A Patient-Specific Closed-Loop Epilepsy Management SoC With One-Shot Learning and Online Tuning

Abstract

Epilepsy treatment in clinical practices with surface electroencephalogram (EEG) often faces training dataset shortage issue, which is aggravated by seizure pattern variation among patients. To facilitate future optimization of the detection accuracy as new datasets are available, a fully programmable patient-specific closed-loop epilepsy tracking and suppression system-on-chip (SoC) is proposed with the first-in-literature oneshot learning and online tuning to the best of our knowledge. The proposed two-cycle analog front end (2C-AFE) obtains a 9.8-b effective number of bits (ENOB) with 8× capacitive digital-toanalog converter (CAPDAC) area reduction and 4× switching energy saving compared to a conventional 10-b SAR with an identical unit capacitor size. The entire SoC with 16 surface EEG recording channels consumes an ultra-low energy of 0.97 µJ/class and occupies a miniaturized area of 0.13 mm 2 /ch. in 40-nm CMOS, achieving real-time concurrent seizure detection and raw EEG recording. Verified with the CHB-MIT database, the guided time-channel averaging (GTCA) neural processor achieves the vector-based sensitivity, the specificity, and the latency of 97.8%, 99.5%, and <1 s, respectively. The initial one-shot learning and follow-up online tuning function is validated with the EEG recording from a local hospital patient, which demonstrates a 1.8× vector-based sensitivity boost.

A review of automatic detection of epilepsy based on EEG signals

“…Nogay [167] CNN + Alexnet STFT spectogram Acc 100% Bonn Yildiz [168] CNN, Alexnet, resnet-18, googlenet STFT spectogram, scalogram Acc 100% Bonn Sui [169] CNN Normalization, STFT Acc 91.8% Bern Barcelona 1.74 mm 2 SoC implemented in 60 nm CMOS consumes 2.06 µJ/classification [174] . Zhang et al proposed a fully programmable patient-specific closed-loop epilepsy tracking and suppression SoC.…”

NeuralTree: A 256-Channel 0.227-$μ$J/Class Versatile Neural Activity Classification and Closed-Loop Neuromodulation SoC

“…2) Spectral features: Spectral energy (SE) in multiple frequency bands of neural oscillations has been a commonly used biomarker in epilepsy [5], [16], [17], [19], [21], [22], [40], Parkinson's disease [10], [45], and BMIs [14], [46]. As a measure of signal power integrated over time, the SE can be defined in the discrete-time domain, as follows:…”

“…The 256-channel SoC achieves an 8× improvement in channel count, 9.3× in per-channel area, and 4.3× in system energy efficiency over the state-of-the-art. The compressed NeuralTree takes up 2.93kB out of the 3.17kB on-chip memory, enabling more efficient memory utilization than SVM classifiers [5], [17], [19], [20], [22]. For instance, the recent SVM classifier in [22] utilized 70kB of memory to store 256 support vectors for 16 channels.…”

“…The compressed NeuralTree takes up 2.93kB out of the 3.17kB on-chip memory, enabling more efficient memory utilization than SVM classifiers [5], [17], [19], [20], [22]. For instance, the recent SVM classifier in [22] utilized 70kB of memory to store 256 support vectors for 16 channels. The proposed SoC achieves better multichannel scalability than previous DT-based SoCs [16], [18] thanks to the end-to-end TDM implementation.…”

“…Moreover, ML intelligence can improve the accuracy of motor intention decoding in brain-machine interfaces (BMIs) for rehabilitation of motor impairments [13], [14]. Despite recent innovations, existing ML-embedded SoCs [5], [15]- [22] are limited in the following aspects:…”