Abstracts

Rationale: Online seizure detection (SD) is important to bring closed-loop devices (e.g. neurostimulation) to clinical practice and is still an open topic of research. Current approaches compute univariate or bivariate statistics from a subset of electrodes spatially distributed over the cortex, and then set thresholds on these statistics to detect seizure onsets. These approaches often perform at chance level and have high false alarm rates because univariate and bivariate statistics fail to capture network dynamics of the brain and/or the thresholds are not optimized. Results would improve if multivariate statistics are exploited and performance measures are explicitly optimized.Methods: We propose a framework for SD that involves (i) constructing a multivariate statistic (?) that captures the connectivity strength of the brain network from iEEGs and distinguishes between non-ictal and ictal states; (ii) modeling the evolution of ? in each state and the state transition probabilities; and, (iii) developing an optimal model-based strategy to detect state transitions from sequential measurements of ?. This strategy is optimal as it minimizes a cost function of the distance between detected and actual onset times, thus decreasing false positives while increasing true positives (with minimal delays). We construct a two-state (non-ictal, ictal) hidden Markov model (HMM) from multisite iEEGs for each subject. The HMM describes the probability of transitioning to an ictal state as a function of the current non-ictal state, and current and past measurements of ?, thus taking into account temporal dependencies that exist in the data. Then, we formulate SD as a Quickest Detection (QD) of hidden state transition. The QD strategy uses sequential measurements of ? to compute an estimate (T_S) of the unequivocal onset time T* (time of transition from non-ictal to ictal state) such that ?T_S?T*? is minimized. QD is solved via optimal control and automatically adapts to each newly acquired measurement based on model predictions and cost function. Experiment: two male Sprague Dawley rats were treated with pentylenetetrazol (PTZ) and generated seizures with activation of the anterior thalamus (AN). Electrodes were located in AN, posterior thalamus, hippocampus, and temporal cortex. Data includes 11 clinical seizures from 7 recording sessions (session duration: 33.0 5.3 min, mean s.e.m.).Results: Our framework detects seizure onsets tens of seconds before the hand-annotated clinical onset (lag: 86.3 19.5 s, mean s.e.m.), which is suitable for stimulation-based seizure suppression therapies. Performance improves (t-test, p<0.05) over non-optimized Bayesian and non-Bayesian approaches (chance level, heuristic threshold-based predictor, CUSUM test).Conclusions: A novel framework is introduced for online SD. Tested on PTZ rat models of epileptic seizure, this framework differs from the state-of-the-art as it is optimal in a precise mathematical sense (it minimizes false alarms and detection delay) and combines information from the whole cortical network by exploiting multivariate statistics computed on multisite iEEGs.