Sepsis is fundamentally characterized by a dysregulated host immune response to infection, initiated by excessive inflammatory stimulation and often progressing to organ dysfunction (1). Although the incidence and mortality of sepsis vary by geographic region and level of national development, ICU mortality remains persistently high, typically ranging from 30% to 46% (2). Accumulating evidence indicates that platelets play a pivotal role in the pathophysiology of sepsis (3). Upon activation during sepsis, platelets release procoagulant and pro-inflammatory mediators—components that contribute to host defense but, when excessively activated, may trigger disseminated intravascular coagulation (DIC), leading to organ dysfunction (4). DIC-induced organ failure primarily arises from fibrin deposition in the microvasculature and suppression of fibrinolysis, both of which impair tissue oxygenation and ultimately result in multi-organ dysfunction (5).

Platelets originate from bone marrow megakaryocytes and are non-nucleated cell fragments generated through cytoplasmic shedding. They are small in size but abundant in number, constituting the second most numerous cellular component in circulating blood after red blood cells (6). Traditionally, platelets are well - known for their hemostatic and thrombotic functions. However, their functions extend far beyond hemostasis and thrombus formation. Their crucial role in host immune regulation is attracting increasingly wide - spread research interest (7). Platelets act as a bridge between hemostasis and immunity, and their immunomodulatory functions are mainly reflected in the following two pathways: On the one hand, platelets can recognize the fibrinogen - binding proteins expressed by pathogens via their surface hemostatic receptor GPIIb/IIIa, induce platelet aggregation, and subsequently promote the occurrence of disseminated intravascular coagulation (DIC), ultimately driving the process of multiple organ failure. On the other hand, platelets also express various immune - related receptors such as FcγRIIa and Toll - like receptors (TLR), and directly participate in pathogen recognition and intracellular signal transduction, thus playing a significant role in the innate immune response and platelet activation process (8). More and more studies have revealed that a decrease in platelet count is clearly associated with a poor prognosis in patients with sepsis (9, 10). In addition, the study by Wang et al. further established a significant non - linear relationship between platelet count and 30 - day in - hospital mortality, with the optimal turning point being 176 × 10⁹/L. On both sides of this turning point, the changes in platelet count have opposite effects on the risk of death: below this value, an increase of 10 × 10⁹/L is associated with a 6% reduction in risk; above this value, the same increase is associated with a 1% increase in risk (11). This intricate nonlinear relationship indicates that distinct patients may follow markedly different dynamic evolution trajectories of platelet counts. Moreover, specific trajectory patterns could predict prognosis more precisely than the values at any single, isolated time point. Recently, a review by Aseman et al. put forward the proposal that in sepsis, platelet production, consumption, and functional status undergo rapid temporal changes. This suggests that static measurements are likely to overlook significant temporal heterogeneity and its clinical implications. Therefore, it is essential to characterize the longitudinal dynamic changes of platelets to ascertain whether distinct platelet trajectory patterns can more effectively reflect the underlying pathophysiological mechanisms, thereby overcoming the limitations of isolated measurements and enhancing prognostic stratification (12).

Group-based trajectory modeling (GBTM) is a finite mixture modeling approach that identifies latent subgroups of patients with similar longitudinal trajectories over time. Unlike dimension-reduction approaches that transform data into abstract latent variables, GBTM directly models repeated measurements on the original scale, thereby preserving clinical interpretability. This allows the identification of distinct temporal patterns, clinically meaningful subphenotypes, and dynamic heterogeneity (13). GBTM has been extensively utilized in fields such as medicine and psychology and is especially well - suited for analyzing longitudinal data (14). Sepsis is a heterogeneous syndrome characterized by extensive and multi - dimensional clinical and biological features. Additionally, the physiological response to infection is dynamic and evolves rapidly within minutes to hours (15, 16). From a systems perspective, sepsis can be conceptualized as a Complex Adaptive System characterized by nonlinear regulation, feedback interactions, and time-dependent progression across multiple biological scales. These properties generate heterogeneous clinical trajectories that cannot be adequately captured by static assessments.

