Development and validation of a predictive model for septic shock in maternal sepsis: a retrospective cohort study in a single referral center

Abstract

Objectives

Maternal sepsis carries a high risk of progression to septic shock. However, existing risk stratification tools, such as qSOFA and the Sepsis in Obstetrics Score, perform suboptimally in pregnant and postpartum populations, and no simple model specifically predicts progression from sepsis to septic shock. This study aimed to identify variables associated with septic shock in maternal sepsis, characterize infection sources and pathogens, and develop and internally validate a practical predictive model.

Methods

A retrospective cohort study was conducted on 139 pregnant and postpartum women with sepsis admitted to a single tertiary center between January 2016 and July 2025. The primary outcome was septic shock within 72 h of sepsis diagnosis, and predictors were collected at the time of diagnosis. Patients were randomly split into training (70%) and validation (30%) cohorts. Predictors were first screened using univariate robust Poisson regression, followed by LASSO regression for feature selection. The final predictive model was fitted using multivariable robust Poisson regression. Model performance was assessed using the area under the receiver-operating-characteristic curve, calibration plots, and decision curve analyses (DCA).

Results

Four independent predictors for septic shock were identified: history of surgery during pregnancy, prothrombin time (PT), procalcitonin (PCT), and arterial partial pressure of oxygen (PaO₂). The model exhibited good discrimination in the training cohort (Area-under-curve [AUC] 0.87, 95% CI: 0.80–0.94) and validation cohort (AUC 0.87, 95% CI: 0.75–0.99), with good calibration and clinical utility. Higher risk scores were significantly associated with adverse maternal and neonatal outcomes. Respiratory infections were the most common source (44.6%), followed by genitourinary (30.9%) and gastrointestinal infections (12.9%). Gram-negative bacteria (55.1%) were the main pathogens. Septic shock occurred more frequently in puerperal sepsis than in pregnancy-onset sepsis (43.3% vs. 19.0%, p = 0.003).

Conclusion

A four-variable model using readily available clinical parameters demonstrated favorable predictive performance for predicting septic shock in maternal sepsis and was associated with adverse perinatal outcomes. This tool may support early risk stratification and clinical decision-making, pending future prospective external validation.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12884-026-09381-x.

Introduction

Maternal sepsis is defined as a life-threatening organ dysfunction due to a dysregulated host response to infection that occurs during pregnancy or the postpartum period [1, 2]. Its progression to septic shock markedly increases mortality and morbidity, posing a major challenge in obstetric practice [1, 2]. Physiological adaptations of pregnancy may mask early organ dysfunction, while immune-metabolic changes can exacerbate the response to infection [3–6]. Furthermore, pregnant women are frequently excluded from sepsis trials, leading to limited evidence to guide management [7]. The Global Maternal Sepsis Study (GLOSS) reported approximately 13 severe adverse outcomes per 1,000 live births, highlighting the substantial disease burden [6, 8].

Existing sepsis screening tools, including the Sepsis in Obstetrics Score (SOS), quick SOFA (qSOFA), and Sequential Organ Failure Assessment (SOFA), have important limitations in obstetric populations. SOS demonstrated a positive predictive value of only 16.7% for ICU admission in the initial study, with a subsequent prospective validation revealing a sensitivity of 64% for the same outcome [9–11]. The qSOFA score, widely used in general medicine, performed even poorer in maternal sepsis, with a sensitivity of merely 50% [12, 13]. The SOFA score, another standard tool, is also problematic as normal pregnancy physiology alters key parameters such as serum creatinine and mean arterial pressure, leading to potential misclassification [10, 13–15].

A critical gap remains: current tools (qSOFA, SOS) are screening instruments for infection‑related deterioration, not risk‑stratification tools at sepsis diagnosis [16]. Their performance in obstetric populations is suboptimal, and they rely heavily on vital signs or subjective criteria, which may be altered by normal pregnancy physiology. Moreover, no simple model specifically predicts the short‑term progression from sepsis to septic shock in pregnant or postpartum women—a time window during which early recognition and intervention could prevent haemodynamic collapse [10, 12].

To address this gap, we conducted a retrospective cohort study with two objectives: 1) primary objective: identify clinical and laboratory variables available at the time of sepsis diagnosis that are associated with progression to septic shock within 72 h, develop and internally validate a practical predictive model for risk stratification in pregnant and postpartum women with sepsis. 2)Secondary objective: describe the infection sources and microbiological pathogen profiles in maternal sepsis, providing contextual data to support clinical decision‑making. Unlike existing screening scores, our model is intended for use after sepsis diagnosis, to refine risk assessment and guide triage of care intensity, resource allocation, and coordinated obstetric‑neonatal preparedness.

Materials and methods

Study population

This single-center retrospective cohort study was conducted at Renji Hospital, Shanghai Jiao Tong University School of Medicine, China—a tertiary care hospital with a dedicated obstetric critical care unit that manages approximately 5,500 deliveries annually and serves as a regional referral center for high-risk pregnancies. Eligible participants were pregnant or postpartum women (defined as within 42 days after delivery, including vaginal delivery and cesarean section) admitted to the hospital between January 2016 and July 2025, who met the Sepsis-3 criteria for sepsis: suspected or confirmed infection accompanied by an increase of ≥ 2 points in the Sequential Organ Failure Assessment (SOFA) score [17]. Exclusion criteria were: (1) incomplete medical records (e.g., missing key laboratory parameters or outcome data); (2) fetal anomaly.

