"","filename":"492ee1fa9713457b81ced6dbc7b70fcc1structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/70bdb3d6e5c7806ee1f71769.xml"},{"id":98780137,"identity":"edb8df61-206e-4473-b2fd-c11242edf430","added_by":"auto","created_at":"2025-12-22 12:31:06","extension":"html","order_by":33,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":96944,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/fda5e0aad6c1f307f52978ca.html"},{"id":98766850,"identity":"4cd062b8-1931-4d31-a199-b10591fd6e36","added_by":"auto","created_at":"2025-12-22 10:19:40","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":249544,"visible":true,"origin":"","legend":"\u003cp\u003eFlowchart of patient selection for the development and validation cohorts.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eAbbreviations\u003c/em\u003e: MIMIC-IV, Medical Information Mart for Intensive Care-IV; MIMIC-III-CareVue, Medical Information Mart for Intensive Care-III CareVue subset; eICU, eICU Collaborative Research Database; ICU, intensive care unit; n, number of patients.\u003c/p\u003e","description":"","filename":"Figure1.png","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/02d9bb05b790358d52f09fbc.png"},{"id":98766853,"identity":"0dbbd76d-57ed-4f84-ae06-6868d8757402","added_by":"auto","created_at":"2025-12-22 10:19:40","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":4702844,"visible":true,"origin":"","legend":"\u003cp\u003e72-hour cumulative fluid balance (CFB) trajectories and associated 28-day mortality.\u003c/p\u003e\n\u003cp\u003e(A) Five distinct CFB trajectories identified in the MIMIC-IV development cohort. (B) Trajectories validated in the MIMIC-III-CareVue temporal validation cohort. (C) Trajectories validated in the eICU multicenter validation cohort. (D) Kaplan-Meier survival curves for the five trajectories in the MIMIC-IV cohort. (E) Kaplan-Meier survival curves in the MIMIC-III-CareVue cohort. (F) Kaplan-Meier survival curves in the eICU cohort.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eAbbreviations\u003c/em\u003e: CFB, cumulative fluid balance; MIMIC-IV, Medical Information Mart for Intensive Care-IV; MIMIC-III-CareVue, Medical Information Mart for Intensive Care-III CareVue subset; eICU, eICU Collaborative Research Database.\u003c/p\u003e","description":"","filename":"Figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/3c5552413736bf26a607f0cc.png"},{"id":98780872,"identity":"80b0d4fc-285e-4a54-833c-48b180d0fa48","added_by":"auto","created_at":"2025-12-22 12:31:47","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":4404973,"visible":true,"origin":"","legend":"\u003cp\u003eSubgroup analysis of the association between fluid balance trajectories and 28-day mortality.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eAbbreviations\u003c/em\u003e: HR, hazard ratio; CI, confidence interval; SOFA, Sequential Organ Failure Assessment; HF, heart failure; Vaso, vasopressor use.\u003c/p\u003e","description":"","filename":"Figure3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/6938b895507cc46c835ca046.jpg"},{"id":98766863,"identity":"b34b5926-ea01-4b94-9169-82744928c9b2","added_by":"auto","created_at":"2025-12-22 10:19:40","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":5244904,"visible":true,"origin":"","legend":"\u003cp\u003ePerformance and interpretability of machine learning models for early prediction of the high-risk Class 5 trajectory.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eAbbreviations\u003c/em\u003e: ROC, receiver operating characteristic; AUC, area under the curve; XGBoost, eXtreme Gradient Boosting; SHAP, SHapley Additive exPlanations; OASIS, Oxford Acute Severity of Illness Score; SOFA, Sequential Organ Failure Assessment; Ca, calcium.\u003c/p\u003e","description":"","filename":"Figure4.png","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/ee1c3b886cf6b148e15f3fc9.png"},{"id":106959166,"identity":"d8ea4d17-6d86-4fae-9bc5-a2f23068ae09","added_by":"auto","created_at":"2026-04-15 08:51:28","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":19761696,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/7f391828-285e-41f0-89c3-1ef8b5f3a5ca.pdf"},{"id":98779831,"identity":"c736b479-17a3-47e5-abe9-25c35230dac1","added_by":"auto","created_at":"2025-12-22 12:30:48","extension":"xlsx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":14078,"visible":true,"origin":"","legend":"","description":"","filename":"Table1.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/fbbb008e66f41418ad9eb7b5.xlsx"},{"id":98766846,"identity":"f4e41f1b-558b-462d-b9d1-7d9940f98f7d","added_by":"auto","created_at":"2025-12-22 10:19:40","extension":"xlsx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":10520,"visible":true,"origin":"","legend":"","description":"","filename":"Table2.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/e08e086066d53b1c584558ad.xlsx"},{"id":98780478,"identity":"e9167a15-6ce8-455d-be29-9bab2f538958","added_by":"auto","created_at":"2025-12-22 12:31:23","extension":"xlsx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":9766,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryTableS1.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/ccbde6c4a825c09f595b0c4e.xlsx"},{"id":98780728,"identity":"19a585f2-8f90-4398-8c62-b0454386863f","added_by":"auto","created_at":"2025-12-22 12:31:35","extension":"xlsx","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":9814,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryTableS2.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/0cd6b767c022432ae3db0355.xlsx"},{"id":98766860,"identity":"dc1e2cc5-2df5-4bb6-ba24-b8dcc0689934","added_by":"auto","created_at":"2025-12-22 10:19:40","extension":"xlsx","order_by":4,"title":"","display":"","copyAsset":false,"role":"supplement","size":10142,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryTableS3.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/b2ef87ab459413a83cd0336b.xlsx"},{"id":98766895,"identity":"a6489e1c-2864-42ed-b121-0f46b33bef93","added_by":"auto","created_at":"2025-12-22 10:19:43","extension":"png","order_by":5,"title":"","display":"","copyAsset":false,"role":"supplement","size":89566095,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFigureS1.png","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/aa3314ba0f98037b336d5682.png"},{"id":98766855,"identity":"682cc9c1-c9b7-440e-9e46-f05eb7cdf35d","added_by":"auto","created_at":"2025-12-22 10:19:40","extension":"png","order_by":6,"title":"","display":"","copyAsset":false,"role":"supplement","size":1931186,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFigureS2.png","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/63ed517cb85352a52caf76de.png"},{"id":98766866,"identity":"a61e4f75-38df-49fe-9e05-71f86ac9d8e1","added_by":"auto","created_at":"2025-12-22 10:19:40","extension":"png","order_by":7,"title":"","display":"","copyAsset":false,"role":"supplement","size":5100585,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFigureS3.png","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/793f6ac43e28d9673dfe501d.png"},{"id":98780852,"identity":"e99594f1-1ec3-4eae-bd0c-ee299207107c","added_by":"auto","created_at":"2025-12-22 12:31:45","extension":"png","order_by":8,"title":"","display":"","copyAsset":false,"role":"supplement","size":3040185,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFigureS4.png","url":"https://assets-eu.researchsquare.com/files/rs-8247018/v1/421823f6c70de855afc13e19.png"}],"financialInterests":"No competing interests reported.","formattedTitle":"Development and Validation of 72-Hour Fluid Balance Trajectory Subphenotypes and Prognosis in Sepsis: A Multicenter Cohort Study","fulltext":[{"header":"Background","content":"\u003cp\u003eSepsis, defined as life-threatening organ dysfunction caused by a dysregulated host response to infection, represents a major global health challenge [1,2]. Before the COVID-19 pandemic, sepsis affected nearly 50\u0026nbsp;million people worldwide annually and continues to be associated with a high risk of death, killing between one in three and one in six of those affected [3,4]. In the United States alone, sepsis is implicated in over one-third of in-hospital deaths, making it the most common cause of in-hospital mortality and the most expensive cause of hospitalization [5]. Despite advances in medical care, mortality from sepsis remains high, highlighting significant knowledge gaps and the persistent need for research to improve patient outcomes [2,6].