Training, Validating, and Testing Machine Learning Prediction Models for Endometrial Cancer Recurrence.

Area of Science:

• Oncology
• Genomics
• Machine Learning

Background:

• Endometrial cancer (EC) is a prevalent gynecologic malignancy in the US with increasing incidence and mortality.
• A significant percentage of EC patients (15%-20%) experience recurrence despite standard treatments.
• Accurate prediction of recurrence risk is crucial for selecting appropriate adjuvant therapies.

Purpose of the Study:

• To develop, validate, and test predictive models for endometrial cancer recurrence.
• Utilize machine learning (ML) and deep learning (DL) analytics for enhanced prediction accuracy.
• Employ a comprehensive dataset encompassing clinical, pathologic, genomic, and genetic information.

Main Methods:

• Data from the Oncology Research Information Exchange Network database were analyzed.
• Patients were stratified into low-risk, high-risk, and non-endometrioid histology groups.
• Multivariate models were trained, validated, and tested using lasso regression, ML (MATLAB), and DL (TensorFlow), incorporating genomic data (microRNA, lncRNA, pseudogene expression) and genetic variations (SNV, CNV).

Main Results:

• Initial clinical recurrence models showed AUCs ranging from 56% to 70%.
• Models achieving AUC >80% were selected for further analysis: five for low-risk, 20 for high-risk, and 20 for non-endometrioid groups.
• Top-performing models incorporated clinical data, copy-number variations (CNVs), pseudogene expression, and single-nucleotide variations (SNVs).

Conclusions:

• ML and DL-based prediction models demonstrate superior performance compared to models relying solely on clinical and pathologic data.
• These advanced models hold promise for improving the prediction of endometrial cancer recurrence.
• Further prospective validation is necessary to establish the clinical utility of these predictive models.