Clinical characteristics and outcomes of critically ill women with gynecologic malignancies: a two-database retrospective study using MIMIC-IV v3.1 and eICU-CRD v2.0

This was a retrospective cohort study using two de-identified public critical care databases. The study was designed to describe ICU characteristics and in-hospital outcomes rather than to develop a prognostic model or estimate causal effects. Reporting was guided by STROBE and RECORD recommendations for observational studies using routinely collected health data [ 11 , 12 ]. Because the databases contain de-identified data, no patient contact occurred, and patient-level data were not redistributed. MIMIC-IV v3.1 was used as one cohort. MIMIC-IV contains de-identified electronic health record data from ICU and hospital encounters at Beth Israel Deaconess Medical Center; its de-identified anchor-year groups span 2008–2022 [ 13 ]. Both MIMIC-IV and eICU-CRD are distributed through PhysioNet, which provides controlled access to credentialed users [ 14 ]. eICU-CRD v2.0 was used as a second cohort. eICU-CRD is a multicenter critical care database containing de-identified ICU data from hospitals participating in the Philips eICU program between 2014 and 2015 [ 15 ]. These databases are referred to as public because they are openly documented and accessible to qualified researchers after credentialing, not because identifiable clinical records are publicly released. The present analysis used MIMIC-IV v3.1 and eICU-CRD v2.0 [ 16 , 17 ]. Eligible encounters were ICU stays among adult female patients with evidence of gynecologic malignancy. In MIMIC-IV, gynecologic malignancy was identified from hospital ICD-9 and ICD-10 diagnosis codes for cervical, uterine or endometrial, ovarian or fallopian tube, vulvar, vaginal, and related gynecologic cancers. In eICU-CRD, gynecologic malignancy was identified using diagnosis strings from the oncology/GU tumors diagnosis hierarchy, including cervical, ovarian, uterine, and uterine adenocarcinoma strings. Phenotype construction used a conservative rule set. Candidate eICU diagnosis strings were reviewed at the aggregate string level, and only strings explicitly indicating gynecologic malignancy were retained. Ambiguous, nonmalignant, or nonspecific gynecologic strings were not used for cohort entry. Six high-priority eICU gynecologic oncology diagnosis strings were retained and none were excluded after aggregate review. Plausibility checks included verification of adult female status, duplicate ICU-stay identifiers, unique patient counts, tumor-group distributions, database-specific metastatic cancer and acute kidney injury prevalence, and the absence of the other-gynecologic-cancer category in eICU-CRD under the predefined string rules. Direct patient-level chart validation was not possible because these de-identified public databases do not provide source chart review beyond structured and coded fields. For MIMIC-IV, the first ICU stay within each hospital admission was selected to reduce repeated ICU-stay clustering within the same hospitalization. In eICU-CRD, the public analytic unit is the ICU stay; a uniform first-ICU-stay-within-hospitalization restriction was not applied in this multicenter extraction. Eligible eICU-CRD ICU stays were retained to preserve sample size and multicenter representation, while unique patients were summarized separately and patient-level clustering was assessed in sensitivity analysis. Lactate and albumin measurements were not required for cohort entry, preserving the broader clinical ICU cohort. Tumor groups were categorized as ovarian/fallopian, uterine/endometrial, cervical, or other gynecologic malignancy. When more than one gynecologic cancer category was present, a prespecified hierarchy was applied in the order ovarian/fallopian, uterine/endometrial, cervical, and other gynecologic malignancy. Ovarian/fallopian cancer was prioritized because it was expected to be the largest clinically recognizable category across both databases and because overlapping ovarian, fallopian, and related adnexal terms may occur in gynecologic oncology records. This hierarchy was used for classification only and was not intended to imply biologic priority, more extensive cytoreductive surgery, or higher clinical severity. Baseline and clinical variables included age, ICU length of stay, database-recorded sepsis, database-recorded acute kidney injury, metastatic cancer, chronic kidney disease, chronic liver disease, congestive heart failure, diabetes, hypertension, early laboratory measurements, physiologic severity scores where available, and first-24-hour organ-support indicators. First-day Sequential Organ Failure Assessment (SOFA) score and SOFA components were extracted from MIMIC-IV derived tables. Acute Physiology and Chronic Health Evaluation (APACHE) IVa score, APACHE IVa acute physiology score, and selected APACHE physiologic variables were extracted from eICU-CRD. Because SOFA and APACHE IVa are different scoring systems and were not both available in both databases, they were analyzed in database-specific sensitivity models rather than forced into a single harmonized pooled score. First-24-hour organ support included mechanical ventilation, invasive ventilation or tracheostomy, vasoactive infusion, and renal replacement therapy (RRT), using database-specific derived tables in MIMIC-IV and APACHE, respiratory care, treatment, infusion, and intake/output tables in eICU-CRD. Source-recorded ICU admission context was summarized using conservative, nonadjudicated categories: postoperative/procedural, infection/sepsis, respiratory failure, hemorrhage, organ failure/shock, or other/unclear. These categories were derived from admission source fields, admission diagnosis strings, diagnosis codes, and early ICU support variables and should not be interpreted as manually adjudicated primary ICU admission indications. In eICU-CRD, ages recorded as greater than 89 years