ObjectiveWe aimed to develop an ovarian cancer-specific predictive framework for clinical use platinum-sensitivity and prognosis using machine learning methods based on multiple biomarkers, including circulating tumor cells (CTCs).Patients and MethodsWe enrolled 156 epithelial ovarian cancer (EOC) patients, randomly assigned into the training and validation cohorts. Eight machine learning classifiers, including Random Forest (RF), Support Vector Machine, Gradient Boosting Machine, Conditional RF, Neural Network, Naive Bayes, Elastic Net, and Logistic Regression, were used to derive predictive information from 11 peripheral blood parameters, including CTCs. Through the advanced CanPatrol CTC-enrichment technique, we detect CTCs and classify them into subpopulations: epithelial, mesenchymal, and hybrids. Survival curves were generated by Kaplan–Meier method and calculated through the Log rank test.ResultsMachine learning techniques, especially the Random Forest classifier, were superior to conventional regression-based analyses in predicting multiple clinical parameters related to EOC. The values for the receiver operating characteristic (ROC) curve for segregating EOC with advanced clinical stages and platinum-sensitivity were 0.796 (95% CI, 0.727–0.866) and 0.809 (95% CI, 0.742–0.876), respectively. Stepwise, we used the unsupervised clustering analysis to identify EOC subgroups with significantly worse overall survival (OS), especially in the advanced-stage group with the p-value of 0.0018 (HR, 2.716; 95% CI, 1.602–4.605) for progression-free survival (PFS) and 0.0037 (HR, 2.359; 95% CI, 1.752–6.390) for overall survival (OS).ConclusionMachine learning systems could provide risk stratification for EOC patients before initial intervention through blood variables, including circulating tumor cells. The predictive algorithms could facilitate personalized treatment options through promising pre-treatment stratification of EOC patients.Trial registrationChiCTR-DDD-16009601 Registered 25 October 2016.