Contest Problems

1.Introduction

房颤（AF）是一种与血栓、中风、心力衰竭和其他心血管疾病有关的心律不齐[1]。随着年龄增长，患房颤的风险也随之增加，全球已有超过3300万人被诊断患有此病[2]。 由于房颤可能偶尔发作，且患者往往无明显症状，因此发现并诊断房颤颇具挑战。通常，医生会推荐疑似房颤患者使用心脏监测设备（如Holter监护仪）进行24小时或48小时的心电图（ECG）记录。 心脏专家通过分析这些心电图数据来确认是否患有房颤[3, 4]。 然而，这一诊断过程耗时费力，如果患者在佩戴Holter期间没有出现房颤，还可能出现漏诊。
Atrial Fibrillation (AF) is an irregular heart rhythm associated with blood clots, stroke, heart failure, and other cardiovascular diseases [1]. The risk of developing AF increases with age, and it is reported that there are more than 33 million people experiencing AF [2]. It is hard to identify the arrhythmia in the time since AF occurs occasionally and has no obvious symptoms even if the person does experience AF. For the patients with suspected AF, cardiac monitors such as the Holter are utilized to record the 24-hour or 48-hour electrocardiogram (ECG) and the cardiologists would confirm the diagnosis by interpreting the recorded heart rhythm [3, 4]. This diagnosis process, however, is time-consuming and could even miss the AF if the patient does not experience AF while wearing the Holter.

为了降低工作负担并提升检测效率，研究者开发了计算机辅助的房颤检测技术。此类技术特别适用于长期监测设备， 例如可植入心脏监测器（ICM）和智能手表。这些检测方法通过从记录的心电图中提取关键特征，并利用一系列带有可调节参数的启发式规则来判定房颤节律。 然而，确定这些特征及其判定逻辑需要广泛的心脏病学知识和大量的临床试验验证。 此外，由于心律特征在不同个体之间的信号动态和形态特征具有显著差异，极大增加了检测的复杂性。
To further reduce the manual workload and improve the detection efficiency, computer-aided AF detection methods have been developed and deployed on the long-term monitoring embedded devices such as the Insertable Cardiac Monitor (ICM) and the smart watch. Those methods extract crucial features from the recorded ECG and utilize a series of heuristic rules with programmable parameters to determine the AF rhythm. However, the extracted features, along with the corresponding determination logic, require great cardiology expertise and numerous clinical trials to be determined. Moreover, the rhythm characteristics are highly variable among people in terms of signal dynamics and morphological characteristics.

基于人工智能算法的房颤检测算法可以进一步提高检测性能，并减少医生在标准设计和参数调整方面的工作量[5, 6]。 本届比赛要求参赛队伍构建端侧人工智能算法，对单通道心电图（ECG）对房颤（AF）发作进行分类，同时满足龙芯开发平台部署后的实时检测要求。 我们将利用大赛公布的ECG数据集，对提交的参赛作品进行精度和性能的检测。
Automatic detection with less expertise in cardiac monitoring can further improve the detection performances and reduce the workload from physicians in criteria design and parameters adjustment [5, 6]. This year’s Challenge is asking the team to build an AI/ML algorithm that can classify AF episodes with the labeled one-channel Electrocardiogram (ECG) while satisfying the requirements of in-time detection on the Loongson development platform. We will test each algorithm on the released ECG dataset in terms of detection performances as well as practical performances, and the comprehensive performances will reveal the utility of the algorithm on real-time AF detection in cardiac monitor.

2.ECG and Atrial Fibrillation

心脏电功能的信息是通过心电图（ECG）信号获取的，该信号记录了每个心跳的心肌去极化和复极化的小电压变化随时间的变化[7]。 如图1所示，它展示了由两个连续的心脏周期（心跳）组成的心电图信号，呈正常窦性节律。正常的心电图反映了P波、QRS波群（包括Q波、R峰和S波）和T波的形态特征， 这些波是由单个心跳的不同心脏电活动产生的。心跳间隔信息，如R-R间隔和T-P间隔，也可以由心电图导出。心脏内科专家通常依据心电图的信息对心律失常做出诊断。
The information of the heart electrical function is obtained through the ECG signal, which records small electrical changes (on voltage) of the cardiac muscle depolarization and repolarization of each heartbeat against time [7]. As shown in Fig. 1, it demonstrates the ECG signal consisting of two consecutive cardiac cycles (heartbeats) with normal sinus rhythm. The normal ECG reflects the morphological characteristics of the P-wave, the QRS-complex (includes the Q-wave, the R-peak, and the S-wave), and the T-wave, which are generated by the different cardiac electrical activity of the single heartbeat. The inter-heartbeat information such as the R-R interval and the T-P interval can also be derived by the ECG. Cardiologists and Electrophysiologist (EPs) primarily reply on the information obtained from the ECG to diagnose arrhythmias.

