Training, Validating, and Testing Machine Learning Prediction Models for Endometrial Cancer Recurrence

JCO Precis Oncol 2025 AI 6 Explanations View Original
Original Paper (PDF)

Unable to display PDF. Download it here or view on PMC.

Plain-English Explanations
Page [1, 2]
Why Predicting Endometrial Cancer Recurrence Matters

After initial treatment, endometrial cancer returns in a meaningful fraction of patients, and when it does, it is often harder to treat. Identifying which patients are at highest risk of recurrence before it happens could allow clinicians to offer more intensive follow-up, additional therapies, or enrollment in clinical trials aimed at prevention.

Current recurrence risk is estimated using clinical factors like tumor stage, grade, and pathology type. Machine learning models that incorporate genomic data - such as mutations and gene expression - could potentially predict recurrence more precisely by capturing molecular characteristics of each individual tumor. This study tested that approach using a large clinical-genomic database.

TL;DR: This study used machine learning with clinical and genomic data to predict which endometrial cancer patients are most likely to experience disease recurrence after treatment.
Page [3, 4]
Study Design and Patient Stratification

The study used data from 892 endometrial cancer patients in the ORIEN (Oncology Research Information Exchange Network) database, a large multi-institution repository linking clinical records with genomic profiling. Patients were divided into three distinct subgroups based on established clinical criteria: low-risk endometrioid (n=329), high-risk endometrioid (n=324), and non-endometrioid tumors (n=239). Analyzing each group separately is important because these groups have very different biology and recurrence patterns.

Multiple machine learning approaches were tested including LASSO regression, random forest, XGBoost, and deep learning models. The models were trained and validated internally within the ORIEN dataset, then externally tested on data from TCGA (The Cancer Genome Atlas), an independent genomic database with separate endometrial cancer patients.

TL;DR: 892 ORIEN patients were divided into three clinical groups and tested with multiple ML models, with external validation attempted in an independent TCGA dataset.
Pages 6-6
Strong Internal Performance but Weak External Generalization

Internally, the best models achieved impressive performance within the ORIEN dataset. For the low-risk group, a model combining clinical features with copy number variation (CNV) data reached an AUC up to 0.99. For the high-risk group, pseudogene expression features were most informative. For non-endometrioid tumors, combining mutations with pseudogene expression worked best.

However, when these models were applied to the TCGA external dataset, performance dropped dramatically to AUCs of approximately 0.50 to 0.58 - essentially no better than random chance for most groups. This sharp decline in external performance is a critical finding and a common problem in AI medical research. A model that performs well on the data it was trained on does not automatically work on data from a different source.

TL;DR: Models reached AUC up to 0.99 internally but performed near chance (AUC 0.50-0.58) on external TCGA data, revealing serious generalization failure.
Page [8, 9]
The Problem of Overfitting and Dataset Differences

The collapse in external performance is likely explained by a combination of overfitting and dataset differences. Overfitting means the model learned patterns specific to the ORIEN dataset - including quirks of how data was collected, what treatments patients received, and how outcomes were recorded - rather than universal biological rules that apply to any endometrial cancer patient.

Dataset differences between ORIEN and TCGA also play a role. The two databases were built using different genomic sequencing platforms, different patient selection criteria, and different follow-up timelines. Features that are informative in one database may not have the same meaning in another. This problem - called domain shift - is one of the biggest unsolved challenges in applying AI medical models to new settings.

TL;DR: The dramatic performance drop on external data reflects overfitting to ORIEN-specific data patterns and inherent differences between the two databases - a common problem in AI medicine.
Page [10, 11]
Pseudogene Expression as a Novel Predictive Signal

One notable finding was the importance of pseudogene expression in predicting recurrence for high-risk patients. Pseudogenes are genomic sequences that resemble functional genes but were long thought to serve no purpose. Recent research has shown that pseudogene-derived RNA transcripts can actually regulate gene activity and influence cancer behavior.

This study provides evidence that pseudogene expression patterns contain real prognostic information for endometrial cancer, which is a novel and underexplored area. If confirmed in larger independent studies, pseudogene profiling could become a valuable addition to standard genomic testing for cancer patients.

TL;DR: Pseudogene expression emerged as an important recurrence predictor in high-risk endometrial cancer, pointing to an underexplored layer of cancer biology.
Page [12, 13]
Lessons Learned and the Path Forward

This study illustrates both the promise and the pitfalls of AI-driven recurrence prediction in endometrial cancer. Machine learning can achieve very high accuracy when trained and tested on the same dataset, but this does not translate to reliable performance in new clinical settings unless the model has been rigorously validated across diverse, independent populations.

Future work must prioritize external validation from the outset rather than as an afterthought. Larger, more harmonized multi-institutional genomic databases would reduce the domain shift problem. The three-group stratification approach used here - analyzing low-risk, high-risk, and non-endometrioid tumors separately - is a methodologically sound framework that future studies should adopt to avoid blending biologically distinct patient groups.

TL;DR: Strong internal but poor external AI performance highlights the critical need for multi-institutional validation from the start, with separate analysis of biologically distinct tumor subgroups.
Citation: Open Access, 2025. Available at: .