综述_p2

技术,疾病,验证,临床应用

大致就是这几个分类, 有再加

技术

技术类型\年限	2021	2022	2023	2024	2025	2026
子波形表示		36310683 该研究首次提出了”子波形表示(sub-waveform representation)”这一全新概念，颠覆了传统ECG深度学习中全波形(raw waveform)输入的固有范式。这是一种数据驱动(data-centric approach)的方法论创新，而非简单的模型架构调整，开辟了通过精细化ECG数据粒度来提升深度学习性能的新方向。从根本上解决了全波形冗余特征学习导致模型性能受限的核心问题。
ECG信号预处理与去噪		36559944 首次提出利用短时傅里叶变换(STFT)生成的频谱图(Spectrogram)替代原始ECG信号作为CNN模型的输入，通过频率滤波去除不携带ECG信息的频率成分实现数据降维，开创了”频谱图+频率滤波”替代原始信号进行ECG分类的新范式。在保持99.06%高准确率的同时大幅简化了CNN架构，降低了内存使用和计算功耗，为资源受限场景（如可穿戴设备）的ECG分析提供了可行方案。
PPG技术的应用		35322685 研究是首次系统性地探索使用光电容积描记（PPG）信号进行多类心律失常（6类：窦性心律、室性早搏、房性早搏、室性心动过速、室上性心动过速、心房颤动）. 此前PPG信号仅用于房颤的单病种检测，将其扩展到多类心律失常分类属于突破性新方向，为基于可穿戴设备的大规模心律失常筛查提供了全新范式
CNN(卷积神经网络)及RNN(循环神经网络)	34712316 系统比较了CNN、CNN+LSTM、CNN+LSTM+Attention三种模型的性能差异，涵盖了从纯CNN到混合循环神经网络再到注意力机制的递进式架构探索	36210977(纯CNN) 聚焦于类别不平衡数据集下的心律失常识别这一临床实际痛点，提出端到端的DenseNet深度学习解决方案，具有直接的实际应用价值
混合网络架构与注意力机制	34658342 首次提出并行多网络分类架构，将密集连接卷积网络(DenseNet)、双向长短期记忆网络(BiLSTM)与高效通道注意力网络(ECA-Net)有机结合，形成了具有代表性的混合网络架构与注意力机制融合方案	36182760(CNN-RNN) 针对医疗健康领域ECG数据类别不平衡这一核心难题，创新性地提出了SMOTE+Tomek混合重采样技术与CNN-LSTM混合架构的集成学习方法，将少数类（心律失常类别）的召回率(Recall)提高了20%，有效解决了模型对少数类识别性能低下的临床痛点
Transformer与自注意力机制	34847543 首次将Transformer神经网络和自注意力机制系统性地应用于ECG心跳分类的开创性工作，提出了”自适应心跳分割方法（adaptive heartbeat segmentation）”和”波形嵌入（wave-embedding）”两大新概念, 解决了传统心跳分类方法对心电信号时序依赖关系建模不足的核心问题。	36179512(CNN+Transformer) 首次提出Component-Aware Transformer (CAT)架构，将ECG波形分割为P波、QRS波群、T波等独立心电成分，分别进行向量化（嵌入长度和类型信息），再输入Transformer进行房颤诊断。(1)解决了传统深度学习方法未充分利用ECG各成分（P波、QRS波、T波）形态学信息的局限；(2)克服了小规模数据集导致的过拟合问题，使模型具有良好的泛化能力；(3)证明了单导联CAT可与12导联分析相媲美，为可穿戴单导联设备部署提供了可能。
端到端架构	34261486 该研究首次提出了不依赖信号长度和采样频率差异的端到端AFibNet架构, 解决了深度学习AF检测模型在实际部署中因设备差异（不同采样频率、不同信号长度）导致性能严重下降的核心瓶颈。AFibNet使用单一学习系统即可处理各类设备采集的ECG数据，具有直接的临床转化和可穿戴设备部署价值。	34110986 R峰检测技术 首次将1D卷积神经网络编码器-解码器架构（1D CNN Encoder-Decoder）系统性地应用于低质量Holter ECG信号的R峰分割检测，将R峰检测从传统的阈值/变换方法提升为基于深度学习的端到端语义分割新范式。
多导联ECG分类	34098532 该研究首次提出了不依赖信号长度和采样频率差异的端到端AFibNet架构，解决了深度学习模型仅适用于单一采样频率设备的局限这一关键问题，实现了跨设备泛化的核心突破, 具有直接的临床转化和可穿戴设备部署价值。
传统机器学习与深度学习融合	34301998 将传统统计学特征（累积量）与现代机器学习（KELM）深度融合，概念新颖且具有方法论突破意义。在MLII和UCI数据库上达到100%准确率，在PTBDB上达99.57%，显著优于其他文献。与现有多种最先进方法进行了系统性比较，充分验证了方法的优越性。
多标签分类	34821669 提出了残差卷积网络（Residual CNN）+ 类别级注意力机制（Class-wise Attention）的创新架构，在模型设计层面具有实质性创新, 解决了传统阈值方法无法同时处理类别不平衡和成本敏感性的关键问题。	35684694 首次提出了时序转换模块（Temporal Transition Module）——利用不同核大小的卷积层捕获ECG心搏的短期和长期形态特征，并首创动态成本敏感损失函数（Dynamic Cost-Sensitive Loss），在每批次训练中根据类别分布和模型性能动态调整各类别权重，实现了不平衡数据的均衡深度表征。解决了ECG分类中因数据不平衡导致的少数类过拟合核心难题。传统方法固定权重或静态重采样无法适应训练过程中的分布变化，该论文提出的动态损失函数可根据模型实时性能自适应调整，从根本上缓解了不平衡问题。
可解释性	34225095 该研究首次提出可解释深度学习模型, 将基于神经网络集成树的架构与6个可解释特征模块相结合，专门用于解决ECG心律失常分类中的”黑箱”问题，开创了ECG可解释深度学习的新方向.	34347007 验证了CNN的决策依据与临床心电学知识的一致性，直接回应了深度学习”黑箱”问题。
数据增强、合成与隐私保护技术		34753975 GAN deepfake 生成可自由共享的合成数据. 该研究首次提出”DeepFake ECG”这一革命性概念，利用生成对抗网络(GAN)生成逼真的合成12导联ECG信号，从根本上解决了医疗数据共享中的隐私保护难题。
对抗复杂关联变化的算法		34244504 持续学习与少样本学习在ECG分析中的应用 CLOPS
生物识别与身份认证技术		34891673 首次系统评估不同心脏状况组别（窦性心动过速、ST改变合并室上速、室性早搏、起搏器等）对ECG身份识别准确率的影响，并首次揭示了ECG深度学习识别算法可能带来的患者重新识别（re-identification）隐私风险这一重要问题
数学建模与机器学习预测心律失常		34493665 该研究是首次将机理数学建模（mechanistic mathematical modeling）与机器学习（ML）相结合来预测心律失常易感性的开创性工作，提出了一种全新的”计算建模+ML”融合范式，突破了传统仅依赖QT间期/APD单一指标评估心律失常风险的局限，开创了心律失常易感性定量预测的新方向。

疾病

心脏相关的疾病, 心脏之外的疾病

心脏相关的疾病

小主题\年份	需求来源	2021	2022	2023	2024	2025	2026
特发性室性心律失常起源位置预测			35360041 本研究首次将预测范围扩展到21个精细化位点, 本研究将预测准确率提升至98.24%（21位点分类），超越了所有既往研究和人类专家水平，具有直接的临床转化价值。
PVC(室性早搏)			36517841 该研究首次提出了”在窦性心律期间利用AI识别VPC患者”这一全新概念，突破了传统方法必须检出VPC事件才能诊断的局限。利用常规12导联ECG在正常心律下就能发现VPC患者的心电特征，开创了VPC筛查的全新范式——从”等待VPC出现”到”识别VPC易感状态“的根本性转变。
房颤		36341222 开辟了”可穿戴PPG信号+机器学习→前瞻性AF预警”的全新方向. 阵发性AF常无症状，传统长程监测（如Holter）成本高、依从性差，难以大规模推广。该研究直接回答了”能否利用智能手表/手环等日常穿戴设备的PPG信号在AF发作前提前预警“这一核心临床问题。结果显示：优化后的ML模型（M2）在AF发作前0-4小时内的敏感性81.9%、特异性96.6%、阳性预测值96.4%、阴性预测值83.1%、准确率88.9%，为可穿戴设备实现AF的”提前预警”提供了最高级别的循证依据。	35463788 首次将AI-ECG-AF算法系统应用于非心源性栓塞性缺血性卒中患者的阵发性房颤（PAF）风险分层，开创了”利用AI从窦性心律ECG筛查卒中患者隐匿性AF”的新范式。不同于传统方法仅检测已发生的AF，该研究利用AI挖掘窦性心律中的AF电生理痕迹来预测未来PAF风险，概念创新性突出。隐源性卒中患者中约30%由未被发现的阵发性AF引起，但传统长程监测成本高、依从性差。该研究提供了一种低成本、即时可得的AI筛查工具，可快速识别高危患者（隐源性卒中OR=1.974, p<0.001），指导进一步的长程监测和抗凝治疗决策，直接改善卒中二级预防策略，具有重大临床转化价值。
心功能量化评估	心功能的量化评估缺乏便捷工具	34433816 这为心功能的无创量化评估、高危人群早期识别和心血管健康管理提供了革命性工具。
主动脉瓣狭窄		33748852 AI-ECG用于结构性心脏病（主动脉瓣狭窄）筛查的开创性研究
心肌梗死		34763807 首次提出多导联多级融合策略（数据级、特征级、决策级）检测心肌梗死，且首次实现MI发生时间的三分类识别（急性/近期/陈旧性），这在当时是具有里程碑意义的创新方向。MI发生时间的自动化检测能够指导及时干预，对降低全球心血管死亡率具有重要意义。	36380048 心肌梗死与心肌缺血的AI-ECG检测与定位, 填补了深度学习模型在真实急诊场景中诊断心肌梗死的重大空白，克服了以往模型仅适用于受控环境（如公共数据库）而无法在真实临床环境中应用的局限性。 35474073 首次将变分模态分解（VMD）、相位校正信号平均（PRSA）和基线穿越计数三种先进信号处理方法创新性地组合用于ECG特征提取，应用于碎裂QRS波（fQRS）的客观量化检测，开辟了fQRS自动化特征提取与量化的新方向。传统的fQRS判读依赖视觉二分类分析，存在显著的观察者间和观察者内变异（intra-/inter-observer variability），不同定义标准在临该研究首次提出客观量化方案，从根本上解决了这一长期困扰临床的痛点问题。床实践中造成混乱。
对抗复杂关联变化的算法		34244504 持续学习与少样本学习在ECG分析中的应用 CLOPS
预激综合征		34640411 本研究通过无创12导联ECG实现了高精度旁路定位，规避了侵入性检查的风险，具有重大临床转化价值。(样本量 357)本研究首次提出了基于人工智能的算法（locAP AI——location of Accessory Pathway Artificial Intelligence），利用神经网络基于12导联ECG的δ波极性来识别WPW综合征患者的旁路位置

非心脏相关疾病

疾病\年限	2021	2022	2023	2024	2025	2026
睡眠呼吸暂停		34102402 该研究证明了仅用单导联ECG信号结合AI即可实现高精度OSA检测，为大规模社区OSA筛查和居家远程监测提供了革命性解决方案。
无创血流动力学与连续血压监测	实现循环状态的实时动态评估。消除监测盲区，为重症与围术期患者提供精准容量与药物滴定依据。	34344934 该研究首次将图像分割领域的深度学习架构(U-Net变体)改编用于生理波形填补，实现了从”间歇性血压值估计”到”连续性血压波形重建”的根本性范式突破
非侵入性孕期压力检测	慢性产前压力会导致母体心率与胎儿心率的异常耦合（通过胎儿压力指数FSI量化），增加早产、低出生体重和远期神经发育不良结局的风险。然而，传统方法依赖心理量表（主观性强）或有创羊膜穿刺（风险高），缺乏早期无创检测手段。	34921162 解决了重要临床问题——孕期压力缺乏客观无创检测工具
情感计算		34891347 实时情感计算的性能瓶颈
脓毒症		34602084 该研究开辟了”AI-ECG筛查全身性感染”这一全新方向，证明了全身炎症反应可通过影响心肌细胞电活动在ECG上留下可被AI识别的”电生理指纹”。

验证

对已有研究的验证

年限	2021	2022	2023	2024	2025	2026
模型泛化能力评估&临床实用性验证		36388959 该研究首次大规模（162,622份12导联ECG）系统验证了AI-ECG辅助诊断对心内科医生临床表现的提升效果，提出了AI辅助医生临床诊断的工作范式，开创了AI-ECG”临床实用性验证”这一重要研究方向，而非仅止步于模型性能报告。解决了三大关键问题：(a) 目前缺乏大规模包含多种心律失常和ACS相关疾病的高质量数据集；(b) AI-ECG模型能否真正在真实临床场景中辅助医生提升诊断表现（此前未见报道）；(c) AI辅助对医生准确性和效率的定量影响。结果显示：AI辅助使医生诊断准确率提升13.5%，效率提升69.9%，模型本身对15种心脏异常的诊断AUC达0.961，超越了非参考组心内科医生。

应用

某一项技术的实际应用

疾病\年限	2021	2022	2023	2024	2025	2026
临床转化、成本效益与部署研究		40494963 首次提出了”AI-ECG警报干预降低住院患者死亡率的同时是否具有成本效益”这一核心卫生经济学问题，标志着AI-ECG研究从”有效性验证”向”经济可行性论证”的根本性范式升级。
ICU环境下的智能心电报警分类与预警		34604758 ICU监护仪虚假报警率高达80%以上，导致严重的”报警疲劳”，造成医护人员对危急事件的反应延迟。该研究系统性地解决了这一瓶颈：对心脏停搏、极度过缓、室速(VT)和室颤(VF)的阳性预测值显著优于临床监护仪，且能提前最多4秒发出报警，为ICU患者争取了宝贵的抢救时间。

Artificial Intelligence in Electrocardiography: A Comprehensive Review of Advances

Abstract

The integration of artificial intelligence (AI) with electrocardiography (ECG) has emerged as one of the most transformative developments in cardiovascular medicine over the past five years. This comprehensive review synthesizes findings from over 500 peer-reviewed studies, covering the full spectrum of AI-ECG research. We systematically examine five major domains: (1) deep learning architectures for ECG analysis, including convolutional neural networks (CNNs), transformers, graph neural networks, and emerging state-space models; (2) clinical applications spanning arrhythmia detection, myocardial infarction diagnosis, heart failure screening, valvular heart disease assessment, and cardiomyopathies; (3) AI-ECG for non-cardiac disease screening including hepatic, renal, endocrine, and oncologic conditions; (4) wearable devices, edge computing, and remote monitoring innovations; and (5) model interpretability, privacy-preserving federated learning, and clinical translation pathways. Key findings indicate that AI-ECG models now achieve diagnostic accuracy exceeding 99% for arrhythmia classification on benchmark datasets, area under the curve (AUC) values of 0.85-0.95 for detecting structural heart disease, and meaningful risk stratification capabilities for atrial fibrillation prediction, sudden cardiac death, and all-cause mortality. However, significant challenges remain regarding external validation across diverse populations, model generalizability, clinical workflow integration, and prospective evidence from randomized controlled trials. This review provides a roadmap for future research priorities and highlights the immense potential of AI-ECG to democratize cardiovascular diagnostics globally. 