房颤通常是基于心电图上反映的快速和不规则的节律特征进行诊断的。如图2所示，有四个心电图信号代表了两位患者的正常窦性节律和房颤。 每个心电图片段长度为10秒（以250赫兹采样）。与图2（a）和图2（b）中正常窦性节律的心电图片段相比，图2（c）和图2（d）中显示的房颤的心电图片段具有一些独特的视觉特征。 心脏病学家通过心电图诊断房颤节律，关键特征由房颤的发病机制引起：
AF is usually diagnosed based on its rapid and irregular rhythm features reflected on the ECG. As shown in Fig. 2, there are four ECG signals representing normal sinus rhythm and AF over two patients. Each ECG episode is with the length of 10 seconds (sampling at 250 Hz). When comparing with the ECG episodes of normal sinus rhythm in Fig. 2(a) and Fig. 2(b), the ECG episodes of AF shown in Fig. 2(c) and Fig. 2(d) have some unique visual features. Cardiologists diagnose the AF rhythm via ECG by two critical features caused by the pathogenesis of AF:

• 在房颤节律的心跳中，P波消失[8]。相反，如图2（c）和图2（d）所示，在图2（a）和图2（b）中明显存在的P波被颤动波（f波）替代，f波的电压值较小，但形态和频率不断变化。f波的幅度在一个小范围内变化。 这个特征是由于没有单一的冲动去极化心房所致。
  The P-wave is absent in the heartbeat of AF rhythm [8]. Instead, as shown in Fig. 2(c) and Fig. 2(d), the P-wave, which is clearly present in Fig. 2(a) and Fig. 2(b), is replaced by the fibrillatory waves (f-waves) which are small in voltage value but with varying morphology and high frequency. The amplitude of the f-waves varies within a small range. This feature is led by the fact that there is no single impulse depolarizing the atria.
• 心室率不规则不规则，通常在每分钟100至180次之间[9, 10]。如图2（c）和图2（d）所示，与正常窦性节律中的R-R间隔相比，每两个连续R峰之间的R-R间隔值彼此不同，并且没有规律的模式。 这是由于只有随机冲动导致心室去极化的发病机制造成的。
  The ventricular rates are irregularly irregular, typically in the range between 100 and 180 beats per minute (BPM) [9, 10]. As shown in Fig. 2(c) and Fig. 2(d), when comparing with the R-R intervals in normal sinus rhythm, the values of the R-R intervals between each of two consecutive R-peak are mutually different and have no regular pattern. This is caused by the pathogenesis that only random impulses result in ventricular depolarization.

这两个特征是心脏病学家区分房颤和其他节律的黄金标准。
Those two features are the golden standards for cardiologists to discriminate between AF and other rhythms, and hence determine the AF rhythm when reading the ECG.

3.Objective

今年挑战赛的目标是在单通道心电图上对房颤进行检测。
The goal of this year's challenge is to discriminate AF (i.e., Atrial Fibrillation) from single-channel ECG recordings.

我们要求团队设计并实现一个可工作的、开源的人工智能算法，可以从心电图片段中区分房颤（即二元分类：房颤或非房颤）， 同时能够部署并在给定的龙芯平台上高效运行。我们将根据检测精度、模型泛化能力、内存占用和推断延迟等综合性能，结合现场答辩，对表现最佳的团队进行奖励。
We ask teams to design and implement a working, open-source AI/ML algorithm that can automatically discriminate AF (i.e., binary classification: AF or non-AF) from ECG episodes while being able to be deployed and efficiently run on the given Loongson platform. We will award prizes to the teams with top comprehensive performances in terms of detection precision, model generalization, memory occupation, inference latency, and presentations.

4.Data

本届大赛采用的数据为单导联心电图片段。每个心电图片段长度为5秒，采样率为250赫兹。记录经过预处理，利用带通滤波器去除噪声。 每个心电图片段有一个标签，描述了心脏异常（或正常窦性）节律。在这里，对于房颤的分类，标签为房颤（Atrial Fibrillation，AF）。对于非心律失常（non-AF），段标记为除房颤外的其他标签。 我们提供了标签列表供参考。
The data contains single-lead ECG episodes. Each episode is 5-second in length with 250 Hz sampling rate. The recordings are pre-processed by applying a band-pass filter to remove noise. Each ECG episode has one label that describes the cardiac abnormalities (or normal sinus rhythm). Here, for the classification of AF, it contains the labels being AF (Atrial Fibrillation). For the non-AF, the segments are labeled with the one other than AF. We have provided the lists of labels for references.

