Identifying spatio-temporal seizure propagation patterns in epilepsy using Bayesian inference

Abstract

Focal drug resistant epilepsy is a neurological disorder characterized by seizures caused by abnormal activity originating in one or more regions together called as epileptogenic zone. Treatment for such patients involves surgical resection of affected regions. Epileptogenic zone is typically identified using stereotactic EEG recordings from the electrodes implanted into the patient’s brain. Identifying the epileptogenic zone is a challenging problem due to the spatial sparsity of electrode implantation. We propose a probabilistic hierarchical model of seizure propagation patterns, based on a phenomenological model of seizure dynamics called Epileptor. Using Bayesian inference, the Epileptor model is optimized to build patient specific virtual models that best fit to the log power of intracranial recordings. First, accuracy of the model predictions and identifiability of the model are investigated using synthetic data. Then, model predictions are evaluated against a retrospective patient cohort of 25 patients with varying surgical outcomes. In the patients who are seizure free after surgery, model predictions showed good match with the clinical hypothesis. In patients where surgery failed to achieve seizure freedom model predictions showed a strong mismatch. Our results demonstrate that proposed probabilistic model could be a valuable tool to aid the clinicians in identifying the seizure focus.

Virtual epileptic patient brain modeling: Relationships with seizure onset and surgical outcome

“…VEP uses personalized brain network models and machine learning methods by integrating epilepsy patient‐specific anatomical with functional data to provide a computational representation of the hypothesized epileptogenic and propagation zone networks in cerebral space. This method has recently been validated by modeling seizures previously recorded in SEEG 24,29 and on synthetic data 30,31 . The pilot study by Proix et al 29 further showed that VEP can reliably predict seizure propagation, suggesting that in the future this approach could improve SEEG analysis and consequently surgical prognosis.…”

Simulation-Based Inference for Whole-Brain Network Modeling of Epilepsy using Deep Neural Density Estimators

“…Taking advantage of recent advances in probabilistic programming languages (PPLs) for automatic Bayesian inference such as Stan (a state-of-the-art platform for statistical modeling), the feasibility of inverting the coupled Epileptor model that best explains patient's data from whole-brain source activity has been shown previously (Hashemi et al, 2020(Hashemi et al, , 2021. In other studies, efficient and robust inversion of seizure propagation on whole-brain intracranial recordings were achieved by either simplifying the seizure dynamics using a threshold model (Sip et al, 2021) which considerably restricts the range of model dynamics, or by embedding Epileptor equations as priors on brain source dynamics and using maximum a posteriori (MAP) techniques, which does not capture the uncertainty in parameters (Vattikonda et al, 2021). These issues could possibly be addressed by sampling the whole posterior of Epileptor parameters using gradient-based and self-tuning MCMC sampling algorithms such as No U-Turn sampler (NUTS; Hoffman and Gelman (2014)).…”

“…These issues could possibly be addressed by sampling the whole posterior of Epileptor parameters using gradient-based and self-tuning MCMC sampling algorithms such as No U-Turn sampler (NUTS; Hoffman and Gelman (2014)). However, MCMC sampling of the posterior density from sparse observations such as SEEG recordings becomes computationally infeasible, in particular when parameters in a high-dimensional model such as VEP show strong non-linear correlations or if the posterior exhibits pathological geometries such as Neal’s funnel with varying curvature (Betancourt et al, 2014; Vattikonda et al, 2021). Recently, a hybrid of NFs (such as inverse autoregressive flows parameterized by neural networks) and HMC called neural-transport (NeuTra) HMC (Hoffman et al, 2019) has been proposed to correct this sort of unfavorable geometry, while it works in the amortized setting.…”

Linking Brain Structure, Activity, and Cognitive Function through Computation

“…This dual approach guides the identification of causal mechanisms, going beyond the estimation of statistical correlations in traditional data mining approaches. Examples include the Perturbation Complexity Index (PCI) used to assess effective connectivity ( Comolatti et al, 2019 ), variants of dynamic causal modeling used in The Virtual Brain (TVB; see below for examples of clinical applications) and uses of generative models in a “digital twin” approach ( Hashemi et al, 2020 ; Vattikonda et al, 2021 ), which optimizes parameters to best explain personalized data as a prelude to characterizing within and between subject variability.…”

“… TVB, a data-driven neuroinformatics tool, fusing individual brain imaging data with atlas data and state-of-the-art brain modeling, for personalized simulations of brain activity and clinical interventions. Generative brain models operationalize a causal hypothesis, which is evaluated against the patient’s own brain imaging data using variants of dynamical causal modeling such as Monte Carlo simulations ( Hashemi et al, 2020 , 2021 ; Sip et al, 2021 ; Vattikonda et al, 2021 ; https://www.humanbrainproject.eu/en/medicine/the-virtual-brain/ ). …”