Neuronal Waveform Classification in Multielectrode Recordings Using Machine Learning Techniques and Multidimensional Analysis

Extracellular recordings of neuronal spikes are crucial for studying brain activity. These signals are typically classified based on firing patterns and waveform shape, particularly trough-to-peak duration. While useful, this method oversimplifies the diversity of cortical neurons and discharge patterns. Recent advances in recording and analysis techniques allow for more precise waveform classification, though the main criteria remain waveform features. We aim to develop an automatic spike waveform classifier using advanced machine learning techniques selected from a range of candidate methods based on their optimized performance, such as Uniform Manifold Approximation and Projection (UMAP), Gaussian Mixture Model (GMM), and Random Forest (RF). The classifier is part of the working progress of a preprocessing pipeline previously developed. For the classifying step, we use all voltage samples that define each waveform, enabling a multi-dimensional analysis. To evaluate our approach, RF model was trained and tested on a subset of electrophysiological recordings from the human visual cortex achieving high [Formula: see text]-scores. The comparison of the classified neurons was carried out between our method and a waveform analysis toolbox described in the literature. Our method improves the characterization of the clusters of waveforms based on statistical measurements that found a third group while the accepted method categorizes just broad and narrow waveforms, labeling some as unclassifiable.