Using 1D-convolutional neural networks to detect and interpret sharp-wave ripples

Sharp-wave ripples (SWR) are high frequency events recorded in the local field potential (LFP) of the hippocampus of rodents and humans. During SWR, the sequential firing of ensembles of neurons act to reactivate memory traces of previously encoded experience. SWR-related interventions can influence hippocampal-dependent cognitive function, making their real-time detection crucial to understand underlying mechanisms. However, existing SWR identification tools mostly rely on using spectral methods, which remain suboptimal. Here, we introduce a 1D convolutional neural network (CNN) operating over high-density LFP recordings to detect hippocampal SWR both offline and on-line. The adapted architecture included seven convolutional layers composed of different kernels to process 8-channel LFP inputs in increasing hierarchical complexity and one output layer delivering the probability of an occurring SWR. We report offline performance on several types of recordings (e.g. linear arrays, high-density probes, ultradense Neuropixels) as well as on open databases that were not used for training. By saturating the operation of different kernels, we examine and interpret their optimal behaviour associated to the ground truth versus a random selection. We then use dimensionality reduction techniques to visualize how the network processes information across layers. Finally, we show how by building a plug-in for a widely used open system such Open Ephys, our method detects SWRs in real time. We conclude with discussion on how this approach can be used as a discovery tool for better understanding the dynamics of SWR.