📄 medicine

Coagulation Trajectory Subphenotypes Identify High-Risk Sepsis Patients Missed by the Sepsis-Induced Coagulopathy Score: A Retrospective Cohort Study with External Validation

This retrospective cohort study demonstrates that analyzing 72-hour trajectories of routine coagulation markers identifies a high-mortality sepsis subphenotype missed by the static Sepsis-Induced Coagulopathy (SIC) score in over 80% of cases, a finding validated in an external dataset.

Original authors: Xueshu Yu, Wen Xu, Yincai Ye, Luwei Xu, Jingye Pan

Published 2026-10-22

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Xueshu Yu, Wen Xu, Yincai Ye, Luwei Xu, Jingye Pan

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Sepsis is a life-threatening reaction to an infection, a condition where the body's own defenses turn against its organs. In the cascade of organ failure that follows, the blood's clotting system plays a central and dangerous role. Normally, clotting stops bleeding, but in sepsis, this system can go haywire, creating tiny clots that block blood flow while simultaneously using up the body's clotting factors, leaving the patient vulnerable to bleeding. Doctors have long relied on a specific scoring system, known as the Sepsis-Induced Coagulopathy score, to identify patients whose blood is in this dangerous state. This score acts like a snapshot, taking a single measurement of blood platelets and clotting times at the moment a patient arrives at the intensive care unit to decide if they are at risk. The logic has been that if the numbers look bad right then, the patient needs attention; if they look okay, the patient is likely safe.

However, a new study suggests that this single moment in time might be missing a crucial part of the story. Just as a photograph cannot capture the speed of a car or the direction it is turning, a single blood test cannot show whether a patient's condition is stabilizing or rapidly deteriorating. Researchers wondered if looking at how blood clotting numbers change over the first three days of a hospital stay could reveal a group of high-risk patients who appear safe at first glance. By tracking the movement of these numbers rather than just their starting point, the team hoped to find a hidden pattern of danger that current tools overlook.

To investigate this, a team of researchers analyzed data from thousands of adult patients with sepsis who had spent at least three days in an intensive care unit. They focused on routine blood tests that are already standard in hospitals: platelet counts, which help blood clot, and two measures of how long it takes blood to clot. Instead of looking at these numbers only when the patient was admitted, the team mapped out their values every eight hours for the first seventy-two hours. They used a computer method to group patients based on the shape of these lines over time, looking for distinct patterns in how the numbers rose or fell. This approach allowed them to see the story of the patient's blood clotting system as it unfolded, rather than just a single frame.

The analysis revealed three distinct groups of patients. The largest group, which the researchers called "Low," showed relatively stable blood clotting numbers and had a lower rate of death. A second group, "Moderate," showed slight changes but had a death rate very similar to the first group. The third group, labeled "High," told a different story. These patients did not necessarily have the worst numbers when they walked into the hospital, but their blood clotting markers showed a persistent pattern of elevation over the first three days. Their blood took longer to clot, and this condition did not improve. This group faced a significantly higher risk of death, with nearly twenty-eight percent of them dying in the hospital, compared to sixteen percent in the low-risk group.

The most striking finding was that the standard screening tool completely failed to catch most of these high-risk patients. The researchers found that eighty-two percent of the patients in the "High" group did not meet the criteria for the standard sepsis coagulation score. According to the existing rules, these patients were considered low risk because their initial blood tests were not bad enough to trigger an alarm. Yet, their mortality rate was nearly double that of the low-risk patients who were correctly identified as safe. In fact, the standard score only flagged about seven percent of all patients as high risk, missing the vast majority of those whose blood clotting systems were quietly failing over time.

To ensure this was not just a fluke of one hospital's data, the researchers tested their findings on a separate database containing patients from over two hundred hospitals across the United States. The results held true. Even in this diverse group, the patients whose blood clotting markers stayed high over three days faced a much higher risk of death, and the standard score continued to miss them. This confirmed that the problem was not with the specific data set but with the method of looking at only a single moment in time. The study suggests that the direction of change in blood tests carries vital information that a snapshot cannot capture.

The researchers also looked at whether treating these patients with blood thinners, a common intervention, made a difference. They found that patients in the high-risk group who received heparin, a blood-thinning medication, appeared to have better survival rates compared to those who did not. However, the researchers were careful to note that this was an observation from past data, not a proven result from a controlled experiment. The link between the treatment and the outcome was strong in the high-risk group but not in the others, suggesting that this specific group of patients might be the ones most likely to benefit from such therapy, though this would need to be confirmed in future clinical trials.

This work does not propose a new machine or a new test. The blood tests used are already part of routine care, and the data is already being collected. The change is in how doctors interpret that data. The study argues that the current practice of checking a patient's blood clotting status once and moving on is insufficient for identifying everyone at risk. By paying attention to the trend of these numbers over the first three days, clinicians could identify a large group of patients who are currently slipping through the cracks. The researchers conclude that future guidelines for sepsis should incorporate the direction and speed of change in these blood markers, turning a static checklist into a dynamic process that reflects the reality of how the disease progresses.

