A 16.4nJ/Class Patient-Independent Prototype-Based Spatio-Temporal CNN Processor with Forward-Inference-Based Adaptation for Robust and Low-Latency Seizure Detection

📋 论文概要

本文提出一种患者独立、基于原型的时空CNN处理器，用于癫痫检测，无需患者特定数据即可实现高精度，能量效率为16.4nJ/类，延迟小于120ms。通过可选的基于前向推理的零样本适应，进一步提高了对信号变异和低质量输入的鲁棒性。

💡 主要创新点

• 提出患者独立的原型学习方法，无需患者特定数据即可实现高精度癫痫检测，解决了传统方法需要大量患者数据训练的痛点。
• 引入基于前向推理的零样本适应机制，在推理过程中动态调整模型，提升对信号变异和低质量输入的鲁棒性。
• 设计低功耗时空CNN处理器架构，在40nm工艺下实现16.4nJ/类的超低能耗和<120ms的实时延迟。

🏷 关键词

📄 原文摘要

Abstract We present a patient-independent, prototype-based spatio-temporal CNN processor for seizure detection, achieving high accuracy without patient-specific data, at an energy of 16.4nJ/class and latency of <120ms. An optional forward-inference-based zero-shot adaptation improves robustness to variability and low-quality inputs. The 40nm IC achieves 94.3%/94.9% sensitivity/specificity without patient data, improving to 95.4%/97.9% with adaptation. For millions of patients worldwide with drug-resistant epilepsy, closed-loop neuromodulation offers a promising solution, providing precise, on-demand intervention upon seizure detection [1-7]. The therapeutic efficacy and practical viability of this approach relies critically on a seizure detection processor capable of simultaneously achieving: 1) high accuracy and minimal latency to exploit the short window (within seconds) between electrical onset and disabling symptoms to intervene seizure evolution; 2) high energy

👥 作者与机构

Yang Wang1, Longyang Lin1, Jerald Yoo2, Jiamin Li1

Southern University of Science and Technology, Shenzhen, China, 2Seoul National University, Seoul, Korea