2.1.

Fiber Bragg Gratings

The sensor is designed using a Bragg grating inscribed into a single-mode optical fiber as the sensing element. FBGs have been one of the most widely researched optical components within the last decade. They have found manifold applications in telecommunications, laser, and sensor technology. The sensor applications of FBGs alone cover many fields, such as structural health monitoring,⁴¹ monitoring of conditions in electrical plants,⁴² and physiological activity monitoring.⁴³ FBGs owe their popularity to their immunity to optical signal intensity modulation, which is possible on account of their spectral encoding as well as excellent multiplexing and self-referencing capabilities, thus allowing for the addressing of several sensors with one optical cable. At the same time, they offer the standard benefits of fiber-optic technology such as EM immunity, intrinsically safe modes of operation, a chemically inert nature, small size, and light weight.

The FBG acts as a stop-band filter, except that part of the incident light, $I_{\mathrm{in}}$, is not stopped but is reflected from periodically arranged layers formed by refractive index modulation (see Fig. 1). The reflected beams from each of the layers destructively interfere with each other unless they are all in phase. This only occurs at one wavelength, known as the Bragg wavelength (or the central wavelength), ${\lambda }_{\mathrm{B}}$, which is given by

Fig. 1

Principle of operation of an FBG.

The light reflected from the grating, $I_{\mathrm{re}}$, assumes the spectral form of a sharp “peak” located at the Bragg wavelength, whereas the transmitted light, $I_{\mathrm{tr}}$, is characterized by a narrow “dip” at the same wavelength. In most sensing applications, the central wavelength is determined on the basis of the peak position, as the reflected signal shape can be more easily demodulated. Thus, FBG-based sensors offer the unique advantages of a single-ended connection to the interrogation system.

2.2.

Sensor Construction

A starting point from which to build the sensor for monitoring respiration and heart activity was the susceptibility of the FBG to strain, which produces a shift in the Bragg wavelength, $\mathrm{\Delta }{\lambda }_{\mathrm{B}}$, and that is described by the following relation:⁴¹

The sensor design is based on the approach in which a springboard is used to convert body movements, including lung- and heart-induced vibrations, to strain, and an FBG attached to the board is used to measure this strain. The instantaneous spectral position of the FBG reflection peak can be calibrated in terms of respiration and heart activity readings. For the sensor to work properly, the board has to be placed between the bed mattress and the back of the patient.

As the base of the sensor, a polymethyl methacrylate (PMMA) (i.e., Plexiglas) board with dimensions of ${220\times 95\mathrm{mm}}^2$ and a thickness of 1.5 mm was used. A springboard of this size and thickness provides proximity to the body without a loss of comfort for the patient lying on it. At the same time, the board size ensures both that the body vibrations are sufficiently transferred to the FBG and that the vibrations can be decoded in the form of FBG elongations/contractions. The FBG, with ${\lambda }_{\mathrm{B}}=1541\mathrm{nm}$ at a temperature of 20°C, was attached with epoxy adhesive to the center part of the board, as shown in Fig. 2. After adhesive curing, the peak position shifted in the direction of higher wavelength values; its full width at half maximum increased slightly, and the peak amplitude decreased. One of the FBG pigtails was spliced with a pigtail that ended with an angled physical contact fiber connector connector capable of connecting the interrogation system. The splice protector was equipped with a ceramic reinforcing rod, instead of a steel one, in order to allow the sensor to be used in an MRI environment. The excess fiber between the FBG and the splice protector was formed into a loop, consisting of several turns, and was taped to the PMMA board. The other FBG pigtail was cut off a few centimeters from the grating and also taped to the board.

Fig. 2

Construction of the sensor.

2.3.

Experimental Setup

A block diagram of the experimental setup is depicted in Fig. 3. The study was conducted at the Military Institute of Aviation Medicine, Warsaw, Poland, equipped with the Achieva 1.5 T MRI scanner by Philips, Andover. The sensor was placed in the MRI examination room between the mattress of the track table in the MRI system and the back of the lying patient in the area of the shoulders to ensure proximity to the heart and the lungs. The sensor lead, in the form of a 10-m optical cable, was threaded through an opening in the wall and connected to the FBG interrogation system in the MRI system operator’s room. To interrogate the sensor, a commercially available sm130-700 unit from Micron Optics, Atlanta was used.⁴⁵ The device consists of two main modules: a broadband light source to illuminate the FBG, and a high-speed scanning Fabry–Pérot (FP) tunable filter to analyze the signal reflected from the grating. Using a FP tunable filter is the most common technique for monitoring an FBG; any changes in the Bragg wavelength located within the free spectral range of the FP filter can be tracked by the spectrally scanning filter.²³ The sm130-700 interrogator allows for scanning a band of 1510 to 1590 nm with a frequency of 1 kHz and a resolution of $1{\mathrm{pm}}_{\mathrm{pk}-\mathrm{pk}}$, which in turn, according to Eq. (3), ensures measurement of strains as low as $0.8\mu {\epsilon }_{\mathrm{pk}-\mathrm{pk}}$. The data from the interrogator were sent via the Ethernet interface to a laptop PC with software developed for signal visualization, data archiving, processing, and analysis.²²

Fig. 3

Experimental setup.