Given its unique edge in capturing dynamic heterogeneity, GBTM offers a novel possibility for the in - depth analysis of the pathological process of sepsis. It can identify clinical subgroups with distinctive evolution patterns, thus revealing patient groups that demand key attention. By categorizing patients into trajectory categories corresponding to different prognostic outcomes, this approach facilitates more accurate risk stratification and provides a foundation for formulating individualized treatment strategies, particularly in heterogeneous syndromes such as sepsis where platelet behavior reflects the evolving balance between immune activation and coagulation dysregulation. Therefore, to more comprehensively explore the association between platelet dynamic evolution and prognosis in patients with sepsis, this study measured platelet counts at multiple time points and applied group-based trajectory modeling to identify sepsis subphenotypes with distinct evolution patterns, followed by comparisons of their clinical characteristics.

Due to its high incidence and high mortality rate, sepsis has emerged as a major global public health challenge. The intrinsic high heterogeneity of sepsis is the primary reason why it is difficult to effectively manage this disease. Meanwhile, this highly complex heterogeneity impels us to categorize patients into more homogeneous groups sharing common biological characteristics. By doing so, it not only facilitates the formulation of individualized treatment strategies but also enables the accurate identification of high - risk patient groups (19, 20).

Although it was previously widely assumed that platelet consumption and activation were associated phenomena of sepsis, a growing body of evidence indicates that platelets play a pivotal role in the occurrence and development of sepsis and are also a core factor in sepsis - induced organ dysfunction (5). A substantial number of studies have verified that the severity of thrombocytopenia is an independent predictor of the prognosis of sepsis patients. Given that platelet detection is straightforward and the results can be obtained rapidly, it has become a highly regarded biomarker in this research field (21–23). For example, Hua et al. evaluated admission platelet levels using conventional logistic regression, while Wang et al. applied nonlinear and threshold-based models to explore associations between initial platelet levels and mortality. Although these approaches improved risk stratification, they remained inherently cross-sectional and were unable to capture the dynamic evolution of platelet changes during disease progression, as single time-point measurements cannot adequately reflect temporal variability (9, 11). Currently, there was still a dearth of in - depth research on the influence of the continuous change pattern of platelets after sepsis patients are admitted to the hospital on clinical outcomes. Therefore, this study aimed to utilize the group trajectory model to perform dynamic typing of platelets and explore its potential mediating factors.

In this study, the Group-Based Trajectory Model (GBTM) was utilized. By analyzing the platelet counts of patients within the first three days after their admission to the Intensive Care Unit (ICU), three distinct dynamic evolution patterns of platelets were successfully identified. After comprehensive analysis, the optimal model was determined to consist of three groups of trajectories: The first group was the “persistent low - level type” (n = 201, accounting for 71.79%). Its characteristic feature was that the platelet count remained at a relatively low level. Although it decreased gradually and then rebounded, the overall trajectory was relatively smooth with minimal fluctuations. The second group was the “high - level decline type” (n = 56, accounting for 20.00%). Initially, the platelet level was high, and subsequently, it exhibited a continuous downward trend. The third group was the “rebound and rise type” (n = 23, accounting for 8.21%). It was characterized by a rapid rebound of the platelet count following an initial decline. One of the significant contributions of this study was that it captured the dynamic evolution of platelets. In comparison to a single static measurement, the dynamic assessment of platelets might offer more advantages. Continuous dynamic assessment of platelets could more sensitively identify high - risk groups that were easily overlooked by traditional indicators, thereby providing a new theoretical basis for early and active clinical interventions targeting this group. After fully adjusting for potential confounding factors, the multivariate Cox regression analysis revealed that, when compared with the “persistent low - level group” which had the highest mortality rate, the 28 - day death risks of the “high - level decline group” and the “rebound and rise group” were significantly reduced by 42% and 93% respectively. This result strongly suggested that, apart from the static platelet count, the dynamic change trend of platelets themselves was an independent factor influencing the prognosis of sepsis. The findings of this study were in line with previous literature reports. Although there were disparities in the optimal number of groups determined according to different cohorts and models, multiple studies had arrived at a common conclusion: patients with an upward trend in platelet count had a higher survival rate, while those with a dynamic trajectory of persistently low - level platelets faced a significantly higher risk of death. This discovery further validated the crucial position of this phenotype in risk stratification (24–26).

