Interpretable Machine Learning Model Supported by Parallel Ensemble Learning to Predict Local Recurrence for Patients with Cervical Cancer

Abstract

Local recurrence in patients with cervical cancer after definitive chemoradiation can be difficult to treat. Early prediction of local recurrence and risk adapted treatment strategies can lead to improved outcomes. Previous machine learning models using clinical features to predict local failure for cervical cancer have achieved modest discriminatory power with reported AUCs ranging from 0.73 - 0.80. In this study, we aim to improve upon a previously developed clinical model at our institution by using an interpretable machine learning classifier, Explainable Boosting Machine (EBM), and evaluate the impact of a parallel ensemble design on overall model performance.A retrospective cohort of 95 patients with cervical cancer (FIGO stage IB2 - IVA, no para-aortic involvement) treated with definitive chemoradiation between 2009 - 2016 were included. Patients had a mean age of 48.1 years (range, 26.2 - 78.2 years). Most patients had either FIGO stage IB2 (41%) or IIIB (38%) cervical cancer. After a median follow-up period of 59.7 months (range, 3.5 - 135.3 months), 14 patients experienced a local recurrence and there were 27 deaths. An EBM classifier was trained using the clinical features of age, race, histology, year treated, nodal status at diagnosis, stage, and clinical tumor size at diagnosis. The model was trained and tested using 5-fold cross validation. This process was repeated with 25 independent runs, and the predicted scores for each patient were averaged to determine the ensemble score. The clinical ensemble EBM model was compared to the single iteration EBM model (at the median AUC achieved) in a statistical software with paired samples t-test with ROC analysis function.The ensemble EBM model achieved an AUC of 0.853 ± 0.090 with the clinical features of age, year treated, stage, and clinical tumor size at diagnosis, which was statistically significantly higher than the single iteration EBM model (AUC 0.752 ± 0.090), P = 0.022.Parallel ensemble learning increases the discriminatory power of the proposed interpretable EBM model. The model's performance compares favorably to previous models predicting local failure for cervical cancer using clinical features. This tool could be potentially used to aid clinicians in prognostication for cervical cancer and warrants further evaluation in a separate validation cohort.