UNVEILING THE SELECTIVITY PATTERNS OF BIVALVES AT THE K-PG MASS EXTINCTION EVENT: A MACHINE LEARNING APPROACH

The past extinction events in the fossil record present a unique opportunity to investigate why some groups are more vulnerable to extinction than others, which is critical for a better understanding of the present biotic crisis. While most studies employed general linear models and the correlation between a specific extrinsic and/or intrinsic variable and extinction intensity, more recent studies argued for a novel multivariate approach, which can simultaneously facilitate understanding the non-linear interactions between predictor variables and the underlying evolutionary consequences. The present study examines the selectivity patterns of the Cretaceous – Paleogene (K-Pg) mass extinction event for the marine bivalve fauna using two machine learning algorithms, Decision Tree and Random Forest, to identify pathways of extinction and the relative importance of predictor variables in a comprehensive multivariate framework. While other algorithms, such as gradient boosting, can result in better-fitting models, these protocols often lead to too complex models that cannot be untied to explain the underlying processes. A global dataset of 87 extinct, 204 surviving bivalve genera and their 11 ecological and physiological traits has been compiled from available databases and previous literature. The species richness, occurrence, geographic range, and environmental breadth of a genus appear to be the most important predictors of extinction vulnerability at the K-Pg boundary. The present study highlights the importance of machine learning methods to detect complex relationships in large paleontological datasets.