Early risk prediction in endometrial cancer using biomarker and machine learning models: a systematic review and meta-analysis protocol Authors: Priya Giri, Rakesh Kumar Saroj Published: 2026 ABSTRACT AbstractBackgroundGlobally, endometrial cancer (EC) is among the most prevalent gynaecological cancers, with rising incidence driven by demographic and metabolic changes. Accurate early prognostic assessment is essential to guide personalized treatment and improve long-term outcomes. Traditional clinicopathological prognostic methods-including FIGO (International Federation of Gynaecology and Obstetrics) staging, tumor grade, and histological subtype remain the current standard but have limitations in detecting early recurrence risk. In recent years, biomarker-based prognostic signatures and machine-learning (ML) guided risk models, including radiomics and multi-omics approaches, have emerged as promising tools for improving prognostic accuracy. However, no systematic review and meta-analysis have comprehensively evaluated whether these biomarker or ML-based prognostic models offer improved early risk prediction compared with conventional methods in EC.MethodsWe will undertake a comprehensive search of MEDLINE, Scopus, IEEE Xplore, and Web of Science from 2010 to 2025. Two reviewers will independently perform title/abstract screening, full-text assessment, data extraction, and quality appraisal, with arbitration by a third reviewer where necessary. Studies developing, validating, or evaluating biomarker-based, radiomics-based and ML-guided prognostic models in human EC patients will be eligible for inclusion. Purely laboratory-based biomarker discovery studies that do not develop or evaluate a prognostic model in human participants will be excluded. Outcomes of interest include measures of prognostic performance such as AUC (area under the curve), C-index (concordance index), time-dependent AUC, calibration metrics, hazard ratios for survival outcomes, and decision analytic measures. Risk of bias will be assessed using PROBAST (Prediction model Risk of Bias Assessment Tool) and QUIPS (Quality in Prognosis Studies), and certainty of evidence will be summarized using GRADE (Grading of Recommendations Assessment, Development and Evaluation). When feasible, meta-analysis of performance metrics will be performed using random-effects models with restricted maximum likelihood (REML) estimation, with subgroup analysis restricted to histological subtype when sufficient data permits.DiscussionThis systematic review will synthesize current evidence on the prognostic performance of biomarker and ML-guided prognostic models in EC, comparing them with conventional clinicopathological risk assessment methods. Findings will help clarify whether emerging model types offer meaningful improvements in the early prognostic accuracy, provide insight into methodological limitations of existing models, and identify gaps in evidence requiring future research. This review will support clinicians, researchers, and policymakers seeking to integrate precision-based prognostic tools into EC care.Systematic review registrationPROSPERO CRD420251230895Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s13643-026-03248-0.Keywords: Endometrial cancer, Biomarkers, Prognostic models, Machine learning, Systematic review, Meta-analysis Background Endometrial cancer (EC) is one of the most prevalent gynaecological cancers globally, representing a significant and growing public health concern. Increasing life expectancy, changes in reproductive patterns, and the rising prevalence of metabolic risk factors - including obesity and type 2 diabetes-contribute to its increasing incidence worldwide [1, 2]. A substantial proportion of patients experience recurrence or progression despite initial treatment, even though many EC cases are diagnosed at an early stage with generally favourable outcomes [1, 3]. Therefore, early and precise prognostic evaluation is essential for directing personalized treatment, identifying recurrence risk, severity of follow-up and recommending decisions for adjuvant therapy [3]. Conventional clinicopathological characteristics, such as FIGO stage, tumor grade, myometrial invasion, lymphovascular space invasion (LVSI), and histological subtype, serve as the core component of the current prognostic methods [4]. Despite being the cornerstone of risk assessment for several decades, the ability of these measures to predict the risk of an early recurrence still remains limited [5]. Recent advances in molecular profiling and computational analytics have led to significant interest in biomarker-based and machine-learning (ML)-guided prognostic models for EC [6–8]. Biomarkers including genomic, transcriptomic, proteomic, epigenetic, and immunohistochemical signatures offer opportunities to capture tumor biology beyond what conventional features provide [6]. In parallel, ML models, including radiomics, deep learning and integrated multi-omics frameworks, allow high-dimensional data to be combined with clinical features to generate patient-specific risk predictions. Several emerging prognostic models claim superior performance to standard methods, reporting improved calibration and clinical utility [6, 7]. However, the extent to which these approaches outperform traditional clinicopathological prognostic systems has not yet been systematically synthesized [8]. The prognostic landscape of EC has also evolved with the introduction of molecular classification systems including The Cancer Genome Atlas (TCGA) Research Network which identified four prognostically distinct molecular subgroups in EC: POLE-ultramutated, microsatellite-instability-high (MSI-H), copy-number-low (CNL), and copy-number-high (CNH), each with distinct molecular features and clinical outcomes [9, 10]. Despite growing evidence supporting the prognostic value of molecular classification, the routine integration of biomarker-based or ML-guided prognostic models into clinical practice remains inconsistent [11]. It remains unclear whether these models offer clinically meaningful improvements in early outcome prediction across the general EC population or within biologically distinct subgroups, including endometrioid and non-endometrioid carcinoma [12, 13]. Previous studies in this field have been heterogenous with respect to biomarker type, modelling methodology, study design, sample size, and validation rigor [14, 15]. Many proposed models suffer from issues such as overfitting, lack of external validation, incomplete reporting