WO2008074151A1 - System and method of assessing the condition of a joint - Google Patents

Abstract

A system for diagnosing the health of a joint is provided. The system comprises: i) at least one device capable of capturing a vibration signal from the joint; iiI) data collection means adapted to extract time-frequency information from the vibration signal and generate a feature vector; iii) means to correlate the feature vector with anatomical data to identify latent variables; and iv) means to classify the vibration signal based on an input of latent variables to produce an output that defines the health of the joint.

Description

SYSTEM AND METHOD OF ASSESSING THE CONDITION OF A JOINT

Field of the Invention

[OOOl]This invention relates to a system for diagnosing the state of health of a joint. In particular, the present invention relates to a system that measures joint vibrations and analyzes this data to determine the health of the joint.

Background of the Invention

[0002]There are three known techniques in diagnosing joint disorders. Arthroscopy is considered to be the best method to assess the state of a joint. It is a procedure in which a camera or arthroscope is inserted into the joint through small incisions to allow the surgeon to visuedize the condition of the joint. The surgeon can also insert instruments through small incisions to make minor repairs or eliminate damaged tissue. The arthroscope uses fiber optics to send pictures of the inside of the joint to a television monitor. The surgeon can then determine what the problems are, and at the same time may decide to insert other surgical instruments through small incisions in the joint to remove or repair damaged tissues. Although it is considered the most accurate approach, the risks and cost associated with this surgical diagnostic are significant

[0003]A second class of techniques in joint diagnosis is the use of radiological imaging such as X-rays, computed tomography (CT), ultrasound, and magnetic resonance imaging (MRI). Magnetic resonance imaging (MRI) is considered to be the gold standard in medical imaging of the joint. It is a reliable technique that has been growing in popularity over the past two decades and allows for non-invasive evaluation of meniscal abnormalities, cartilage lesions and other features of knee osteoarthritis (OA). Clinically, however, X-ray remains the conventional imaging technique used to diagnose OA and monitor disease progression. Unfortunately, the use of plain radiographs generated using X-ray limits the evaluation of OA to bony features such as joint space narrowing and osteophyte formation. In this context, the primary advantage of MRI is its ability to identify cartilaginous, bony and soft tissue lesions not visualized on plain radiographs. Unfortunately, MRI is relatively expensive and doesn't always reveal accurate information. [0004JA third technique, auscultation, has been used by physicians for centuries and relies on their interpretation of the sounds and vibrations emitted when moving the joint. As early as the 17^th century, Robert Hooke suggested that joint noise could be used as a diagnostic tool in patients suffering from painful joints. A century later, the experiments of Rene Laexuαec with a paper cylinder resulted in the stethoscope, a potential application of which was believed to be the clinical evaluation of joint disorder. Modified versions of the stethoscope were specifically designed for the examination of joints; however, despite the promising concept, results remained very subjective in nature. This drawback was recognized by a Karl Erb, German researcher who, being skeptical of the stethoscope as a robust method, used a sensitive microphone suspended close to the joint. After recognizing the effect of background noise he reverted to a contact microphone where the output was fed through an amplifier and displayed on an oscilloscope. His methodology and findings revolutionized this diagnostic tool mainly by electronically recording the noise. He also hypothesized that by measuring the vibration emitted by the joint directly, one could improve the robustness of the measurement, a speculation that could not be further investigated due to the unavailability of sensitive accelerometers and computers.

[0005JIn 1937, Steindler performed a large scale examination of knee joints using a stethoscope connected to a crystal microphone which picked up joint noises via a short ball-piece with a diaphragm. Over the next two decades, many researchers continued to use the same approach with very interesting results about the nature of noise from normal, degenerative, and rheumatoid joints. In 1976, aware of the problem of background noise, Chu, Gradisar, Railey and Bowling developed a setup with a double microphone to enable noise cancellation. Despite their innovative approach, their work was criticized chiefly due the lack of correlating waveform plots with clinical observations.

[0006]The year 1980 marked a significant departure in the methodology by the use of accelerometers rather than microphones. Mang et al. (1980) used an accelerometer on the patella and recorded signals by passing them through an analogue to digital convertor. Working independently and unaware of Mang's work, Mollan and co-workers (1982) reached a conclusion that microphones were poor transducers when it came to frequency response. McCrea improved the technique further by using accelerometers with built-in amplifiers to eliminate cable noise that Mollan found to be a robustness issue. More recently, Rangayyan et al (1997) published several papers aiming to improve the method by introducing adaptive filtering and other approaches.

[0007]Desρite some success in reporting results, there remains a need for a portable device that can readily be used to provide an initial diagnosis of the health of a joint. Currently, a device capable of providing a robust measurement system that yields a repeatable signal and proper analysis of a non-stationary signal given its time-frequency behaviour does not exist.

Summary of the Invention

[0008JA system has now been developed

[0009]In one aspect of the present invention, there is provided a system for assessing the health of a joint comprising: i) at least one device adapted to receive/capture a vibration signal from the joint; ii) data collection means adapted to extract time-frequency information from the vibration signal and generate a feature vector; ϋi) correlation means configured to correlate the feature vector with anatomical data corresponding to the joint to identify latent variables;, and iv) classification means configured to classify the vibration signal based on an input of said latent variables to produce an output relating to the health of the joint.

[001O]In another aspect, a method for assessing the health of a joint is provided comprising the steps of:

i) capturing one or more vibration signals of the joint; ii) extracting time and frequency information from the captured vibration signal to generate a feature vector; iii) correlating the feature vector with anatomical data corresponding to the joint to identify latent variables; and iv) analyzing the latent variables of the feature vector as input to produce an output which, indicates at least one classification of the joint.

