Design of Smart Brain Oxygenation Monitoring System for Estimating Cardiovascular Disease Severity

Clinically, cardiovascular disease (CVD) patients need physicians to suggest different exercise training and rehabilitation procedures to improve their cardiopulmonary function (CPF). In previous studies, several approaches, such as cardiopulmonary exercise testing (CPET), echocardiography and computed tomography angiography (CTA), were proposed to indirectly estimate the rehabilitation effect on CPF. However, the above approached require experienced operators and complex equipment. In this study, a smart and wearable brain oxygenation monitoring system without motion artifact and crosstalk is proposed to estimate the blood circulation state of brain tissue directly during incremental exercise. Moreover, the technique of neural network is also used for classifying different CPF groups from the indexes extracted from the measured hemoglobin parameters. The experimental results show that the defined indexes extracted from the hemoglobin parameters can present the state of CPF, and the proposed smart brain oxygenation monitoring system can also effectively and automatically classify different CPF groups from these indexes via artificial intelligence. The proposed system therefore may assist physicians in the clinical evaluation of the CVD severity and rehabilitation effect on CPF in the future.


I. INTRODUCTION
Cardiovascular diseases (CVD), including cerebrovascular disease, heart diseases and vascular diseases, are one of top ten causes of death globally. World health organization (WHO) estimated that approximate 17.3 million people died from CVD per year, and the expected number of people died from CVD might increase to over 23.6 million by 2030 [1]. The previous study indicated that one of current methods of CVD prevention and treatment was exercise training and rehabilitation, which could reduce the morbidity of cardiovascular (CV) events and improve the athletic ability, quality of life and life expectancy of patients [2]. However, The associate editor coordinating the review of this manuscript and approving it for publication was Yue Zhang . patients with different CVD severities require different prescriptions of training and rehabilitation programs. For this reason, how to estimate the CVD severity of the patients is useful for the treatment procedure.
In clinical practice, several approaches are used for estimating the CV function of the patients, including cardiopulmonary exercise testing (CPET), echocardiography, and computed tomography angiography (CTA). CPET is a clinical test by exercise that measures the heart ability in a controlled clinical condition and provides valuable diagnostic and prognostic information for patients with CV and pulmonary disease by using various parameters, such as oxygen uptake (V O 2 ), carbon dioxide output (V CO 2 ) and ventilation (V E ) of gas exchange, electrocardiogram (ECG) findings and symptoms [3]. However, the above method contains some disadvantages, including the inconvenience of using the gas mask to measure gas parameters, and the influence of ECG electrode sliding caused by sweat during exercise. Echocardiography is a noninvasive imaging technique used for detecting cardiac structures and cardiac function assessment via ultrasound reflection, and calculates cardiac blood flow and tissue velocities through Doppler equation to provide a cardiac diagnosis of the patients [4]. However, this method requires more expensive cost and professional clinical staff to operate. Moreover, tissue Doppler imaging is angle dependent. If the angle of interrogation exceeds 20 degrees, the tissue velocity may be underestimated [5]. CTA is a noninvasive angiography technique for coronary arteries that assesses obstructive coronary artery diseases (CAD) via electron beam tomography. CTA results will be classified as significant CAD (50% luminal narrowing), mild CAD (<50% stenosis) and normal coronary arteries. However, asymptomatic patients are generally not suggested to do screening test via CTA due to its significant radiation and contrast administration [6].
Near infrared spectroscopy (NIRS) is a noninvasive technique used for estimating the change of human tissue components, such as the hemoglobin concentrations, and has been widely applied in many biomedical applications [7]. In 2009, Waxman et al. proposed a novel catheter-based NIRS system to detect lipid core coronary plaques in the coronary [8]. In 2012, Rao et al. used NIRS to monitor blood oxygen parameters of patients during CPET to predict the anaerobic threshold (AT) obtained by CPET [9]. The previous study also indicated that NIRS could provide non-invasive continuous assessment of the regional circulations of the venous side, and approximate global measurement and organ-specific to help the detection of circulatory abnormalities [10]. The technique of NIRS might contain the potential of estimating the CVD severity via real-time monitoring the state of blood circulation in the brain tissue. Based on the technique of NIRS, a smart brain oxygenation monitoring system for estimating cardiovascular disease severity of CVD patients is proposed in this study. Here, the cardiovascular disease severity is classified according to the peak metabolic equivalents (METs) of patients under CPET [11], [12].
In this system, a wearable optical headband is designed to non-invasively assess the blood circulation state of brain tissue during incremental exercise. From the changes of blood circulation and the peak METs, several meaningful indexes are defined and obtained to estimate the severity of CVD patient indirectly. Moreover, the technique of radial basis function neural network (RBFNN) is used for classifying patients with different CVD severity. Here, the cardiopulmonary function (CPF) estimated from CPET is used for defining the groups of different CVD severity. From the experimental results, it shows the defined parameters are significantly related to the CPF state and this proposed system can also evaluate the CPF state automatically during incremental exercise from these indexes via artificial intelligence. In the blind test stage, it shows the proposed system can provide a good performance on classification of different groups. Therefore, it may be a novel and simple method to classify different CPF groups automatically in the clinical literature.