Analysis of the clinical characteristics across the three groups revealed that patients in the persistent low-level group exhibited significantly higher SOFA scores, along with the highest levels of total bilirubin, creatinine, and blood lactate, indicating a more severe clinical condition characterized by pronounced organ dysfunction and microcirculatory impairment. In our 28-day follow-up analysis, distinct platelet trajectory phenotypes were significantly associated with survival outcomes (p < 0.01), whereas baseline WBC counts showed no significant variation across these phenotype groups (p = 0.27). Moreover, leukocyte-related variables, including differential counts and inflammatory markers, were simultaneously adjusted for in the fully multivariable model, and platelet trajectories remained independently associated with 28-day mortality. These findings suggest that platelet dynamics capture prognostic information complementary to leukocyte-based inflammatory indices rather than merely reflecting concurrent leukocyte states. WBC counts primarily reflect the instantaneous intensity of acute inflammatory responses and are highly susceptible to short-term fluctuations. In contrast, longitudinal changes in platelet counts integrate multiple processes involved in sepsis—including endothelial injury, coagulation activation, and impaired hematopoiesis— and therefore better reflect the cumulative physiological burden of disease over time. Consistent with this interpretation, the strong association between platelet trajectory dynamics and organ dysfunction observed in our study aligns closely with established pathological mechanisms reported in prior research. Vardon-Bounes et al. demonstrated that in murine models of sepsis, platelets are rapidly activated and accumulate within organ microvessels, forming platelet-rich microthrombi that lead to microvascular occlusion and tissue ischemia. Furthermore, activated platelets release substantial quantities of inflammatory mediators and procoagulant factors, amplifying systemic inflammation and coagulopathy, thereby contributing to multi-organ failure (27). Notably, the interaction between platelets and neutrophil extracellular traps (NETs) represents a pivotal mechanism exacerbating organ injury and microcirculatory dysfunction. Studies by McDonald and colleagues have shown that NETs are extensively formed in the microvasculature during sepsis, where they effectively activate thrombin and promote platelet aggregation, accelerating the development of platelet-rich microthrombi and resulting in impaired microcirculatory flow and inadequate organ perfusion. Platelets not only form direct aggregates with neutrophils but also bind to the DNA scaffold and protein components of NETs, further propagating thrombus formation. Moreover, NET-derived molecules such as histone H4 and polyphosphate act synergistically to enhance thrombin generation and platelet activation, establishing a self-amplifying “inflammation-thrombosis” loop that drives widespread microvascular dysfunction and end-organ damage. Importantly, experimental inhibition of NET formation or disruption of their structural integrity has been shown to markedly attenuate disseminated intravascular coagulation, restore microcirculatory perfusion, and mitigate organ dysfunction (28).

Although there were alternating changes in some coagulation indicators between the high-level decline group and the rebound increase group, the continuous low-level group showed the most significant abnormalities in multiple key indicators. The minimum platelet count (minPLT), activated partial thromboplastin time (APTT), and D-dimer level of this group were 83.46 ± 43.36 × 10⁹/L, 43.2 (36.5–50.1) s, and 4.22 (2.44–8.59) μg/ml, respectively, which were the highest among the three groups. This suggested the most severe coagulation dysfunction. At the same time, the procalcitonin (PCT) level, which reflected the severity of systemic inflammation and infection, also showed a similar inter-group distribution trend, further supporting that the patients in this group were in a more severe pathological physiological state. Previous study had shown that the duration of thrombocytopenia is closely related to poor prognosis (29). This finding was highly consistent with the dynamic trajectory of platelets in the continuous low-level group. This phenotype essentially reflected the imbalance in immune thrombosis regulation, characterized by excessive activation of the coagulation system (such as the significantly elevated D-dimer and prolonged APTT in this group) and uncontrolled inflammatory response (such as the significantly increased PCT) (8). These pathological processes drove microcirculation failure and multi-organ damage through mechanisms such as enhanced platelet-neutrophil interaction and extensive microthrombosis, thereby explaining the poorest clinical prognosis of this group of patients.

In the cohort comparison of patients with different prognosis, this study found that many key clinical indicators were significantly increased in patients in the non-survival group. Specifically, the coagulation function parameters (MPV, APTT, PT, FIB, D-dimer), infection and inflammation markers (procalcitonin), and tissue perfusion indexes (lactic acid) were all at high levels, and their platelet counts were difficult to return to baseline levels. At the same time, liver and kidney function indicators (AST, creatinine, urea nitrogen, etc.) reflecting organ dysfunction also showed the same trend. This coexistence of multiple critical markers strongly suggests that this type of patient may be trapped in a vicious cycle driven by “immune thrombosis”, in which the uncontrolled inflammatory response, diffuse microthrombosis, and the consequent organ ischemia combine to lead to the final adverse outcome. The platelet trajectory model constructed in this study just captures the overall impact of this vicious circle from the dynamic dimension - the persistent low-level platelet state is a comprehensive embodiment of the persistence and difficulty of reversing this pathological process.

