Power and Area Optimized Deep Learning Framework for Accurate Automatic Epileptic Seizure Detection

Abstract

Objectives: To design Hardware based system for detecting epileptic seizures automatically by processing the EEG signals, which could help the patients in protecting from adverse effects. Methods: In this paper, an efficient system for detecting the occurrence of epileptic seizures based on Novel features SCAR (Sum to Cumulative Average Ratio) and RHER (Rescaled Hurst Exponent Range). These features are then used to train a Multi-Layer Perceptron (MLP) network which employs 6 layers with 8 nodes in each layers. The MLP network is then optimized for power and area using Post Trained Quantization (PQT) technique which reduces the memory footprint of the network. Clock Gating and Multiple Operating Frequencies are used for optimizing Power. Findings: The classification accuracy in case of seizure detection is 100% for the MLP network. The system is implemented on a Artix7 FPGA for performing real time analysis. The total power consumption of the system is 35.628mW while total number of 1744 LUTS were required for the system. The overall utilization is just 12% of the available resources. The achieved results outperform all the existing works for the case of seizure detection using EEG signal processing. Improvement: A Low power, area optimized and highly accurate system is proposed in the proposed work. The power and Device utilization analysis showed the utility of the proposed system for practical applications.