A. DESIGN OF SMART BRAIN OXYGENATION MONITORING SYSTEM
The hardware architecture of the proposed smart brain oxygenation monitoring system is shown in Fig. 1(a). It contains two parts, including a wearable optical headband and a smart platform. Here, the wearable optical headband consists of a pair of optical probes, a wireless optical signal acquisition module, and a mechanical design of headband, as shown in Fig. 1(b). These optical probes contain multi-wavelengths light emitting diodes and photodiodes to provide a light source to transmit the near-infrared light into the brain tissue and receives penetrating light respectively. When the incident light transmits through the medium, the medium will cause the light attenuation due to light absorption and scattering. In this case, the relative concentrations of multiple absorbers may be estimated from the intensity change of penetrating light, caused by absorption and scattering, via the modified Beer-Lambert law (MBLL) [13], [14]. The change of the absorber concentration will cause the optical density change of the penetrating light. The change of the optical intensity OD(λ) caused from the change of absorber can then be expressed by where I o (λ) and I i (λ) denote the intensity of the penetrating and incident light corresponding to wavelength λ respectively, ε λ is the molar extinction coefficient corresponding to wavelength λ [15], C denotes the change of molar concentration of the absorber in medium, and L is the distance between the light source and the light detector. The parameter B(λ) is the pathlength factor corresponding to wavelength λ [16]- [19], which is a correction factor used for correcting L.
Consequently, L · B(λ) can be viewed as the real light path in the medium. Here, both oxyhemoglobin (HbO 2 ) and deoxyhemoglobin (Hb) are main absorbers in human tissues in the wavelength range of near-infrared light [18], and (1) can be simplified as followings  [13], [14]. Finally the change of molar concentration of the total hemoglobin ( HbT ) and tissue oxygen saturation (StO 2 ) can be obtained from [HbO 2 ] and [Hb].
[HbT ] = [HbO 2 ] + [Hb] (3) In this proposed system, the 735 nm and 850 nm wavelengths [13], [20], that straddle the isosbestic point being the superposition of the absorption spectra of HbO 2 and Hb, are used in the proposed system.
In consideration of motion artifact and crosstalk between the probes [21]- [23], a time multiplexing is implemented in order to avoid crosstalk between the probes. Moreover, these optical probes are fixed on an adjustable and flexible rubber headband. By using the flexibility of the rubber headband, these optical probes can effectively maintain a good contacting condition between the optical probes and the head to reduce the influence of motion artifact.
The block diagram of the wireless optical signal acquisition module is shown in Fig. 2. It contains a wireless transmission circuit, a microprocessor, a LED driving circuit and a PD amplification circuit. The LED drive circuit is designed to drive LEDs by the control of the microprocessor to provide multi-wavelength light. The received light signal is then converted into a voltage signal by the PD and amplified by the PD amplifying circuit. The optical signal is then digitized by an analog-to-digital converter built in the microprocessor with the sampling rate of 200 Hz. Finally, the data of the optical signal are sent to the wireless transmission circuit to transmit to the smart platform. After receiving these data of optical signal, the smart platform will real-time estimate the change of relative hemoglobin concentration to evaluate the physiological state of the patients. Here, the transmission through the wireless transmission circuit fits the specification of Bluetooth v4.2. The wireless optical signal acquisition module can be powered by a 3.7V 400mAh Li-ion battery.
The design of the smart platform is based on a laptop with the operational system of Windows 10. In this smart platform, a real-time monitoring application is developed by Microsoft Visual C#, to provide the basic functions of displaying analysis results and storing the raw data obtained by wearable optical headband. Three event buttons in this application are also designed to record the times and states of activity events.

