2.1.

Sensor Design and System Configuration

The working principle behind this sensor is based on the theory of microbending optical fibers, which has been proven both theoretically and experimentally in previous studies.^12–15 Figure 1 shows the microbend fiber optic sensor system. This system consists of a sensor mat embedded with microbending multimode fiber, a transceiver, and a computer.

Fig. 1

Microbend fiber sensor system configuration.

The battery-powered transceiver weighs about 100 g and is comprised of a light-emitting diode (LED) light source operating at 1310 nm, a photodetector that is able to detect light signal in the range of 1100 to 1650 nm, amplifiers, simple low-pass filter operating at 250 Hz, an analog-to-digital converter, a microprocessor, and interface circuits for connecting the transceiver to a computer via Bluetooth. The sensor mat was constructed of two microbenders, a section of the multimode fiber, and plastic cover materials as shown in Fig. 2. Figure 3 shows the typical microscope examination of the micobenders (made of polyester fibers).

Fig. 2

Design of the microbend fiber optic sensor mat.

Fig. 3

Microscopic view of a typical microbender.

The optical fiber used for fabrication of the sensor is a standard graded-index multimode fiber with a core diameter of 100 μm and numerical aperture (NA) of 0.272 as the sensing fiber. The mat was constructed to a dimension of $25\times 20\mathrm{cm}$ for characterization. The thickness of the sensor mat is $<2\mathrm{mm}$ and only weighs $\sim 65\mathrm{g}$, enabling its portability and easily embedment into medical textiles or cushion, pillow, chair, bed, etc. The sensor mat, in response to breathing-induced body movement and heart beating-induced body movement, applies a force $\mathrm{\Delta }F$ or pressure $\mathrm{\Delta }P$ to the bent multimode fiber and cause the amplitude of the fiber deformation $X$ to change by an amount $\mathrm{\Delta }X$. The transmission coefficient $T$ for light travelling along the bent multimode fiber is changed by an amount $\mathrm{\Delta }T$ such that¹⁵

The microbending loss occurs as the guided modes are coupled to radiation modes. Effective coupling between the guided modes and radiation modes can be achieved with the spatial frequency $\mathrm{\Lambda }$ given by Eqs. (3) and (4) for graded-index and step-index multimode fibers, respectively. However, graded-index fibers are better than step-index fibers because graded-index fibers have resonance condition, where the microbending loss is sharply peaked while step-index fibers do not have the resonance condition.¹² So, we choose graded-index multimode fibers for our sensor mat in this article.

The graded-index multimode fiber we used is glass optical fibers. One potential problem with using glass optical fibers is that there is a risk to the users or patients when glass optical fibers break. The pieces of the glass optical fiber may puncture skin causing bodily injury. This is not acceptable in the application. Patient’s safety is the first priority. Another problem was that glass optical fibers can be easily broken. In biomedical applications, plastic optical fibers are preferred because plastic optical fibers are biocompatible. More importantly they are more rugged and enable greater safety than glass optical fibers. In this article, however, we use glass graded-index multimode fiber for proof of concept demonstration only. In product manufacturing, the use of plastic optical fibers is preferred.

The sensor design is based on microbending effect created through a “sandwich” microbender structure as shown in Fig. 2. Under mechanical perturbation $\mathrm{\Delta }F$ such as periodic body movements induced by breathing and cardiac biomechanical forces, the microbenders squeeze the multimode fiber and induce a series of microbends along the fiber axis. Microbending causes light coupling from core-guided modes into radiation modes, resulting in irreversible light loss and modulation in the light intensity detected by the photodetector inside the transceiver. By extracting the modulated light intensity, the BR, breathing waveforms, HR, BCG waveforms of the patient, and other body movement information within a specific time can be processed and measured using the digital signal processing (DSP) algorithm we have developed. The electrical output of the transceiver was wirelessly connected to a computer via Bluetooth sampled at a rate of 50 Hz. There was no crosstalk between this frequency and the grid frequency in the lab and hospital environment in Singapore. It is easy to change the sampling rate to other sampling rate whenever needed.

2.2.

Key DSP Algorithms

Figure 4 shows the general workflow in processing a measured signal using the algorithm. Smoothing and filtering are applied after receiving raw data from the transceiver followed by data quality check and peak detection in the time domain for calculation of the BR and HR. The final BR, breathing waveforms, HR, BCG waveforms were then displayed on the graphical user interface (GUI). LabVIEW (National Instruments Corporation, Austin, Texas) programming language was used to design this application software.

