Prediction models for gynecological cancers: an assessment from a statistical perspective.

Abstract

Objective: To systematically evaluate the methodological quality and statistical rigor of recent prediction model studies (2020-2025) for ovarian, cervical, and endometrial cancers.

Methods: We performed a systematic assessment of PubMed literature (January 2020-April 2025), including studies developing, validating, or updating diagnostic/prognostic models for these cancers. Methodological quality and risk of bias were assessed using the Prediction Model Risk Of Bias Assessment Tool across 4 domains (participant selection, predictors, outcome, and analysis). Sub-group analyses compared studies by publication period and Journal Citation Report quartile.

Results: Among 192 included studies, Prediction Model Risk Of Bias Assessment Tool assessment revealed a high overall risk of bias in 96.9% (n = 189). Key issues included a high risk of bias in the analysis domain (89.1%, n = 171) and participant selection (85.9%, n = 165), primarily due to flawed methods and use of unsuitable cohorts (eg, public databases). External validation was critically lacking (62.5% performed none, only 6.8% performed ≥3), and statistician involvement was minimal (2.6%). While baseline reporting improved significantly from 2020-2022 (39.6%) to 2023-2025 (59.7%, p = .02), deficiencies in sample size, outcome definition, analytical methods, and validation practices showed no significant improvement.

Conclusions: Current gynecological cancer prediction models exhibit widespread methodological shortcomings and a high risk of bias, severely limiting clinical utility. Urgent adherence to Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis standards, prioritized multi-center external validation, integration of statisticians, and reduced reliance on single public data sets are essential for developing reliable and applicable models.