Bayesian method boosts neural recording efficiency by 17.2% over static selection

A team led by Kazushi Takehana and Hirokazu Takahashi at the University of Tokyo has introduced a novel Bayesian approach to dynamically select which electrodes to monitor during long-term neural recordings. Traditional high-density microelectrode arrays (HD-MEAs) can contain hundreds or thousands of electrodes, but practical hardware constraints often limit the number of readout channels to a fixed budget. The researchers formulated this as a sequential subset-selection problem and applied a discounted Poisson-Gamma model combined with Thompson sampling—a technique commonly used in reinforcement learning for balancing exploration and exploitation.

In offline tests using nine 34-hour recordings from dissociated neuronal networks, the method selected 100 electrodes out of 529 candidates at each time step. The algorithm consistently outperformed static and heuristic selection strategies, capturing the largest fraction of spikes available to an ideal oracle. By the end of the recordings, the active electrode set had turned over by 47.8%, and the Bayesian method exceeded static selection by 17.2 percentage points in spike capture. In an online demonstration with 1,024 routed electrodes, the adaptive method successfully identified the first synchronized burst and enabled center-of-activity trajectory analysis. The work shows that uncertainty-aware exploration and temporal discounting can dramatically improve the efficiency of long-term neural recordings under fixed readout constraints.