Multiple Clustering Algorithms for Ovarian Cancer Subtyping (MCAOCS): identifying clinical phenotypes and treatment response in critical care Authors: Yamin Qiu, Zhengwen Qin, Huixian Zheng, Tao Wang, Runhong Liu, Jing Yang Published: 2026 ABSTRACT BackgroundOvarian cancer (OC) patients admitted to the Intensive Care Unit (ICU) represent a critically ill population with significant heterogeneity, posing challenges for prognosis and personalized management. A robust subtyping system integrating multidimensional clinical data is urgently needed.ObjectivesThis study aimed to identify and characterize distinct clinical phenotypes of OC in critical care settings and investigate their associations with outcomes and subtype-specific treatment responses.MethodsThis retrospective cohort study analyzed data from 448 ICU patients with OC across two critical care databases (MIMIC-IV and eICU). We employed a Multiple Clustering Algorithms for OC Subtyping (MCAOCS) approach to identify phenotypes. The robustness of subtypes was assessed through internal and external application. Survival analysis and Inverse Probability of Treatment Weighting (IPTW) were used to evaluate mortality and treatment-exposure association.ResultsWe identified two distinct clinical phenotypes from the MIMIC-IV discovery cohort (n = 259): Critical OC (n = 118) characterized by hemodynamic instability and high disease severity, and Stable OC (n = 141) with preserved organ function. The robustness and reproducibility of these phenotypes were assessed in the external eICU application cohort (n = 189). The Critical OC subtype demonstrated significantly higher mortality from 28 days to one year. Exploratory treatment-exposure analyses suggested subtype-specific associations: several ICU support and medication exposures were associated with lower 28-day mortality in the Critical OC subtype, whereas in the Stable OC subtype, only insulin and phenylephrine exposures retained statistical significance after multiple-testing correction. These findings should be interpreted as hypothesis-generating rather than causal.ConclusionsThe MCAOCS-based binary phenotype system is robust and may have potential clinical utility, effectively stratifying ICU OC patients by mortality risk and suggesting subtype-specific treatment associations, thereby potentially facilitating precision critical care.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s13048-026-02167-x. Introduction Ovarian cancer remains one of the most lethal malignancies of the female reproductive system. According to GLOBOCAN 2022, there were approximately 324,000 new cases and 207,000 deaths globally, with marked regional variations in incidence and mortality [1]. About 70% of patients are diagnosed at an advanced stage, contributing to a five-year survival rate of only about 40% [2]. Patients with ovarian cancer admitted to the Intensive Care Unit (ICU) represent a particularly critical population, exhibiting significant clinical heterogeneity in disease severity, organ dysfunction, and treatment response [3]. Traditional assessment systems based solely on “tumor staging” are inadequate for capturing this complexity, necessitating a refined stratification strategy that integrates multidimensional clinical data for precision critical care. Unlike stable oncology outpatients, ICU-admitted ovarian cancer patients face an acute-on-chronic pathophysiology where tumor-related complications (e.g., malignant effusion, infection, thromboembolism) interact with critical illness to produce highly variable clinical trajectories [4]. Therefore, focusing on ICU-specific heterogeneity is essential for guiding time-sensitive, life-saving interventions. Cluster analysis is particularly valuable in critical care settings, where patients present with complex, overlapping physiological disturbances and high dimensional clinical data [5]. By integrating multiple physiological, laboratory, and severity indicators simultaneously, clustering can uncover hidden patient subgroups that traditional single-score systems fail to identify, enabling more precise risk stratification and subtype‑specific intervention [6]. In this context, cluster analysis, as a method capable of integrating multi-dimensional data and uncovering potential population stratification structures, provides a powerful tool for precision cancer subtyping [7, 8]. Compared to traditional subtyping methods reliant on pathological morphology or single indicators, cluster analysis can synthesize multi-dimensional information such as demographic characteristics, laboratory parameters, and molecular expression profiles, quantify similarities and differences between samples, and identify clinically relevant subtypes. For example, cluster analysis based on senescence-related genes identified three subtypes with distinct prognostic features in hepatocellular carcinoma [9]; Dyikanov et al. successfully distinguished five cancer immune phenotypes through clustering of multi-parameter flow cytometry and RNA-seq data [10]. We acknowledge that while molecular and pathological characteristics are important sources of tumor heterogeneity, they are not available in the critical care databases used in this study. Therefore, our subtyping system focuses on clinically accessible variables to ensure rapid bedside applicability in time-sensitive ICU settings. However, clustering studies integrating multi-dimensional clinical data for critically ill ovarian cancer patients in the ICU remain relatively scarce, and existing subtyping approaches generally lack robust cross-center and cross-cohort validation, limiting their clinical translation and application. To address these gaps, the present study aimed to develop and characterize a robust clinical subtyping system for critically ill ovarian