Characterizing hippocampal replay using hybrid point process state space models

In the hippocampus, replay sequences are temporally compressed patterns of neural spiking that resemble patterns that occur when the animal is moving through the environment. Because replay sequences typically occur when the animal is at rest, replay is hypothesized to be part of an internal cognitive process that enables the retrieval of past spatial memories and the planning of future movement. Traditionally, replay sequences have been discovered by identifying sharp wave ripples (SWRs)—high frequency oscillations that occur in association with replay—and then looking within SWRs for spatially continuous patterns of neural spiking. This does not fully account for the content or timing of replay sequences, however. Replay sequences do not always co-occur with sharp wave ripples, have more complex dynamics than spatially continuous movement, have different temporal ordering than during movement, and change based on task. In this work, we introduce a hybrid state space framework to describe the richness of replay sequences. We show how defining discrete latent states associated with continuous latent dynamics and point process observations allows us to identify when non-local replay sequences occur, categorize the type of sequence based on their inferred continuous dynamics, and decode the spatial trajectory corresponding to the replay sequence.