Low-latency neuromorphic closed-loop control of hippocampal ripples in vivo

Technical Summary: Low-latency neuromorphic closed-loop control of hippocampal ripples in vivo

Problem Statement
Real-time closed-loop neuromodulation holds transformative potential for treating neurological disorders and probing neural circuit function by timing stimulation to ongoing brain dynamics. However, effective intervention requires the low-latency, energy-efficient processing of high-bandwidth neural signals. Conventional computing architectures often struggle to meet these constraints, particularly for short-lived, high-frequency events like hippocampal ripples (30–100 ms, 100–250 Hz). Deep learning models, while powerful, often demand computational resources and energy that are incompatible with embedded, real-time deployment, frequently necessitating external processors and wireless telemetry that introduce latency overheads. Consequently, there is a critical need for low-power, low-latency computational substrates capable of edge processing for real-time closed-loop operation. While neuromorphic computing offers an event-driven, energy-efficient alternative, a fully integrated framework validated in vivo for the rapid manipulation of transient oscillations has not yet been demonstrated.

Methodology
The authors developed a fully integrated neuromorphic framework comprising four main stages: signal acquisition, spike encoding, neuromorphic processing, and feedback stimulation.

1. Signal Encoding: Continuous hippocampal local field potentials (LFPs) were bandpass filtered (100–250 Hz) and converted into discrete UP and DOWN spike trains using an asynchronous delta-modulation (step-forward) scheme. These spike trains were downsampled from 30 kHz to 1 kHz to match the simulation timestep.
2. Spiking Neural Network (SNN) Architecture: A compact, feed-forward SNN was designed with two input channels (UP/DN), two hidden layers of Leaky Integrate-and-Fire (LIF) neurons (24 and 16 units), and a single output neuron acting as the ripple detector. The network contained only 41 neurons and 530 trainable parameters.
3. Training: The SNN was trained using surrogate-gradient backpropagation through time (SG-BPTT) on a dataset of manually annotated ripples from 23 recording sessions. The objective minimized Time-to-First-Spike (TTFS) error with a 20 ms tolerance window.
4. Hardware Deployment: The trained network was deployed on SpiNNaker neuromorphic hardware (SpiNN-3 board) and integrated with the open-source Open Ephys platform. A Python bridge facilitated bidirectional communication, streaming encoded spikes to the SpiNNaker board and returning detection events to trigger stimulation.
5. In Vivo Validation: The system was tested in awake, head-fixed mice. Upon ripple detection, the system triggered an Arduino MEGA 2560 to deliver 100 ms optogenetic inhibition pulses to GABAergic interneurons in the CA1 hippocampus.

Key Results

• Detection Performance: The compact SNN achieved a median F1-score of 0.56 across 23 sessions, with high recall (0.89) but lower precision (0.50), consistent with the difficulty of distinguishing ripples from artifacts in noisy biological data. The SNN's performance was comparable to larger deep learning models (e.g., 1D-CNN, LSTM) and significantly outperformed standard, unoptimized conventional filter-based detectors.
• Hardware Efficiency: When deployed on SpiNNaker, the system operated in biological real-time. Power consumption was approximately 3.6 W at the board level (0.35 W attributed to simulation), representing a 16–200-fold reduction in energy consumption compared to GPU/CPU-based simulations.
• Latency and Intervention: Hardware-in-the-loop validation demonstrated that optimizing the feedback path (direct Python triggering vs. Open Ephys plugins) and minimizing buffer sizes (3 ms) reduced total closed-loop latency to approximately 50 ms. This latency enabled intra-event stimulation (before ripple termination) in up to 80% of detected ripples.
• In Vivo Manipulation: In the in vivo experiments, neuromorphic-triggered optogenetic inhibition significantly altered ripple dynamics. Specifically, "light ON" sessions showed a significant reduction in ripple energy (p=0.048) and trends toward reduced duration, lower peak frequency, and higher spectral entropy, particularly in the longest detected ripples.

Significance and Claims
The paper claims to establish the first practical and accessible neuromorphic framework for low-latency closed-loop control of fast brain dynamics in vivo. The authors emphasize that their work bridges the gap between compact SNNs, neuromorphic hardware, and widely adopted open-source neuroscience infrastructure.

The significance of this work lies in demonstrating that:

1. Compact SNNs can detect transient, high-frequency neural biomarkers with performance competitive with deep learning while maintaining a hardware-compatible footprint.
2. Neuromorphic systems can process live neural data and trigger feedback stimulation within the strict temporal windows required to manipulate fast brain events (specifically hippocampal ripples).
3. A complete sensing-processing-stimulation pipeline can be validated in vivo, successfully altering the dynamics of ongoing neural oscillations.

The authors position this not as a system with superior offline detection accuracy compared to optimized classical methods, but as a solution that combines competitive detection with the essential properties of event-driven computation, low power, and stable real-time operation. They acknowledge limitations, including the preliminary nature of the in vivo latency (which was higher than the optimized hardware-in-the-loop setup) and the need for future work to address false positives and generalizability across different recording systems and behavioral states. However, the study provides a proof-of-concept that neuromorphic computing can enable causal, state-dependent manipulation of fast brain rhythms.