Endometrial cancer is the sixth most common cancer in women globally. While mortality has declined in recent decades, incidence rates continue to rise. Most patients are diagnosed at an early stage, but a troubling minority will experience disease recurrence even after apparently successful treatment - often without warning from existing prognostic tools.
Current treatment planning for endometrial cancer relies on the ESMO-ESGO-ESTRO (European Society for Medical Oncology - European Society of Gynaecological Oncology - European Society for Radiotherapy and Oncology) consensus guidelines. These guidelines stratify patients into risk classes based on FIGO stage, tumor grade, histologic subtype, and molecular biomarkers (POLE mutation status, TP53 status, and microsatellite instability).
Despite these established risk classifications, some patients categorized as low risk still relapse, while others classified as high risk never do. This misclassification leads to both undertreatment (missing patients who need more intensive therapy) and overtreatment (subjecting low-risk patients to unnecessary, toxic adjuvant therapies).
A key missing piece in current risk models is the immune landscape of the tumor - the types, quantities, and states of immune cells within and around the tumor. Emerging evidence suggests these immune features independently predict prognosis, but they have not yet been integrated into clinical guidelines or standard risk stratification tools.
The study used the publicly available TCGA-UCEC (The Cancer Genome Atlas - Uterine Corpus Endometrial Carcinoma) dataset, selecting 230 patients with endometrioid EC, non-null FIGO staging, and a minimum 2-year follow-up. The cohort had a 1:3 ratio of patients who relapsed to those who did not, reflecting the real-world distribution of outcomes.
To estimate the immune cell composition of each tumor, the researchers used three independent computational deconvolution tools (CIBERSORTx, xCell, and quanTIseq). These algorithms analyze bulk RNA sequencing data - which captures gene expression from millions of cells at once - to estimate the proportions of different immune cell types in the sample, similar to separating the voices in a recorded choir.
In addition to immune cell abundances, the study computed immune signaling scores including the interferon gamma (IFN-gamma) signature - a measure of active anti-tumor immune activity. Molecular features including POLE mutation status, TP53 abnormality, microsatellite instability, tumor mutational burden (TMB), L1CAM expression, and CTNNB1 mutation were also included.
A Random Forest machine learning algorithm was used to build two models: a baseline model using only guideline-recommended features (FIGO stage, TP53, POLE mutation, microsatellite instability), and an immune-integrated model that added selected immune features. Model performance was evaluated using balanced accuracy and true positive rate.
Variable importance analysis identified nine features for the final model. FIGO stage was the strongest predictor overall, with stages IIIB and IV associated with steep drops in disease-free survival probability. However, among the molecular markers, POLE mutation status and microsatellite instability were both strongly associated with better prognosis, while TP53 abnormality did not show significant predictive power in this dataset.
Among immune features, higher levels of Natural Killer (NK) cells (up to 2.5% abundance) and CD4+ T cells (up to 24.3% abundance) were both associated with increased disease-free survival. NK cells are first-responder immune cells that kill cancer cells without prior sensitization; CD4+ T cells coordinate adaptive immune responses. Their presence in tumors reflects an active, competent immune environment.
The interferon gamma (IFN-gamma) signature was one of the strongest predictors in the model, with a variable importance coefficient above 2 - indicating a stronger effect than most other features. IFN-gamma is a cytokine (signaling molecule) produced primarily by activated T cells and NK cells that promotes anti-tumor immunity. High IFN-gamma signaling suggests an active immune assault on the tumor.
In contrast, higher levels of monocytes (immune cells that can be co-opted by tumors to suppress anti-cancer immunity) were correlated with decreased disease-free survival whenever their relative abundance exceeded 0.6%, consistent with the known immunosuppressive role of certain monocyte-derived cells in the tumor microenvironment.
The baseline model, trained on guideline-standard features only (FIGO stage, TP53, POLE mutation, microsatellite instability), achieved a balanced accuracy of 53.7% - barely better than chance. This confirms what clinicians already suspect: standard risk factors alone are insufficient for reliable recurrence prediction, especially in early-stage disease.
Adding the immune landscape features to build the immune-integrated model raised balanced accuracy to 68.6% - an improvement of approximately 15 percentage points - and improved the true positive rate (correctly identifying high-risk patients who will relapse) to 73.7%. This is a clinically meaningful improvement that could change treatment decisions for many patients.