cancer patients admitted to the ICU. Using two large critical care databases (MIMIC-IV and eICU), we applied a multi-algorithm clustering framework (MCAOCS) to identify clinically distinct phenotypes. We then evaluated the robustness and generalizability of these subtypes in internal and external application cohorts, and further explored their prognostic implications and subtype-specific treatment responses. This framework may facilitate improved risk stratification and individualized critical care decision-making for ovarian cancer patients admitted to the ICU. Materials and methods Data sources and patient selection Data sources Two critical care databases were analyzed in this study. The Medical Information Mart for Intensive Care IV version 3.1 (MIMIC-IV v3.1) database (2008–2019, single-center) served as the discovery cohort [11]. The eICU Collaborative Research Database version 2.0 (eICU v2.0) (2014–2015, multi-center, 208 hospitals) served as the external application cohort [12]. Study population Female patients with a diagnosis of ovarian cancer (OC) were identified using International Classification of Diseases codes: ICD-9 183.0 and ICD-10 C56 (malignant neoplasm of the ovary). Exclusion criteria The following exclusion criteria were applied: (1) male patients or patients with missing sex information, (2) age < 18 years, (3) ICU length of stay < 24 h, (4) patients with multiple ICU admissions (only the first ICU stay was retained), and (5) patients with incomplete or missing critical clinical variables (defined as > 30% missingness across key variable categories). The detailed patient selection process, including the number of patients excluded at each step, is presented in Fig. 1. Fig. 1Patient selection flowchart Patient selection flowchart Data processing and variable extraction Variable selection and categorization Variables were categorized into several groups: demographics [age, race, weight, height, and body mass index (BMI)], vital signs [mean arterial pressure (MAP), systolic blood pressure (SBP), diastolic blood pressure (DBP), heart rate, temperature, respiratory rate, and oxygen saturation], laboratory values [white blood cell (WBC) count, hemoglobin, platelet count], biochemistry [creatinine, blood urea nitrogen (BUN), electrolytes, glucose], severity scores [Simplified Acute Physiology Score II (SAPS II), Oxford Acute Severity of Illness Score (OASIS), Sequential Organ Failure Assessment (SOFA), Glasgow Coma Scale (GCS), and Charlson Comorbidity Index], interventions (mechanical ventilation, renal replacement therapy), and medications (vasopressors, sedatives, opioids, and other supportive medications). Additionally, ovarian cancer-specific variables were extracted, including CA-125 levels (categorized as normal [< 35 U/mL], elevated [35–500 U/mL], and high [> 500 U/mL]), presence of ascites, malignant pleural effusion, and tumor-related anemia severity (categorized by hemoglobin levels as severe [< 8 g/dL], moderate [8–10 g/dL], mild [10–12 g/dL], and normal [≥ 12 g/dL]). Treatment-related oncologic information was also collected from the MIMIC-IV database, including pre-ICU surgical procedures (cytoreductive surgery, total abdominal hysterectomy with bilateral salpingo-oophorectomy [TAH-BSO], other surgery, or no surgery), in-hospital chemotherapy regimen (carboplatin/paclitaxel, other regimen, or no chemotherapy), surgery-to-ICU time interval (≤ 24 h, 1–3 days, 3–7 days, or > 7 days), and postoperative complications (infection, hemorrhage, venous thromboembolism [VTE], or none). These oncologic variables were available only in the MIMIC-IV cohort and were therefore analyzed exclusively within this dataset. Data cleaning and imputation Variables with > 30% missing values were excluded. The remaining missing values were imputed using multiple imputation by chained equations with the R package ‘mice’, as this approach preserves statistical power and reduces bias compared to complete-case analysis or single imputation. Continuous variables were imputed using predictive mean matching, binary variables using logistic regression, and categorical variables with more than two levels using polytomous regression. Five imputed datasets were generated, a standard recommendation that balances robustness against computational efficiency. Because clustering algorithms cannot directly pool results across multiple imputed datasets, we collapsed the five imputed datasets into a single complete dataset by averaging continuous values and taking the mode for categorical values. To quantitatively assess whether the imputation process altered the clustering structure, we performed clustering independently on each of the five imputed datasets and compared the resulting cluster assignments to the reference clustering solution (derived from the collapsed dataset) using multiple agreement metrics: Adjusted Rand Index (ARI), Normalized Mutual Information (NMI), Jaccard index, and proportion of consistently assigned patients. The comparisons demonstrated high agreement across all imputed datasets (mean ARI = 0.82 ± 0.02, mean NMI = 0.78 ± 0.02, mean Jaccard = 0.88 ± 0.02, mean consistency = 93.1% ± 1.2%), indicating that multiple imputation did not substantially alter the identified cluster structure (Table S2). Identification of ovarian cancer subtypes To identify distinct OC subtypes, we implemented Multiple Clustering Algorithms for OC Subtyping (MCAOCS) using the MOVICS methodology. We used the MIMIC-IV database as the discovery cohort and employed consensus clustering to ensure robust subtype discovery. The MCAOCS approach synthesizes results from multiple clustering algorithms including hierarchical clustering, k-means, Partitioning Around Medoids (PAM), and spectral clustering to reduce algorithmic bias and random