A Noninvasive, Economical, and Instant-Result Method to Diagnose and Monitor Type 2 Diabetes Using Pulse Wave: Case-Control Study

Introduction

Background

Diabetes is becoming one of the most severe health problems in China. The World Health Organization indicated that type 2 diabetes accounts for around 90% of all the diabetes cases worldwide. In 2016, 9.4% of the population of China was diabetic, and 2% of all deaths in the country were because of diabetes and its chronic complications [1,2]. Type 2 diabetes and a variety of chronic complications in the middle and late stages of the disease develop over a long time. Therefore, early diagnosis, prevention, and treatment of type 2 diabetes are vital for reducing the medical burden and mortality rate. Diabetes is a long-standing example of a disease whose patients have been positively impacted by traditional Chinese medicine (TCM) [3]. Accurate diagnosis through the noninvasive, convenient, and economical techniques of TCM allows effective prevention and treatment of type 2 diabetes.

In TCM diagnosis, pulse taking is an important skill for diagnosing diseases by touching and sensing radial pulsations, but it frequently depends on the subjective consciousness and experience accumulation of the doctors. It is the lack of objective criteria that reduces the accuracy and repeatability of diagnosis. To overcome the shortcomings of subjectivity of traditional pulse diagnosis, in recent years, the importance of objective pulse-taking diagnosis has gained more and more attention. In addition, the study of diagnosis of some chronic diseases such as coronary heart disease and lung cancer by using the objective pulse parameters had some progress [4,5]. In the study of type 2 diabetes, some researchers have found that the changes of radial artery pulse wave are related to elevated blood glucose levels and major adverse cardiovascular events caused by diabetes mellitus [6,7]. Therefore, we hope that the noninvasive, convenient, and objective pulse information detection can be used to assist conventional methods to diagnose and monitor the occurrence and development of type 2 diabetes.

Objectives

The purpose of this study was to use a TCM pulse informatics analysis system to measure people’s objective pulse information and use that data to develop a model to diagnose type 2 diabetes. This instrument uses a pressure sensor to record pulse beats and display this information within a pulse wave. The pressure sensor is the most commonly used device for recording pulse wave of people’s radial artery at present. The parameters of the pulse wave are extracted and analyzed using statistical methods. This approach can quantitatively analyze pulse signals and provide more objective results than traditional pulse-taking diagnosis by TCM doctors.

This is the first study to use this method to analyze the differences in pulse wave parameters between healthy individuals and diabetic patients with hypertension and hyperlipidemia. The purpose of our research was (1) to find the association of the objective pulse information with type 2 diabetes and the disease with hypertension or hyperlipidemia and (2) to establish the prediction models of diabetes and its complications. We hypothesized that diabetes and its combination with hypertension and hyperlipidemia can be diagnosed reliably by using the pulse wave parameters.

Methods

Patients

This is a case-control study. Patients with type 2 diabetes (referred to as diabetes from now on) were recruited from outpatient services in 4 hospitals in Shanghai, including Yueyang Hospital of Integrated Traditional Chinese and Western Medicine, Shuguang Hospital, Shanghai Municipal Hospital of TCM, and Shanghai Qiangsheng Worker’s Hospital, between April 2012 and December 2018. The individuals in the healthy control group (group 1, n=50) were healthy staff and graduate students from the Shanghai University of TCM. The subjects were of Han ethnicity.

On the basis of blood pressure and serum lipid levels, diabetic patients were divided into 4 groups, including diabetic patients without hypertension and hyperlipidemia (group 2, n=139), diabetic patients with hypertension (group 3, n=133), diabetic patients with hyperlipidemia (group 4, n=70), and diabetic patients with hypertension and hyperlipidemia (group 5, n=75).

There was no significant difference in the age (P=.13) or gender (P=.59) between healthy control individuals (group 1) and all diabetic patients (group 2-5) (Table 1).

Ethics Approval

The study was approved by the Ethics Committee of Shanghai University of TCM in China in January 2012 and performed in accordance with the Declaration of Helsinki. All subjects had signed informed consent agreements.

Criteria

Diagnostic Criteria

The diagnostic criteria of diabetes were referred to as the Standards of medical care in diabetes [8].

Inclusion Criteria

The inclusion criteria were as follows:

1. Meet the diagnostic standard of diabetes
2. Age range is from 40 to 75 years
3. Hypertension or hyperlipidemia occur after diabetes.

Exclusion Criteria

The exclusion criteria were as follows:

1. Women during their pregnancy and lactation period
2. Mentally ill individuals
3. Those complicated with other diseases
4. Individuals who had acute metabolic disorders such as diabetic ketoacidosis or inflammatory complications.

Elimination Criteria

Patients without complete clinical data either because of incomplete collection or missing data were eliminated.

