SciCNN: A 0-Shot-Retraining Patient-Independent Epilepsy-Tracking SoC

Abstract

Patient-specific seizure-detection SoCs targeting ambulatory seizure treatment [1-8] achieve outstanding accuracy and low energy consumption for monitoring over an extended period powered by a small battery, energy harvesting or body-coupled powering [9]. However, they must collect each patient's seizure episode EEG to train a classifier before the actual deployment, which requires patients to undergo costly and time-consuming hospitalizations, as there is no guarantee a single hospitalization can capture the event. In contrast, a patient-independent seizure detection can address these issues by training the classifier with pre-existing databases, then directly deploying to new patients (Fig. 32.6.1). Traditional classifiers, such as Logistic Regression (LR), Support Vector Machines (SVM) and Decision Trees (DT) [1-7] are not suitable for patient-independent detection as they have difficulty capturing all the possible seizure patterns across patients without firing too many false alarms; this is due to their computational structures and the nature of inter-patient seizure pattern variation. In contrast, Neural Networks (NN) could perform better by mining the features automatically [8]. However, the inter-patient seizure pattern variation could still be vague to the trained NN. Thus, we present a patient-independent (non-specific) seizure-detection SoC with a Seizure-Cluster-Inception Convolutional Neural Network (SciCNN) to first be trained offline with the pre-existing EEG database for auto feature extraction of the neural patterns, then to be further tuned online for fine calibration after being deployed to an unseen patient.