Benchmarking spike source localization algorithms in high density probes

Estimating neuron location from extracellular recordings is essential for developing advanced brain-machine interfaces. Accurate neuron localization improves spike sorting, which involves detecting action potentials and assigning them to individual neurons. It also assists in monitoring probe drift, which affects long-term probe reliability. Although several localization algorithms are currently in use, the field is nascent and arguments for using one algorithm over another are largely theoretical or based on visual inspection of clustering results. We present a first-of-its-kind benchmarking of commonly used neuron localization algorithms. We assess these algorithms using two ground truth datasets: a biophysically realistic simulated dataset, and an experimental dataset pairing patch-clamp and extracellular Neuropixels recording data. We systematically evaluate the accuracy, robustness, and runtime of these algorithms in ideal recording conditions and long-term recording conditions with electrode degradation. Our findings highlight significant performance differences; while more complex and physically realistic models perform better in ideal conditions, models relying on simpler heuristics demonstrate superior robustness to noise and electrode degradation, making them more suitable for long-term neural recordings. This work provides a framework for assessing localization algorithms and developing robust, biologically grounded algorithms to advance the development of brain-machine interfaces.Author summary: Accurately estimating neuron locations from extracellular recordings is critical to building reliable brain–machine interfaces. This spatial information enhances spike sorting and enables long-term monitoring of neural activity, especially in the presence of probe drift and electrode degradation. Despite the availability of several spike source localization algorithms, their comparative long-term performance has not been systematically benchmarked against ground truth data. In this study, we benchmark three widely used algorithms—center of mass (COM), monopolar triangulation (MT), and grid convolution (GC)—using both simulated and experimental ground truth datasets. We assess their accuracy, runtime, and robustness under ideal and degraded recording conditions. Our results reveal that while MT demonstrates higher accuracy in ideal conditions, GC and COM demonstrate superior resilience to noise and electrode degradation, making them more suitable than MT for long-term recordings. These findings provide a foundational framework for evaluating and improving spike localization algorithms and highlight the importance of robustness in real-world neural interface applications.