The experimental setup was equipped with respiration and electrocardiogram (ECG) recorders from Invivo, Gainesville,⁴⁶ designed to operate in the MRI environment, as a source of signals used to determine the reference RR and HR. The respiration and ECG signals were acquired from the patient using pneumatic bellows placed on the patient’s chest and carbon electrodes fixed to his or her chest, respectively, and transmitted from the recorders to a PC wirelessly via a 2.4 GHz radio link. As the data were transmitted in two different ways (i.e., by Ethernet and radio), to minimize the possible occurrence of a shift between the signals, they were synchronized by means of motion artifacts introduced in the reference signals and the sensor signal at the beginning and the end of the recording.

2.4.

Subjects

Twelve adult volunteers were involved in the experimental studies (four females, eight males); age: $33.25\pm 6.06$ ($\text{mean}\pm \mathrm{SD}$)years, weight: $78.42\pm 17.43\mathrm{kg}$, and height: $177.17\pm 8.54\mathrm{cm}$. When the sensor was placed under the back of each patient in the same manner, no significant correlations were observed between the quality of the received signal and the age, weight, or height of the subjects. Any differences in the representation of the sensor signal are instead a result of individual features of the subjects, such as the size of internal organs or the pattern of blood vessels.

3.1.

Frequency Response

For the purpose of demonstrating the system’s frequency response, the sensor was vibrated by a modified high-power loudspeaker driven by a signal generator, as shown in Fig. 4. A reference FBG was mounted on the flexible cone suspension of the loudspeaker to provide a strain measurement that is proportional to the cone excursion. This allowed the sinusoidal vibration to be set to a constant amplitude within the measurement range. The signal amplitude recorded by the sensor that was referred to the signal amplitude measured by the reference FBG is shown in Fig. 5 in the form of transmittance corresponding with the frequency response of the measurement system. The sensor performance within the range of the ballistocardiographic (BCG) signal being measured (i.e., 1 to 20 Hz^17,47) shows that signals are sensed without any degradation of the physiological information. The BCG frequency range lies clearly below the resonant frequency of the sensor.

Fig. 4

Setup for determining the system’s frequency response.

Fig. 5

Frequency response of the measurement system.

3.2.

Determining RR

The main phase of the experiments covered clinical evaluation of the sensor’s RR and HR measurement capabilities, in real MRI conditions. Each of the subjects was asked to lie down on the track table of the MRI scanner, and the physiological activities of their bodies were recorded during a 5-min exposure that was part of a thoracic spine examination. Thus, the total recording time was 60 min. A 14-s portion derived from the entire recording made on a 36-year-old man (weight: 86 kg; height: 183 cm) is shown in Fig. 6. The upper trace represents the respiration curve recorded with the reference module for respiration monitoring, and the lower one shows the corresponding part of the signal coming from the sensor (shown in gray). Additionally, the original signal was denoised through an averaging of 250 samples (with the result shown by the black line).²³ The slowly changing “waving” pattern in the sensor signal reflects the activity of the lungs. The arrows indicate the maximum values of the signal during subsequent breaths used by the authors as distinctive points for determining an RR trace. In this way, $n$ local maxima were extracted for the processed 5-min sensor signal. The RR value at the time $t_n$, at which the $n$’th maximum occurred, is defined by the formula

Fig. 6

14-s portion of a recording with visible respiration wave.

Figure 7 shows the 5-min RR traces obtained for each of the subjects, successively, using both (a) readings from the reference recorder and (b) the set of local maxima extracted from the sensor signal. In both cases, the reference RR trace and that obtained on the basis of the sensor signal are of similar shape, and therefore the Bland–Altman plot was used for quantitatively assessing the sensor’s RR measurement capability.^48,49 In this graphical method, the differences between the two techniques (i.e., $x_1-x_2$) are plotted against the averages of the two techniques [i.e., $(x_1+x_2)/2$]. The better the reproducibility is considered to be, the narrower the $\pm 1.96$ SD range containing approximately 95% of the results is. The accuracy of determining the characteristic points from the sensor signal in relation to those determined from the reference signal is shown in Fig. 8(a) for the total pool of data (all subjects). Among the values, 96.10% lie within the limits of agreement (LoA); the results show no systematic error, have no proportional character, and do not depend on the frequency of the distinctive point occurrence. Going further, the sensor’s RR determination capability is estimated in Fig. 8(b), in which 95.02% of the values lie within the LoA of 0.89 respirations per minute (rpm). In contrast to the Bland–Altman assessment of the characteristic point determination, the sample distributions for the RR determination clearly show the dependence on the RR values. This can be explained by the fact that the error in determining characteristic points that is independent on the RR translates into the error in determining the RR, which increases with RR, according to Eq. (4).