[0011]In another aspect of the present invention, a method of quantitative determination of a characteristic of a joint is provided comprising the steps of:

i) capturing one or more vibration signals of the joint; ii) extracting time and -frequency information from the captured vibration signal to generate a feature vector; iϋ) correlating the feature vector witb anatomical data corresponding to the joint characteristic to identify latent variables; and iv) analyzing the latent variables of the feature vector as input to produce an output which quantifies the characteristic of the joint.

[0G12]In a further aspect of the invention, a computer program product having computer readable code embodied therein for execution by a processor of a computer system for configuring the computer to assess the health of a joint is provided. The computer program product comprises instructions and data for configuring the processor to:

i) extract time and frequency information from a received vibration signal of the joint to generate a feature vector; ϊi) correlate the feature vector with anatomical data corresponding to the joint to identify latent variables; and iii) analyze the latent variables of the feature vector as input to produce an output which indicates at least one classification of the joint.

[0013]In another aspect of the invention, a method for tracking the effectiveness of a treatment on the transient state of health of a joint is provided comprising the steps of:

i) capturing one or more vibration signals of the joint at a first time period and at a second time period following said first time period;

ii) extracting time and frequency information from the vibration signals to generate a feature vector for said first time period and a feature vector for said second time period; iii) computing the latent variables for said first and second time periods using the feature vectors; and iv) analyzing the latent variables of each feature vector to produce an output which quantifies the transient state of health of the joint.

[0014]In another aspect of the invention, a method to quantify a joint characteristic is provided comprising the steps of:

i) introducing vibration signals to various parts of the joint by an external source of excitation; ii) extracting time and frequency information from the vibration signals to generate a feature vector; iii) correlating the feature vector with anatomical data of the joint characteristic to identify latent variables; and iv) analyzing the latent variables of the feature vector as input to produce an output winch quantifies the joint characteristic.

[0015]These and other aspects of the invention will become apparent from the following detailed description and by reference to the figures in which:

Brief Description of the Figures

Figure 1 is a flowchart generally illustrating a method in accordance with the invention;

Figure 2 illustrates a typical placement of accelerometer sensors at sites around a knee joint;

Figure 3 illustrates a typical time domain signal from three accelerometers (10 cycles) on a knee joint;

Figure 4 shows a plot of the latent variable scores, from a PCA analysis, for patients exhibiting various states of joint health including healthy patients, OA patients, and patients with a meniscus tear;

Figure 5 shows the structure of X and Y data matrices; Figure 6 shows the wavelet decomposition of a vibration signal from an accelerometer placed on the patella of a healthy knee;

Figure 7 shows the structure of the X-matrix after performing wavelet transform in which, for each sensor, Dl-9 and AP denote the features extracted from 9 wavelet detail scales and the approximation;

Figure 8 shows the structure of the X and Y matrix for the PLS model;

Figure 9 illustrates the structure of the SVM network;

Figure 10 graphically illustrates the result of the SVM classification in the latent variable space between healthy knees and osteoarthritic knees;

Figure 11 graphically illustrates the result of the S VM classification in the latent variable space between healthy knees and knees with meniscus tear,

Figure 12 illustrates the ability of the method to provide quantitative prediction of cartilage thickness using signals obtained on pat cnts during external excitation of the patella;

Figure 13 illustrates a joint assessment environinent including a diagnosis system in accordance with an aspect of the invention;

Figure 14 illustrates a diagnosis system;

Figure 15 illustrates a feature reduction system for generating latent variables from an X- matrix and a Y -matrix; and

Figure 16 generally illustrates a computing device useful to implement systems in accordance with aspects of the invention.

Detailed Description of the Invention

[0016]A method for diagnosing the health of a joint is provided comprising the stops of: capturing one or more vibration signals of the joint ; pretreating or conditioning the signals; extracting time and frequency information from the vibration signal to generate a feature vector; correlating the feature vector with anatomical data to identify latent variables; and analyzing the latent variables of the feature vector as input to produce an output which provides a diagnosis of the health of the joint. A system employing the present method is also provided. The present inven ion provides, thus, a method by which an accurate, non-invasive diagnosis of the state of health of a joint may be obtained. A system employing the method can readily be provided in a portable package for use in an office or other non-hospital setting, thereby obviating disadvantages associated with other diagnostic methods such as invasive surgical methods and imaging diagnostic methods (MRI and X-ray) which can be expensive and generally entail long wait times.

[0017] A schematic flowchart generally illustrating the steps of the present method is shown in FIG. 1

[0018] A system for assessing the health of a joint is also provided by which the method can be conducted. The system comprises at least one device capable of capturing a vibration signal from the joint; data collection means adapted to extract time-frequency information from the vibration signal and generate a feature vector; means to correlate the feature vector with anatomical data to identify latent variables; and means to classify the signal based on an input of latent variables to produce an output that defines the health of the joint.

[0019]The first component of the system is a device, or aplurality of devices, capable of detecting a vibration signal(s) from the joint. Thus, such device is used to carry out the first step of the present method. Examples of suitable devices include accelerometers and microphones.

[0020] In one embodiment, the system comprises at least one accelerometer, or optionally, an array of accelerometers, as shown in FIG. 2, for placement at different locations around a joint to be diagnosed. Preferably, the accelerometer also includes retaining means that applies a force against the accelerometer to retain it firmly against the skin at the contact site. The retaining means applies a constant force on the accelerometer at the contact site to minimize damping of the signal through the skin and to improve reproducibility of the measurement. Examples of suitable retaining means include a mechanical device which attaches to the knee and holds the accelerometers(s) in place throughout the flexing of the joint, and an air cuff filled with an amount of air that provides the desired force on the accelerometer. The accelerometer(s) may also be manually held in position using an appropriate grip. In this regard, the accelerometer(s) may be embedded in the fingers of a glove and positioned appropriately on the joint by the wearer of the glove, e.g. a technician. FIG. 2 illustrates the placement of multiple accelerometer sensors around a knee joint.

