*3.1.3 Acoustic emission (AE)*

*Data Acquisition - Recent Advances and Applications in Biomedical Engineering*

Whilst phonoarthrography is based on the sound produced during the flexion or extension of joints, in VAG all vibrations produced during movement are considered [62]. Consequently, it is more common for a single accelerometer to be used as the sensor rather than a microphone [71]. It is also very common for signals in a frequency range below 1000 Hz to be of primary focus [72], with sampling rates of the order 1–4 kHz. A key advantage of the low sampling rate is that it allows for wireless data acquisition and processing using simple microcontrollers or singleboard computers [73, 74]. That said, it has been suggested [71, 75] that single signal processing may be limited and multi-channel recordings may lead to better discrimination of the severity and location of joint injury or disorder. In many cases noise mitigation is achieved through prefiltering (commonly using a bandpass filter from 10 Hz to 1 kHz) and amplification prior to digitization at a specified sampling rate [76, 77]. The digital signal may go through additional filtering, such as that conducted by Andersen et al. [78] who used a Kaiser-windowed finite impulse

There are other rationales for using multiple sensors during VAG as it has been

Therefore, the VAG signal is inherently nonstationary and potentially multicomponent in nature. The nature of the VAG signal means that it is not easily analysed using common signal processing techniques. This coupled with the difficulty in ascertaining the biological origin of the source of the signal is the main barrier to its use as a common diagnostic tool. As a result, much of the recent research activity has been focussed on feature extraction and statistical pattern classification [60]. Adaptive segmentation using least-square, linear prediction and autoregression algorithms is common [81, 82]. A host of statistical measures has been considered to characterise the VAG signal, including the form factors, skewness, kurtosis and entropy [71, 76]. It has also been shown that time-frequency distribution (TFD) [81, 83] and wavelet decomposition [84] are potentially powerful techniques for analysis and may negate the need for segmentation [83] but may be susceptible to noise [85]. These advancements have mostly been driven by developments in digital signal processing technologies that sped up analysis time as well as nonstationary signal analysis techniques developed for other biological signals like EEGs [84]. Using these techniques, spectral features such as frequency, energy and their respective spreads can be classified and linked to joint position, loading and pathology. The commonly used classifiers are neural network-based classifiers and support vector machines (SVM), as well as logistic regression and rule-based techniques [62, 71]. These neural networks and SVMs are supervised learning algorithms which search for a number of independent training data patterns taken from signals measured from participants with known pathologies to characterise new signals. These classification algorithms are increasingly dependable and can perform well with a limited amount of data. A number of different variants of these algorithms and classifiers have been investigated [60, 62]. Wu et al. [73] used an SVM based on the entropy and envelope amplitude features and achieved an overall accuracy of 83.56%. Nalband et al. [86] utilised an a priori algorithm with

observed that VAG may pick up vibrations not necessarily just due to the joint directly or to external interference [79]. For instance, the 10 Hz signal generated by the rectus femoris muscle which activates during the extension of the leg could interfere with the VAG signal recorded from the skin surface over the patella [80]. As this signal may vary in a similar fashion to the VAG signal, simple bandpass filtering may not be sufficient. It may be necessary to record the vibromyogram at the rectus femoris at the same time as the VAG signal and use adaptive filtering and

noise cancellation techniques to isolate the VAG signal [79].

*3.1.2 Vibroarthrography (VAG)*

response (FIR) bandpass filter.

**38**

AE for biomedical applications is derived from non-destructive techniques developed for detecting damage in engineering materials, such as metals and composites [88]. AE occurs when materials locally under stress emit energy in the form of transient elastic waves. This allows for the monitoring of microcrack initiation and propagation in the bones and joints [89]—essential parts of bone remodelling [90], and wear [91, 92]. Other characteristic sounds in joints, such as the bursting of gas bubbles in synovial joints during movement, can also be detected using AE [93]. AE frequencies are usually in the ultrasonic range and so detection often involves the use of ultrasonic sensors.

A number of researchers have proposed AE sensor-based joint monitoring systems using piezoelectric films, electret or MEMS-based microphones.

