Prediction models for endometriosis-associated ovarian malignancy: A systematic review and meta-analysis

Xulian Tu: Writing – original draft, Data curation, Conceptualization. Wan Xie: Writing – review & editing, Data curation. Shuai Han: Writing – original draft, Formal analysis. Mingyue Leng: Formal analysis. Xueqing Yang: Formal analysis. Xia Liu: Writing – review & editing, Supervision, Project administration.

This systematic review was conducted in accordance with PRISMA 2020[ 6 ] and informed by the CHARMS checklist [ 7 ] for reviews of prediction-model studies. The protocol was registered in PROSPERO (CRD420251052389). The review was conducted in accordance with the registered protocol, with no material deviations. We organized the evidence by clinical purpose: longitudinal risk prediction, preoperative discrimination, external validation of an existing tool, and other diagnostic or modeling studies. Performance from model development, internal validation, and external validation was summarized separately. Predictor estimates were pooled only when definitions, units, reference categories, outcomes, and adjustment status were compatible. PubMed, Embase, Web of Science, the Cochrane Library, CINAHL, Scopus, China National Knowledge Infrastructure (CNKI), Wanfang, VIP, and SinoMed were searched from inception to 24 May 2025. Search concepts covered endometriosis or ovarian endometrioma, malignant transformation or EAOC, and prediction models, risk scores, nomograms, machine learning, discrimination, calibration, or validation. The complete database-specific search strategies and applied limits are provided in Supplementary Appendix 1 . Retrieved records were imported into Zotero for deduplication and then screened in Rayyan. No eligible grey-literature reports were identified. Two reviewers independently screened titles and abstracts and subsequently assessed potentially eligible full texts. Disagreements were resolved by discussion and, when required, consultation with a third reviewer. Reasons for exclusion were recorded at the full-text stage and summarized in the PRISMA flow diagram. A study-level list of reports excluded after full-text assessment, with reasons, is provided in Supplementary Appendix 2 . Studies were eligible if they: (1) included women with endometriosis, ovarian endometrioma, or a clinically relevant population undergoing evaluation for EAOC; (2) developed, updated, or evaluated a multivariable model intended to estimate future malignant transformation or to discriminate EAOC from benign endometriosis; and (3) reported sufficient information on model specification or performance. Development and validation studies, including regression-based and machine-learning models, were eligible. Studies evaluating a single biomarker or imaging measurement without a multivariable model were considered separately and were not pooled with multivariable models. We excluded non-original reports, case reports, narrative reviews, studies limited to univariable risk-factor associations without a usable multivariable model, models for non-specific ovarian lesions without EAOC-specific performance, and reports with insufficient information for appraisal. Using a structured extraction form based on CHARMS, two reviewers independently extracted study design, setting, eligibility criteria, participant characteristics, outcome definition and timing, candidate and final predictors, sample size and number of outcome events, modeling method, predictor-selection strategy, missing-data handling, model presentation, and validation approach. Performance measures included the AUC or C-statistic with 95% CI, calibration plots, calibration slope or intercept, observed-to-expected ratio, Brier score, sensitivity, specificity, and decision-curve analysis where reported. Adjusted and unadjusted predictor estimates were recorded separately. Disagreements in data extraction were resolved through discussion and, when consensus could not be reached, adjudication by a third reviewer. Risk of bias and applicability were assessed with PROBAST [ 8 ] across the participants, predictors, outcome, and analysis domains. Each domain was rated as low, high, or unclear risk of bias according to the signaling questions; applicability concerns were rated as low, high, or unclear. Overall