心电图片段根据患者进行分割，分为训练集和测试集。80%的受试者心电图片段被释放为训练数据集，其余20%的受试者心电图片段将用于评估提交算法的检测性能。 训练数据集将发送至参赛队伍注册的通信邮箱，用于最终评估的测试数据集将不会发布。
The ECG episodes are partitioned patient-wisely in the training and testing set. 80% of the subjects’ ECG episodes are released as training material. The rest 20% of the subjects’ ECG episodes will be utilized to evaluate the detection performances of the submitted algorithm. The training dataset will be sent to the register email address of each team. The dataset for final evaluation will remain private and will not be released.

5.Data Format

所有数据都以文本格式进行格式化。每个文本文件包含一个5秒钟的心电图片段，每行采样点总长度为1,250个。每个采样点的值以浮点数形式存储。
All data is formatted in text. Each text file contains a 5-second ECG episode with the sampling point on each row of 1,250 in total length. The value of each sampling point is stored as a floating-point number.

每个文本文件的名称包含患者ID、标签和索引，用破折号（即“-”）分隔。例如，对于文件“S0-AFIB-0.txt”，“S0”表示受试者编号，“AFIB”表示心律失常或节律的类型 （参考提供的标签列表），“0”表示片段的索引。
The name of each text file contains the patient ID, labels, and index separated by a dash (i.e., "-"). For example, for the file "S0-AFIB-0.txt", "S0" represents the subject number, "AFIB" represents the type of arrhythmias or rhythm (refer to the provided label list), "0" represents the index of the segments.

6.Loongson Platform

本次比赛采用的开发板为龙芯2K500先锋板，广东龙芯2K500先锋板采用龙芯2K0500芯片，是LoongArch架构首款面向嵌入式应用的开发板，兼容行业生态。先锋板集成LCD/以太网/USB等基本接口，扩展支持（插针形式）2个SPI，2个I2C，6路串口，2路CAN，4路PWM, 8个GPIO等接口。 更多资料详见： [用户手册], [先锋板详细资料], [用户试用报告等板卡使用资料]。 完成注册并核验资质成功的队伍，龙芯科技将提供实验平台，根据队伍注册时提供的收件地址，邮寄开发板至队伍。
The development board required in the Challenge is Loongson 2K500 Pioneer Board. The data sheet of the board can be accessed [datasheet], [other materials], [user reports]。 For registered team which completes the identity verification, Loongson will mail the board by the registered mailing address.

7.Scoring

队伍最终得分由算法分数和答辩分数构成。 对于算法分数，我们将使用综合指标对提交作品进行评估打分，指标包括模型精度、模型泛化性、推理延时和存储占用。指标的具体定义如下：
The final score of the submitted solution consists of two part, algorithmic score and presentation score. As for algorithmic score, we will evaluate the submitted algorithm with the scoring metric that evaluates the comprehensive performances in terms of detection performances and practical performances. It is defined as follows:

For detection performances,

• score: This score is set to measure the AF detection performance of the model. We will obtain the confusion matrix of the classification generated by the submitted design for each testing subject, where the case positive is AF. Then, the score of each individual subject from the testing dataset will be calculated as follows: with =2. This setting gives a higher weight to recall since the detection accuracy of AF is the most important metric for cardiac monitor. It is expected to discriminate as many AF recordings as possible to avoid missed detection on AF while achieving a high detection accuracy on non-AF recordings to reduce inappropriate shock rate. The final score will be obtained by averaging of all testing subjects as follows: . The score will be normalized by , where , .
• Generalization score: This score is set to evaluate the generalization of the submitted model for AF detection on different subjects. A well-generalized model could perform well on different subjects. We will first calculate for each subject from the testing dataset and obtain , which is the number of the surpassing a preset threshold 0.9. The generalization score is defined as follows: , where is the total number of the testing subjects.
  The generalization score will be normalized by , where , .

For practical performances,

• Inference latency: The average latency (in ms) of inference executed on the Loongson development board over recordings from testing dataset will be measured. The latency score will be normalized by , where , .
• Memory occupation: The memory occupation (in KiB) will be measured based on Flash occupation of Loongson development board for the storage of the deep neural network model and the program. To be more specific, is the memory occupation by the developed program when building the project. will be normalized by , where , .