effects, thereby producing more stable and reliable classifications. Our analysis incorporated comprehensive clinical data including demographics (age, weight, height, BMI), vital signs (mean arterial pressure, SBP, diastolic blood pressure, heart rate, temperature, respiratory rate, oxygen saturation), laboratory values (WBC count, hemoglobin, platelet count), biochemistry parameters (creatinine, BUN, electrolytes, glucose, bicarbonate), severity scores (SAPS II, OASIS, SOFA, GCS, Charlson Comorbidity Index), interventions (sepsis-3 criteria, mechanical ventilation, renal replacement therapy), comorbidities [hypertension, diabetes, renal disease, chronic obstructive pulmonary disease (COPD), coronary artery disease, stroke, malignancy], and medication usage (vasopressors, sedatives, analgesics, and other supportive medications). The MCAOCS process followed four major steps: First, all variables were standardized to have a mean of 0 and standard deviation of 1 to ensure equal contribution to the clustering. Second, multiple clustering algorithms were applied simultaneously to capture different aspects of data structure, with cluster numbers ranging from 2 to 5 evaluated for each algorithm. Third, a consensus matrix was constructed to quantify how frequently pairs of patients clustered together across all algorithm runs, with consensus clustering performed using 1000 iterations with 80% subsampling. Finally, based on consensus matrices, silhouette analysis, gap statistics, and clinical interpretability, we determined that two distinct OC subtypes (designated Subtype 1 and Subtype 2, hereafter referred to as Critical OC and Stable OC, respectively) optimally represented the patient population. The robustness of the identified subtypes was assessed using the eICU database as an external application cohort. External application of subtypes The robustness of the identified subtypes was assessed using two approaches: Internal validation: Within the MIMIC-IV cohort, we compared results from different clustering algorithms, including Nearest Template Prediction (NTP) and Prediction Analysis of Microarrays (hereinafter abbreviated as PAMR to distinguish it from the Partitioning Around Medoids clustering algorithm); External application: The identified subtypes were applied to the eICU database using the same classification algorithms trained on the MIMIC-IV cohort. Clinical characterization and outcome analysis Baseline characteristics Baseline characteristics were compared between subtypes and across databases. Continuous variables were reported as median (interquartile range) and compared using the Kruskal-Wallis test. Categorical variables were reported as counts (percentages) and compared using the chi-square test or Fisher’s exact test when appropriate. Survival analysis Survival outcomes were analyzed for each subtype using Kaplan-Meier curves and compared with the log-rank test. Cox proportional hazards models were fitted to estimate the hazard ratios (HRs) with 95% confidence intervals (CIs). Outcomes assessed included ICU mortality, hospital mortality, and mortality at various time points (28-day, 30-day, 60-day, 90-day, 180-day, and 1-year). Treatment-exposure association analysis using inverse probability of treatment weighting To evaluate the subtype-specific treatment-exposure associations with 28-day mortality while controlling for confounding by indication, we implemented an Inverse Probability of Treatment Weighting (IPTW) approach [13]. This method addresses the inherent bias in observational data where sicker patients are more likely to receive intensive treatments, potentially masking estimated associations. For each medication and OC subtype combination, we calculated propensity scores using logistic regression models that included baseline confounders: demographics (age, BMI), disease severity scores (SAPS II, OASIS, SOFA, GCS, Charlson Comorbidity Index), vital signs (MAP, heart rate, respiratory rate), key laboratory values (creatinine, hemoglobin), and cancer-related clinical factors (CA-125 level category, pre-ICU surgical procedure type, in-hospital chemotherapy status, and surgery-to-ICU time interval). Given that 9 medications were evaluated across 2 subtypes (yielding 18 comparisons), p-values from the IPTW-weighted Cox regression analyses were adjusted for multiple testing using the Benjamini-Hochberg (BH) method [14]. Both raw and BH-adjusted p-values are reported, with statistical significance determined based on the BH-adjusted p-values. These propensity scores were converted to stabilized inverse probability weights to maintain effective sample size. The IPTW analysis proceeded as follows: First, covariate balance was assessed using standardized mean differences (SMD), with SMD < 0.1 indicating adequate balance. Second, weighted Cox proportional hazards regression with robust standard errors was performed to estimate hazard ratios (HRs) and 95% confidence intervals for 28-day mortality. Third, we identified subtype-specific associations where medications demonstrated differential associations between subtypes. Statistical analysis All statistical analyses were performed using R version 4.3.0 (R Foundation for Statistical Computing, Vienna, Austria). Continuous variables were compared using the Kruskal-Wallis test, and categorical variables were compared using the chi-square test or Fisher’s exact test. Survival analyses were conducted using the ‘survival’ and ‘survminer’ packages. Clustering analyses were performed using the ‘MOVICS’ package. P-values < 0.05 were considered statistically significant. For multiple comparisons, p-values were adjusted using the Benjamini-Hochberg method [14]. Results Baseline characteristics This study included a total of 448 ICU patients with OC, comprising 189 from