B. CLASSIFICATION OF CARDIOVASCULAR DISEASE SEVERITY
In order to estimate the CVD severity, the technique of neural network is used in this study. Here, the radial basis function neural network, which is a nonlinear data modeling and contains the advantages of a faster training procedure, a better approximation capability, and a simpler network configuration, is used in this study [25], [26]. The basic structure of RBFNN contains three layers (input, hidden, and output layers), as shown in Fig 3, which contains N 0 input neurons, N 1 hidden neurons, and an output neuron. Here, standardized hemoglobin parameters are used as the input of RBFNN. Then, the output Y (n) of RBFNN at n iteration can be expressed by where Z (n) = Z 1 (n), Z 2 (n), . . . , Z N 1 (n) T is the output vector of the hidden neurons, and the output Z k (n) of the k-th neuron can be calculated by the Gaussian basis function 98424 VOLUME 8, 2020 where I (n) and σ (n) denote the vector and its standard deviation of the input neurons and CT k (n) is the center vector of the k-th hidden neuron at iteration n. The operator · denotes the Euclid norm between two vectors. W (n) = W 1 (n), W 2 (n), . . . , W N 1 (n) T is a weight vector, and here, W k (n) is the weight between the k-th hidden neuron and the output neuron. The machine learning procedure of RBFNN contains the training and blind test stage. In the training stage, the machine learning of the center vector CT k (n) is an unsupervised learning (k-means clustering algorithm [26], [27]). The center vectors in the hidden neurons can be viewed as the prototype characteristic vectors taken from the training sets. The machine learning of the weight vector W (n) is a supervised learning (normalized least mean square algorithm [26]). In the normalized least mean square algorithm, the desired signal for the better CPF group and the poorer CPF group are defined as 1 and 0, respectively. In the blind test stage, if the output of RBFNN is larger than the given threshold, it will be classified into the better CPF group; otherwise, it will be classified into the poorer CPF group.  Table 1. In this experiment, the patients are instructed to wear the smart brain oxygenation monitoring system and perform CPET [3]. CPET is divided into three parts, namely warm-up exercise, incremental exercise and cool-down exercise [28]. According to their CPET results, these patients are divided into two groups of the better CPF group (peak METs 5; 23 males and 1 females) and the poorer CPF group (peak METs < 5; 14 males and 11 females).

A. CHANGES OF HEMOGLOBIN PARAMETERS CORRESPONDING DIFFERENT GROUPS
The change of hemoglobin parameters (StO 2 , and relative HbT concentrations) under incremental exercise are shown in Fig. 4. In this study, several indexes related to the change of hemoglobin parameters under incremental exercise are defined to describe the state of blood circulation in the brain tissue of CVD patients, and their definitions are illustrated in Fig. 5. Index I is defined as the time duration from the start to the maximum value of the hemoglobin parameter under incremental exercise. Index II is defined as the variation between the initial value and the maximum value of the hemoglobin parameter under incremental exercise. Index III is defined as the average variation of the hemoglobin parameter under incremental exercise, i.e. the sum of the hemoglobin parameter variation is divided by the time duration between the initial value and the maximum value of VOLUME 8, 2020 the hemoglobin parameter. Index IV is defined as the average variation of the hemoglobin parameter under first oneminute incremental exercise. Index V is defined as the average variation of the hemoglobin parameter between 60% of the whole incremental exercise duration and the time of the maximum hemoglobin parameter. Fig. 6(a) - Fig. 6(e) show the experimental results of Indexes I-V for different hemoglobin parameters corresponding to different groups. Except the hemoglobin parameters of Index IV for the better CPF group are lower than that of the poorer CPF group, most of Indexes I, II, II and V related to the change of StO 2 , and relative HbT concentrations for the better CPF group are significantly higher than that of the poorer CPF group.

B. CLASSIFICATION PERFORMANCE OF CARDIOVASCULAR DISEASE SEVERITY
In this study, most of Indexes I, II, III and V related to the change of StO 2 and relative HbT concentration are significantly different between the better CPF group and the poorer CPF group. In this case, Indexes I, II, III and V of StO 2 and relative HbT concentrations for both of left and right foreheads are used as the input of RBFNN to classify different CPF groups. Before the classification, the optimal threshold has to be determined. Here, the threshold is set from 0 to 1 by the step of 0.1, and the numbers of the input and hidden neurons are set to 8 and 64, respectively. In order to evaluate the classification performance, several parameters for binary test have to be defined first. True-positive (TP) means better CPF is correctly identified as the better CPF group; Falsepositive (FP) means poorer CPF is incorrectly identified as the better CPF group; True-negative (TN) means poorer CPF is correctly identified as the poorer CPF group; and Falsenegative (FN) means better CPF is incorrectly identified as the poorer CPF group. Moreover, f-measure, which is the harmonic mean of recall and precision, and accuracy are used for finding the optimal threshold and can be calculated by where recall = TP TP+FN is also called sensitivity, and precision = TP TP+FP is also called positive predictive value (PPV).
In the training stage, 39 trials are used for training, and the classification performance corresponding to different thresholds are shown in Fig. 7. The optimal performance (f-measure = 79.07%, sensitivity = 70.83%, PPV = 89.47%, accuracy = 81.63%) can be obtained when the highest value of f-measure occurs and the threshold is set to 0.4 respectively. In the blind test stage, 10 trials are used, where the proposed system can provide a good performance on the classification 98426 VOLUME 8, 2020  of different groups. The performance in the blind test is shown as Table 2. The accuracy, sensitivity and PPV of this proposed system are 80%, 75% and 75%, respectively. The output values of RBFNN corresponding to different groups are shown in Fig. 8. The output value of the better CPF group (0.504 ± 0.158) is significantly higher than that of the poorer CPF group (0.296 ± 0.049; p = 1.566E-4).