were assigned a value of 90 years. Database-recorded sepsis and database-recorded acute kidney injury were derived from diagnosis codes or diagnosis strings according to database-specific data structures, with clinical interpretation informed by established consensus definitions [ 18 , 19 ]. These variables should be interpreted as database-recorded clinical diagnoses rather than uniformly adjudicated Sepsis-3 or KDIGO-defined events. We did not reconstruct guideline-adjudicated sepsis or acute kidney injury because harmonized source elements required for full adjudication, including baseline creatinine, urine output, infection adjudication, vasopressor context, and SOFA component timing, were not consistently available across both databases. Contemporary sepsis guidance was considered when interpreting sepsis as an acute critical illness marker rather than a causal exposure [ 20 ]. Metastatic cancer and comorbidities were derived from diagnosis codes or diagnosis strings according to database-specific data structures. Laboratory values were defined as the first available measurement within 24 h after ICU admission. Lactate-to-albumin ratio availability was summarized when both lactate and albumin were available during this window; it was treated as a descriptive laboratory-completeness measure rather than as the primary exposure. Tumor stage, histology, treatment status, recurrence, recent surgery, performance status, code status, and goals-of-care documentation were not reliably available in the harmonized analytic dataset. The primary outcome was in-hospital mortality. In-hospital mortality was selected because it was consistently available across the two databases and is clinically relevant for ICU outcomes research. ICU length of stay was summarized descriptively. ICU mortality was available in eICU-CRD but was not used as the primary endpoint because the corresponding variable was not harmonized across both databases. Thirty-day mortality and post-discharge survival were not available in a consistent manner across MIMIC-IV and eICU-CRD. The unit of analysis was the ICU stay. Unique patient counts were reported to describe the degree of repeated ICU-stay contribution. Continuous variables were summarized as medians and interquartile ranges. Categorical variables were summarized as counts and percentages. Characteristics were summarized separately for MIMIC-IV, eICU-CRD, and the pooled cohort. Mortality was summarized overall, by tumor group, by source-recorded admission context, by first-24-hour organ support, and by selected clinical factors. Logistic regression models were intentionally parsimonious because the number of deaths was modest, especially in database-specific analyses. No formal sample size calculation was performed because this was a retrospective analysis of all eligible ICU stays in the two databases. The primary pooled model had 103 deaths and 8 non-intercept parameters, corresponding to approximately 12.9 events per parameter. Database-specific main models included age, tumor group, database-recorded sepsis, database-recorded acute kidney injury, and metastatic cancer. Tumor group was modeled categorically, with ovarian/fallopian cancer serving as the reference category because it was the largest tumor group and clinically recognizable across both databases. The pooled model additionally included database cohort, with MIMIC-IV as the reference. Variables used in the primary model had complete data in the analytic cohort after database extraction; no imputation was performed. Multiple imputation was not used because the primary model variables were complete after cohort construction and because laboratory variables with substantial missingness were either reported descriptively or used only in complete-case sensitivity analyses, avoiding model-dependent assumptions about unmeasured laboratory ordering. Results are reported as odds ratios (ORs) with 95% confidence intervals (CIs) and P values. Additional sensitivity analyses were performed after re-extraction of severity and organ-support variables. First, a MIMIC-IV database-specific model added first-day SOFA score to the primary database-specific covariates. Second, an eICU-CRD database-specific model added APACHE IVa score to the primary database-specific covariates. Third, a pooled organ-support sensitivity model added first-24-hour mechanical ventilation, vasoactive infusion, and RRT to the primary pooled model. Fourth, a complete-case pooled severity-proxy model added the organ-support indicators plus an early laboratory abnormality count based on creatinine, blood urea nitrogen, white blood cell count, hemoglobin, and platelet count. Fifth, patient-level clustering from repeated ICU stays was assessed using generalized estimating equations with a logit link, binomial family, and exchangeable working correlation structure. Sixth, hospital/center-cluster robust standard errors were estimated using eICU hospital identifiers plus one MIMIC center, and an eICU APACHE IVa model was additionally estimated with hospital-cluster robust standard errors. Seventh, inter-database heterogeneity was explored using cohort-by-risk-marker interaction terms for database-recorded sepsis, database-recorded acute kidney injury, metastatic cancer, mechanical ventilation, and vasoactive infusion. Eighth, multicollinearity was assessed using variance inflation factors (VIFs), and the age-mortality association was explored using a restricted cubic spline for age. P values were interpreted descriptively, and no causal effects were claimed. Analyses were performed using SQL extraction and Python-based workflows with pandas and statsmodels. AI-assisted tools were used to support code drafting, workflow documentation, and language polishing. The tools were not listed as authors. All analyses, interpretations, and final manuscript content were reviewed, verified, and approved by the authors, who take full responsibility for the work.