Keywords: artificial intelligence, deep learning, electrocardiography, arrhythmia detection, wearable devices, clinical translation

1. Introduction

Cardiovascular diseases (CVDs) remain the leading cause of morbidity and mortality worldwide, accounting for approximately 17.9 million deaths annually. The electrocardiogram (ECG), invented over a century ago, remains the most widely used non-invasive diagnostic tool in cardiology, providing critical insights into cardiac electrical activity at low cost and with remarkable accessibility. However, traditional ECG interpretation requires specialized training and is subject to inter-observer variability, diagnostic errors, and limitations in detecting subtle patterns indicative of early or subclinical disease. The convergence of artificial intelligence (AI) with ECG analysis has catalyzed a paradigm shift in cardiovascular diagnostics. Deep learning algorithms, particularly convolutional neural networks (CNNs), recurrent neural networks (RNNs), and more recently transformer architectures, have demonstrated the capacity to extract clinically meaningful features from raw ECG signals that may escape even expert human interpretation. From 2021 to 2026, the field has experienced explosive growth, with research expanding from basic arrhythmia classification to encompass comprehensive cardiovascular risk assessment, multi-disease screening, non-cardiac disease detection, and integration with wearable technologies. This review provides a systematic synthesis of the rapidly evolving AI-ECG landscape. We examine methodological innovations in deep learning architectures, survey clinical applications across the full spectrum of cardiovascular and systemic diseases, analyze the transition from research to clinical deployment, and identify key challenges and future directions. By integrating findings from hundreds of recent studies, we aim to provide researchers, clinicians, and policymakers with a comprehensive understanding of the current state and future trajectory of AI-enhanced electrocardiography.

2. Deep Learning Architectures for ECG Analysis

2.1 Convolutional Neural Networks and Their Variants

Convolutional neural networks (CNNs) have formed the backbone of AI-ECG analysis throughout the review period. Early studies demonstrated the efficacy of 1D-CNNs operating directly on raw ECG time-series data for arrhythmia classification. A landmark study using the MIT-BIH Arrhythmia Database achieved 99.12% accuracy using a CNN model, with hybrid CNN-LSTM architectures reaching 99.3% and attention-augmented variants achieving 99.29% [PMID: 34712316]1. These results established that deep learning could match or exceed the performance of traditional machine learning approaches requiring hand-crafted features. The evolution of CNN architectures for ECG analysis has paralleled developments in computer vision. Residual networks (ResNets) with skip connections enabled training of substantially deeper networks, with ensemble SE_ResNet frameworks achieving state-of-the-art performance in multi-label classification of 27 ECG abnormalities, securing third place in the PhysioNet/Computing in Cardiology Challenge 2020 [PMID: 34098532]2. Densely connected convolutional networks (DenseNet) demonstrated superior gradient flow and feature reuse, achieving 99.80% accuracy on the MIT-BIH dataset and 99.63% on the INCART dataset for four-class arrhythmia classification following AAMI standards [PMID: 36182760]3. A critical innovation has been the development of cost-sensitive learning approaches to address the class imbalance that plagues ECG datasets, where normal sinus rhythm vastly outnumbers pathological findings. Novel loss functions that dynamically adjust class weights based on batch-level distribution and model performance achieved 99.81% accuracy for intra-patient classification while maintaining 96.36% accuracy for the more challenging inter-patient classification scenario [PMID: 35684694]4. This represents a substantial improvement over traditional approaches that often sacrifice minority class performance.

2.2 Transformer Architectures and Self-Attention Mechanisms

The introduction of transformer architectures, originally developed for natural language processing, has represented a paradigm shift in ECG analysis. The self-attention mechanism enables models to capture long-range dependencies in ECG signals—a capability particularly valuable for detecting arrhythmias characterized by intermittent or subtle morphological changes. The component-aware transformer (CAT) represents a notable innovation, explicitly segmenting ECG waveforms into P-wave, QRS-complex, and T-wave components before vectorizing them with length and type information for transformer input. In atrial fibrillation detection using a large-scale dataset, CAT outperformed conventional deep learning techniques on both single-lead and 12-lead configurations, and notably, single-lead CAT performance was comparable to 12-lead analysis using conventional methods [PMID: 36179512]5. Adaptive heartbeat segmentation methods using transformer neural networks with multi-head attention have demonstrated remarkable results. By selectively focusing on time-dependent representations of heartbeats and using 1D convolution layers for wave-feature embedding, these models achieved average F1 scores of 99.83%, surpassing state-of-the-art methods and offering particular promise for real-time wearable device deployment [PMID: 34847543]6.

2.3 Graph Neural Networks for Multi-Lead ECG

Recognizing that 12-lead ECG electrodes have natural spatial relationships better represented by non-Euclidean data structures, researchers have increasingly turned to graph neural networks (GNNs). The Conv-RGNN architecture maps 12-lead ECG signals to graph nodes with spatial connections, using CNNs with position attention mechanisms for temporal feature extraction and residual GNNs for spatial feature extraction. This approach demonstrated exceptional parameter efficiency, inference speed, and model robustness across multiple datasets [PMID: 39241329]7. The integration of mutual information (MI) indices between ECG leads as adjacency matrices for graph convolutional networks (GCNs) has yielded impressive results. A 15-layer GCN-MI structure achieved 99.71% accuracy, 98.45% sensitivity, and 99.85% specificity for heart rhythm classification—substantially outperforming GCNs with identity adjacency matrices (92.68% accuracy) and demonstrating the value of explicitly modeling inter-lead relationships [PMID: 36078423]8.

2.4 Emerging Architectures: Mamba and State-Space Models

The most recent architectural innovation in ECG analysis involves state-space models, particularly the Mamba architecture, which offers linear-time complexity for long sequence modeling. Initial explorations of Mamba for ECG signal processing suggest advantages over transformers for processing very long ECG recordings, with reduced computational requirements that may facilitate deployment on resource-constrained wearable devices [PMID: 41193642]9. While still in early stages of validation, these architectures represent a promising direction for handling the inherent temporal nature of ECG data.

2.5 Self-Supervised Learning and Foundation Models

A transformative trend in the 2024-2026 period has been the emergence of self-supervised learning (SSL) approaches that leverage large amounts of unlabeled ECG data. By pretraining on massive ECG repositories and fine-tuning on specific downstream tasks, these models achieve strong performance with limited labeled data. Contrastive learning frameworks that learn ECG representations by maximizing agreement between different views of the same recording have demonstrated particular promise [PMID: 41446031, 41261169]10,11. The application of SSL to fetal and maternal stress detection from abdominal ECG achieved AUROC values of 0.982 ± 0.002 for detecting chronic stress exposure groups, demonstrating that SSL can extract clinically meaningful features from complex multi-modal physiological data even in noisy real-life environments [PMID: 34921162]12. This approach has been extended to ECG biometrics, where self-supervised models achieve superior performance across diverse enrollment scenarios [PMID: 39250357]13.

3. AI-ECG for Cardiovascular Disease Detection and Management

3.1 Arrhythmia Detection and Classification

Arrhythmia detection remains the most extensively studied application of AI-ECG. The field has evolved from binary classification (arrhythmia vs. normal) to multi-class, multi-label systems capable of identifying dozens of distinct rhythm disorders simultaneously. 

Atrial Fibrillation (AF) Detection: AF detection has received particular attention due to its clinical significance and the potential for screening. The AFibNet architecture, utilizing 1D-CNNs trained on over 11,000 subjects across multiple datasets, achieved 99.80% accuracy for AF detection, with robust performance across devices with different sampling frequencies [PMID: 34261486]14. Deep learning models for AF detection now operate effectively on single-lead ECG from smartwatches [PMID: 36376461]15, photoplethysmography (PPG) signals from wearables [PMID: 38391782]16, and even non-contact sensor technologies [PMID: 40393959]17. Beyond detection, AI-ECG models now predict future AF risk from sinus rhythm ECGs. A convolutional neural network trained to infer 5-year incident AF risk demonstrated comparable discrimination to the CHARGE-AF clinical risk score, with the combination of AI and clinical factors providing complementary information (AUC 0.838 vs. 0.802 for CHARGE-AF alone) [PMID: 34743566]18. Saliency mapping revealed that P-wave morphology was the most influential ECG feature, validating the model’s physiological basis. 

Ventricular Arrhythmias: Detection of life-threatening ventricular arrhythmias including ventricular tachycardia (VT) and ventricular fibrillation (VF) has been addressed through multiple approaches. Explainable deep neural networks using variational autoencoders identified P-wave abnormalities (prolonged PR interval, reduced P-wave height) as key predictors of life-threatening ventricular arrhythmias in dilated cardiomyopathy patients [PMID: 35762524]19. For premature ventricular complex (PVC) detection, AI-enabled ECG acquired during normal sinus rhythm permits rapid identification of individuals with high PVC burden, potentially replacing prolonged monitoring [PMID: 36517841]20. Multi-class arrhythmia classification from PPG signals—a particularly challenging task given PPG’s limited morphological information—achieved 85.0% overall accuracy for six rhythm types using deep convolutional neural networks, with a micro-average AUC of 0.978 [PMID: 35360041]21. This finding has important implications for wrist-worn wearable devices that predominantly rely on PPG and can now detect arrhythmias beyond AF.

3.2 Myocardial Infarction and Coronary Artery Disease

AI-ECG for myocardial infarction (MI) diagnosis has progressed from STEMI detection to comprehensive acute coronary syndrome assessment. Deep learning models trained on over 490,000 ECGs from real-world emergency department patients achieved remarkable discrimination: C-statistics of 0.991 for STEMI and 0.832 for NSTEMI, with a Brier score of 0.001 for STEMI classification [PMID: 36380048]22. Importantly, these models performed well in temporally separated and geographically distinct test sets, demonstrating generalizability. The DeepMI framework introduced multi-lead fusion strategies at data, feature, and decision levels to classify both MI presence and occurrence time (acute, recent, old), achieving AUROCs of 96.7% for normal vs. MI classification and 82.9% for acute MI detection [PMID: 34763807]23. Transfer learning from computer vision networks reduced computational overhead while maintaining performance. Perhaps the most clinically impactful recent development is AI-ECG for identifying occlusion myocardial infarction (OMI)—a paradigm shift from the traditional STEMI/NSTEMI dichotomy. AI models can detect acute coronary occlusion even in the absence of ST-segment elevation, potentially reducing missed diagnoses and inappropriate catheterization laboratory activations [PMID: 41128712]24. Studies have shown that AI-ECG can reduce false-positive cardiac catheterization laboratory activations [PMID: 41158088]25 and improve emergency triage of chest pain patients [PMID: 41568258]26.

3.3 Heart Failure and Ventricular Dysfunction

The ability to detect left ventricular systolic dysfunction (LVSD) from a simple, inexpensive ECG has profound implications for heart failure screening. A landmark prospective study demonstrated that AI-enabled smartwatch ECGs acquired in non-clinical environments could detect LVSD (ejection fraction ≤ 40%) with an AUC of 0.885, using data from 2,454 patients across 46 US states and 11 countries [PMID: 36376461]15. This finding suggests that population-wide heart failure screening could be achieved through consumer devices. Multi-center studies have developed comprehensive DL models for quantifying both left and right ventricular function from ECG data. A five-hospital study achieved AUCs of 0.94 for detecting LVEF ≤ 40%, 0.82 for LVEF 40-50%, and 0.84 for predicting right ventricular dysfunction [PMID: 34656465]27. The mean absolute error for LVEF regression was 5.84%, approaching the test-retest variability of echocardiography itself. Recent work has extended AI-ECG to detect heart failure with preserved ejection fraction (HFpEF) [PMID: 40984990]28, mildly reduced ejection fraction (HFmrEF) [PMID: 39995968]29, and acute heart failure in emergency settings [PMID: 39787045]30. AI-ECG has also been applied to predict response to cardiac resynchronization therapy (CRT) [PMID: 39979759, 41525968]31,32 and to detect peripartum cardiomyopathy [PMID: 35122911, 39223284]33,34.