Fig. 4

Workflow in processing a measured signal using the algorithm.

Below are more detailed descriptions on the algorithms used for extraction of breathing component and heart-beating component from measured signals and for noise suppression.

2.3.

Sensor Characteristics

The proposed design of the microbend sensor is shown in Fig. 2. A graded-index multimode fiber with a core diameter of 100 μm and NA of 0.272 was used as the sensing fiber for construction of this sensor. The mat was constructed to a dimension of $25\times 20\mathrm{cm}$. Figure 5 shows a typical plot of sensor output $P$ as a function of weight $F$, where the optical signals have been converted into electrical signals. The force is applied uniformly over the area of $0.05{\mathrm{m}}^2$ within the sensor mat. This is a typical microbend sensor response characteristics similar to the results shown in Ref. 17. The key difference in the sensor response with the load is that the output of our microbend fiber sensor decreases slower than the sensors in Ref. 17. Our experimental data fits a Gauss curve perfectly, in which

Fig. 5

Sensor output as a function of weight of the sensor mat.

Figure 6 shows the sensor sensitivity curve as a function of weight based on the fitting data according to Eq. (6). For this specific sensor design, there are three distinct regions of sensitivity. The first region is the nonlinear region from 0 to 3.7 kg and the sensor sensitivity hits peak of $0.05\mathrm{V}/\mathrm{kg}$ at about 3.7 kg. The second region is the approximately linear region from 3.7 to 15 kg wherein the sensor sensitivity decreases. The third region is from 15 to 17 kg and then becomes zero after 17 kg. This zero sensitivity results mainly from no signals detected by the detector due to large optical loss induced. The force sensitivity of current design is good enough for detection of very weak signals, e.g., changes in cardiac force due to heart beating. No excessive bending is needed for low values of load.

Fig. 6

Sensor sensitivity as a function of weight of the sensor mat.

2.4.

Sensor Optimization for Various Applications

To avoid zero force sensitivity and to extend the working range of the sensor mat so as to cater for various body weights for various applications as shown in Fig. 7, sensor designs can be optimized to handle different ranges of applied force. Key design parameters are core diameter and NA of the multimode fiber as well as spatial frequency of mechanical perturbation along the fiber axis. The sensor performance shown in Figs. 5 and 6 is good for bed application in the lying position. The sensor mat is placed under the upper thoracic region (head-shoulder area) of the monitored subject. In this case, only part of the body weight is applied to the sensor mat. It should be noted that if the sensor mat is placed under a mattress with a 7-inch thickness, e.g., heart beat signal cannot be detected using current sensor mat whereas breathing signal is still detectable. This means that a very thick mattress placed in between our sensor mat and the subject can considerably obstruct the propagation of body vibrations by heart beating to the sensor mat, as our sensor requires body contact for measurement of biomechanical signals.

Fig. 7

Various applications of sensor mats: bed, pillow, chair, and standing application.

As for seating and standing application, significant body weight is applied to sensor mat. As can be seen from Figs. 5 and 6, the relationship between sensor output and applied force is not linear. The dynamic working range is $<13\mathrm{kg}$ in this case. To overcome this problem, we propose a “spacer technique” to extend the working range of the sensor mat. A dummy plastic tube is used as a spacer to lessen the impact of the applied force to the fiber so as to increase the sensor mat’s dynamic working range. The spacer is located with the sensing multimode fiber in between microbenders. Figure 8 shows the sensor output as a function of weight. It can be seen from Fig. 8 that with the spacer, the sensor’s dynamic range of operation is extended beyond 100 kg while the sensor could not work beyond 15 kg without the spacer despite using the same fiber. Another simple way to extend the working range of the sensor mat so as to cater for large applied force is to use 50- or 62.5-μm multimode fiber instead of 100-μm multimode fiber for given microbenders. We have demonstrated that a sensor mat made of 62.5-μm multimode fiber is good for the seating application.

Fig. 8

Sensor output as a function of weight.

A sensor mat with an area of $0.05{\mathrm{m}}^2$ was built for standing application (see the first photo from right in Fig. 7). In this application, the whole body weight is applied to the sensor mat. This sensor mat uses 62.5-μm multimode fiber. Figure 9 shows BCG waveform (acquired over a period of 6 s) in the standing application. Only the HR can be measured for this standing application using our current DSP algorithm. The impact to the sensor mat from breathing-induced body vibration is so weak that BR cannot be measured using our current algorithm. However, BR may be derived based on breathing-induced modulation on the BCG according to the phenomenon of RSA. We are currently in the process of developing a new DSP algorithm for BR measurement for the standing application.