Patients were stratified into two groups based on the primary outcome: septic shock group (development of septic shock within 72 h of sepsis diagnosis) and non-shock group (no septic shock during hospitalization). Septic shock was defined per Sepsis-3 criteria as persistent hypotension requiring vasopressor therapy to maintain a mean arterial pressure (MAP) ≥ 65 mmHg, plus a serum lactate level > 2 mmol/L (18 mg/dL) despite adequate fluid resuscitation (≥ 30 mL/kg crystalloid within 6 h) [18] Additionally, patients were categorized by sepsis onset timing: pregnancy-onset sepsis (sepsis occurring from conception to immediately before delivery) and puerperal sepsis (sepsis occurring from delivery to 42 days postpartum). Secondary analyses included maternal clinical outcomes and neonatal outcomes.

During the study period, all eligible patients meeting the inclusion criteria were consecutively identified through the hospital electronic medical record system. A total of 151 patients were screened, of whom 12 were excluded based on predefined exclusion criteria. The final analysis included 139 patients. Figure 1 shows the patient flow diagram. Given the retrospective design and the single-center setting, there is a potential risk of selection bias, particularly as our institution serves as a tertiary referral center for high-risk pregnancies.

Fig. 1.

Flowchart of patient inclusion and cohort allocation. Pregnant and postpartum women (within 42 days after delivery) who met the Sepsis-3 criteria for maternal sepsis between January 2016 and July 2025 were screened. After applying the exclusion criteria, 139 patients were included in the final analysis. Patients were categorized into two outcome groups based on whether septic shock occurred: septic shock group and non–septic shock group. The cohort was randomly split into a training cohort (70%) and a validation cohort (30%) for model development and internal validation

Variables and outcome measures

Data collected from medical records included maternal general information (i.e., age, body mass index [BMI], gestational age, gravidity, parity, and mode of conception, including in vitro fertilization–embryo transfer [IVF-ET]); clinical characteristics included pre-existing comorbidities, pregnancy- and delivery-related complications, surgical procedures during pregnancy, and infection sites. A history of surgery during pregnancy was defined as any non-obstetric surgical procedure performed after conception and before the onset of sepsis, including but not limited to appendectomy, cervical cerclage, and urologic interventions (e.g., ureteral stent placement). While these procedures differ in their indications and physiological impact, they were grouped a priori as a potential marker of increased vulnerability to infection or systemic insult during pregnancy. To address heterogeneity within this composite variable, a post-hoc sensitivity analysis was performed by categorizing surgeries into two groups: (1) surgery for potentially infectious causes (e.g., appendectomy) and (2) surgery for non-infectious causes (e.g., cervical cerclage). Associations with septic shock were then compared between groups (see Supplementary Material).

Laboratory parameters included complete blood count, liver and renal function tests, coagulation profile, arterial blood gas analysis, and microbiological culture results, all measured on the day of sepsis diagnosis. Maternal outcomes included length of hospitalization, need for mechanical ventilation, completion of sepsis bundle therapy, in-hospital mortality. Neonatal outcomes included gestational age at delivery, birth weight, Apgar score at 1 min, incidence of low birth weight [< 2500 g], neonatal asphyxia [Apgar score ≤ 7 at 1 min], stillbirth. Adverse perinatal outcomes were defined as follows: sepsis patients whose pregnancy was terminated before 36 weeks of gestation; neonates with a birth weight below 2500 g; or Apgar score ≤ 7. All data were collected at the time of diagnosis and independently reviewed by two investigators to ensure completeness and accuracy. Discrepancies were resolved through consensus with a third senior investigator. A complete-case analysis was performed, and patients with missing key variables were excluded.

Exploratory subgroup analyses were performed to evaluate the stability of model performance across clinically relevant subgroups, including maternal age (≥ 35 vs. <35 years), obesity status (BMI ≥ 30 vs. <30 kg/m²), mode of conception (IVF vs. spontaneous), parity (nulliparous vs. multiparous), and timing of sepsis onset (during pregnancy vs. puerperium). Model discrimination within each subgroup was assessed using the area under the receiver operating characteristic curve (AUC). These analyses were considered exploratory due to limited sample size and were not adjusted for multiple comparisons.

Statistical analysis

All statistical analyses were performed using R software (version 4.5.1; R Foundation for Statistical Computing, Vienna, Austria). Continuous variables with a normal distribution were presented as mean (standard deviation, SD), and non-normally distributed data as median (interquartile range, IQR). Categorical variables were summarized as counts (percentages). Missing values were imputed by median for continuous variables with missing less than 20%. Differences between the shock and non-shock groups were compared using the independent-samples t test or the Mann-Whitney U test for continuous variables, and the χ² test or Fisher’s exact test for categorical variables.

Log-transformation was applied to variables with highly skewed distributions (i.e., neutrophil-to-lymphocyte ratio, NLR) to improve normality, reduce the influence of extreme values, and enhance model stability. Since no zero values were present in these variables, no continuity correction was required before transformation. Univariate robust Poisson regression models with robust variance estimators were used to evaluate associations between potential predictors and septic shock, reporting the association with risk ratios (RRs).