Collecting Methods for the Objective Parameters of Traditional Chinese Medicine Diagnosis

All objective parameters of TCM pulse diagnosis were collected by 2 MD doctors using the TCM pulse informatics analysis system (type: Smart TCM-I, product by: Shanghai Asia & Pacific Computer Information System CO, Ltd, Shanghai, China; Figure 1). In the analysis system, the instrument of detecting pulse wave information is a wristband acquisition terminal (Figure 2). The device consists of a pulse sensor, adapter, and acquisition software. The pulse sensor is attached to the arm by a wristband and connected to the adapter by a cable. The adapter has a universal serial bus (USB) connector, which can be directly connected to the computer through the USB interface. The acquisition software runs on the computer and realizes the acquisition and data management function of the pulse wave image by cooperating with the hardware (Figure 3).

The indoor temperature was 18°C to 25°C during pulse wave information collection. The subjects were either sitting or lying down for at least 3 min before they were tested. For pulse-taking diagnosis data collection, the forearms of the patients were extended forward naturally and at the same height as the heart. Their wrists were kept straight, with palms upward and fingers slightly bent. A small, soft pillow was placed under the wrist joint for better data collection. The pulse was measured at the radial artery corresponding to the inside of the styloid process of the radius in the left hand, the best position to feel the pulse. The frequency of the pulse acquisition device was 720 Hz. The sensor was tied to the position of the radial artery, and the pressure was adjusted by a knob. The acquisition was realized through observation and operation on the software interface. The software interface displayed the dynamic pulse waveform and pressure value in real time. The software recorded the pulse wave parameters for 1 min when the amplitude of the pulse wave reaches the maximum value. Afterward, the parameters of the pulse wave were extracted through the built-in software.

The data collected could be presented as a time-domain pulse wave (Figure 4), which includes the height of dominant wave (h₁), height of tidal wave (h₃), height of dicrotic notch (h₄), height of dicrotic wave (h₅), time distance between the start point of pulse wave and dominant wave (t₁), time distance between the start point of pulse wave and dicrotic notch (t₄), time distance between dicrotic notch and the end point of pulse wave (t₅), and width of percussion wave in its one-third height position (W).

The pulse wave parameters, including h₁, h₃, h₄, h₅, t₁, t₄, t₅, and W, have different physiological significances (Table 2).

Disease Prediction Methods and Statistical Analysis

The first objective was to analyze the differences in pulse wave parameters between healthy individuals and diabetic patients, and among the 4 subgroups of diabetic patients. One-way analysis of variance (ANOVA) was applied to the pulse wave parameters that follow the normal distribution, and non-parametric methods including the Mann-Whitney U test (for 2 groups) and the Kruskal-Wallis one-way ANOVA test (for 4 groups) were used for non-normal data.

The second objective was to classify diabetic patients from healthy individuals and to detect hypertension and hyperlipidemia among diabetic patients by using the pulse wave parameters. The following machine learning algorithms were applied:

• Logistic regression using all the pulse wave parameters as predictors
• Linear discriminant analysis (LDA) using all the pulse wave parameters as predictors
• Random forests with 500 decision trees, each tree samples two-thirds of the data, and each split in a decision tree randomly samples 3 out of 11 pulse wave parameters.
• Support vector machines (SVMs): We applied SVM with linear kernel (SVM-Linear) and SVM with 3-degree polynomial kernel (SVM-Poly). The cost and gamma parameters are determined through cross-validation, with trial values for cost being 0.001, 0.01, 0.1, 1, 5, 10, and 100 and trial values for gamma being 0.001, 0.01, and 0.1.

To estimate the prediction accuracy of the models (out-of-sample accuracy), we applied the 10-fold cross-validation technique.

‎View this figure

‎View this figure

‎View this figure

‎View this figure

Results

Difference in Pulse Wave Parameters Between Groups

Tables 3 and 4 show the ANOVA test and nonparametric test results on the pulse wave parameters between groups.

The results indicated that all diabetic patients (groups 2-5) had significantly higher h₃, t₁, and W than those in the healthy control group, whereas h₅ was significantly lower in diabetics when compared with healthy individuals. The parameters of pulse wave height, h₁, h₃, h₄, and h_5, were significantly different in groups 2 to 5. On the contrary, there were no significant differences in parameters of time distance, t₁, t₄, t₅, and W, among these groups. Pairwise comparisons indicated that the h₁ values of diabetic patients with hypertension (groups 3 and 5) were significantly higher than those of diabetic patients with hyperlipidemia (group 4). In addition, the h₁ values of diabetic patients with hypertension (group 3) were also significantly higher than those of diabetic patients without hypertension and hyperlipidemia (group 2). Diabetic patients without hypertension (groups 2 and 4) showed a significantly lower h₃ parameter than those of diabetic patients with hypertension (groups 3 and 5). The h₄ parameter of group 2 was significantly lower than that of groups 5. Similar to h₁ and h₃, h₄ values of groups 3 and 5 were also significantly higher than those of group 4. The value of parameter h₅ in group 5 was significantly lower than that of parameter h₅ in groups 2, 3, and 4.

Classification Algorithms

The accuracies of the classification algorithms are presented in Table 5.