the eICU database and 259 from the MIMIC-IV database (Fig. 1). The age of the MIMIC-IV group was slightly lower than that of the eICU group, with a statistically significant difference (P = 0.042, SMD = 0.205). Regarding disease severity scores, the MIMIC-IV cohort had significantly lower SAPS II, OASIS, GCS scores, and Charlson Comorbidity Index compared to the eICU cohort (P < 0.01), while the SOFA score was higher in the eICU cohort than in the MIMIC-IV cohort (P < 0.001). Although the Charlson Comorbidity Index differed significantly between the two cohorts, the distributions of specific comorbidities such as hypertension (P = 0.360), diabetes (P = 0.799), and chronic kidney disease (P = 0.976) were relatively consistent. For physiological and pathological indicators, diastolic blood pressure (DBP) and oxygen saturation (SpO₂) showed statistically significant differences between the cohorts (P < 0.05), whereas systolic blood pressure (SBP), hemoglobin, and creatinine levels showed no significant differences (P > 0.05). In terms of clinical interventions and outcomes, the MIMIC-IV cohort had significantly higher vasopressor use and longer hospital stays than the eICU cohort (P < 0.05), but no significant differences were observed in interventions such as mechanical ventilation, sedative use, or ICU mortality (P > 0.05). Overall, the two cohorts of ICU patients with OC were comparable in terms of race, common comorbidities, and mortality, but exhibited some heterogeneity in disease severity, physiological/pathological indicators, and treatment patterns (Table 1). Additionally, we provided a comprehensive comparison of all available clinical variables across the databases (Table S1). These baseline differences in disease severity (SAPS II, OASIS, SOFA), physiological parameters (DBP, SpO2), and treatment patterns (vasopressor use) should be considered when interpreting subsequent subtype analyses, as they may influence the generalizability of findings across cohorts. Table 1Baseline characteristics of ovarian cancer patients across databasesCharacteristicsLevelOverall(n = 448)eICU (n = 189)MIMIC-IV(n = 259)P-valueSMDAge (years) (median [Q1,Q3])61.00[47.75,72.00]62.00[52.00,73.00]59.00[44.00,71.00]0.0420.205RaceAsian20(4.5)10(5.3)10(3.9)0.8160.093Black48(10.7)18(9.5)30(11.6)Other83(18.5)35(18.5)48(18.5)White297(66.3)126(66.7)171(66.0)SAPS II score (median [Q1,Q3])33.00[22.00,43.00]28.00[19.00,39.00]36.00[25.00,45.00]< 0.0010.465SOFA score (median [Q1,Q3])4.00[2.00,7.00]5.00[3.00,8.00]4.00[2.00,6.00]< 0.0010.344Charlson comorbidity index (median [Q1,Q3])5.00[2.00,8.00]4.00[2.00,7.00]6.00[2.00,8.00]0.0050.279Heart rate (bpm) (median [Q1,Q3])95.00[80.00,109.00]93.00[80.00,112.00]95.00[80.00,108.50]0.6400.044SBP (mmHg) (median [Q1,Q3])119.00[101.00,138.00]119.00[100.00,139.00]119.00[102.50,137.00]0.9380.044DBP (mmHg) (median [Q1,Q3])66.00[55.00,77.25]63.00[54.00,75.00]67.00[57.00,80.00]0.0340.252MAP (mmHg) (median [Q1,Q3])82.00[70.00,96.00]82.00[69.00,94.00]81.00[70.00,97.00]0.3960.064Hemoglobin (g/dL) (median [Q1,Q3])10.40[8.90,11.80]10.40[8.80,11.70]10.40[9.00,11.90]0.4470.080Creatinine (mg/dL) (median [Q1,Q3])0.80[0.60,1.10]0.81[0.65,1.10]0.80[0.60,1.15]0.0930.080Mechanical ventilationNo277(61.8)112(59.3)165(63.7)0.3910.091Yes171(38.2)77(40.7)94(36.3)HypertensionNo234(52.2)104(55.0)130(50.2)0.3600.097Yes214(47.8)85(45.0)129(49.8)Diabetes mellitusNo378(84.4)158(83.6)220(84.9)0.7990.037Yes70(15.6)31(16.4)39(15.1)Chronic kidney diseaseNo402(89.7)169(89.4)233(90.0)0.9760.018Yes46(10.3)20(10.6)26(10.0) Table 1Baseline characteristics of ovarian cancer patients across databasesCharacteristicsLevelOverall(n = 448)eICU (n = 189)MIMIC-IV(n = 259)P-valueSMDAge (years) (median [Q1,Q3])61.00[47.75,72.00]62.00[52.00,73.00]59.00[44.00,71.00]0.0420.205RaceAsian20(4.5)10(5.3)10(3.9)0.8160.093Black48(10.7)18(9.5)30(11.6)Other83(18.5)35(18.5)48(18.5)White297(66.3)126(66.7)171(66.0)SAPS II score (median [Q1,Q3])33.00[22.00,43.00]28.00[19.00,39.00]36.00[25.00,45.00]< 0.0010.465SOFA score (median [Q1,Q3])4.00[2.00,7.00]5.00[3.00,8.00]4.00[2.00,6.00]< 0.0010.344Charlson comorbidity index (median [Q1,Q3])5.00[2.00,8.00]4.00[2.00,7.00]6.00[2.00,8.00]0.0050.279Heart rate (bpm) (median [Q1,Q3])95.00[80.00,109.00]93.00[80.00,112.00]95.00[80.00,108.50]0.6400.044SBP (mmHg) (median [Q1,Q3])119.00[101.00,138.00]119.00[100.00,139.00]119.00[102.50,137.00]0.9380.044DBP (mmHg) (median [Q1,Q3])66.00[55.00,77.25]63.00[54.00,75.00]67.00[57.00,80.00]0.0340.252MAP (mmHg) (median [Q1,Q3])82.00[70.00,96.00]82.00[69.00,94.00]81.00[70.00,97.00]0.3960.064Hemoglobin (g/dL) (median [Q1,Q3])10.40[8.90,11.80]10.40[8.80,11.70]10.40[9.00,11.90]0.4470.080Creatinine (mg/dL) (median [Q1,Q3])0.80[0.60,1.10]0.81[0.65,1.10]0.80[0.60,1.15]0.0930.080Mechanical ventilationNo277(61.8)112(59.3)165(63.7)0.3910.091Yes171(38.2)77(40.7)94(36.3)HypertensionNo234(52.2)104(55.0)130(50.2)0.3600.097Yes214(47.8)85(45.0)129(49.8)Diabetes mellitusNo378(84.4)158(83.6)220(84.9)0.7990.037Yes70(15.6)31(16.4)39(15.1)Chronic kidney diseaseNo402(89.7)169(89.4)233(90.0)0.9760.018Yes46(10.3)20(10.6)26(10.0) Baseline characteristics of ovarian cancer patients across databases Identification of ovarian cancer subtypes Based on the MIMIC-IV database, we identified OC subtypes using the MCAOCS algorithm and compared the performance of different clustering algorithms. Despite the diversity of algorithms, their classification patterns were broadly similar, ultimately identifying two subtypes (Fig. 2A). The consensus matrix demonstrated stable clustering results for the two subtypes, characterizing the reproducibility of subtype clustering patterns (Fig. 2B). Furthermore, silhouette analysis indicated that the silhouette widths for individual clusters and the overall clustering were at or near good standards, with