Patients were randomly divided into a training cohort (70%) and a validation cohort (30%) using a random seed of 123, while maintaining a balanced proportion of septic shock cases between cohorts. Variables with a P value < 0.10 in the univariate analyses were fitted into least absolute shrinkage and selection operator (LASSO) regression for feature selection in the training cohort (Table S1). Variables retained by LASSO were then fitted into a multivariable robust Poisson regression model to derive the final predictive model. Internal validation and optimism correction were performed using bootstrap resampling with 2,000 repetitions. The final shock prediction model was defined based on the multivariable robust Poisson regression model. The shock risk score was calculated using the robust Poisson regression coefficients as follows: -4.88 + 0.53 * surgery (No = 0/ Yes = 1) + 0.06 * PT (s) + 0.004 * PCT (ng/ml) + 1.40 * PaO2(< 60 mmHg = 1/ ≥ 60 mmHg = 0).

Model discrimination was assessed using receiver operating characteristic (ROC) curve analysis and the area under the curve (AUC) in both the training and validation cohorts. The optimal cutoff value of the risk score was determined by maximizing the Youden index. Sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were subsequently calculated. Calibration performance was evaluated using calibration plots, and clinical utility was assessed by decision curve analysis (DCA).

To further evaluate the clinical relevance of the model, robust Poisson regression models were additionally used to assess the associations between the predicted risk score and maternal/neonatal outcomes, with results reported as risk ratios (RRs) and 95% confidence intervals (CIs), with or without the adjustment of age, BMI, IVF, first pregnancy, onset time, bundleTx, vp, ventilation, and steroid usage.

Exploratory subgroup analyses were performed according to maternal age (≥ 35 vs. <35 years), obesity status (BMI ≥ 30 vs. <30 kg/m²), mode of conception (IVF vs. spontaneous conception), parity (nulliparous vs. multiparous), and timing of sepsis onset (during pregnancy vs. puerperium). Model discrimination within each subgroup was assessed using AUC values. These analyses were considered exploratory and were not adjusted for multiple comparisons.

All statistical tests were two-sided, and P < 0.05 was considered statistically significant. The study was reported in alignment with TRIPOD guidelines (Table S6).

Results

Study population and baseline characteristics

Between January 2016 and July 2025, 139 pregnant or postpartum women with sepsis were enrolled in the study. Of these, 60 (43.2%) progressed to septic shock within 72 h of sepsis onset. Table 1 summarizes the baseline demographic, clinical, and laboratory characteristics of the study population. Women who developed septic shock were more likely to have undergone non-obstetric surgery during pregnancy (including appendectomy, cervical cerclage and ureteral double-J stent placement) than those who did not develop shock (28.3% vs. 10.1%, p = 0.01). Septic shock was also more frequent among women with postpartum-onset sepsis (43.3% vs. 19%, p = 0.003). Maternal and neonatal outcomes according to septic shock status are summarized in Table 2.

Table 1.

Baseline characteristics of patients with maternal sepsis according to septic shock status

	Non-shock (N = 79)	Shock (N = 60)	P value
Age, mean (SD), years	31.7 (4.90)	31.3 (4.87)	0.61
Age			0.23
≥ 35 years	27 (34.2%)	14 (23.3%)
< 35 years	52(65.8%)	46(76.7%)
BMI, mean (SD)	24.2 (4.18)	24.9 (5.34)	0.41
obesity	13 (17.1%)	14 (28.0%)	0.22
Gravidity, mean (SD)	2.15 (1.54)	2.37 (1.35)	0.39
Parity, mean (SD)	0.43 (0.57)	0.67 (0.82)	0.06
Nulliparous	48 (60.8%)	31 (51.7%)	0.37
Onset time group			0.003
During Pregnancy	64 (81.0%)	34 (56.7%)
Puerperium	15 (19.0%)	26 (43.3%)
Diabetes (including gestational diabetes and pregestational diabetes)	6 (7.59%)	9 (15.0%)	0.26
Hypertensive disorders of pregnancy	13 (16.5%)	8 (13.3%)	0.79
Cardiovascular disease	4 (5.06%)	8 (13.3%)	0.16
Renal disease	12 (15.2%)	6 (10.0%)	0.52
Immune system disease	8 (10.1%)	8 (13.3%)	0.75
hematological system diseases	7 (8.86%)	1 (1.67%)	0.14
History of surgery during pregnancy	8 (10.1%)	17 (28.3%)	0.01
White Blood Cell Count, mean (SD)	12.6 (6.64)	15.6 (10.8)	0.06
Percentage of neutrophils, mean (SD)	83.6 (7.81)	86.3 (12.7)	0.15
Count of neutrophils, mean (SD)	11.7 (10.7)	13.9 (10.1)	0.21
Percentage of eosinophils mean (SD)	0.39 (1.85)	0.42 (1.40)	0.92
Count of eosinophils, mean (SD)	0.14 (0.65)	0.05 (0.18)	0.25
Percentage of lymphocytes, mean (SD)	9.06 (6.25)	7.66 (11.3)	0.39
Count of lymphocytes, mean (SD)	1.00 (0.61)	0.88 (0.71)	0.31
logarithm of Neutrophil-to-Lymphocyte Ratio, mean (SD)	2.36 (0.75)	2.73 (1.10)	0.03
Red Blood Cell Count, mean (SD)	3.28 (0.61)	3.30 (0.84)	0.88
Hematocrit, mean (SD)	0.30 (0.05)	0.30 (0.07)	0.90
Hemoglobin, mean (SD)	99.8 (16.3)	99.1 (23.4)	0.83
Platelet Count, mean (SD)	164 (81.8)	139 (82.0)	0.08
Prothrombin Time, mean (SD)	11.9 (1.50)	14.5 (3.64)	< 0.001
Activated Partial Thromboplastin Time, mean (SD)	29.9 (4.45)	36.4 (12.9)	< 0.001
Fibrinogen, mean (SD)	4.75 (1.72)	4.06 (1.89)	0.03
D-dimer, mean (SD)	4.42 (12.4)	14.5 (35.8)	0.04
International Normalized Ratio, mean (SD)	1.02 (0.11)	1.26 (0.32)	< 0.001
Lactate, mean (SD)	1.71 (1.83)	3.27 (2.43)	< 0.001
Total Bilirubin, mean (SD)	14.0 (23.4)	16.0 (15.5)	0.56
Prealbumin, mean (SD)	118 (58.1)	116 (59.2)	0.80
Albumin, mean (SD)	30.0 (5.50)	29.5 (5.50)	0.56
Aspartate Aminotransferase, mean (SD)	39.2 (44.9)	89.6 (192)	0.05
Alanine Aminotransferase, mean (SD)	30.7 (30.3)	55.1 (127)	0.15
Alkaline Phosphatase, mean (SD)	107 (69.8)	96.4 (85.5)	0.43
Albumin/Globulin Ratio, mean (SD)	1.28 (0.32)	1.33 (0.41)	0.41
Creatinine, mean (SD)	61.8 (71.5)	94.2 (89.7)	0.02
Uric Acid, mean (SD)	277 (144)	304 (157)	0.32
B-type Natriuretic Peptide, mean (SD)	289 (470)	525 (692)	0.03
C-reactive Protein, mean (SD)	73.6 (68.5)	96.3 (66.6)	0.06
Procalcitonin, mean (SD)	1.48 (3.38)	21.1 (34.7)	< 0.001
Interleukin-6, mean (SD)	77.7 (85.4)	850 (1662)	0.01

