Insights and limitations of endometrial cancer risk prediction models for clinical applicability: a systematic review

TL;DR

This review evaluates endometrial cancer risk models and finds they have moderate accuracy but lack validation and diversity, limiting their use in clinical practice.

Contribution

The study systematically reviews and highlights the limitations of current endometrial cancer risk models, emphasizing the need for diverse data and improved validation.

Findings

Nine EC risk models were identified, with most showing moderate discrimination (AUROC 0.64–0.77) and limited external validation.

Models were primarily developed in postmenopausal White women, limiting generalizability to non-White populations.

Calibration issues and lack of diversity in datasets hinder clinical applicability and equity in cancer care.

Abstract

Endometrial cancer (EC) is the most common gynecologic cancer in high-income countries, with rising incidence rates. Risk prediction models can identify high-risk individuals, enabling targeted prevention and early intervention. Despite the development of several multivariable risk models aimed at stratifying EC risk, none have yet been adopted for clinical use in cancer prevention. This systematic review critically examines the performance, validation, and clinical applicability of existing EC risk prediction models.

We systematically searched online search engines PubMed and Ovid MEDLINE for EC risk model publications written in English from January 1, 2000, to October 9, 2024. Studies were selected based on the inclusion of multivariable models for EC risk estimation. Data extraction focused on cohort characteristics, predictors included, validation efforts, and model performance metrics such as discrimination (C-statistics or AUROC) and calibration (E/O ratio or calibration slopes). The quality of model reporting was assessed using the TRIPOD-AI guidelines.

Nine risk prediction models were identified, predominantly based on epidemiological factors, with four incorporating polygenic risk scores, and one using blood biomarkers. Most models were developed in datasets of postmenopausal women of White or European ancestry from Western countries. Only five models were externally validated; most exhibited moderate discrimination (AUROC ranging from 0.64 to 0.77). Calibration varied, with some models showing significant overestimation of risk. Importantly, the lack of racial and ethnic diversity in the development datasets limits their generalizability, particularly for non-White populations.

Current EC risk prediction models show moderate performance but suffer from limited external validation, homogeneity in demographics, and exclusion of diverse populations. Future research should focus on broadening participant diversity and incorporating new risk factors, such as hormonal intrauterine device use, hysterectomies, environmental exposures, and socio-economic status. Developing dynamic models that account for these factors and model outcomes that span various forms of the disease can improve clinical relevance. Personalized, risk-based approaches targeting high-risk groups may offer a viable path forward for EC screening and prevention strategies, ensuring more equitable cancer care and improving patient outcomes.

The online version contains supplementary material available at 10.1186/s12885-025-15200-x.

Figures