[0021]Using an accelerometer, for example, vibration signals from a joint are captured. FIG.3 shows typical vibration signals captured at 3 locations during multiple knee flexing cycles, e.g. non-stationary cycles. The vibration signals are fed into a data collection and processing means which is adapted to extract time-frequency information from the signal to generate a feature vector. As one of skill in the art will appreciate, any one of several methods may be employed to transform the vibration signal from the time domain to the frequency domain including Fourier transform (FT), the Fast Fourier transform (FFT), and Wavelet Transform (WT) such as Discrete Wavelet Transform (DWT), the Continuous Wavelet transform (CWT) and Wavelet Packets (WP) which each provide good time-frequency resolution for signals that are non-stationary in intensity and frequency.

[0022] In one embodiment of the present system, feature extraction is conducted by wavelet transform (WT) which provides a number of features. For example, wavelet analysis can reveal aspects like trends, breakdown points, discontinuities in higher derivatives, and self-sirnilarity not readily observed with other techniques, and can compress or de-noise a signal without appreciable signal degradation. Wavelet analysis also offers a windowing technique with variable-sized regions, allowing the use of long time intervals for more precise low-frequency information and shorter regions for high- frequency information. The wavelet transform provides wavelet coefficients relating to the signal strength for various times and frequencies. Various transformations of these coefficients provide features including the power or variance of the coefficients in each detail, and the histogram of the coefficients in each detail.

[0023] Further processing of the frequency components of the decomposed signal to perform feature reduction is then conducted by correlation means such as multivariate analysis. The goal of performing feature reduction is to condense the feature vector into a reduced number of features (latent variables) while still capturing all the valuable information. ID this context, multivariate projection methods, such as principal component analysis (PCA), SVD (singular value decomposition) and PLS (partial least squares) may be used.

[0024]In order to calculate the number of latent variables that best summarize the feature vector, a model that consists of two datasets, the X-marrix and the Y-matrix, is used for the correlation. The X-matriχ comprises the feature vector of the joint vibrations, while the Y-matrix consists of anatomical data obtained from a joint using MRI and arthroscopy. Examples of anatomical data that may be obtained include:

i. Total area of subchondral bone ii. Area of subchondral bone covered with cartilage iii. Percent of subchondral bone covered with cartilage iv. Area of subchondral bone, denuded, eroded, full thickness defect; percent of subchondral bone area that is denuded v. Area of cartilage surface vi. Volume of cartilage vii. Volume of cartilage divided by total area of subchondral bone viii. Number of (separate) cartilage denuded areas ix. Cartilage thickness over total subchondral bone area; denuded areas counting as 0 mm thickness x. Cartilage thickness over cartilaginous area of subchondral bone; denuded areas not included xi. Arthroscopy/surgery

[0025]The number of latent variables may be even further reduced to refine the information to the necessary or most important information using techniques well established in the art, including for example, principle component analysis (PCA) and partial least squares (PLS). In addition, uncorrelated noise may be reduced, for example, using orthogonal signal correctioxx (O-PLS) methods.

[0026]Plots of the reduced latent variable features can then be used to visually diagnose the state of health of the joint FIG. 4 shows the visual separation among patients provided by the latent vectors of PCA. Different patients with similar states of joint health (e.g. severe OA, mild OA, healthy or meniscus tear) will have latent variable values (scores) that fall in similar regions in the plot, while patients with different states of joint health will fall in different regions.

[0027]Classification of the data into defined classes of joint health (healthy, severe OA, etc) may then be employed. Based on an input of latent variables processed as described above, the system produces an output which provides a classification of the input data relative to standard classifications. Classification means may be any means capable of determining the placement of input data based on standards developed from the training data which is comprised of anatomical data obtained by imaging techniques and corresponding vibration data. Examples of classification means in accordance with the invention include feed-forward neural networks such as support vector machines (SVM) and radial basis function (RBF); recurrent neural networks; discriminant PLS; and partial least squares - discriminant analysis (PLS-DA).

[0028]The data collection means, the correlation means and the classification means described above may execute within a computer system. The computer system may include a visual display, a keyboard, and one or more auxiliary user interfaces, each of which is coupled to a computer processor. The computer processor may include one or more general purpose processors and/or special purpose processors (e.g., ASICs, FPGAs, DSPs, etc.). The computer system further includes a memory module which is coupled to the processor through a BUS or other suitable means. Each of the data collection means, the correlation means and the classification means provide a set of computer instructions which are fetched by the processor from the memory for execution by the processor. The processor also reads data stored on the memory and writes data to the memory. For example, the computer system may be coupled to an accelerometer which captures a vibration signal from the joint and said vibration signal is stored on the memory for subsequent access by the processor and execution of the computer instructions thereon. [0029] The processor can further interface with the visual display to show received information, stored information, user inputs and the like in accordance with the computer instructions.

[0030] The memory described above can include any type of computer memory such as, but not limited to, random access memory (RAM), read-only memory (ROM), and computer readable mediums which can include hardware and/or software such as, by way of example only, magnetic disks, magnetic tape, optically readable medium such as CD/DVD ROMS, and memory cards.

[0031] It will further be understood by a person skilled in the art that the data collection means, the correlation means and the classification means may execute in the same computer system, or can be distributed among separate computer systems. It will be understood that the system and methods described herein may be implemented by any hardware, software, or a combination of hardware and software having the above described functions. As described above, the software code, either in its entirety or a part thereof, may be stored in a computer readable memory,

[0032] The present system may also provide a visual display of the output in which the output for a given joint is compared with training data or standards, e.g. results obtained by imaging techniques. This feature of the invention advantageously minimizes the subjectivity generally associated with analyzing a vibration signal. Thus, the output, as well as the training data, is transmitted to a display unit, such as a LED screen, for display in the conventional manner, thereby providing an immediate diagnosis of the health of a joint.