All variables represent measurements obtained at the time of sepsis diagnosis

Table 2.

Maternal and neonatal outcomes according to septic shock status

	Non-shock (N = 79)	Shock (N = 60)	P value
Gestational age of delivery < 36w	34 (48.6%)	24 (64.9%)	0.16
Apgar			0.003
0–3	2 (2.99%)	8 (22.2%)
4–7	7 (10.4%)	6 (16.7%)
8–10	58 (86.6%)	22 (61.1%)
Low birth weight infant	32 (42.1%)	34 (75.6%)	0.001
Duration, median (interquartile range)	11.0 (8.0,15.5)	15(11,27.8)	0.002
Bundle therapy	56 (70.9%)	56 (93.3%)	0.002
Invasive ventilator	9 (11.4%)	37 (61.7%)	< 0.001
High-flow	18 (22.8%)	20 (33.3%)	0.23
Steroid	23 (29.1%)	14 (23.3%)	0.57

Maternal and neonatal outcomes were assessed during hospitalization and at delivery

Laboratory parameters measured at sepsis diagnosis showed markedly higher inflammatory and coagulation biomarkers in the septic-shock group, including log-transformed neutrophil-to-lymphocyte ratio (log-NLR) (2.73 vs. 2.36, p = 0.03), procalcitonin (PCT) (21.1 vs. 1.48, p < 0.001), interleukin-6 (IL-6) (850 vs. 77.7, p = 0.01), prothrombin time (PT) (15.5 vs. 11.9, p < 0.001), activated partial thromboplastin time (APTT) (36.4 vs. 29.9, p < 0.001), and international normalized ratio (INR) (1.26 vs. 1.02, p < 0.001). Additionally, serum creatinine (Cr) levels were higher in the shock group (94.2 vs. 61.8, p = 0.02), and hypoxemia (PaO₂ < 60 mmHg) was significantly more common in the shock group (90.0% vs. 48.1%, p < 0.001).

The median treatment duration in the shock group was 15 (11, 27.8) days. Mechanical ventilation was required in 37 cases (61.7%), and 56 patients (93.3%) received bundled therapy. Compared with the non-shock group, the shock group had significantly longer hospitalization (15 vs. 11, p = 0.002), higher rates of invasive ventilation (61.7% vs. 11.4%, p < 0.001) and sepsis-bundle completion (93.3% vs. 70.9%, p = 0.002). No significant differences were observed in high-flow oxygen therapy or corticosteroid administration between groups. Regarding neonatal outcomes, the incidence of low-birth-weight infants (75.6% vs. 42.1%, P = 0.001) and neonatal asphyxia (38.9% vs. 13.4%, P = 0.003) were significantly higher in the shock group.

Sepsis occurred during pregnancy in 98 patients (70.5%) and postpartum in 41 patients (29.5%); detailed baseline characteristics are presented in Table S2.

Infection sources and microbiology

Respiratory infections were the most common source of maternal sepsis (44.6%), followed by genitourinary infections (30.9%), digestive system infections (12.9%), and hematologic/neurologic infections (11.5%). Among septic shock patients, genitourinary and hematologic/neurologic systems were more frequent, while respiratory and digestive system infections were less common (P < 0.05). No significant difference in infection-source distribution was observed between pregnancy-onset and the puerperal period (Table S3).

