wECGdb: An ECG database acquired using a wrist-worn device from patients with acute myocardial infarction and controls

Rationale According to Eurostat [1], timely interventions could prevent two-thirds of deaths in individuals under 75, with myocardial infarction being the leading cause [2]. Myocardial infarction is typically diagnosed with a 12-lead electrocardiogram (ECG), often unavailable outside the hospital when early symptoms like chest pain occur. Accessible technology for home-based multilead ECG acquisition, followed by automatic ECG analysis or specialist review, may be a promising solution to this problem. Several smartwatches on the market offer ECG functionality; however, they cannot acquire chest ECG leads, which is a crucial drawback, as anterior, septal, and lateral infarctions cannot be diagnosed without such leads. To address the lack of ECG leads, a wrist-worn wearable device, featuring three electrodes, can be used to enable the simultaneous acquisition of two ECG leads with a single touch (hereafter referred to as a wECG) [3]. One lead is standard lead I, while the other is acquired from chest location, thereby enhancing the amount of cardiac information. However, since the latter lead is non-standard, it poses challenges for interpretation. Therefore, adopting the standard ECG configuration, e.g., through 12-lead ECG synthesis, is essential to present the information in a format that is clinically interpretable. The wECGdb database contains ECGs simultaneously acquired using a wrist-worn device and a 12-lead ECG for reference. Therefore, it is particularly suitable for the development and testing algorithms for 12-lead ECG synthesis from wECG. Subjects The database consists of data from 92 participants, divided into three groups: healthy participants, patients with acute myocardial infarction, and patients with other cardiovascular disease (CVD). To be eligible for inclusion, participants had to be at least 18 years old, without an implanted cardiac device, and without cognitive or linguistic impairments. The acute myocardial infarction group consisted of patients diagnosed with either STEMI or NSTEMI, with ECGs taken within 24 hours of percutaneous coronary intervention. The other CVD group included patients with heart conditions that caused infarction-like changes in the ECG. The healthy group consisted of individuals with no history of heart disease. The patients were recruited from the inpatient wards of the Cardiology Department at Vilnius University Hospital Santaros Klinikos, Lithuania. All eligible participants provided signed, written informed consent in accordance with the ethical principles outlined in the Declaration of Helsinki. The study was approved by the regional bioethics committee, under reference number 158200-18/7-1052-557. wECG acquisition The wrist-worn wearable device, developed at the Biomedical Engineering Institute of Kaunas University of Technology [3, 4], equipped with three bio-potential electrodes, was used to acquire two wECG leads at a single touch. The wECG was acquired at a sampling rate of 500 Hz. For all participants, the device was positioned