judgments were made separately for model development and model evaluation when a report contained more than one analysis. Disagreements were resolved by consensus. Study and model characteristics were summarized in tables. AUCs were pooled only for studies with comparable clinical purposes, populations, outcomes, and evaluation stages. We used the logit transformation to stabilize the variance and keep back-transformed estimates between 0 and 1. Standard errors were derived from reported 95% CIs on the transformed scale. If a 95% CI was not reported, an approximate standard error was derived from the numbers of participants with and without the outcome using the Hanley-McNeil method; confidence intervals based on these approximations are explicitly identified in the Results. Because clinical and methodological heterogeneity was expected, AUCs were analyzed with restricted maximum-likelihood random-effects models and then back-transformed. S E A U C = A U C ( 1 − A U C ) + ( n 1 − 1 ) ( Q 1 − A U C 2 ) + ( n 0 − 1 ) ( Q 2 − A U C 2 ) n 1 ⋅ n 0 Heterogeneity was described with tau-squared, I-squared, and Cochran's Q. Pooling decisions were based on clinical and methodological comparability rather than an I-squared cutoff. We stratified analyses by clinical purpose and validation stage. Further subgroup analyses by design, region, or model type were not feasible because few studies were compatible. Studies were not removed solely to lower heterogeneity. Analyses were performed in Python 3.12.13 using reproducible custom scripts. Predictor ORs were analyzed on the log scale with inverse-variance fixed-effect or DerSimonian-Laird random-effects models, as appropriate. Adjusted ORs were preferred for predictor meta-analysis. Estimates were pooled only when the predictor definition, measurement scale, reference category, clinical outcome, and adjustment structure were judged sufficiently compatible. Continuous and categorical effects, different measurement units, and adjusted and unadjusted estimates were not combined in the same analysis. Results were back-transformed to the OR scale. Predictor associations were not interpreted as evidence of causality or independent predictive contribution.

The searches identified 2999 records, of which 1475 were removed as duplicates. Of 1524 records screened by title and abstract, 1444 were excluded and 80 full-text reports were assessed. Sixty-three full-text reports were excluded, and one additional report was identified by manual reference searching, resulting in 18 included reports ( Fig. 1 ). The study-level exclusions are listed in Supplementary Appendix 2 . Fig. 1 Flow diagram of the study-selection process. Flow diagram of the study-selection process. The 18 reports were published from 2008 to 2024. Thirteen were from China, three from Japan, one from Israel, and one from the United States [ [9] , [10] , [11] , [12] , [13] , [14] , [15] , [16] , [17] , [18] , [19] , [20] , [21] , [22] , [23] , [24] , [25] , [26] ]. Most used retrospective case-control sampling; prospective and longitudinal cohort studies were uncommon. Thus, the evidence mainly concerns preoperative discrimination of existing malignancy rather than future malignant transformation. Table 1 summarizes the study characteristics. Table 1 Characteristics of the included studies. Study Country Study design Data source Sample size (training/validation) Participants Outcome Prediction time window Guo (2023) [ 9 ] China Case-control study Single center 286/72 Patients with EAOC and OEM Predict risk factors for malignant transformation in endometriosis Not reported Wang et al. (2024) [ 10 ] China Case-control study Multicenter 194/83 Perimenopausal patients with OE and EAOC The risk of malignant transformation in perimenopausal patients with ovarian endometriosis Before surgery Li et al. (2021)[ 11 ] China Case-control study Single center 195 Patients with EAOC and OEM Risk prediction for EAOC Preoperative assessment Wang et al. (2022) [ 12 ] China Case-control study Single center 296/148 Patients with OEC and OCCC, grouped according to the presence of endometriosis Prediction of the 5-year risk of malignant transformation in endometriosis Over the next 5 years Wang (2024) [ 13 ] China Case-control study Multicenter 148 Patients with ENDOMETRIOSIS and EAOC