Abstracts

Rationale: Seizure onset detection at the focus of the ictal event is hypothesized to provide enough time for corrective action such as neurostimulation to suppress the event before it spreads to the cortical regions. Seizure detection is aided by the use of linear and non-linear metrics extracted from depth electrode recordings obtained using implanted microelectrodes. In order to design a miniature implantable epilepsy prosthesis, it is important to consider hardware feasibility while designing real-time seizure detection algorithms. In this study, we develop a novel hardware feasibility analysis to compare commonly used metrics reported to efficiently detect seizures and trade off algorithmic efficiency with hardware feasibility in an implantable application. Methods: Kainate treated rats are used to model human temporal lobe epilepsy. The animals are implanted with 2-channel depth electrodes in the hippocampus and administered Kainic acid doses periodically until they reach a convulsive state of status epilepticus. Local field potential data is recorded that is bandpass filtered from 10-500Hz at a rate of 1526 samples/s. Seizures are marked out by a combination of visual and electrographic inspection by our collaborating team of neurologists at Indiana University school of medicine. The seizure data is then analyzed in software and the metrics under study are evaluated for efficacy in identifying the onset. Hardware equivalents required to compute the same metrics are implemented in Verilog code, and the circuits are synthesized with a given technology library to get estimates of chip area, power, delay and other hardware parameters. Results: Metrics compared in the study include energy, variance, line length, power spectral density, wavelet based onset detection metric and a novel event based detection scheme developed by our group. Preliminary results indicate that the event based detection scheme can be implemented in hardware with power consumption levels as low as 350nW, having significant implications for battery life. A wavelet based filtering scheme is also presented that significantly improves selectivity of the compared metrics. A digital CMOS implementation of the wavelet filtering mechanism is discussed that utilizes ultra-low power FIR filters optimized for low-frequency applications such as this one. Conclusions: The presented hardware-software co-design and analysis techniques would have a significant clinical impact if seizure detection algorithms are to be translated into responsive neurostimulation devices to treat epilepsy. The study also aims to trade-off mathematical complexity with hardware feasibility to arrive at an optimal operating solution. It is planned to utilize the results presented in this study to develop a multi-modal detection scheme employing more than one metric with shared hardware resources to still be feasible in a miniaturized implantable prosthesis.