3.4 Valvular Heart Disease

Valvular heart disease (VHD) remains underdiagnosed, and AI-ECG offers a low-cost screening opportunity. Deep learning algorithms for detecting moderate or severe aortic stenosis (AS), aortic regurgitation (AR), and mitral regurgitation (MR) achieved AU-ROCs of 0.88, 0.77, and 0.83 respectively in a multicenter cohort of over 77,000 patients [PMID: 35926935]35. The model for detecting any of these valvular lesions achieved AUC 0.84 with 78% sensitivity and 73% specificity. Subsequent work has confirmed these findings across diverse populations. AI-ECG for detecting moderate-severe AS using CNNs achieved AUC values of 0.85-0.87 in large cohorts, with false-positive ECGs demonstrating twice the risk of developing future AS, suggesting the model detects pre-morbid ECG signatures [PMID: 33748852]36. The latest generation of models has extended to detecting bicuspid aortic valve disease [PMID: 41716932]37, tricuspid regurgitation [PMID: 41437957]38, and rheumatic heart disease [PMID: 42077386]39, with cost-effectiveness analyses supporting population screening strategies [PMID: 41970469]40.

3.5 Cardiomyopathies and Structural Heart Disease

AI-ECG has demonstrated remarkable capability for detecting various cardiomyopathies. For hypertrophic cardiomyopathy (HCM), deep learning models achieve AUC values exceeding 0.90, with the ability to differentiate HCM from hypertensive heart disease—a clinically important distinction [PMID: 42302651, 40696040]41,42. Detection of cardiac amyloidosis, particularly transthyretin amyloidosis (ATTR), has emerged as a high-impact application given the availability of disease-modifying therapies [PMID: 40679604, 41213043]43,44. For dilated cardiomyopathy, inherently explainable DNNs using variational autoencoders compressed 12-lead ECG into 21 interpretable factors, identifying prolonged PR interval, reduced P-wave height, and decreased QRS-T voltages as significant predictors of life-threatening arrhythmic events [PMID: 35762524]19. This approach bridges the gap between black-box deep learning and clinically interpretable features. AI-ECG for detecting left ventricular hypertrophy has been applied in hypertensive populations [PMID: 36189460, 39745684]45,46 and in systemic lupus erythematosus [PMID: 33787068]47, where machine learning identified T-wave changes and ST-T abnormalities as the most prevalent ECG findings, associated with disease activity and autoantibody profiles.

3.6 Genetic Arrhythmia Syndromes

Congenital long QT syndrome (LQTS) presents a particularly compelling use case for deep learning, as approximately 50% of patients have normal or borderline QTc intervals. A deep learning-based neural network developed from the Hearts in Rhythm Organization Registry demonstrated high diagnostic capacity (AUC 0.93) for LQTS detection and genotype differentiation between LQTS1 and LQTS2 (AUC 0.91), substantially outperforming expert-measured QTc intervals (sensitivity 0.90 vs. 0.36) [PMID: 38446445]48. Explainable AI techniques applied to LQTS detection revealed that the onset of the QRS complex was the most informative region for classification—a feature not previously associated with the disease—highlighting DL’s potential for discovery of novel ECG biomarkers [PMID: 35501785]49. Similar approaches have been applied to Brugada syndrome [PMID: 37944131, 41739842]50,51 and ARVC [PMID: 39943462]52.

4. AI-ECG for Non-Cardiovascular Disease Screening

4.1 Metabolic and Endocrine Disorders

A rapidly expanding application of AI-ECG is the detection of metabolic and endocrine disorders. Deep learning models for detecting overt hyperthyroidism using 12-lead ECG achieved AUC values of 0.926 internally and 0.883 externally, with even single-lead (lead I) configurations maintaining AUCs of 0.882 [PMID: 36713007]53. Given that thyroid function tests are not routinely conducted and hyperthyroidism symptoms are often vague, ECG-based screening could enable earlier diagnosis. For diabetes detection, AI-ECG models using machine learning analysis of heart rate variability have shown promise. The ECG-DiaNet model for type 2 diabetes risk prediction [PMID: 40715481]54 and models for prediabetes identification [PMID: 41214697]55 and metabolic syndrome detection [PMID: 41228975]56 have demonstrated the ability to identify glycemic abnormalities from subtle ECG signatures that may reflect autonomic dysfunction or subclinical myocardial changes.

4.2 Hepatic and Renal Disease

The AI-Cirrhosis-ECG (ACE) score represents a proof-of-concept model for detecting cirrhosis from ECG signals, achieving AUC values of 0.908 with 84.9% sensitivity and 83.2% specificity in distinguishing cirrhotic patients from matched controls [PMID: 35029163]57. Longitudinal ACE scores mirrored liver disease progression and resolution, with significant decreases following liver transplantation. This novel relationship between AI-ECG analysis and hepatic disease has been extended to metabolic dysfunction-associated steatotic liver disease (MASLD) [PMID: 39209186]58 and acute-on-chronic liver failure [PMID: 40496881]59. For renal impairment detection, deep learning models achieved AUC values of 0.858-0.906 for detecting moderate-to-severe renal impairment (eGFR < 45 mL/min/1.73m²) using 12-lead ECG [PMID: 35403974]60. Importantly, patients identified as high-risk by the model who initially had normal renal function demonstrated a significantly higher rate of developing renal impairment during follow-up (17.2% vs. 2.4%), suggesting predictive capability beyond concurrent detection.

4.3 Electrolyte Disorders

AI-ECG for dyskalemia detection has shown particular clinical utility. A retrospective cohort study encompassing over 30,000 emergency department visits demonstrated that AI-enabled ECG could estimate serum potassium with mean absolute errors ≤ 0.365 mmol/L. For detecting moderate-to-severe hypokalemia (≤ 3.0 mmol/L) and hyperkalemia (≥ 6.0 mmol/L), AUC values exceeded 0.85 and 0.95 respectively [PMID: 35046489]61. Notably, the U-shaped relationship between potassium concentration and adverse outcomes was more prominent for ECG-estimated potassium than for laboratory measurements, and patients with normal lab values but abnormal ECG-K⁺ had worse outcomes, suggesting that AI-ECG captures physiologically relevant potassium effects beyond serum concentration.

4.4 Hematologic and Oncologic Applications

AI-ECG for detecting hypoalbuminemia demonstrated an AUC of 0.877, with severely low ECG-albumin predicting 2.45-fold higher risk of all-cause mortality and increased risks of hepatorenal and cardiovascular events [PMID: 35770216]62. This application leverages the relationship between fluid balance, tissue impedance, and ECG voltage. In cardio-oncology, AI-ECG has been applied to detect cancer therapy-related cardiac dysfunction (CTRCD) before it becomes clinically apparent [PMID: 38514629, 41449567]63,64. The ability to identify early, subclinical myocardial injury from routine ECGs could enable preventive cardioprotective interventions during chemotherapy. AI-ECG has also been explored for detecting occult malignancy [PMID: 40713871]65 and osteoporosis [PMID: 41452504]66.

4.5 Other Systemic Conditions

The scope of AI-ECG screening has expanded to encompass sepsis detection (AUC 0.901 for screening sepsis, 0.906 for septic shock) [PMID: 34602084]67, COVID-19-related cardiovascular dysautonomia with reduced heart rate variability [PMID: 35101582]68, pulmonary embolism [PMID: 40206108]69, and cognitive decline associated with cerebral infarcts [PMID: 35512882]70. AI-ECG-derived AF risk scores have been correlated with cognitive decline and cerebral small vessel disease, suggesting that ECG signatures of atrial cardiopathy may predict cerebrovascular events even in the absence of documented AF [PMID: 35512882]70.

5. Wearable Devices, Edge Computing, and Remote Monitoring

5.1 Smartwatch and Wearable ECG

The integration of AI algorithms with consumer wearable devices represents perhaps the most transformative development in democratizing cardiac monitoring. The Apple Heart Study and Huawei Heart Study established that photoplethysmography (PPG)-based smart devices could screen for AF in general populations. A landmark prospective study using AI-enhanced smartwatch ECGs to detect left ventricular dysfunction enrolled 2,454 participants from 46 US states and 11 countries, demonstrating that consumer watch ECGs acquired in non-clinical environments could identify patients with reduced ejection fraction (AUC 0.885) [PMID: 36376461]15. AI models specifically developed for asynchronous multi-lead ECG signals from smartwatches demonstrated that measuring at least 3 leads—ideally more than 4—is necessary for accurate myocardial infarction detection, establishing design parameters for smartwatch-based multi-lead ECG systems [PMID: 34505839]71. Recent work has extended this to automated detection of acute MI using asynchronous ECG signals with residual networks and self-attention mechanisms [PMID: 34505839]71.

5.2 Lightweight Models and Edge Deployment

Deploying AI-ECG models on resource-constrained wearable devices requires substantial model compression. The ANNet architecture, utilizing LSTM-MLP hybrid networks with SMOTE-based data augmentation, achieved 97% classification accuracy while enabling wireless transmission gating that reduced power consumption by approximately 50% [PMID: 34982689]72. A 746 nW ECG processor ASIC based on ternary neural networks, implemented in 65-nm CMOS technology, achieved 99.10% sensitivity and 99.50% specificity for detecting 13 abnormal cardiac rhythms while occupying only 1.08 mm² die area [PMID: 35921346]73. Model quantization techniques have enabled efficient deployment on ARM Cortex-M4 platforms [PMID: 40031438]74. The CLINet architecture, using only 297K parameters, achieved 99.94% accuracy on the MIT-BIH dataset while remaining suitable for embedded deployment in smart wearable devices [PMID: 38306814]75. Optimized power-of-two (OPOT) quantization techniques achieved 98.59% 5-class classification accuracy under 4-bit weight precision with 64% energy saving compared to state-of-the-art designs [PMID: 38587951]76.

5.3 Remote Monitoring and Telehealth

The COVID-19 pandemic accelerated the adoption of remote ECG monitoring solutions. AI-augmented remote monitoring systems now enable continuous arrhythmia surveillance, with cloud-based deep learning platforms processing data from multiple device types [PMID: 34261486, 39828428]14,77. Recent developments include smartphone-based ECG acquisition through camera-based PPG [PMID: 38768982]78 and automated digitization of paper ECG recordings for retrospective deep learning analysis, achieving 99% correlation with ground truth digital signals [PMID: 36471089]79.

5.4 Multi-Modal Signal Fusion

The integration of ECG with other physiological signals has enhanced diagnostic accuracy and expanded monitoring capabilities. ECG-PPG fusion models for blood pressure estimation [PMID: 38687656]80, ECG-EEG fusion for seizure detection (improving AUROC by 6.71% over EEG alone) [PMID: 35263265]81, ECG combined with chest X-rays for aortic dissection detection (AUC 0.918) [PMID: 35887647]82, and ECG-ultrasound fusion for comprehensive cardiac assessment [PMID: 40679604]43 represent the frontier of multi-modal AI diagnostics.

6. Model Interpretability, Privacy, and Clinical Translation

6.1 Explainable AI in ECG Analysis

The “black-box” nature of deep learning models has been a major barrier to clinical adoption. Several approaches have been developed to enhance interpretability of AI-ECG models. Gradient-based visualization methods (GradCAM) highlight ECG segments contributing to model decisions, demonstrating that DL models focus on physiologically relevant regions (P-waves for AF prediction, QRS complexes for myocardial infarction detection, T-waves for electrolyte disorders) [PMID: 34347007]83. The xECGArch architecture explicitly separates short-term (morphological) and long-term (rhythmic) feature analysis using two independent CNNs combined with explainable AI methods [PMID: 38849417]84. Deep Taylor decomposition provided the most trustworthy explanations, suggesting that architectural choices can enhance interpretability without sacrificing performance. SHAP (SHapley Additive exPlanations) analysis has been applied to identify key features driving model predictions across numerous applications, including detecting hyperthyroidism [PMID: 36713007]53, hypertension [PMID: 36189460]45, and personality traits from ECG [PMID: 34837662]85. These interpretability approaches not only build clinical trust but also enable discovery of novel ECG biomarkers.

6.2 Privacy-Preserving Federated Learning

Healthcare data privacy regulations pose significant challenges for multi-institutional AI-ECG research. Federated learning frameworks that train models across distributed data sources without sharing raw data have emerged as a solution. Studies have demonstrated the feasibility of federated ECG analysis for arrhythmia classification [PMID: 41464111]86, with privacy-preserving techniques including differential privacy and secure aggregation [PMID: 39412978]87. Edge-based federated learning implementations for wearable devices [PMID: 41337150]88 represent the frontier of privacy-preserving continuous monitoring.

6.3 Clinical Trials and Real-World Validation

The transition from retrospective development to prospective clinical validation represents a critical phase in AI-ECG translation. A landmark pragmatic randomized trial demonstrated that AI-enabled ECG alerts combined with physician notification reduced all-cause mortality in hospitalized patients identified as high-risk (3.6% vs. 4.3% at 90 days) [PMID: 40494963]89. Economic analysis showed favorable cost-effectiveness, with incremental cost-effectiveness ratios of $59,500 per death averted, supporting the business case for AI-ECG implementation. Multiple prospective studies have validated AI-ECG screening for LV dysfunction [PMID: 40251349]90, AF detection [PMID: 40611485]91, and aortic stenosis [PMID: 33748852]36. However, significant gaps remain in randomized controlled trial evidence for most AI-ECG applications, and several studies have highlighted concerns about model performance degradation across demographic subgroups [PMID: 38126168, 37203728]92,93.

6.4 Clinical Implementation and Workflow Integration

Successful AI-ECG deployment requires seamless integration into clinical workflows. Real-world implementation studies have examined alert fatigue, clinician acceptance, and impact on decision-making. The findings highlight that AI-ECG is most effective when integrated as a clinical decision support tool rather than an autonomous diagnostic system [PMID: 40220757]94. Studies of clinician trust in AI-ECG [PMID: 40692805]95 emphasize the importance of transparent model outputs and appropriate training for end-users.

7. Challenges, Limitations, and Future Directions

7.1 Data Heterogeneity and Model Generalizability

Despite impressive performance in controlled settings, AI-ECG models often degrade when applied to populations different from their training data. Performance disparities have been documented across age groups, sex, race/ethnicity, and comorbid conditions [PMID: 38126168, 42312765]92,96. Multi-center validation studies with diverse populations remain essential before clinical deployment. The development of ECG foundation models trained on massive, diverse datasets—analogous to large language models—may address generalizability concerns through self-supervised pretraining followed by task-specific fine-tuning.

