Hardware-Efficient Seizure Detection

Abstract

Hardware-efficient classification is essential for applications such as medical implants, wearables, and IoT devices, with severe energy and resources constraints. Here, we propose a hardware-efficient machine learning algorithm based on gradient boosted decision trees. Specifically, we train our model to minimize the energy cost associated with feature extraction, and reduce the model size by employing a fixed point quantization method. Hardware parameters such as filter order and coefficient resolution are further optimized for seizure detection task to achieve a reasonable trade-off between performance and hardware cost. Testing this model on the intracranial EEG (iEEG) recordings from 10 patients with epilepsy, we are able to reduce the energy cost by 68.4% compared to the base model, and quantize the tree parameters with 3b (for leaf weights) and 10b (for thresholds), while maintaining the classification performance.