The final cohort included 711 ICU stays (Fig.  1 ). MIMIC-IV contributed 248 ICU stays from 219 unique patients, with 35 in-hospital deaths (14.1%). eICU-CRD contributed 463 ICU stays from 400 unique patients across 103 hospitals, with 68 in-hospital deaths (14.7%). Overall, there were 103 in-hospital deaths, corresponding to a pooled mortality of 14.5%. Baseline and clinical characteristics are summarized in Table  1 . Fig. 1 Study cohort flow diagram. Adult female ICU stays screened from MIMIC-IV v3.1 and eICU-CRD v2.0 are shown with database-specific phenotype rules, ICU-stay selection approach, unique patient counts, hospital counts where applicable, and in-hospital deaths Study cohort flow diagram. Adult female ICU stays screened from MIMIC-IV v3.1 and eICU-CRD v2.0 are shown with database-specific phenotype rules, ICU-stay selection approach, unique patient counts, hospital counts where applicable, and in-hospital deaths Table 1 Cohort characteristics Characteristic MIMIC-IV eICU-CRD Pooled ICU stays, n 248 463 711 Unique patients, n 219 400 619 Age, median (IQR), years 63.0 (51.8–70.0) 65.0 (56.0–75.0) 64.0 (54.0–73.0) In-hospital death, n (%) 35 (14.1%) 68 (14.7%) 103 (14.5%) ICU length of stay, median (IQR), days 1.9 (1.1-3.0) 1.8 (1.0-3.2) 1.9 (1.0-3.1) Ovarian/fallopian cancer, n (%) 104 (41.9%) 238 (51.4%) 342 (48.1%) Uterine/endometrial cancer, n (%) 85 (34.3%) 163 (35.2%) 248 (34.9%) Cervical cancer, n (%) 31 (12.5%) 62 (13.4%) 93 (13.1%) Other gynecologic cancer, n (%) 28 (11.3%) 0 (0.0%) 28 (3.9%) Database-recorded sepsis, n (%) 58 (23.4%) 95 (20.5%) 153 (21.5%) Database-recorded acute kidney injury, n (%) 104 (41.9%) 71 (15.3%) 175 (24.6%) Metastatic cancer, n (%) 156 (62.9%) 41 (8.9%) 197 (27.7%) Chronic kidney disease, n (%) 28 (11.3%) 88 (19.0%) 116 (16.3%) Chronic liver disease, n (%) 19 (7.7%) 4 (0.9%) 23 (3.2%) Congestive heart failure, n (%) 45 (18.1%) 24 (5.2%) 69 (9.7%) Diabetes, n (%) 48 (19.4%) 52 (11.2%) 100 (14.1%) Hypertension, n (%) 142 (57.3%) 74 (16.0%) 216 (30.4%) Lactate available within 24 h, n (%) 147 (59.3%) 118 (25.5%) 265 (37.3%) Albumin available within 24 h, n (%) 87 (35.1%) 224 (48.4%) 311 (43.7%) LAR available within 24 h, n (%) 64 (25.8%) 85 (18.4%) 149 (21.0%) Cohort characteristics The pooled median age was 64.0 (54.0–73.0) years. Median ICU length of stay was 1.9 (1.0-3.1) days. Database-recorded sepsis was identified in 153 (21.5%), database-recorded acute kidney injury in 175 (24.6%), and metastatic cancer in 197 (27.7%). Lactate and albumin were both available within the early ICU window in 149 (21.0%). Several database-recorded variables differed substantially between cohorts. Metastatic cancer was recorded in 156 of 248 MIMIC-IV stays (62.9%) and 41 of 463 eICU-CRD stays (8.9%). Database-recorded acute kidney injury was identified in 104 MIMIC-IV stays (41.9%) and 71 eICU-CRD stays (15.3%), whereas database-recorded sepsis prevalence was more similar (58 [23.4%] vs. 95 [20.5%]). These differences motivated database-specific severity models and formal inter-database interaction analyses. Severity-score availability differed by database. First-day SOFA was available for all 248 MIMIC-IV ICU stays; the local MIMIC-IV APSIII derived table did not contain nonmissing APSIII values. APACHE IVa score was available for 380 of 463 eICU-CRD ICU stays (82.1%). Among nonsurvivors versus survivors, median SOFA was 6.0 (3.5–8.5) versus 3.0 (2.0–5.0) in MIMIC-IV, and median APACHE IVa score was 77.0 (56.0-101.0) versus 50.0 (39.0–65.0) in eICU-CRD. First-24-hour mechanical ventilation was recorded in 71 MIMIC-IV stays (28.6%) and 145 eICU-CRD stays (31.3%). Vasoactive infusion was recorded in 62 MIMIC-IV stays (25.0%) and 96 eICU-CRD stays (20.7%). RRT was recorded in 7 MIMIC-IV stays (2.8%) and 13 eICU-CRD stays (2.8%). Source-recorded admission context was heterogeneous. Postoperative/procedural context was identified in 52 MIMIC-IV stays (21.0%; mortality 3.8%) and 227 eICU-CRD stays (49.0%; mortality 5.3%), whereas infection/sepsis context was identified in 58 (23.4%; mortality 25.9%) and 96 (20.7%; mortality 28.1%), respectively. Respiratory failure context was recorded in 44 MIMIC-IV stays (17.7%; mortality 18.2%) and 59 eICU-CRD stays (12.7%; mortality 27.1%). Detailed severity-score coverage, organ-support summaries, and admission-context summaries are provided in Additional file 2. Ovarian/fallopian malignancy was the most common tumor group in both