Interpretable Machine Learning Identifies an Emergent Absence Seizure Mechanism
Absence epilepsy is a generalized seizure disorder marked by widespread spike-and-wave oscillations and sudden lapses in consciousness. Although no consensus exists on mechanisms of consciousness, several frameworks highlight dynamic processes relying on cortical and thalamocortical feedback loops. What those dynamics entail is an open question. Here, using interpretable machine learning, we identified a system of dynamical equations reproducing absence seizure dynamics directly from electrocorticogram recordings. The data-generated and human-interpretable model entailed multiarea synchronization on a chaotic attractor, challenging the idea of a single cortical or thalamic origin. The model contained several interconnected feedback loops which are stabilized at particular phase offsets to drive the emergence of the large amplitude seizure oscillations. High-density multielectrode recordings revealed a synchronized seizure network linking cortical layer 5 neurons of the somatosensory and motor systems with the posterior thalamic nucleus (PO), a higher order thalamic nucleus driving cortical and thalamic bursting which corresponded to the model coupling functions. Multisite optogenetics and multielectrode recordings then showed PO acts as timing-dependent gate of cortico-cortical L5 connectivity to promote the seizures. Overall, our work identifies absence seizures as corresponding to the confinement of dynamics on an attractor which relies on the same circuit substrate as has been identified for loss of consciousness in general anesthesia. This work introduces a unified approach to identify, explain, and test how distributed network instabilities contribute to a disorder of consciousness and provides an explicit dynamical framework describing the underlying multiarea feedback loops.