A challenge in applying machine learning algorithms to pathological speech classification is the labelled data shortage problem. Labelled data acquisition often requires significant human effort and time-consuming experimental design. Further, for medical applications, privacy and ethical issues must be addressed where patient data is collected. While labelled data are expensive and scarce, unlabelled data are typically inexpensive and plentiful. In this paper, we propose a semi-supervised learning approach that employs a generative adversarial network to incorporate both labelled and unlabelled data into training. We observe a promising accuracy gain with this approach compared to a baseline convolutional neural network trained only on labelled pathological speech data.