\u003c/p\u003e \u003cp\u003eEarly fluid resuscitation is considered a cornerstone of sepsis management [6\u0026ndash;9]. Initial strategies, such as Early Goal-Directed Therapy (EGDT), showed early promise in improving outcomes [10]. However, subsequent large-scale randomized controlled trials (RCTs), namely the ProCESS, ARISE, and ProMISe studies, failed to demonstrate the superiority of EGDT over usual care in terms of survival [11\u0026ndash;13]. These findings challenged the notion of \"more is better\" or the strict adherence to specific resuscitation protocols. Consequently, despite the critical importance of fluid resuscitation, the optimal fluid management strategy\u0026mdash;encompassing the initial volume, the timing and methods for subsequent fluid administration, and appropriate resuscitation targets\u0026mdash;remains a central controversy and a key unresolved issue in sepsis research and clinical practice [2,6]. Indeed, \"how to individualize fluid resuscitation initially and beyond\" has been identified as a top clinical research priority in the field of sepsis [6].\u003c/p\u003e \u003cp\u003eFluid therapy in sepsis presents a \"double-edged sword\": while essential, excessive administration leading to fluid overload or persistent positive fluid balance is linked to increased mortality, acute kidney injury, and prolonged ventilation [14,15]. This underscores the need for precise fluid stewardship beyond initial resuscitation. Given the substantial heterogeneity among sepsis patients, individualized strategies that move beyond fixed protocols are increasingly sought. Therapeutic approaches ideally should be tailored to the patient's physiological state and the specific phase of resuscitation, as conceptualized by models like the Resuscitation-Optimization-Stabilization-Evacuation (ROSE) framework [16].\u003c/p\u003e \u003cp\u003eHowever, the limitations inherent in static fluid assessments and snapshot phenotyping approaches highlight a critical gap: the prognostic significance of temporal patterns in fluid balance during the initial 72 hours of sepsis remains largely unexplored [16,17]. It is unclear whether classifying patients based on these early fluid trajectories provides additional prognostic information beyond traditional risk stratification tools, such baseline severity scores or static cumulative fluid balance assessments at fixed time points. We hypothesized that identifying distinct subgroups of sepsis patients based on their 72-hour cumulative fluid balance trajectories, using Group-Based Trajectory Modeling (GBTM) [18], would reveal clinically meaningful phenotypes associated with significantly different risks for adverse outcomes, particularly 28-day mortality. Therefore, this study aimed to delineate these early fluid trajectories in large, multicenter cohorts of sepsis patients and evaluate their independent association with prognosis.\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eData sources\u003c/h2\u003e \u003cp\u003eThis multi-center retrospective cohort study utilized data from three large-scale, publicly available critical care databases. The development cohort, used for identifying fluid trajectory subphenotypes, was derived from the Medical Information Mart for Intensive Care IV (MIMIC-IV) database (version 3.1), representing patients admitted to a single academic medical center (Beth Israel Deaconess Medical Center, BIDMC) between 2008 and 2022 [19]. To assess temporal robustness, the MIMIC-III CareVue subset (v1.4), containing data from the same institution during an earlier period (2001\u0026ndash;2008), served as the temporal validation cohort [20]. Generalizability was assessed using the eICU Collaborative Research Database (eICU-CRD) (version 2.0) as the multi-center validation cohort, which includes data from over 200 hospitals across the United States (2014\u0026ndash;2015) [21]. Data access was granted and data were extracted by an authorized researcher (Kangxing Wang, certification number: 13024213) after obtaining necessary institutional approvals and completing required ethical training.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eStudy population\u003c/h3\u003e\n\u003cp\u003eWe included adult patients (age\u0026thinsp;\u0026gt;\u0026thinsp;18 years) admitted to the ICU who met the Sepsis-3 criteria, defined as suspected infection (indicated by concurrent administration of antibiotics and sampling of body fluids) combined with an acute increase in the Sequential Organ Failure Assessment (SOFA) score of \u0026ge;\u0026thinsp;2 points [1]. Only the first ICU admission was included for patients with multiple admissions during a single hospitalization. The exclusion criteria were: (1) ICU length of stay\u0026thinsp;\u0026lt;\u0026thinsp;72 hours; (2) multiple ICU admissions (only the first was retained, as mentioned); and (3) patients with fewer than two recorded fluid balance data points during the first 72 hours. The detailed patient selection process for all three databases is illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e\n\u003ch3\u003eOutcome\u003c/h3\u003e\n\u003cdiv class=\"Heading\"\u003eOutcome\u003c/div\u003e \u003cp\u003eThe primary outcome was all-cause mortality within 28 days following ICU admission. The key secondary outcomes were the duration of mechanical ventilation (in hours) and the utilization of renal replacement therapy (RRT) during the ICU stay.\u003c/p\u003e\n\u003ch3\u003eFeature extraction\u003c/h3\u003e\n\u003cp\u003eData extraction was performed using structured query language (SQL) in PostgreSQL (version 8.2). For each patient, a comprehensive set of baseline variables was obtained, primarily reflecting their clinical status within the first 24 hours after ICU admission unless otherwise specified. The extracted variables included demographics (age, sex, race); ICU type; severity of illness scores (Sequential Organ Failure Assessment [SOFA], Oxford Acute Severity of Illness Score [OASIS], Glasgow Coma Scale [GCS], and Charlson Comorbidity Index [CCI]); comorbidities identified by ICD-9/10 codes (hypertension, diabetes mellitus, chronic obstructive pulmonary disease [COPD], heart failure [HF], and stroke); \u003cb\u003einfection site\u003c/b\u003e (lung, gastrointestinal, genitourinary, or other) [22]; the most abnormal vital signs within 24 hours (minimum mean arterial pressure, maximum heart rate, maximum respiratory rate, and maximum temperature); initial laboratory values (first recorded hemoglobin, white blood cell count, platelets, creatinine, lactate, PaO₂, PaCO₂, pH, potassium, sodium, chloride, and calcium); and interventions within 24 hours (mechanical ventilation, renal replacement therapy [RRT], and vasopressor use).\u003c/p\u003e\n\u003ch3\u003eFluid balance trajectory construction\u003c/h3\u003e\n\u003cp\u003eThe primary fluid balance trajectory variable was derived as follows: hourly net fluid balance data were extracted for the first 72 hours following the diagnosis of sepsis and aggregated into 24 consecutive 3-hour intervals. For each interval, the net fluid balance was normalized by the patient\u0026rsquo;s admission body weight (kg) to yield a value in mL/kg. These 3-hourly, weight-normalized net fluid balance values were winsorized at the 1st and 99th percentiles\u0026mdash;based on the development cohort (MIMIC-IV)\u0026mdash;to reduce the influence of extreme outliers.