Toreyin et al. [94, 95] used an off-the-shelf low-noise MEMS microphone in conjunction with gyroscope and accelerometer pairs in order to monitor sounds generated during various complex motions. The microphone used had a sensitivity range of 100 Hz to 10 kHz, and the researchers suggested that the MEMS-based microphone had a similar performance to an electret microphone [94]. The acoustic data were sampled at 100 kHz, and the inertial data (monitoring joint angle and limb movement) at 1 kHz, with the data being collected by a field programmable gate array (FPGA)-based real-time processor. It was noted that air microphones do not exhibit signal losses due to motion artefacts, but they are sensitive to ambient noise.

Teague et al. [96] compared a piezoelectric film-based contact microphone to two air microphones: one electret and one MEMS-based. The air microphones were used with a 15 Hz high-pass filter and a second-order low-pass filter with a cut-off frequency of 21 kHz and sampled at 44.1 kHz using an acoustic recorder. The piezoelectric microphone was used with a 100 Hz high-pass filter followed by a fourth order low-pass filter with a 10 kHz cut-off frequency. It was sampled at 50 kHz using custom circuits. The 100 Hz high-pass filter was chosen to attenuate the motion artefact noise. It was noted that the electret and MEMS microphones performed similarly in detecting joint sounds, although the electret sensor was significantly more expensive. They were both sensitive to ambient and interface noise, including rubbing of the tape securing the sensors. It was noted that the air microphones did not need to be in contact with the skin. Experiments with sensors positioned 5 cm off the skin captured similar acoustic signals, albeit with lower amplitude. The piezoelectric sensor was more sensitive to interface noise but less sensitive to background noise. Importantly, the contact microphone did not pick up higher frequency vibrations as distinctly as the air microphones which provided higher quality recordings as indicated by higher SNIRs.

Jeong et al. [97] used a low-noise electret microphone with a frequency range of 50 Hz to 20 kHz recorded by an audio recorder at a rate of 44.1 kHz. Signals were digitally filtered using a finite impulse response bandpass filter with a bandwidth from 1 to 15 kHz to prioritise short duration joint sounds whilst supressing interface noise.

Feng and Chen [98] developed a piezoelectric sensor comprised of a lead zirconium titanate (PZT) film deposited on titanium cantilever arrays as an acoustic sensing layer. This sensor uses a 1-mm-tall SU8 cylindrical probe on each cantilever to be in direct contact with the skin of the participant and transmit vibrations to the sensor. A thermoresponsive poly(N-isopropylacrylamide) (PNIPA) film was

integrated into the sensor to apply a force to the cantilever and hence improve contact between the probe and the skin when a current is applied across it. The sensor achieved a frequency range of up to 100 kHz, with at least one strong resonant peak at 390 Hz. A sampling rate of 2 MHz was used with a 1 kHz high-pass digital filter to remove low-frequency noise signals. Testing of the sensor on a butchered porcine leg during repeated joint flexure cycles revealed the presence of well-defined peaks located between 30 and 40 kHz, 60 and 70 kHz and 70 and 80 kHz. Similar trends to that observed with commercial AE sensors (the same used in the studies by Mascaro et al. in the JAAS system described later [99]) were noted during overuse of the joint.

Choi et al. [93] developed the bone joint acoustic sensor (BJAS). This has a pin-type probe on a disk-shaped piezoceramic supported by a damped metal plate. The structure is in a metal case with the probe in direct contact with the skin. The system used in conjunction with IMUs seems to have a frequency range of 100 Hz to 25 kHz and is sampled at 50 kHz.

Shark and Goodacre developed the joint acoustic analysis system (JAAS) [99, 100]. This system uses commercial piezoelectric contact ultrasonic acoustic sensors (with high sensitivity in the range 50–200 kHz but monitored over 20–400 kHz at a 1–5 MSPS sampling rate) [100] and electro-goniometers to provide joint anglebased AE during knee joint movement (see **Figure 1**). These commercial AE sensors use relatively thick piezoelectric bulk blocks for AE sensing and are housed in metal shells. The housing is fixed to the skin with surgical tape to maintain a rigid contact. The AE data acquisition operates in a non-continuous recording mode to minimise data volume. When the AE PCI data acquisition board is triggered by a signal value above a pre-set threshold, a 'hit' is recorded corresponding to an acoustic event. Each AE hit is recorded with a set of characteristic waveform features (i.e. dominant frequency, maximum amplitude and duration), and in addition the full waveforms were also stored, digitalized at a 1 MHz sampling frequency over a maximum duration of 15 ms [99]. The number of hits during each joint motion was used to determine a correlation with OA severity defined by KL scores determined using MRI data. It was noted that the frequency response of the acoustic sensor data is characterised by two peaks with a high probability of occurrence during

