33.1 A High-Accuracy and Energy-Efficient Zero-Shot-Retraining Seizure-Detection Processor with Hybrid-Feature-Driven Adaptive Processing and Learning-Based Adaptive Channel Selection

Seizure-detection processors using machine learning have been proposed to detect the seizure onset of patients for alert or stimulation purposes [1–4]. Existing designs can achieve high accuracy when large amounts of seizure data from a patient is available for the training. However, unlike the collection of non-seizure data, the collection of seizure data with low occurrence requires patients to undergo time-consuming and costly hospitalization, which is difficult in practice. To address this issue, [5] proposed a zero-shot-retraining seizure-detection processor achieving relatively high accuracy without seizure data from the patient for retraining (the zero-shot here means zero seizure data [5]). Instead, only 2-minute non-seizure data from the patient is required to calibrate the clustered features extracted with a neural network (NN) pre-trained on the public seizure dataset. Although this addresses the aforementioned issue, the accuracy (sensitivity 90.3% & specificity 93.6%) of this design is still limited for practical use, and the energy consumption is large for wearable EEG monitoring devices like other seizure-detection processors using NN, as shown in Fig. 33.1.1. In this work, we propose a zero-shot-retraining seizure-detection processor requiring no seizure data from the patient for retraining as in [5] but with much higher accuracy and energy efficiency. It has two major features: 1) a hybrid-feature-driven adaptive processing architecture with on-chip learning requiring no seizure data from the patient to achieve ultra-low energy consumption and high accuracy, and 2) a learning-based adaptive channel-selection technique to further reduce the energy consumption while maintaining high accuracy.