databases, accounting for 104 ICU stays (41.9%) in MIMIC-IV and 238 ICU stays (51.4%) in eICU-CRD. Uterine/endometrial cancer accounted for 85 stays (34.3%) in MIMIC-IV and 163 stays (35.2%) in eICU-CRD. Cervical cancer accounted for 31 stays (12.5%) in MIMIC-IV and 62 stays (13.4%) in eICU-CRD. Tumor-group mortality is shown in Fig.  2 . In MIMIC-IV, mortality was 17.3% for ovarian/fallopian cancer, 11.8% for uterine/endometrial cancer, 9.7% for cervical cancer, and 14.3% for other gynecologic cancers. In eICU-CRD, mortality was 13.9% for ovarian/fallopian cancer, 13.5% for uterine/endometrial cancer, and 21.0% for cervical cancer. The eICU-CRD cervical-cancer estimate was based on 13 deaths among 62 ICU stays, with a wide 95% CI of 12.7%-32.6%. No eICU-CRD ICU stay met the operational definition for other gynecologic cancer because the retained eICU gynecologic oncology strings were cervical, ovarian, uterine, or uterine adenocarcinoma terms; pooled estimates for this category were therefore driven by MIMIC-IV. Tumor-group comparisons should therefore be interpreted cautiously because subgroup sizes were modest and detailed stage or treatment information was unavailable. Fig. 2 In-hospital mortality by tumor group and database. Whiskers show Wilson 95% confidence intervals; labels show mortality percentage and deaths/ICU stays In-hospital mortality by tumor group and database. Whiskers show Wilson 95% confidence intervals; labels show mortality percentage and deaths/ICU stays Unadjusted mortality patterns for database-recorded sepsis, database-recorded acute kidney injury, and metastatic cancer are summarized in Fig.  3 . In MIMIC-IV, database-recorded sepsis was associated with higher in-hospital mortality (OR 2.47, 95% CI 1.10–5.55; P  = 0.029). Database-recorded acute kidney injury had a similar direction of association (OR 2.20, 95% CI 1.00-4.85; P  = 0.051), and metastatic cancer was directionally associated with mortality but with wide confidence intervals (OR 1.75, 95% CI 0.75–4.09; P  = 0.194). In eICU-CRD, database-recorded sepsis (OR 2.06, 95% CI 1.14–3.74; P  = 0.017), database-recorded acute kidney injury (OR 2.21, 95% CI 1.17–4.18; P  = 0.015), and metastatic cancer (OR 2.25, 95% CI 1.04–4.86; P  = 0.038) were each associated with higher in-hospital mortality. Fig. 3 Unadjusted in-hospital mortality by selected clinical factors. Bars compare ICU stays with and without each database-recorded factor within each database; whiskers show Wilson 95% confidence intervals, and labels show mortality percentage and deaths/ICU stays Unadjusted in-hospital mortality by selected clinical factors. Bars compare ICU stays with and without each database-recorded factor within each database; whiskers show Wilson 95% confidence intervals, and labels show mortality percentage and deaths/ICU stays Adjusted associations are shown in Fig.  4 ; Table  2 . In the primary pooled model, database-recorded sepsis (OR 2.17, 95% CI 1.35–3.49; P  = 0.001), database-recorded acute kidney injury (OR 2.27, 95% CI 1.38–3.71; P  = 0.001), and metastatic cancer (OR 2.03, 95% CI 1.14–3.59; P  = 0.016) were associated with higher in-hospital mortality. After adjustment for selected clinical factors, the eICU-CRD cohort was also associated with higher adjusted odds of mortality (OR 2.01, 95% CI 1.10–3.69; P  = 0.024), despite similar crude mortality. This database-cohort coefficient should not be interpreted as a biological, quality-of-care, or causal effect because database-specific coding structures, diagnosis capture, ICU admission thresholds, and unmeasured case-mix differences may strongly influence adjusted comparisons. Fig. 4 Adjusted associations with in-hospital mortality. Odds ratios are shown on a logarithmic scale. Models were adjusted for age, tumor group, database-recorded sepsis, database-recorded acute kidney injury, and metastatic cancer; the pooled model was additionally adjusted for database cohort Adjusted associations with in-hospital mortality. Odds ratios are shown on a logarithmic scale. Models were adjusted for age, tumor group, database-recorded sepsis, database-recorded acute kidney injury, and metastatic cancer; the pooled model was additionally adjusted for database cohort Table 2 Main adjusted logistic regression for in-hospital mortality Cohort Variable OR (95% CI) P value ICU stays Deaths MIMIC-IV Cervical cancer 0.46 (0.12–1.80) 0.263 248 35 MIMIC-IV Other gynecologic cancer 0.70 (0.20–2.39) 0.569 248 