or limited applicability to diverse clinical settings [15]. In addition, variation in reported performance metrics including AUC, C-index, hazard ratios, calibration measures, net reclassification improvement, and decision curve analysis makes cross-study comparison challenging. A systematic evaluation is therefore needed to synthesize available evidence, assess methodological quality and determine whether biomarker-based and ML-guided prognostic models deliver notable improvements over established clinicopathological approaches [14]. The management of EC has been guided by several evidence-based clinical risk classification systems. The ESGO/ESTRO/ESP 2021 guidelines [3] introduced a comprehensive risk stratification framework incorporating FIGO stage, histological subtype, tumor grade, myometrial invasion, and LVSI. The ESMO 2022 Clinical Practice Guidelines [13] further refined adjuvant treatment recommendations based on molecular and pathological risk groups. Most recently, the updated ESGO/ESTRO/ESP 2025 guidelines [12] incorporate molecular classification (TCGA subgroups) into the risk stratification framework, reflecting significant advances in precision oncology. Additionally, surgical staging in EC has transitioned from FIGO 2009 to FIGO 2023, which introduced sub-staging and incorporated molecular classification. Studies included in this systematic review may use either staging system; we will record the staging system used for each study and address this in the data synthesis, using sensitivity analysis where necessary to assess the impact of staging system differences on pooled performance estimates. The rapid development of precision oncology as well as the rising availability of high-throughput data, highlights the significance of a comprehensive synthesis of prognostic studies in EC [7]. Identifying which models demonstrate reliable improvements in early prognostic accuracy has the potential to facilitate clinical translation and enable truly risk-tailored management of EC. This encompasses multiple dimensions of clinical decision-making: de-escalation of adjuvant therapy in patients confirmed to be at genuinely low recurrence risk (e.g., omission of vaginal brachytherapy or chemotherapy in low-risk endometrial carcinoma); intensification of treatment and surveillance for those identified as high-risk; informed decisions regarding the extent of surgical staging, including sentinel lymph node mapping vs. systematic lymphadenectomy; and personalization of follow-up protocols. Precision risk stratification is therefore directly relevant to patient safety, quality of life, and healthcare resource allocation. This protocol outlines the rationale and methodology for a systematic review and meta-analysis that will evaluate the prognostic accuracy and comparative performance of biomarker-based and ML-guided prognostic models in EC relative to conventional clinicopathological methods [16]. Methods Objectives This systematic review and meta-analysis seek to evaluate whether biomarker-based or machine-learning guided prognostic models improve early risk prediction and prognostic accuracy among patients with endometrial cancer compared with conventional clinicopathological methods. The primary objective is to synthesize evidence on discrimination (AUC, C-index, time-dependent AUC) and time-to-event prognostic performance (hazard ratios for recurrence or survival) for these models. Secondary objectives include evaluating calibration, reclassification measures (NRI, IDI), clinical decision utility (decision curve analysis), and identifying methodological factors associated with improved model performance. One subgroup analysis is planned, comparing prognostic model performance between endometrioid and non-endometrioid endometrial cancer subtypes. The secondary analysis will be undertaken only if sufficient and compatible data are reported across studies. Protocol and registration This protocol has been developed in accordance with PRISMA-P 2015 recommendations for systematic review protocols. A completed PRISMA-P checklist is provided as Supplementary file 1, mapping each item to the relevant section of this protocol. The protocol was prospectively registered in PROSPERO prior to data extraction [17]. The PROSPERO registration number is CRD420251230895. The full protocol record is available from https://www.crd.york.ac.uk/PROSPERO/view/CRD420251230895. Eligibility criteria Population We will include human studies involving patients diagnosed with endometrial cancer of any histological subtype. Subgroup analysis will compare outcomes between endometrioid and non-endometrioid carcinomas. There will be no restrictions based on country, ethnicity, or clinical setting (primary, secondary or tertiary care). Studies focusing exclusively on animal models, cell lines, or in-vivo assays without human prognostic evaluation will be excluded. Intervention Eligible studies must describe the development, validation, or evaluation of prognostic models incorporating one or more of the following intervention types. Biomarker-based models—encompassing genomic, transcriptomic and proteomic signatures, non-coding RNA markers including miRNA and lncRNA, DNA methylation and other epigenetic markers, immunohistochemical (IHC) markers, and circulating biomarkers such as cell-free DNA, circulating tumor cells, and serum or plasma protein markers—will be included where they are applied to predict patient outcomes. Radiomics or pathomics-based models, which extract and analyze quantitative features from medical imaging or digitized histopathological slides respectively, will be eligible where they incorporate a structured prognostic modelling framework. Machine-learning or deep-learning prognostic algorithms—including but not limited to classifiers, regression-based models, risk scores and nomograms developed using supervised, unsupervised, or ensemble learning approaches—will be included where they are applied to prognostic prediction in endometrial cancer patients. Finally, multi-omics integrated frameworks that combine two or more data modalities—such as genomic, transcriptomic, proteomic, radiomic, or clinical data—within a unified prognostic model will be eligible. Models must be designated to predict recurrence, progression, survival, or risk stratification in endometrial cancer patients. Both internally validated and externally validated models will be included. Studies reporting model development alone, without internal or external validation, will also be eligible provided