[0033] The present system and method may be applied to achieve, for example, determinations such as discriminating between healthy and non-healthy joints including knees, elbows, ankles, wrists and hips; assessing the degree of osteoarthritis in a joint, classifying menisci tears, evaluating the effectiveness of different treatments and tracking patients after treatment to determine the nature and timing of subsequent treatments, if required. The system and method may also be used to predict quantitative features of a joint, for example, cartilage thickness and osteophyte length. As one of skill in the art will appreciate, other determinations may also be made using the present system and method.

[0034]A computer program product is also provided comprising computer readable code embodied therein for execution by a processor of a computer system which configures the system to assess the health of a joint according to the method herein described. The computer program product comprises a general user interface (GUI) providing instructions and a set of algorithms to extract time and frequency information from received vibration signals of the joint to generate a feature vector; to calculate the latent variables from the feature vector; and then to analyze the latent variables of the feature vector as input to produce an output which indicates at least one classification of the joint

[0035]A detailed description of the system and its components follows.

Joint Assessment Environment

[0036]Referring to Figure 13, shown is a joint assessment environment 10. The environment 10 has a diagnosis system 14 for receiving a plurality of vibration data (e.g. frequency data of the joint vibration signal amplitude or magnitude for a specific frequency) signals 12 from a vibration measurement and data capture device 13 (e.g. an accelerometer, microphone, etc.). The signals 12 are processed by the diagnosis system 14 for producing the feature/frequency data X (e.g. the X matrix or the feature vector). For example, the feature/frequency data X could represent a temporally dependent sequence of discrete/continuous vibration data measurements for various physical positioning of a patient's joint (e.g. knee), such as one or more joint vibration signal data value(s) for each position of the joint (such as a plurality of positions from fully bent to fully extended). Accordingly, it is recognised that the vibration data measurement captured from device 13 could be correlated to specific times that the frequency is detected by the device 13 throughout the range of motion of the joint and/or can be correlated to the relative position of the joint (e.g. the vibration data measurement is a function of position). In any event, it is recognised that the vibration data measurement sequence is correlated with (e.g. is dependent upon) an independent variable (e.g. time and/or positioning).

[0037]The diagnosis system 14 also has access to raw anatomical data 16 (e.g. such as patient images and patient comments with respect to pain or other observations obtained from the patient and/or examiner during the joint positioning). The anatomical data 16 (e.g. MRJ of the patient's joint) is recorded during the different joint positioning, such that the raw anatomical data measurements (also patient comments such as their indication of pain in a specific position) are correlated to the same independent variable (e.g. time and/or positioning) as the joint vibration data measurement, thus providing for synchronization between the anatomical data 16 and the feature/frequency data X. The anatomical data 16 has image characteristics extracted therefrom (either by the diagnosis system 14 or a third party processor - not shown) in order to produce a temporally dependent sequence of discrete/continuous anatomical characteristics Y (e.g. a matrix) fox various anatomical positioning of the patient' s j oint. It is recognised that the independent basis (e.g. as a function of time) for the feature/frequency data X is synchronized with the independent basis (e.g. same function of time) for the anatomical characteristics Y.

[0038]The diagnosis system 14 then combines the feature/frequency data X and the anatomical characteristics Y in order to produce a resultant dataset 19 (e.g. LV ~ latent variables) that is then compared with one or more classification templates/standards 18, in order to produce output data 20 that is representative of the health of the patient's joint. It is recognised that the dataset 19 is preferably dimensionally reduced in comparison to the data X, Y. For example, in statistics, latent variables (as opposed to observable variables), are variables that are not directly observed but are rather inferred (through a mathematical model) from other variables that are observed and directly measured. One advantage of using latent variables is that it reduces the dimensionality of data. A large number of observable variables can be aggregated in the model, Further, in statistics, one example for calculation of the latent variables is the method of partial least squares regression (PLS-regression), which bears some relation to principal component analysis. Instead of finding the hyperplanes of maximum variance, PLS can find a linear model describing some predicted variables in terms of other observable variables. PLS can be used to find the fundamental relations between the two data sets (X and Y), i.e. a latent variable approach is to model a covariancε structures in the two spaces of an X matrix and a Y matrix. For example, the PLS model will try to find the multidimensional direction in the X space that explains the maximum multidimensional variance direction in the Y space. It is recognised that an alternative long form for PLS is projection to latent structures.

Diagnosis System

[0039]Referring to Figure 14, the diagnosis system 14 contains a data collection and processing module 110 (e.g. data collection means), a correlation and extraction module 114 (e.g. correlation means), an image characteristic extractor module 112 for converting the anatomical image data 16 to the anatomical characteristics Y, a user interface 402 for monitoring/controllmg conversion operation of the characteristic extractor module 112 and monitoring/controlling classification operation of the module 114 for template (e.g. standard) generation, a classification module 116 (e.g. classification means), and one or more memories 410 for storing the data 16, 18, X, Y, as desired. It is recognised that the characteristic extractor module 112 can be part of or separate from the diagnosis system 14, and is configured so as to convert the anatomical image data 16 to the anatomical characteristics Y prior to sending the data Y to the correlation module 114. It is also recognised that the correlation module 114 can obtain the data Y directly from the memory 410. It is shown in Figure 14 that the various modules 110-116 can be used to process individual data X,Y,LV (for an individual joint of a patient) and/or group data X,Y,LV (for joints of a number of patients used for template generation and training).