We further performed detailed subgroup analyses based on the patient's baseline (age, sex, and comorbidities) and disease severity (SOFA, APACHE II, GCS score). The results showed that the association between platelet trajectories and mortality risk was particularly prominent in patients older than 75 years, females, APACHE II > 20 points, SOFA > 8 points, and GCS scores ≤ 12 points. In other subgroups, although the risk ratios of some groups were not statistically significant compared with the reference group, the risks were consistent from high to low, showing the trend of “sustained low level group> high level decline group > rebound rising group”. In addition, no significant interaction between the above factors and platelet trajectory grouping was detected. This result suggested that the prognostic value of platelet trajectories may be more clinically important in vulnerable patient groups with more severe disease and limited physiological reserves. And together, they confirmed the robustness of the association between platelet trajectories and prognosis. The consistent risk trend across subgroups and the absence of significant interactions suggested that the association might be a universal pattern that is not easily modified by the effects of common confounders such as age, gender, underlying disease, and initial disease severity. This provided a basis for generalizing the model to a wider population of sepsis. Our mediation analysis revealed that the degree of organ dysfunction represented by the SOFA score was an important mediating pathway connecting initial platelets to the risk of death. This finding was consistent with previous studies that platelets cause organ dysfunction during sepsis through a variety of mechanisms, including activation and aggregation to form blood clots, interaction with immune cells to form “danger alliances”, induction of NET formation, release of inflammatory mediators and procoagulants, and circulatory disorders (4). These mechanisms work together to eventually triggered microvascular dysfunction and organ damage. From a clinical point of view, the effect of prolonged coagulation time on prognosis might not be simply linear, but showed a “multiplication” effect, that was, when APTT breaks through a certain critical point, the risk of death associated with its small increase will increase sharply. This finding strongly suggested that the use of log (APTT) might be more risk-differentiated than the original value in future prognostic models or clinical decisions.

This study classified the dynamics of platelets in sepsis patients using a population trajectory model. Although the use of growth mixture models to analyze longitudinal data is not novel, the core value of this study lies in moving beyond methodological repetition, providing new clinical evidence and deeper mechanistic insights. First, we independently validated the strong association between platelet dynamic phenotypes and outcomes in a specific critical care cohort, offering new support for the generalizability of this theory across different populations. More importantly, the unique trajectory identified by us, termed “rebound rise,” provides a dynamic phenotype with a positive prognostic significance, enriching the understanding of the pathophysiological evolution in such patients. Furthermore, this study did not stop at classification itself but, through systematic mediation analysis, revealed the specific mediating role of organ dysfunction (SOFA score) and nonlinear coagulation activation (log (APTT)) in linking initial platelet dynamics to mortality risk. This finding advances the association from mere phenotype correlation to potential causal mechanisms, offering a new perspective on the complex role of coagulation parameters in sepsis. Therefore, the core innovation of this study does not lie in the first application of the method but in systematically addressing the progressive questions of “what are the characteristics of the phenotypes” and “what are the potential underlying mechanisms” using this method, accumulating richer evidence for the heterogeneity and precise prognostic judgment of sepsis.

In addition to circulating biomarkers, non-invasive imaging techniques are increasingly recognized as valuable tools for functional assessment and prognostic stratification in critical illness. Advanced MRI-based approaches, including hyperpolarized gas imaging and functional ventilation–perfusion assessment, enable real-time visualization of regional lung function and tissue perfusion without radiation exposure. These modalities provide spatial and physiological information that cannot be captured by blood-based indicators alone. Recent studies have demonstrated the feasibility of hyperpolarized gas MRI for quantitative evaluation of pulmonary ventilation and gas exchange, highlighting its potential role in risk prediction and disease monitoring (30, 31). Therefore, integrating imaging-derived functional metrics with laboratory trajectories may further enhance personalized prognosis assessment in critically ill patients.