Breaking down performance by FIGO stage, the model performed best for Stage IV disease (77.8% correctly predicted) and Stage II (68.2%), but showed the lowest accuracy for Stage III (59.1%). The false negative rate - patients who will relapse but are incorrectly classified as low risk - was kept low across stages (under 8%), prioritizing the detection of true high-risk cases.
Many of the model's false positives (predicting high risk when patients did not relapse) in stages II-IV likely reflect patients who received effective adjuvant therapy that prevented relapse. The model predicts the inherent tumor biology, not the treatment outcome - patients classified as high risk who received aggressive therapy may still achieve favorable outcomes, which the model cannot account for without treatment data.
Among early-stage patients (Stage IA-IB), the researchers identified a critical subgroup: patients predicted by the model to be low risk but who still relapsed. Examining these patients' gene expression profiles revealed three distinct immunological categories of early-stage disease with fundamentally different recurrence mechanisms.
"Hot" tumors (no-relapse early stages) displayed an active pro-inflammatory anti-tumor immune microenvironment - high immune cell infiltration, active immune signaling, and effective cancer cell killing. These tumors have favorable outcomes because the immune system is successfully containing tumor growth.
"Cold" tumors (characterized relapse early stages) showed a pro-tumoral immune-escaping environment - the tumor has successfully suppressed local immunity, creating a microenvironment where cancer cells can proliferate unchecked. These patients are correctly identified as high risk by the model.
"Exhausted hot" tumors (unexplained relapse early stages) represent a novel and clinically dangerous profile - they initially appear immunologically active (similar to hot tumors) but show signs of chronic exhaustion from prolonged immune activation. These tumors have evolved mechanisms to evade even a potent immune response, allowing them to relapse despite apparent low-risk features. Crucially, this group was enriched for genes involved in chemokine/cytokine activity, antigen presentation, and translation - suggesting active immune-tumor co-evolution.
The identification of the "exhausted hot" tumor profile is particularly important because these patients are currently assigned to the low-risk category and may not receive sufficient follow-up or adjuvant treatment. Recognizing them as a distinct high-risk group could change clinical management - potentially warranting more intensive monitoring, earlier re-staging, or enrollment in clinical trials of immunotherapy.
The findings suggest that immunotherapy may be particularly promising for the exhausted hot profile. These tumors have triggered immune responses but have evolved to overcome them - exactly the situation where immune checkpoint blockade (drugs that reinvigorate exhausted immune cells) might restore effective anti-tumor immunity. The authors explicitly suggest exploring this group for putative immunotherapy biomarkers.
More broadly, the model supports a vision of precision oncology for EC in which every patient's treatment is tailored to their specific combination of clinical stage, molecular features, and immune profile. This could mean de-escalating treatment for truly low-risk patients (reducing side effects and cost) while intensifying or personalizing therapy for patients who appear low risk but carry high-risk immune signatures.
For younger patients in whom fertility preservation is a priority, better risk stratification is especially critical. Unnecessary radical surgery can permanently impact quality of life; conversely, inadequate treatment carries recurrence risk. An immune-integrated model could help identify young, early-stage patients who are truly low risk and suitable for fertility-sparing approaches versus those who need more aggressive treatment despite early staging.
The study's main limitations include the relatively small and class-imbalanced dataset (3:1 ratio of non-relapse to relapse cases), which makes it challenging to train models that are equally sensitive to both classes. The heterogeneity of adjuvant treatments received by patients across the 1990-2016 time period further complicates model interpretation.
Since the immune features were estimated computationally from bulk RNA sequencing (a process called in silico deconvolution), rather than measured directly from tissue samples using immunohistochemistry or flow cytometry, the results require validation in prospective clinical cohorts where immune cell populations are directly measured.
The authors call for prospective clinical trials to validate the early-stage stratification model in real-world clinical settings. Additional validation on external datasets and testing on other data modalities - such as hematoxylin and eosin stained tissue scans (a standard pathology technique) and molecular diagnostic panels - would increase confidence in the model's clinical utility.
If validated, this integrated clinical-molecular-immune model represents a significant advance toward true precision oncology in endometrial cancer - one that moves beyond population-based risk categories toward individualized predictions of recurrence risk, treatment response, and follow-up intensity for each patient.