LV Calculation

[0040]Referring to Figure 15, shown is the feature reduction system (e.g. module 114) for generating latent variables from an X-raatrix and a Y-matrix (or other forms of the X, Y data other than matrix format). As illustrated, for example, the X-matrix is generated by obtaining one or more vibration signals 12 from a patient, the vibration signals 12 are processed by a data collection modules 110, which extracts vibration signal information at various times and frequency ranges. As described earlier, the data collection module 110, may for example perform Wavelet analysis to extract the time frequency information and generate the X-matrix representing the vibration signal 12. The Y-matrix represents the anatomical data such as that obtained from imaging techniques (X-Ray, MRI). As illustrated, the Y-matrix has image characteristic information (e.g. shown as the variable o), which can be for example, various physical properties of the joint assessed from the image (including for example bone density information). As also illustrated, the Y- matrix is a function of time, which corresponds approximately to the time reference of the X-matrix. In this way, anatomical data (Y-matrix) and vibration signal data (X- tnatrix) occurring during approximately the same time period may be correlated via the correlation module 114. If there is no corresponding data (e.g. of negligible value) in one of the Y-matrix or the X-matrix for me corresponding data entry in the other one of the X-matrix or the Y-matrix then, for example, the date is deemed not correlated and the latent variables (LV) generated either do not include the latent variables for the time period where the X-matrix is not correlated to the Y-matrix or a zero/negligible value is contained for that time period. For example, if the X2 data does not exist or docs not correspond to the Y2 data set, the latent variables may either be lacking the LV(X2) or LV(X2) may be equivalent to a null value.

[0041]Similarly, if the anatomical data Y4 has features or characteristics that do not correspond or correlate with the X4 values, then the latent values may either be lacking the LV(XA) ox a null value for LV(X4). Other methods such as the goodness of fit, and the sum of squares may be used to determine which values of the X and Y matrices are correlated such as to generate a latent variable set denoting the correlated values. In this manner, the latent variables provide a significant reduction in the dimension space. In this manner, a reduction in dimension of the LV values is realized in view of the operation of the module 114.

Computing Devices

[0042]Rcferring to Figures 13 and 16, each of the above-described components of the environment 10, i.e. the signal collection device 13, the diagnosis system 14, and the anatomical data collector 18, can be implemented on one or more respective computing device{s) 101. The devices 101 in general can include a network connection interface 400, such as a network interface card or a modem, coupled via connection 418 to a device infrastructure 404. The connection interface 400 is connectable during operation of the devices 101 to the network 11 (e.g. an intranet/extranet for making available the anatomical data 16 and/or the signal data 12), which enables the devices 101 to communicate with each other as appropriate. The network 11 can support the communication of the data signals 12 between the components of the environment 10.

[0043]Referring again to Figure 16, the devices 101 can also have a user interface 402, coupled to the device infrastructure 404 by connection 422, to interact with a user (e.g. user diagnosis system 14). The user interface 402 is used by the user of the device 101 to interact with the non-stationary frequency information contained within the signals 12 (e.g. the X matrix) and the anatomical data 17 (e.g. the Y matrix) to operate the diagnosis system 14. The user interface 402 can include one or more user input devices such as but not limited to a QWERTY keyboard, a keypad, a trackwheel, a stylus, a mouse, a microphone and the user output device such as an LCD/LED screen display and/or a speaker. If the screen is touch sensitive, then the display can also be used as the user input device as controlled by the device infrastructure 404. The user interface 402 is used to interact with the output results 20 of the diagnosis system 14, for example in graphical and/or textual form.

[0044]Referring again to Figure 16, operation of the devices 101 is facilitated by the device infrastracrure 404. The device infrastructure 404 includes one or more computer processors 408 and can include an associated memory 410 (e.g. a random access memory). The computer processor 408 facilitates performance of the device 101 configured for the intended task through operation of the network interface 400, the user interface 402 and other application programs/hardware 407 (eg. the diagnosis system 14) of the device 101 by executing task related instructions. These task related instructions can be provided by an operating system, and/or software applications 407 located in tho memory 410, and/or by opcτability that is configured into the electronic/digital circuitry of the processors) 408 designed to perform the specific task(s). Further, it is recognized that the device infrastructure 404 can include a computer readable storage medium 412 coupled to the processor 408 for providing instructions to the processor 408 and/or to load/update application programs 407. The computer readable medium 412 can include hardware and/or software such as, by way of example only, magnetic disks, magnetic tape, optically readable medium such as CD/DVD ROMS, and memory cards. In each case, the computer readable medium 412 may take the form of a small disk, floppy diskette, cassette, hard disk drive, solid-state memory card, or RAM provided in the memory module 410. It should be noted that the above listed examples of computer readable mediums 412 can be used either alone or in combination. The device memory 410 and/or computer readable medium 412 can be used to store the protocols and associated plug-in identifications of the device 101.

[0045]Further, it is recognized that the computing devices 101 can include the executable applications 407 comprising code or machine readable instructions for implementing predetermined functions/operations including those of an operating system and the diagnosis system 14, for example. The processor 408 as used herein is a configured device and/or set of machine-readable instructions for performing operations as described by example above. As used herein, the processor 408 may comprise any one or combination of, hardware, firmware, and/or software (e.g. modules 110, 112, U 4, 116). The processor 408 acts upon information by manipulating, analyzing, modifying, converting or transmitting information for use by an executable procedure or an information device, and/or by routing the information with respect to an output device. The processor 408 may use or comprise the capabilities of a controller or microprocessor, for example. Accordingly, any of the functionality of the diagnosis system 14 (e.g. modules 110-116, and subset/submodules thereof) may be implemented in hardware, software or a combination of both. Accordingly, the use of a processor 408 as a device and/or as a set of machine-readable instructions is hereafter referred to genericaliy as a processor/module for sake of simplicity. Further, it is recognised that the diagnosis system 14 can include one or more of the computing devices 101 (comprising hardware and/or software) for implementing the modules 110-116 or functionality subset thereof, as desired. [0046]It will be understood that the computing devices 101 of the measurement device 13 and the diagnosis system 14 may be, for example, personal computers, personal digital assistants and the like. Further, it is recognised that the measurement device 13 may only comprise an electronic data collection mechanism, such as a sensor (e.g. accelerometer).

