In recent years, researchers have established that endometrial cancer (EC) can be divided into four molecular subtypes - POLEmut, MMR-deficient, copy-number high, and copy-number low - each with different prognoses and treatment implications. However, identifying these subtypes requires expensive genomic testing that is unavailable in many clinical settings worldwide.
At the same time, clinical pathologists routinely perform immunohistochemistry (IHC) on tumor samples - a much cheaper and widely available test that measures protein expression. Key IHC markers like p53, Ki67, MMR proteins, ER, and PR are already standard in EC pathology reports.
This study asked: can a machine learning model built from routine IHC markers, combined with standard clinical variables, achieve molecular-level prognostic accuracy? The result was the RF16 model - a random forest classifier using 16 variables that captures the molecular characteristics of EC without expensive gene sequencing.
The study enrolled 849 EC patients, split 698 for training and 151 for testing. The 16 variables in the RF16 model include patient characteristics (age, BMI), pathological staging features (FIGO stage, depth of myometrial invasion, lymphovascular space invasion), and IHC markers (p53, Ki67, MMR protein status, estrogen receptor ER, progesterone receptor PR).
A random forest (RF) was chosen as the machine learning method because it handles both categorical and continuous variables well, tolerates missing data, and provides variable importance rankings that help interpret which factors drive predictions. Random forests build multiple decision trees and combine their votes to produce a robust final prediction.
The model was developed to predict overall survival (OS) and progression-free survival (PFS). Key IHC markers like p53 mutational status and MMR deficiency were included because they act as surrogates for the molecular subtypes - p53 aberrant expression marks copy-number high tumors, and MMR IHC detects MMR-deficient tumors, the same subtypes identified by expensive DNA sequencing.
In the independent test cohort, RF16 achieved an AUC of 0.991 and a C-index of 0.938. The C-index measures how well the model ranks patients by survival risk - a score of 0.938 means the model correctly ordered survival outcomes in 93.8% of patient pairs, a remarkably high level of discrimination.
The model also generated accurate nomograms - visual tools that translate a patient's combination of risk factors into an individualized probability of surviving 1, 3, or 5 years. Calibration curves confirmed that the predicted probabilities closely matched actual observed survival rates at each time point.
Variable importance analysis showed that FIGO stage, p53 status, MMR status, and lymphovascular space invasion (LVSI) were the most influential predictors. These findings align with known prognostic factors in EC and validate that the model is capturing clinically meaningful biological signals rather than spurious statistical relationships.
The central contribution of this study is demonstrating that IHC-based markers can effectively replace expensive molecular testing for prognostic stratification. MMR IHC detects MMR-deficient tumors with high accuracy - the same tumors identified as MSI-H by microsatellite instability testing or NGS. Similarly, aberrant p53 IHC identifies the copy-number high subtype that carries the worst prognosis.
The clinical implications are most significant for low-resource settings - hospitals in developing countries or rural areas that routinely perform IHC but cannot afford molecular profiling. The RF16 model could give these institutions access to molecular-level risk stratification using equipment and workflows they already have.
The very high AUC (0.991) warrants some caution - models with near-perfect training performance sometimes reflect overfitting or patient overlap issues. The independent test cohort of 151 patients helps address this concern, but external validation in diverse cohorts from different institutions would strengthen confidence in the model's generalizability.
In practice, the RF16 model could be integrated into pathology reporting workflows: after the standard IHC panel is run on an EC biopsy, the results are entered into the model to generate a risk score and predicted survival probabilities. This adds minimal cost and time to existing workflows while providing substantially more prognostic information.
The model's individualized survival probability output - delivered through a nomogram - supports shared decision-making between physicians and patients. Rather than grouping patients into coarse categories like 'low risk' or 'high risk', the model gives specific probability estimates that can be discussed with patients when weighing aggressive treatment versus observation.
Future directions include prospective trials testing whether RF16-guided treatment allocation improves outcomes compared to standard clinical staging, and external multicenter validation to confirm performance across diverse populations, imaging protocols, and IHC staining methods that may vary between institutions.