An Energy Efficient AdaBoost Cascade Method for Long-Term Seizure Detection in Portable Neurostimulators

Responsive neurostimulation (RNS) is becoming a promising therapy in refractory epilepsy control. In a RNS system, a critical challenge is how to detect seizure onsets accurately with low computational costs. In this study, an energy efficient AdaBoost cascade method for robust long-term seizure detection from local field potential (LFP) signals was proposed and evaluated in a portable neurostimulator. The AdaBoost cascade method included two stages: a seizure candidate detection stage (stage1) and a false alarm rejection stage (stage2). Since seizure activities occurred occasionally in most cases, most normal signal segments can be efficiently classified in stage1. A small percent of suspicious segments were fed into stage2 for more strict examination, where more sophisticated features were extracted to precisely identify seizure activities and reduce false alarms. To further optimize energy efficiency for hardware implementation, we proposed a soft-cascade algorithm for stage2 to reduce the computational costs. Our method was implemented in a generalized neurostimulator and evaluated online with four rats with chronic temporal lobe epilepsy (TLE). A total of 2280.1 hours of LFP signals were recorded and analyzed. Our approach achieves a detection rate of 91.6%, 3.85/h false alarm rate, and 2.31 second detection delay. With the two-stage cascade approach, 98.13% of computational costs could be reduced, with respect to the time of calculation of all features. Our method can detect seizure onsets precisely with high energy efficiency, which is suitable for hardware implementation in portable neuro-stimulators. Therefore, this proposed approach is promising to provide effective and robust performances in long-term seizure detection in neurostimulators for responsive seizure control.