key: cord-0067385-2s7bn24e
authors: Li, Jing; Lu, Yunhang; Xiao, Ziyi
title: Sports Training Health Analysis Algorithm Based on Heart Rhythm Feature Extraction and Convolutional Neural Network
date: 2021-08-30
journal: J Healthc Eng
DOI: 10.1155/2021/2946044
sha: 893e193cece36ab78950222ada73c00b6609c4da
doc_id: 67385
cord_uid: 2s7bn24e

Intelligent sports equipment and software have emerged in the field of sports as a result of the advancement of information technology, allowing professional athletes to collect and visually display the movement and physical signs of the human body to aid in the planning of sports strategies. Intuitive data, on the other hand, cannot assist ordinary people who lack professional knowledge in exercising correctly. As a result, in the field of intelligent sports and health, effective use of collected exercise and physical sign data to analyze the user's personal physical condition and generate reasonable exercise suggestions has emerged as a research direction. In humans, the heart sound signal is a biological signal. It can help people detect and monitor heart health problems by analyzing the characteristics of heart sound signals. The goal of this paper is to use heart sound to identify and analyze athletes' training health. It provides a revolutionary health analysis algorithm based on heart rhythm feature extraction and convolutional neural networks, which is based on exercise training. It greatly improves the accuracy of the recognition and prediction of the athlete's training health status.

e economic status of today's society is improving all the time, but fast-paced work and rising living costs have had a significant negative impact on people's health [1] [2] [3] . Young and middle-aged adults' adult physical fitness indexes dropped by an average of 0.5 percent. As a result, there are still flaws in people's physical health [4, 5] , particularly among adults who are under a lot of work and life stress. Sports are an effective way to keep the human body healthy on a daily basis [6] , and numerous studies have shown that using the right exercise methods and getting enough exercise can improve physical fitness and health in people of all ages.

As a part of the sports health market, sports fitness has also developed rapidly in recent years. According to estimates, the scale of China's online sports fitness market reached 174 billion yuan in 2019, and the number of users reached 195 million. e huge user scale brings a huge amount of information and data, which contains great value. At present, sports health applications [7] [8] [9] [10] are more focused on data collection [11] and data visualization display [12] [13] [14] and lack further analysis and mining of user health data. Reasonable advice and help is the main direction of sports and health applications in the intelligent age.

In recent years, technologies such as the Internet of ings [15, 16] , big data [17, 18] , cloud computing [19, 20] , and artificial intelligence [21] [22] [23] have gradually matured.

e Internet of ings technology represented by 5G has promoted the in-depth development of the mobile Internet and provided an infrastructure for large-scale real-time user data acquisition. In order to improve resource utilization, cloud computing integrates storage and computing resources into a scalable and expandable resource pool, which has large-scale computing and storage capabilities. Customers can uniformly manage and schedule resource pools through the network, with great flexibility. erefore, cloud computing has the advantages of reducing hardware costs, enhancing independent development capabilities, and improving resource utilization. Big data technology can quickly store and calculate massive amounts of data, provide enterprises with more accurate decision-making assistance, increase productivity, and reduce production costs. Artificial intelligence technologies such as deep learning and machine learning have aided the industry's evolution toward unmanned and intelligent operations, allowing customers to obtain more precise unmanned services, lower societal production costs, and increase service quality. Integrating technologies such as the Internet of ings, cloud computing, big data, and artificial intelligence to achieve a low-cost, intelligent, and precise sports health cloud platform can promote the development of the sports health industry and provide advanced technical assistance for improving the physical health of the people. e goal of this paper is to look at the problem of sports health prediction [24] [25] [26] and recognition from the standpoint of heart sound periodic feature extraction methods and heart sound neural network structure design. To provide accurate services for athletes and promote the development of sports health and intelligence, a sports training health analysis algorithm based on heart rhythm feature extraction and convolutional neural network is proposed.