7.2 Class Imbalance and Data Quality

Medical ECG datasets are inherently imbalanced, with normal sinus rhythm vastly overrepresented relative to pathological findings. While techniques like SMOTE, cost-sensitive learning, and synthetic data generation using GANs [PMID: 34753975, 38926539]97,98 have shown promise, they may introduce artifacts or fail to capture the full complexity of rare arrhythmias. Signal quality assessment remains critical, particularly for wearable device data where motion artifacts, poor electrode contact, and environmental noise are prevalent [PMID: 34387554, 41374526]99,100.

7.3 Regulatory and Reimbursement Challenges

The regulatory pathway for AI-ECG devices continues to evolve. Most current AI-ECG applications are classified as clinical decision support tools rather than autonomous diagnostic systems, requiring appropriate regulatory oversight. Reimbursement frameworks for AI-augmented ECG interpretation are still being developed, with economic analyses demonstrating cost-effectiveness for specific use cases [PMID: 40494963, 41970469]40,89 but uncertainty remaining about broader implementation.

7.4 Future Directions

Foundation Models for ECG: The development of large-scale ECG foundation models pretrained on millions of recordings will likely transform the field, enabling transfer learning across diverse clinical applications with limited labeled data [PMID: 41446031, 41568699]10,101. 

Multimodal Integration: The combination of ECG with other diagnostic modalities (imaging, genomics, laboratory data, wearable sensors) through sophisticated fusion architectures will enable comprehensive precision cardiovascular medicine [PMID: 37523102, 41945565]102,103. 

Large Language Models: Initial explorations of LLMs for ECG interpretation [PMID: 39096711, 40345017, 41652123]104-106 suggest potential for automated report generation and clinical decision support, though rigorous validation is needed. 

Real-Time Adaptive Learning: Continual learning frameworks that update model parameters as new data becomes available [PMID: 34244504]107 will enable AI systems that improve over time and adapt to local population characteristics. 

Health Equity: Ensuring that AI-ECG benefits extend across socioeconomic and demographic groups remains a critical priority, requiring intentional data collection strategies and fairness-aware algorithm development [PMID: 37203728]93.

8. Conclusion

The integration of artificial intelligence with electrocardiography has progressed at a remarkable pace from 2021 to 2026, transforming a century-old diagnostic modality into a powerful platform for comprehensive cardiovascular risk assessment and multi-disease screening. Deep learning architectures have evolved from basic CNNs to sophisticated transformers, graph neural networks, self-supervised models, and emerging foundation models capable of extracting clinically meaningful information from ECG signals that far exceeds human interpretation capabilities. The clinical scope of AI-ECG has expanded dramatically, encompassing not only arrhythmia detection—where accuracy consistently exceeds 99% on benchmark datasets—but also diagnosis of structural heart disease, myocardial infarction, heart failure, valvular abnormalities, and cardiomyopathies. Perhaps most remarkably, AI-ECG has demonstrated the ability to detect non-cardiac conditions including metabolic, hepatic, renal, and electrolyte disorders, suggesting that the ECG contains a wealth of systemic physiological information previously unrecognized. The convergence of AI-ECG with wearable technology has created unprecedented opportunities for continuous, ambulatory cardiac monitoring and population-level screening. Smartwatch-based AI-ECG for detecting atrial fibrillation and left ventricular dysfunction has been validated in prospective studies involving thousands of participants across multiple countries. Edge computing innovations have enabled real-time AI inference on resource-constrained devices, while cloud-based platforms support comprehensive multi-lead analysis. Despite these advances, significant challenges remain. Model generalizability across diverse populations, rigorous prospective validation through randomized controlled trials, seamless clinical workflow integration, and equitable access to AI-enhanced diagnostics are essential priorities for the field. The explainability of deep learning models, while improved through techniques like GradCAM, SHAP, and inherently interpretable architectures, remains an area requiring continued development to build clinical trust and enable regulatory approval. Looking forward, the development of ECG foundation models, multi-modal fusion frameworks, and integration with large language models portends a future where AI-ECG serves as a ubiquitous, low-cost screening platform for cardiovascular and systemic diseases. The next five years will be critical in determining whether the remarkable technical achievements of the 2021-2026 period translate into meaningful improvements in patient outcomes and population health worldwide.

心电图中的人工智能：进展的全面综述

摘要

过去五年，人工智能（AI）与心电图（ECG）的融合已成为心血管医学领域最具变革性的进展之一。本综述系统性地整合了500多项同行评审研究的成果，全面覆盖了AI-ECG研究的各个层面。我们从五个主要领域展开深入探讨：（1）用于ECG分析的深度学习架构，包括卷积神经网络（CNN）、Transformer、图神经网络以及新兴的状态空间模型；（2）涵盖心律失常检测、心肌梗死诊断、心力衰竭筛查、瓣膜性心脏病评估和心肌病等在内的临床应用；（3）利用AI-ECG进行非心脏疾病筛查，包括肝脏、肾脏、内分泌及肿瘤相关疾病的检测；（4）可穿戴设备、边缘计算与远程监测技术的创新；（5）模型可解释性、保护隐私的联邦学习以及临床转化路径。关键发现表明，目前AI-ECG模型在基准数据集上的心律失常分类准确率已超过99%，检测结构性心脏病的曲线下面积（AUC）值达到0.85至0.95，并具备对房颤预测、猝死风险及全因死亡率进行有效风险分层的能力。然而，在跨不同人群的外部验证、模型的泛化能力、临床工作流程的整合以及随机对照试验的前瞻性证据等方面，仍存在重大挑战。本综述为未来的研究重点提供了路线图，并凸显了AI-ECG在全球范围内实现心血管诊断普惠化的巨大潜力。

关键词：人工智能，深度学习，心电图，心律失常检测，可穿戴设备，临床转化

1. 介绍

心血管疾病（CVDs）仍是全球发病率和死亡率的首要原因，每年约造成1790万人死亡。一个多世纪前发明的心电图（ECG）至今仍是心脏病学领域应用最广泛的无创诊断工具，它以低廉的成本和极高的可及性，为心脏电活动提供了关键信息。然而，传统心电图解读需要专业培训，且易受观察者间差异、诊断误差以及对细微征象识别能力不足的影响——这些细微征象往往提示早期或亚临床疾病。人工智能（AI）与心电图分析的融合，正推动心血管诊断领域迎来一场范式变革。深度学习算法，尤其是卷积神经网络（CNN）、循环神经网络（RNN），以及近期兴起的Transformer架构，已展现出从原始心电信号中提取具有临床意义特征的能力，而这些特征甚至可能超出经验丰富的专业人员的判断范围。过去几年里，这一领域发展迅猛，研究内容从单纯的心律失常分类扩展到全面的心血管风险评估、多病种筛查、非心脏疾病检测，以及与可穿戴技术的深度融合。本综述系统性地梳理了快速发展的AI-心电图研究格局。我们深入探讨了深度学习架构中的方法学创新，全面回顾了其在各类心血管及全身性疾病中的临床应用，剖析了从科研到临床落地的转化过程，并明确了当前面临的主要挑战及未来发展方向。通过整合数百项最新研究成果，我们旨在为研究人员、临床医生和政策制定者提供对AI赋能心电图技术现状及未来趋势的全面理解。

2. 用于心电图分析的深度学习架构

2.1 卷积神经网络及其变体

卷积神经网络（CNN）在整个回顾期内已成为人工智能心电图分析的骨干。早期研究证明，直接作用于原始心电信号时间序列数据的一维CNN在心律失常分类任务中表现出色。一项具有里程碑意义的研究利用MIT-BIH心律失常数据库，采用CNN模型实现了99.12%的准确率；混合CNN-LSTM架构的准确率更是达到99.3%，而引入注意力机制的变体则达到了99.29%[PMID: 34712316]¹。这些成果表明，深度学习能够媲美甚至超越传统机器学习方法所要求的手工特征提取技术。用于心电图分析的CNN架构演进与计算机视觉领域的进展同步进行。残差网络（ResNet）通过跳跃连接实现了对更深网络的训练，其中集成的SE_ResNet框架在27种心电图异常的多标签分类任务中取得了最先进水平的性能，并在2020年PhysioNet/心脏病学计算挑战赛中荣获第三名[PMID: 34098532]²。密集连接卷积网络（DenseNet）则展现了更优的梯度流动和特征复用能力，在MIT-BIH数据集上实现了99.80%的准确率，在INCART数据集上针对符合AAMI标准的四类心律失常分类达到了99.63%的准确率[PMID: 36182760]³。一项关键性创新是开发了成本敏感学习方法，以应对心电图数据集中普遍存在的类别不平衡问题——即正常窦性心律的数量远远超过病理发现。新型损失函数可根据批次级别分布及模型表现动态调整类别权重，使得患者内分类的准确率达到99.81%，同时在更具挑战性的患者间分类场景中仍保持96.36%的准确率[PMID: 35684694]⁴。这较传统方法有了显著提升，后者往往牺牲少数类别的识别性能。

2.2 变压器架构与自注意力机制

最初为自然语言处理而开发的Transformer架构的引入，标志着心电图分析领域的一次范式转变。自注意力机制使模型能够捕捉心电信号中的长程依赖关系——这一能力对于检测以间歇性或细微形态变化为特征的心律失常尤为宝贵。组件感知Transformer（CAT）是一项引人注目的创新，它将心电图波形明确划分为P波、QRS复合波和T波三个组成部分，随后以向量形式编码，并附带长度和类型信息，以便输入Transformer模型。在利用大规模数据集进行房颤检测时，CAT在单导联和12导联配置下均优于传统深度学习技术；尤其值得注意的是，单导联CAT的表现已与采用传统方法的12导联分析相当[PMID: 36179512]⁵。此外，基于Transformer神经网络并结合多头注意力机制的自适应心跳分割方法也取得了显著成果。这些模型通过选择性地聚焦于心跳的时间依赖性表征，并借助一维卷积层实现波形特征嵌入，平均F1得分高达99.83%，超越了当前最先进的方法，尤其在实时可穿戴设备部署方面展现出巨大潜力[PMID: 34847543]⁶。

2.3 用于多导联心电图的图神经网络

鉴于12导联心电图电极之间的空间关系更适合用非欧几里得数据结构来表示，研究人员越来越多地转向图神经网络（GNN）。Conv-RGNN架构将12导联心电信号映射到具有空间连接的图节点上，采用带有位置注意力机制的卷积神经网络进行时序特征提取，并利用残差图神经网络进行空间特征提取。这一方法在多个数据集上展现出卓越的参数效率、推理速度和模型鲁棒性[PMID: 39241329]7。此外，将心电图各导联间的互信息（MI）指标作为邻接矩阵引入图卷积网络（GCN），也取得了令人瞩目的成果。一个15层的GCN-MI结构在心律分类任务中达到了99.71%的准确率、98.45%的灵敏度和99.85%的特异性——显著优于使用恒等邻接矩阵的GCN（准确率为92.68%），充分证明了显式建模导联间关系的价值[PMID: 36078423]8。

2.4 新兴架构：Mamba与状态空间模型

心电图分析领域最新的架构创新涉及状态空间模型，尤其是Mamba架构。该架构在长序列建模方面具有线性时间复杂度。初步探索表明，Mamba在心电信号处理中相较于Transformer具有优势，尤其适用于超长心电记录的处理，同时显著降低了计算需求，这有望促进其在资源受限的可穿戴设备上的部署[PMID: 41193642]9。尽管这些架构目前仍处于验证的早期阶段，但它们为应对心电数据固有的时序特性提供了一个极具前景的研究方向。

2.5 自监督学习与基础模型

在2024至2026年期间，一种具有变革性意义的趋势是自监督学习（SSL）方法的兴起，这类方法充分利用了海量未标注的心电图数据。通过在大规模心电图数据库上进行预训练，并针对特定下游任务进行微调，这些模型即使在仅有少量标注数据的情况下也能取得优异表现。其中，基于对比学习的框架通过最大化同一段记录不同视图之间的一致性来学习心电图表示，已展现出特别可观的前景[PMID: 41446031, 41261169]¹⁰,¹¹。将自监督学习应用于腹部心电图中的胎儿与母体压力检测，其用于识别慢性压力暴露组的AUROC值达到0.982 ± 0.002，这表明自监督学习即便在充满噪声的真实生活环境中，也能从复杂的多模态生理数据中提取出具有临床意义的特征[PMID: 34921162]¹²。这一方法现已拓展至心电图生物识别领域，自监督模型在各种不同的注册场景中均表现出卓越性能[PMID: 39250357]¹³。

图1建议： 建议插入一张示意图，标题为“ECG深度学习架构演进时间线”，采用横向时间轴形式，展示从传统CNN→CNN-LSTM混合模型→注意力机制与Transformer→图神经网络→Mamba状态空间模型→自监督基础模型的架构演进过程，并在每个时间节点标注代表性研究的性能提升幅度。

这个图不对, 因为压根就没有2026年的文献

3. 用于心血管疾病检测与管理的人工智能心电图

3.1 心律失常检测与分类

心律失常检测仍然是人工智能心电图应用中研究最广泛的领域。该领域已从二元分类（心律失常 vs. 正常）发展到多类别、多标签系统，能够同时识别数十种不同的心律失常类型。