minimal difference in the average silhouette width between the two subtypes, suggesting the reliability of the clustering results (Figure S1). Fig. 2Identification of ovarian cancer subtypes. A Comparison of clustering results using different algorithms for ovarian cancer patients in the MIMIC-IV database. Each column represents a patient, and each row represents a clustering algorithm. Colors indicate cluster assignment. B Consensus clustering heatmap showing the two identified ovarian cancer subtypes. The heatmap represents the consensus matrix where darker blue indicates higher probability of patients being clustered together across multiple clustering runs Identification of ovarian cancer subtypes. A Comparison of clustering results using different algorithms for ovarian cancer patients in the MIMIC-IV database. Each column represents a patient, and each row represents a clustering algorithm. Colors indicate cluster assignment. B Consensus clustering heatmap showing the two identified ovarian cancer subtypes. The heatmap represents the consensus matrix where darker blue indicates higher probability of patients being clustered together across multiple clustering runs Clinical phenotypes of the two ovarian cancer subtypes Further comparative analysis of the clinical characteristics of the two OC subtypes in the MIMIC-IV cohort was conducted (Fig. 3). Critical OC exhibited a critical high-risk phenotype, characterized by manifestations such as circulatory instability (hypotension), multi-organ involvement (SOFA ≥ 5), concomitant sepsis, requirement for vasoactive drugs, as well as features indicating poor prognosis including SAPS II > 40, high Charlson Comorbidity Index, and very low albumin levels. Stable OC showed normal vital signs (blood pressure, body temperature), no significant abnormalities in electrolytes or organ function (renal, metabolic), absence of sepsis or circulatory failure, along with SAPS II < 30, SOFA < 4, and fewer comorbidities, representing a relatively stable low-risk phenotype. The study further compared the two OC subtypes in the eICU cohort, which also demonstrated highly consistent clinical logic with the MIMIC-IV cohort subtype classification: the two subtypes exhibited critical high-risk and stable low-risk characteristics, respectively (Figure S2), further revealing the stability of the subtype features. Fig. 3Clinical characteristics of ovarian cancer subtypes in MIMIC-IV Heatmap showing the clinical features of the two ovarian cancer subtypes (Critical OC and Stable OC) in the MIMIC-IV cohort. Features include demographics, vital signs, laboratory values, severity scores, interventions, comorbidities, and medications. Z-scores are shown for continuous variables, and binary indicators (Yes/No) are shown for categorical variables. Asterisks indicate statistical significance: *p < 0.05, **p < 0.01, ***p < 0.001 Clinical characteristics of ovarian cancer subtypes in MIMIC-IV Heatmap showing the clinical features of the two ovarian cancer subtypes (Critical OC and Stable OC) in the MIMIC-IV cohort. Features include demographics, vital signs, laboratory values, severity scores, interventions, comorbidities, and medications. Z-scores are shown for continuous variables, and binary indicators (Yes/No) are shown for categorical variables. Asterisks indicate statistical significance: *p < 0.05, **p < 0.01, ***p < 0.001 In addition, ovarian cancer-specific clinical features differed significantly between the two subtypes in the MIMIC-IV cohort (Figure S3). The Critical OC subtype had a markedly higher proportion of patients with CA-125 levels > 500 U/mL (55.9% vs. 27.0%), malignant pleural effusion (43.2% vs. 12.8%), and severe tumor-related anemia (13.6% vs. 4.3%) (all p < 0.001), whereas no significant difference was observed for ascites (p = 0.130). Treatment-related oncologic characteristics also differed between subtypes (Figure S4): the Critical OC subtype had higher rates of cytoreductive surgery (31.4% vs. 14.2%, p = 0.004) and chemotherapy exposure (63.6% vs. 44.0%, p = 0.006); among patients who underwent surgery, a greater proportion of Critical OC patients were admitted to the ICU within 24 h postoperatively (46.4% vs. 25.9%, p = 0.008), and postoperative complications were more frequent (69.0% vs. 38.3%, p = 0.001). These oncologic variables were available only in the MIMIC-IV cohort and could not be assessed in the eICU database. Survival analysis of ovarian cancer subtypes Survival outcomes of the two subtypes in the MIMIC-IV cohort were analyzed using Kaplan-Meier curves and the log-rank test (Fig. 4). The survival curves of the two subtypes were clearly separated, indicating significant differences in mortality between the two clustered patient groups. From 28 days to one year of follow-up, the mortality of Critical OC was significantly higher than that of Stable OC (P < 0.05). The significant survival difference indicates prognostic variation between the two OC subtypes and demonstrates the reliability and stability of the classification. Fig. 4Survival Analysis of Ovarian Cancer Subtypes Kaplan-Meier curves showing differences in survival between ovarian cancer subtypes Critical OC and Stable OC. A 28-day mortality B 60-day mortality (C) 90-day mortality (D) 1-year mortality Survival Analysis of Ovarian Cancer Subtypes Kaplan-Meier curves showing differences in survival between ovarian cancer subtypes Critical OC and Stable OC. A 28-day mortality B 60-day mortality (C) 90-day mortality (D) 1-year mortality External application of ovarian cancer subtypes External application was performed using two established classification algorithms to further evaluate the reproducibility of the classification. We assessed the consistency of the MCAOCS, NTP, and PAMR classification