Spatial Modulation (SM) offers a good balance between energy and spectral efficiency of interest for next generation networks. This, together with the need for only one Radio Frequency (RF) chain, makes SM a good proposal for Internet of Things (IoT) devices. In this work, we present a method based on Deep Learning to select the optimum Modulation and Coding Scheme (MCS) in an adaptive SM system. The deep neural network is trained with supervised learning to perform a mapping between the channel conditions and the MCS from a given set. We provide simulations results for a 4 × 4 SM link which uses several coding rates and three different constellations: QPSK, 8PSK and 16QAM. Results show how the adaptive system has a throughput close to its maximum value and how the outage probability can be reduced easily by applying a back-off margin to the neural network output.