Following are the main innovations of this paper:

(1) e proposed novel method can use the current mainstream classifier with the image and biological signal-based convolution neural network (CNN) and further combine CNN and heart sounds feature extraction method. It can be used for recognizing the heart sounds of the athletes training. (2) In this paper, heart sound convolutional neural network is designed to solve the problem of heart sound recognition and classification based on the extraction of heart sound periodic features. is paper studies the heart sound signal preprocessing, the feature extraction method, and the structure design of the heart sound neural network, starting from the mathematical definition and deducing the experimental method. e rest of the article is organized as follows. In Section 2, background study and literature review are elaborated. e methodology is discussed in Section 3, followed by experimental setup and results in Section 4. Finally, Section 5 concludes the paper.

Sports health services [27] include exercise advice, physical fitness monitoring, etc. Among them, exercise advice is also called exercise prescription. It was proposed by physiologist Kapović in the 1950s. Due to the difference in human physique, each person's suitable exercise method and amount of exercise vary. ere are big differences. Just like the form of medical consultation, exercise advice is customized according to the user's own physical characteristics, forming the concept of exercise prescription. e traditional sports generation is a oneto-one method, which is limited by the scale of sports prescription experts and cannot meet large-scale needs. With the emergence of big data and artificial intelligence technology, especially the mature and unmanned artificial intelligence technology, intelligently generating exercise prescriptions and providing precise services for users has become a major direction of sports health services.

With the development of data information technology such as big data and machine learning [28, 29] , using data algorithms to mine valuable information from large amounts of data has become a hot topic of data research [30, 31] . e sports health application, which collects a large number of users' sports and physical signs data, can use data information technology to dig out valuable information, so that the sports assistant service can be more intelligent and personalized.

Data analysis can assist people in comprehending the various types and structures of data, but such a superficial understanding of data is insufficient to perform the data's intended function. Data also assist humans in focusing on the information contained in data analysis, which is the true value of data. Such information can be classified as biological or nonbiological. Genomic data, microbiome data, personalized medicine data, and disease data are examples of biological type data, and this topic focuses on the types of data found in personalized medicine and disease data. It is well known that a certain biological signal data contains the physiological characteristics of the organism, and we can analyze the biological information from such biological signal data. Like the uniqueness of the organism, its biological signal also has certain uniqueness.

ese data have become a treasure trove for many researchers. By studying the biological signals of the same organism, we can understand the differences between these data, so that we can draw some quantitative or regular conclusions. When we have data like the amount of lactic acid in the muscle at rest and the amount of lactic acid in the muscle at exercise, it can help us draw some accurate conclusions. It can be said that the development of data collection and analysis methods has helped us understand the laws of life more and more. ere are also many biosignal data in humans, which are beneficial to the researchers studying the physiological structure of human beings and the pathologic causes of human beings.

In recent years, convolutional neural network (CNN) has emerged in the field of neural networks, especially in the field of image recognition, which has an unprecedented effect. In the classification and recognition of images, CNN has a wide range of application prospects. Not only that, but for medical images, it is also within the scope of CNN. At present, in the research on heart sound signals in academia, many researchers use deeper and more complex networks and more computing resources in pursuit of higher accuracy, and the methods used by these researchers to improve accuracy are of very high standard. Special consideration is seldom given to storage space and computing resources. e included platforms have limited computing power in real-world applications, such as hand-held heart sound filters, wearable life suits, and other tasks, due to production costs and actual market demand, and these platforms are all real time. Inspectors must be able to obtain test results quickly while keeping costs low, so these platforms must be able to process classification and recognition tasks quickly. As a result, in the engineering of heart sound classification and recognition, a neural network with fewer parameters and calculations is more feasible.

In the process of heart sound acquisition, attention should be paid to the following three aspects in the design of heart sound acquisition device.

First, ambient noise should be considered. is is a problem that will be encountered in all the processes of heart sound signal acquisition, which requires extra attention and efforts to reduce the impact of environmental noise.

