Spike sorting biases and information loss in a detailed cortical model

Abstract

Sorting electrical signals (spikes) from extracellular recordings of large groups of connected neurons is essential to understanding brain function. Despite transformative advances in dense extracellular recordings, the activity of most cortical neurons remains undetected. Small simulations with known neuron spike times offer critical ground truth data to improve spike sorting. Yet, current simulations underestimate neuronal heterogeneity and connectivity, which can potentially make spike sorting more challenging. We simulated recordings in a detailed large-scale cortical microcircuit model to link spike sorting accuracy to neuronal heterogeneity, evaluate the performance of state-of-the-art spike sorters and examine how spike sorting impacts the retrieval of information encoded in the cortical circuit. We found that modern spike sorters accurately isolated about 15% of neurons within 50 µm of the electrode shank, which contrasts with previous simulated yields but agrees with experiments. Neurons were unresolved because their spike trains were either missed (undersampling) or, when detected, incomplete or merged with other units’ (assignment biases). Neuron isolation quality was influenced by both anatomical and physiological factors (selection bias), improving with increased neuron firing rate, spike spatial extent, for neurons in layer 5, and excitatory neurons. We exposed the network to various stimuli to dissociate the impact of these biases on its stimulus discrimination ability. Surprisingly, undersampling did not affect discrimination capacity, but selection and assignment biases nearly reduced it by half. These findings posit realistic models as a complementary method to evaluate and improve spike sorting and, hence, brain activity representations.