35 MIMIC-IV Uterine/endometrial cancer 0.70 (0.29–1.67) 0.418 248 35 MIMIC-IV Age, per year 1.00 (0.97–1.03) 0.907 248 35 MIMIC-IV Database-recorded sepsis 2.47 (1.10–5.55) 0.029 248 35 MIMIC-IV Database-recorded acute kidney injury 2.20 (1.00-4.85) 0.051 248 35 MIMIC-IV Metastatic cancer 1.75 (0.75–4.09) 0.194 248 35 eICU-CRD Cervical cancer 1.39 (0.66–2.93) 0.391 463 68 eICU-CRD Uterine/endometrial cancer 0.81 (0.44–1.48) 0.491 463 68 eICU-CRD Age, per year 1.00 (0.98–1.03) 0.644 463 68 eICU-CRD Database-recorded sepsis 2.06 (1.14–3.74) 0.017 463 68 eICU-CRD Database-recorded acute kidney injury 2.21 (1.17–4.18) 0.015 463 68 eICU-CRD Metastatic cancer 2.25 (1.04–4.86) 0.038 463 68 Pooled Cervical cancer 1.03 (0.54–1.97) 0.921 711 103 Pooled Other gynecologic cancer 0.81 (0.25–2.65) 0.727 711 103 Pooled Uterine/endometrial cancer 0.77 (0.47–1.26) 0.303 711 103 Pooled eICU-CRD cohort 2.01 (1.10–3.69) 0.024 711 103 Pooled Age, per year 1.00 (0.99–1.02) 0.586 711 103 Pooled Database-recorded sepsis 2.17 (1.35–3.49) 0.001 711 103 Pooled Database-recorded acute kidney injury 2.27 (1.38–3.71) 0.001 711 103 Pooled Metastatic cancer 2.03 (1.14–3.59) 0.016 711 103 Main adjusted logistic regression for in-hospital mortality In the pooled organ-support sensitivity model, first-24-hour mechanical ventilation (OR 1.78, 95% CI 1.11–2.85; P  = 0.017) and vasoactive infusion (OR 2.62, 95% CI 1.60–4.30; P  < 0.001) were associated with mortality. In this model, database-recorded sepsis attenuated (OR 1.60, 95% CI 0.95–2.69; P  = 0.074), whereas database-recorded acute kidney injury (OR 2.18, 95% CI 1.30–3.64; P  = 0.003) and metastatic cancer (OR 2.19, 95% CI 1.21–3.93; P  = 0.009) remained associated with mortality. In the complete-case organ-support plus laboratory-burden model (616 ICU stays, 80 deaths), the early laboratory abnormality count was strongly associated with mortality (OR 1.82 per abnormality, 95% CI 1.45–2.29; P  < 0.001), and sepsis and acute kidney injury estimates further attenuated. Database-specific severity models showed strong associations between physiologic severity scores and mortality. In MIMIC-IV, first-day SOFA was associated with mortality after adjustment for age, tumor group, database-recorded sepsis, database-recorded acute kidney injury, and metastatic cancer (OR 1.30 per point, 95% CI 1.14–1.49; P  < 0.001). After SOFA adjustment, sepsis (OR 1.20, 95% CI 0.48–3.03) and acute kidney injury (OR 1.87, 95% CI 0.82–4.26) were attenuated. In eICU-CRD, APACHE IVa score was associated with mortality (OR 1.04 per point, 95% CI 1.03–1.05; P  < 0.001); sepsis (OR 1.32, 95% CI 0.60–2.92) and acute kidney injury (OR 1.03, 95% CI 0.43–2.48) were attenuated, while metastatic cancer remained directionally associated (OR 2.55, 95% CI 0.95–6.88; P  = 0.064). Patient-level GEE analyses showed the same general direction for database-recorded sepsis (OR 2.49, 95% CI 1.23–5.07; P  = 0.011), database-recorded acute kidney injury (OR 2.28, 95% CI 1.09–4.77; P  = 0.029), metastatic cancer (OR 2.15, 95% CI 1.01–4.58; P  = 0.048), and the eICU-CRD cohort (OR 2.06, 95% CI 0.99–4.26; P  = 0.052). Hospital/center-cluster robust standard errors, using eICU hospital identifiers plus one MIMIC center, yielded similar primary-model inferences. Multicollinearity diagnostics did not indicate serious collinearity in the pooled model: the maximum VIF was 1.67, with VIFs of 1.09 for sepsis and 1.19 for acute kidney injury. The primary pooled model included 103 deaths and 8 non-intercept parameters, corresponding to approximately 12.9 events per parameter; more parameter-rich sensitivity models were treated as exploratory. Exploratory restricted cubic spline modeling did not provide evidence that age had a nonlinear association with in-hospital mortality (likelihood-ratio chi-square 1.25, df = 2, P  = 0.537). Cohort-by-risk-marker interaction tests did not show statistically clear inter-database interaction signals for database-recorded sepsis (P for interaction = 0.846), database-recorded acute kidney injury ( P  = 0.941), metastatic cancer ( P  = 0.673), mechanical ventilation ( P  = 0.731), or vasoactive infusion ( P  = 0.066). Detailed sensitivity outputs, model-stability diagnostics, database-interaction tests, severity-score coverage, admission-context summaries, organ-support summaries, and phenotype-review summaries are provided in Additional file 2, and the exploratory spline plot is provided in Additional file 3.