slightly above the wrist on the left arm. Lead I, between the left arm (LA) and right arm (RA), was acquired by touching one electrode with the right index finger. The other lead was obtained by placing the electrode on the strap against a specific part of the body. Acquisition of a standard 12-lead ECG The standard 12-lead ECG was acquired using disposable Ag/AgCl electrodes at 200 Hz with the Euroholter 12view recorder (Lumed, Italy) and resampled to 500 Hz to match the sampling rate of the wECG. ECG preprocessing Both the wECG and the 12-lead ECG were filtered using a high-pass Butterworth filter with a cutoff frequency of 0.5 Hz and a low-pass Parks-McClellan filter with a cutoff frequency of 100 Hz. Some elderly patients had difficulty maintaining consistent pressure on the electrode, resulting in fewer high-quality wECGs. The database includes only wECGs with acceptable signal quality, as assessed by the consensus beat detection signal quality index [5]. To improve wECG quality, each beat in each lead was replaced by an amplitude-scaled average beat. Technical details The wECGdb database is divided into development and test datasets. The development dataset was collected by asking participants to touch specific body sites under clinician guidance: an abdominal site (A), located 2 cm to the left of the umbilicus, and two precordial sites corresponding to the conventional V3 and V5 electrode sites, positioned just below those used to acquire the standard 12-lead ECG. This protocol produced three leads: LA-A, LA-V3, and LA-V5. The test dataset includes only the LA-A lead, which was self-acquired by participants without clinician assistance. Each recording lasted approximately one minute, with at least a one-minute interval between recordings. The acquired signals are provided in MAT-files labeled as follows: dataset_XXX, where XXX represents the participant ID. Each dataset_XXX file contains five structures: info – Contains participant information, including age, sex, diagnosis (infarction type and location, other CVD), and sampling frequency (Fs); LA_A – ECG recordings from the abdomen area with clinician assistance; LA_A_self – ECG recordings from the abdomen area without clinician assistance; LA_V3 – ECG recordings from the V3 electrode placement area with clinician assistance; LA_V5 – ECG recordings from the V5 electrode placement area with clinician assistance. The LA_A, LA_A_self, LA_V3, and LA_V5 structures include the following variables: data_availability – “yes” indicates that the recording meets the quality criteria and is therefore available. “no” indicates that the recording did not meet the quality criteria and is not available. wECG – Two ECGs leads acquired using the wrist-worn device; reference_12_lead – Simultaneously acquired standard 12-lead ECG; ecg_leads – Labels of the ECG leads. Additionally, the LA_A, LA_V3, and LA_V5 structures contain subdivided recordings into training and testing segments. These segments each include the variables: data, reference_12_lead, and wECG. The file participant_info.xlsx contains details about the participants, including age, infarction type (STEMI/NSTEMI), presence of bundle branch block (left or right), and information about other CVD. wECGdb application example: ECG synthesis using an echo state network The utility of the wECGdb database is demonstrated through the development of a person-specific model for synthesizing the standard 12-lead ECG from the two-lead wECG. The 12-lead ECG is synthesized using an echo state network – a fixed, sparsely connected, recurrent neural network that functions as a random nonlinear excitable medium to the input wECG [6, 7]. The MATLAB code for designing an echo state network is available in [8].研究依据 根据欧盟统计局（Eurostat）[1]的统计数据，及时的临床干预可避免75岁以下人群中三分之二的死亡病例，而心肌梗死（myocardial infarction）是此类死亡的首要诱因[2]。心肌梗死的常规诊断依赖12导联心电图（electrocardiogram, ECG），但当患者出现胸痛等早期症状时，院外往往无法获取该检查设备。基于家用的多导联心电图采集技术，配合自动心电图分析或专业医师审阅，或是解决该问题的可行方案。目前市场上多款智能手表已支持心电图功能，但无法采集胸导联心电图，这一关键缺陷导致其无法诊断前壁、间隔壁及侧壁心肌梗死，因为这类梗死的诊断依赖胸导联数据。 为弥补胸导联采集的不足，一款搭载3个生物电位电极的腕戴式可穿戴设备可通过单次触摸同步采集两路心电图信号（下文简称腕部心电图（wECG））[3]。其中一路为标准导联I，另一路采集自胸部区域，从而提升心脏电活动的信息采集量。但由于该非标准导联的信号解读存在挑战，因此需采用标准心电图的配置格式——例如通过12导联心电图合成技术——将采集到的信号转换为临床可解读的标准形式。 wECGdb数据集包含腕戴式设备采集的心电图与同步采集的12导联参考心电图，因此特别适用于基于腕部心电图合成12导联心电图的算法开发与测试。 研究对象 本数据集共纳入92名受试者，分为三组：健康受试者、急性心肌梗死患者以及其他心血管疾病（cardiovascular disease, CVD）患者。入组标准为：年龄≥18岁，无心脏植入式设备，且无认知或语言障碍。急性心肌梗死组纳入经确诊为ST段抬高型心肌梗死（STEMI）或非ST段抬高型心肌梗死（NSTEMI）的患者，其心电图采集于经皮冠状动脉介入治疗（percutaneous coronary intervention, PCI）后24小时内。其他心血管疾病组纳入心电图呈现类似梗死样改变的心脏病患者；健康组则纳入无心脏病病史的个体。 所有受试者均招募自立陶宛维尔纽斯大学医院圣塔罗斯诊所心内科住院病房。所有符合入组标准的受试者均签署了书面知情同意书，研究严格遵循《赫尔辛基宣言》规定的伦理原则。本研究已获得地区生物伦理委员会批准，审批编号为158200-18/7-1052-557。 腕部心电图采集方案 本研究采用的腕戴式可穿戴设备由考纳斯大学生物医学工程研究所研发[3,4]，搭载3个生物电位电极，可通过单次触摸同步采集两路腕部心电图信号，采样率为500 Hz。 所有受试者均将设备佩戴于左臂腕部稍上方位置。标准导联I采集于左臂（LA）与右臂（RA）之间，受试者通过右手食指触摸设备上的对应电极即可获取该导联信号；另一路导联则通过将设备绑带上的电极贴合至身体特定区域采集得到。 标准12导联心电图采集方案 标准12导联心电图采用一次性Ag/AgCl电极采集，使用Euroholter 12view记录仪（意大利Lumed公司），采样率为200 Hz，后续重采样至500 Hz以匹配腕部心电图的采样率。 心电图预处理流程 腕部心电图与12导联心电图均经过滤波处理：采用截止频率为0.5 Hz的高通巴特沃斯（Butterworth）滤波器，以及截止频率为100 Hz的低通Parks-McClellan滤波器。部分老年受试者难以维持电极的稳定按压，导致合格的腕部心电图样本量较少。本数据集仅纳入经一致性搏动检测信号质量指数[5]评估为合格的腕部心电图信号。为进一步提升腕部心电图质量，每一路导联的每个心动周期均被替换为经过幅度缩放的平均心动周期波形。 数据集技术细节 wECGdb数据集分为开发集与测试集两部分。开发集的采集流程为：在临床医师指导下，让受试者触摸特定身体部位，包括脐部左侧2 cm处的腹部区域（A），以及对应常规V3、V5电极位置的胸前区域（该位置略低于标准12导联心电图的电极安放点），由此可采集三路导联信号：LA-A、LA-V3与LA-V5。测试集仅包含LA-A导联信号，由受试者在无医师指导的情况下自行完成采集。每段记录时长约为1分钟，且相邻记录之间至少间隔1分钟。 采集到的信号以MAT文件（MAT-files）格式存储，文件命名规则为：dataset_XXX，其中XXX为受试者编号。每个dataset_XXX文件包含5个结构体： - info：存储受试者基本信息，包括年龄、性别、诊断结果（梗死类型与部位、其他心血管疾病情况）以及采样频率（Fs）； - LA_A：临床医师指导下采集的腹部区域心电图信号； - LA_A_self：受试者自行采集的腹部区域心电图信号； - LA_V3：临床医师指导下采集的V3区域心电图信号； - LA_V5：临床医师指导下采集的V5区域心电图信号。 LA_A、LA_A_self、LA_V3与LA_V5结构体包含以下变量： - data_availability：取值为"yes"时表示该段记录符合质量标准，可纳入分析；取值为"no"时表示该段记录未达到质量标准，不可用； - wECG：腕戴式设备采集的两路心电图导联信号； - reference_12_lead：同步采集的标准12导联心电图信号； - ecg_leads：心电图导联的标签信息。 此外，LA_A、LA_V3与LA_V5结构体还将记录划分为训练段与测试段，每个分段均包含data、reference_12_lead与wECG三个变量。 文件participant_info.xlsx存储了受试者的详细信息，包括年龄、梗死类型（STEMI/NSTEMI）、束支传导阻滞（左束支或右束支）情况以及其他心血管疾病相关信息。 wECGdb数据集应用示例：基于回声状态网络的心电图合成 本示例通过构建受试者专属模型，从两路腕部心电图信号合成标准12导联心电图，以此验证wECGdb数据集的实用性。该12导联心电图合成采用回声状态网络（echo state network）——一种固定且稀疏连接的循环神经网络（recurrent neural network），可将输入的腕部心电图信号转换为随机非线性可激发介质[6,7]。回声状态网络的MATLAB代码可参考文献[8]获取。