methods, with Kappa coefficients of 0.693, 0.860, and 0.632, respectively (all P < 0.001), indicating high agreement in the classification of OC subtypes among the three clustering algorithms and reflecting the reliability of the classification (Fig. 5A, B, and C). For external application, the NTP and PAMR classifiers trained exclusively on the MIMIC-IV discovery cohort were applied to the independent eICU cohort, yielding a Kappa coefficient of 0.683 (P < 0.001), supporting the cross-database reproducibility of the subtype classification (Fig. 5D). Fig. 5External Application of Ovarian Cancer Subtypes. A Agreement between MCAOCS and NTP clustering algorithms in the MIMIC-IV database. Kappa statistic indicates the level of agreement. B Agreement between MCAOCS and PAM clustering algorithms in the MIMIC-IV database. C Agreement between NTP and PAM clustering algorithms in the MIMIC-IV database. D Agreement between NTP and PAMR classifiers in the external application cohort (eICU database) External Application of Ovarian Cancer Subtypes. A Agreement between MCAOCS and NTP clustering algorithms in the MIMIC-IV database. Kappa statistic indicates the level of agreement. B Agreement between MCAOCS and PAM clustering algorithms in the MIMIC-IV database. C Agreement between NTP and PAM clustering algorithms in the MIMIC-IV database. D Agreement between NTP and PAMR classifiers in the external application cohort (eICU database) Associations of subtype-specific treatments with 28-day mortality The association between treatment modalities and 28-day mortality was evaluated using IPTW-adjusted Cox regression, with p-values corrected for multiple comparisons using the Benjamini-Hochberg (BH) method across 18 medication–subtype combinations (Fig. 6). In the Critical OC subtype, six interventions were associated with lower 28-day mortality after BH correction: fentanyl (HR 0.16, 95% CI 0.04–0.72, BH-adjusted p = 0.039), insulin (HR 0.33, 95% CI 0.14–0.76, BH-adjusted p = 0.029), mechanical ventilation (HR 0.24, 95% CI 0.08–0.72, BH-adjusted p = 0.029), midazolam (HR 0.21, 95% CI 0.07–0.60, BH-adjusted p = 0.018), and propofol (HR 0.24, 95% CI 0.10–0.56, BH-adjusted p = 0.016). Sedatives were also associated with lower 28-day mortality (HR 0.22, 95% CI 0.07–0.71, BH-adjusted p = 0.029). Norepinephrine (HR 0.11, 95% CI 0.01–0.93, raw p = 0.043) and vasopressors (HR 0.37, 95% CI 0.13–0.99, raw p = 0.048) showed trends toward association with lower mortality in the Critical OC subtype but did not retain significance after BH correction (BH-adjusted p = 0.071 and 0.072, respectively). In the Stable OC subtype, only two interventions retained significance after BH correction: insulin (HR 0.30, 95% CI 0.14–0.68, BH-adjusted p = 0.018) and phenylephrine (HR 0.25, 95% CI 0.10–0.60, BH-adjusted p = 0.017). Fig. 6Subtype-specific medication associations with 28-Day Mortality in Ovarian Cancer Patients. Forest plot showing the subtype-specific treatment-exposure associations with 28-day mortality using inverse probability of treatment weighting (IPTW) analysis in the MIMIC-IV cohort. Hazard ratios (HRs) with 95% confidence intervals (CIs) are displayed for each medication stratified by ovarian cancer subtype (Critical OC vs. Stable OC). The vertical dashed line at HR = 1.0 represents the null value (no association). HRs < 1.0 (left side) indicate associations with lower estimated 28-day mortality, whereas HRs > 1.0 (right side) indicate associations with higher estimated 28-day mortality. Statistical significance is denoted by asterisks (*p < 0.05). The analysis was adjusted for baseline confounders including demographics, disease severity scores, and clinical characteristics through IPTW methodology. These associations should be interpreted as hypothesis-generating rather than causal Subtype-specific medication associations with 28-Day Mortality in Ovarian Cancer Patients. Forest plot showing the subtype-specific treatment-exposure associations with 28-day mortality using inverse probability of treatment weighting (IPTW) analysis in the MIMIC-IV cohort. Hazard ratios (HRs) with 95% confidence intervals (CIs) are displayed for each medication stratified by ovarian cancer subtype (Critical OC vs. Stable OC). The vertical dashed line at HR = 1.0 represents the null value (no association). HRs < 1.0 (left side) indicate associations with lower estimated 28-day mortality, whereas HRs > 1.0 (right side) indicate associations with higher estimated 28-day mortality. Statistical significance is denoted by asterisks (*p < 0.05). The analysis was adjusted for baseline confounders including demographics, disease severity scores, and clinical characteristics through IPTW methodology. These associations should be interpreted as hypothesis-generating rather than causal Discussion This study established a binary subtype system comprising Critical OC and Stable OC in critically ill ovarian cancer patients. Cross-database application and multi-dimensional quality assessment confirmed the robustness of this subtyping system, and the two subtypes exhibited distinctly different associations with mortality and treatment responses. The binary system complements previous multi-omics-based molecular subtyping research [15, 16]. While those studies revealed the biological origins of ovarian cancer heterogeneity at the genomic level, acquiring and analyzing complex molecular data is often impractical in the time-sensitive ICU environment. Therefore, this study aimed to develop a subtyping system grounded in clinical feasibility that enables rapid translation and application in resource-constrained ICU settings, thereby advancing precision medicine concepts to the frontline of critical care decision-making. Previous studies have established that inflammatory markers such as the