Risk prediction for EAOC Preoperative risk Wang (2023) [ 14 ] China Case-control study Multicenter 110 Patients with OCCC and OEC, grouped according to the presence of endometriosis Differential diagnosis of EAOC Preoperative diagnosis Wang et al. (2023) [ 15 ] China Case-control study Single center 90 patients with OCCC and OEC Differential diagnosis of EAOC Before surgery Chen (2023) [ 16 ] China Case-control study Multicenter 468/134 Patients with EAOC and OEM Risk prediction for EAOC Preoperative prediction Yan et al. (2022) [ 17 ] China Case-control study Single center 357 Patients with EAOC and OEM Predict risk factors for malignant transformation in endometriosis Before surgery Xu et al. (2024) [ 18 ] China Case-control study Single center 304/131 Patients with EAOC and OEM Differential diagnosis of EAOC Before surgery Jing et al. (2023) [ 19 ] China Case-control study Multicenter 740/410 Patients with ovarian lesions (including endometriosis) Predicting endometriosis and other ovarian lesions Before surgery Cohen et al. (2021)[ 20 ] Israel Retrospective cohort study Single center 181 Women of reproductive age with suspected endometriosis Distinguish between benign and malignant ovarian masses Before surgery Kawahara et al. (2024) [ 21 ] Japan Prospective cohort study Multicenter 31 Patients over 20 years old with suspected ovarian endometriosis cysts or malignant transformation Distinguish EAOC from OE Before surgery Kobayashi et al. (2008) [ 22 ] Japan Retrospective cohort study Multicenter 6398 Patients with ovarian endometriosis (US diagnosis), aged 20–59 years Predict the risk of ovarian cancer in patients with endometriosis Follow-up of up to 17 years (median 12.8 years Zhou et al. (2018) [ 23 ] China Case-control study Multicenter 208 Patients with EAOC and OEM Predict the risk of malignant transformation in OEM Before surgery Kadan et al. (2015) [ 24 ] USA Case-control study Single center 138 Patients with EAOC and OEM Predict the risk of malignant transformation in OEM Preoperative diagnosis Chao et al. (2022) [ 25 ] China Case-control study Multicenter 6809/1806 Patients with EAOC and those with endometriosis but without ovarian cancer Risk prediction for EAOC Before surgery Kawahara et al. (2022) [ 26 ] Japan Case-control study Single center 150 Patients with OE and EAOC Distinguish EAOC from OE Before surgery EAOC: Endometriosis-associated ovarian cancer; OE: Ovarian endometriosis; OEM: Ovarian endometrioma; OCCC: Ovarian clear cell carcinoma; OEC: Ovarian endometrioid carcinoma Characteristics of the included studies. EAOC: Endometriosis-associated ovarian cancer; OE: Ovarian endometriosis; OEM: Ovarian endometrioma; OCCC: Ovarian clear cell carcinoma; OEC: Ovarian endometrioid carcinoma Logistic regression and nomograms were most common, although the studies also used decision trees, rule-based tools, magnetic resonance measures, and machine learning. Only one report described missing-data handling. Predictor selection often relied on univariable screening, and penalization or shrinkage was uncommon. Internal validation was inconsistent, and independent external validation was rare. Six reports presented a calibration plot, none reported a calibration intercept or slope, two reported a Brier score, and seven presented decision-curve analysis. Five reports gave a complete equation or explicit scoring algorithm, and one provided a web-based dynamic nomogram ( Table 2 ). Table 2 Characteristics and performance of the included prediction models. Study Predictors Model type Variable-selection method Model presentation Missing data handling Validation method AUC (development/validation) Guo (2023) [ 9 ] Gastrointestinal symptoms, tumor diameter, solid components of masses, D-dimer Logistic regression Univariate analysis, multivariate analysis Nomogram Not reported Not reported 0.938 Wang et al. (2024) [ 10 ] Postmenopausal status, progressive dysmenorrhea, infertility, severe endometriosis, masses with blood flow signals detected by ultrasound Logistic regression Univariate analysis, multivariate analysis Nomogram Not reported Internal 0.848/0.880 Li et al. (2021)[ 11 ] Age, body mass index (BMI), mass volume, blood flow signals Logistic regression Univariate analysis, multivariate analysis Nomogram Not reported Not reported 0.925 Wang et al. (2022) [ 