Causally Constrained Latent Graph Dynamical Models for Seizure Network Inference and Intervention Planning under Partial Observability

Abstract

Epileptic seizures emerge from distributed, time-varying interactions among neural populations, and clinical decisions increasingly depend on understanding how these interactions evolve across space, time, and modality. Contemporary monitoring spans minimally invasive intracranial recordings and ambulatory scalp systems, yet both regimes face partial observability, heterogeneous noise, and shifting dynamical regimes across pre-ictal, ictal, and post-ictal phases. This paper develops a unified technical framework for inferring directed seizure networks and translating them into intervention-relevant targets while explicitly accounting for incomplete sensor coverage and nonstationary dynamics. We propose a causally constrained latent graph dynamical model in which a hidden neural state evolves on a sparse, directed interaction graph whose edges are regularized by stability, hemispheric symmetry priors, and phase-dependent modulation. Observations from intracranial and scalp electrodes are treated as linear, bandwidth-limited projections with sensor-specific noise and artifact processes, allowing principled fusion and consistent uncertainty quantification. A variational inference procedure with structured sparsity yields posterior distributions over edge directionality and strength, enabling intervention metrics that balance controllability, robustness to model mismatch, and expected energy required for stimulation. We introduce a targetability functional that connects inferred directed connectivity to predicted downstream suppression under constrained actuation, and we derive identifiability conditions highlighting when partial coverage still permits reliable directional inference. The result is a clinically aligned yet methodologically general approach for network-driven planning of resection and neuromodulation.