neutrophil-to-lymphocyte ratio and platelet-to-lymphocyte ratio are independent prognostic factors in ovarian cancer, predicting long-term survival [17–19]. Our study extends this literature by demonstrating that systemic inflammatory hyperactivation—reflected in these same markers—characterizes the Critical OC subtype within the ICU setting, suggesting that inflammation-driven heterogeneity exists even among critically ill ovarian cancer patients. However, our study’s contribution lies in transcending the limitations of single indicators or simple combinations by integrating these disparate inflammatory markers with hemodynamic and organ function data into a comprehensive clinical phenotype framework, thereby providing a more holistic and systematic assessment of patient status. In critical care medicine, several pioneering studies have attempted to subtype conditions such as sepsis [20] and acute respiratory distress syndrome [21] to overcome heterogeneity barriers and achieve precision therapy. Our study successfully applies this advanced “subtyping thinking” to critically ill ovarian cancer patients, filling a gap in precision critical care management for this specific population. Our findings complement and extend a previous study by Gao et al. [22], which applied unsupervised clustering to general ICU cancer patients using the same databases. While Gao et al. [22] identified three phenotypes across multiple cancer types, our study focuses specifically on ovarian cancer and incorporates ovarian cancer‑specific clinical features (e.g., CA-125, ascites, pleural effusion, surgical and chemotherapy history). Moreover, our MCAOCS framework integrates multiple clustering algorithms to reduce algorithmic bias, whereas Gao et al. [22] used k-means alone. The binary “Critical OC-Stable OC” system provides a more tailored stratification for ovarian cancer patients. We propose that the origin of this subtype differentiation lies in the complex interaction between local tumor-related damage (e.g., necrosis, infection, malignant effusion) and the host’s systemic inflammatory response [23, 24]. The Critical OC subtype presents an uncontrolled state with circulatory failure, multi‑organ injury, and high inflammatory burden [25, 26]. The underlying pathophysiological basis is likely an excessive, decompensated systemic inflammatory response to tumor burden and concurrent infections [27, 28]. While intrinsic molecular heterogeneity (e.g., pro-coagulant features of clear cell carcinoma [29–32], abnormal metabolic pathways [33–35], and immune microenvironment dysregulation [36]) may contribute to the observed clinical phenotypes, our study did not have histologic or molecular data to directly test these mechanisms. These mechanistic interpretations remain speculative and require further independent clinical and experimental investigation in future studies. In contrast, the Stable OC subtype exhibits a compensated clinical phenotype. Its “low-stress–organ compensation” characteristics may be related to well-controlled tumor burden and preserved basic immune regulatory capacity [37]. The relative stability in these patients likely benefits from their relatively lower tumor burden and reduced risk of co-infection [38]. These factors collectively restrict the intensity and scope of the downstream systemic inflammatory response, allowing the body’s compensatory mechanisms to function effectively [39, 40]. In summary, through the “Critical OC-Stable OC” binary system, this study confirms that heterogeneity in critically ill ovarian cancer is reflected not only in clinical manifestations but also in deep-seated mechanisms ranging from systemic inflammatory responses to tumor molecular characteristics. From the perspective of comprehensive clinical phenotypes, this study reveals a distinction between uncontrolled and compensated states in the systemic inflammatory response of critically ill ovarian cancer patients. This finding moves the clinical understanding of heterogeneity toward subtype differentiation based on objective indicators and underlying pathophysiological logic. Importantly, the subtype-specific treatment responses identified here are consistent with and also extend existing concepts in critical care medicine. Our analytical results show that for the Critical OC subtype, multiple interventions—including fentanyl, insulin, mechanical ventilation, midazolam, propofol, sedatives, and vasopressors—were associated with lower 28-day mortality. This suggests that active, comprehensive multi-organ support may be important for this patient subtype. In contrast, the Stable OC subtype showed associations with lower mortality only for a limited number of interventions (insulin and phenylephrine). These findings suggest that different treatment intensities might be considered according to subtype, these subtype-linked statistical associations only provide exploratory clues for subsequent clinical investigation, rather than serving as direct therapeutic recommendations. The present study has several clinical and methodological advantages. First, we employed a multi-center database design, incorporating patient data from two independent databases (MIMIC-IV and eICU), and confirmed the robustness of the subtype classification through internal and external application. This cross-cohort reproducibility assessment is particularly important yet often overlooked in clinical subtyping research. Second, we integrated multi-dimensional clinical data, including demographics, vital signs, laboratory tests, severity scores, and treatment information, providing a comprehensive assessment of patient status. Third, we used a consensus clustering approach, combining multiple clustering algorithms to reduce bias inherent in any single algorithm and improve