This was a single-center, retrospective, observational cohort study approved by the University of Florida Health (UF Health) Institutional Review Board prior to the collection of any data. An integrated data repository (IDR) of all clinical data (structured and unstructured) in the EHR for all infants admitted to the UF Health, level IV NICU between January 1st 2012 to March 1st 2020 was created. Clinical variables that were unavailable in the IDR were collected via chart review. Demographic variables and outcomes were defined as previously reported (12–19). Maximum and hourly neonatal sequential organ failure assessment (nSOFA) scores and vasoactive-inotropic scores (VIS) were calculated as previously described (13, 19, 20).

An episode of sepsis was defined as a new episode of bacteremia/fungemia or meningitis with bacteria or fungus that was associated with ≥5 days of antimicrobial treatment or mortality while receiving antimicrobial treatment if duration was <5 days. Episodes were excluded if there was parenteral antimicrobial exposure within 48 h of the time of the reference culture. Coagulase negative Staphylococcus (CoNS) episodes were included only if there was effective antimicrobial therapy prescribed for ≥5 days or death <5 days. Infants with positive blood cultures with organisms that are frequently considered contaminants, including Bacillus, Micrococcus, and Corynebacterium species, were excluded.

A coagulation evaluation was included if it included simultaneous measurement of at least two parameters of the following three tests: fibrinogen, PT, and PTT. Reference values measured at discrete chronologic ages where testing has been performed in this population vary by gestational age at birth and chronological age (21–24) (Supplementary Table S1). Because daily normative values for these tests from birth are not available for this population, we applied normative ranges for each coagulation measure by gestational age at birth and discrete postnatal age in a fill-forward manner (until the next tested interval where normative values have been described).

Non-parametric continuous variables were summarized as medians with quartiles (25th and 75th percentiles). Categorical variables were presented as percentages. Demographic variables were compared using Chi-square or Mann–Whitney U test. The nSOFA scores at discrete time intervals between multiple groups were compared using a Kruskal–Wallis test with Dunn's multiple comparisons test. Shapley Additive Explanation (SHAP) values were used to determine which variables were most important in regression modeling for patients who received treatment vs. no treatment for coagulopathy. Machine learning modeling was performed for specific variables and composite outcome for death and/or treatment to prevent over-fitting. The SHAP value is the average marginal contribution of a feature value across all possible coalitions. The threshold for statistical significance was less than 0.05 for two-sided tests. Analyses were performed using GraphPad Prism (version 10) and Python (Version 3.7; SHAP).

Among 7,186 patients cared for in the UF NICU during the study period, we identified 910 positive blood or cerebrospinal fluid cultures in 367 patients. After applying our sepsis definition, 225 sepsis episodes in 216 patients were studied (Table 1). Episodes associated with coagulation evaluation (n = 96) occurred at an earlier age (13 vs. 23 days old), were more often associated with gram-negative (52% vs. 20%) and less often gram-positive (38% vs. 74%) pathogens, and showed greater nSOFA scores and greater mortality as compared with episodes without coagulation evaluation.

Among the 225 episodes studied, 220 were associated with bacteremia and 5 were episodes of meningitis without bacteremia. Gram-positive organisms were most common (132/225; 59%), followed by gram-negative organisms (76/225; 34%), polymicrobial infections (14/225; 6%), and minimal fungus (3/225; 1%). The most common single pathogens recovered were CoNS (68/225; 30%), Escherichia coli (34/225; 15%), Staphylococcus aureus (23/225; 10%), Group B Streptococcus (20/225; 9%), and Enterococcus faecalis (15/225; 7%). Antimicrobial resistance was common and did not differ between episodes with and without coagulopathy evaluation by pathogen class (Supplementary Table S2). Among non-CoNS episodes, 61/150 (41%) were associated with pathogens resistant to one or more common antimicrobial interventions (ampicillin, penicillin + beta-lactamase inhibitor, gentamicin, oxacillin, and third generation cephalosporin). Among episodes with recovery of non-CoNS pathogens, resistance to these common antimicrobial interventions was more common among those with a coagulopathy evaluation as compared with those without an evaluation [36/76 (47%) vs. 25/74 (34%)], but this did not reach significance (p = 0.10).

