Background Alzheimer's disease (AD) is a difficult to diagnose pathology of the brain that progressively impairs cognitive functions. Computer-assisted diagnosis of AD based on image analysis is an emerging tool to support AD diagnosis. In this article, we explore the application of Supervised Switching Autoencoders (SSAs) to perform AD classification using only one structural Magnetic Resonance Imaging (sMRI) slice. SSAs are revised supervised autoencoder architectures, combining unsupervised representation and supervised classification as one unified model. In this work, we study the capabilities of SSAs to capture complex visual neurodegeneration patterns, and fuse disease semantics simultaneously. We also examine how regions associated to disease state can be discovered by SSAs following a local patch-based approach. Methods Patch-based SSAs models are trained on individual patches extracted from a single 2D slice, independently for Axial, Coronal, and Sagittal anatomical planes of the brain at selected informative locations, exploring different patch sizes and network parameterizations. Then, models perform binary class prediction - healthy (CDR = 0) or AD-demented (CDR > 0) - on test data at patch level. The final subject classification is performed employing a majority rule from the ensemble of patch predictions. In addition, relevant regions are identified, by computing accuracy densities from patch-level predictions, and analyzed, supported by Atlas-based regional definitions. Results Our experiments employing a single 2D T1-w sMRI slice per subject show that SSAs perform similarly to previous proposals that rely on full volumetric information and feature-engineered representations. SSAs classification accuracy on slices extracted along the Axial, Coronal, and Sagittal anatomical planes from a balanced cohort of 40 independent test subjects was 87.5%, 90.0%, and 90.0%, respectively. A top sensitivity of 95.0% on both Coronal and Sagittal planes was also obtained. Conclusions SSAs provided well-ranked accuracy performance among previous classification proposals, including feature-engineered and feature learning based methods, using only one scan slice per subject, instead of the whole 3D volume, as it is conventionally done. In addition, regions identified as relevant by SSAs' were, in most part, coherent or partially coherent in regard to relevant regions reported on previous works. These regions were also associated with findings from medical knowledge, which gives value to our methodology as a potential analytical aid for disease understanding.