This investigation adhered to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines (11). The study protocol received approval following registration with the International Prospective Register of Systematic Reviews (PROSPERO) (CRD42024605113).

Inclusion criteria:

Exclusion criteria:

We conducted a systematic retrieval of the Cochrane Library, Web of Science, Embase, and PubMed up to October 11, 2024. We adopted a subject term plus free term search method, and we did not restrict the search by region or period. We employed the keywords ‘Endometriosis’ and ‘Machine learning’ as well as their synonyms. All eligible articles underwent peer review. A comprehensive search strategy for all databases is detailed in Supplementary Table S1.

We imported the retrieved literature into EndNote and read the titles and abstracts after eliminating duplicates. Then, we screened the original studies that met the criteria for our systematic review and downloaded the full texts. Finally, we reviewed the full texts and selected the eligible studies. Before data extraction, we created a standard spreadsheet that included the following: title, first author, publication year, author’s country, research type, patient source, diagnostic criteria for endometriosis, number of endometriosis cases, total number of cases, number of endometriosis cases in the training set, total number of cases in the training set, generation method of validation set, overfitting method, number of endometriosis cases in the validation set, number of cases in the validation set, missing value processing method, variable screening/feature selection method, type of model used, modeling variables, AUC, area under the receiver operating characteristic curve (AUC), diagnostic 2x2 tables, sensitivity, specificity, and precision.

Two researchers carried out the above literature screening and data extraction independently and then cross-checked. If there was any dispute, the third researcher helped to decide.

The risk of bias (ROB) in eligible studies was appraised using the Prediction Model Risk of Bias Assessment Tool (PROBAST) (12). PROBAST is a framework that assesses overall ROB and applicability across four domains: participant selection, predictor variables, outcome assessment, and statistical analysis. Each domain contains specific questions that are evaluated as having a low, high, or unclear ROB. A domain is considered to be at high risk if any of its questions are at high risk. A domain is considered to be of low risk if all of its questions are of low risk. If a domain lacks high risk but includes unclear risks, it is categorized as having an unclear risk.

Two researchers evaluated the ROB independently based on PROBAST, and then cross-checked their findings. Any disagreements were resolved with the help of a third researcher.

A meta-analysis of the AUC, a measure of overall MLM accuracy, was performed. For studies lacking the AUC or its 95% confidence interval (CI) and standard error, their standard error and 95% CI were estimated following Debray TP et al. (13). Heterogeneity among studies was assessed using the I² statistic. A random-effects model was employed for meta-analysis when I² > 50%, while a fixed-effects model was used when I² < 50%.

Sensitivity and specificity were also meta-analyzed using a bivariate mixed-effects model. Diagnostic 2x2 tables are essential for meta-analysis. When these tables were unreported in original studies, calculations were derived from case numbers combined with sensitivity, specificity, and precision. Subgroup analyses were conducted based on dataset, model type, and modeling variable types. Meta-analyses were executed using Stata software. A P-value below 0.05 indicated statistical significance.

The database retrieval yielded 2,380 articles. After removing 391 duplicates based on titles, 1,989 articles were screened by title and abstract. Of those, 1,926 were excluded for being irrelevant, reviews, letters, case reports, experiments, registered protocols, non-English publications, or other reasons. The remaining 63 articles were downloaded for a full-text review. Eighteen articles were excluded for risk factor analysis or missing outcome metrics. Finally, 45 articles passed the full-text evaluation. Figure 1 illustrates the study selection process according to PRISMA guidelines.