Fig. 9

A 6-s acquisition of ballistocardiogram (BCG) waveform from a healthy subject in the standing position.

3.1.

Calibration Phase

In this phase, body weight, $W_{\mathrm{b}}$, is not applied to the sensor mat. Data is acquired in the preset period of 20 s. The mean of the curve, $W_{\mathrm{c}}$, is calculated.

3.2.

Measurement Phase

In this phase, body weight is applied to the sensor mat. Data is acquired in the preset period of 20 s. The mean of the curve, $W_{\mathrm{m}}$, is calculated. At the same time, arithmetic mean of the IJ and JK amplitude, $W_{ijk}$, is calculated.

3.3.

Estimation of Heart Beating Induced Force

In this phase, the heart beating-induced force, $F=W_{\mathrm{b}}/(W_{\mathrm{c}}-W_{\mathrm{m}})*W_{ijk}$. We have estimated the heart beating-induced force of two subjects using this method. The heart beating-induced force of a 51-year-old man and a 32-year--old man are 0.18 and 0.3 kg, respectively. A similar measurement method is applied for breathing-induced force measurement. As the acquisition of breathing signal is indirect in the standing position, we chose the seating position for the measurement. The breathing-induced force applied to the sensor mat is about 1 to 2 kg depending on shallow and deep breathings.

The local maxima in breathing and BCG waveforms showed in Fig. 12 are used for BR and HR calculation. The BR and HR are defined as

Table 1 shows the simultaneous measurement of the BR and HR from a 32-year-old healthy male subject (body weight: 65 kg) using our microbend sensor. The sensor mat was placed on top of a mattress and under a very thick bed sheet. The subject was asked to lie still (see the second photo from the left in Fig. 7). Measurements were recorded within 1 min (measurement every 6 s) and compared with a commercial physiologic monitoring device (BioHarness™, Zephyr Technology Corporation, Maryland) and a ${\mathrm{SpO}}_2$ device (OxyWatch™ Finger Pulse Oximeter). The maximum error detected was only $\pm 2$ beats per minute (bpm) for both BR and HR measurements.

Table 1

Experimental results comparing the breathing rate (BR) and heart rate (HR) measurements obtained from a healthy male subject within 1 min using our microbend sensor with commercial physiologic monitoring devices.

In the second experiment, the second subject is a healthy male of age, 28-year-old and weight 70 kg. Figure 15 shows BR and HR plots of the subject acquired using our sensor mat and a commercial device, Zephyr sensor. Zephyr sensor is a wearable sensor, which can be worn on the chest for BR and HR measurements. Key experimental results are summarized in Table 2. The recording time is 3495 s. Our sensor recorded 10 samples each second, while the Zephyr sensor recorded one sample for each second. The mean values of BR measured from both sensors are comparable. The absolute error of the BR mean values is $<2\mathrm{bpm}$ and the relative error is 9.7%. An accuracy of $\pm 2\mathrm{bpm}$ for the BR measurement should be acceptable for monitoring physiologic application. As for HR measurement, the mean values of HR measured from the two sensors are quite similar; the absolute error for HR mean values is $<0.5\mathrm{bpm}$ and the relative error is $<1\%$. The standard deviations for BR measurements are small, 0.95 bpm for Zephyr sensor and 1.38 bpm for our sensor mat, respectively. The standard deviations for HR measurement are about 3 bpm for both our fiber optic sensor and the reference Zephyr sensor. As can be seen in Fig. 15, our sensor and Zephyr sensor have different waveform shapes for BR. This is because our sensor takes a longer time or more number of local maxima for BR calculation. Zephyr sensor is aimed for sports application and warranted short time or less number of local maxima for BR calculation. However, in HR measurement, both our sensor and Zephyr sensor have similar waveform shapes. We have used four local maxima for HR calculation using our fiber optic sensor. It is not justified to calculate the “real time” difference between measured values from our sensor and Zephyr sensor because both sensor systems use different algorithms and average time for BR and HR calculations although we have synchronized both devices for recording. We have observed that there is a delay in displaying the same results between two sensors.

