Site-specific recurrence prediction in patients with endometrial cancer using multiclass machine learning models

Abstract Objective This study investigated site-specific differences in clinical factors for recurrence in patients who were newly diagnosed and treated for endometrial cancer. A model for predicting recurrence sites was generated by various machine learning technique. Methods Electronic medical records’ data were retrieved from January 2006 to December 2018 for patients who were diagnosed with endometrial cancer at the National cancer center in Korea. Recurrence sites were classified as local, regional, or distant. We used multinomial logistic regression models that modeled the log-odds for the three recurrence sites relative to non-recurrence as a linear combination of possible risk factors for the recurrence of endometrial cancer. And then, we employed several machine learning algorithms, including multinomial logistic regression models, random forest, support vector machine and artificial neural network and assess the prediction accuracy. Results The data of 631 patients were selected for analysis; there were 20, 13, and 26 cases of local, regional, and distant recurrence, respectively, and 572 patients had no recurrence. High-grade disease was associated with local recurrence; non-endometrioid histology and parametrial invasion were risk factors for regional recurrence; additionally, parametrial invasion and no lymphadenectomy were associated with distant metastasis. Random forest showed the best performance (0.9133) in prediction accuracy. (Fig 1) Other algorithms followed with 0.8674 (artificial neural network) and 0.8776 (logistic regression models) where support vector machine always predicted non recurrence (0.9082). The most important variables in Random forest were invasion depth, age, tumor size, parametrial invasion, grade, in order. Conclusion We identified different risk factors specific for each type of recurrence site and searched for recurrence predictors using various machine learning methods. If the measurement accuracy can be increased through refinement and a combination of various techniques, it would be useful in clinical practice.