Explanation and evaluation of the random forest (RF) classifier. (A) Schematic illustration of the classification of samples by the RF. (B, C) Representative classification trees from the discrimination between malignant and benign tumors. These trees are only representations out of 4,000 trees constructed in the RF classifier. The final class is determined as a result of voting by all 4,000 trees. (D, E) The highest accuracy of prediction (D) and the AUC (E) using different numbers of samples in RF classification between malignant and benign tumors. The mean accuracy and AUC with these 95% confidence intervals were presented for 10 independent sets of randomly selected data of 20%, 40%, 60%, and 80% of patients from the training and test cohorts.
ARTICLE ABSTRACT
We aimed to develop an ovarian cancer–specific predictive framework for clinical stage, histotype, residual tumor burden, and prognosis using machine learning methods based on multiple biomarkers.
Overall, 334 patients with epithelial ovarian cancer (EOC) and 101 patients with benign ovarian tumors were randomly assigned to “training” and “test” cohorts. Seven supervised machine learning classifiers, including Gradient Boosting Machine (GBM), Support Vector Machine, Random Forest (RF), Conditional RF (CRF), Naïve Bayes, Neural Network, and Elastic Net, were used to derive diagnostic and prognostic information from 32 parameters commonly available from pretreatment peripheral blood tests and age.
Machine learning techniques were superior to conventional regression-based analyses in predicting multiple clinical parameters pertaining to EOC. Ensemble methods combining weak decision trees, such as GBM, RF, and CRF, showed the best performance in EOC prediction. The values for the highest accuracy and area under the ROC curve (AUC) for segregating EOC from benign ovarian tumors with RF were 92.4% and 0.968, respectively. The highest accuracy and AUC for predicting clinical stages with RF were 69.0% and 0.760, respectively. High-grade serous and mucinous histotypes of EOC could be preoperatively predicted with RF. An ordinal RF classifier could distinguish complete resection from others. Unsupervised clustering analysis identified subgroups among early-stage EOC patients with significantly worse survival.
Machine learning systems can provide critical diagnostic and prognostic prediction for patients with EOC before initial intervention, and the use of predictive algorithms may facilitate personalized treatment options through pretreatment stratification of patients.
