Logo image
Back
Machine learning approach for risk-based prediction of endometrial cancer recurrence
Abstract   Peer reviewed

Machine learning approach for risk-based prediction of endometrial cancer recurrence

Andrew Polio, Erin George, Ahmad Tarhini, Casey Cosgrove, Marilyn Huang, Bradley Corr, Aliza Leiser, Lisa Landrum, Robert Rounbehler and Jesus Gonzalez-Bosquet
Gynecologic oncology, Vol.200(Suppl 1), pp.189-189
09/2025
DOI: 10.1016/j.ygyno.2025.04.238

View Online

Abstract

Objectives Endometrial cancer is the most common gynecologic cancer in the United States, with rising incidence and mortality. Despite optimal surgical and adjuvant treatment, approximately 15–20 % of all patients will recur. Studies have identified multiple clinical and pathologic factors associated with recurrence, but models of disease relapse remain inconsistent. Therefore, to better select patients for adjuvant therapy, it is important to accurately predict patients at risk for recurrence. This study aimed to train, validate and test models of endometrial cancer recurrence using machine (ML) and deep learning (DL) analytics in a large, comprehensive dataset. Methods Data from endometrial cancer patients was downloaded from the Orien Consortium database and stratified into low-risk (FIGO grade 1 and 2, stage I, n = 329), high-risk (FIGO grade 3 or stages II, III, IV, n = 324) and non-endometrioid histology (n = 239) groups. Clinical, pathologic, genomic and genetic data were used for the analysis. Genomic data from RNA sequencing included microRNA, long noncoding RNA, isoforms and pseudogene expressions. Genetic variations from whole exome sequencing included single nucleotide variations (SNV) and copy number variations (CNV). Sequencing was performed in Illumina platforms. In the discovery phase, we used univariate analysis of variance (ANOVA) with cross-validation to identify variables informative for recurrence (P < 0.05). Then, we created multivariate lasso regression models. The models' performance was measured using the area under the curve (AUC). We selected the best-performing models for training, validation and testing using ML analytics with MATLAB and DL methods with TensorFlow. Results Recurrence clinic models for low-risk, high-risk and non-endometrioid histology had an AUC of 56 %, 81 % and 65 %, respectively. To improve performance, models with AUC >80 % were selected: five models were identified for low-risk, 20 for high-risk and 20 for non-endometrioid groups. For the low-risk group, the two best models identified with ML and DL analytics included clinical data and CNVs; the only clinical variable informative about recurrence in this group was body mass index (BMI). For the high-risk group, three of the five best-performing models included pseudogene expression. Finally, for the non-endometrioid group, pseudogene expression and SNV were overrepresented in the best-performing models. Conclusions Prediction models of endometrial cancer recurrence built with ML and DL analytics performed better than classical methods using clinical and pathologic data alone. Further study is required in independent and prospective datasets to determine clinical utility.

Details

Metrics

1 Record Views
Logo image