\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eStatistical analysis\u003c/h2\u003e \u003cp\u003eBaseline characteristics were summarized as mean\u0026thinsp;\u0026plusmn;\u0026thinsp;standard deviation (SD) or median [interquartile range (IQR)] for continuous variables, and as counts (percentages) for categorical variables. Comparisons were made using the Student\u0026rsquo;s t-test, Mann-Whitney U test, or Chi-square test, as appropriate. We assessed variables for missingness, and those with \u0026lt;\u0026thinsp;30% missing data were imputed using multiple imputation by chained equations (MICE). We utilized Group-Based Trajectory Modeling (GBTM) to identify distinct patient subphenotypes based on their cumulative fluid balance trajectories during the first 72 hours after sepsis diagnosis. To determine the optimal number of trajectory classes, we fitted models with a varying number of classes (e.g., 2\u0026ndash;7) in the development cohort. The final model selection was guided by a combination of statistical criteria and clinical interpretability, including the Bayesian Information Criterion (BIC), average posterior probability (APP\u0026thinsp;\u0026gt;\u0026thinsp;0.90), relative class size (\u0026gt;\u0026thinsp;10%), and the clinical plausibility of the trajectories. For external validation, the trained model was applied to the validation cohorts. Each patient was assigned to the trajectory class that yielded the lowest Mean Squared Error (MSE) when fitting their individual data to the class-specific polynomial functions [23]. Kaplan-Meier curves with the log-rank test were used for unadjusted survival comparisons. We employed multivariable Cox proportional hazards regression to evaluate the independent association between fluid trajectories and outcomes. Three sequential models were constructed: Model 1 (Crude Model), which included only the trajectory class as the predictor; Model 2 (Demographics-Adjusted Model), which adjusted for trajectory class, age, gender, and race; and Model 3 (Fully-Adjusted Model), which adjusted for all variables in Model 2 plus ICU type, severity of illness scores (SOFA, OASIS, GCS, Charlson Comorbidity Index), infection site (lung, GI, GU, other), baseline comorbidities (hypertension, diabetes, COPD, heart failure, stroke), 24-hour vital signs (minimum MAP, maximum heart rate, maximum respiratory rate, maximum temperature), initial laboratory values (hemoglobin, WBC, platelet, creatinine, lactate, PO2, PCO2, pH, potassium, sodium, chloride, calcium), and 24-hour interventions (mechanical ventilation, RRT, and vasopressor use). Several sensitivity analyses were performed to test the robustness of our findings. First, to further control for confounding, we applied an Inverse Probability of Treatment Weighting (IPTW) analysis. Second, to test the robustness of our missing data imputation method, we repeated the analysis utilizing random forest imputation for missing data instead of MICE. All analyses were conducted using R, version 4.2.2.\u003c/p\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003eStudy Cohort and Trajectory Identification\u003c/h2\u003e \u003cp\u003eBased on the Group-Based Trajectory Modeling (GBTM) analysis, a five-class model was determined to be the optimal fit for the 72-hour cumulative fluid balance data. This model was selected based on the lowest Bayesian Information Criterion (BIC\u0026thinsp;=\u0026thinsp;2317109), excellent average posterior probabilities (APP\u0026thinsp;\u0026gt;\u0026thinsp;0.90 for all classes), and all class proportions being greater than 12%. The detailed model fit statistics are available in Supplementary Table S1-S2. The five distinct fluid trajectories are visualized in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eA and were clinically characterized as follows: Class 1 (Blue line): \"Persistent Negative Balance\" (n\u0026thinsp;=\u0026thinsp;3,772, 23.5%), which started with a low-positive balance and consistently trended downwards, ending near or below a 0 mL/kg cumulative balance by 72 hours; Class 3 (Yellow line): \"Low-Stable Positive Balance\" (n\u0026thinsp;=\u0026thinsp;2,538, 15.8%), which maintained a stable, low-level positive fluid balance throughout the 72-hour period; Class 2 (Orange line): \"Medium-Stable Positive Balance\" (n\u0026thinsp;=\u0026thinsp;3,372, 21.0%), which maintained a stable, medium-level positive fluid balance throughout the 72-hour period; Class 4 (Green line): \"High, Rapid Decline\" (n\u0026thinsp;=\u0026thinsp;4,417, 27.5%), the largest group, characterized by a high initial cumulative balance that rapidly and steadily declined over 72 hours; and Class 5 (Red line): \"Persistently High Balance\" (n\u0026thinsp;=\u0026thinsp;1,970, 12.3%), which started with the highest cumulative balance and maintained this persistently high positive balance with little to no decline over the 72-hour period. Critically, these five distinct trajectory patterns were consistently identified in both the MIMIC-III-CareVue (temporal validation) and eICU (external validation) cohorts, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eB and \u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eC.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eClinical Characterization of Subphenotypes\u003c/h2\u003e \u003cp\u003eThe baseline characteristics of the five subphenotypes in the MIMIC-IV cohort are presented in Table\u0026nbsp;1. A critical comparison emerges between Class 5 (\"Persistently High Balance\") and Class 4 (\"High, Rapid Decline\"), which represented the two most acutely ill profiles at baseline. Patients in Class 5 demonstrated the highest severity, with the highest median SOFA score (9.00), OASIS score (40.00), initial lactate (2.40 mmol/L), and the highest rates of 24-hour vasopressor use (70.91%) and mechanical ventilation (72.34%). In contrast, while Class 4 patients also presented with high initial illness severity (SOFA 7.00, Vaso 64.89%), they were distinguished by near-normal renal function (Creatinine 1.00 mg/dL) and a minimal 24-hour RRT rate (1.47%). This was in stark opposition to Class 5, which had significantly higher creatinine (1.40 mg/dL) and a dramatically higher RRT rate of 17.01%. The other three trajectories, Class 1 (\"Persistent Negative\"), Class 2 (\"Medium-Stable\"), and Class 3 (\"Low-Stable\"), generally represented patients with lower baseline severity scores.\u003c/p\u003e \u003cp\u003eThese static baseline differences were mirrored in the 72-hour dynamic physiological trends. As shown in Supplementary Figure S1, patients in Class 5 exhibited persistent shock, maintaining the lowest mean arterial pressure (MAP) and the highest heart rate over the 72-hour period. Furthermore, the 72-hour SOFA score trajectories (Supplementary Figure S1) revealed that Class 5 patients not only started with the highest organ dysfunction but also experienced a progressive worsening or failure to improve, particularly in the cardiovascular and renal sub-scores. This aligns perfectly with their high RRT rate and persistently high fluid balance observed in Table\u0026nbsp;1.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eAssociation with Primary Outcome\u003c/h2\u003e \u003cp\u003eUnadjusted analyses demonstrated a strong association between fluid trajectories and 28-day mortality. Class 5 (\"Persistently High Balance\") had the highest mortality rate (38.58%), while Class 1 (\"Persistent Negative\") and Class 4 (\"High, Rapid Decline\") had the lowest (16.54% and 17.70%, respectively) (Table\u0026nbsp;1). Kaplan-Meier survival analysis confirmed this finding, showing a significant separation in survival probabilities across all five groups (Log-rank p\u0026thinsp;\u0026lt;\u0026thinsp;0.001) (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eD). This clear prognostic stratification was robustly replicated in both the MIMIC-III-CareVue (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eE) and eICU external validation cohorts (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eF).