The spike classification is a critical step in the implantable neural decoding. The energy efficiency issue in the sensor node is a big challenge for the entire system. Compressive sensing (CS) theory provides a potential way to tackle this prob- lem by reducing the data volume on the communication channel. However, the constant transmission of the compressed data is still energy-hungry. On the other hand, the feasibility of direct analysis in compression domain is mathematically demonstrated. This advance empowers the in-sensor light-weight signal analysis on the compressed data. In this paper, we propose a novel selective CS architecture for energy-efficient wireless implantable neural decoding based on compression analysis and deep learn- ing. Specifically, we develop a two-stage classification procedure, including a light-weight coarse-grained screening module in the sensor and an accurate fine-grained analysis module in the server. To achieve better energy efficiency, the screening module is designed by the Softmax regression, which can complete the low-effort classification task at the sensor end and screen the high-effort task to transmit their compressed measurements to the remote server. The fine-grained analysis located in server end is constructed by the customized deep residual neural network. It can not only promote the spike classification accuracy, but also benefit the model quality of in-sensor Softmax model. The extensive experimental results indicate that our proposed selective CS architecture can gain more than 60% energy savings than the conventional CS architecture, yet even improve the accuracy of state-of-the-art CS architectures.