Uncertainty-Aware Latent Neural-Field Modeling on Patient-Specific Graphs for Seizure Dynamics Estimation and Interventional Planning

Abstract

Stereo- and scalp-electroencephalography provide multichannel time series that partially observe complex, spatially distributed neural dynamics implicated in focal seizure generation and spread. Clinical decision-making increasingly seeks computational support that can handle sparse electrode coverage, nonstationary dynamics, and heterogeneous patient anatomy, while remaining robust to artifacts and modeling mismatch. This paper develops an uncertainty-aware framework that treats seizures as emergent transitions of a latent neural field evolving on a patient-specific graph, where nodes represent localized cortical or deep structures and edges encode directed, time-varying effective influence. We propose a stochastic latent neural-field model that couples continuous-time dynamics with discrete event modulation to capture pre-ictal drift, ictal recruitment, and post-ictal recovery within a single probabilistic program. Inference is performed via structured variational methods that yield calibrated posterior distributions over connectivity, latent states, and regime parameters, enabling principled uncertainty quantification for clinical interpretation. Building on these posteriors, we formulate intervention planning as a constrained optimal control problem over feasible ablation or stimulation operators, incorporating safety-aware penalties and sparsity constraints to encourage minimally disruptive strategies. The resulting algorithm produces individualized intervention scores, target sets, and confidence measures without assuming a single mechanistic cause of seizures. Experiments on retrospective intracranial and scalp recordings, complemented by anatomically informed simulations, demonstrate improved stability of inferred networks under subsampling and noise, as well as consistent ranking of intervention candidates across initialization and hyperparameter sweeps. The framework is designed to be modular, supporting different sensing modalities and operational constraints, and emphasizes transparent uncertainty reporting alongside predictive performance.