all other inclusion criteria are met. Comparator Studies must include at least one conventional clinicopathological comparator for the prognostic model under evaluation. Eligible comparators encompass FIGO surgical staging, with both FIGO 2009 and FIGO 2023 accepted given the study period of 2010–2025; the specific staging system used will be recorded for each study and its potential contribution to between-study heterogeneity will be assessed through sensitivity analysis. Histological subtype, classified according to edition of the WHO Classification of Tumors of the Female Genital Tract published during 2010–2025, will be accepted as a comparator, with the specific edition recorded. The shift in WHO classification—including the reclassification of carcinosarcoma as a metaplastic endometrial carcinoma in the 2020 edition—will be noted as a potential source of heterogeneity in histological composition across studies and addressed in the narrative synthesis. Tumor grade and depth of myometrial invasion will be accepted as standard pathological comparators and extracted as reported. LVSI will be extracted as reported by each study, whether classified using a binary scheme (absent/present) or three-tier scheme (absent/focal/substantial); any stated cut-off between focal and substantial LVSI will be recorded verbatim. For studies using a three-tier classification without an explicitly stated cut-off, LVSI will be recorded as reported without retrospective imposition of an external threshold; heterogeneity in LVSI reporting will be addressed in the narrative synthesis. Where a sufficient number of eligible studies report three-tier LVSI classification (absent/focal/substantial), a pre-specified sensitivity analysis will be conducted comparing pooled performance estimates derived from studies using binary LVSI classification against those using three-tier classification, to assess whether LVSI reporting heterogeneity represents a source of variability in pooled estimates. Standard clinical risk stratification systems, including the ESGO/ESTRO/ESP 2021 guidelines, the ESMO 2022 Clinical Practice Guidelines, and the ESGO/ESTRO/ESP 2025 guidelines, as described in the Background—will also be eligible comparators; the specific framework and version used in each included study will be recorded in the data extraction form. Studies comparing biomarker-based or ML-based prognostic models against these conventional methods will be included. Outcomes The primary outcomes of interest are measures of early risk prediction and prognostic accuracy. Discrimination metrics—including AUC, C-index and time-dependent AUC—will constitute the principal performance measures of interest, alongside effect estimates in the form of hazard ratios (HRs) with 95% confidence intervals for recurrence, progression, disease-free survival (DFS), progression-free survival (PFS), and overall survival (OS). Sensitivity, specificity, and classification performance will additionally be extracted where reported. Secondary outcomes will encompass calibration measures—including calibration slope, intercept, and calibration curves - reclassification metrics such as the net reclassification improvement (NRI) and integrated discrimination improvement (IDI) and clinical utility measures including decision curve analysis and net benefit estimates. Studies will be eligible if they report sufficient data to extract or calculate at least one relevant performance metric. Study design/setting Eligible study designs will include prospective and retrospective cohort studies, case-control prognostic studies, and studies reporting model development, internal validation or external validation of prognostic models. Prognostic modelling studies of any design will be considered, provided they meet all other eligibility criteria. Cross-sectional studies of any design without follow-up data will be excluded, as will conference abstracts lacking full methodological detail and purely in-vitro or laboratory biomarker discovery studies that do not develop or evaluate prognostic models in human participants. Limits The search will be restricted to English-language publications spanning the period 2010–2025, reflecting the emergence and growing clinical relevance of biomarker-based and ML-based prognostic modelling approaches in endometrial cancer during this period. The review will be limited to studies involving human participants; non-human studies will be excluded during the screening phase. No database-level human species filter was applied at the search stage, as manual screening was considered more reliable for ensuring consistent application of this criterion across all included databases. Exclusion criteria This review aims to include studies that reflect real-world prognostic model performance in endometrial cancer and provide extractable quantitative performance metrics. Accordingly, studies will be excluded if they lack human patient data, including purely in-vitro or animal studies, as well as biomarker discovery studies that do not develop or evaluate a prognostic model in human participants. Studies will also be excluded if they report no outcome related to survival, recurrence, or risk prediction, or if they provide insufficient statistical information to derive at least one relevant performance metric. Case reports, case series with fewer than ten patients, editorials, letters, and expert opinion pieces will not be eligible. Conference abstracts lacking complete methodological details will similarly be excluded. Non-comparative interventional trials focusing solely on treatment effects rather than prognostic performance will not be considered, nor will studies that evaluate individual prognostic factors without constructing or validating a prognostic model. Information sources and literature search A comprehensive search will be conducted across four electronic databases: MEDLINE, Scopus, Web of Science Core Collection, and IEEE Xplore. All databases will be searched for publications from January 2010 to November 2025. The search strategy will combine EC related MeSH terms and free-text keywords with terms pertaining to prognostic modelling, biomarkers and ML-based approaches, and will be adapted appropriately for each database interface. Reference lists of all included studies will be hand-searched to identify any additional eligible studies not captured through the database search. Where clarifications are required or relevant data are missing from published reports, study authors will be contacted directly. Grey