提交作品的算法得分将为上述指标的加权平均分，加权方式如下。
The final algorithmic score will be calculated based on the weighted average of aforementioned metrics achieved by the submitted solution by follow equation.

算法得分部分成绩优异的参赛队伍将会收到ICMC大会邀请，受邀的参赛队伍将进行现场答辩和专家提问。评审专家将对每项作品实行分项打分并集体讨论。 最终获奖队伍的判定将由算法得分和答辩得分共同决定。
Invited participants will conduct on-site presentation and expert questioning. Evaluation experts will score each work separately and hold collective discussions. The final ranking is determined by the scores of the project and the presentation.

获奖规则：

最终获奖队伍的判定将由算法得分和答辩得分共同决定，即以获奖为目标的参赛队伍必须参加现场答辩以获得20%的答辩分数。

Award Rules:

The final ranking is determined by the scores of the project and the presentation, i.e., teams aiming to win a prize must participate in the on-site presentation to get the presentation score which occupied 20% of totals.

8.Example Code

我们提供了[pytorch]和 [tensorflow] 两种框架在PC端的模型训练例程。并提供了训练后模型在先锋板上部署和推理计算的 [例程]。 此外，我们公布了用于最终测试的[例程]。
例程部署与测试的演示视频链接如下：https://www.bilibili.com/video/BV1gf421X7f2/
We have provided [pytorch] and [tensorflow] example code to illustrate how to train the model. The example code of model deployment and evaluation can be accessed [here]. In addition, we have published [example code] for the final evaluation.

9.References

[1] A. H. Association, “What is atrial fibrillation (afib or af)?” https://www.heart.org/en/health-topics/atrial-fibrillation/what-is-atrial-fibrillation-afib-or-af, 2016.

[2] M. Jiao, C. Liu, Y. Liu, Y. Wang, Q. Gao, A. Ma, “Estimates of the global, regional, and national burden of atrial fibrillation in older adults from 1990 to 2019: insights from the Global Burden of Disease study 2019”, Front. Public Health, vol. 11, 2023.

[3] C. Ferguson, S. C. Inglis, P. J. Newton, S. Middleton, P. S. Macdonald, and P. M. Davidson, “Atrial fibrillation: stroke prevention in focus,” Australian Critical Care, vol. 27, no. 2, pp. 92–98, 2014.

[4] L. Toner, A. Al-Kaisey, A. Koshy, F. Ha, R. Spencer, J. Sajeev, A. Teh, O. Farouque, and H. Lim, “The accuracy of smartwatches compared to holter monitors for heart rate monitoring in atrial fibrillation: A pilot study,” Heart, Lung and Circulation, vol. 28, p. S229, 2019.

[5] Z. Jia, Y. Shi, S. Saba, J. Hu, "On-device prior knowledge incorporated learning for personalized atrial fibrillation detection." ACM Transactions on Embedded Computing Systems (TECS) vol. 20, no. 5s, pp. 1-25, 2021.

[6] T. Zhou, H. Li, Z. Shen and Z. Jia, “End-Edge Coordinated Multiview Deep Learning for Time-Efficient Atrial Fibrillation Detection.” Proceedings of the IEEE 24th International Conference on High Performance Computing & Communications; 8th Int Conf on Data Science & Systems; 20th Int Conf on Smart City; 8th Int Conf on Dependability in Sensor, Cloud & Big Data Systems & Application (HPCC/DSS/SmartCity/DependSys) (pp. 2003-2010). IEEE.

[7] K. Richard, “Cardiac electrophysiology: normal and ischemic ionic currents and the ECG.”, Advances in physiology education 41, 1 (2017), 29–37.

[8] H. Pürerfellner, E. Pokushalov, S. Sarkar, J. Koehler, R. Zhou, L. Urban, and G. Hindricks, “P-wave evidence as a method for improving algorithm to detect atrial fibrillation in insertable cardiac monitors.” Heart Rhythm 11, 9 (2014), 1575–1583.

[9] J. Lian, L. Wang, and D. Muessig. “A simple method to detect atrial fibrillation using RR intervals.”, The American journal of cardiology 107, 10 (2011), 1494–1497.

[10] S. Sarkar, D. Ritscher, and R. Mehra. “A detector for a chronic implantable atrial tachyarrhythmia monitor.”, IEEE Transactions on Biomedical Engineering 55, 3 (2008), 1219–1224.