Background: We have recently conducted a study to evaluate the predictive value of proteins associated with extracellular vesicles (EVs) isolated using an alternating current electrokinetic (ACE) system, for the detection of early stage (I & II) ovarian cancer cases. Ovarian cancer patients, typically, have few symptoms in stages I & II; the disease is often undetected until it is advanced. Currently, only 15% of cases are detected at early stages when clinical treatment outcomes are improved. Methods: We tested the feasibility of detecting the presence of early stage (I-II) ovarian cancer using protein biomarkers carried by EVs in patient blood. Using an ACE microelectrode array, EVs were purified from ovarian cancer patient plasma (N = 70; Stage I = 51, Stage II = 19) and female non-cancer donors (N = 290) and tested for the presence of cancer biomarker proteins with multiple bead-based multiplex immunoassay kits. An adaptive boosted trees machine learning algorithm with 100x cross-validation was developed to identify the most important features for differentiation between cases and controls. Results: The machine learning algorithm identified relevant biomarkers for the detection of ovarian cancer with a receiver-operator characteristic area under the curve (AUC) of 0.952 (95% CI: 0.918 – 0.985). Overall sensitivity was 72.9% at 99% specificity, rising to 84.3% at 95% specificity. When further stratified, Stage I sensitivity was 70.6% and stage II sensitivity was 78.9% at the 99% specificity level, as expected. We also separated the cases into endometroid carcinoma (N = 24) and serous carcinoma (N = 36). At the 99% specificity level, the sensitivity was 87.5% and 63.9%, respectively. Conclusions: These results suggest that it is possible to detect early-stage ovarian cancer by analyzing the EV-bound proteins isolated with the ACE technology. EVs may carry the earliest signals of cancer development and a blood-based EV-protein biomarker test can yield high sensitivity and specificity. Stable biomarker detection with small input volumes may be possible. These results require further investigation in larger datasets to validate the current observations with the potential application for general, routine screening, as well as for populations at elevated disease risk.