literature will generally be excluded, given the methodological depth required for rigorous prognostic model evaluation and the limited feasibility of systematically appraising non-peer-reviewed modelling studies. However, a targeted exception will be made for conference papers and abstracts from IEEE Xplore that provide complete methodological descriptions of the prognostic model, as these represent a specialized and relevant source for machine-learning-based approaches that may not always be indexed in traditional medical databases. Studies lacking complete methodological detail, irrespective of source, will be excluded in accordance with the pre-specified eligibility criteria. The full MEDLINE search strategy is provided in the Appendix 1. Study selection All records retrieved from the search databases will be imported into reference management software (e.g., Zotero) for deduplication. Following deduplication, two reviewers will independently screen all titles and abstracts (Level 1) against predefined eligibility criteria. Full texts of potentially eligible articles will then be assessed independently by the same reviewers (Level 2). Disagreements will initially be addressed through discussion; and any unresolved issues will be adjudicated by a third reviewer. Reasons for exclusion at the full-text stage will be documented. All extracted data from included studies will be recorded in a pre-piloted standardized extraction form developed in MS Excel. A PRISMA flow diagram will be presented in the final review to illustrate the screening and selection process. Full eligibility criteria are detailed in Eligibility criteria section. Data items and data collection process Data will be extracted from all included studies using a pilot-tested standardized extraction form developed and refined prior to full-scale data collection. The following categories of information will be systematically collected. Study characteristics For each included study, the following bibliographic and contextual details will be recorded: author names, publication year, and country of origin; study design and clinical setting; total sample size and duration of follow-up; histological subtypes represented in the study population; funding source and declared conflicts of interest, which will be noted as contextual information relevant to the assessment of potential bias in reported findings. Population characteristics Participant-level characteristics will be extracted to describe the study populations and facilitate assessment of generalizability. These will include the age distribution of participants, disease stage, and tumor grade at the time of study entry; histological subtype breakdown, with particular attention to the proportion of endometrioid versus non-endometrioid cases; ethnicity or racial composition of the cohort where reported; and details of any treatment received, including surgical approach and adjuvant therapy, where these are described. Prognostic model characteristics The characteristics of each prognostic model will be extracted in detail to enable systematic categorization and comparison. This will include the model category—classified as biomarker-based, radiomics or pathomics-based, machine learning or deep learning, or multi-omics integrated—along with the specific predictors or features incorporated into the model. Details of internal and external validation procedures will be recorded, including the validation method employed, the size of the validation cohort, and the source of validation data. Where reported, the software environment or computational platform used for model development will also be noted. Comparator information Information on the conventional comparator methods used in each study will be extracted to enable direct performance comparison. This will include the specific clinicopathological prognostic methods employed as comparators—such as FIGO stage, tumor grade, histological subtype, myometrial invasion, LVSI, or established risk stratification systems—as well as the baseline clinicopathological variables included in any reference or comparator model. Outcomes Quantitative performance metrics will be extracted for both the biomarker or ML-based prognostic model and any conventional comparator. Discrimination metrics—including AUC, C-index, and time-dependent AUC – will be the primary extracted performance measures. Calibration metrics will be recorded where reported, including calibration slope, intercept, and graphical calibration curves. Hazard ratio (HRs) with 95% confidence intervals (CIs) will be extracted for time-to-event outcomes including overall survival (OS), disease-free survival (DFS), recurrence-free survival (RFS), and progression-free survival (PFS). Sensitivity and Specificity will be recorded where available. Reclassification and clinical utility metrics—including net reclassification improvement (NRI), integrated discrimination improvement (IDI), and decision curve analysis—will be extracted where reported. Additional items Several supplementary data items will also be recorded to support methodological appraisal and contextual interpretation. The approach used to handle missing data—including imputation methods, complete case analysis, or exclusion—will be documented for each study. Where reported, information on model interpretability such as feature importance rankings, or variable contribution measures will be extracted. The ethical approval status of each study will also be recorded. Data extraction will be performed independently by two reviewers, with any discrepancies resolved through discussion and consensus. A third reviewer will be consulted to adjudicate where agreement cannot be reached. Inter-rater agreement will be quantified using Kappa statistics. No individual patient-level data will be collected at any stage of this review. Methodological appraisal and risk of bias Risk of bias will be assessed using the Prediction model Risk of Bias Assessment Tool (PROBAST), which evaluates four domains: participants, predictors, outcome, and analysis. PROBAST is specifically designed for prediction model studies and will be applied to all eligible prognostic model studies [18]. For observational prognostic factor studies, the Quality in Prognostic Studies (QUIPS) tool will additionally be applied, evaluating study participation, attrition, prognostic factor measurement, confounding, outcome measurement, and statistical analysis [19]. The overall certainty of evidence will be summarized using the Grading of