Pathogens were identified in 65 patients, with 11 cases of polymicrobial infections. Gram-negative bacteria were the most common (55.1%), mainly including Escherichia coli (17.9%), Klebsiella pneumoniae (16.7%), and Acinetobacter baumannii (14.1%). Fungal infections (all Candida albicans) accounted for 19.2%, followed by Gram-positive bacteria (17.9%), and Mycoplasma species (7.7%). No significant differences in pathogen distribution were found between the shock and non-shock groups, or between pregnancy-onset and postpartum sepsis (Table S4).

Among the 25 patients with a history of surgery during pregnancy, 16 (64%) underwent procedures for potentially infectious indications (e.g., appendectomy) and 9 (36%) for non-infectious indications (e.g., cervical cerclage). Both subgroups showed a higher incidence of septic shock compared to patients without surgery (infectious-cause surgery: 43.8% vs. 10.1%; non-infectious-cause surgery: 33.3% vs. 10.1%). The association remained significant when the composite variable was used in the model. Detailed results are presented in Supplementary Table S5.

Primary outcome: prediction of septic shock

Potential predictors of progression to septic shock are summarized in Table 3. On univariate analysis, a history of surgery during pregnancy, log-transformed Neutrophil-to-lymphocyte Ratio (log-NLR), PCT, Interleukin-6 (IL-6), PT, APTT, fibrinogen (Fg), INR, lactate (Lac), aspartate aminotransferase (AST), Cr, B-type natriuretic peptide (BNP), and arterial partial pressure of oxygen (PaO₂) were significantly associated with septic shock. Variables with a p < 0.10 were subsequently entered into a LASSO regression model for further variable selection. Four predictors were ultimately retained: non-obstetric surgery during pregnancy, PCT, PT, and PaO₂. A nomogram was constructed to predict the risk of septic shock in patients with maternal sepsis (Fig. 2).

Table 3.

Univariate Analysis of Risk Factors for Septic Shock

variable	RR(95%CI)	P value
age	0.73 (0.45, 1.17)	0.19
obesity	1.40 (0.83, 2.35)	0.21
Gravidity	1.06 (0.93, 1.19)	0.39
Parity	1.29 (1.04, 1.59)	0.02
IVF	1.12 (0.69, 1.84)	0.64
PROM	1.34 (0.85, 2.10)	0.21
Diabetes (including gestational diabetes and pregestational diabetes)	1.46 (0.92, 2.32)	0.11
Hypertensive disorders of pregnancy	0.86 (0.48, 1.55)	0.62
Cardiovascular disease	1.63 (1.04, 2.56)	0.03
Renal disease	0.75 (0.38, 1.48)	0.40
Immune system disease	1.18 (0.70, 2.01)	0.54
hematological system diseases	0.28 (0.04, 1.75)	0.17
History of surgery during pregnancy	1.80 (1.26, 2.58)	0.001
Onset time	1.83 (1.28, 2.61)	0.001
Duration	1.02 (1.01, 1.03)	< 0.001
White Blood Cell Count	1.02 (1.01, 1.03)	0.01
Count of neutrophils	1.01 (0.99, 1.03)	0.25
Count of lymphocytes	0.84 (0.58, 1.23)	0.37
Count of eosinophils	0.69 (0.38, 1.26)	0.23
logarithm of Neutrophil-to-Lymphocyte Ratio	1.30 (0.98, 1.72)	0.06
Red Blood Cell Count	1.02 (0.78, 1.34)	0.88
Platelet Count	0.998 (0.995, 1.000)	0.10
Prothrombin Time	1.12 (1.07, 1.16)	< 0.001
Activated Partial Thromboplastin Time	1.02 (1.01, 1.04)	< 0.001
Fibrinogen	0.89 (0.79, 0.99)	0.03
D-dimer	1.006 (1.003, 1.008)	< 0.001
International Normalized Ratio	3.60 (2.27, 5.71)	< 0.001
Lactate	1.12 (1.05, 1.20)	0.001
Total Bilirubin	1.00 (0.99, 1.01)	0.58
Albumin	0.99 (0.96, 1.03)	0.56
Aspartate Aminotransferase	1.001 (1.000, 1.001)	< 0.001
Alanine Aminotransferase	1.001 (1.001, 1.002)	< 0.001
Alkaline Phosphatase	0.999 (0.995, 1.003)	0.57
Albumin/Globulin Ratio	1.26 (0.76, 2.08)	0.38
Creatinine	1.002 (1.000, 1.004)	0.03
B-type Natriuretic Peptide	1.000 (1.000, 1.001)	0.003
C-reactive Protein	1.002 (1.000, 1.005)	0.06
Procalcitonin	1.01 (1.01, 1.02)	< 0.001
Interleukin-6	1.000 (1.000, 1.000)	< 0.001
Infection Sites
Digestive system	0.82 (0.36, 1.87)	0.64
Genitourinary system	1.58 (1.01, 2.47)	0.04
Hematologic/Neurologic system	2.03 (1.26, 3.28)	0.004
Infectious Pathogens
Gram-negative bacteria	1.49 (0.75, 2.98)	0.26
Gram-positive bacteria	1.00 (0.40, 2.50)	1.00
others/mixed	1.47 (0.68, 3.14)	0.32
PaO2	4.60 (2.14, 9.90)	< 0.001

IVF In Vitro Fertilization, PROM Premature Rupture of Membranes, PaO₂ arterial partial pressure of oxygen

Fig. 2.