\u003c/p\u003e \u003cp\u003eIn the multivariable Cox proportional hazards analysis (Table\u0026nbsp;2), using the highest-risk Class 5 as the reference, all other trajectories were associated with significantly lower mortality in the fully-adjusted Model 3. Specifically, in the MIMIC-IV cohort, the adjusted hazard ratios (aHR) were 0.50 (95% CI 0.44\u0026ndash;0.56) for Class 1, 0.51 (95% CI 0.46\u0026ndash;0.57) for Class 4, 0.73 (95% CI 0.64\u0026ndash;0.83) for Class 3, and 0.87 (95% CI 0.78\u0026ndash;0.96) for Class 2. These findings were confirmed in the Inverse Probability of Treatment Weighting (IPTW) model, which controlled for baseline confounding. This independent association remained highly consistent across the external validation cohorts, particularly for the protective associations seen in Class 1 and Class 4.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003eAssociation with Secondary Outcomes\u003c/h2\u003e \u003cp\u003eThe prognostic association of the fluid trajectories extended to secondary ICU outcomes (Supplementary Figure S2). The association with mechanical ventilation was highly robust across all cohorts: in the MIMIC-IV development cohort, Class 5 (\"High-Persist\") had a significantly higher ventilation burden (both hours and proportion) compared to all other groups (Kruskal-Wallis p\u0026thinsp;\u0026lt;\u0026thinsp;0.001; all post-hoc p.adj\u0026thinsp;\u0026lt;\u0026thinsp;0.001). This finding was robustly validated in both the eICU and MIMIC-III cohorts, which showed a nearly identical pattern of significant differences. In contrast, the association with CRRT utilization was strong in the development cohort (Class 5 vs. all others, p.adj\u0026thinsp;\u0026lt;\u0026thinsp;0.001), partially validated in the eICU cohort (where Class 1 showed significantly lower use than other groups), and not observed in the MIMIC-III temporal validation cohort (Kruskal-Wallis p\u0026thinsp;=\u0026thinsp;0.78).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003eSubgroup Analyses\u003c/h2\u003e \u003cp\u003eTo assess the consistency of the primary outcome, we performed subgroup analyses on the association between fluid trajectories and 28-day mortality, using Class 5 as the reference (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e). The protective effect of the other trajectories (especially Class 1 and Class 4) was consistent across most subgroups, including age (\u0026lt;\u0026thinsp;65 vs. \u0026gt;=65 years, P for interaction\u0026thinsp;=\u0026thinsp;0.531) and baseline SOFA score (\u0026thinsp;\u0026lt;\u0026thinsp;=\u0026thinsp;5 vs. \u0026gt;5, P for interaction\u0026thinsp;=\u0026thinsp;0.755). However, we observed significant interactions for patients with baseline heart failure (HF, P for interaction\u0026thinsp;=\u0026thinsp;0.001), stroke (P for interaction\u0026thinsp;=\u0026thinsp;0.019), and vasopressor use (Vaso, P for interaction\u0026thinsp;=\u0026thinsp;0.001). Specifically, in patients with pre-existing stroke, the protective effects of Class 1, Class 2, and Class 3 were attenuated and lost statistical significance (p\u0026thinsp;=\u0026thinsp;0.06, 0.32, and 0.32, respectively). Furthermore, the protective associations of Class 1 and Class 4 appeared to be even stronger in patients with HF or those receiving vasopressors, while the effects of Class 2 and 3 were attenuated in these same high-risk groups.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eEarly Prediction of the High-Risk (Class 5) Trajectory\u003c/h2\u003e \u003cp\u003eFinally, to explore the feasibility of early identification, we developed and validated several machine learning models using baseline data to predict patient membership in the high-risk Class 5 (\"Persistently High Balance\") trajectory. Among the six models tested, the XGBoost model demonstrated the best and most robust performance (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eA). It achieved a high Area Under the Receiver Operating Characteristic (AUC) of 0.811 in the MIMIC-IV development (internal validation) cohort. Critically, this high predictive performance was well-sustained during external validation, achieving an AUC of 0.793 in the multi-center eICU cohort and 0.809 in the MIMIC-III-CareVue temporal validation cohort. Model interpretability using SHAP analysis (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eB) identified baseline severity scores (OASIS, SOFA), Ca, Infection, and Lactate as the most important features for predicting this high-risk fluid trajectory. As shown in the SHAP plot, high initial values of OASIS, SOFA, and Lactate, and a lower level of Ca, were all strongly associated with an increased likelihood of being classified into Class 5.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003eSensitivity Analyses\u003c/h2\u003e \u003cp\u003eTwo key sensitivity analyses were conducted to confirm the robustness of our primary findings. First, as shown in the primary analysis (Table\u0026nbsp;2), the associations remained consistent after applying an Inverse Probability of Treatment Weighting (IPTW) model to account for baseline confounding. Second, to test the robustness of our missing data imputation method, we repeated the entire analysis using Random Forest imputation instead of MICE. The results were nearly identical to our primary findings (Supplementary Table S3); for example, in the MIMIC-IV Model 3, the aHR for Class 1 vs. Class 5 was 0.49 (95% CI 0.43\u0026ndash;0.56), compared to 0.50 (95% CI 0.44\u0026ndash;0.56) in the main analysis.\u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this large, multicenter retrospective cohort study utilizing three distinct critical care databases (MIMIC-IV, eICU, and MIMIC-III-CareVue), we successfully identified and validated five distinct and reproducible subphenotypes based on 72-hour cumulative fluid balance trajectories following sepsis diagnosis. Our primary finding is that these dynamic trajectories are a robust and independent predictor of 28-day mortality, providing prognostic information beyond traditional static baseline severity scores. As hypothesized, the Class 5 (\"Persistently High Balance\") trajectory was associated with the worst outcomes, including the highest mortality, longest duration of mechanical ventilation, and greatest CRRT utilization. However, the central and most significant finding of this study is that the Class 4 (\"High, Rapid Decline\") trajectory\u0026mdash;a pattern perfectly aligning with the successful \"Evacuation\" (de-resuscitation) phase of the ROSE framework\u0026mdash;was associated with a favorable prognosis nearly identical to that of the Class 1 (\"Persistent Negative Balance\") trajectory. This observation, which remained robust after rigorous adjustment for baseline confounding using an IPTW model, strongly suggests that the dynamic process of successfully reversing a high positive fluid balance may be as prognostically protective as a primary state of persistent negative balance.\u003c/p\u003e \u003cp\u003eA central finding of this study is the stark contrast in outcomes between Class 5 and Class 4. Admittedly, the Class 5 subphenotype was characterized by more severe baseline renal impairment, as evidenced by higher creatinine and a 17.01% 24-hour RRT rate in Table\u0026nbsp;1. However, a critical finding is that even after rigorously adjusting for these baseline confounders (including renal function and SOFA score) using an IPTW model, the Class 5 trajectory remained independently associated with significantly higher mortality compared to Class 4. This strongly implies that the trajectory itself\u0026mdash;the process of persistent high fluid balance\u0026mdash;carries a prognostic significance independent of the baseline state. The 72-hour dynamic data in Supplementary Figure S3 provides a clear physiological explanation for this persistent risk. The vital sign trajectories demonstrate that Class 5 patients remained in a state of persistent shock (lowest MAP, highest HR) that failed to reverse. Moreover, their SOFA sub-score trajectories (Supplementary Figure S4) revealed progressive or non-improving organ dysfunction, particularly in the Cardiovascular (Panel B) and Renal (Panel E) scores. Therefore, the Class 5 trajectory likely represents a phenotype of \"refractory shock and failed de-resuscitation,\" where persistent fluid administration is both a consequence of, and a contributor to, a vicious cycle of non-reversing organ failure, such as AKI [24\u0026ndash;26]. In contrast, Class 4, whose dynamic SOFA scores stabilized or improved, represents a \"successful resuscitation\" cohort with the capacity to be de-resuscitated\u0026mdash;a capability increasingly recognized as a critical determinant of survival in sepsis [27] .