12 ] Menstrual irregularity, tumor diameter, CA125, follicle-stimulating hormone (FSH) Decision tree analysis, Gail model Univariate analysis, Decision tree analysis Decision tree Not reported Internal validation 0.783 Wang (2024) [ 13 ] Age, tumor diameter, menopause, ultrasound showing solid components and blood flow signals, HE4, CYFRA21-1 Logistic regression Univariate analysis, multivariate analysis forest plot Not reported Not reported — Wang (2023) [ 14 ] Age at menarche, initial presentation of symptoms of “dysmenorrhea or irregular bleeding,” and the ROMA index Logistic regression Univariate analysis, multivariate analysis nomogram Not reported Not reported 0.842 Wang et al. (2023) [ 15 ] Menopausal status, presence of uterine fibroids or polyps, dysmenorrhea or irregular bleeding, ROMA index Logistic regression Univariate analysis, multivariate analysis Nomogram Not reported Not reported 0.896 Chen (2023) [ 16 ] Age at onset, number of deliveries, menopausal status, maximum tumor diameter, intratumoral blood flow Logistic regression Univariate analysis, multivariate analysis Nomogram Not reported Internal/external 0.9717/0.9361 Yan et al. (2022) [ 17 ] Age, disease duration, menopause, menstrual irregularity, maximum cyst diameter, cyst wall thickening, papillary structures, solid components, blood flow signals, ascites, HE4, ROMA Logistic regression Univariate analysis, multivariate analysis Nomogram Not reported Training/test 0.904/0.912; overall ROC 0.982 Xu et al. (2024) [ 18 ] Age, tumor size, CA19-9, ROMA Logistic regression LASSO regression Nomogram Not reported Internal 0.858/0.870 Jing et al. (2023) [ 19 ] ROMA_post, HE4, ROMA_pro, CA125, AFP, CA199, CEA, age, MLR and RBC LLRM, RFM, XGBM, SVCM, KNM, MLPM LASSO regression Nomogram Not reported Internal/reported external 0.968/0.944 (RFM) Cohen et al. (2021)[ 20 ] Irregular solid mass, with ascites, ≥ 4 papillary structures within the cyst, irregular multilocular cystic or cystic-solid mass with maximum diameter > 100 mm, with abundant blood flow signals Expert rule model , international ovarian tumor analysis（IOTA） IOTA rules were predefined with no variable selection steps IOTA rules classification Not reported external validation — Kawahara et al. (2024) [ 21 ] Age, maximum tumor diameter, tumor volume, CA125, CA19–9, R2 value MR relaxation measurement method Univariate analysis MR relaxation measurement method Not reported Not reported 0.750 Kobayashi et al. (2008) [ 22 ] Age, tumor size, and menopausal status Logistic regression, Cox regression Univariate analysis, multivariate analysis Logistic and Cox regression models Not reported Not reported — Zhou et al. (2018) [ 23 ] Pregnancy status, tumor size, presence of uterine fibroids, and multiple endometriotic lesions Logistic regression Univariate analysis, multivariate analysis Logistic regression model Not reported Not reported 0.89 Kadan et al. (2015) [ 24 ] Age, cyst size, CA125, solid complexity Logistic regression Univariate analysis, multivariate analysis Logistic regression model Not reported Not reported — Chao et al. (2022) [ 25 ] Age at diagnosis, dysmenorrhea history, preoperative use of GnRHa, family history of endometrial cancer, history of benign cystectomy, Mirena, preoperative serum CA125 level, ultrasound maximum tumor diameter, ultrasound suggestive of leiomyoma or adenomyoma, pathologically confirmed pelvic endometriosis Logistic regression, Gradient Boosting Decision Tree (GBDT) Machine learning feature selection Logistic regression model, GBDT Not reported Internal/temporal 0.942/0.876 (GBDT) Kawahara et al. (2022) [ 26 ] Age, cyst size, R2 value Logistic regression Univariate analysis, multivariate analysis e-NARA index scoring system Inclusion of complete data Not reported 0.928 Characteristics and performance of the included prediction models. All included model studies were judged at high overall risk of bias. The main concerns arose in the analysis domain: small effective sample sizes, data-driven predictor selection, unclear or complete-case handling of missing data, inadequate control of overfitting, and limited validation. Retrospective case-control sampling also raised concerns in the participant domain. Applicability differed across studies. Models developed in selected cancer populations, general