Recommendations Assessment, Development and Evaluation (GRADE) framework [20]. Data synthesis Given the anticipated methodological heterogeneity in this field—spanning model category, biomarker type, algorithmic approach, outcome definition, and validation strategy—all analysis will be organized according to a pre-specified model taxonomy. Studies will be grouped into four primary strata: (A) biomarker-based models (genomic, transcriptomic, proteomic, epigenetic, IHC, or circulating biomarker-based); (B) radiomics and pathomics models; (C) machine learning and deep learning prognostic algorithms, including classifiers, regressors, risk scores, and nomograms; and (D) multi-omics integrated prognostic frameworks. Within each stratum, performance metrics will be further stratified by outcome type: discrimination metrics (AUC, C-index, time-dependent AUC) and time-to-event measures (hazard ratios). Meta-analysis will be considered feasible within a given stratum-outcome cell when at least four studies report a compatible, extractable metric using comparable populations and outcome definitions, consistent with recommended minimum study thresholds for meaningful quantitative synthesis in systematic reviews [21, 22]. Meta-analysis will be deemed infeasible when: (i) fewer than 4 studies contribute to a given cell; (ii) I2 exceeds 75% and no clear sources of heterogeneity can be identified or addressed through pre-specified subgroup analysis; or (iii) outcome definitions, metric scales, or reporting approaches are too heterogenous for valid statistical combination. For studies that cannot be included in quantitative synthesis—due to heterogenous outcome reporting, incompatible metrics, or insufficient comparable data—findings will be summarized in structured narrative paragraphs organized by model type and outcome category, following SWiM (synthesis without meta-analysis) reporting guidance as specified in Narrative synthesis section. These criteria are pre-specified and will be applied consistently across all outcomes and model strata. All analysis will be conducted using R (packages: meta, metafor, metamisc) or Python statistical libraries. Summary tables will present study characteristics, prognostic model details, validation approaches and key outcome measures. Where sufficient homogeneity in outcomes and reporting exists we will conduct quantitative meta-analysis following recommended methodology for prognostic model evaluation. Meta-analysis of model performance Meta-analysis of prognostic model performance will be conducted where a sufficient number of studies report compatible, extractable metrics across comparable populations and outcome definitions; the specific metrics considered for pooling are described in the subsections that follow. Discrimination metricsDiscrimination metrics—specifically AUC, C-index and time-dependent AUC—will constitute the primary performance measures considered for quantitative pooling. To account for differences in reporting scales and distributional properties across studies, AUCs may be transformed to the logit scale prior to pooling where appropriate. Given the anticipated variability across modelling strategies, histological cohort compositions, and validation approaches, all pooled analysis will be conducted using random-effects models with restricted maximum likelihood (REML) estimation. Discrimination metricsDiscrimination metrics—specifically AUC, C-index and time-dependent AUC—will constitute the primary performance measures considered for quantitative pooling. To account for differences in reporting scales and distributional properties across studies, AUCs may be transformed to the logit scale prior to pooling where appropriate. Given the anticipated variability across modelling strategies, histological cohort compositions, and validation approaches, all pooled analysis will be conducted using random-effects models with restricted maximum likelihood (REML) estimation. Discrimination metrics—specifically AUC, C-index and time-dependent AUC—will constitute the primary performance measures considered for quantitative pooling. To account for differences in reporting scales and distributional properties across studies, AUCs may be transformed to the logit scale prior to pooling where appropriate. Given the anticipated variability across modelling strategies, histological cohort compositions, and validation approaches, all pooled analysis will be conducted using random-effects models with restricted maximum likelihood (REML) estimation. Discrimination metrics—specifically AUC, C-index and time-dependent AUC—will constitute the primary performance measures considered for quantitative pooling. To account for differences in reporting scales and distributional properties across studies, AUCs may be transformed to the logit scale prior to pooling where appropriate. Given the anticipated variability across modelling strategies, histological cohort compositions, and validation approaches, all pooled analysis will be conducted using random-effects models with restricted maximum likelihood (REML) estimation. Time-to-event measuresHazard ratios (HRs) with 95% confidence intervals (CIs) will be pooled using a random-effects model (REML) for time-to-event outcomes including overall survival (OS), disease-free survival (DFS), recurrence-free survival (RFS), and progression-free survival (PFS). Where necessary, log-HRs and standard errors will be derived from reported statistics, KM curves or Cox model outputs. Time-to-event measuresHazard ratios (HRs) with 95% confidence intervals (CIs) will be pooled using a random-effects model (REML) for time-to-event outcomes including overall survival (OS), disease-free survival (DFS), recurrence-free survival (RFS), and progression-free survival (PFS). Where necessary, log-HRs and standard errors will be derived from reported statistics, KM curves or Cox model outputs. Hazard ratios (HRs) with 95% confidence intervals (CIs) will be pooled using a random-effects model (REML) for time-to-event outcomes including overall survival (OS), disease-free survival (DFS), recurrence-free survival (RFS), and progression-free survival (PFS). Where necessary, log-HRs and standard errors will be derived from reported statistics, KM curves or Cox model outputs. Hazard ratios (HRs) with 95% confidence intervals (CIs) will be pooled using a random-effects model (REML) for time-to-event outcomes including overall survival (OS), disease-free survival (DFS), recurrence-free survival (RFS), and progression-free survival (PFS). Where necessary, log-HRs and standard errors will be derived from reported statistics, KM curves or Cox model outputs. Clinical utility measuresDecision curve analysis, NRI and IDI will be extracted where reported in the eligible studies. Given the inconsistent availability of these metrics and of calibration measures in the prognostic modelling literature, they will not be prioritized for quantitative synthesis. Where such data are reported across a sufficient number of studies in comparable form, they will be summarized narratively. Where data are sparse or reported incompatibly across studies, these outcomes will be listed descriptively where extractable and their underreporting will be explicitly acknowledged as a limitation of the available evidence base. Clinical utility measuresDecision curve analysis, NRI and IDI will be extracted where reported in the eligible studies. Given the inconsistent availability of these metrics and of calibration measures in the prognostic modelling literature, they will not be prioritized for quantitative synthesis. Where such data are reported across a sufficient number of studies in comparable form, they will be summarized narratively. Where data are sparse or reported incompatibly across studies, these outcomes will be listed descriptively where extractable and their underreporting will be explicitly acknowledged as a limitation of the available evidence base. Decision curve analysis, NRI and IDI will be extracted where reported in the eligible studies. Given the inconsistent availability of these metrics and of calibration measures in the prognostic modelling literature, they will not be prioritized for quantitative synthesis. Where such data are reported across a sufficient number of studies in comparable form, they will be summarized narratively. Where data are sparse or reported incompatibly across studies, these outcomes will be listed descriptively where extractable and their underreporting will be explicitly acknowledged as a limitation of the available evidence base. Decision curve analysis, NRI and IDI will be extracted where reported in the eligible studies. Given the inconsistent availability of these metrics and of calibration measures in the prognostic modelling literature, they will not be prioritized for quantitative synthesis. Where such data are reported across a sufficient number of studies in comparable form, they will be summarized narratively. Where data are sparse or reported incompatibly across studies, these outcomes will be listed descriptively where extractable and their underreporting will be explicitly acknowledged as a limitation of the available evidence base. Assessment of heterogeneity Heterogeneity will be quantified using the I2 statistic [23], which estimates the proportion of total variability attributable to between-study heterogeneity rather than sampling error, and τ2, the between-study variance estimated using the restricted maximum likelihood (REML) method [24]. Subgroup analysis The primary planned subgroup analysis will compare prognostic model performance between endometrioid and non-endometrioid endometrial carcinoma subtypes. This distinction was selected on the basis of the fundamental, biological and prognostic differences between these two broad categories: endometrioid carcinoma follows estrogen-driven carcinogenetic pathway and is generally associated with more favourable outcomes, while non-endometrioid subtypes—including serous, clear cell, and carcinosarcoma—arise via distinct molecular mechanisms and carry substantially poor prognosis, factors that are expected to influence prognostic model performance. The subgroup classification was intentionally kept broad, informed by a preliminary assessment of the available literature which indicated that the majority of published prognostic modelling studies in endometrial cancer report findings predominantly in endometrioid-type cohorts, with limited and inconsistent reporting of subgroup-specific data by tumor grade or individual non-endometrioid histological subtype. A finer stratification—such as by tumor grade within endometrioid carcinomas (Grade 1/2 vs. Grade 3) or by individual non-endometrioid subtype (serous, clear, carcinosarcoma)—was therefore not formally pre-specified, as it was anticipated that insufficient comparable data would be available to support a meaningful stratified analysis. This secondary analysis will be undertaken only if sufficient and compatible data are reported across eligible studies. Regarding alternative subgroup frameworks, the clinical relevance of disease stage (early-stage FIGO I/II vs. advanced-stage FIGO III/IV) and TCGA molecular subgroups (POLE-ultramutated, microsatellite instability-high, copy-number-low, copy-number-high) is fully acknowledged. Stage-based and molecular subgroup analysis were considered during protocol development but were not formally pre-specified, as the preliminary literature assessment indicated that performance metrics disaggregated by disease stage or molecular subgroup are infrequently reported in published prognostic modelling studies in endometrial cancer, making stratified pooling unlikely to be feasible across a sufficient number of eligible studies. Should eligible studies reporting adequate subgroup-specific data encompassing stage-specific, molecular subgroup classifications or individual non-endometrioid histological subtypes—be identified during the review, exploratory subgroup analysis will be conducted and reported transparently. Molecular subgroup analysis is acknowledged as a particularly important future direction as standardized molecular classification becomes more uniformly adopted and reported in primary prognostic modelling studies in endometrial cancer. Sensitivity analysis Sensitivity analysis will be conducted to test the robustness of pooled estimates: Excluding studies rated as high overall risk of bias according to PROBAST, provided sufficient studies remain for meaningful comparison Excluding studies rated as high overall risk of bias according to PROBAST, provided sufficient studies remain for meaningful comparison Using alternative meta-analytic models (fixed-effects vs random-effects) to assess the stability of estimates, although the random-effects model will be retained as the primary approach. Using alternative meta-analytic models (fixed-effects vs random-effects) to assess the stability of estimates, although the random-effects model will be retained as the primary approach. Assessment of publication bias Publication