\u003c/p\u003e \u003cp\u003eA key limitation of trajectory analysis is that it is a descriptive, post-hoc classification. To bridge this gap for clinical practice, we demonstrated that the high-risk Class 5 trajectory can be accurately predicted within the first 24 hours. While many machine learning models have been developed to predict sepsis mortality [28], our focus on predicting a dynamic trajectory aligns with the strategic move toward identifying actionable subphenotypes [29]. The SHAP analysis revealed that this prediction was driven by intuitive clinical features. The importance of features like OASIS, SOFA, and Lactate is well-documented in prognostication, as exemplified by the Sepsis-3 definitions [1], but the inclusion of Ca (Calcium) is notable, as hypocalcemia is increasingly recognized as a marker of severity and poor prognosis in sepsis [30]. This model provides a proof-of-concept for a crucial clinical tool: an early warning system, a type of intervention that has shown promise for improving sepsis outcomes by facilitating earlier intervention [31].\u003c/p\u003e \u003cp\u003eOur subgroup analyses confirmed the robustness of the primary findings across most key subgroups, including age and baseline SOFA score (P for interaction\u0026thinsp;\u0026gt;\u0026thinsp;0.05). However, we observed several clinically significant interactions (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). Most notably, in patients with baseline heart failure (HF, P-interaction\u0026thinsp;=\u0026thinsp;0.001), the protective effects of all other trajectories (Classes 1\u0026ndash;4) relative to Class 5 were significantly magnified compared to non-HF patients. This suggests that for patients with compromised cardiac function, the survival benefit of avoiding the \"Persistently High\" (Class 5) trajectory is even more pronounced, adding critical nuance to the controversy surrounding the 30mL/kg bolus in this population [6,32,33]. Furthermore, in patients with stroke (P-interaction\u0026thinsp;=\u0026thinsp;0.019), only the Class 4 (\"High, Rapid Decline\") trajectory remained significantly protective (p\u0026thinsp;\u0026lt;\u0026thinsp;0.05), while the benefits of other trajectories were attenuated. This may reflect the clinical \"tightrope walk\" of neurocritical care, where an \"aggressive-followed-by-rapid-removal\" strategy (Class 4) may represent the optimal balance between cerebral perfusion and avoiding cerebral edema [34].\u003c/p\u003e \u003cp\u003eAn important and honest finding of our study was the heterogeneity in secondary outcomes. While the association with mechanical ventilation was remarkably consistent across all three databases, the association with CRRT was strong in MIMIC-IV, partially validated in eICU, and absent in the MIMIC-III cohort (p\u0026thinsp;=\u0026thinsp;0.78). We believe this does not invalidate the finding, but rather reflects a critical real-world variance in clinical practice over time. The MIMIC-III cohort (2001\u0026ndash;2008) represents an older era of critical care. During that period, the optimal timing of RRT initiation was a major debate, with many observational studies and meta-analyses suggesting a potential survival benefit for \"early\" RRT [35]. However, several landmark randomized controlled trials (RCTs) published after this period, most notably the AKIKI trial (2016) [36] and the STARRT-AKI trial (2020) [37], failed to show a survival benefit for an \"accelerated\" or \"early\" strategy compared to a \"standard\" or \"delayed\" approach. These trials have fundamentally shifted global practice toward a more conservative threshold for RRT initiation [38]. This evolution of practice likely explains the null finding in the older MIMIC-III cohort, while the clear association we observed in the more contemporary MIMIC-IV and eICU cohorts more accurately reflects current practice.\u003c/p\u003e \u003cp\u003ePrevious literature on sepsis fluid management has often been contentious, largely focusing on static, 24-hour cumulative fluid balance or snapshot assessments [39]. Numerous meta-analyses have now robustly confirmed that a high static positive fluid balance at 24, 48, or 72 hours is strongly linked to increased mortality [40]. Our work significantly extends this knowledge. By using a 72-hour trajectory-based approach, we move beyond a simple \"positive vs. negative\" dichotomy. We demonstrate that a dynamic pattern of \"High, Rapid Decline\" (Class 4) is highly protective\u0026mdash;a critical nuance entirely missed by static analysis, which might have simply labeled both Class 4 and Class 5 as \"high CFB.\" This methodology is similar to how trajectory analysis of biomarkers like lactate, C-reactive protein, or even SOFA scores has provided new prognostic insights in sepsis [41,42]. Our finding that Class 4 and Class 1 share a similar, favorable prognosis strongly supports a paradigm shift away from the rigid \"liberal vs. restrictive\" debate and toward the more modern, personalized concepts of \"fluid stewardship\" and de-escalation, as encapsulated by the ROSE framework [43].\u003c/p\u003e \u003cp\u003eThis study has several significant strengths. First, to our knowledge, it is the largest and most rigorously validated study of its kind, using a robust \"development\u0026thinsp;+\u0026thinsp;temporal validation\u0026thinsp;+\u0026thinsp;multicenter external validation\" design across three massive databases. This ensures the high generalizability and robustness of our identified trajectories. Second, we used advanced statistical methods, including GBTM for novel phenotype discovery, IPTW-adjusted Cox models to rigorously mitigate confounding by indication, and machine learning to provide a clear pathway for clinical translation. Finally, our focus on dynamic trajectories, rather than static snapshots of fluid balance, provides a more profound insight into the heterogeneity of fluid management in sepsis.\u003c/p\u003e \u003cp\u003eWe must also acknowledge several limitations. First, this is a retrospective observational study; therefore, we can only report associations, not causation. Although we used IPTW, unmeasured confounders may still exist. Second, our fluid balance data is derived from EHRs and can be prone to inaccuracies, a well-documented limitation of all large database research. Third, GBTM statistically assigns patients to their \"most likely\" trajectory, which may not perfectly capture every individual's clinical course. Finally, our ML model, while robust, requires further prospective validation before it can be implemented in clinical practice.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eIn conclusion, we identified and validated five distinct 72-hour fluid balance trajectories in sepsis patients following their diagnosis. These trajectory subphenotypes are strongly and independently associated with 28-day mortality, with the \"Persistently High Balance\" (Class 5) trajectory representing a unique high-risk phenotype uniquely associated with profound organ failure and death. The \"High, Rapid Decline\" (Class 4) trajectory, however, is associated with a highly favorable prognosis, equivalent to that of a persistent negative balance. These findings challenge static assessments of fluid balance and provide a new dynamic framework for prognosticating sepsis. Furthermore, the ability to predict the high-risk Class 5 trajectory early using machine learning offers a tangible pathway toward individualized fluid stewardship in the ICU.