ovarian-lesion populations, or specialist centers may not apply directly to women with benign ovarian endometriosis. Table 3 presents the domain-level judgments. Table 3 PROBAST assessment of the included studies. Study Risk of bias Applicability Overall Study Participants Predictors Outcome Analysis Participants Predictors Outcome Risk of bias Applicability Guo (2023) [ 9 ] + ? + − + + + − + Wang et al. (2024) [ 10 ] + ? + − + + + − + Li et al. (2021)[ 11 ] ? ? + − + + + − + Wang et al. (2022) [ 12 ] + ? + − + + + − + Wang (2024) [ 13 ] + ? + − + + + − + Wang (2023) [ 14 ] ? ? + − ? + + − ? Wang et al. (2023) [ 15 ] − + + − − + + − − Chen (2023) [ 16 ] + + + − + + + − + Yan et al. (2022) [ 17 ] + ? + − + + + − + Xu et al. (2024) [ 18 ] + + + − + + + − + Jing et al. (2023) [ 19 ] + + + − + + + − + Cohen et al. (2021)[ 20 ] ? + + − + + + − + Kawahara et al. (2024) [ 21 ] + + ? − + + + − + Kobayashi et al. (2008) [ 22 ] + ? ? − + + + − + Zhou et al. (2018) [ 23 ] + + + − + + + − + Kadan et al. (2015) [ 24 ] + ? ? − + + + − + Chao et al. (2022) [ 25 ] + + + − + + + − + Kawahara et al. (2022) [ 26 ] + + ? − + + + − + Note: + indicates low risk of bias/low applicability concern; − indicates high risk of bias/high applicability concern;? indicates unclear risk of bias/applicability concern. PROBAST assessment of the included studies. Note: + indicates low risk of bias/low applicability concern; − indicates high risk of bias/high applicability concern;? indicates unclear risk of bias/applicability concern. Predictor estimates differed in definition and scale. Two adjusted, threshold-defined maximum-diameter estimates were sufficiently similar to pool (OR 1.85, 95% CI 1.64–2.08). Three valid adjusted estimates for intratumoral blood flow yielded a pooled OR of 3.90 (95% CI 1.73–8.80). Solid components and coexisting uterine fibroids also showed positive pooled associations, whereas dysmenorrhea did not. Tumor size, age, menopausal status, ROMA, and HE4 were not assigned a single pooled effect when units, thresholds, coding, outcomes, or study independence were not sufficiently comparable ( Table 4 ). Table 4 Meta-analysis of predictor associations. Predictor Included references (n) Heterogeneity test Effect model Combined effect size I² (%) P OR (95% CI) Z P Maximum tumor diameter (threshold-defined) 2 0.0 0.67 Fixed-effect 1.85 (1.64, 2.08) 10.20 <0.001 Other tumor-size measures 5 — — Not pooled Not pooled — — Solid component of the tumor 3 74.5 0.02 Random-effects 6.90 (2.14, 22.22) 3.24 0.001 Blood flow within the tumor 3 61.5 0.07 Random-effects 3.90 (1.73, 8.80) 3.28 0.001 Coexisting uterine fibroids 2 10.1 0.29 Fixed-effect 6.55 (2.61, 16.47) 4.00 <0.001 Age 10 — — Not pooled Not pooled — — Dysmenorrhea 4 87.7 < 0.001 Random-effects 1.61 (0.53, 4.87) 0.85 0.40 Menopausal status 5 — — Not pooled Not pooled — — ROMA 4 — — Not pooled Not pooled — — HE4 2 — — Not pooled Not pooled — — Meta-analysis of predictor associations. For maximum tumor diameter, Chen 2023[ 16 ] and Yan 2022[ 17 ] reported threshold-defined adjusted ORs using broadly similar cutoffs (>8.75 cm and ≥10 cm). Their subgroup estimate was OR 1.85 (95% CI 1.64–2.08; I² = 0%). Chao 2022 reported a continuous effect (OR 1.027 per unit increase, 95% CI 1.001–1.053) and was displayed separately rather than combined with threshold-defined effects ( Fig. 2 ). Fig. 2 Forest plot of maximum tumor diameter. The pooled diamond represents only the two threshold-defined adjusted estimates; the continuous Chao 2022 estimate was not pooled. Forest plot of maximum tumor diameter. The pooled diamond represents only the two threshold-defined adjusted estimates; the continuous Chao 2022 estimate was not pooled. Five reports provided estimates related to tumor size, but the effect measures included ORs and an HR, the cutoffs varied, and the studies addressed different clinical outcomes. Estimates were therefore displayed individually without an overall pooled effect ( Fig. 3 ), allowing the direction and uncertainty of each study-specific estimate to be visualized without implying numerical comparability. Fig. 3 Forest plot of reported tumor-size effects, displayed without an overall estimate because effect measures, cutoffs, and outcomes were not sufficiently comparable. Forest plot of reported tumor-size effects, displayed