Second, errors from the acquisition device and the operator's misoperation should be considered. is can be improved or avoided during the collection process, such as handling lightly and selecting sensors with better performance for the collector.

ird, other biological signals from the gatherers themselves should be considered. In the human body, there are other biological signals besides a heart sound, for example, lung sounds, so if the collector picks up heart sounds right after strenuous exercise, they will be different from the heart sounds that are picked up at rest. erefore, this section reviews the current situation of heart sound acquisition equipment and looks forward to more intelligent heart sound acquisition devices in the future from the composition to the specific structure. is is shown in Table 1 .

e original heart sound signal is typically sampled at 22050 Hz. e sampling point is relatively large when compared to the signal's frequency. e difference between the two is an order of magnitude, so there is a lot of original sampling data. e frequency range of the heart sound signal is 101000 Hz, so there is a lot of original sampling data. It can be said that the majority of the data in these data are other irrelevant signals and may even be interference signals.

erefore, such a sampling rate increases the possibility of noise, and the huge amount of data also increases the signal. e time complexity and space complexity of processing will affect the speed of processing heart sound signals. So, it is necessary to resample the original heart sound signal. is section uses a sampling rate of 2000 Hz, and the sampling effect is shown in Figures 1(a)  and 1(b) .

Denoising. Because the heart sound signal is a biological signal with a low signal strength, it is easily disturbed by noise in signal collection or processing. If the noise in the signal is not removed, it will have a significant impact on the use of heart sound in disease diagnosis. As a result, the heart sound signal must be denoised. For example, the aforementioned heart sound signal collection device can recognize that current technology has yet to achieve zero noise for heart sound signal collection. erefore, it is necessary not only to rely on hardware measures to solve the interference problem but also to have a specific signal filtering technology to participate in it, which lays a good foundation for various studies of heart sound signals.

e selection of wavelet basis function and the determination of the number of signal decomposition layers are two important aspects that affect the performance of wavelet denoising. If the selection of wavelet basis is not good, the whole performance will suffer loss. erefore, both the small heart sound signal and the noise signal are separated, if you choose too many levels of decomposition, but if you choose too few levels of decomposition, you filter out some of the useful signals. It is very important to choose the wave basis function or to determine the number of decomposition layers. e denoising process of heart sound signal using wavelet can be summarized as the following basic steps:

(1) Select the wavelet basis function most suitable for heart sound signal. (2) Determine the number of decomposition layers N.

(3) According to the decomposition level N, the heart sound signal is decomposed in N layers. (4) For the high-frequency coefficient of each layer obtained after decomposition, soft 1}7 value or hard 1}7 value is selected for quantitative processing. (5) If the effect is not good, return until the best effect is obtained.

Because the heart sound is a quasiperiodic signal and this paper is a periodic study of the heart sound signal, this section employs the envelope extraction method without indicating the heart sound's position. In fact, the height of the peak in the time domain for heart sounds contains the characteristics of the heart sound signal, which can be reflected in the heart sound signal's envelope. In signal analysis tools, Hilbert transform is an important method. Hilbert transform is generally used to extract the phase information of the signal, calculate the envelope of the signal, and analyze the spectrum information of the signal. e definition of Hilbert transform is given below. Suppose there is a signal x(t), after Hilbert transformation, x(t) can be obtained from the following equation:

(1) 

It can be seen from equation (1) that the output signal is linearly related to the input signal divided by time. Just like the relationship between voltage and current, a filter is added in between. erefore, this transformation can be turned into a filter, and the output response of the filter is set to h(t); then:

e above formula can be seen in the form of the signal in the time domain. But the Hilbert transform is mainly to analyze the frequency-domain relationship of the signal, so we need to know the frequency response H(f) of the Hilbert transform.

erefore, after the Hilbert transform of the signal x(t), the result in the frequency domain is the spectrum shift. e signal with positive frequency rotates 90°counterclockwise in the frequency domain, that is, the phase shift is reduced by 90°. e signal with negative frequency rotates 90°clockwise in the frequency domain, which increases the phase shift by 90°.