Technical Summary: Coagulation Trajectory Subphenotypes in Sepsis

Problem Statement
Current tools for assessing sepsis-associated coagulopathy, such as the Sepsis-Induced Coagulopathy (SIC) score, rely on static, single-time-point laboratory assessments (e.g., admission platelet count, INR, and SOFA subscores). The authors argue that these "snapshot" evaluations fail to capture the dynamic evolution of coagulation profiles during the critical first days of sepsis. Consequently, patients whose coagulation status is deteriorating (e.g., rising INR or falling platelets over time) may be misclassified as low-risk if their baseline values at admission do not yet meet the threshold for coagulopathy. While machine learning models have been developed to predict SIC status or mortality, they often treat the SIC score as a ground truth label rather than investigating whether the SIC construct itself systematically misses high-risk subgroups defined by the trajectory of serial measurements.

Methodology
This retrospective cohort study utilized two large, de-identified critical care databases: MIMIC-IV (derivation cohort, 2008–2022) and eICU-CRD (external validation, 2014–2015).

Population: Adults (≥18 years) meeting Sepsis-3 criteria with a first ICU stay of at least 72 hours. Patients with liver disease, metastatic cancer, traumatic brain injury, or major admission bleeding were excluded to isolate sepsis-driven coagulopathy.
Data Extraction: Routine coagulation markers (platelet count, INR, PTT) were extracted for the first 72 hours of ICU admission. The timeline was divided into nine 8-hour windows. Within each window, the platelet nadir and INR/PTT peaks were recorded, resulting in 27 features per patient.
Clustering: A Gaussian Mixture Model (GMM) with diagonal covariance was applied to the 27 features to identify unsupervised trajectory subphenotypes. The optimal number of clusters ( $k$ ) was determined via Bayesian Information Criterion (BIC) and clinical interpretability, resulting in a 3-cluster solution.
Validation: The eICU-CRD cohort underwent independent refitting of the GMM (rather than direct model transport) to account for differences in laboratory density and timing between databases.
Comparative Analysis: The study compared the GMM-derived subphenotypes against the standard SIC score (defined as a total score ≥4 with a platelet-plus-INR subtotal >2). Outcomes included hospital mortality, with adjustments for SOFA scores and age.
Treatment Effect Analysis: A target trial emulation using inverse probability weighting and causal forest analysis was conducted to explore the association between heparin use (within 72 hours) and 28-day mortality across the identified subphenotypes.

Key Results

Subphenotype Identification: The GMM identified three reproducible subphenotypes: Low ( $n=1,814$ $n = 1, 814$ ), Moderate ( $n=633$ $n = 633$ ), and High ( $n=691$ $n = 691$ ).
- Mortality Gradient: Hospital mortality was 16.2% (Low), 17.7% (Moderate), and 27.8% (High) ( $p < 0.001$ ). The High group was characterized by persistently elevated INR and PTT ratios, whereas platelet trajectories overlapped significantly across groups.
- Independence from Static Scores: After adjusting for SOFA scores, the High group had a significantly higher hazard ratio for mortality compared to the Low group (HR 1.62, 95% CI 1.35–1.96).
The "SIC Blind Spot": The study found that the SIC score failed to flag the majority of high-risk patients identified by trajectory analysis.
- 81.9% of patients in the High-mortality GMM group (566/691) were classified as SIC-negative at admission.
- Despite being SIC-negative, these patients had a mortality rate of 26.3%, which was 1.7 times higher than SIC-negative patients in the Low group (15.8%).
- Among all SIC-negative patients, a significant mortality gradient persisted (15.8% to 26.3%, $p < 0.001$ ).
External Validation: The independent refitting in eICU-CRD ( $n=12,828$ ) reproduced the mortality gradient (Low 15.6%, Moderate 21.5%, High 24.7%) and the SIC blind spot (86.3% of High-group patients were SIC-negative).
Predictive Performance: Adding the trajectory cluster to a model containing SOFA and age improved the AUC modestly (from 0.645 to 0.656). The GMM approach (grouping Moderate + High as positive) offered higher sensitivity (50.9%) compared to the SIC score (11.2%) but with lower specificity.
Heparin Association: In the High subphenotype, heparin use within 72 hours was associated with lower 28-day mortality (adjusted HR 0.67). However, the formal interaction test between treatment and subphenotype was not statistically significant ( $p=0.49$ ), suggesting the finding is an effect gradient rather than definitive proof of differential benefit.

Significance and Claims
The authors claim that the primary contribution of this work is the identification of a reproducible high-mortality subphenotype that is systematically missed by static screening instruments like the SIC score.

Clinical Implication: The study suggests that serial coagulation measurements, which are already routinely collected in ICU flowsheets, should be interpreted as a dynamic process rather than a static snapshot. Specifically, the direction of change (e.g., persistently elevated INR/PTT) carries prognostic information invisible to admission-only scoring.
Limitations of Current Tools: The paper posits that the SIC score has a structural blind spot for patients whose risk emerges through the trajectory of serial measurements rather than baseline values.
Future Directions: The authors propose that trajectory-defined subgroups could serve as an enrichment stratum for future prospective trials of anticoagulation in sepsis. They explicitly state that their findings warrant prospective evaluation to determine if incorporating the direction of change into coagulopathy screening improves patient outcomes.
Modesty: The authors acknowledge that the discrimination improvement (ΔAUC) is modest and that the GMM does not currently yield a simple, deployable bedside rule. They emphasize that the Moderate group was statistically indistinguishable from the Low group in terms of mortality, suggesting the clinically actionable partition is effectively "High vs. Non-High." They also note that the heparin findings are observational and hypothesis-generating, not causal.

Technical Summary: Coagulation Trajectory Subphenotypes in Sepsis

More like this