Fig. 15

Breathing rate (BR) and heart rate (HR) curves obtained from a healthy subject at rest in this study using our microbend fiber optic sensor system and Zephyr sensor.

Table 2

Summary of the experimental results for the second experiment.

Zephyr sensor is a device that requires direct body skin contact for physiologic monitoring, thus has less motion noise issues. Our sensor does not require direct skin contact, i.e., it can be embedded in bed and chair, etc. Thus, our sensor is very sensitive to motion noises. For example, the other body movement, such as changes in body position, can distort both the breathing waveform and the BCG waveform. Even very small body movements like muscle tremor, swallowing, etc. can also distort the BCG waveform. Although we have applied the algorithm for motion noise removal, we can still see the spikes in our results as shown in Fig. 15. Fortunately, these spike amplitudes are very high and their duration is short. These spikes did not affect our final results and simultaneous measurement of BR and HR.

We did the same experiment on another day. The results for the third experiment are shown in Table 3. Similar results were found.

Table 3

Summary of the experimental results for the third experiment.

We have also performed a preliminary clinical trial in the 3.0 Tesla (T) MRI environment to compare the performance of our microbend fiber optic sensor with conventional MR-compatible monitoring system (Respiratory bellows and Pulse Oximeter, Physiological Measurement Unit, Siemens Medical Solutions, Erlangen, Germany) (Fig. 16). The experiment was conducted during the MRI liver scans of 11 nonanesthetized healthy subjects of ages 26 to 62-years old (6 males and 5 females). The body weights of these subjects were in the range 46.6 to 80.0 kg (mean 60.3 kg). In this clinical trial, we were not allowed to interface with any MRI machine and devices. So, we had to manually record our experimental results. Our results demonstrated good agreement in simultaneous BR and HR measurements between our microbend sensor and the conventional physiologic monitoring system (Table 4, Figs. 17 and 18). Figure 19 shows the axial and coronal images obtained during MRI liver scan of a healthy subject under physiologic monitoring using our sensor. An examination and evaluation of all MR images obtained during physiologic monitoring using our microbend sensor and hospital equipment were conducted by using the same method presented in Ref. 19. They were quantitatively evaluated for the liver signal-to-noise ratio and liver-to-spleen contrast-to-noise ratio. In addition, MR image quality was qualitatively analyzed based on the five-point image scoring scale. Results showed no obvious image distortion and RF artifacts in the clinical report. This study suggested the feasibility in using our sensor mat for physiologic monitoring during MRI. Our sensor was able to detect a comparable BR and HR to the predicate devices and produce liver MRI images of good and comparable image quality to the prospective acquisition correction technique navigator-acquired scans in 3.0-T MRI environment.

Fig. 16

Experimental setup of the microbend fiber optic sensor mat in the magnetic resonance imaging (MRI) room. The sensor mat was placed under the thoracic region of all trial subjects for simultaneous monitoring of the BR and HR during liver MRI scan. Measurements obtained at every 5-min intervals were compared with conventional respiratory bellows (for the BR) and finger pulse oximeter (for the pulse rate).

Table 4

A good agreement between our microbend sensor and conventional MR-compatible devices in the mean measurements of the BR and HR recorded for every 5-min interval during a 30-min MRI scan session on 11 clinical trial subjects.

Fig. 17

BR detected and recorded for all 11 healthy subjects by each respiratory monitoring device during MRI showed maximum error of $\pm 1.7\mathrm{bpm}$ (standard deviation is 0.5 bpm).

Fig. 18

HR detected and recorded for all 11 healthy subjects by each HR monitoring device during MRI showed maximum error of $\pm 1.2\mathrm{bpm}$ (standard deviation is 0.4 bpm).

Fig. 19

(a) and (b) the axial and coronal images obtained during MRI liver scan of a healthy subject under physiologic monitoring using our microbend optical fiber sensor. No image distortion or radiofrequency artifacts were seen in all images obtained from the trial subjects, despite the presence of the sensor mat inside the MRI gantry, suggesting the MR-compatibility of the sensor.

As demonstrated in this article, our device works very well for normal subjects. As for abnormal breathing and heart beating cases, we believe our device can still apply. The key is that our sensor can recover correct breathing waveform and BCG waveform. We clearly observed the phenomenon of RSA, where BCG J-J interval was shortened during inspiration and prolonged during expiration. Peak detection in the algorithm is independent of regular or irregular beat–beat intervals. This shows our device can be potentially used for abnormal breathing and heart beating subjects.