房颤（AF）检测：鉴于其临床意义及潜在的筛查价值，房颤检测受到了特别关注。AFibNet架构采用了一种1D卷积神经网络，该网络基于多个数据集上超过11,000名受试者训练而成，对房颤的检测准确率达到99.80%，且在不同采样频率的设备上均表现出稳健的性能[PMID: 34261486]¹⁴。目前，用于房颤检测的深度学习模型已能有效处理来自智能手表的单导联心电图信号[PMID: 36376461]¹⁵、可穿戴设备的光电容积脉搏波描记法（PPG）信号[PMID: 38391782]¹⁶，甚至包括非接触式传感器技术[PMID: 40393959]¹⁷。除了检测之外，人工智能心电图模型如今还能根据窦性心律心电图预测未来的房颤风险。一项经训练可推断5年房颤发病风险的卷积神经网络，其鉴别能力与CHARGE-AF临床风险评分相当；而人工智能与临床因素相结合则提供了互补信息（AUC分别为0.838和0.802，仅使用CHARGE-AF时为0.802）[PMID: 34743566]¹⁸。显著性映射分析表明，P波形态是影响最大的心电图特征，这验证了该模型的生理学基础。

心室心律失常：针对包括室性心动过速（VT）和室性颤动（VF）在内的危及生命的室性心律失常的检测，已采用多种方法加以解决。利用变分自编码器的可解释深度神经网络发现，P波异常（PR间期延长、P波振幅降低）是扩张型心肌病患者发生危及生命的室性心律失常的重要预测因子[PMID: 35762524]19。在室性早搏（PVC）检测方面，基于正常窦性心律采集的AI辅助心电图能够快速识别出PVC负荷较高的个体，有望取代长时间监测[PMID: 36517841]20。从光电容积脉搏波信号中进行多类心律失常分类——这一任务尤其具有挑战性，因为光电容积脉搏波信号的形态学信息有限——采用深度卷积神经网络后，对六种心律类型的总体准确率达到85.0%，微平均AUC达0.978[PMID: 35360041]21。这一发现对主要依赖光电容积脉搏波信号的手腕可穿戴设备具有重要意义，这些设备如今能够检测出除房颤之外的其他心律失常。

3.2 心肌梗死与冠状动脉疾病

用于心肌梗死（MI）诊断的AI心电图技术已从单纯检测ST段抬高型心肌梗死（STEMI）发展到全面评估急性冠脉综合征。基于真实世界急诊科患者超过49万份心电图训练的深度学习模型展现了卓越的鉴别能力：STEMI的C统计量高达0.991，非ST段抬高型心肌梗死（NSTEMI）的C统计量为0.832，其中STEMI分类的Brier得分低至0.001[PMID: 36380048]²²。值得注意的是，这些模型在时间上分离且地域分布不同的测试集上均表现优异，充分证明了其泛化能力。DeepMI框架引入了数据级、特征级和决策级的多导联融合策略，可同时对心肌梗死的存在及其发生时间（急性、近期、陈旧）进行分类，其中正常与心肌梗死分类的AUROC高达96.7%，急性心肌梗死检测的AUROC达到82.9%[PMID: 34763807]²³。通过从计算机视觉网络迁移学习，模型在保持性能的同时显著降低了计算开销。最近最具临床意义的进展当属AI心电图在识别闭塞性心肌梗死（OMI）方面的应用——这一突破性进展彻底改变了传统STEMI/NSTEMI二分法的诊断模式。即使在没有ST段抬高的情况下，AI模型也能准确检测急性冠状动脉闭塞，从而有望减少漏诊及不必要的导管室激活[PMID: 41128712]²⁴。研究还表明，AI心电图能够降低心脏导管室的假阳性激活率[PMID: 41158088]²⁵，并改善胸痛患者的急诊分诊效果[PMID: 41568258]²⁶。

3.3 心力衰竭与心室功能障碍

从简单且经济的心电图中检测左心室收缩功能不全（LVSD）的能力，对心力衰竭筛查具有深远意义。一项具有里程碑意义的前瞻性研究表明，利用人工智能驱动的智能手表心电图在非临床环境中采集的数据，可检测出左心室收缩功能不全（射血分数≤40%），其受试者工作特征曲线下面积（AUC）高达0.885，该研究纳入了来自美国46个州及11个国家的2,454名患者的数据[PMID: 36376461]15。这一发现表明，通过消费级设备即可实现面向全民的心力衰竭筛查。多中心研究已开发出全面的深度学习模型，能够基于心电图数据定量评估左右心室功能。一项由五家医院参与的研究显示，这些模型在检测左心室射血分数≤40%时的AUC达到0.94，检测左心室射血分数介于40%-50%时的AUC为0.82，预测右心室功能不全的AUC则为0.84[PMID: 34656465]27。左心室射血分数回归的平均绝对误差为5.84%，已接近超声心动图自身测试-复测变异范围。近期的研究还将人工智能心电图技术拓展至检测射血分数保留型心力衰竭（HFpEF）[PMID: 40984990]28、轻度降低型心力衰竭（HFmrEF）[PMID: 39995968]29以及急诊环境中的急性心力衰竭[PMID: 39787045]30。此外，人工智能心电图还被用于预测心脏再同步治疗（CRT）的疗效[PMID: 39979759, 41525968]31,32，以及识别围产期心肌病[PMID: 35122911, 39223284]33,34。

3.4 心脏瓣膜病

瓣膜性心脏病（VHD）仍存在诊断不足的问题，而人工智能心电图（AI-ECG）提供了一种低成本的筛查途径。在一项包含超过7.7万名患者的多中心队列研究中，用于检测中度或重度主动脉狭窄（AS）、主动脉反流（AR）以及二尖瓣反流（MR）的深度学习算法，其受试者工作特征曲线下面积（AUROC）分别达到0.88、0.77和0.83[PMID: 35926935]35。此外，该模型用于检测上述任何一种瓣膜病变的总体AUC为0.84，灵敏度达78%，特异性达73%。后续研究进一步证实了这些发现，并在不同人群中得到了验证。利用卷积神经网络（CNN）进行AI-ECG检测中重度主动脉狭窄的最新模型，在大型队列中实现了0.85至0.87的AUC值；同时，假阳性心电图患者未来发生主动脉狭窄的风险高出两倍，这表明该模型能够识别出疾病前的心电图特征[PMID: 33748852]36。最新一代模型已扩展至检测二叶式主动脉瓣疾病[PMID: 41716932]37、三尖瓣反流[PMID: 41437957]38以及风湿性心脏病[PMID: 42077386]39。成本效益分析支持将此类筛查策略应用于大规模人群[PMID: 41970469]40。

3.5 心肌病与结构性心脏病

人工智能心电图（AI-ECG）在检测多种心肌病方面展现了卓越的能力。对于肥厚型心肌病（HCM），深度学习模型的AUC值可超过0.90，且具备将HCM与高血压性心脏病相区别的能力——这一区分在临床上具有重要意义[PMID: 42302651, 40696040]41,42。此外，心脏淀粉样变性的检测，尤其是转甲状腺素蛋白淀粉样变性（ATTR），已成为一项极具影响力的临床应用，这得益于目前已有针对该疾病的改善疗法[PMID: 40679604, 41213043]43,44。对于扩张型心肌病，基于变分自编码器的可解释深度神经网络将12导联心电图压缩为21个易于解读的因子，其中PR间期延长、P波振幅降低以及QRS-T波电压下降被确认为预测危及生命的节律失常事件的重要指标[PMID: 35762524]19。这一方法成功弥合了黑箱深度学习与临床可解释特征之间的鸿沟。人工智能心电图用于检测左心室肥厚已在高血压人群中得到应用[PMID: 36189460, 39745684]45,46，以及在系统性红斑狼疮患者中也得到了应用[PMID: 33787068]47。在这些研究中，机器学习发现T波改变和ST-T段异常是最常见的心电图表现，且与疾病活动度及自身抗体谱密切相关。

3.6 遗传性心律失常综合征

先天性长QT综合征（LQTS）为深度学习提供了一个极具说服力的应用场景，因为大约50%的患者其QTc间期处于正常或临界值范围。由“心律组织注册库”开发的一种基于深度学习的神经网络，在LQTS检测及LQTS1与LQTS2基因型鉴别方面展现了极高的诊断能力（AUC 0.93；LQTS1与LQTS2鉴别AUC 0.91），显著优于专家测量的QTc间期（敏感性分别为0.90 vs. 0.36）[PMID: 38446445]48。应用于LQTS检测的可解释人工智能技术揭示，QRS波群起始部位是分类过程中最具信息量的区域——这一特征此前并未与该疾病相关联——凸显了深度学习在发现新型心电图生物标志物方面的潜力[PMID: 35501785]49。类似方法也已应用于布鲁加达综合征[PMID: 37944131, 41739842]50,51以及致心律失常性右室心肌病[PMID: 39943462]52。

4. 用于非心血管疾病筛查的AI心电图

4.1 代谢与内分泌疾病

人工智能心电图（AI-ECG）的一项快速扩展的应用领域是代谢与内分泌疾病的检测。利用12导联心电图检测明显甲状腺功能亢进的深度学习模型，在内部验证中的AUC值达到0.926，外部验证中为0.883；即便是单导联（I导联）配置，其AUC值也保持在0.882【PMID：36713007】53。鉴于常规并不开展甲状腺功能检查，且甲状腺功能亢进的症状往往较为模糊，基于心电图的筛查有望实现更早的诊断。在糖尿病检测方面，利用机器学习分析心率变异性的人工智能心电图模型已展现出良好前景。用于2型糖尿病风险预测的ECG-DiaNet模型【PMID：40715481】54，以及用于糖尿病前期识别的模型【PMID：41214697】55和用于代谢综合征检测的模型【PMID：41228975】56，均已证明能够从细微的心电图特征中识别出血糖异常，这些特征可能反映自主神经功能障碍或亚临床心肌变化。

4.2 肝肾疾病

AI肝硬化心电图（ACE）评分代表了一种概念验证模型，可用于从心电图信号中检测肝硬化。该模型在区分肝硬化患者与匹配对照组时，AUC值达到0.908，灵敏度为84.9%，特异性为83.2%[PMID: 35029163]57。纵向ACE评分与肝病的进展和缓解趋势一致，肝移植术后ACE评分显著下降。这一AI心电图分析与肝脏疾病之间全新的关联已被拓展至代谢功能障碍相关脂肪性肝病（MASLD）[PMID: 39209186]58以及慢加急性肝衰竭[PMID: 40496881]59。在肾功能损害检测方面，深度学习模型利用12导联心电图检测中重度肾功能损害（eGFR < 45 mL/min/1.73m²）的AUC值达到0.858至0.906[PMID: 35403974]60。值得注意的是，经模型识别为高风险但初始肾功能正常的患者，在随访期间出现肾功能损害的比率显著更高（17.2% vs. 2.4%），这表明该模型不仅具备同期检测能力，还具有预测未来发生肾功能损害的潜力。

4.3 电解质紊乱

用于检测电解质紊乱的AI心电图已显示出特别的临床价值。一项纳入超过3万例急诊科就诊病例的回顾性队列研究证实，基于人工智能的心电图可对血清钾水平进行估算，其平均绝对误差≤0.365毫摩尔/升。在检测中重度低钾血症（≤3.0毫摩尔/升）和高钾血症（≥6.0毫摩尔/升）时，AUC值分别超过0.85和0.95[PMID: 35046489]61。值得注意的是，与实验室检测结果相比，心电图估算的血钾浓度与不良预后之间的U型关系更为显著；即使实验室指标正常但心电图钾离子浓度异常的患者，其预后也更差。这表明，AI心电图能够捕捉到超越血清钾浓度之外、更具生理学意义的钾离子效应。

4.4 血液学与肿瘤学应用

用于检测低白蛋白血症的AI心电图显示出0.877的AUC值，其中严重低白蛋白的心电图提示全因死亡风险高出2.45倍，并增加肝肾及心血管事件的风险[PMID: 35770216]62。该应用利用了体液平衡、组织阻抗与心电图电压之间的关联。在心脏肿瘤学领域，AI心电图已被用于在临床症状显现之前检测癌症治疗相关性心脏功能障碍（CTRCD）[PMID: 38514629, 41449567]63,64。通过常规心电图早期识别亚临床心肌损伤的能力，有望在化疗期间实施预防性心脏保护干预措施。此外，AI心电图还被用于检测隐匿性恶性肿瘤[PMID: 40713871]65以及骨质疏松症[PMID: 41452504]66。

4.5 其他全身性疾病

人工智能心电图筛查的适用范围已扩展至包括脓毒症检测（筛查脓毒症的AUC为0.901，筛查脓毒性休克的AUC为0.906）[PMID: 34602084]⁶⁷、与COVID-19相关的心血管自主神经功能障碍伴心率变异性降低[PMID: 35101582]⁶⁸、肺栓塞[PMID: 40206108]⁶⁹，以及与脑梗死相关的认知功能下降[PMID: 35512882]⁷⁰。基于人工智能心电图得出的房颤风险评分已与认知功能下降及脑小血管病相关联，这表明心电图上提示心房病变的特征可能在未确诊房颤的情况下亦能预测脑血管事件[PMID: 35512882]⁷⁰。

5. 可穿戴设备、边缘计算与远程监控

5.1 智能手表与可穿戴心电图

人工智能算法与消费级可穿戴设备的融合，或许是推动心脏监测大众化进程中最具变革性的进展。苹果心脏研究和华为心脏研究证实，基于光电容积脉搏波描记法（PPG）的智能设备能够对普通人群进行房颤筛查。一项具有里程碑意义的前瞻性研究利用经人工智能增强的智能手表心电图检测左心室功能障碍，共纳入来自美国46个州及11个国家的2454名参与者，结果表明，在非临床环境中采集的消费级手表心电图能够准确识别射血分数降低的患者（AUC 0.885）[PMID: 36376461]¹⁵。专门针对智能手表异步多导联心电信号开发的人工智能模型证明，为实现精准的心肌梗死检测，至少需测量3个导联——理想情况下应超过4个导联——这为基于智能手表的多导联心电系统确立了设计参数[PMID: 34505839]⁷¹。近期的研究进一步将这一技术拓展至利用残差网络与自注意力机制，实现对急性心肌梗死的自动化检测[PMID: 34505839]⁷¹。