#### **Figure 1.**

*Output from the joint acoustic analysis system (JAAS). Recording is made as the participant performs five sit-stand-sit movements. A: Acoustic 'hits' from a single knee recorded using a piezoelectric contact ultrasonic acoustic sensors. Each square indicates one acoustic emission captured by the system. For each 'hit' a waveform is also captured [D] from which waveform characteristics are calculated by the software. Alongside the acoustic emissions, joint angle [B] and weight through the leg [C] are also recorded.*

**41**

*Acoustic Monitoring of Joint Health*

**4. Conclusion**

*DOI: http://dx.doi.org/10.5772/intechopen.92868*

to a peak of sensitivity of the sensor used [99].

robust and statistically significant way [78].

knee measurements using a sit-stand-sit protocol, one in the low-frequency range (20–50 kHz) and the other one around 150 kHz. The latter frequency is mainly due

Using radiographic techniques to monitor variations in joint structure and morphology is the classic method of quantifying OA. However, this technique is ionising, often requires multiple measurements as only the plane perpendicular to the radiation is observed and cannot monitor soft tissue directly. MRI can measure the thickness and volume of cartilage, but there are limitations with respect to time and cost. Ultrasound can monitor joint effusion and the thickness of cartilage, but it is not possible for ultrasound to penetrate thick bone tissue and observe the whole joint. There is the additional issue of subjectivity and the large difference in reproducibility based on the skill of those analysing the image. The use of invasive cameras in arthroscopy and joint endoscopy necessitate recovery after diagnosis. These techniques also do not facilitate measurements using dynamic movements. The use of acoustic sensors has the potential to quantify and classify joint pathology whilst removing the subjectivity of classic imaging techniques. Despite progress in detecting differences between type and severity of joint disorders, questions remain about the true origin and form of acoustic signals generated by joint structural changes. Thus, a significant part of the challenge linked to acoustic signal analysis resides in the retrieval of pertinent parameters from irrelevant information in a

As yet, whilst several protocols, sensor types and data analysis techniques have been developed, to date there is no consensus on the most adequate way to record and process vibration data [60]. The methodological aspects of acoustic assessments, such as sensor placement and outcomes measures have not been thoroughly investigated allowing doubt in the technique to remain. For instance, for knee investigations, many studies [73, 81, 101] favour what may be called an open kinematic chain configuration [102] whereby participants sit in a chair and lift their legs in a repetitive fashion, perhaps with weights attached. This has the advantage of being able to vary the load on the joint and allow for the inclusion of participants with advanced degenerative conditions or injuries affecting the limitation of the range of motion in the joint. A common alternative protocol involves repeated sit-stand-sit movements [103–105], creating a closed kinematic chain. This latter configuration perhaps has the advantage of forming a more natural loading of the knee joint. It potentially has the consequence of being inconsistent over time, as people can have the tendency of adjusting their movement to compensate for restricted or painful movement, thus changing the distribution of forces and moments acting on the knee [106]. Data comparing the protocols is limited, and there is no strong evidence for favouring one protocol over the other or indeed over alternatives, such as squatting [94, 102]. Given the protocol affects the loading of the joint and the frequency response of the vibration data generated, it also affects the potential consistency of the statistics derived therefrom and their subsequent interpretation for diagnostic and prognostic purposes. This suggests the necessity of a standard protocol if such techniques are to be used for monitoring the development of OA in an individual over time for clinical or research purposes. Similarly, it is unclear what sort of vibrations and which frequency range is the most pertinent range to measure. In phonoarthrography acoustic waves in the audible range are of most interest. In VAG, focus is on low-frequency (<1000 Hz) vibrations, the cause and nature of which is more general. In AE, acoustic signals are of primary focus, albeit generally of a higher frequency than that used

knee measurements using a sit-stand-sit protocol, one in the low-frequency range (20–50 kHz) and the other one around 150 kHz. The latter frequency is mainly due to a peak of sensitivity of the sensor used [99].
