发表机构
Universidade Federal de São João del-Rei; Federal University of Ouro Preto（圣若昂德尔雷伊联邦大学; 欧鲁普雷图联邦大学）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
针对设备端ECG个性化，提出仅原型头自适应方法，在PSoC 6上实现少样本学习，无需反向传播，显著提升心律失常检测的宏F1，并降低注释负担。
AI 中文摘要
可穿戴和床旁心电图（ECG）监护仪必须适应患者特定的形态，以保持跨用户的心律失常检测准确性，然而个性化通常在离线进行，无法考虑个体生理、电极放置或记录漂移。通过反向传播进行设备端自适应对于微控制器级医疗设备来说成本高昂，因为它需要优化器状态、通过卷积层的重复反向传播，以及部署时可能不可用的标记心律失常搏动。本文提出仅原型头自适应作为TinyML ECG系统的紧凑个性化原语。一个一维卷积神经网络（1-D CNN；1,314个参数，每次搏动72.6k乘加运算）在MIT-BIH心律失常数据库上以患者间协议离线训练，冻结为特征提取器，并导出到PSoC 6微控制器。患者特定自适应随后简化为在32维嵌入空间中计算闭式类均值，无需卷积反向传播，无需迭代优化，每个支持搏动仅需一次前向传播。原型自适应将患者间宏F1从0.635/0.639/0.646提高到1/5/10样本下的0.731/0.771/0.797，在目标小型骨干网络的每个样本数上均优于线性随机梯度下降（SGD）头微调。在PSoC 6 Cortex-M4F上对18个单样本情节的设备端回放，原型头的宏F1与主机匹配（0.798），每次搏动11.39毫秒，闪存5.2 KB，SRAM 22.2 KB。一种受限变体仅从被动缓冲的窦性搏动更新正常类原型，产生一致的+0.05宏F1增益，减少了初始注释负担。
英文摘要
Wearable and bedside electrocardiogram (ECG) monitors must adapt to patient-specific morphology to maintain arrhythmia detection accuracy across users, yet personalization is typically performed offline and cannot account for individual physiology, electrode placement, or recording drift. On-device adaptation by backpropagation is expensive for microcontroller-class medical devices because it requires an optimizer state, repeated backward passes through convolutional layers, and labeled arrhythmic beats that may not be available at deployment time. This letter proposes prototype-only head adaptation as a compact personalization primitive for TinyML ECG systems. A one-dimensional convolutional neural network (1-D CNN; 1,314 parameters and 72.6k multiply-accumulate operations per beat) is trained offline on the MIT-BIH Arrhythmia Database under an inter-patient protocol, frozen as a feature extractor, and exported to a PSoC 6 microcontroller. Patient-specific adaptation then reduces to computing closed-form class means in a 32-dimensional embedding space, requiring no convolutional backward pass, no iterative optimization, and only one forward pass per support beat. Prototype adaptation improves inter-patient macro-F1 from 0.635/0.639/0.646 to 0.731/0.771/0.797 at 1/5/10-shot, outperforming linear stochastic-gradient-descent (SGD) head fine-tuning at every shot count for the target tiny backbone. On-device replay over 18 one-shot episodes on a PSoC 6 Cortex-M4F matches the host macro-F1 for the prototype head (0.798), with 11.39 ms per beat, 5.2 KB flash, and 22.2 KB SRAM. A restricted variant that updates only the normal-class prototype from passively buffered sinus beats yields a consistent +0.05 macro-F1 gain, reducing the annotation burden during initial
