Efficient implementation of a ultra-low-power seizure detection algorithm based on vagus nerve signals

Details

Supervisors

Bol, David

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil électricien, à finalité spécialisée

Abstract

The first known records of epilepsy date back to 2000 BCE., making it one of the world’s oldest known diseases. It is a neurological disorder caused by excessive electrical discharges in the brain, leading to recurrent seizures, affecting around 50 million people worldwide. One method of treating it is through vagus nerve stimulation (VNS), which consists of stimulating the vagus nerve at the seizure onset to prevent or alleviate them. The current downside of the methods that use this treatment is the frequency at which current devices operate. They either function at fixed time intervals or are user activated. This leads to unnecessary stimulation of the patient, possibly causing or not stimulating at the right moment and not preventing seizures. To tackle this issue, researchers at the Université Libre de Bruxelles have developed a signal/patient specific seizure detection algorithm based on template matching on VENG signals. Following this, a real-time version of the algorithm has been implemented in C for the Apollo 3 Blue microcontroller. On one hand, this work looked at automating seizure detection through discriminative spectral features of VENG data, however positive results were not obtained. On the other hand, the main bottlenecks of the implementation are identified and different methods for bypassing them are implemented and tested. These methods lead to execution time improvements ranging from 87% to 91% for the most computationally intensive parts of the algorithm, and 71.3% to 86.6% for the whole algorithm. Memory optimization, through bit-width reduction from 32 to 16 bits was implemented to allow to leave space for potential functionality additions. This produced significant up to 44% of memory savings, however at the cost of a 10% decrease in speed and precision which still produces great results. Overall, the different optimizations were a success, allowing for a more efficient implementation of the algorithm, leaving room for future improvements and additions.