BayesENDS: Bayesian Electrophysiological Neural Dynamical Systems for Alzheimer’s Disease Diagnosis

ICLR 2026 Conference Submission25469 Authors

Keywords: Bayesian Neural Dynamical Systems, Alzheimer’s Disease Diagnosis, EEG

Abstract: Alzheimer’s disease (AD) alters Electroencephalogram (EEG) through slowed oscillations and diminished neural drive, yet most AD-EEG pipelines are black-box classifiers, lacking a unifying mathematical account of how both neural activity and its interaction dynamics evolve over time. We introduce BayesENDS, a Bayesian electrophysiological neural dynamical system that explores the possibility of incorporating neuron spiking mechanisms into a Bayesian neural dynamical system. By introducing a differentiable leaky-integrate-and-fire (dLIF) prior, BayesENDS is capable of inferring population events and interaction dynamics directly from EEG—without spike or interaction annotations. The dLIF prior encodes membrane dynamics, rate/refractory constraints, and physiologically plausible frequency ranges, improving identifiability while yielding biologically plausible, subject-level biomarkers alongside AD predictions. Across synthetic event-sequence benchmarks and real AD EEG datasets, BayesENDS delivers superior performance to state-of-the-art baseline methods.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 25469