Gynecologic malignancies, including ovarian, fallopian tube, uterine, endometrial, cervical, vaginal, and vulvar cancers, account for a substantial burden of morbidity and mortality among women worldwide [ 1 ]. Patients with these cancers may require ICU admission for perioperative complications, hemorrhage, sepsis, respiratory failure, kidney injury, treatment-related toxicity, or progressive metastatic disease [ 2 ]. As cancer survival improves and oncologic treatments become more complex, clinicians increasingly encounter patients with active or prior cancer in the ICU [ 3 ]. Contemporary ICU oncology literature emphasizes that outcomes in this population are heterogeneous and depend on both cancer-related and acute critical illness factors [ 4 , 5 ]. Large ICU cohorts and database studies have shown that short-term outcomes among patients with cancer are strongly influenced by acute organ dysfunction, infection, baseline cancer burden, and treatment goals [ 6 , 7 ]. National trend data also suggest that ICU outcomes in patients with cancer have changed over time, likely reflecting evolving admission practice and supportive care [ 8 ]. However, many ICU oncology studies combine hematologic malignancies, solid tumors, surgical monitoring, and medical deterioration across cancer sites, while gynecologic oncology patients may have distinct ICU contexts. These include major abdominal or pelvic surgery, extensive cytoreductive procedures, hemorrhage, urinary or bowel obstruction, treatment-related complications, and advanced pelvic or peritoneal disease [ 9 , 10 ]. The knowledge gap is therefore not whether sepsis, acute kidney injury, or metastatic cancer are generally relevant to ICU mortality; those relationships are well established in critical care oncology. The gap is the limited multicohort description of ICU characteristics, admission context, database-recorded severity markers, and short-term outcomes specifically among women with gynecologic malignancies. De-identified critical care databases provide an opportunity to study clinically important subgroups that are difficult to assemble in single-center cohorts. In this manuscript, the term public database refers to a de-identified research database made available to credentialed users through PhysioNet after required training and data use agreements, not to unrestricted release of patient-level data. MIMIC-IV and eICU-CRD are complementary ICU databases with different data structures, time periods, and healthcare settings. Using both databases may improve transparency and help distinguish reproducible clinical patterns from database-specific findings. However, database studies in gynecologic oncology must be interpreted carefully because detailed tumor stage, histology, treatment status, recurrence, surgery timing, chemotherapy, radiotherapy, performance status, code status, and goals of care are often unavailable, and because database heterogeneity may reflect coding practice as much as clinical case mix. We therefore conducted a two-database retrospective cohort study of adult female ICU patients with gynecologic malignancies in MIMIC-IV v3.1 and eICU-CRD v2.0. The objectives were to describe clinical characteristics and in-hospital outcomes, summarize tumor-group-specific mortality, and explore selected factors associated with in-hospital mortality. The study was designed as a descriptive clinical epidemiology analysis rather than a causal inference or prediction-model study. Phenotype definitions and database-specific extraction rules are provided in Additional file 1.

In this two-database retrospective cohort study, critically ill women with gynecologic malignancies had substantial in-hospital mortality, with a pooled mortality rate of 14.5%. Mortality was similar in MIMIC-IV and eICU-CRD in crude analyses despite differences in database structure, time period, hospital setting, and diagnosis-capture approach. In the primary pooled model, database-recorded sepsis, database-recorded acute kidney injury, and metastatic cancer identified patients at higher risk of in-hospital mortality. However, additional severity and organ-support analyses showed that the sepsis and acute kidney injury estimates attenuated after adjustment for SOFA in MIMIC-IV, APACHE IVa in eICU-CRD, and pooled organ-support/laboratory-burden proxies. These findings support interpreting sepsis and acute kidney injury primarily as database-recorded markers of acute illness burden and organ dysfunction, rather than as fully independent causal effects. The added severity analyses directly address a major limitation of the original analysis. SOFA was available for all MIMIC-IV stays, and APACHE IVa was available for most eICU-CRD stays. Both scores were substantially higher among nonsurvivors and remained strongly associated with mortality in database-specific models. Because SOFA and APACHE IVa are not the same scale and were not jointly available across both databases, a single harmonized pooled severity-score model would have been methodologically misleading. The database-specific severity models are therefore best viewed as robustness analyses showing that acute physiologic severity is a central driver of mortality in this population. The organ-support and laboratory-burden analyses provide additional clinical context. Mechanical ventilation and vasoactive infusion within the first 24 h were associated with mortality in the pooled sensitivity model, and early laboratory abnormality count was strongly associated with mortality in complete-case analysis. These findings are consistent with the idea that the observed mortality signal reflects acute physiologic deterioration, circulatory/respiratory support needs, and laboratory evidence of organ dysfunction. They also explain why database-recorded sepsis and acute kidney injury should not be overinterpreted as stand-alone causal exposures. Between-database heterogeneity was clinically meaningful. MIMIC-IV had much higher database-recorded metastatic cancer and acute kidney injury prevalence than eICU-CRD, suggesting differences in diagnosis capture, coding practice, ICU selection, or case mix. Interaction tests did not identify statistically clear effect modification for the main risk markers, but absence of statistically significant interaction does not prove that the two databases are exchangeable. The pooled results should therefore be read alongside database-specific and sensitivity analyses. The observed in-hospital mortality rate of 14.5% should be interpreted against several comparator contexts. It appears lower than many reports of