[0047]Further, it is recognised that the modules 110-116 can be configured to operate interactively as shown, the operations/functionality of the selected modules 110-116 can be combined or the operations/functionality of the selected modules 110-116 can be further subdivided, as desired. Further, it is recognised that the modules 110-116 can communicate or otherwise obtain their calculated results from one another or can store their respective calculated results in the storage 410 for subsequent retrieval by another module 110-116 therefrom. It is also recognised that there may be more than one memory 410 used by associated modules 110-116, as desired. Further, it is recognised that the extractor module 112 may be implemented external to the diagnosis system 14 and thereby only supply completed training sets/templates to the diagnosis system 14 for use by the correlation module 114.

[0048] Embodiments of the present invention are described in the following specific examples which are not to be construed as limiting.

Example 1 — Development of a vibration diagnostic method

[0049]Development of the present methodology consisted of two main steps: (i) building a database, and (ii) development of a statistical classifier.

[0050]In the first step, the objective was to build a database consisting of subjects (observations) belonging to three different groups: (i) Healthy individuals, (ii) Patients with known history of knee problems, and (iii) Patients that received treatments. For each subject, X-ray, MRI, and vibration measurement were taken on the non-dominant knee.

[0051]In the second step, the objective was to build a supervised classification model that defined classes based on observed results from the MRI and X-rays using vibration measurements. This was accomplished in three steps: 1) wavelet analysis applied to the time-domain signals from the accelerometers to construct sets of parameters that describe the vibration pattern in terms of time and frequencies; 2) the vector containing features computed from, the coefficients from the wavelet decomposition was further processed using multivariate projection methods (PCA or PLS) to reduce its dimensions to few latent variables that contained most of the significant information in the signal. In a sense, this step can be seen to de-noise the data and improve the computation; and 3) Support Vector Machine (SVM) was employed to design a classifier.

Structure of the database

[0052]The structure of an X- matrix and Y-matrix used in a method according to the present invention are described below. Figure 5 is a schematic showing the structure of the database based on this X-matrix and Y-matrix.

[0053]Data used to develop the X-matrix and Y-matrix were obtained from the following patient groups. Each patient consented to have a single X-ray and MRl scan of his/her non-dominant knee and to have the vibration of the Jknee while in motion recorded using accelerozneters.

Group 1: Male and female between 20 and 69 years of age were recruited provided they were free of knee pain, had never sustained a knee injury nor been previously diagnosed with a bone or joint disease (i.e. rheumatoid arthritis, osteoporosis, etc.). This group consists of people with normal knees, mild OA, and mild degrees of menisci tears. (n=88 )

Group 2: Male and female patients with diagnosed knee problems were recruited. Although this group of patients consisted dominantly of subjects diagnosed with advanced stage of Osteoarthritis (OA), patients with varying degrees of menisci tears were also included . (o= 52)

Group 3: Patients with advanced of OA that had received a treatment in the form of injection were also recruited Measurements were performed on the subject before and after the treatment (n-1)

The X-matrix: [0054]The vibration of the knee was measured in the time domain on patients in the above identified groups by using three accelerometers placed on the patella and the femur (medial & lateral) as illustrated in Figure 2, to generate a time domain signal as shown, in Fig. 3. These measurements were acquired under two different conditions. Io the first condition, the subject is seated and swinging their leg in a repetitive motion. On a small subset of group 1, active vibration was applied while the patient was asked to go up and down three steps.

The Y-matrtx:

Group 1:

[0055]Radiσgraphy: X-rays were acquired using the fixed-flexion technique as described in Eckstein et al., 2006, the relevant contents of which arc hereby incorporated by reference, particularly pages A46-A75, and evaluated independently by two radiologists according to the K-L scale for OA (Kellgren and Lawrence, 1957, the contents of pages 494-502 of which are incorporated herein by reference). Features considered in this scoring system included the presence and/or severity of osteophytes, joint space narrowing, sclerosis of the bone and deformity of bone ends.

[0056]Periρheral MRI: MR scans were acquired using a Tesla pMRI system. Subjects were positioned such that their non-dominant knee was centered within the iso-centre of the magnet. A complete description of this particular MR system can be found in a review by Shellock and Hollister (2002), the contents of which at pages 261-287 are incorporated herein by reference. The following measurements were evaluated from the scan:

i. Total area of subchondral bone

ii. Area of subchondral bone covered -with cartilage

iii. Percent of subchondral bone covered with cartilage

iv. Area of subchondral bone, denuded, eroded, full thickness defect Percent of subchondral bone area that is denuded v. Area of cartilage surface vi. Volume of cartilage vii. Volume of cartilage divided by total area of subchondral bone

[0057]In addition to the above measurements, features were graded on a categorical scale as indicated in Table 1. This grading is used as the basis for the classification.

Groups 2 and 3:

[0058]These patients had X-rays and MRI scans on the knee that was diagnosed with

OA. The classification was then performed in accordance with Table 1.

Table 1. Evaluation template for diagnostic pMRI

Development of a Classifier

[0059]The formulation of the classification problem includes three distinct steps; (i) feature extraction (ii) feature reduction, and (iii) classifier development

[0060]Feαture extraction: A data collection means was used to extract important features from the time domain signal A typical signal in the time domain is shown in Figure 3 obtained on a knee using three accelerometers.

