REAL-TIME DETECTION AND CLOSED-LOOP PERTUBATION OF HIPPOCAMPAL SHARP-WAVE RIPPLES USING BONSAI-RX AND ONIX

Sharp-wave ripples (SWRs) are synchronous bursts in hippocampal CA1 that are hypothesized to orchestrate memory consolidation, yet causal tests demand sub-cycle (< 10 ms) closed-loop control. We therefore built an open pipeline that pairs the Open Ephys ONIX acquisition system with a flexible deep-learning inference engine as a plug-and-play Bonsai-RX package. To ensure robustness, we benchmarked various model architectures on publicly available CA1 data—comprising both expert- and auto-annotated datasets—demonstrating the comparative performance of different lightweight temporal-convolutional networks (TCNs). We have also recorded local-field potentials spanning the CA1 pyramidal layer in freely moving mice. Off-line, our optimized TCN models achieved state-of-the-art ripple detection, systematically outperforming conventional thresholded ripple-band power detection or other machine learning based methods. Crucially, the developed Bonsai-RX package allows users to not only combine multiple models for ensemble inference but also flexibly adapt models to novel experimental data via human-in-the-loop supervision. On-line, continuous LFPs were fed directly to the network to produce a ripple-probability trace that triggered electrical stimulation via ONIX with hardware-to-pulse latencies below 5 ms—well within a few ripple cycles from event onset. On-line accuracy in freely moving animals matched off-line performance, demonstrating robust generalization across datasets. This openly distributed Bonsai-RX package thus provides millisecond-scale, state-of-the-art SWR detection and the flexibility to customize models to distinct experimental needs, enabling new closed-loop experiments to dissect the causal role of ripples during sleep and memory processing.