unselected oncology ICU cohorts, particularly cohorts enriched for acute respiratory failure, septic shock, hematologic malignancy, or late-stage cancer, but higher than would be expected for purely planned postoperative monitoring populations. We did not extract matched comparator cohorts of non-gynecologic malignancy ICU patients or nonmalignant ICU admissions; therefore, this study cannot estimate a within-database excess-risk contrast between gynecologic malignancy and nonmalignant critical illness. Direct comparison is also limited by differences in cancer mix, active versus historical cancer status, ICU admission thresholds, postoperative monitoring practice, organ support, and goals-of-care selection. Prior ICU oncology studies have reported heterogeneous outcomes among patients with solid tumors [ 6 , 7 ]. Gynecologic oncology patients represent a clinically distinct subgroup because ICU admission may follow major abdominal or pelvic surgery, infection, hemorrhage, urinary or bowel obstruction, treatment-related toxicity, or complications of advanced pelvic or peritoneal disease [ 2 , 9 , 10 ]. In the present cohort, source-recorded admission context suggested two clinically relevant patterns: postoperative/procedural ICU use, which was common and associated with relatively low mortality, and medical deterioration contexts such as infection/sepsis or respiratory failure, which had higher mortality. This pattern is more informative for gynecologic oncology practice than the presence of sepsis, acute kidney injury, or metastatic disease alone. It also suggests that gynecologic oncology ICU cohorts may differ from both general oncology ICU cohorts and nonmalignant ICU cohorts because the denominator includes a mixture of planned perioperative escalation, potentially reversible complications, and advanced cancer-related deterioration. These categories were not manually adjudicated and should be interpreted as source-recorded descriptors rather than definitive ICU admission indications. Several findings are informative for gynecologic oncology practice. Tumor site alone did not show a stable independent association with mortality after adjustment, and tumor-group estimates were imprecise. Prognostic discussions for women with gynecologic malignancies admitted to the ICU should therefore consider the acute critical illness episode, physiologic severity, organ support, metastatic or advanced disease burden, and the clinical trajectory rather than tumor site alone. Code status, treatment intent, prior recurrence, recent surgery, performance status, and goals-of-care discussions are especially important in this population, but these variables were not reliably available in MIMIC-IV or eICU-CRD. Their absence matters because ICU admission is partly a clinical decision about whether escalation is appropriate, not only a marker of physiologic deterioration. We cannot assume that all ICU-admitted patients were full code or had not previously had goals-of-care conversations; conversely, patients with advanced recurrence, poor performance status, or comfort-focused goals may have been less likely to enter the ICU cohort at all. This selection process could make observed ICU mortality appear lower than mortality among all patients with progressive or recurrent gynecologic malignancy facing acute deterioration. At the same time, a 14.5% in-hospital mortality rate remains clinically meaningful if many admitted patients were considered candidates for aggressive or reversible critical care. The absence of these decision-making variables may confound reported associations in several ways: metastatic cancer coding may reflect advanced stage, recurrence, poor performance status, or palliative treatment intent; recent surgery may identify planned postoperative monitoring with lower short-term mortality; and recent systemic therapy may influence infection, kidney injury, and organ-support risk. These limitations underscore why database-derived mortality estimates should support, rather than replace, individualized multidisciplinary discussions among gynecologic oncologists, intensivists, patients, and families. The short median ICU length of stay should also be interpreted in context. A median ICU stay of approximately two days may reflect a mixture of postoperative monitoring, early stabilization and transfer, short admissions for acute complications, and early deaths. It should not be used as a direct proxy for illness severity. The newly added severity-score and organ-support analyses provide more clinically meaningful context than length of stay alone, but they still cannot capture detailed duration, dose, reversibility, or treatment goals. Age was not associated with in-hospital mortality in the adjusted models, and exploratory spline analysis did not support a nonlinear age-mortality relationship. This finding should not be interpreted as evidence that age is clinically unimportant. The age distribution was relatively compressed among adult ICU patients, and acute physiologic severity, organ support, metastatic cancer coding, treatment selection, and unmeasured care-goal factors may have dominated the short-term hospital outcome signal in this selected ICU cohort. The use of two independent ICU databases provides replication across data sources but should not be considered external clinical validation. Both databases largely reflect United States critical care practice and may not generalize to hospitals, countries, or resource settings with different ICU admission thresholds, surgical pathways, oncology access, palliative-care integration, or coding practices. Future external validation should evaluate contemporary institutional gynecologic oncology cohorts, non-ICU gynecologic oncology populations, and general ICU populations to clarify how much of the observed risk is specific to gynecologic malignancy rather than critical illness alone. The limited availability of paired early lactate and albumin measurements highlights an important reporting issue in database research. Lactate-to-albumin ratio has biologic and clinical appeal, but paired measurement was available for only 21.0% of ICU stays. We therefore minimized its role, reported it as a laboratory-completeness measure, and avoided making it a primary exposure. This approach reduces selection pressure and preserves a broader, clinically recognizable ICU population. Clinically, these findings support early attention to physiologic severity, organ support needs, infection, kidney injury, and advanced disease burden in ICU care for women with gynecologic malignancies, while