[0061]In this example, the data collection means included discrete wavelet transform (DWT) analysis to extract the appropriate features from the signal since it exhibited a non- stationary behavior. The rational behind this choice was due to the fact that Wavelet Transform (WT) provides local information about the signal at different frequencies. Thus, for example, to detect features in the vibration signal that arise at a certain time during the cycle, the wavelet transform at a fine scale would be used, whereas to capture several low-frequency cycles (which occur over a broad interval of time), the transform at a coarser scale would be used. The mother wavelet used in this example was the Daubechies; however, any one of several may be used including the Daubechies, Symlets, and Haar. The wavelet decomposition obtained (from a healthy knee) is shown in Figure 6. The structure of the X-matrix after performing the wavelet transform is illustrated in Figure 7.

[0062]Feature reduction: In this step, a correlation means was used to reduce the size of the dataset by performing multivariate analysis. Two approaches are commonly used; principal component analysis (PCA) and partial least square (PLS). Since prior knowledge existed on each observation in the dataset, PLS was preferable due to the fact that the scores were calculated in a way that correlated with the Y-matrix. PLS calculates then "new X- variables", t₈, as linear combinations of the old matrix (containing the wavelet decomposition from three sensors). The structure of the dataset obtained is illustrated in Figure 8. The benefits of performing this step are as follows:

i. Significant reduction in the dimension space, ii. Since the number of latent variables are calculated using the concept of "cross validation", an optimal balance between fit and prediction is achieved, iii. Since the score matrix captures only the dominant latent variables that summarize the X-matrix and correlate with the Y-matriχ, this step can be seen as a mean to denoise the x-matrix and therefore improve the classification.

[0063]First, as an exploratory step, a PCA was built on the X-matrix. Figure 4 shows that there exists evident separation between healthy knees, OA knees, and tears in the meniscus. Second, a PLS model was built between the X and Y matrices using the structure shown in Figure 8. The model yielded nine components that were significant by cross validation. The goodness of fit (R²γcum; the % sum of squares explained by the fitted model) was 89% and the goodness of prediction (Q² _ycum t;he % sum of squares predicted by the model in. cross validation) was 85%.

[0064]Classiflcation: Ib this step, a classification means, such as SVM (Support Vector Machine) was used to build linear and nonlinear decision boundaries. In addition, the design of the SVM network offers flexibility in that it can be tuned (by the choice of the inner-product kernel and the regularization parameter, c) The structure of the SVM network is illustrated in Figure 9. The inputs to the network are the t-scores (nine latent variables) identified in the previous step by the PLS model. Radial-basis function -was used in the inner-product kernel.

[0065]Cross- validation was used to determine the optimum value for the inner-product kernel and the regularization parameter. The result of the different types of classifiers are summarized in Table 2.

Table 2: Summary of training and testing error for different classifiers

Training error Testing error

Feature Classification Normal OA Posterior Bucket-handle Normal OA Posterior Bucket- reduction method knee Knee nom tear tnr knee Knee horn tear handle tear

None SVM 0.02% 10.53% 11.39% 16.74% 9.39% 12 64% 17.77% 26.11%

PLS SVM 2.23% 3.99% 4.22% 6.20% 3.48% 4.68% 6.58% 9.67%

PLS RBF network 2.45% 4.29« 4.64% 6.82« 3.83% 5 15% 7.24« 10.54«

PLI Discriminant PLS 2.97« S.19% 5.61« 8.25« 4.63« 6.22« 8.76« 12.86«

PLS Recurrent network 2 21« 3.00* 4.2:1% 6.20« 6.62« 11.70« 9.70« 14.20«

[0066]Figure 10 shows graphically the result of the SVM classification in the latent variable space (t₁..t₂) between healthy knee and knee diagnosed with OA, while Figure 11 shows the separation between healthy knee and knees diagnosed with meniscus tear. The separation between different types of tear was not clear due to the small sample size. Figure 10 also shows the projection of one patient diagnosed with OA before and after the treatment by a synovial fluid injection. This sequence of measurements shows that this tool may be used to assess the effect of treatments. [0067]Although the disclosure herein has been drawn to one or more exemplary systems and methods, many variations will be apparent to those knowledgeable in the field, and such variations are within the scope of the application.

Example 2 - Quantitative Prediction of knee joint features

[0068]The methodology can also be used to predict quantitative features (e.g. those listed in Table 1) obtained from the analysis MRI images or from observations made during surgical or arthroscopic examinations.

[0069]On a small subset (n=18) from group 1 , a study was performed by vibrating the knee joint on the patella while the subject went up and down three steps. Three accelerometers recorded the response signal from the patella and the femur (medial & lateral) as previously illustrated in Figure 2. Subjects were chosen to represent different degrees of cartilage thickness and volume with which they were highly correlated. A micro-vibrator, placed on the patella, was used to generate a Gaussian vibration signal at low frequency (2k-30k) at mid range magnitude.

[0070]The analysis of the signal was done in two steps. In a first step, feature extraction was performed followed, in a second step, by a linear PLS. The X-matrix consisted of features extracted from the wavelet coefficients, while cartilage thickness and volume were used in the Y-matrix. Y-observed versus Y-predicted showed good correlation with a prediction error of 8% as shown in Figure 12,

References:

1. Jones WHS Hippocrates: De Morbis, ii sec 59 - 61, iii sec 7 Cambridge (USA), Harvard Press 1952.

2. Hooke R The Postumous Work of Robert Hookc London, Smith and Walford 1705.

3. Laennec RTH Traite de L'Auscultation Mediate et des Maladies des Poumons et du Coeur 4th Ed. Vol. 3 Paris. JS Chande 1837.