Several limitations of this study should be acknowledged. First, study design and generalizability: This single-center observational study had a limited sample size—especially in the “rebound up” group—limiting generalizability. Selection and information biases are inherent to this design and cannot be fully eliminated. Although rigorous statistical modeling adjusted for many known confounders, residual confounding remains possible. Key clinical factors—including pathogen type, antimicrobial timing/selection, and vasoactive agent dosing—were not fully captured and may affect both platelet dynamics and outcomes. In this study, we did not perform statistical analyses on the pathogenic agents (such as bacteria and viruses) in sepsis patients. As is widely recognized, viral pathogens can modulate the host immune response via mechanisms like immune dysregulation, endothelial damage, and coagulation abnormalities, which might influence the severity and prognosis of the disease. In the future, stratified analysis of pathogens should be conducted to further investigate the differences in the prognostic impacts between bacterial sepsis and viral sepsis. Second, data quality and trajectory estimation. Modeling population trajectories involves subjective decisions—especially in selecting the optimal number of trajectory groups, which requires balancing statistical fit with clinical interpretability. All lab variables (e.g., platelet counts) came from routine clinical tests and thus carry inherent measurement error. Sampling times followed clinical practice—not standardized research protocols—potentially adding noise to trajectory identification. Third, limitations in analytical and causal inference. Mediation analysis offers insights for mechanistic exploration, but observational data cannot establish definitive causality—such studies are better suited for generating hypotheses than testing them. While variable selection and log (APTT) transformation are justified biologically and statistically, this mediation model remains just one plausible approach. Furthermore, “persistently low” platelet trajectory was strongly linked to higher admission SOFA scores and mortality. However, because severe organ dysfunction can itself cause ongoing platelet consumption and suppressed production, this trajectory may reflect underlying disease severity rather than independently drive poor outcomes. Its independent prognostic value—beyond initial disease severity—requires further validation in rigorously designed studies. Fourth, GBTM identifies temporal patterns—not mechanism-based subphenotypes—summarizing common trends statistically, not as discrete biological entities. Relying solely on platelet count is oversimplified: it ignores platelet function, production kinetics (e.g., reticulated platelets), and interactions with endothelial or coagulation pathways. Thus, the resulting phenotypes are descriptive only—useful for hypothesis generation, not comprehensive biological characterization. Although we applied established criteria to select clinically interpretable trajectories, results remain model - dependent. Future work should use higher-resolution longitudinal sampling, sensitivity analyses under alternative distributional assumptions, more flexible frameworks (e.g., growth mixture models), and multi-omics or systems-level approaches to validate trajectory stability and uncover underlying biology. Finally, trajectory construction used only platelet counts from the first three days after admission. While this window captures early disease dynamics, it may miss later platelet changes under treatment. Therefore, the identified trajectories represent an early acute-phase phenotype and require validation—using longer follow-up and denser sampling—for predictive stability and long-term applicability.

However, these limitations have not detracted from the distinct contribution of this study. Our study incorporated repeated longitudinal measurements and applied trajectory modeling to characterize temporal phenotypes, thereby identifying clinically meaningful patterns of platelet dynamics rather than relying solely on single baseline values. This longitudinal phenotyping framework may better reflect underlying pathophysiological heterogeneity and provide additional insights into disease mechanisms and prognostic pathways. By integrating dynamic trajectories with clinical outcomes and mechanistic exploration, we established a systematic analytical framework that offers a complementary perspective for understanding sepsis heterogeneity. The identified mediating effect and nonlinear relationship of log (APTT) have delineated the direction for future research. Although external validation in larger multi-center studies is warranted, this dynamic phenotypic classification strategy is anticipated to offer novel tools for more precise prognostic assessment and patient stratification in clinical trials.

This study successfully identified three clinical phenotypes with marked differences in prognosis by modeling and analyzing the early admission dynamic trajectories of platelet counts in sepsis patients. The study verified that the platelet dynamic trajectories based on continuous measurements can effectively identify high - risk populations. Among them, the “persistently low - level” phenotype is strongly associated with the highest degree of organ dysfunction and death risk. The mediating analysis further disclosed the possible mediating role of organ dysfunction and non - linear changes in coagulation parameters. Despite the constraints of this single - center retrospective study, it provides a novel dynamic perspective for the risk stratification of sepsis. The constructed trajectory model is expected to become a practical instrument for clinical prognostic assessment and has also indicated an important direction for subsequent mechanism research.