erefore, there is a difference of 180°between the positive frequency signal and the negative frequency signal. If the output signal is described as x(t) and the input signal is described as x(t), the envelope of the signal can be obtained. erefore, if the envelope signal is set as z(t), the equation can be obtained:

e original signal is shown in Figure 2 . e heart sound envelope obtained by the Hilbert transform method is shown in Figure 3 . e original signal is represented by a blue curve, and the envelope curve is represented by a red curve. In the Hilbert transform, an envelope curve of the same length of the signal can be obtained, that is, the envelope characteristic of the signal can be completely expressed.

e architecture of the depth model can be determined by the total number of layers n and the number of structures in each stage a i . For example, n � 3 and a 1 , a 2 , a 3 � 4, 3, 2 indicate that the model has 3 stages, and the number of convolutional layers in the first, second, and third stages is 4, 3, and 2, respectively. Before deriving the formula, you need to define some required parameter symbols, which are the input spectrogram size z, the convolution kernel size k, and the minimum c value t. is paper requires that the c value of all layers is not less than the minimum c value t. As the receptive field of the convolutional layer continues to grow and the size of the convolution kernel in each convolutional layer remains unchanged, it can be seen that the c value of the last convolutional layer is the smallest. erefore, it is equivalent to ensuring that the c value of the last layer in each convolutional layer is not less than t, which can be converted into a set of inequalities:

where 2 l k is the size of the receptive field of the last layer of the convolutional layer l i�1 2 i−1 (k − 1)a i of the l-th layer. e receptive field at the top of the convolutional layer should be smaller than the area of one heart sound cycle.

at is, the formula is expressed as follows:

e objective function can be expressed in the form of the total number of maximum pooling layers, and its final equation is

, l � 1, 2, . . . , n,

e model of the heart sound convolutional neural network designed in this paper is shown in Figure 4 

is experiment uses the PyCharm compiler and the TensorFlow deep learning framework in the Windows environment and uses the constructed heart sound convolutional neural network to predict and recognize the athlete's training health status. Using a small batch learning method, Adam optimizer, the learning rate is 0.0001, and a total of 1000 iterations are performed. Journal of Healthcare Engineering

is article uses the PhysioNet/CinC Challenge 2016 database. All recordings have been resampled to 2000 Hz, and the duration of the recordings ranges from 5 seconds to more than 120 seconds. e ultimate goal of this study is to divide them into two types of normal and abnormal heart sounds. e training set consists of five data folders (A to E), containing a total of 3240 heart sound recordings, and the test set contains 301 heart sound recordings. In this study, by extracting 5 cardiac cycles as the time length of a spectrogram, a total of 28,000 training samples are obtained. ere are 4800 test samples.

Relative error (RE) takes the actual value as the reference value and judges by comparing the relative difference between the predicted result of the target data and the actual value. We usually express it as a percentage. e smaller the value of the relative error is, the closer the prediction result is to the actual value. In other words, the better the prediction effect of the constructed model. Generally speaking, the relative error can better reflect whether the prediction result is credible. e relative error expression is as follows:

On the basis of the designed basic heart sound neural network structure, the complexity of the model is optimized by changing the number of convolutional layers, and the model is verified experimentally; the parameters of the heart sound classification model, training time, and accuracy of training are shown in Table 2 . Some conclusions can be drawn from Table 2 . When there is only one convolutional layer, the amount of model parameters is huge, important features cannot be accurately obtained, and the training time is long; when there are three convolutional layers in the model, the training time is short and the number of parameters is small, but the accuracy rate is reduced, and in the case of two convolutional layers, the training time is acceptable and the accuracy is the highest.

is paper realizes the design of adapting to the heart sound signal by improving the structure. In order to illustrate the effectiveness of the model designed by the method proposed in this paper, the heart sound challenge database dataset is used as the data source. e fixed number of training steps is 10,000 steps. We change the batch_size of the model in this section to different values for comparative analysis. e training process is shown in Figure 5 .