bias will be assessed using funnel plots for performance metrics with sufficient contributing (≥10) studies to allow visual interpretation. Trim-and-fill analysis will be performed as an exploratory correction for potential missing studies. Narrative assessment will accompany any observed asymmetry. Risk of bias-adjusted synthesis Where feasible, risk-of-bias assessments (PROBAST) will be incorporated into the quantitative synthesis (meta-analysis) through sensitivity analysis, primarily by excluding studies judged at high risk of bias to evaluate the robustness of pooled estimates. Narrative synthesis Where quantitative synthesis is not feasible due to heterogeneity in model types, outcome definitions, or reporting approaches, findings will be presented through narrative synthesis. This narrative synthesis will systematically describe the model development strategies employed across included studies, the predictors and feature sets incorporated into each model type, the validation approaches used and their methodological rigor, and the key findings and limitations identified within and across studies. The clinical applicability of reported models—including their potential for translation into routine prognostic practice in endometrial cancer care—will also be addressed. The narrative synthesis will be conducted and reported in accordance with SWiM (synthesis without meta-analysis) guidance, which provides a structured framework for transparently describing and integrating evidence when statistical pooling is not appropriate. Discussion To our knowledge, no prior systematic review and meta-analysis has comprehensively evaluated whether biomarker-based or ML-guided prognostic models improve early risk prediction in endometrial cancer compared with conventional clinicopathological methods. Given the rapid expansion of multi-omics technologies, radiomics and advanced computational approaches, synthesizing existing prognostic evidence is essential for informing clinical decision-making and providing directions for future research. This SRMA seeks to clarify whether emerging prognostic models presents significant enhancement over established methods and to provide a structured overview of the strengths, limitations and clinical implications of current models. Improved prognostic accuracy has the potential to influence multiple aspects of EC management, including the tailoring of surgical staging, selection of adjuvant therapies and appropriate follow-up strategies. Precision risk stratification is especially crucial for EC, where most patients present with early-stage disease but a proportion; particularly those with high-risk histologies experience early recurrence that may not be adequately predicted by conventional features alone. Biomarkers and ML-guided models may offer earlier detection of recurrence risk by capturing molecular, imaging-based, or high-dimensional patterns that traditional staging systems cannot. By comparing the prognostic performance of these models with standard approaches, this SRMA may help identify which strategies hold the greatest potential for clinical integration. This SRMA also aims to evaluate the methodological quality of existing prognostic studies. Many proposed biomarker and ML models suffer from limitations such as inadequate sample size, incomplete reporting, and lack of external validation. The use of PROBAST, QUIPS, and GRADE will allow us to assess the internal validity, risk of bias, and certainty of evidence in a structured manner. Furthermore, evaluating calibration, clinical utility and decision-analytic measures will provide a more complete understanding of whether improved discrimination translates into meaningful clinical benefit. Potential limitations of this SRMA include anticipated heterogeneity in biomarkers, modelling approaches, study designs, validation techniques and reported outcomes. Variability in histological composition, sample size, molecular assays, radiomic feature extraction and ML algorithms may make synthesis challenging. Differences in reporting standards such as incomplete calibration metrics or missing confidence intervals for performance measures may limit the feasibility of meta-analysis for some outcomes. In addition, publication bias in prognostic modelling research, particularly in high-dimensional ML studies, may overestimate model performance. These issues will be addressed through sensitivity analysis, subgroup analysis and structured narrative synthesis where quantitative pooling is not feasible. Several inherent limitations of this protocol warrant acknowledgement. The restriction to English-language publications, while pragmatically necessary to ensure rigorous quality appraisal using PROBAST, may introduce language bias by excluding relevant prognostic modelling studies published in non-English languages. However, it is recognized that a substantial proportion of research groups from East Asian countries—including China, Japan, and Korea, where endometrial cancer prognostic modelling research is increasingly prolific—publish their findings directly in English-language peer-reviewed journals and such studies would be fully captured by the search strategy. The practical impact of this restriction on the comprehensiveness of the evidence base is therefore anticipated to be limited, though it cannot be entirely excluded and is acknowledged as a potential source of bias. Additionally, the exclusion of general grey literature—including preprints, theses and non-peer-reviewed reports—was a deliberate methodological decision informed by the requirement for complete and transparent reporting of model development and validation methodology for PROBAST-based quality assessment, and by the recognition that broader grey literature searches, such as those conducted via Google Scholar, return an unmanageable volume of heterogenous and unevaluable results that would compromise the feasibility and consistency of the synthesis. A targeted exception was made for IEEE Xplore conference proceedings providing complete methodological descriptions, to capture machine learning-based prognostic studies not indexed in traditional medical databases. Studies lacking complete methodological detail were excluded as per the pre-specified eligibility criteria. These restrictions may nonetheless result in some underrepresentation of unpublished work and should be considered when interpreting the findings of the completed review. Despite these challenges, this SRMA is expected to make a significant contribution to