Variable selection and construction of the septic shock prediction nomogram. A Selection of the optimal tuning parameter λ in the LASSO regression model by cross-validation. B LASSO coefficient profiles of candidate predictors. C Nomogram constructed using the final multivariable robust Poisson regression model to estimate the risk of septic shock within 72 h after sepsis diagnosis. The example patient had no hypoxemia, a PCT level of 1.31 ng/mL, a PT level of 14.1 s, and no history of surgery during pregnancy, corresponding to an estimated septic shock risk of 12.8%

At the optimal cutoff probability of 0.184, determined by maximizing the Youden index, the predictive model demonstrated good discrimination and calibration. In the training cohort, the model achieved an AUC of 0.87 (95% CI: 0.80–0.94), with a sensitivity of 0.93 (95% CI: 0.83–1.00) and specificity of 0.68 (95% CI: 0.55–0.80). The positive predictive value (PPV) was 0.69 (95% CI: 0.61–0.78), and the negative predictive value (NPV) was 0.93 (95% CI: 0.85–1.00). In the validation cohort, the AUC was 0.87 (95% CI: 0.75–0.99), with sensitivity of 0.89 (95% CI: 0.72–1.00) and specificity of 0.65 (95% CI: 0.44–0.83). The PPV and NPV were 0.68 (95% CI: 0.55–0.81) and 0.89 (95% CI: 0.74–1.00), respectively. Calibration analysis demonstrated good agreement between predicted and observed probabilities. DCA showed favorable clinical utility and net benefit across a wide range of threshold probabilities (Fig. 3). The shock group had significantly higher risk scores than the non-shock group (p < 0.001; Fig. S1). Given the short 72-hour time window, the model performance should be interpreted as reflecting early identification of patients at high imminent risk of septic shock at sepsis diagnosis.

Fig. 3.

Calibration curves and decision curve analysis of the final robust Poisson prediction model in the training and validation cohorts. A, B Calibration plots in the training and validation cohorts. C, D Decision curve analysis in the training and validation cohorts. The y-axis represents net benefit, and the x-axis represents the threshold probability

Subgroup analysis showed stable predictive performance across age, obesity status, mode of conception (spontaneous vs. IVF), parity, and sepsis onset timing (during pregnancy vs. postpartum), with AUC values ranging from 0.71 to 0.91. Notably, higher AUCs were observed in subgroups of advanced maternal age > 35 years (AUC 0.91), obesity (AUC 0.91), spontaneous conception (AUC 0.90), primiparous women (AUC 0.87), and pregnancy-onset sepsis (AUC 0.88). These findings require external validation before the model can be applied in broader clinical settings.

Secondary outcomes: maternal and neonatal outcomes

Secondary outcomes, including maternal and neonatal outcomes, are presented in Table 2. To further evaluate the clinical relevance of the predictive model, robust Poisson regression analyses were conducted to examine the association between the model-derived shock risk score and pregnancy as well as neonatal outcomes (Table 4). Higher risk scores were significantly associated with adverse pregnancy outcomes (RR = 1.62, 95% CI: 1.19–2.21, p = 0.002). Similarly, the shock risk score was independently associated with lower Apgar scores at birth (RR = 1.77, 95% CI: 1.10–2.86, p = 0.02) and reduced neonatal birth weight (RR = 1.53, 95% CI: 1.21–1.93, p < 0.001).

Table 4.

Associations between the predicted septic shock risk score and maternal/neonatal outcomes using robust Poisson regression

	RR（95%CI）	P value
Gestational age of delivery < 36w
Crude	1.54 (1.20,1.99)	<0.001
Model1	1.54 (1.19,1.98)	<0.001
Model2	1.60 (1.22,2.02)	<0.001
Model3	1.62 (1.19,2.21)	0.002
Apgar
Crude	1.88 (1.17,3.03)	0.01
Model1	1.77 (1.10,2.86)	0.02
Model2	1.53 (0.96,2.43)	0.07
Model3	0.97 (0.60,1.53)	0.91
Low birth weight infant
Crude	1.56 (1.24,1.97)	<0.001
Model1	1.58 (1.26,1.97)	<0.001
Model2	1.53 (1.21,1.93)	<0.001
Model3	1.23 (0.93,1.63)	0.15

Crude model represents the unadjusted association between the predicted septic shock risk score and each maternal/neonatal outcome. Model 1 adjusts for maternal age and BMI. Model 2 additionally adjusts for nulliparous and Onset time. Model 3 further adjusts for bundle therapy, invasive ventilator, vasopressor therapy and steroid. P value <0.05 was considered statistically significant. Apgar score was analyzed as a binary outcome (≤7 vs. >7 at 1 minute)

Discussion

In this retrospective cohort study of 139 pregnant and postpartum women with sepsis, we developed and internally validated a pragmatic risk-stratification model for septic shock using four readily available variables: history of non-obstetric surgery during pregnancy, prothrombin time (PT), procalcitonin (PCT), and arterial partial pressure of oxygen (PaO₂). The model showed good discrimination in both the training and validation cohorts (AUC 0.87 and 0.87, respectively), along with acceptable calibration performance. These findings suggest that the model may have potential utility in early risk stratification of septic shock in maternal sepsis.

