researchsquare.comSource publication:
Loading content…
Sepsis coagulopathy screening still rests on a single admission lab set, with platelets, INR and SOFA combined into one cross-sectional threshold. Among 3,138 sepsis patients in MIMIC-IV, clustering the joint 72-hour trajectories of platelet count, INR and PTT yielded a high-risk subphenotype with 27.8% mortality, of whom the admission SIC score flagged only 18.1%. The gradient reproduced in an independent refit of 12,828 eICU-CRD patients, where 86.3% of the high group remained SIC-negative.
Sepsis patients can be split into subphenotypes with different death risk by how their coagulation labs move over the first 72 ICU hours, rather than by their values at admission alone. The SIC score, the ISTH overt DIC criteria and the SOFA coagulation component all read a single time point, so a patient with falling platelets or rising INR is judged the same as one with stable values when the two match at measurement. Among 3,138 sepsis patients in MIMIC-IV, clustering produced low, moderate and high groups with hospital mortality of 16.2%, 17.7% and 27.8% (p<0.001); after SOFA adjustment the high group's hazard ratio versus low was 1.62 (95% CI 1.35–1.96).
Among patients the SIC score calls negative, the trajectories still separate those at higher risk. SIC is positive at a total score of at least 4 with a platelet-plus-INR subtotal above 2, which covered only 7.3% of patients and left the other 92.7% indistinguishable by the score. Among 2,908 SIC-negative patients, mortality across low, moderate and high was 15.8%, 17.7% and 26.3% (p<0.001); 566 of the 691 high-group patients (81.9%) were SIC-negative, and their 26.3% mortality was 1.7 times the 15.8% of SIC-negative low patients.
The method lays routine coagulation labs out as a time series and lets the data form the groups. Each parameter is cut into nine 8-hour windows, taking the platelet nadir and the INR and PTT peaks, giving 27 features fitted with a diagonal-covariance Gaussian mixture model and labeled low, moderate and high by ascending mortality. Assignments had a mean posterior probability of 0.998, with 99.7% of patients at or above 0.7 and an adjusted Rand index of 0.998–1.0; separation came mainly from INR and PTT, since 72-hour platelet nadirs were 171–192 ×10⁹/L in all three groups.
The same gradient reappears independently in a second, multicenter database. Transporting the MIMIC scaler and parameters directly to eICU-CRD gave discrimination near chance (C≈0.50), so the model was refitted independently in eICU-CRD instead. Among 12,828 eICU-CRD patients, mortality across low, moderate and high was 15.6%, 21.5% and 24.7% (p<0.001); 2,722 of the 3,153 high-group patients (86.3%) were SIC-negative, with 23.4% mortality.
The findings apply to adults with sepsis admitted to an ICU and surviving at least 72 hours, after excluding liver disease, metastatic cancer, traumatic brain injury and major admission bleeding; derivation came from single-center MIMIC-IV and external validation was an independent refit in eICU-CRD.
Requiring at least 72 ICU hours excludes early deaths and discharges, so the findings cover only patients who survive the first 72 hours; retrospective data may carry unmeasured confounding; bleeding was defined by ICD codes rather than chart confirmation; labs are not drawn on fixed schedules and missing windows were filled with cohort medians; the moderate group did not differ significantly from low in mortality (17.7% vs 16.2%, p=0.40), leaving high versus the rest as the usable partition; the heparin analysis is observational and the treatment-by-subphenotype interaction was not significant (p=0.49).