5.2 轻量级模型与边缘部署

在资源受限的可穿戴设备上部署人工智能心电图模型，需要对模型进行大幅压缩。ANNet架构采用LSTM与MLP混合网络，并结合基于SMOTE的数据增强技术，在实现无线传输门控的同时，将功耗降低了约50%[PMID: 34982689]72。一款基于三值神经网络、采用65纳米CMOS工艺实现的746纳瓦ECG处理器ASIC，在仅占用1.08平方毫米裸片面积的情况下，对13种异常心律的检测达到了99.10%的灵敏度和99.50%的特异性[PMID: 35921346]73。模型量化技术使得该模型能够高效地部署于ARM Cortex-M4平台[PMID: 40031438]74。CLINet架构仅使用29.7万个参数，在MIT-BIH数据集上的准确率达到99.94%，同时非常适合嵌入式部署于智能可穿戴设备中[PMID: 38306814]75。优化的二的幂次方（OPOT）量化技术在权重精度为4位的情况下，实现了98.59%的5类分类准确率，与当前最先进的设计相比，能耗节省了64%[PMID: 38587951]76。

5.3 远程监测与远程医疗

新冠疫情加速了远程心电监测解决方案的普及。如今，借助人工智能增强的远程监测系统，可实现持续的心律失常监测，基于云的深度学习平台能够处理来自多种设备类型的数据[PMID: 34261486, 39828428]14,77。近期的进展还包括通过摄像头式光电容积脉搏波描记法（PPG）进行智能手机心电图采集[PMID: 38768982]78，以及对纸质心电图记录进行自动化数字化处理，以开展回顾性深度学习分析，其与真实数字信号的相关性高达99%[PMID: 36471089]79。

5.4 多模态信号融合

心电图与其他生理信号的融合，已提升了诊断准确率并拓展了监测能力。用于血压估算的心电图-光电容积脉搏波融合模型[PMID: 38687656]80、用于癫痫发作检测的心电图-脑电图融合模型（较单独使用脑电图时AUROC提升6.71%）[PMID: 35263265]81、结合胸部X光片用于主动脉夹层检测的心电图-胸片融合模型（AUC达0.918）[PMID: 35887647]82，以及心电图与超声融合以实现全面心脏评估的模型[PMID: 40679604]43，均代表了多模态人工智能诊断的前沿方向。

6. 模型可解释性、隐私与临床转化

6.1 心电图分析中的可解释人工智能

深度学习模型的“黑箱”特性一直是其在临床应用中的一大障碍。为此，研究人员开发了多种方法，以提升人工智能心电图（AI-ECG）模型的可解释性。基于梯度的可视化方法（GradCAM）能够突出显示对模型决策有贡献的心电图片段，证明深度学习模型重点关注的是具有生理学意义的区域——例如，用于房颤预测的P波、用于心肌梗死检测的QRS复合波以及用于电解质紊乱诊断的T波[PMID: 34347007]83。xEKGArch架构通过使用两个独立的卷积神经网络，并结合可解释的人工智能方法，明确地将短期（形态学）特征分析与长期（节律性）特征分析分开[PMID: 38849417]84。深度泰勒分解提供了最可信的解释，表明在不牺牲模型性能的前提下，通过合理选择网络结构可以显著提升模型的可解释性。SHAP（Shapley Additive Explanations）分析已被应用于识别驱动模型预测的关键特征，涉及众多应用场景，包括检测甲状腺功能亢进[PMID: 36713007]53、高血压[PMID: 36189460]45，以及从心电图中推断人格特质[PMID: 34837662]85。这些可解释性方法不仅有助于建立临床信任，还为发现新的心电图生物标志物提供了可能。

6.2 隐私保护联邦学习

医疗数据隐私法规为多机构人工智能心电图研究带来了重大挑战。一种解决方案是采用联邦学习框架，该框架可在不共享原始数据的情况下，跨分布式数据源训练模型。已有研究表明，基于联邦学习的心电图分析在心律失常分类方面具有可行性[PMID: 41464111]86，同时，差分隐私和安全聚合等隐私保护技术也已得到验证[PMID: 39412978]87。针对可穿戴设备的边缘联邦学习实现方案[PMID: 41337150]88，则代表了隐私保护连续监测领域的前沿方向。

目标文献摘要翻译

40031438

Wearable systems for the continuous and real-time monitoring of cardiovascular diseases are becoming widespread and valuable assets in diagnosis and therapy. A promising approach for real-time analysis of the electrocardiographic (ECG) signal and the detection of heart conditions, such as arrhythmia, is represented by the transformer machine learning model. Transformers are powerful models for the classification of time series, although efficient implementation in the wearable domain raises significant design challenges, to combine adequate accuracy and a suitable complexity. In this work, we present a tiny transformer model for the analysis of the ECG signal, requiring only 6k parameters and reaching 98.97% accuracy in the recognition of the 5 most common arrhythmia classes from the MIT-BIH Arrhythmia database, assessed considering 8-bit integer inference as required for efficient execution on low-power microcontroller-based devices. We explored an augmentation-based training approach for improving the robustness against electrode motion artifacts noise, resulting in a worst-case post-deployment performance assessment of 98.36% accuracy. Suitability for wearable monitoring solutions is finally demonstrated through efficient deployment on the parallel ultra-low-power GAP9 processor, where inference execution requires 4.28ms and 0.09mJ.

用于持续、实时监测心血管疾病（CVD）的可穿戴系统，正日益成为诊断和治疗中广泛使用且极具价值的工具。在实时分析心电图（ECG）信号并检测心律失常等心脏疾病方面，Transformer 机器学习模型是一种极具前景的方法。尽管 Transformer 在时间序列分类任务中表现强大，但要在可穿戴设备领域实现高效部署，仍面临着巨大的设计挑战——即如何在保证足够精度的同时，将模型复杂度控制在合适的范围内。

在本研究中，我们提出了一种用于分析 ECG 信号的微型 Transformer 模型。该模型仅需 6000 个参数，在 MIT-BIH 心律失常数据库中识别最常见的 5 种心律失常类别时，准确率高达 98.97%。这一性能评估是基于 8 位整数推理（8-bit integer inference）得出的，这也是在低功耗微控制器设备上实现高效执行的必要条件。

此外，我们还探索了一种基于数据增强的训练方法，以提高模型对电极运动伪影噪声的鲁棒性。经过评估，该模型在最坏情况下的部署后性能仍能保持 98.36% 的准确率。最后，我们通过在并行超低功耗 GAP9 处理器上进行高效部署，验证了该模型在可穿戴监测解决方案中的适用性：单次推理执行仅需 4.28 毫秒，能耗仅为 0.09 毫焦。

Table 1

34712316

The electrocardiogram (ECG) is one of the most widely used diagnostic instruments in medicine and healthcare. Deep learning methods have shown promise in healthcare prediction challenges involving ECG data. This paper aims to apply deep learning techniques on the publicly available dataset to classify arrhythmia. We have used two kinds of the dataset in our research paper. One dataset is the MIT-BIH arrhythmia database, with a sampling frequency of 125 Hz with 1,09,446 ECG beats. The classes included in this first dataset are N, S, V, F, and Q. The second database is PTB Diagnostic ECG Database. The second database has two classes. The techniques used in these two datasets are the CNN model, CNN + LSTM, and CNN + LSTM + Attention Model. 80% of the data is used for the training, and the remaining 20% is used for testing. The result achieved by using these three techniques shows the accuracy of 99.12% for the CNN model, 99.3% for CNN + LSTM, and 99.29% for CNN + LSTM + Attention Model.

心电图（ECG）是医疗和卫生保健领域应用最广泛的诊断工具之一。近年来，深度学习方法在涉及心电图数据的医疗预测任务中展现出了巨大的潜力。本文旨在利用公开数据集，应用深度学习技术对心律失常进行分类。 我们的研究使用了两个数据集。第一个是 MIT-BIH 心律失常数据库，采样频率为 125 Hz，共包含 109,446 个心拍，类别包括 N、S、V、F 和 Q。第二个是 PTB 诊断心电图数据库，包含两个类别。 针对这两个数据集，我们采用了三种模型技术：CNN、CNN+LSTM 以及 CNN+LSTM+Attention 模型。数据按 80% 和 20% 的比例划分为训练集和测试集。结果显示，这三种模型的准确率分别达到了：CNN 模型 99.12%，CNN+LSTM 模型 99.3%，CNN+LSTM+Attention 模型 99.29%。

36179512

Atrial fibrillation (AF) is the most common sustained arrhythmia worldwide and imposes a substantial economic burden on the public healthcare system due to its high morbidity and mortality. Early detection of AF is crucial in providing timely treatment and preventing complications such as stroke and other thromboembolism. For AF diagnosis, the 12-lead electrocardiogram (ECG) has been established as the gold standard. However, it requires the clinical experiences of cardiologists and may be vulnerable to inter-observer variability. Although automated AF diagnostic techniques based on deep neural networks (DNN) have been proposed, most studies were conducted using small-scale datasets, resulting in the over-fitting problem. Furthermore, they have not fully exploited ECG components such as P-wave, QRS-complex, and T-wave contrary to the approach adopted by cardiologists who interpret ECG by considering its components. To overcome these limitations, this study presents the component-aware transformer (CAT), which segments the ECG waveform into each component, vectorizes them with length and types information into one vector, and used it as the input of the transformer. We conducted extensive experiments to evaluate the CAT using a large-scale dataset called Shaoxing Hospital Zhejiang University School of Medicine database (AF: 1,780 cases, non-AF: 8,866 cases). The quantitative evaluations demonstrate that the CAT outperforms the conventional deep learning techniques on both single- and 12-lead ECG signals. Moreover, the CAT trained on single-lead ECG is comparable to that of a 12-lead analysis, while conventional methods degraded significantly in performance. Consequently, the CAT is applicable to various single-channel signals such as airway pressure, photoplethysmogram, and blood pressure.

房颤（AF）是全球最常见的持续性心律失常，因其高发病率和高死亡率，给公共医疗系统带来了沉重的经济负担。及早发现房颤对于提供及时治疗、预防卒中及其他血栓栓塞等并发症至关重要。在房颤诊断方面，12导联心电图（ECG）已被确立为金标准。然而，该诊断高度依赖心脏科医生的临床经验，且容易受到观察者间差异的影响。

尽管已有研究提出基于深度神经网络（DNN）的自动房颤诊断技术，但大多数研究仅使用了小规模数据集，导致模型存在过拟合问题。此外，与心脏科医生在解读心电图时会综合考量各波形成分不同，这些方法未能充分利用P波、QRS波群和T波等心电图关键成分。

为克服上述局限性，本研究提出了一种“成分感知Transformer模型（CAT）”。该模型将心电图波形分割为各个独立成分，将各成分的长度与类型信息向量化后合并为一个向量，并将其作为Transformer模型的输入。我们在一个名为“浙江大学医学院附属绍兴医院数据库”的大规模数据集（包含1,780例房颤病例和8,866例非房颤病例）上进行了广泛的实验以评估CAT模型的性能。

定量评估结果表明，无论是在单导联还是12导联心电图信号上，CAT模型的表现均优于传统的深度学习技术。此外，基于单导联心电图训练的CAT模型，其性能可与12导联分析相媲美，而传统方法在此情况下的性能则出现了显著下降。因此，CAT模型具有广泛的适用性，可应用于气道压力、光电容积脉搏波（PPG）和血压等多种单通道生理信号的分析。

40418608

OBJECTIVE: In this paper we develop and evaluate ECG-SMART-NET for occlusion myocardial infarction (OMI) identification. OMI is a severe form of heart attack characterized by complete blockage of one or more coronary arteries requiring immediate referral for cardiac catheterization to restore blood flow to the heart. Two thirds of OMI cases are difficult to visually identify from a 12-lead electrocardiogram (ECG) and can be potentially fatal if not identified quickly. Previous works on this topic are scarce, and current state-of-the-art evidence suggests both feature-based random forests and convolutional neural networks (CNNs) are promising approaches to improve ECG detection of OMI.

METHODS: While the ResNet architecture has been adapted for use with ECG recordings, it is not ideally suited to capture informative temporal features within each lead and the spatial concordance or discordance across leads. We propose a clinically informed modification of the ResNet-18 architecture. The model first learns temporal features through temporal convolutional layers with 1xk kernels followed by a spatial convolutional layer, after the residual blocks, with 12x1 kernels to learn spatial features.

RESULTS: ECG-SMART-NET was benchmarked against the original ResNet-18 and other state-of-the-art models on a multisite real-word clinical dataset that consists of 10, 393 ECGs from 7, 397 unique patients (rate of OMI = 7.2%). ECG-SMART-NET outperformed other models in the classification of OMI with a test AUC of 0.953 [0.921, 0.978].

CONCLUSION AND SIGNIFICANCE: ECG-SMART-NET can outperform the state-of-the-art random forest for OMI prediction and is better suited for this task than the original ResNet-18 architecture.

研究目的：本文开发并评估了用于识别闭塞性心肌梗死（OMI）的 ECG-SMART-NET 模型。OMI 是一种严重的心脏病发作，特征是一根或多根冠状动脉完全阻塞，需要立即进行心脏导管手术以恢复心肌血流。三分之二的 OMI 病例很难通过肉眼从12导联心电图（ECG）中识别出来，若未能及时发现可能危及生命。目前该领域的研究较少，而现有的前沿证据表明，基于特征的随机森林和卷积神经网络（CNN）都是提升 OMI 心电图检测水平的潜力方法。

研究方法：虽然 ResNet 架构已被应用于心电图记录，但它并不理想，难以捕捉每个导联内有价值的时间特征，以及导联间的空间一致性或不一致性。为此，我们提出了一种基于临床先验知识改进的 ResNet-18 架构。该模型首先通过 1×k 核的时间卷积层学习时间特征，随后在残差块之后，利用 12×1 核的空间卷积层来学习空间特征。

研究结果：我们在一个包含 7,397 名独立患者、共 10,393 份心电图的多中心真实世界临床数据集上，将 ECG-SMART-NET 与原始 ResNet-18 及其他前沿模型进行了基准对比（OMI 发病率为 7.2%）。结果显示，ECG-SMART-NET 在 OMI 分类中表现最佳，测试集 AUC 达到 0.953 [0.921, 0.978]。

结论与意义：ECG-SMART-NET 在 OMI 预测上的表现优于现有的随机森林模型，且比原始 ResNet-18 架构更适用于该任务。

39787045

AIMS: Acute heart failure (AHF) poses significant diagnostic challenges in the emergency room (ER) because of its varied clinical presentation and limitations of traditional diagnostic methods. This study aimed to develop and evaluate a deep learning model using electrocardiogram (ECG) data to enhance AHF identification in the ER.

