AUTHOR=Manzouri Farrokh , Zöllin Marc , Schillinger Simon , Dümpelmann Matthias , Mikut Ralf , Woias Peter , Comella Laura Maria , Schulze-Bonhage Andreas 

TITLE=A Comparison of Energy-Efficient Seizure Detectors for Implantable Neurostimulation Devices

JOURNAL=Frontiers in Neurology

VOLUME=Volume 12 - 2021

YEAR=2022

URL=https://www.frontiersin.org/journals/neurology/articles/10.3389/fneur.2021.703797

DOI=10.3389/fneur.2021.703797

ISSN=1664-2295

ABSTRACT=Introduction:
About 30 percent of epilepsy patients are resistant to treatment with antiepileptic drugs and only a minority of them are surgical candidates. A more recent therapeutic approach is the application of electrical stimulation in the early phases of a seizure to interrupt its spread across the brain. To accomplished this energy-efficient seizure detectors are required, which are able to detect a seizure in its early stages. 
Methods:
Three patient-specific energy-efficient seizure detectors are proposed in this work: (a) Random Forest (RF); (b) long short-term-memory (LSTM)-based Recurrent Neural Network (RNN); and (c) Convolutional Neural Network (CNN). Performance was investigated on EEG data (n=50 patients) from a selected set of surface EEG electrodes, which mimic the electrode layout of an implantable neurostimulation system. As input for the RF, 16 features in time and frequency domain were selected. For CNN as well as for RNN, the raw EEG data was used. Energy consumption was estimated by a platform independent model based on the number of arithmetic operations and memory accesses. To validate the estimation of the energy consumption, the RNN classifier was implemented on an ultra-low power microcontroller. 
Results
The RNN seizure detector achieved the best performance with a median area under the precision-recall curve (PR-AUC) score of 0.88 in comparison with those of the CNN (0.76) and RF (0.46). With respect to energy consumption, RF is the most efficient algorithm with a total number of 67k arithmetic operations (AO) and 67k memory accesses (MA) per classification, followed by CNN (488k AOs and 963k MAs) and RNN (772k AOs and 978k MAs) with memory accesses contributing more to the total energy consumption. Measurements from the hardware implementation of the RNN algorithm demonstrated profound correlation between estimations and measurements.
Discussion
Promising results for seizure detection were obtained using only a few channels with a limited spatial distribution. All three proposed seizure detection algorithms are suitable for application in implantable devices. The method for a platform-independent energy estimation prove to be accurate by hardware implementation of the RNN algorithm. The proposed methodology can be applied for designing new models for responsive neurostimulation.