A seizure-detection IC employing machine learning to overcome data-conversion and analog-processing non-idealities

This paper presents a seizure-detection system wherein the accuracy required of the analog frontend is substantially relaxed. Typically, readout of electroencephalogram (EEG) signals would dominate the energy of such a system, due to the precision (noise, linearity) requirements. The presented system performs data conversion and analog multiplication for EEG feature extraction via simple circuits to demonstrate that feature errors can be overcome by appropriate retraining of a classification model, using a machine-learning algorithm. This precludes the need to design a high-precision frontend. The prototype, in 32nm CMOS, results in features whose RMS error normalized to their ideal values is 1.16 (i.e. errors are larger than ideal values). An ideal implementation of the seizure detector exhibits sensitivity, latency, false alarms of 5/5, 2.0 sec., 8, respectively. The feature errors degrade this to 5/5, 3.6 sec., 443, causing high false alarms; but retraining of the classification model restores this to 5/5, 3.4 sec., 4.