Table 4 indicates that SVM with polynomial kernel (tuned parameters are cost=100 and gamma=0.01) has the highest prediction accuracy for diabetes, LDA has the highest prediction accuracy for hypertension among diabetic patients, and random forests has the highest prediction accuracy for hyperlipidemia among diabetic patients. The maximum accuracy, sensitivity, and specificity for each diagnosing task are presented in Table 6.

Discussion

Principal Findings

Pulse wave is the track of radial artery pulsation. It integrates a large amount of useful information about heart ejection activity and how the pulse wave travels along the vascular tree [9]. In this paper, time-domain analysis, one of the most frequently used techniques for pulse research [9], was applied to study the association between pulse parameters (including the height of pulse wave and the time of pulse phase) and type 2 diabetes to predict the attack of type 2 diabetes and its complications (hypertension and hyperlipidemia) for the first time.

Hyperglycemia is a leading cause of cardiovascular disease [10]. Therefore, the incidence of atherosclerosis is increased 2 to 4 folds in diabetic patients compared with nondiabetic individuals [11]. These diabetic patients experience hardened arterial blood vessels and decreased arterial elasticity, which increases peripheral resistance and the duration of high pressure in blood vessels. In this paper, we found that the parameters of h₃, t₁, and W increased significantly in diabetic patients compared with healthy individuals. From the result, we know that the high value of h₃ gives rise to weak elasticity and high peripheral resistance of arterial blood vessels. The high t₁ value reflects the weak elasticity of large vessels and high tensioning of small vessels, and the high W value denotes that high pressure in the artery is maintained for a longer period. Diversely, the value of h₅ was significantly lower in diabetic patients when compared with their healthy counterparts. The high h₅ value reflects the strong elasticity of the large artery. Our results from pulse time-domain analysis agreed well with the findings on patients and rats with diabetes [12,13].

Our analysis revealed that diabetic patients with hyperlipidemia had lower h₅ values compared with patients without hyperlipidemia, indicating that the arterial blood vessels were more rigid and less elastic in hyperlipidemia patients. These results are consistent with previous findings of high arteriosclerosis incidence and reduced hemodynamic functions in hyperlipidemia patients [14].

Our results also showed that h₁, h₃, and h₄ were relatively high in diabetic patients with hypertension compared with patients without hypertension, which can be explained by elevated arterial pressure because of high systolic and diastolic blood pressure in hypertension patients.

The prediction accuracy for diabetes is 96.35% (SVM-Poly). The algorithm can be used in mobile devices to conveniently and reliably diagnose diabetes. Predicting hypertension and hyperlipidemia among diabetic patients has low accuracy, 59.44% (LDA) and 69.77% (random forests), respectively. This may be explained by 2 reasons. First, the sample size for classifying hypertension and hyperlipidemia is relatively small (n=417). Second, hyperglycemia, hypertension, and hyperlipidemia can aggravate the microvascular and macrovascular lesions in diabetic patients [15], such as the decrease in vascular elasticity and the increase of intravascular pressure. Therefore, the differences between diabetic patients with hypertension/hyperlipidemia and patients without hypertension/hyperlipidemia may not be obvious. We plan to collect more samples, and we hope the accuracy of the models for identifying diabetic patients with hypertension/ hyperlipidemia may increase with the larger sample size.

Strengths and Limitations

With the growing maturity of sensors, chips, mobile internet, and other technologies, people’s health awareness is increasing and the demand for health services is greatly improved. The demand for wearable technology is becoming more and more high. Wearable and portable devices have shown great potential in the field of medical health. Pulse-taking diagnosis has several advantages such as its noninvasive nature and convenience. In our research, the pulse wave information acquisition element is a pressure sensor, which is the most commonly used device. Although there is another new type of sensor called an ultrasonic sensor used to detect pulse, the traditional pressure sensor is more cost-effective. In recent years, mobile phone apps are becoming more and more useful for the self-management of diabetes [16-18]. The wristband acquisition terminal we used also provides the possibility of integrating pulse parameter collection to mobile devices (such as sport bands and watches) and data analysis through mobile apps in the future.

The results from our study have confirmed the findings from previous studies [10-15]. Our methods showed a very high diagnostic accuracy for type 2 diabetes, whereas the diagnosis of hypertension and hyperlipidemia is not yet reliable. If we continuously improve the accuracy of using objective pulse information to detect type 2 diabetes in future research, the noninvasive and convenient pulse-taking diagnosis technique may have the potential to become a low-cost and accurate method to help people monitor the occurrence and development of type 2 diabetes and its complications more conveniently in daily life. It is important to note that in our research, when we used this device to collect pulse wave information, the subjects were asked to keep quiet and relaxed to measure the pulsation of the radial artery of their left hands under the appropriate temperature (as mentioned in the section Collecting Methods for the Objective Parameters of Traditional Chinese Medicine Diagnosis). Therefore any mobile device that uses our model will rely on the user being able to replicate the same condition. In the future, we hope to improve the convenience and stability of the device so that people can use it on both the left and right hands when moving. We are currently developing such devices and programs for the improvement of diabetes monitoring and remote diagnosis.