avoiding tumor-site-only prognostication. The consistent signal for database-recorded acute kidney injury in the primary and organ-support models suggests that kidney dysfunction should be treated as an important bedside warning marker in this population, although severity-adjusted analyses indicate that it should not be interpreted as an isolated causal exposure. Prior ICU-oncology studies have reported that acute kidney injury is common in critically ill patients with cancer and is associated with higher mortality [ 21 , 22 ]. Clinical reviews have emphasized cancer-specific mechanisms and management challenges of acute kidney injury in this population [ 23 , 24 ]. Broader oncology cohorts also support kidney injury as an adverse prognostic marker in patients with cancer [ 25 ]. The source-recorded admission-context patterns also suggest that planned postoperative/procedural ICU monitoring should be interpreted differently from infection, respiratory failure, or shock contexts. For multidisciplinary decision-making, the most important missing variables are likely treatment intent, disease recurrence, performance status, code status, and goals-of-care documentation. These variables are needed to distinguish patients admitted for a time-limited trial of intensive care or planned postoperative monitoring from patients admitted during progressive disease with limited oncologic options. Future studies should integrate structured ICU data with tumor registries or institutional oncology records to distinguish active versus historical cancer, perioperative monitoring versus medical deterioration, curative versus palliative treatment intent, and recurrent or progressive disease. This study has several strengths. It used two independent de-identified critical care databases, focused on a clinically important subgroup often diluted within broader ICU oncology cohorts, provided database-specific phenotype definitions with aggregate diagnosis-string review and plausibility checks, and avoided overstating causal inference. The analysis includes actual database-derived severity and organ-support variables, source-recorded admission-context summaries, patient- and hospital/center-clustering sensitivity analyses, model-stability diagnostics, database-interaction tests, and confidence intervals for mortality estimates in updated figures and supplementary tables. The limitations are mainly related to design, data availability, and database structure. First, this was an ICU-stay-based retrospective study and does not represent all women with gynecologic malignancies, particularly patients not admitted to ICU because of referral patterns, survivorship selection, or treatment limitations. Second, active versus historical cancer, perioperative versus medical ICU use, and advanced or recurrent disease could not be reliably distinguished. Key oncologic and decision-making variables, including tumor stage, histology, treatment line, recent therapy, recurrence, performance status, code status, and goals of care, were not reliably available. Third, although we added SOFA in MIMIC-IV and APACHE IVa in eICU-CRD, no single harmonized severity score was available across both databases, and organ-support variables did not capture duration, dose, reversibility, or treatment intent. Fourth, gynecologic malignancy, sepsis, acute kidney injury, metastatic cancer, and comorbidities were based on database-recorded codes or diagnosis strings and may be misclassified differently across databases. Finally, patient-level GEE and hospital/center-cluster robust analyses addressed some nonindependence, but ICU stay remained the unit of analysis; tumor-subgroup mortality was underpowered, external validation in non-US or contemporary institutional cohorts was not performed, and 30-day, post-discharge, and longitudinal oncologic outcomes could not be assessed.

In MIMIC-IV v3.1 and eICU-CRD v2.0, critically ill women with gynecologic malignancies had substantial in-hospital mortality. Database-recorded sepsis, database-recorded acute kidney injury, and metastatic cancer identified higher-risk patients in the primary pooled model. After adding severity and organ-support analyses, first-day SOFA in MIMIC-IV, APACHE IVa in eICU-CRD, mechanical ventilation, vasoactive infusion, and early laboratory abnormality burden provided important context and attenuated the sepsis and acute kidney injury estimates. These findings should be interpreted as descriptive risk-marker associations rather than causal effects. Future studies should incorporate harmonized physiologic severity scores, detailed oncologic variables, treatment intent, and goals-of-care information.

Additional file 1. Phenotype definitions and database-specific extraction rules. This Excel workbook contains MIMIC-IV ICD code patterns, eICU-CRD diagnosis-string rules, database-recorded sepsis, database-recorded acute kidney injury, metastatic cancer definitions, and tumor-group hierarchy rules. Additional file 1. Phenotype definitions and database-specific extraction rules. This Excel workbook contains MIMIC-IV ICD code patterns, eICU-CRD diagnosis-string rules, database-recorded sepsis, database-recorded acute kidney injury, metastatic cancer definitions, and tumor-group hierarchy rules. Additional file 2. Supplementary tables. This Excel workbook contains tumor-group outcomes, mortality estimates with confidence intervals, patient-cluster GEE sensitivity analysis, multicollinearity diagnostics, exploratory age spline results, model-stability diagnostics, database-interaction tests, severity-score coverage, source-recorded admission-context summaries, first-24-hour organ-support summaries, enhanced severity and organ-support sensitivity models, and phenotype-review summaries. Additional file 2. Supplementary tables. This Excel workbook contains tumor-group outcomes, mortality estimates with confidence intervals, patient-cluster GEE sensitivity analysis, multicollinearity diagnostics, exploratory age spline results, model-stability diagnostics, database-interaction tests, severity-score coverage, source-recorded admission-context summaries, first-24-hour organ-support summaries, enhanced severity and organ-support sensitivity models, and phenotype-review summaries. Additional file 3. Supplementary Figure S1. Exploratory restricted cubic spline for age. Predicted mortality was standardized over the pooled cohort covariate distribution. Additional file 3. Supplementary Figure S1. Exploratory restricted cubic spline for age. Predicted mortality was standardized over the pooled cohort covariate distribution.