the field. It will bring clarity landscape of EC prognostic modelling, identify high-quality models that demonstrate consistent improvement over conventional risk system and highlight frequent methodological shortcomings that should be addressed in future research. The findings may assist clinicians and researchers understand the utility and limitations of existing biomarker and ML-based prognostic tools and encourage standardization in future model development and reporting. The results may also aid in translational research priorities, clinical guidelines and the development of robust, externally validated prognostic tools for routine patient care. This SRMA represents a timely and necessary step towards evidence-based adoption of precision prognostic tools in EC. The results will be disseminated through peer-reviewed publications and presentations to academic, clinical and patient communities. The review aligns with ongoing efforts to integrate molecular and computational approaches into gynaecologic cancer care and may ultimately contribute to more accurate risk stratification, personalized care and improved outcomes for patients with endometrial cancer. Supplementary Information Supplementary Material 1. Abbreviations Area under the curve Concordance Index Disease-free survival Endometrial cancer European Society of Gynaecological Oncology European Society for Medical Oncology European Society of Pathology European Society for Radiotherapy and Oncology International Federation of Gynaecology and Obstetrics Grading of Recommendations Assessment, Development and Evaluation Hazard ratio Immunohistochemistry Integrated discrimination improvement Lymphovascular space invasion Machine learning Microsatellite instability-high Net reclassification improvement Overall survival Progression-free survival Prediction model Risk of Bias Assessment Tool Preferred Reporting Items for Systematic Reviews and Meta-Analysis Protocols Quality in Prognosis Studies Restricted maximum likelihood Recurrence-free survival The Cancer Genome Atlas Systematic review and meta-analysis Synthesis without meta-analysis World Health Organization Appendix 1 Table 1Search strategy for MEDLINE Results from the search run on October 6, 2025S.NoSearch strategyResults1exp *Endometrial Neoplasms/29,0602endometrial cancer.ti,ab.26,4593endometrial carcinoma.ti,ab.11,5344endometrioid.ti,ab.9,1005uterine cancer.ti,ab.4,0366uterine neoplasm*.ti,ab.3,69971 or 2 or 3 or 4 or 5 or 651,9388exp *Machine Learning/1,11,2059exp *Artificial Intelligence/2,63,88210machine learning.ti,ab.1,76,15911artificial intelligence.ti,ab.95,54512deep learning.ti,ab.1,01,47113radiomics.ti,ab.15,74914neural network*.ti,ab.1,45,00015support vector machine*.ti,ab.36,35716random forest*.ti,ab.42,30817xgboost.ti,ab.10,71518prediction model*.ti,ab.60,58619predictive model*.ti,ab.52,38020prognostic model*.ti,ab.17,10521nomogram*.ti,ab.37,05722biomarker*.ti,ab.5,62,09023genomic*.ti,ab.4,36,91524proteomic*.ti,ab.1,37,00425transcriptomic*.ti,ab.1,09,34726multi-omic*.ti,ab.22,006278 or 9 or 10 or 11 or 12 or 13 or 14 or 15 or 16 or 17 or 18 or 19 or 20 or 21 or 22 or 23 or 24 or 25 or 2617,22,68128prognos*.ti,ab.9,74,28929predict*.ti,ab.24,81,18630risk stratification.ti,ab.67,05731risk prediction.ti,ab.25,82032recurrence.ti,ab.4,29,13433survival.ti,ab.13,76,83434overall survival.ti,ab.2,89,18235disease free survival.ti,ab.60,52736progression free survival.ti,ab.87,08637hazard ratio.ti,ab.1,69,39138C-index.ti,ab.13,96839AUC.ti,ab.1,78,01740time-dependent AUC.ti,ab.2814128 or 29 or 30 or 31 or 32 or 33 or 34 or 35 or 36 or 37 or 38 or 39 or 4044,82,06242validation.ti,ab.4,22,60643model performance.ti,ab.18,42544external validation.ti,ab.24,1184542 or 43 or 444,33,924467 and 27 and 41 and 4560447mouse.ti,ab.8,10,48648mice.ti,ab.11,89,07249rat.ti,ab.9,48,14650cell line.ti,ab.2,87,98551in vitro.ti,ab.16,03,52252case reports.pt.25,23,48853review.pt.37,85,8495447 or 48 or 49 or 50 or 51 or 52 or 5397,92,4015546 not 5452656Limit 55 to English language52557Limit 56 to yr = “2010–2025”479 Search strategy for MEDLINE Results from the search run on October 6, 2025 Author contribution The concept for this review was developed by the primary investigator. The search strategy was drafted by the investigator with methodological input from academic supervisors and domain experts. Screening, data extraction, and risk of bias assessments will be conducted independently by two reviewers, with arbitration by a third reviewer where necessary. Statistical analysis will be supported by a qualified statistician. All authors will review and approve the final manuscript before publication. Funding This systematic review and meta-analysis will be conducted as the part of primary investigator’s doctoral research work. The investigator is supported through the DST INSPIRE fellowship which provides personal fellowship support but does not constitute dedicated project funding. No additional external funding has been obtained for this SRMA. Institutional sources will support database access, data management, and statistical analysis. The views expressed in this protocol are those of the authors and do not necessarily reflect the views of the host institution or the funding agency. Data availability Data sharing is not applicable to this article at the protocol stage, as no datasets have yet been generated or analyzed. All data extracted during the review will be derived from published studies. Any supplementary materials, including extraction templates will be made available upon reasonable request following publication of the completed systematic review. Declarations Ethics approval and consent to participate Not applicable. This study involves secondary analysis of previously published data and does not involve human participants or identifiable patient information. Consent for publication Not applicable. Competing interests The authors declare that they have no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Priya Giri, Email: priyag13_sit@jnu.ac.in. Rakesh Kumar Saroj, Email: rakeshsaroj@jnu.ac.in. References Associated Data Supplementary Materials Supplementary Material 1. Data Availability Statement Data sharing is not applicable to this article at the protocol stage, as no datasets have yet been generated or analyzed. All data extracted during the review will be derived from published studies. Any supplementary materials, including extraction templates will be made available upon reasonable request following publication of the completed systematic review.