The frequency of any coagulation assessment [PT, International normalized ratio (INR), fibrinogen, PTT] at any time among all patients cared for in the UF NICU during the study period was 23% (1,654/7,186). Coagulation evaluations during sepsis events were performed at the discretion of the clinical team. Reasons for coagulation evaluation were abstracted by chart review and included concerns for or documentation of bleeding (n = 39), thrombocytopenia (n = 19), sepsis (n = 19), preoperative or unit protocol (e.g., hypoxic-ischemic encephalopathy) (n = 15), or abnormal liver function tests. An analysis of available measures for study revealed that greater than 90% of sepsis episodes were associated with ≤10 total coagulation evaluations and a median of 2 evaluations per episode (IQR 1, 5). Therefore, we restricted our analyses to the first 10 coagulation measures for each episode of sepsis, which yielded 339 episode-specific coagulation evaluations performed during 96 sepsis episodes (Table 2). The median time of the first coagulation evaluation relative to the reference culture was 14 h (IQR 4, 45) and at a median age of 14 days (IQR 6, 40). Abnormal coagulation evaluation results were similarly common in the first testing of the episode (78/96, 81%) and among all testing during the episode (271/339, 80%). The most common abnormal parameter was PT (230/336, 68%) and the median time to any abnormal result was 43 h (IQR 17, 111). An abnormal coagulation evaluation result during the episode was very common and occurred in 86/96 (90%). Episode mortality among those that had a coagulation evaluation with any abnormal result was 18/86 (21%).

Among episodes where a coagulation evaluation was performed, 40/96 (42%) were associated with an intervention (transfusion of fresh frozen plasma or cryoprecipitate). Among all coagulation evaluations, 84/271 (31%) abnormal coagulation evaluation test results were associated with subsequent intervention and 12/68 (18%) normal test results were associated with intervention (Table 3). The median timing of the first intervention was 14 h after the reference culture (IQR 4, 38). Interventions primarily occurred following the first (24/40, 60%) or second (10/40, 25%) coagulation evaluation conducted during the episode. Most episodes (29/40, 73%) were associated with >1 intervention. Repeat testing was common. Among all normal test results, 36/68 (53%) were repeated. Among abnormal test results, 214/271 (79%) were repeated.

Among the 56 episodes where a coagulation evaluation was performed without associated intervention, 34 (61%) were associated with mild thrombocytopenia (<150,000/µL), 29 (52%) with moderate thrombocytopenia (<100,000/µL), and 14 (25%) with severe thrombocytopenia (<50,000/µL) (Table 3). Among the 40 episodes where coagulation evaluation was performed with associated intervention, 37 (93%) were associated with mild thrombocytopenia, 29 (73%) with moderate thrombocytopenia, and 19 (48%) with severe thrombocytopenia. Out of the 96 episodes where a coagulation evaluation was performed (regardless of intervention), 48 episodes were associated with ≥1 platelet transfusions (27/40 (68%) episodes with intervention; 21/56 (38%) episodes without intervention).

We examined nSOFA scores during episodes with (lived, n = 77; died n = 19) and without (lived, n = 125) a coagulation evaluation at hourly resolution (Figure 1A) and compared scores at multiple discrete time points (Figure 1B) relative to the reference blood culture. Median nSOFA scores were 0 for survivors that did not have a coagulation evaluation at all nine discrete time points we evaluated (IQR for all nine time points: 0, 1). Survivors that had a coagulation evaluation also demonstrated a median nSOFA of 0 prior to evaluation (range of all IQR: 0, 2) but increased to 2 at evaluation (IQR 0, 4) and remained 2 at all future timepoints (range IQR 0–1, 4). Episodes with a coagulation evaluation that ended in death demonstrated a median nSOFA of 3 prior to evaluation (range IQR 0–2, 3–6) that increased to 4 (IQR 2, 8) at evaluation, 6 (IQR 3, 9) at 6 h, 8 (IQR 3, 10) at 12 h, 10 (IQR 6, 12) at 18 h, and 8 (IQR 5, 12) at 24 h.

Logistic regression modeling and SHAP analysis were conducted to evaluate the significance of clinical variables and abnormal coagulation values in predicting outcomes (Supplementary Table S3 and Figure 2). The SHAP modeling demonstrated strong predictive performance for the composite outcome of death and/or treatment for coagulopathy in neonates (f1 score 0.8, area under receiver operating characteristic curve 0.83 for those with abnormal coagulation values). The three most important features influencing death and/or treatment for coagulopathy included administration of vasoactive medications, quantified by the VIS, hematologic dysfunction assessed by the maximum nSOFA platelet score, and early sepsis (within ≤72 h after birth).