Abstract
Background: Endometriosis is benign, but it is associated with an increased risk of clear-cell and endometrioid ovarian carcinoma. Several multivariable prediction models have been developed to identify malignant transformation or differentiate endometriosis-associated ovarian cancer (EAOC) from benign ovarian endometrioma. It remains unclear whether these models are sufficiently validated and reported for clinical use. Methods: We searched PubMed, Embase, Web of Science, the Cochrane Library, CINAHL, Scopus, CNKI, Wanfang, VIP, and SinoMed from inception to 24 May 2025. We included reports that developed or evaluated a multivariable model for endometriosis-associated ovarian malignancy. Two reviewers independently selected studies, extracted data with the CHARMS framework, and assessed risk of bias with PROBAST. Studies were grouped by clinical purpose and by evaluation stage. AUCs were pooled on the logit scale with restricted maximum-likelihood random-effects models only within clinically comparable groups. Adjusted ORs were pooled on the log scale when predictor definitions and reference categories were compatible. Results: Eighteen reports describing heterogeneous model-development or model-evaluation studies were included. Logistic regression was the predominant modeling approach, whereas several studies evaluated nomograms, rule-based models, and machine-learning algorithms. In the exploratory stratified analyses, the pooled AUC was 0.907 (95% CI 0.874–0.932) for model-development or apparent-performance datasets and 0.915 (95% CI 0.880–0.941) for internal validation datasets. The two independent external or temporal evaluations were not pooled. Threshold-defined maximum tumor diameter (OR 1.85, 95% CI 1.64–2.08) and intratumoral blood flow (OR 3.90, 95% CI 1.73–8.80) showed positive pooled associations, whereas dysmenorrhea was not statistically significant. Calibration was inconsistently reported, and all included studies had an overall high risk of bias. Six reports presented calibration plots, two reported a Brier score, and seven presented decision-curve analysis; none reported a calibration intercept or slope. Conclusion: Current models often report high discrimination, but the supporting evidence is limited by high risk of bias, inconsistent calibration reporting, few independent validations, and geographic concentration. Routine use is premature. The most promising models should first undergo prospective external validation, calibration assessment, and evaluation of clinical net benefit in diverse populations.