Early seizure detection for closed loop direct neurostimulation devices in epilepsy

Abstract

Current treatment concepts for epilepsy are based on continuous drug delivery or electrical stimulation to prevent the occurrence of seizures, exposing the brain and body to a mostly unneeded risk of adverse effects. To address the infrequent occurrence and short duration of epileptic seizures, intelligent implantable closed-loop devices are needed which are based on a refined analysis of ongoing brain activity with highly specific and fast detection algorithms to allow for timely, ictal interventions. Since the development and FDA approval of a first closed loop neurostimulation device relying on simple threshold-based approaches, machine learning approaches became widely available, probably outperformed in the near future by deep convolutional neural networks, which already showed to be extremely successful in pattern recognition in images and partly in signal analysis. Handcrafted features or rules defined by experts become replaced by systematic feature selection procedures and systematic hyperparameter search approaches. Training of these classifiers augments the need of large databases with intracranial EEG recordings, which is partly given by existing databases but potentially can be replaced by continuously transferring data from implanted devices and their publication for research purposes. Already in early design states, the final target hardware must be taken into account for algorithm development. Size, power consumption and, as a consequence, limited computational resources given by low power microcontrollers, FPGAs and ASICS limit the complexity of feature computation, classifier complexity, and the numbers and complexity of layers of deep neuronal networks. Novel approaches for early seizure detection will be a key module for new generations of closed-loop devices together with improved low power implant hardware and will provide together with more efficient intervention paradigms new treatment options for patients with difficult to treat epilepsy.