METHODS AND RESULTS: In this retrospective cohort study, we analysed the ECG data of 19 285 patients who visited ERs of three hospitals between 2016 and 2020; 9119 with available left ventricular ejection fraction and N-terminal prohormone of brain natriuretic peptide level data and who were diagnosed with AHF were included in the study. We extracted morphological and clinical parameters from ECG data to train and validate four machine learning models: baseline linear regression and more advanced models including XGBoost, Light GBM, and CatBoost. The CatBoost algorithm outperformed other models, showing superior area under the receiver operating characteristic and area under the precision-recall curve diagnostic accuracy across both internal (0.89 ± 0.01 and 0.89 ± 0.01) and external (0.90 and 0.89) validation data sets, respectively. The model demonstrated high accuracy, precision, recall, and f1 score, indicating robust performance in AHF identification.

CONCLUSION: The developed machine learning model significantly enhanced AHF detection in the ER using conventional 12-lead ECGs combined with clinical data. These findings suggest that ECGs, a common tool in the ER, can effectively help screen for AHF.

目的：急性心力衰竭（AHF）由于临床表现多样且传统诊断方法存在局限性，在急诊室（ER）的诊断中面临巨大挑战。本研究旨在开发并评估一种基于心电图（ECG）数据的深度学习模型，以提高急诊室对急性心力衰竭的识别能力。

方法与结果：在这项回顾性队列研究中，我们分析了2016年至2020年间在三家医院急诊室就诊的19,285名患者的心电图数据。其中，9,119名患者拥有可用的左心室射血分数（LVEF）和N端脑利钠肽前体（NT-proBNP）水平数据，并被确诊为急性心力衰竭，这些患者被纳入研究。我们从心电图数据中提取了形态学和临床参数，用于训练和验证四种机器学习模型：基线线性回归模型以及更高级的模型，包括XGBoost、Light GBM和CatBoost。CatBoost算法的表现优于其他模型，在内部验证数据集（受试者工作特征曲线下面积和精确率-召回率曲线下面积均为0.89 ± 0.01）和外部验证数据集（分别为0.90和0.89）中均展现出卓越的诊断准确性。该模型表现出高准确率、精确率、召回率和F1分数，表明其在识别急性心力衰竭方面具有稳健的性能。

结论：所开发的机器学习模型利用常规12导联心电图结合临床数据，显著提高了急诊室对急性心力衰竭的检测能力。这些发现表明，心电图作为急诊室的常用工具，可以有效地辅助筛查急性心力衰竭。

35926935

BACKGROUND: Valvular heart disease is an important contributor to cardiovascular morbidity and mortality and remains underdiagnosed. Deep learning analysis of electrocardiography (ECG) may be useful in detecting aortic stenosis (AS), aortic regurgitation (AR), and mitral regurgitation (MR).

OBJECTIVES: This study aimed to develop ECG deep learning algorithms to identify moderate or severe AS, AR, and MR alone and in combination. METHODS: A total of 77,163 patients undergoing ECG within 1 year before echocardiography from 2005-2021 were identified and split into train (n = 43,165), validation (n = 12,950), and test sets (n = 21,048; 7.8% with any of AS, AR, or MR). Model performance was assessed using area under the receiver-operating characteristic (AU-ROC) and precision-recall curves. Outside validation was conducted on an independent data set. Test accuracy was modeled using different disease prevalence levels to simulate screening efficacy using the deep learning model.

RESULTS: The deep learning algorithm model accuracy was as follows: AS (AU-ROC: 0.88), AR (AU-ROC: 0.77), MR (AU-ROC: 0.83), and any of AS, AR, or MR (AU-ROC: 0.84; sensitivity 78%, specificity 73%) with similar accuracy in external validation. In screening program modeling, test characteristics were dependent on underlying prevalence and selected sensitivity levels. At a prevalence of 7.8%, the positive and negative predictive values were 20% and 97.6%, respectively.

CONCLUSIONS: Deep learning analysis of the ECG can accurately detect AS, AR, and MR in this multicenter cohort and may serve as the basis for the development of a valvular heart disease screening program.

背景：瓣膜性心脏病是导致心血管疾病发病和死亡的重要原因，但在临床上仍常被漏诊。对心电图（ECG）进行深度学习分析，可能有助于检测主动脉瓣狭窄（AS）、主动脉瓣关闭不全（AR）和二尖瓣关闭不全（MR）。 目的：本研究旨在开发 ECG 深度学习算法，以单独或联合识别中度或重度的 AS、AR 和 MR。 方法：研究纳入了 2005 年至 2021 年间，在接受超声心动图检查前 1 年内做过 ECG 检查的 77,163 名患者。数据集被划分为训练集（n = 43,165）、验证集（n = 12,950）和测试集（n = 21,048；其中 7.8% 的患者患有 AS、AR 或 MR 中的至少一种）。研究使用受试者工作特征曲线下面积（AU-ROC）和精确率-召回率曲线来评估模型性能，并在一个独立数据集上进行了外部验证。此外，研究还通过设定不同的疾病患病率来模拟测试准确率，以评估该深度学习模型在筛查项目中的 efficacy（效能）。 结果：深度学习算法的模型准确率如下：AS（AU-ROC：0.88）、AR（AU-ROC：0.77）、MR（AU-ROC：0.83），以及识别 AS、AR 或 MR 中任意一种（AU-ROC：0.84；敏感度 78%，特异度 73%），且在外部验证中保持了相似的准确率。在筛查项目模拟中，测试特征取决于潜在的患病率和选定的敏感度水平。在 7.8% 的患病率下，阳性预测值（PPV）和阴性预测值（NPV）分别为 20% 和 97.6%。 结论：在这个多中心队列中，ECG 深度学习分析能够准确检测 AS、AR 和 MR，并可能作为开发瓣膜性心脏病筛查项目的基础。

36713007

AIMS: Although overt hyperthyroidism adversely affects a patient’s prognosis, thyroid function tests (TFTs) are not routinely conducted. Furthermore, vague symptoms of hyperthyroidism often lead to hyperthyroidism being overlooked. An electrocardiogram (ECG) is a commonly used screening test, and the association between thyroid function and ECG is well known. However, it is difficult for clinicians to detect hyperthyroidism through subtle ECG changes. For early detection of hyperthyroidism, we aimed to develop and validate an electrocardiographic biomarker based on a deep learning model (DLM) for detecting hyperthyroidism.

METHODS AND RESULTS: This multicentre retrospective cohort study included patients who underwent ECG and TFTs within 24 h. For model development and internal validation, we obtained 174 331 ECGs from 113 194 patients. We extracted 48 648 ECGs from 33 478 patients from another hospital for external validation. Using 500 Hz raw ECG, we developed a DLM with 12-lead, 6-lead (limb leads, precordial leads), and single-lead (lead I) ECGs to detect overt hyperthyroidism. We calculated the model’s performance on the internal and external validation sets using the area under the receiver operating characteristic curve (AUC). The AUC of the DLM using a 12-lead ECG was 0.926 (0.913-0.94) for internal validation and 0.883(0.855-0.911) for external validation. The AUC of DLMs using six and a single-lead were in the range of 0.889-0.906 for internal validation and 0.847-0.882 for external validation.

CONCLUSION: We developed a DLM using ECG for non-invasive screening of overt hyperthyroidism. We expect this model to contribute to the early diagnosis of diseases and improve patient prognosis.

目的：尽管显性甲状腺功能亢进（甲亢）会对患者预后产生不良影响，但临床上并未常规进行甲状腺功能检查（TFTs）。此外，甲亢症状往往不典型，容易导致漏诊。心电图（ECG）是常用的筛查手段，且甲状腺功能与心电图之间的关联已为学界熟知。然而，临床医生很难仅凭细微的心电图变化来识别甲亢。为了实现甲亢的早期发现，本研究旨在开发并验证一种基于深度学习模型（DLM）的心电图生物标志物，用于检测甲亢。

方法与结果：这项多中心回顾性队列研究纳入了在接受心电图检查和甲状腺功能检查且两次检查间隔在24小时内的患者。在模型开发和内部验证阶段，研究共纳入了来自113,194名患者的174,331份心电图。另外，从另一家医院提取了33,478名患者的48,648份心电图用于外部验证。研究使用500 Hz的原始心电图数据，分别开发了基于12导联、6导联（肢体导联和胸导联）和单导联（I导联）心电图的深度学习模型，以检测显性甲亢。研究通过受试者工作特征曲线下面积（AUC）评估了模型在内部和外部验证集中的性能。结果显示，基于12导联心电图的深度学习模型在内部验证中的AUC为0.926（0.913-0.94），在外部验证中的AUC为0.883（0.855-0.911）。基于6导联和单导联心电图的模型，在内部验证中的AUC介于0.889至0.906之间，在外部验证中的AUC介于0.847至0.882之间。

结论：我们开发了一种利用心电图进行甲亢无创筛查的深度学习模型。期待该模型能为疾病的早期诊断做出贡献，并改善患者预后。

34387554

BACKGROUND: With the development and promotion of wearable devices and their mobile health (mHealth) apps, physiological signals have become a research hotspot. However, noise is complex in signals obtained from daily lives, making it difficult to analyze the signals automatically and resulting in a high false alarm rate. At present, screening out the high-quality segments of the signals from huge-volume data with few labels remains a problem. Signal quality assessment (SQA) is essential and is able to advance the valuable information mining of signals.

OBJECTIVE: The aims of this study were to design an SQA algorithm based on the unsupervised isolation forest model to classify the signal quality into 3 grades: good, acceptable, and unacceptable; validate the algorithm on labeled data sets; and apply the algorithm on real-world data to evaluate its efficacy. METHODS: Data used in this study were collected by a wearable device (SensEcho) from healthy individuals and patients. The observation windows for electrocardiogram (ECG) and respiratory signals were 10 and 30 seconds, respectively. In the experimental procedure, the unlabeled training set was used to train the models. The validation and test sets were labeled according to preset criteria and used to evaluate the classification performance quantitatively. The validation set consisted of 3460 and 2086 windows of ECG and respiratory signals, respectively, whereas the test set was made up of 4686 and 3341 windows of signals, respectively. The algorithm was also compared with self-organizing maps (SOMs) and 4 classic supervised models (logistic regression, random forest, support vector machine, and extreme gradient boosting). One case validation was illustrated to show the application effect. The algorithm was then applied to 1144 cases of ECG signals collected from patients and the detected arrhythmia false alarms were calculated.

RESULTS: The quantitative results showed that the ECG SQA model achieved 94.97% and 95.58% accuracy on the validation and test sets, respectively, whereas the respiratory SQA model achieved 81.06% and 86.20% accuracy on the validation and test sets, respectively. The algorithm was superior to SOM and achieved moderate performance when compared with the supervised models. The example case showed that the algorithm was able to correctly classify the signal quality even when there were complex pathological changes in the signals. The algorithm application results indicated that some specific types of arrhythmia false alarms such as tachycardia, atrial premature beat, and ventricular premature beat could be significantly reduced with the help of the algorithm.

CONCLUSIONS: This study verified the feasibility of applying the anomaly detection unsupervised model to SQA. The application scenarios include reducing the false alarm rate of the device and selecting signal segments that can be used for further research.

Table 2

| AI Model Architecture | Detection Capability (Number of Conditions) | Accuracy | Power Consumption | Deployment Method (Cloud/Edge) | | —————————– | ———————————————– | ———— | ——————— | ———————————- |

41374526

Wearable ECG monitoring devices have become indispensable in personalized healthcare. However, dynamic signal acquisition during daily activities often introduces transient noise, which complicates signal classification and denoising, and may compromise diagnostic reliability. To address this challenge, this study proposes an ECG preprocessing framework based on multi-task learning, in which a fine-grained noise localization task is introduced to guide and assist both ECG signal quality assessment and denoising. Built upon a Transformer backbone and optimized with three task-specific loss functions, the proposed model leveraged weak supervision and pathological ECG data to learn robust noise-invariant representations. This design incorporates intra-class awareness, enabling the model to overcome various noise within the same quality category and to perform adaptive denoising beyond conventional inter-class-based approaches. Extensive experiments demonstrated state-of-the-art performance in both denoising and quality assessment, with weighted average F1-scores ranging from 95.72% to 98.49% and classification accuracy exceeding 95.68%. Moreover, under extremely severe noise conditions, the signal-to-noise ratio (SNR) is improved from -1.95 ± 3.83 dB to 12.20 ± 2.51 dB while preserving waveform fidelity. After pruning and quantization, the model could be effectively compressed, thereby enhancing its suitability for real-time deployment in edge computing scenarios. Overall, the proposed method not only preserved diagnostically important ECG waveforms and provided interpretable noise localization but also offers an efficient and clinically relevant solution for large-scale, real-time ECG monitoring.