Surgery during pregnancy imposes additional physiological stress on an already immune‑modulated host, potentially amplifying systemic inflammation and susceptibility to septic deterioration [18]. Moreover, even seemingly non‑infectious procedures (e.g., emergency cerclage) are frequently complicated by subclinical intra‑amniotic infection [19], which may explain the sustained predictive value observed in our sensitivity analysis across both infectious and non‑infectious surgical indications. The surgical event itself may therefore represent an important risk amplifier in the setting of sepsis. However, our broad definition of surgery is a limitation, and future studies with larger samples could investigate whether specific procedure types confer differential risk.

Procalcitonin (PCT) — a well‑established biomarker of bacterial infection and sepsis severity, showed a strong association with septic shock in our cohort. Our findings extend prior work that established the prognostic utility of PCT in general and obstetric sepsis populations [20]. While previous studies focused on mortality prediction, our data support PCT as an early indicator of impending haemodynamic deterioration in pregnant women with established sepsis. The discriminatory capacity of PCT in our model reinforces its role beyond diagnosis, as a tool for dynamic risk refinement at the bedside.

Prothrombin time (PT) independently predicted septic shock, corroborating emerging evidence that coagulation abnormalities are integral to the pathophysiology of sepsis‑induced organ dysfunction. In a large mixed‑ICU cohort, PT/INR > 1.5 was independently associated with shock and 30‑day mortality [21]. To our knowledge, our study is the first to validate this association specifically in an obstetric sepsis population, where pregnancy‑induced hypercoagulability may mask early coagulopathic changes. PT therefore may represent a clinically accessible marker of sepsis severity in pregnant patients.

Arterial partial pressure of oxygen (PaO₂) was the strongest individual predictor in our model. This finding aligns with large‑scale critical care studies demonstrating a U‑shaped relationship between PaO₂ and mortality, with both hypoxemia and extreme hyperoxemia conferring increased risk [22, 23]. In the obstetric population, where physiological respiratory alkalosis and reduced functional residual capacity are common, even modest reductions in PaO₂ may indicate limited cardiopulmonary reserve and a heightened risk of imminent clinical deterioration. Our results suggest that PaO₂, measured at sepsis diagnosis, may serve as an early physiological warning sign rather than simply reflecting established shock.

Direct comparison of our model with existing tools such as qSOFA and SOS should be interpreted with caution, as these instruments were designed for different purposes and time points. qSOFA and SOS primarily serve as screening tools for infection-related deterioration and rely on vital signs and subjective criteria and have demonstrated suboptimal sensitivity in obstetric populations [10, 12, 24]. In contrast, our model is specifically designed for risk stratification after sepsis diagnosis using objective laboratory and blood gas parameters. Although sensitivity was moderate, the model demonstrated moderate PPV and an acceptable NPV. From a clinical perspective, this model is primarily intended for risk stratification rather than broad population screening. Therefore, a higher PPV is desirable to accurately identify patients at genuinely elevated risk and to avoid unnecessary interventions in false-positive cases. At the same time, the maintained NPV suggests that the model retains reasonable exclusion capability without markedly compromising screening utility. Rather than replacing existing scores, we view our model as complementary—it may be most useful as a secondary triage tool to refine risk assessment in patients already identified as septic, helping to distinguish those at imminent risk of hemodynamic collapse.

Infection sources and microbiology. Consistent with global epidemiological data [8], respiratory and genitourinary tract infections were the most common sources of maternal sepsis in our cohort. The higher proportion of respiratory infections likely reflects our center’s role as a tertiary referral facility for severe pneumonia in pregnancy. Although no significant differences in pathogen distribution were observed between shock and non‑shock groups, the predominance of Gram‑negative organisms (55.1%) — including E. coli, K. pneumoniae, and A. baumannii — underscores the need for local antimicrobial stewardship and early targeted therapy [25, 26]. The lack of association between specific pathogens and shock progression supports the concept that host response, rather than microbial type per se, drives the transition to septic shock. Thus, while the fundamental treatment of sepsis remains unchanged, the model enables precision in the operationalization of that treatment, ensuring that the right intensity of care is delivered to the right patient at the right time. This represents a shift from protocol‑driven care to risk‑adapted, personalized management.

An additional consideration raised by recent literature is the distinction between community‑acquired and healthcare‑associated infections, which may have different implications for prognosis and management. In a study of critically ill COVID‑19 patients, superinfections—particularly ventilator‑associated pneumonia, bloodstream infections, and catheter‑associated urinary tract infections—were associated with a > 15‑fold increase in mortality odds, highlighting the substantial burden of nosocomial infections in ICU settings. Although our study focused on maternal sepsis rather than COVID‑19, a similar distinction may be relevant: infections acquired during hospitalization (e.g., postpartum endometritis, surgical site infections, catheter‑related infections) may carry a different risk profile for septic shock compared with community‑onset infections (e.g., pyelonephritis, pneumonia). In our cohort, puerperal sepsis (which is more likely to be healthcare‑associated) showed a significantly higher rate of septic shock than pregnancy‑onset sepsis (43.3% vs. 19.0%, p = 0.003), indirectly supporting this concept. However, we did not systematically classify infection sources as community versus healthcare‑associated, and this represents an important area for future investigation. Future studies should explicitly differentiate these categories to better understand their respective contributions to septic shock risk and to guide targeted prevention strategies [27].