The 45 eligible articles were published between 2003 and 2024, with 28 (62%) published in the last five years. Fourteen investigations were conducted in China, while others originated from 17 countries, including France, the United Kingdom, Poland, Germany, Russia, the United States, Spain, Israel, Italy, the Netherlands, Canada, Kazakhstan, Austria, Turkey, Belgium, and Portugal. All studies employed a case-control design. The age of women was within the reproductive range. Eight studies (17.8%) focused on individuals with infertility (14–21). Thirty-one studies incorporated women with relevant symptoms but unclear endometriosis diagnosis as controls. Conversely, two studies utilized completely healthy women as controls (22, 23). Participant numbers ranged from 39 to 612,777 (17, 24). Seventeen studies (38%) underwent internal validation, such as random sampling or k-fold cross-validation. Only six studies (13%) performed external validation. A total of 79 models demonstrating the most effective outcome prediction were identified. Sixty-nine models (87%) reported the AUC, and 74 models (94%) reported diagnostic 2x2 tables or sensitivity and specificity metrics. Twenty-one ML methods were adopted to construct diagnostic models. Logistic regression (LR) was the most frequent (39%), followed by random forest (RF) (14%). Feature selection identified the most relevant features for effective and interpretable models. The current research explored target features for endometriosis diagnosis, including clinical characteristics, genetic regulators, imaging, and omics data. Clinical features, such as questionnaires, medical records, and serum/urine biomarkers, constituted the majority of features in all studies. Genetic regulator information primarily came from salivary microRNAs (miRNAs), with Sofiane Bendifallah and colleagues contributing multiple sequential studies (25–28). Other feature types were less prevalent (Supplementary Tables S2–S4).

PROBAST was utilized to appraise the ROB of predictive diagnostic models. Of the 45 eligible studies, most were case-control or prospective cohort studies from registry databases. Among 118 models, 44 exhibited low ROB concerning study subjects, 42 regarding predictor variables, 107 regarding outcomes, and 10 regarding statistical analysis (Figure 2). Overall, certain models presented a high ROB, particularly in statistical analysis. This suggests that future research should prioritize optimization and validation of statistical analysis methods.

This meta-analysis thoroughly evaluated the efficacy of ML and DL algorithms in detecting endometriosis. Initially, 2,380 studies were considered, and 45 reports were ultimately included after the screening process. Of those, 30 (67%) explored clinical characteristics as variables for diagnostic models, six (13%) focused on genetic regulators, three (7%) on imaging data, and six (13%) on other omics data. The results revealed that these approaches achieved relatively favorable AUCs, sensitivities, and specificities in identifying endometriosis from healthy women or those with similar symptoms. Additionally, our analysis indicated that MLMs based on genetic regulators displayed an even higher range of AUC, sensitivity, and specificity: 0.982 (95% CI: 0.975-0.990), 0.94 (95% CI: 0.90-0.97), and 0.99 (95% CI: 0.94-1.00), respectively, in the training set, supporting a previous study (39). These findings underscore the potential role of genetic regulators in endometriosis development.

Previous systematic reviews highlighted the potential of AI in diagnosing endometriosis. Sivajohan et al.’s scoping review indicated that AI models utilizing diverse data types, including biomarkers, imaging, and clinical variables, achieved pooled sensitivity from 81.7% to 96.7% and specificity from 70.7% to 91.6%, suggesting robust diagnostic performance in controlled settings (3). Furthermore, current reviews emphasized significant methodological heterogeneity, noting variations in algorithms (e.g., LR, SVM, RF), diagnostic targets (e.g., ovarian endometriosis versus deep infiltrating endometriosis), and evaluation metrics, which impede direct comparison across studies (40). These reviews consistently identified a critical gap: the lack of quantitative synthesis to establish standardized diagnostic benchmarks. While prior syntheses offered valuable narrative insights, none conducted meta-analyses to derive pooled accuracy estimates. The present meta-analysis addresses this gap by systematically synthesizing existing evidence and providing pooled diagnostic accuracy estimates for ML-based endometriosis diagnostic models.

Feature selection played a central role in shaping the performance of the models included in the study. Input variables used for endometriosis diagnosis primarily consisted of clinical features, genetic regulators, radiomics, and other omics data. CA-125 levels, visual analog scale (VAS), history of dysmenorrhea, body mass index, and age were the most frequently used clinical inputs and formed the core of the interpretable models in the eligible studies. CA-125, a serum biomarker commonly used in diagnosing endometriosis, correlates with disease severity, particularly in advanced stages (41). Pain is widely regarded as a key diagnostic factor for endometriosis, and the VAS is the most commonly used tool for assessing it, demonstrating good validity, test-retest reliability, and consistency (42). Since clinical features are easily accessible during routine gynecological evaluations, they are practical for use in clinical practice, particularly in early population screening.

