Deciphering Molecular Subtypes and Histological Patterns in Uterine Corpus Endometrial Carcinoma

Uterine corpus endometrial carcinoma (UCEC) remains a significant clinical challenge, necessitating early and precise diagnostic approaches to improve patient outcomes. The study proposes a multimodal machine learning framework capable of integrating demographic, clinical, and molecular data to develop predictive models for precise classification of diseases. Multiple algorithms were systematically evaluated using nested cross-validation and hyperparameter optimization. The one-dimensional Convolutional Neural Network (1D-CNN) dominated the performance across the board with an accuracy of 97.2% and an AUC of 0.98 in the internal validation cohort. The major issue of data imbalance in the analyzed dataset was approached through applying synthetic minority oversampling techniques, meanwhile, dimensionality reduction through Uniform Manifold Approximation and Projection (UMAP) exhibited separate latent structures for molecular profiles. Model interpretability was facilitated through SHapley Additive exPlanations (SHAP) and Integrated Gradients, which consistently identified biologically relevant molecular features contributing to classification. Finally, decision-curve analysis showed how 1D-CNNs enjoyed a greater net clinical benefit for a range of threshold probabilities, thus emphasizing the possibility of this technology’s adaptation into real-world clinical decision support systems.