classification reliability. Finally, we analyzed subtype-specific treatment associations using IPTW with multiple comparison correction, providing information that may be useful for clinical translation. However, this study also has some limitations. First, it is based on retrospective data. Although a multi-center design and internal-external application were used, prospective studies are needed to further verify the subtype classification. Second, the study primarily relied on clinical data and lacked tumor molecular characteristic information (such as BRCA mutation status, homologous recombination deficiency), preventing a complete revelation of the biological basis of phenotypic differences. Third, treatment allocation was non-randomized. Although we used IPTW to control for confounding, residual confounding may still exist. Fourth, while we refer to eICU as an “external application” cohort, the classification algorithms were trained on MIMIC-IV and then applied to eICU. This constitutes external application rather than true independent validation. Future studies should perform true external validation using an entirely independent cohort. Fifth, detailed oncologic information that could help interpret the identified phenotypes—including FIGO stage, histologic subtype (e.g., high-grade serous, clear cell, endometrioid), prior lines of chemotherapy, exposure to PARP inhibitors or bevacizumab, and hormone therapy—was not available in the MIMIC-IV or eICU databases. These factors may influence both the clinical course of ovarian cancer and the response to ICU interventions, and their absence is an important limitation. Sixth, the ICU patients included may not represent all ovarian cancer patient populations, particularly those with advanced disease not admitted to the ICU. Future work should involve prospective cohort studies to validate this classification in independent populations, explore simplified bedside scoring systems, incorporate dynamic monitoring indicators, and integrate multi-omics data to better understand the biological basis of the observed phenotypes. Conclusions In summary, this study identified and characterized two clinically distinct subtypes, Critical OC and Stable OC, in critically ill ovarian cancer patients through multi-dimensional cluster analysis. This classification system demonstrates good robustness and reproducibility, and may support risk stratification for critically ill ovarian cancer patients, providing a potentially useful framework for precision critical care. Future research should focus on validating this classification in prospective cohorts, exploring its biological basis, and generating hypotheses for future subtype-stratified treatment studies, ultimately aiming to improve clinical outcomes for this vulnerable patient population. Supplementary Information Supplementary Material 1: Table S1. Comprehensive Comparison of All Variables in Ovarian Cancer Patients Between Databases. Supplementary Material 1: Table S1. Comprehensive Comparison of All Variables in Ovarian Cancer Patients Between Databases. Supplementary Material 2: Table S2. Clustering stability across multiply imputed datasets. Supplementary Material 2: Table S2. Clustering stability across multiply imputed datasets. Supplementary Material 3: Figure S1. Silhouette Analysis of Ovarian Cancer Subtypes Silhouette plot showing the quality of clustering for each patient in the identified subtypes. Higher silhouette width indicates better clustering quality. The average silhouette width for each cluster is shown at the bottom. Supplementary Material 3: Figure S1. Silhouette Analysis of Ovarian Cancer Subtypes Silhouette plot showing the quality of clustering for each patient in the identified subtypes. Higher silhouette width indicates better clustering quality. The average silhouette width for each cluster is shown at the bottom. Supplementary Material 4: Figure S2. Clinical Characteristics of Ovarian Cancer Subtypes in eICU. Supplementary Material 4: Figure S2. Clinical Characteristics of Ovarian Cancer Subtypes in eICU. Supplementary Material 5: Figure S3. Ovarian Cancer-Specific Clinical Features by Subtype in the MIMIC-IV Cohort. Comparison of tumor-related clinical characteristics between Critical OC (n=118) and Stable OC (n=141), including CA-125 levels, ascites, malignant pleural effusion, and tumor-related anemia severity. P-values were calculated using chi-square tests. Supplementary Material 5: Figure S3. Ovarian Cancer-Specific Clinical Features by Subtype in the MIMIC-IV Cohort. Comparison of tumor-related clinical characteristics between Critical OC (n=118) and Stable OC (n=141), including CA-125 levels, ascites, malignant pleural effusion, and tumor-related anemia severity. P-values were calculated using chi-square tests. Supplementary Material 6: Figure S4. Treatment-Related Oncologic Characteristics by Subtype in the MIMIC-IV Cohort. Comparison of pre-ICU surgical procedures, in-hospital chemotherapy regimens, surgery-to-ICU time interval, and postoperative complications between Critical OC and Stable OC. Surgery-to-ICU time interval and postoperative complications were analyzed among patients who underwent surgical procedures (Critical OC: n=84; Stable OC: n=81). P-values were calculated using chi-square tests. Supplementary Material 6: Figure S4. Treatment-Related Oncologic Characteristics by Subtype in the MIMIC-IV Cohort. Comparison of pre-ICU surgical procedures, in-hospital chemotherapy regimens, surgery-to-ICU time interval, and postoperative complications between Critical OC and Stable OC. Surgery-to-ICU time interval and postoperative complications were analyzed among patients who underwent surgical procedures (Critical OC: n=84; Stable OC: n=81). P-values were calculated using chi-square tests.