In contrast, an increasing number of studies have incorporated genetic regulators, primarily derived from miRNA datasets. miRNAs are believed to play a vital role in endometriosis pathogenesis by regulating inflammation, cell proliferation, angiogenesis, and tissue remodeling. In our study, models based on miRNA expression profiles achieved relatively high AUC values (>0.85), supporting the potential value of transcriptomic features in differential diagnosis. Prior systematic reviews have also reported the moderate to high diagnostic accuracy of circulating miRNAs in endometriosis. However, their clinical application has been limited by differences in sample size and detection platforms (43). Individual studies have reported promising results for models utilizing other omics, indicating the potential value of integrating molecular-level data. However, their heterogeneity and limited external validation reduce their general applicability. These findings should be interpreted with caution, as most studies lack standardized pipelines and reproducibility assessments.

In the field of imaging, research has primarily focused on related diseases, such as endometrial cancer, while paying relatively little attention to endometriosis. For instance, a systematic review by Lecointre et al. indicated that imaging-based methods for identifying endometrial cancer lack sufficient evidence and are in the early stages of development (29). High-quality prospective studies and reliable external validation are essential for advancing the clinical application of these methods. Our analysis suggests that, although multi-omics and imaging methods show great potential, models based on routinely available clinical data are the most practical and scalable for real-world diagnostic settings.

Furthermore, the type of ML model chosen determines the balance between accuracy and interpretability. The studies included in this research predominantly relied on LR, especially for clinical features. This implies that models combining clinical biomarkers prioritize interpretability over other types. However, our study found that ensemble and boosting algorithms, such as LightGBM, XGBoost, and AdaBoost, performed better. Except for the training set within the clinical features, the combined AUC, sensitivity, and specificity were higher than the metrics obtained using LR models alone in each feature type. Zhang et al. (30) reported consistent ML prediction results for gestational diabetes, which further supports the idea that some non-LR algorithms may offer higher diagnostic value, particularly in complex or high-dimensional datasets.

This meta-analysis has several strengths. First, it covered a broad and systematic range of the most recent research findings. Second, the study was conducted and reported strictly according to established guidelines. Third, the quantitative evaluation, which was based on input feature grouping, revealed heterogeneity across studies and trends that could inform future research. However, this study also has certain limitations. First, endometriosis exhibits substantial phenotypic diversity, including superficial peritoneal lesions, ovarian endometriomas, and deep-infiltrating forms with varying disease severity. However, many of the included primary studies did not sufficiently explore the diagnostic performance of ML across these subtypes. This limits the interpretation of the diagnostic value of ML for specific endometriosis classifications. Future research should aim to improve endometriosis detection overall and develop robust multi-class ML models that can aid in subtype diagnosis. Second, the definitions of endometriosis cases and control groups varied across the included studies. Some studies relied on laparoscopic confirmation for diagnosis, while others used imaging or symptom-based criteria. Similarly, control groups ranged from healthy women to patients with symptoms but no confirmed diagnosis, and even included cases of non-endometriosis confirmed by surgery. The primary data did not sufficiently detail these criteria to allow for targeted subgroup analysis, which could introduce bias into the pooled diagnostic performance estimates and distort the overall accuracy assessment. Future studies should standardize the reporting of phenotypes, severity, diagnostic criteria, and control group definitions to facilitate stratified analyses in meta-analyses and enhance the precision and clinical applicability of results. Third, two studies (by Ulan Tore et al. and Krystian Zieliński et al.) exhibited severe class imbalance among the included literature. Imbalanced data can impact model development and pose challenges to model robustness (31). However, the number of such studies was limited, and their specific influence on the overall outcomes was not investigated further via subgroup analysis. Future research should employ strategies such as oversampling techniques to mitigate the effects of class imbalance on modeling results. Additionally, the limited number of ML modeling studies involving certain omics features and imaging data affected the robustness of the results. Lastly, most studies used retrospective designs and lacked independent validation, affecting the interpretation of our findings. We recommend that future research focus on refining the data standardization process to enhance the models’ external validity and clinical applicability.