Sub-microwatt KNN Classifier for Implantable Closed-loop Epileptic Neuromodulation System

The implantable closed-loop system for epileptic seizure detection and neuromodulation is getting more attention in recent years. The architecture design for seizure signal sensing and analyzing has been proposed, but the implementation of the classifier for unsupervised seizure detection is still strongly desired. The k-nearest neighbor (KNN) classification algorithm is one commonly used classifiers in previous researches, yet it needs the training data from both non-seizure and seizure EEG/ECoG states, which are difficult to be collected. Also, the large size of the training set and the concept of the exhaustive search for nearest neighbors make the classification procedure power-consuming. In this paper, we propose a sub-microwatt KNN classifier which only requires the non-seizure EEG/ECoG for training. The size of the training set memory as well as the leakage power is saved by 50%. The processing dynamic power is further reduced by 93.9% due to the early termination scheme. This work achieves the sensitivity of 98.04% and the false alarm rate of 1.97% with optimized power consumption at sub-microwatt, and is suitable for the implantable devices.