without an overall estimate because effect measures, cutoffs, and outcomes were not sufficiently comparable. Four studies reported adjusted estimates for dysmenorrhea. Random-effects analysis yielded a pooled OR of 1.61 (95% CI 0.53–4.87), with substantial heterogeneity (I² = 87.7%; τ² = 1.001). The association was not statistically significant, and the direction and magnitude of the study-specific estimates varied considerably ( Fig. 4 ). Fig. 4 Random-effects meta-analysis of adjusted ORs for dysmenorrhea. Random-effects meta-analysis of adjusted ORs for dysmenorrhea. The reported Chen 2023[ 16 ] estimate for intratumoral blood flow was internally inconsistent (OR 0.086 with a 95% CI of 1.120–1.292) and was therefore not included in the quantitative synthesis. Across the remaining three compatible adjusted estimates, random-effects analysis on the log-OR scale yielded a pooled OR of 3.90 (95% CI 1.73–8.80), with substantial heterogeneity (I² = 61.5%; τ² = 0.324) ( Fig. 5 ). Fig. 5 Random-effects meta-analysis of intratumoral blood-flow ORs after exclusion of the internally inconsistent Chen 2023[ 16 ] estimate. Random-effects meta-analysis of intratumoral blood-flow ORs after exclusion of the internally inconsistent Chen 2023[ 16 ] estimate. Age, menopausal status, ROMA, and HE4 could not be summarized with a single effect estimate because sampling designs, scales, cutoffs, adjustment sets, and target populations differed. We therefore describe them as recurring candidate predictors rather than transportable effects. Eight studies contributed AUCs from model-development or apparent-performance datasets for preoperative discrimination. The exploratory random-effects pooled AUC was 0.907 (95% CI 0.874–0.932; I² = 47.5%; τ² = 0.108), with a 95% prediction interval of 0.824–0.953. This is a descriptive summary of apparent performance and should not be interpreted as external validation ( Fig. 6 ). Fig. 6 Exploratory random-effects meta-analysis of AUCs from model-development or apparent-performance datasets. Exploratory random-effects meta-analysis of AUCs from model-development or apparent-performance datasets. Four clinically comparable studies contributed split-sample or testing-set AUC estimates that were classified as internal-validation performance. The exploratory random-effects pooled AUC was 0.915 (95% CI 0.880–0.941; I² = 20.7%; τ² = 0.037), with a 95% prediction interval of 0.862–0.949 ( Fig. 7 ). Because these samples came from the same study settings as model development, the estimates may still be optimistic. Fig. 7 Exploratory random-effects meta-analysis of internally validated or testing-set AUCs. Exploratory random-effects meta-analysis of internally validated or testing-set AUCs. Egger regression and trim-and-fill analyses were not retained for the stratified AUC syntheses. These procedures would have limited power and uncertain validity given the small number of clinically comparable studies, differing evaluation stages, and possible dependence among performance estimates. Only two evaluations met our definition of independent external or temporal validation. Other reports described validation procedures of different types, but these were not considered sufficiently independent or methodologically comparable for inclusion in this external/temporal validation stratum. Chen 2023[ 16 ] reported an AUC of 0.936 (95% CI 0.890–0.982) in participants from independent hospitals. Chao 2022 reported an AUC of 0.876 in a later prospective cohort from the same institution. Its displayed interval was approximated from the reported outcome counts because no CI was given. We did not pool the two evaluations because they assessed different models and used different validation designs and case mixes ( Fig. 8 ). Fig. 8 AUCs reported in independent external or temporal validation datasets; estimates were displayed without pooling. AUCs reported in independent external or temporal validation datasets; estimates were displayed without pooling. With only two independent evaluations, between-study heterogeneity could not be estimated reliably and broad transportability could not be assessed. We therefore report these results separately and consider cohort independence, outcome prevalence, number of malignant cases, and similarity to the intended population.

