Endometrial cancer is the most common gynecologic malignancy in developed countries, and its incidence is rising due to increasing obesity rates and an aging population. While most patients with early-stage disease are cured by surgery, approximately 10-15% will experience disease recurrence - return of the cancer after initial treatment. Identifying these patients in advance would allow clinicians to intensify adjuvant (post-surgery) treatment for high-risk patients while sparing low-risk patients from unnecessary treatment toxicity.
Current risk stratification relies primarily on traditional pathological and clinical factors such as tumor grade, stage, lymphovascular invasion, and histological subtype. Molecular profiling has added new tools, including the TCGA classification system and the ProMisE classifier, but these leave up to 59% of patients in an unclassifiable category with uncertain prognosis. Better models are clearly needed.
This pilot study investigated whether integrating genomic data - specifically from RNA sequencing of tumor tissue - with clinical data could build prediction models of recurrence superior to models using clinical data alone. The focus was specifically on endometrioid endometrial carcinoma (EEC), the most common histological subtype, where most recurrences occur despite generally favorable prognosis.
The study performed a retrospective case-control analysis at a single institution, including 62 endometrioid endometrial cancer patients: 7 who experienced recurrence (cases) and 55 who did not (controls). RNA was extracted from frozen tumor specimens collected from 1991 to 2010 and processed using RNA sequencing (RNAseq), a comprehensive technique that measures the activity level of virtually every gene in the tumor simultaneously.
From the RNAseq data, the researchers extracted multiple types of genomic features: transcriptome gene expression (how active each gene is), lncRNA (long non-coding RNA) expression (activity of non-protein-coding RNA genes), single exon expression, and various forms of genomic variation including single nucleotide variations (SNV), copy number variations (CNV), and structural variations such as fusion transcripts and retained introns.
Feature selection was performed using ANOVA (analysis of variance) with 10-fold cross-validation to identify which genomic and clinical variables most differed between recurrent and non-recurrent patients. Selected features were then used in LASSO regression models - a technique that builds predictive equations while automatically selecting the smallest set of variables that captures the most predictive information, reducing the risk of overfitting with small sample sizes.
Among all genomic and clinical features tested, lncRNA (long non-coding RNA) expression consistently emerged as the most predictive data type for EEC recurrence. In total, over 170 different prediction models were built combining different feature types, and all of the best-performing models contained lncRNA data. This consistent pattern across hundreds of model variants provides strong evidence that lncRNAs carry unique recurrence-related information.
The best model contained just five specific lncRNAs, achieving an AUC (area under the ROC curve) of 0.90 (95% CI: 0.75-1.0). An AUC of 0.90 represents excellent discriminative ability - the model correctly distinguished recurrent from non-recurrent patients in 90% of cases. Critically, adding more clinical or genomic variables to the five-lncRNA model did not improve its performance, suggesting these five molecules capture the key biological signal.
The clinical-only model, using pathological and demographic factors with FIGO stage as the only informative clinical variable, achieved an AUC of only 0.75. The lncRNA model's AUC of 0.90 thus represents a substantial 15-point improvement over the best available clinical prediction approach. This gap quantifies the added value that genomic data brings to recurrence prediction beyond what clinicians can assess from standard diagnostic information.
The best lncRNA model was validated using two independent machine learning platforms to confirm the findings were not specific to the LASSO regression methodology. Using TensorFlow with Keras (a deep learning framework), the model achieved an AUC of 1.00 with 92% accuracy when lncRNA data was combined with FIGO stage, and AUC of 1.00 with 85% accuracy using only lncRNA data. These near-perfect results in the machine learning platform validation provide strong internal confirmation of the model's discriminative power.
The second validation used MATLAB's machine learning application, which offers over 30 different classification algorithms. Again, both the lncRNA-plus-stage model and the lncRNA-only model achieved 100% accuracy and AUC of 1.00 in this platform. The consistency across multiple analytical methods - LASSO regression, TensorFlow neural network, and MATLAB classifiers - is a hallmark of a robust finding rather than an artifact of any single method.
These machine learning results must be interpreted cautiously given the small sample size - only 62 patients, with just 7 recurrences. The machine learning platforms used resampling techniques to account for the severely imbalanced classes (recurrent vs. non-recurrent), but with such small numbers, even excellent internal validation results can reflect overfitting. The critical test is external validation in an independent dataset.
The most rigorous test of a prediction model is performance in a completely independent dataset. The five-lncRNA model was validated using data from TCGA (The Cancer Genome Atlas), which includes 406 endometrioid endometrial cancer patients - 346 without recurrence and 60 with recurrence. This is dramatically larger than the original 62-patient study dataset.
In the TCGA validation, the model achieved accuracies of 78% (TensorFlow) and 86% (MATLAB), with AUC values of 0.68 and 0.78 respectively. These are good but not exceptional results - notably lower than the near-perfect performance in the internal validation. This performance drop is expected and actually reassuring: it suggests the results are real rather than inflated by overfitting to the training data.
An AUC of 0.78 in an independent external dataset represents a meaningful improvement over unaided clinical judgment for predicting EEC recurrence. The somewhat lower performance in TCGA compared with the internal cohort may reflect differences in patient population, RNA sequencing methodologies, treatment protocols across decades (the UI cohort spanned 1991-2010), and differences in how recurrence was defined and documented.
Long non-coding RNAs (lncRNAs) are RNA molecules transcribed from the genome that do not encode proteins. Once considered genomic noise, they are now recognized as important regulators of gene expression, chromatin organization, and cell behavior. Their frequent deregulation in cancer suggests they may play important roles in tumor development and progression - and, as this study demonstrates, in predicting which tumors will recur.
The five lncRNAs identified in this study have connections to known cancer biology. ENSG00000274840, ENSG00000240137, and ENSG00000253622 have been reported in other cancers, while ENSG00000250137 has been linked to obesity - the primary risk factor for endometrial cancer. This biological plausibility strengthens confidence that these lncRNAs are capturing real cancer biology rather than statistical artifacts.
The authors emphasize that this is a pilot study requiring prospective validation in larger, diverse populations before clinical use. The model should not be used to guide treatment decisions until such validation is complete. However, the findings represent an important step toward a molecular recurrence risk test for EEC that could identify the approximately 15% of early-stage patients destined to fail standard treatment, enabling earlier intervention with more intensive adjuvant therapies to improve their outcomes.