A Cost-Effective, Machine Learning-Based New Unified Risk-Classification Score (NU-CATS) for Patients with Endometrial Cancer

Abstract

Treatment for endometrial cancer (EC) with radiotherapy is increasingly guided by molecular risk classifications. Derived from genomic profiling of The Cancer Genome Atlas (TCGA) project, several EC risk classification systems, including ProMisE and Leiden/TransPORTEC, have been developed. However, the current systems were developed on a relatively homogeneous population. Black or African American (BOAA) patients have consistently been demonstrated to have worse stage-adjusted prognosis than Caucasians. Given this, we intended to develop a new unified risk classification system (NU-CATS) for EC patients using machine learning (ML) utilizing datasets with demographically diverse populations. TCGA-Uterine Corpus Endometrial Carcinoma (n = 596), Memorial Sloan Kettering-Metastatic Events and Tropisms (MSK-MET, n = 1,315) and the American Association for Cancer Research Project Genomics Evidence Neoplasia Information Exchange (AACR-GENIE, n = 4,561) were used to identify genetic alterations and clinicopathological features, including age, race, stage, histologic grade and features, and distribution of metastatic disease. Software packages including Keras, Pytorch, and Scikit Learn were tested to build artificial neural networks (ANNs) with a binary output as either intra-abdominal metastatic lesions vs. non-metastatic. A 5-layered ANN (5-6-4-2-1) using 5 inputs ('age at surgery', 'histology', 'race', 'mismatch repair status' and 'TP53'). The optimal performing ANN was selected and cross validated. The weights and biases of the trained ANN were used to reconstruct the algorithm. BOAA patients with EC have worse prognosis than Caucasians, adjusting for TP53 or POLE mutation status. TP53 is the most common gene differentially altered by race in EC. Over 75% of BOAA patients carry TP53 mutations as compared to approximately 40% of Caucasians. Older age is associated with an increasing likelihood of TP53 mutations, high risk histology, and distant metastasis. For patients above age 70, 91% of BOAA and 60% of Caucasian EC patients carry TP53 mutations. The NU-CATS that incorporates age, race, histology, mismatch repair (MMR) status, and TP53 mutation status showed 75% accuracy in prognosticating intra-abdominal metastasis. A higher NU-CATS (>50) is associated with about 2-fold increased risk of having positive pelvic or para-aortic lymph nodes (LNs) and distant. NU-CATS was shown to outperformed TransPORTEC model for estimating risk of FIGO Stage I/II disease progression and survival in BOAA EC patients. Despite adjusting for molecular classification, race and age retain prognostic importance in EC. NU-CATS, a ML-based, cost-effective algorithm, incorporates diverse clinicopathologic and molecular variables of EC, and yields superior prognostication of the risk of nodal involvement, distant metastasis, disease progression, and overall survival as compared to other classification systems.