Emerging biomarkers beyond traditional laboratory parameters, such as circulating microRNAs (miRNAs) and plasma gelsolin (pGSN), have recently gained attention for their potential role in sepsis diagnosis, prognosis, and even therapeutic targeting [28]. For instance, miR-150 and miR-122 have been associated with disease severity and mortality in septic patients, while low pGSN levels at admission predict worse outcomes and may serve as a therapeutic target. Although our model relies on routinely available variables (PCT, PT, PaO₂, and surgical history), these emerging biomarkers could potentially complement our risk stratification approach. Future studies should explore whether adding these novel biomarkers improves the predictive performance of our model, particularly in distinguishing septic shock risk in equivocal cases.

This study has prompted four proposed changes at our center: (1) revised sepsis bundle criteria to include model‑based triggers; (2) a pilot electronic medical record (EMR)‑integrated automated risk alert; (3) early ICU referral policy for high‑risk patients; and (4) earlier neonatology involvement in antenatal planning. Developed through multidisciplinary consensus, these initiatives offer a feasible pathway for translating risk models into bedside decision support. Other tertiary centers may adapt these strategies to their local context. We plan to evaluate the impact of these implementation efforts in future studies.

Limitations

This study has several limitations. First, as a single‑center retrospective study conducted at a tertiary referral hospital, our cohort may not represent the general population of pregnant or postpartum women with sepsis, and referral bias is possible as our center manages a higher proportion of severe cases. Second, all data were extracted from electronic medical records without standardized prospective data collection protocols. Third, despite using LASSO regression for variable selection and bootstrap internal validation, the multi‑stage model development process in a modest sample (N = 139, with only 60 events) carries a substantial risk of overfitting; the observed AUCs of 0.87 in both the training and validation cohorts may still be optimistic, and the model’s performance could decrease in new datasets. Fourth, due to the lack of external validation, the generalizability of our model to other populations and healthcare settings remains unknown. Fifth, the 72‑hour time window for septic shock development was chosen based on clinical relevance but may be arbitrary, and the composite variable “history of surgery during pregnancy” grouped heterogeneous procedures (infectious vs. non‑infectious indications); although a sensitivity analysis showed consistency, residual confounding by indication cannot be excluded. Finally, unmeasured confounders (e.g., timing of antibiotic administration, source control interventions, fluid resuscitation volumes) were not included in the model. Notwithstanding these limitations, the model’s predictors are routinely available, biologically plausible, and showed consistent performance in internal validation, supporting its potential utility as a hypothesis‑generating tool pending future multi‑center prospective validation.

Conclusion

This four‑variable model offers a potential simple tool for early risk stratification at sepsis diagnosis. It may facilitate timely care escalation and resource allocation. However, given the retrospective, single‑centre design and lack of external validation, these findings should be interpreted cautiously. The model reflects risk based on physiological status at diagnosis, rather than prediction independent of baseline severity. Pending rigorous multi‑center prospective validation, the model should be considered hypothesis‑generating and not yet ready for routine clinical use. It may serve as a decision‑support tool to improve maternal sepsis management only after further validation.

Abbreviations

Abbreviation

Full Term

LASSO

Least Absolute Shrinkage and Selection Operator

DCA

Decision Curve Analysis

PT

Prothrombin Time

PCT

Procalcitonin

PaO₂

Arterial Partial Pressure of Oxygen

AUC

Area Under Curve

CI

Confidence Intervals

GLOSS

Global Maternal Sepsis Study

SOS

Sepsis in Obstetrics Score

ICU

Intensive Care Unit

qSOFA

quick Sequential Organ Failure Assessment

MAP

Mean Arterial Pressure

BMI

Body Mass Index

IVF-ET

In Vitro Fertilization-Embryo Transfer

SD

Standard Deviation

IQR

Interquartile Range

log-NLR

Log-transformed Neutrophil-to-lymphocyte Ratio

IL-6

Interleukin-6

APTT

Activated Partial Thromboplastin Time

INR

International Normalized Ratio

Cr

Creatinine

Fg

Fibrinogen

Lac

Lactate

AST

Aspartate Aminotransferase

BNP

B-type Natriuretic Peptide

PPV

Positive Predictive Value

NPV

Negative Predictive Value

MIMIC-II

Medical Information Mart for Intensive Care II

SAPS-I

Simplified Acute Physiology Score I

Authors’ contributions

Conceived and designed the study: CW, XL, NZ, CL; Collected data: CW, XL, SY C, SB H, MJ; Analyzed data: XL, HT S, NZ; Figures preparation: XL, HT S, SC; Interpreted the results: XL, CW, NZ, CL; Contributed to the writing of the manuscript: XL, CW, SY C, NZ. The author(s) read and approved the final manuscript.

Funding

This work was supported by National Natural Science Foundation of China (No. 82201846) and National Natural Science Foundation Promotion Project of Ren Ji Hospital Affiliated to Shanghai JiaoTong University School of Medicine(RJTJ25-MS-022, RJTJ25-QN-103, RJTJ25-MS-067).

Data availability

The datasets generated and/or analyzed during the current study are available from the corresponding author upon reasonable request. Contact information: Prof. Ning Zhang (Email: ningning1723@126.com).

Declarations

Ethics approval and consent to participate

This observational retrospective cohort study was approved and informed consent was waived by the Ethics Committee on Human Research of Renji Hospital, Shanghai Jiao Tong University, Shanghai, China (Approval number: LY2023-199-C). All procedures were performed in accordance with the Declaration of Helsinki and relevant institutional guidelines.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.