A 16-Channel GRU based On-Chip Seizure Classifier for Closed Loop Neuromodulation

The advancement of implantable on-chip machine learning (ML) classifiers enables the treatment of neurological disorders through closed-loop responsive neurostimulation. However, the hardware realization of the ML algorithm demands area and power efficiency along with low processing latency. This paper explores the time-series nature of the ECoG signal and introduces a scalable machine-learning model based on the Unidirectional Gated Recurrent Unit (GRU) for onset seizure detection. A 16-channel time-division multiplexed (TDM) GRU classifier is implemented on 65nm CMOS technology. The proposed SoC uses only 3 spectral features of different physiological frequency bands which were extracted using an FIR filter. Evaluated on the electroencephalography (EEG) data from the CHB-MIT scalp EEG database, the proposed implementation achieved a sensitivity of 93.1%, specificity of 89.1%, f1-score of 93.8% and an energy efficiency of 2 µJ/classification at 1.2 V supply voltage.