Kemere, C. T. | bioRxiv Science

bioRxiv2024

RealtimeDecoder: A fast software module for online clusterless decoding

Decoding algorithms provide a powerful tool for understanding the firing patterns that underlie cognitive processes such as motor control, learning, and recall. When implemented in the context of a real-time system, decoders also make it possible to deliver feedback based on the representational content of ongoing neural activity. That in turn allows experimenters to test hypotheses about the role of that content in driving downstream activity patterns and behaviors. While multiple real-time systems have been developed, they are typically implemented in C++ and are locked to a specific data acquisition system, making them difficult to adapt to new experiments. Here we present a Python software system that implements online clusterless decoding using state space models in a manner independent of data acquisition systems. The parallelized system processes neural data with temporal resolution of 6 ms and median computational latency <50 ms for medium- to large-scale (32+ tetrodes) rodent hippocampus recordings without the need for spike sorting. It also executes auxiliary functions such as detecting sharp wave ripples from local field potential (LFP) data. Performance is similar to state-of-the-art solutions which use compiled programming languages. We demonstrate this system use in a rat behavior experiment in which the decoder allowed closed loop neurofeedback based on decoded hippocampal spatial representations . This system provides a powerful and easy-to-modify tool for real-time feedback experiments.

neuroscience↗

bioRxiv2019

Unsupervised Clusterless Decoding using a Switching Poisson Hidden Markov Model

Spike sorting is a standard preprocessing step to obtain ensembles of single unit data from multiunit, multichannel recordings in neuroscience. However, more recently, some researchers have started doing analyses directly on the unsorted data. Here we present a new computational model that is an extension of the standard (unsupervised) switching Poisson hidden Markov model (where observations are time-binned spike counts from each of N neurons), to a clusterless approximation in which we observe only a d-dimensional mark for each spike. Such an unsupervised yet clusterless approach has the potential to incorporate more information than is typically available from spike-sorted approaches, and to uncover temporal structure in neural data without access to behavioral correlates. We show that our approach can recover model parameters from simulated data, and that it can uncover task-relevant structure from real neural data.

neuroscience↗