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cdiv class=\"DefinitionList\"\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eAKI\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eAcute kidney injury\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eAPP\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eAverage Posterior Probability\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eAUC\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eArea Under the Receiver Operating Characteristic\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eBIC\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eBayesian Information Criterion\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eCCI\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eCharlson Comorbidity Index\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eCFB\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eCumulative Fluid Balance\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eCI\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eConfidence Interval\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eCNS\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eCentral Nervous System\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eCOPD\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eChronic Obstructive Pulmonary Disease\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eCRRT\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eContinuous Renal Replacement Therapy\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eDBP\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eDiastolic Blood Pressure\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eeICU\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eeICU Collaborative Research Database\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eGBTM\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eGroup-Based Trajectory Modeling\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eGCS\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eGlasgow Coma Scale\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eGI\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eGastrointestinal\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eGU\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eGenitourinary\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eHb\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eHemoglobin\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eHF\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eHeart failure\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eHR\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eHazard Ratio\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eICU\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eIntensive Care Unit\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eIPTW\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eInverse Probability of Treatment Weighting\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eMAP\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eMean Arterial Pressure\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eMICE\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eMultiple Imputation by Chained Equations\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eMIMIC-III\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eMedical Information Mart for Intensive Care-III CareVue subset\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eMIMIC-IV\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eMedical Information Mart for Intensive Care IV\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eMSE\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eMean Squared Error\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eOASIS\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eOxford Acute Severity of Illness Score\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eRRT\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eRenal Replacement Therapy\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eSBP\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eSystolic Blood Pressure\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eSHAP\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eSHapley Additive exPlanations\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eSOFA\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eSequential Organ Failure Assessment\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eWBC\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eWhite blood cell count\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eXGBoost\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eeXtreme Gradient Boosting\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe study was conducted in accordance with the Declaration of Helsinki. Ethical approval and individual patient consent were waived for all three databases (MIMIC-IV, MIMIC-III-CareVue, and eICU-CRD) because they contain de-identified health information that is publicly available for research purposes. The use of the MIMIC-IV and MIMIC-III databases for this study was specifically approved by the Massachusetts Institute of Technology Institutional Review Board. The authorized researcher (certification number 13024213) completed the required data user training, which granted access approval for all three publicly available cohorts, including eICU-CRD.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe datasets utilized in this study were obtained from three large-scale, publicly available critical care databases: the Medical Information Mart for Intensive Care IV (MIMIC-IV, version 3.1), the MIMIC-III CareVue subset (v1.4), and the eICU Collaborative Research Database (eICU-CRD, version 2.0). Data access was granted and data were extracted by an authorized researcher (Kangxing Wang, certification number: 13024213) after obtaining necessary approvals and completing ethical training. The datasets are publicly accessible for research purposes via PhysioNet.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting Interests:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare no competing financial interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFundings:\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe study was approved by\u0026nbsp;National Key R\u0026amp;D Program of China（2022YFC2504500）and 1.3.5 project for disciplines of excellence, West China Hospital, Sichuan University（ZYGD23012）\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor Contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWKX, XHY, and YYK contributed equally to this work as co-first authors. WKX was the primary contributor, responsible for the study conception, essential data acquisition (coding and extraction), performing the main statistical analysis, and drafting the manuscript. XHY refined the methodology, provided technical support for the analysis, and assisted with data management. YYK assisted with the statistical modeling, interpretation of results, and critical revision of the manuscript. KY and ZYF served as co-corresponding authors, secured funding, and supervised the study. Both KY and ZYF critically revised the manuscript for intellectual content. All authors read and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgments:\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eSinger M, Deutschman CS, Seymour CW, Shankar-Hari M, Annane D, Bauer M, et al. The third international consensus definitions for sepsis and septic shock (sepsis-3). JAMA. 2016;315:801. https://doi.org/10.1001/jama.2016.0287\u003c/li\u003e\n\u003cli\u003eMeyer NJ, Prescott HC. Sepsis and septic shock. Hardin CC, editor. N Engl J Med. 2024;391:2133–46. https://doi.org/10.1056/NEJMra2403213\u003c/li\u003e\n\u003cli\u003eRudd KE, Johnson SC, Agesa KM, Shackelford KA, Tsoi D, Kievlan DR, et al. Global, regional, and national sepsis incidence and mortality, 1990–2017: analysis for the global burden of disease study. Lancet. 