Currently, diabetes mellitus and breast cancer have become more widespread than ever before. Those suffering from these two types of diseases usually need a blood test or biopsy, where both extract fluids or tissues from the human body, which brings pain and a sense of discomfort. With the rise of medical biometrics, it is possible to perform non-invasive detection according to the biometric identifiers from the face of the patients. However, it is still difficult to simultaneously perform disease detection on both diabetes mellitus and breast cancer accurately. To resolve this issue, in this paper, we propose a progressive stack face-based network (PF-Net) to perform multi-class classification on diabetes mellitus, breast cancer, and healthy control using facial information. To perform diagnosis in a progressive way, a latent facial representation is first generated from a stacked sparse autoencoder. Later, the representation is fed into an ensemble layer containing several classifiers. Finally, only the effective classifiers are activated in the classification layer to make the final decision. The experiments showed our proposed method achieved an overall Accuracy of 92.94%, which outperforms a number of classification methods.