Fig. 7

Respiration rate extracted from (a) reference signal and (b) sensor signal shown successively for all of the subjects.

Fig. 8

Determination of (a) distinctive points and (b) RR assessed using the Bland–Altman method.

The sensor’s RR measurement capability is also estimated in Fig. 9 by calculating the differences between the reference and measured RR values in relation to the average value between them. The relative error span is below 7% [mean relative error (MRE) of RR detection: $5.24\%\pm 1.02\%$], which in relation to the number of inhalations per minute gives a value of approximately $\pm 1.2\mathrm{rpm}$ [mean absolute error (MAE) of intervals between respiration events: $0.65\%\pm 0.30\%$]. The results of the sensor’s RR measurement capability evaluation are summarized in Table 1 for each of the subjects as well as for the total pool of data.

Fig. 9

Relative error of RR determination as a fraction of the average value.

Table 1

Results of the sensor’s RR measurement capability evaluation.

3.3.

Determining Heart Rate

The cyclic disturbances seen on the envelope of the sensor signal shown earlier in Fig. 6 represent the activity of the heart. Hence the following stage of experimental studies included sensor signal processing in the context of the sensor’s HR measurement capability. For a better visualization of the heart activity and due to the need for HR detection, the sensor signal was normalized by reducing it to the reference level of zero through low-pass filtering with a cutoff frequency of 60 Hz, which did not cause a loss of physiological information. Figure 10 shows an enlarged 2.5-s portion of a recording made with the same male subject whose respiration traces were shown in the previous section. The upper trace represents the reference ECG signal (CM5 lead), whereas the lower one shows the corresponding part of the signal recorded by the sensor (gray line). In addition to the low-pass filtering, the signal was denoised by averaging over 25 samples (black line).²³

Fig. 10

2.5-s portion of a filtered signal with visible wave reflecting heart activity.

According to the literature,^50–53 the general waveform of the BCG consists of an early peak (H, Fig. 10) related to the motion of the heart early in systole, and the main IJK complex related to the ventricular ejection and aortic flow. It can be assumed approximately that the highest peak (J) occurs at the maximum value of the ECG T-wave. It is easy to see that the BCG signal acquired by the sensor is also a sort of a kinetocardiogram^54,55 in which the movement of the subject’s back in relation to the mattress of the track table is registered rather than the movement of the chest. However, the signal shape can also reflect the flow of blood in the organs of the body.

The arrows shown in Fig. 10 indicate the J peaks of the BCG signal during subsequent heartbeats used for determining the HR trace, in a manner similar to the RR detection procedure. Thus, $m$ J peaks (local maxima) were extracted for the processed 5-min signal from the sensor. The HR value at the time $t_m$, at which the $m$’th maximum occurred, is defined as follows:

Figure 11 shows 5-min HR traces obtained for each of the subjects using both (a) the ECG recorder and (b) the HR detection procedure performed on the sensor signal. Again, the sensor and reference traces have similar shapes, and the differences between them are better depicted in Fig. 12 by means of the Bland–Altman plot. First, the accuracy of determining the J peaks from the BCG signal in relation to the R waves determined from the ECG signal is shown in Fig. 12(a) for the total pool of data. 95.89% of the considered samples lie within the $\pm 1.96$ SD range. Subsequently, the sensor’s HR determination capability is estimated in Fig. 12(b). 95.11% of the values lie within the LoA, and this range is equal to 3.64 beats per minute (bpm). In the same way as the respiration data shown previously, the data for the distinctive point determination are relatively normally distributed (i.e., there are no systematic or proportional errors), and no dependence on the frequency of the distinctive point occurrence is observed, whereas the HR sample collections strongly depend on the HR value. Again, the error in determining characteristic points that is independent on the HR translates into the error in determining the HR (where the error increases with HR), according to Eq. (5).

Fig. 11

Heart rate waveforms for each of the subjects in sequence extracted from

Fig. 12

Determination of (a) J peaks and (b) HR assessed using the Bland–Altman method.