Models for endometriosis-associated ovarian malignancy often show high apparent discrimination, but current evidence is insufficient for routine use. Future research should prioritize prospective, multicenter external validation of the most promising existing models in geographically diverse populations, with model updating undertaken where calibration or transportability is inadequate. These studies should report the full model, handle missing data appropriately, assess calibration, use clinically relevant decision thresholds, and evaluate clinical net benefit.

The reviewed models often reported high AUCs: 0.907 in development or apparent-performance datasets and 0.915 in internal-validation datasets. These values do not demonstrate external validity. Most studies addressed preoperative discrimination, whereas few examined future malignant transformation. The central finding is therefore a marked gap between encouraging apparent discrimination and the much weaker evidence for calibration, transportability, and clinical utility. Most studies used retrospective case-control designs. Although efficient for studying rare outcomes, this design selected cases and an artificially high event proportion may inflate discrimination, and conventional case-control sampling does not directly yield unbiased estimates of absolute risk without appropriate correction or external information. A high development AUC therefore does not ensure accurate risk estimates in routine practice. Methodological shortcomings were common during model development. Univariable screening was common, missing-data methods were rarely described, and shrinkage, penalization, or adequate resampling was infrequent. These choices increase overfitting and produce unstable predictor effects. Future studies should prespecify candidate predictors, avoid unnecessary categorization of continuous variables, use multiple imputation when appropriate, apply bootstrap or cross-validation for internal validation, and then test a locked model in an independent population. Maximum tumor diameter and intratumoral blood flow had pooled ORs of 1.85 and 3.90, respectively, although heterogeneity remained for blood flow. Dysmenorrhea was not associated with malignancy in the pooled analysis (OR 1.61, 95% CI 0.53–4.87). Age, menopausal status, tumor size, HE4, and ROMA appeared in several models, but their effects were not sufficiently comparable for a single pooled estimate; solid components showed a positive pooled association but with substantial heterogeneity. An association with malignant status does not show that adding a predictor improves discrimination, calibration, or net benefit. ROMA combines CA125, HE4, and menopausal status and is already familiar in the assessment of ovarian malignancy [ 27 ]. Its inclusion in several models may improve face validity, but it also creates potential redundancy when its component variables are entered simultaneously. Future model-development studies should assess collinearity, avoid double counting related information, and report whether ROMA provides incremental value beyond standard clinical and imaging features. Age and menopausal status may help identify a higher-risk clinical phenotype, while tumor size and solid components capture suspicious imaging features. Nevertheless, cutoffs chosen from individual datasets may not transport across settings. Continuous modeling with appropriate non-linear terms is preferable to dichotomization unless a clinically established threshold exists. Biomarker measurements also require assay standardization before coefficients or cutoffs can be transferred between institutions. Calibration reporting was limited. Six reports presented a calibration plot, none reported a calibration intercept or slope, two reported a Brier score, and seven included decision-curve analysis. A model may rank patients well while giving systematically inaccurate risk estimates. Future validation studies should report calibration and overall prediction error and assess net benefit across clinically relevant thresholds. Independent validation remains the main evidence gap. Four clinically comparable studies contributed internal split-sample or testing-set AUCs, but only two evaluations were independent external or temporal validations. They assessed different models and were not pooled. We identified no convincing evidence that any reviewed model had undergone prospective implementation and evaluation in routine clinical care. Before clinical adoption, a fully specified model or accessible calculator should be tested in populations that differ in institution, geography, ethnicity, referral pattern, and malignancy prevalence. This review has several limitations. We included only English- and Chinese-language reports. Reporting was incomplete in several studies, and the models differed in clinical purpose, outcome definition, predictor measurement, and validation strategy. Some AUC variances had to be approximated, and calibration or adjusted predictor estimates could not always be reconstructed. Most studies were conducted in Asian populations, particularly in China, which limits geographic and ethnic generalizability.

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.