2020;395:200–11. https://doi.org/10.1016/S0140-6736(19)32989-7\u003c/li\u003e\n\u003cli\u003eFleischmann-Struzek C, Mellhammar L, Rose N, Cassini A, Rudd KE, Schlattmann P, et al. Incidence and mortality of hospital- and ICU-treated sepsis: results from an updated and expanded systematic review and meta-analysis. Intensive Care Med. 2020;46:1552–62. https://doi.org/10.1007/s00134-020-06151-x\u003c/li\u003e\n\u003cli\u003eRhee C, Jones TM, Hamad Y, Pande A, Varon J, O’Brien C, et al. Prevalence, underlying causes, and preventability of sepsis-associated mortality in US acute care hospitals. JAMA Netw Open. 2019;2:e187571. https://doi.org/10.1001/jamanetworkopen.2018.7571\u003c/li\u003e\n\u003cli\u003eDe Backer D, Deutschman CS, Hellman J, Myatra SN, Ostermann M, Prescott HC, et al. Surviving sepsis campaign research priorities 2023. Crit Care Med. 2024;52:268–96. https://doi.org/10.1097/CCM.0000000000006135\u003c/li\u003e\n\u003cli\u003eNishie H. Guidelines for management of severe sepsis and septic shock. Okayama Igakkai Zasshi (j Okayama Med Assoc). 2013;125:153–7. https://doi.org/10.4044/joma.125.153\u003c/li\u003e\n\u003cli\u003eRhodes A, Evans LE, Alhazzani W, Levy MM, Antonelli M, Ferrer R, et al. Surviving sepsis campaign: international guidelines for management of sepsis and septic shock: 2016. Intensive Care Med. 2017;43:304–77. https://doi.org/10.1007/s00134-017-4683-6\u003c/li\u003e\n\u003cli\u003eEvans L, Rhodes A, Alhazzani W, Antonelli M, Coopersmith CM, French C, et al. Surviving sepsis campaign: international guidelines for management of sepsis and septic shock 2021. Intensive Care Med. 2021;47:1181–247. https://doi.org/10.1007/s00134-021-06506-y\u003c/li\u003e\n\u003cli\u003eRivers E, Muzzin A. Early goal-directed therapy in the treatment of severe sepsis and septic shock. N Engl J Med. 2001;\u003c/li\u003e\n\u003cli\u003eMouncey PR, Osborn TM, Power GS, Harrison DA, Sadique MZ, Grieve RD, et al. Trial of early, goal-directed resuscitation for septic shock. N Engl J Med. 2015;372:1301–11. https://doi.org/10.1056/NEJMoa1500896\u003c/li\u003e\n\u003cli\u003eThe ARISE Investigators and the ANZICS Clinical Trials Group. Goal-directed resuscitation for patients with early septic shock. N Engl J Med. 2014;371:1496–506. https://doi.org/10.1056/NEJMoa1404380\u003c/li\u003e\n\u003cli\u003eThe ProCESS Investigators. A randomized trial of protocol-based care for early septic shock. N Engl J Med. 2014;370:1683–93. https://doi.org/10.1056/NEJMoa1401602\u003c/li\u003e\n\u003cli\u003eLiu VX, Morehouse JW, Marelich GP, Soule J, Russell T, Skeath M, et al. Multicenter implementation of a treatment bundle for patients with sepsis and intermediate lactate values. Am J Respir Crit Care Med. 2016;193:1264–70. https://doi.org/10.1164/rccm.201507-1489OC\u003c/li\u003e\n\u003cli\u003eLat I, Coopersmith CM, De Backer D. The surviving sepsis campaign: fluid resuscitation and vasopressor therapy research priorities in adult patients. Crit Care Med. 2021;49:623–35. https://doi.org/10.1097/CCM.0000000000004864\u003c/li\u003e\n\u003cli\u003eMonnet X, Lai C, Teboul J-L. How I personalize fluid therapy in septic shock? Crit Care. 2023;27:123. https://doi.org/10.1186/s13054-023-04363-3\u003c/li\u003e\n\u003cli\u003eRuste M, Schweizer R, Fellahi J-L, Jacquet-Lagrèze M. Fluid removal tolerance during the de-escalation phase: is preload unresponsiveness the best guiding candidate? Crit Care. 2023;27:154. https://doi.org/10.1186/s13054-023-04444-3\u003c/li\u003e\n\u003cli\u003eNagin DS, Odgers CL. Group-based trajectory modeling in clinical research. Annu Rev Clin Psychol. 2010;6:109–38. https://doi.org/10.1146/annurev.clinpsy.121208.131413\u003c/li\u003e\n\u003cli\u003eJohnson AEW, Bulgarelli L, Shen L, Gayles A, Shammout A, Horng S, et al. MIMIC-IV, a freely accessible electronic health record dataset. Sci Data. 2023;10:1. https://doi.org/10.1038/s41597-022-01899-x\u003c/li\u003e\n\u003cli\u003eJohnson A, Pollard T, Mark R. MIMIC-III clinical database CareVue subset [Internet]. PhysioNet; [cited 2025 Oct 20]. https://doi.org/10.13026/8A4Q-W170\u003c/li\u003e\n\u003cli\u003eMoukheiber D, Temps W, Molgi B, Li Y, Lu A, Nannapaneni P, et al. Northwestern ICU (NWICU) database [Internet]. PhysioNet; [cited 2025 Sep 20]. https://doi.org/10.13026/S84W-1829\u003c/li\u003e\n\u003cli\u003eHu W, Chen H, Ma C, Sun Q, Yang M, Wang H, et al. Identification of indications for albumin administration in septic patients with liver cirrhosis. Crit Care. 2023;27:300. https://doi.org/10.1186/s13054-023-04587-3\u003c/li\u003e\n\u003cli\u003eWang Z, Wang W, Xu J, He Q, Sun C, Xie S, et al. Development and validation of dynamic clinical subphenotypes in acute pancreatitis patients using vital sign trajectories in intensive care units: a multinational cohort study. Signal Transduction Targeted Ther. 2025;10:180. https://doi.org/10.1038/s41392-025-02261-4\u003c/li\u003e\n\u003cli\u003eMalbrain MLNG, Van Regenmortel N, Saugel B, De Tavernier B, Van Gaal P-J, Joannes-Boyau O, et al. Principles of fluid management and stewardship in septic shock: it is time to consider the four D’s and the four phases of fluid therapy. Ann Intensive Care. 2018;8:66. https://doi.org/10.1186/s13613-018-0402-x\u003c/li\u003e\n\u003cli\u003eNandhabalan P, Ioannou N, Meadows C, Wyncoll D. Refractory septic shock: our pragmatic approach. Crit Care. 2018;22:215. https://doi.org/10.1186/s13054-018-2144-4\u003c/li\u003e\n\u003cli\u003eOstermann M, Straaten HMO, Forni LG. Fluid overload and acute kidney injury: cause or consequence? Crit Care. 2015;19:443, s13054-15-1163–7. https://doi.org/10.1186/s13054-015-1163-7\u003c/li\u003e\n\u003cli\u003eSilversides JA, Major E, Ferguson AJ, Mann EE, McAuley DF, Marshall JC, et al. Conservative fluid management or deresuscitation for patients with sepsis or acute respiratory distress syndrome following the resuscitation phase of critical illness: a systematic review and meta-analysis. Intensive Care Med. 2017;43:155–70. https://doi.org/10.1007/s00134-016-4573-3\u003c/li\u003e\n\u003cli\u003eFleuren LM, Klausch TLT, Zwager CL, Schoonmade LJ, Guo T, Roggeveen LF, et al. Machine learning for the prediction of sepsis: a systematic review and meta-analysis of diagnostic test accuracy. Intensive Care Med. 2020;46:383–400. https://doi.org/10.1007/s00134-019-05872-y\u003c/li\u003e\n\u003cli\u003eSeymour CW, Kennedy JN, Wang S, Chang C-CH, Elliott CF, Xu Z, et al. Derivation, validation, and potential treatment implications of novel clinical phenotypes for sepsis. JAMA. 2019;321:2003. https://doi.org/10.1001/jama.2019.5791\u003c/li\u003e\n\u003cli\u003eNiu D, Bai H, Zong Y. The association between ionized calcium level and 28-day mortality in patients with sepsis: a cohort study. Sci Rep. 2025;15:22761. https://doi.org/10.1038/s41598-025-05090-1\u003c/li\u003e\n\u003cli\u003eAl-Juhani A, Desoky R, Iskander Z, Alshehri KT, Alshehri AA, Almuhaimid A, et al. Advances in data-driven early warning systems for sepsis recognition and intervention in emergency care: a systematic review of diagnostic performance and clinical outcomes. Cureus [Internet]. 2025 [cited 2025 Nov 11]; https://doi.org/10.7759/cureus.89882\u003c/li\u003e\n\u003cli\u003eWiss AL, Doepker BA, Hoyte B, Olson LM, Disney KA, McLaughlin EM, et al. Impact of initial fluid resuscitation volume on clinical outcomes in patients with heart failure and septic shock. J Intensive Med. 2023;3:254–60. https://doi.org/10.1016/j.jointm.2023.05.001\u003c/li\u003e\n\u003cli\u003eJones TW, Smith SE, Tuyl JSV, Newsome AS. Sepsis with preexisting heart failure: management of confounding clinical features. 