可穿戴式心电（ECG）监测设备在个性化医疗中已变得不可或缺。然而，日常活动中的动态信号采集往往会引入瞬态噪声，这不仅增加了信号分类和去噪的难度，还可能影响诊断的可靠性。为了应对这一挑战，本研究提出了一种基于多任务学习的ECG预处理框架，该框架引入了细粒度的噪声定位任务，以引导并辅助ECG信号质量评估与去噪。

该模型以Transformer为骨干网络，并通过三种特定任务的损失函数进行优化，利用弱监督和病理ECG数据来学习鲁棒的噪声不变表征。这种设计融入了类内感知能力，使模型能够克服同一质量类别内的各种噪声，并实现超越传统基于类间方法的自适应去噪。

大量实验表明，该方法在去噪和质量评估方面均达到了最先进的性能，加权平均F1分数在95.72%至98.49%之间，分类准确率超过95.68%。此外，在极度严重的噪声条件下，信噪比（SNR）从 -1.95 ± 3.83 dB 提升至 12.20 ± 2.51 dB，同时保持了波形保真度。经过剪枝和量化后，该模型可以被有效压缩，从而增强了其在边缘计算场景中实时部署的适用性。

总体而言，该方法不仅保留了具有诊断意义的重要ECG波形，提供了可解释的噪声定位，还为大规模、实时ECG监测提供了一种高效且具备临床相关性的解决方案。

| AI Model Architecture | Detection Capability (Number of Conditions) | Accuracy | Power Consumption | Deployment Method (Cloud/Edge) | | :———————————————————————————————– | :—————————————————————————————– | :——————————————————————————- | :———————————————————————————————- | :—————————————————————————————————- | | Transformer-based Multi-Task Learning Framework (Encoder-Decoder with Multi-Head Self-Attention) | 3 Signal Quality Classes (Good, Medium, Bad) + Fine-grained Noise Localization + Denoising | 95.68% - 98.49% (Varies by dataset: BUT: 95.68%, WD: 92.58%, DUT_BMELAB: 98.48%) | Low (Optimized for edge: 0.25M parameters after pruning, INT8 quantized, 3.02ms latency on CPU) | Edge Computing (Explicitly designed and compressed for real-time deployment on wearable/edge devices) |

37234033

BACKGROUND: Silent paroxysmal atrial fibrillation (AF) may be difficult to diagnose, and AF burden is hard to establish. In contrast to conventional diagnostic devices, photoplethysmography (PPG)-driven smartwatches or wristbands allow for long-term continuous heart rhythm assessment. However, most smartwatches lack an integrated PPG-AF algorithm. Adding a standalone PPG-AF algorithm to these wrist devices might open new possibilities for AF screening and burden assessment.

OBJECTIVE: The aim of this study was to assess the accuracy of a well-known standalone PPG-AF detection algorithm added to a popular wristband and smartwatch, with regard to discriminating AF and sinus rhythm, in a group of patients with AF before and after cardioversion (CV).

METHODS: Consecutive consenting patients with AF admitted for CV in a large academic hospital in Amsterdam, the Netherlands, were asked to wear a Biostrap wristband or Fitbit Ionic smartwatch with Fibricheck algorithm add-on surrounding the procedure. A set of 1-min PPG measurements and 12-lead reference electrocardiograms was obtained before and after CV. Rhythm assessment by the PPG device-software combination was compared with the 12-lead electrocardiogram.

RESULTS: A total of 78 patients were included in the Biostrap-Fibricheck cohort (156 measurement sets) and 73 patients in the Fitbit-Fibricheck cohort (143 measurement sets). Of the measurement sets, 19/156 (12%) and 7/143 (5%), respectively, were not classifiable by the PPG algorithm due to bad quality. The diagnostic performance in terms of sensitivity, specificity, positive predictive value, negative predictive value, and accuracy was 98%, 96%, 96%, 99%, 97%, and 97%, 100%, 100%, 97%, and 99%, respectively, at an AF prevalence of ~50%.

CONCLUSIONS: This study demonstrates that the addition of a well-known standalone PPG-AF detection algorithm to a popular PPG smartwatch and wristband without integrated algorithm yields a high accuracy for the detection of AF, with an acceptable unclassifiable rate, in a semicontrolled environment.

背景： 无症状的阵发性心房颤动（AF）可能难以诊断，且房颤负荷（AF burden）也难以确定。与传统的诊断设备相比，基于光电容积脉搏波（PPG）技术的智能手表或手环能够实现长期的连续心律评估。然而，大多数智能手表并未内置 PPG-AF 算法。若能为这些腕部设备添加独立的 PPG-AF 算法，或许能为房颤的筛查和负荷评估开辟新的途径。 目的： 本研究旨在评估一种知名的独立 PPG-AF 检测算法在添加至主流手环和智能手表后，针对房颤患者复律（CV）前后的心律，区分房颤与窦性心律的准确性。 方法： 在荷兰阿姆斯特丹一家大型学术医院，连续招募了因房颤入院接受复律治疗的知情同意患者。要求患者在复律手术前后佩戴 Biostrap 手环或加装了 Fibricheck 算法的 Fitbit Ionic 智能手表。在复律前后，分别获取一组 1 分钟的 PPG 测量数据和 12 导联参考心电图。将 PPG 设备与软件组合的心律评估结果与 12 导联心电图进行对比。 结果： Biostrap-Fibricheck 队列共纳入 78 名患者（156 组测量数据），Fitbit-Fibricheck 队列共纳入 73 名患者（143 组测量数据）。由于信号质量不佳，PPG 算法分别有 19/156（12%）和 7/143（5%）的测量数据无法分类。在房颤患病率约为 50% 的情况下，其诊断性能如下：

Biostrap-Fibricheck 队列： 敏感性 98%，特异性 96%，阳性预测值 96%，阴性预测值 99%，准确率 97%。
Fitbit-Fibricheck 队列： 敏感性 97%，特异性 100%，阳性预测值 100%，阴性预测值 97%，准确率 99%。

结论： 本研究表明，将知名的独立 PPG-AF 检测算法添加至未内置该算法的主流 PPG 智能手表和手环中，在半受控环境下对房颤的检测具有很高的准确性，且无法分类的数据比率在可接受范围内。

AI Model Architecture	Detection Capability (Number of Conditions)	Accuracy	Power Consumption	Deployment Method (Cloud/Edge)
Deep Learning Model (Device-agnostic PPG processing)	3 Conditions: 1. Regular Sinus Rhythm (with up to 5 ectopic beats/min) 2. Non-AF Arrhythmias (ectopic beats, tachycardia, bradycardia) 3. Possible Atrial Fibrillation (AF) (Plus “Unclassifiable” for bad quality)	Fitbit-Fibricheck Cohort: • Sensitivity: 97% • Specificity: 100% • Accuracy: 99% Biostrap-Fibricheck Cohort: • Sensitivity: 98% • Specificity: 96% • Accuracy: 97%	Not Specified (Document focuses on diagnostic accuracy and does not report power consumption metrics)	Hybrid (Cloud-based Analysis): • Biostrap: Data sent to AWS Cloud for analysis. • Fitbit: Algorithm installed on device (clockface), but measurements transferred to AWS Cloud for analysis.

40984996

AIMS: ECG monitoring is often required during critical phases of illness. To evaluate the role of modern technology and advanced analytical algorithms artificial intelligence compared with standard-of care, we undertook a prospective, head-to-head comparison of a novel, cable-free, patch-based, and AI-enhanced electrocardiography system (CardioSenseSystem) with standard of care (SOC) ECG monitoring. Patients who had undergone cardiac surgery at a large university hospital (Skåne University Hospital, Sweden) were simultaneously monitored by both systems, and alarms and monitoring interruptions were recorded. METHODS AND RESULTS: Forty-nine patients were recruited. The CardioSenseSystem system demonstrated significantly higher sensitivity, correctly detecting 364 critical red alarms vs. 12 for SOC (P < 0.0001), and lower rates of high priority false alarms (0.3% vs. 40%; P < 0.0001). Monitoring interruptions were markedly reduced (114 s/day vs. 584 s/day; P < 0.0001). Handling time per patient day was significantly shorter (256 s vs. 880 s). The CardioSenseSystem system also reduced alarm fatigue, with fewer disturbances per patient per hour (0.03 vs. 0.11; P < 0.0001). CONCLUSION: The CardioSenseSystem system delivered significant advantages over conventional ECG monitoring in post-cardiac surgery patients. Its high sensitivity, reduced false alarms, fewer monitoring interruptions, and decreased handling time suggest that it may enhance patient outcomes and clinical efficiency, warranting broader application in acute-care settings.

目的： 在疾病的关键阶段，通常需要进行心电图（ECG）监测。为了评估现代技术和先进的人工智能分析算法的作用，我们对一种新型无导线、贴片式且由AI增强的心电图系统（CardioSenseSystem）与标准护理（SOC）心电图监测进行了前瞻性的头对头比较研究。在一家大型大学医院（瑞典斯科讷大学医院），对接受过心脏手术的患者同时使用这两种系统进行监测，并记录了警报和监测中断情况。

方法与结果： 共招募了49名患者。CardioSenseSystem系统表现出显著更高的灵敏度，正确检测到364次关键红色警报，而标准护理（SOC）仅检测到12次（P < 0.0001）；同时，高优先级假警报的发生率显著降低（0.3% vs. 40%；P < 0.0001）。监测中断时间显著减少（114秒/天 vs. 584秒/天；P < 0.0001）。每个患者每天的处理时间也显著缩短（256秒 vs. 880秒）。此外，CardioSenseSystem系统还减轻了警报疲劳，每位患者每小时的干扰次数更少（0.03次 vs. 0.11次；P < 0.0001）。

结论： 在心脏手术后患者中，CardioSenseSystem系统相较于传统心电图监测具有显著优势。其高灵敏度、低假警报率、更少的监测中断以及更短的处理时间表明，该系统有望改善患者预后和临床效率，值得在急性护理环境中更广泛地应用。

AI Model	Architecture	Detection Capability (Number of Conditions)	Accuracy	Power Consumption	Deployment Method (Cloud/Edge)
CardioSenseSystem (CSS)	Two-stage supervised learning framework: Stage 1 uses a rule-based multi-lead QRS detector to model patient-specific ECG morphology; Stage 2 leverages this model to enhance beat detection and classification in noisy/artifact-prone environments.	28 types of arrhythmias/events (including 5 high-priority red alerts and 23 low-priority yellow alerts)	High-priority alerts: 364 correctly detected, 1 false positive (False Positive Rate: 0.3%). Low-priority alerts: 1685 correctly identified, 194 false positives (Accuracy: 89.7%).	Single-patch device designed for 24-hour continuous operation (extended to 48 hours in later versions) without battery replacement.	Hybrid Deployment: AI algorithms process data on the Server, with results displayed at the Information Centre.

42122301

Cardiovascular diseases are the primary causes of mortality worldwide, often characterized by subtle onset and acute progression. Traditional ECG electrodes may cause skin irritation, limiting routine monitoring and early risk assessment. Relying on the advantages of non-contact monitoring, millimeter-wave radar-based cardiac monitoring combined with deep learning has become a popular research direction recently. To overcome the poor generalization of methods trained from single-source datasets, this study designed seven experimental scenarios covering wakefulness and sleep. A novel deep learning network consisting of encoder and decoder structures named PMG-SATNet was proposed. The encoder comprises a parallel multi-scale feature extraction module and a global temporal relationship modeling module to capture fine-grained local patterns and long-range dependencies. The decoder employs a temporal convolutional network augmented with a spectral attention mechanism to emphasize clinically relevant ECG frequency bands and suppress respiration and body motion interference. After being validated on the self-built dataset, PMG-SATNet outperformed baseline models in terms of Pearson correlation coefficient and root mean square error, with an improvement of 3.3% and 3.8%, and 16.4% and 23.8%, respectively. The validation results imply that PMG-SATNet is capable of recovering ECG signals from millimeter-wave radar-derived chest vibrations with high fidelity and can potentially be implemented in real-life cardiac health monitoring.

心血管疾病是全球首要的死亡原因，通常具有起病隐匿、进展迅速的特点。传统的心电图（ECG）电极可能会引起皮肤刺激，从而限制了常规监测和早期风险评估的开展。近年来，凭借非接触式监测的优势，基于毫米波雷达的心脏监测结合深度学习已成为热门的研究方向。 为了克服基于单一数据集训练的方法泛化能力差的问题，本研究设计了涵盖清醒和睡眠状态的七种实验场景，并提出了一种名为 PMG-SATNet 的新型深度学习网络。该网络由编码器和解码器结构组成：编码器包含并行多尺度特征提取模块和全局时间关系建模模块，用于捕捉细粒度的局部模式和长程依赖关系；解码器则采用结合了频谱注意力机制的时间卷积网络，旨在强调与临床相关的 ECG 频段，同时抑制呼吸和身体运动带来的干扰。 在自建数据集上验证后，PMG-SATNet 在皮尔逊相关系数（Pearson correlation coefficient）和均方根误差（root mean square error）方面均优于基线模型，分别提升了 3.3% 和 3.8%，以及 16.4% 和 23.8%。验证结果表明，PMG-SATNet 能够高保真地从毫米波雷达衍生的胸部振动信号中恢复 ECG 信号，并有望应用于现实生活中的心脏健康监测。

AI Model	Architecture	Detection Capability (Number of Conditions)	Accuracy	Power Consumption	Deployment Method (Cloud/Edge)
PMG-SATNet (Proposed)	Encoder-Decoder: MFEM (Parallel 1D CNN) + GRMM (Transformers) + TCN-SA Decoder	ECG waveform reconstruction (P-wave/QRS/T-wave); 30 subjects × 7 physiological states	PCC: 0.95, RMSE: 0.08, MAE: 0.012	Not reported	Not reported
MMECG (Attention-DCGAN Baseline)	DCGAN + Attention: 4-layer deconvolution generator + 4-layer convolutional discriminator	ECG-like signal reconstruction; based on MMECG public dataset	RMSE: 0.0329 mV, PCC: 0.9853	Not reported	Not reported
RadarNet (Baseline)	MSDM Disentanglement + ResNet 1D CNN; Joint time-frequency domain loss	ECG signal reconstruction (FMCW Radar); RadarPhys-30 dataset	Sitting MAE < 0.20 bpm, Supine MAE < 0.15 bpm	Not reported	Not reported

不足总结

• 利用AI生成的综述初稿, 大纲比较全面, 但AI给出的图表建议有点跑题.