Causal-Consistent Budgeted State-Space Identification for Cross-Modal Seizure Forecasting and Closed-Loop Intervention

Abstract

Epileptic seizures reflect transient transitions in neural population dynamics that can unfold over seconds to minutes, yet practical monitoring and therapy selection must operate under partial observability, sensor heterogeneity, and strict computational limits. Wearable scalp EEG offers long-horizon coverage but coarse spatial resolution and frequent artifacts, while intracranial recordings provide localized access to propagating activity but only for short clinical windows and with patient-specific sampling geometries. These constraints motivate algorithms that are not only accurate, but also resource-aware, uncertainty-calibrated, and consistent across sensing modalities. This paper proposes a unified technical framework for seizure forecasting and intervention planning based on causal-consistent, budgeted identification of latent state-space models with directed interaction structure. The core contribution is a learning and inference principle that enforces invariances across modalities and recording sessions via counterfactual consistency regularization, while simultaneously optimizing a constrained objective that accounts for energy, latency, and memory budgets. The resulting model yields a belief state over seizure regimes and a sparse, time-varying directed graph that supports both low-cost wearable inference and high-resolution intracranial interpretation. We further introduce a safety-aware action selection mechanism for closed-loop neurostimulation that operates in belief space and incorporates conservative uncertainty penalties to reduce the risk of inappropriate intervention. The framework is designed to support principled personalization, robust operation under artifacts and missing channels, and explicit trade-offs between predictive performance and computational footprint.