IV. DISCUSSION
From the experimental results in Fig. 6(a) and Fig. 6(b), Indexes I and II related to the change of StO 2 and relative HbT concentration of the better CPF group are higher than those of the poorer CPF group. These results indicate that the poorer CPF group more rapidly reaches the maximum value of these hemoglobin parameters, and the variations of these hemoglobin parameters of the poorer CPF group are less than the better CPF group. The relative HbT concentration is related to the changes in blood volume in tissue and provides an indirect indication of blood flow and perfusion [29].
This phenomenon may be explained by that the cerebral blood flow of the poorer group is relatively lower. In 2017, Smith et al. collected 11 studies related to the investigation of cerebral blood flow during normoxia, acute hypoxia and chronic hypoxia, and the results showed that average relative of cerebral blood flow response was the greatest during exercise at moderate (40-70% W max ) intensity [30]. For this reason, the time duration between 60% of the whole incremental exercise and the time of maximum hemoglobin parameter might correspond to the intense stage of incremental exercise. From the experimental results in Fig. 6(c) and Fig. 6(e), the average variation of hemoglobin parameters of the better CPF group during the whole stage and the intense stage of incremental exercise, is higher than that of the poorer CPF group. This phenomenon of the poorer CPF group may be explained by that the cerebral circulation is relatively insufficient, especially in cardiac patients whose cardiac output fails to increase normally [31] and blood flow velocity is relatively slow. Moreover, Fig. 6(d) shows that the average variation of some hemoglobin parameters for the better CPF group is significantly lower than that of the poorer CPF group. This phenomenon may be explained by that the difference of the cerebral blood flow for different groups in the early stage of incremental exercise is unobvious.
In previous studies, several approaches were also proposed to evaluate CPF of the patients, and the comparison between the proposed system and other systems is summarized in Table 3. Here, CPET, which contains multi-sensors, including gas mask, ECG electrodes and blood pressure monitoring, to non-invasively measure V O 2 , V CO 2 , V E , ECG, heart rate and blood pressure, can be used for evaluating CPF of the patients [3]. Before each exercise test, CPET has to be calibrated the gas exchange system by the experienced operators. Moreover, using the gas mask to measure gas parameters would easily affect the breathing activity of the user, and the ECG quality is also affected by ECG electrode sliding caused by sweat during exercise. The larger system volume, the setup of multi-sensors and the requirement of the experienced operators will cause the inconvenience of use. Echocardiography can non-invasively assess the 2D ultrasonic reflection images of cardiac structures and cardiac function via ultrasound detector, and estimates the velocity of blood flow via Doppler effect [4].
The setup is not complicated but this approach also requires the professional clinical staff to operate. Moreover, it is easily influenced by bones, and tissue Doppler imaging is angle dependent. If the angle of interrogation exceeds 20 degrees, the velocity of tissue blood flow may be underestimated [5]. CTA can be used for non-invasively assessing obstructive CAD via electron beam tomography, but the system is huge and the setup convenience is also lower. Most important of all, asymptomatic patients are not suggested to perform screening test via CTA due to its significant radiation and contrast administration [6]. Different from the above systems, the proposed system contains the advantages of small size, wearability, low cost and ease of use and can access the local blood circulation state of brain tissue directly and evaluate the CPF state automatically during incremental exercise. The experimental results show that the defined parameters are significantly related to the CPF state and the proposed system can also effectively classify different CPF groups.

V. CONCLUSIONS
In this study, a smart brain oxygenation monitoring system without motion artifact and crosstalk is proposed to estimate the CVD severity of the patients. By using the technique of NIRS, the proposed system can non-invasively and real-time access the change of relative hemoglobin concentrations in the brain tissue directly. Here, because of a limited measuring depth from this proposed system, it may be slightly affected by the individual difference of brain structure. From the change of relative hemoglobin concentration under exercise, several indexes are also defined to describe the state of blood circulation in the brain tissue. From the experimental results, most of Indexes I, II, III and V related to the change of hemoglobin parameters (StO 2 and relative HbT concentration) for the better CPF group are significantly different from that of the poorer CPF group. Moreover, the CPF state can be evaluated automatically during incremental exercise from the defined indexes via artificial intelligence. In the future, the proposed system may contain the potential of assisting the physicians in the clinical evaluation of CVD severity and rehabilitation effect on CPF. Moreover, the proposed system may provide more information of rehabilitation effect on brain activation besides clinical observation in other rehabilitation applications.