A Compositional Framework for Interpreting Clustering of Time Series from Neural Cultures

This work addresses the problem of identifying patterns in the electrical activity of neural cultures recorded using high-density MicroElectrode Arrays (MEAs). Pattern identification is typically performed through a two-step procedure: first, a large set of time windows is extracted from the recorded time series around each detected spike; second, clustering methods are applied to group similar time windows across all MEA sensors. We propose a compositional framework for interpreting the outcomes of this procedure. In particular, we suggest quantifying similarities between clusters and between sensors by computing distances between the row and column profiles, respectively, of an appropriate contingency table. Several distance measures are considered, including those originally developed for compositional data analysis, as well as measures commonly used in information theory and machine learning. Our results highlight the importance of carefully selecting the distance metric, especially when the aim is to assess similarities across sensors.