In addition to the Bland–Altman analysis, a fractional error was shown in Fig. 13 to characterize the usefulness of the proposed measurement method. The error varies within the range of 6.5% (MRE of HR detection: $5.71\%\pm 0.61\%$), which corresponds with approximately $\pm 3\mathrm{bpm}$ (MAE of beat-to-beat intervals: $2.40\%\pm 0.46\%$). Table 2 summarizes the results of the sensor’s HR measurement capability evaluation, both for individual subjects and for the total pool of data.

Fig. 13

Relative error of HR determination as a fraction of the average value.

Table 2

Results of the sensor’s HR measurement capability evaluation.

3.4.

Patient Comfort and Safety

At the end of each test, the subject was asked about any negative feelings caused by the presence of the sensor in his or her back, such as excessive rigidity or size of the board, leading to galls or bruises on the subject. None of the subjects reported any symptoms of discomfort caused by the sensor location; no one suffered any prejudice. Finally, the same male subject, whose respiratory and cardiac recordings are shown above in Figs. 6 and 10, was asked to undergo a complete thoracic spine MRI examination for the purpose of assessing the comfort of a subject remaining in the same position for a long time with his back in contact with the sensor. Figure 14 shows (a) RR and (b) HR traces obtained on the basis of the sensor signal acquired during a 39-min exposure. The subject does not need to lie down at a designated position to enable the sensor to acquire his physiological signals. He was able to simply remain in the supine position identical to that taken by patients during a normal spine MRI examination. The physiological signals were correctly acquired from the subject during the entire duration of the scan. The subject did not complain of any discomfort associated with the presence of the sensor in his back.

Fig. 14

(a) RR and (b) HR during a complete thoracic spine MRI examination.

High values of the EM field generated during the examination did not affect the optical signal propagated in the fiber and did not disturb the sensor operation even once. The clinical evaluation proved that the signal was identical to that obtained without exposure to the EM field. Moreover, the presence of the sensor during the MRI examination did not pose a threat to the patient and had no influence over the quality of the imaging. Figure 15 shows an image of the region of the body under which the sensor was placed; the proposed transducer is transparent to the MRI system. It was verified that the sensor does not introduce any artifacts to the spin-echo (SE) T1-weighted and T2-weigtened or to the gradient-echo (GE) T2-weighted imaging sequences.

Fig. 15

MRI image acquired from the sensor location: (a) sagittal view, SE T2-weighted, (b) axial view, GE T2-weighted.

3.5.

Signal Quality

The quality of the sensor signal depends to a large extent on the thickness of the clothing the monitored person is wearing over his or her torso. The thicker the layer of clothing is, the more distorted is the transmission of lung- and heart-induced vibrations to the sensing element. During the experimental studies, the subjects were wearing one or two layers of clothing (i.e., a shirt or T-shirt and a sweatshirt), and no difficulties in the physiological signal acquisition were observed. Although MRI patients usually wear not more than two layers of light, indoor torso garments (room temperature), there are plans for improvement of the system’s reliability by locating several FBGs at various positions inside the mattress that the patient rests on.^40,56 This treatment will also reduce the influence of other factors that can deteriorate the quality of the signal, such as the sensor arrangement under the patient, the force exerted by the lying patient on the sensor located behind him, and individual features of the patient. All of these elements determine the value of the signal-to-noise ratio (SNR), which for the respiration wave presented in Fig. 6 is specified as 18.27 dB with respect to the signal obtained during the 0.5-min period of apnea. Similarly, the SNR level for the BCG signal shown in Fig. 10 is equal to 3.37 dB, estimated on the basis of the normalized signal in relation to the same signal but with heart-induced disturbances removed.⁵⁷

One of the main features of ballistocardiography-based systems is their sensitivity to any body movements, which disturbs the signal carrying physiological information.⁵⁸ Figure 16 shows an example of a signal artifact induced by a short displacement of the track table. The artifact amplitude is five times higher than the signal amplitude produced by the heartbeat. Other movements such as coughs or changes to the body position may cause more significant distortions of the sensor signal. Therefore, BCG devices designed for clinical use usually have to be equipped with additional software or hardware modules that can eliminate or compensate for high-amplitude body movements. Software filtering is the simplest means by which to compensate, at least partially, for the disturbances. Special arrangements of equipment that eliminate the influence of high vibrations can be based, for example, on FBGs or accelerometers not registering vibrations produced by the activity of the lungs or heart. The output signals in these arrangements were used for compensating the influence of disturbances to the sensor output signal. However, the sensor in the presented version can be successfully used to monitor respiration and heart rate during an MRI examination, where the patient remains still for tens of minutes.

Fig. 16

Signal artifact caused by a short displacement of the track table.