It can be clearly seen from Figure 5 that when the batch_size is 32, the optimal effect can be obtained. However, because the training of 10,000 steps is too small for the general CNN, the following fixes the batch_size size and changes the number of trainings in order to obtain the best model.

We verify the influence of the custom parameter learning rate on the performance of the proposed algorithm. We set up a set of ablation experiments and selected 4 different learning rates for simulation experiments. e experimental results are shown in Table 3 .

It can be seen from Table 3 that if the learning rate is reduced, the performance is increased as a whole, until it reaches the highest value at 0.0001. Secondly, after the learning rate is reduced again, the performance decreases. Journal of Healthcare Engineering erefore, this proves that the learning rate setting of the proposed algorithm is correct.

Limitation. Although the model based on heart rhythm feature extraction and convolutional neural network proposed in this paper has achieved effective results, the heart sound signal is a kind of time series data, which involves the dependence of long-distance information. ere are certain limitations in feature extraction. In the next study, we will investigate the effectiveness of recurrent neural networks in extracting features of heart sound signals.

In this paper, we believe that the effective use of collected exercise and physical sign data to analyze the user's personal physique and generate reasonable exercise suggestions has become the research direction in the field of intelligent exercise and health. e heart sound signal is a biological signal that exists in humans. By analyzing the characteristics of heart sound signals, it can help people detect and monitor heart health problems. e purpose of this paper is to identify and analyze the training health of athletes by heart sound. It uses the current mainstream convolutional neural network (CNN) on images and biosignals in the classifier and combines CNN with heart sound feature extraction methods to propose a novel method based on sports training health analysis algorithm based on heart rhythm feature extraction and convolutional neural network. It greatly improves the accuracy of the recognition and prediction of the athlete's training health status.

e data used to support the findings of this study are included within the article.

e authors declare that they have no conflicts of interest. 

How the negative and positive aspects of partner relationships affect the mental health of young married people

Negative impact of the first COVID-19 lockdown upon health-related behaviours and psychological wellbeing in people living with severe and complex obesity in the UK

e negative impact of poverty on the health of women and children

e myth of "older LGBT+" people: research shortcomings and policy/practice implications for health/care provision

Complaints as indicators of health care shortcomings: which groups of patients are affected?

A review of the relationship between sports training and medical health

A review of settingsbased health promotion with applications to sports clubs

Design and characterization of a textile extension sensor for sports and health applications

Activity recognition using inertial sensing for healthcare, wellbeing and sports applications: a survey

iPhone sensor platforms: applications to sports monitoring

Lessons on data collection and curation from the NFL injury surveillance program

State of the art of sports data visualization

A survey of competitive sports data visualization and visual analysis

A visualization template for the graphical representation of sport injury antecedents and consequences models and data

Design and implementation on patchouli growth environment monitoring system based on IoT

Self-monitoring method for improving health-related quality of life: data acquisition, monitoring, and analysis of vital signs and diet

Prediction of cardiovascular diseases based on machine learning

A review of research on cotraining. Concurrency and computation: practice and experience

Robotic grasp detection using a novel two-stage approach

Immature apple detection method based on improved Yolov3

A biomimetic covering learning method based on principle of homology continuity

Deep learning based cell classification in imaging flow cytometer

Polynomial fitting algorithm based on neural network

Evaluation of the functional movement screen as an injury prediction tool among active adult populations

Secure prediction and assessment of sports injuries using deep learning based convolutional neural network

Sports participation of children with or without developmental delay: prediction from child and family factors

A big data analytical framework for sports behavior mining and personalized health services

Faster Mean-shift: GPUaccelerated clustering for cosine embedding-based cell segmentation and tracking

Cloud shape classification system based on multi-channel cnn and improved fdm

Alzheimer's disease multiclass diagnosis via multimodal neuroimaging embedding feature selection and fusion

EEG-based driver drowsiness estimation using an online multi-view and transfer tsk fuzzy system