2022;\u003c/li\u003e\n\u003cli\u003eMalbrain MLNG, Wong A, Nasa P, Ghosh S, editors. Rational use of intravenous fluids in critically ill patients [Internet]. Cham: Springer International Publishing; 2024 [cited 2025 Nov 11]. https://doi.org/10.1007/978-3-031-42205-8\u003c/li\u003e\n\u003cli\u003eLuo K, Fu S, Fang W, Xu G. The optimal time of initiation of renal replacement therapy in acute kidney injury: A meta-analysis. Oncotarget. 2017;8:68795–808. https://doi.org/10.18632/oncotarget.17946\u003c/li\u003e\n\u003cli\u003eGaudry S, Hajage D, Schortgen F, Martin-Lefevre L, Pons B, Boulet E, et al. Initiation strategies for renal-replacement therapy in the intensive care unit. N Engl J Med. 2016;375:122–33. https://doi.org/10.1056/NEJMoa1603017\u003c/li\u003e\n\u003cli\u003eThe STARRT-AKI Investigators. Timing of initiation of renal-replacement therapy in acute kidney injury. N Engl J Med. 2020;383:240–51. https://doi.org/10.1056/NEJMoa2000741\u003c/li\u003e\n\u003cli\u003eBarbar SD, Clere-Jehl R, Bourredjem A, Hernu R, Montini F, Bruyère R, et al. Timing of renal-replacement therapy in patients with acute kidney injury and sepsis. N Engl J Med. 2018;379:1431–42. https://doi.org/10.1056/NEJMoa1803213\u003c/li\u003e\n\u003cli\u003ePfortmueller CA, Dabrowski W, Wise R, Van Regenmortel N, Malbrain MLNG. Fluid accumulation syndrome in sepsis and septic shock: pathophysiology, relevance and treatment—a comprehensive review. Ann Intensive Care. 2024;14:115. https://doi.org/10.1186/s13613-024-01336-9\u003c/li\u003e\n\u003cli\u003eMessmer AS, Zingg C, Müller M, Gerber JL, Schefold JC, Pfortmueller CA. Fluid overload and mortality in adult critical care patients—a systematic review and meta-analysis of observational studies*. Crit Care Med. 2020;48:1862–70. https://doi.org/10.1097/CCM.0000000000004617\u003c/li\u003e\n\u003cli\u003eYang J, Ma B, Tong H. Lymphocyte count trajectories are associated with the prognosis of sepsis patients. Crit Care. 2024;28:399. https://doi.org/10.1186/s13054-024-05186-6\u003c/li\u003e\n\u003cli\u003eXu Z, Mao C, Su C, Zhang H, Siempos I, Torres LK, et al. Sepsis subphenotyping based on organ dysfunction trajectory. Crit Care. 2022;26:197. https://doi.org/10.1186/s13054-022-04071-4\u003c/li\u003e\n\u003cli\u003eZampieri FG, Bagshaw SM, Semler MW. Fluid therapy for critically ill adults with sepsis: a review. JAMA. 2023;329:1967. https://doi.org/10.1001/jama.2023.7560\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp\u003eTable 1 and 2 are available in the Supplementary Files section.\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Sepsis, Fluid Therapy, Group-Based Trajectory Modeling (GBTM), Subphenotypes","lastPublishedDoi":"10.21203/rs.3.rs-8247018/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8247018/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e \u003cp\u003eFluid management in sepsis is controversial, and while persistent positive fluid balance is linked to adverse outcomes, the prognostic value of the temporal patterns (trajectories) of cumulative fluid balance (CFB) during the first 72 hours after sepsis diagnosis remains unclear. This study aimed to identify distinct subphenotypes based on 72-hour CFB trajectories in adult sepsis patients and evaluate their independent association with 28-day mortality.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003eThis multicenter retrospective cohort study utilized three large-scale critical care databases (MIMIC-IV, MIMIC-III-CareVue, and eICU). We included adult patients meeting Sepsis-3 criteria with an ICU stay of over 72 hours. Group-Based Trajectory Modeling (GBTM) identified five distinct 72-hour CFB (mL/kg) subphenotypes in the MIMIC-IV development cohort (n\u0026thinsp;=\u0026thinsp;16,069). Trajectories were validated in the MIMIC-III (n\u0026thinsp;=\u0026thinsp;2,162) and eICU (n\u0026thinsp;=\u0026thinsp;13,805) cohorts. We used Kaplan-Meier analysis and multivariable Cox proportional hazards models, including Inverse Probability of Treatment Weighting (IPTW), to assess the association with 28-day mortality, adjusting for baseline confounders.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eGBTM identified five reproducible CFB subphenotypes, notably a \"Persistent Negative Balance\" (Class 1), a \"High, Rapid Decline\" (Class 4), and a \"Persistently High Balance\" (Class 5). Class 5 patients exhibited the highest illness severity (e.g., highest SOFA scores). Kaplan-Meier analysis showed significant differences in 28-day survival (P\u0026thinsp;\u0026lt;\u0026thinsp;0.001), with Class 1 and Class 4 having the best survival and Class 5 the worst. In the MIMIC-IV cohort, compared to the highest-risk Class 5 (reference) after full multivariable adjustment, all other trajectories were associated with significantly lower 28-day mortality: Class 1 (HR: 0.50, 95% CI: 0.44\u0026ndash;0.56) and Class 4 (HR: 0.51, 95% CI: 0.46\u0026ndash;0.57). These protective findings were consistent across all validation cohorts.\u003c/p\u003e\u003ch2\u003eConclusion\u003c/h2\u003e \u003cp\u003eSepsis patients exhibit five distinct, reproducible CFB trajectories strongly associated with 28-day mortality, independent of baseline severity. The finding that the \"High, Rapid Decline\" (Class 4) trajectory shares an excellent prognosis with the \"Persistent Negative\" (Class 1) trajectory challenges static fluid assessments. The ability to predict the high-risk Class 5 phenotype early using machine learning (AUC\u0026thinsp;\u0026asymp;\u0026thinsp;0.81) provides a tangible pathway toward individualized fluid stewardship.\u003c/p\u003e","manuscriptTitle":"Development and Validation of 72-Hour Fluid Balance Trajectory Subphenotypes and Prognosis in Sepsis: A Multicenter Cohort Study","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-12-22 10:19:35","doi":"10.21203/rs.3.rs-8247018/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-01-09T08:12:26+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-01-02T11:58:16+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-01-02T10:38:10+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-12-23T12:32:53+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-12-23T08:54:40+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"171036063897027728137048084643230561686","date":"2025-12-19T06:59:15+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"32690859890550695502775192770024445887","date":"2025-12-18T17:53:19+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"167250079065031257693879743428616795991","date":"2025-12-18T09:50:20+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"62650160885788064062955240858078097467","date":"2025-12-18T06:23:08+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-12-18T00:29:54+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2025-12-08T12:06:25+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-12-02T13:05:18+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-12-02T13:04:41+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2025-12-01T06:53:39+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"f6b59d3f-7591-4c03-8cb3-107ceff11720","owner":[],"postedDate":"December 22nd, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[{"id":59908564,"name":"Health sciences/Biomarkers"},{"id":59908565,"name":"Health sciences/Diseases"},{"id":59908566,"name":"Health sciences/Health care"},{"id":59908567,"name":"Health sciences/Medical research"},{"id":59908568,"name":"Health sciences/Risk factors"}],"tags":[],"updatedAt":"2026-03-30T16:31:50+00:00","versionOfRecord":{"articleIdentity":"rs-8247018","link":"https://doi.org/10.1038/s41598-026-46063-2","journal":{"identity":"scientific-reports","isVorOnly":false,"title":"Scientific Reports"},"publishedOn":"2026-03-26 16:10:00","publishedOnDateReadable":"March 26th, 2026"},"versionCreatedAt":"2025-12-22 10:19:35","video":"","vorDoi":"10.1038/s41598-026-46063-2","vorDoiUrl":"https://doi.org/10.1038/s41598-026-46063-2","workflowStages":[]},"version":"v1","identity":"rs-8247018","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8247018","identity":"rs-8247018","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}