4. Heuter C Grundriss der Chirurgie 3rd Ed. Leipzig FCW Voge 1885.

5. Blodgett WE Auscultation of the Knee Joint Boston Medical and Surgical Journal 1902: 146; 63 - 66.

6. Birchcr E Zur Diagnose der Meniscusluxation und des Meniscusabrisses Zentralbl. f. Chumrgie 1913: 40; 1852 - 1853.

7. Walter CF The Value of Joint Auscultation Lancet 1929: 1; 920 - 921.

8. Erb KH Registrierung von Gelenkgerauschen Archiv. Klin, Chur. 1933: 177; 475 - 482.

9. Erb KH uber die Moglichkeit der Registrierung von Gelenkgerauschen Deut. Z. f. Chur. 1933: 241; 237 - 245.

10. Steindler A Auscultation of Joints J. Bone Joint Swg 1937: 19; 121 - 136.

11. Peylan A Direct Auscultation of the Joints Rheunatism 1953: 9; 77 - 81.

12. Ekensten B Jamforande Undersokningar av 400 stKSkleden LKliniskt ROntgenologiskt och Fonogrammetriskt Hanseede Odontol. Tidskrift 1955: 63;18 - 23.

13. Fischer H Johnson EW Analysis of Sounds from Normal and Pathologic Knee Joints 3rd Int. Congress - Phys. Med. 1960; 50 - 57.

14. Radochay L Bohensky G Tarjanyi J Das Schallphanomen und siene Andwendung in der Differentialdiagnostik der Meniscusverletzungen Archiv. Orthop. Unfall. 1965:

57; 200 - 204.

15. Radochay L Bohensky G Tarjanyi J Weitere Untersungen uber das Schallphanomen Beitr. Orthop. Traumat 1966: 13; 662 - 664.

16. Szabό E Danis L Tõrõk Z A térd hangjeleπsége nek vizsgalata Magyar Traumatol. 1972: 15; 118 - 127.

17. Danis L Szabό E Tβrδk Z Genophonographias gϋrbέk osztalyozasa meniscus serUleseknέl Magyar Traumatol. 1972: 15; 202 - 209.

18. Yasinovsky MS Lozin LL Rasftkovich MP Volkhov AA Shlovsky BI Novi metod diagnostiki sustavnikh zabolevani - Arthrophonographia Klin Med (Moskva) 1973: 7; 25 - 28.

19. Oehl R Bobnenberger J Hinkelmann W Petrowicz O Zur Tecknik der Phonoarthrographie Med Welt 1974: 25; 1984 - 1989.

20. Chu ML Gradisar IA Railey MR Bowling GF An. Electroacoustical Technique for the Detection of Knee Joint Noise Med. Res. Engineer 1976: 12; 18 - 20.

21. Chu ML Gradisar IA Railey MR Bowling GF Detection of Knee Joint Noises using Acoustical Pattern Recognition Technique J. Biomech. 1976: 9; 111 - 114.

22. Chu ML Gradisar IA Zavodney LD Possible Clinical Application of aNon- invasive Monitoring Technique of Cartilage Damage in Pathological Knee Joints J. Clin. Engineer. 1978: 3; 19 - 27.

23. Mang. W Birk M Blumen G Praktische Aπwendbarkeit der Phonoarthrographie in der Kniegelenksdiagnostik Z. Orthop. 1980: 118; 85 - 90.

24. Mollan RAB McCullagh GC Wilson RI A Critical Appraisal of the Auscultation of Human Joints Clin. Orthop. 1982: Ϊ70; 231 - 237.

25. Schmidt H Schallsignalle zur Diagnostik von Kniebinnenschaden Beitr. Orthop. Traumatol. 1980: 27; 226 -230.

26. Kotani K Suzuki K Acoustical Analysis of Joint Sound Through Passive Motion ( in Japanese) J. Jpn. Orthop. Ass. 1983: 57; 1869 - 1880.

27. McCrea JD Vibration Emission in Normal Knee Joints MD Thesis The Queen's University of Belfast. 1985.

27. Z. M. K. Moussavi, R. M. Rangayyan, G. D. Bell, C. B. Frank, K. O. Ladly, and Y. T. Zhang, "Screening of vibroartbxograpbic signals via adaptive segmentation and linear predict on modeling," IEEE Trans. Biomed. Eng., vol. 43, pp. 15-23, Jan. 1996.

28. Y. Shen, R. M. Rangayyan, G. D. Bell, C. B. Frank, Y. T. Zhang, and K. O. Ladly, "Localization of knee joint cartilage pathology by multichannel vibroarthrography," Med. Eng. Phys., vol. 17, no. 8, pp. 583-594, 1995. 28. R. M. Rangayyan, S. Kxishnan, G. D. Bell, and C. B. Frank, "Computeraided auscultation of knee joint vibration signals," in Proceedings of the European Medical and Biological Engineering Conference, Vienna, Austria, November 1999, pp. 464— 465.

29. S. Krishnan, R. M. Rangayyan, G. D. Bell, and C. B. Frank, " Bonification of knee- joint vibration signals," in CDROM Proceedings of World Congress on Medical Physics and Biomedical Engineering, Chicago, IL, My 2000

30. R. M. Rangayyan, S. Krishnan, G. D. Bell, C. B. Frank, and K. O. Ladly, "Parametric representation and screening of knee joint vibroarthrographic signals," IEEE Trans. Biomed. Eng., vol. 44, pp. 1068-1074, Nov. 1997.

31. Peterfly C, Li J, Zaim S, Duryea J, Lynch J, and Miaux Y. Skeletal Radiol 2003; 32:128-132.

32. Kellgren JH, and Lawrence JS. Radiological Assessment of Osteoarthrosis. Ann Rheum Dis 1957; 16:494-501.

33. Shellock FG, and Hollister, MC. In-office MR imaging. Clin Sports Med 2002; 21:261-87, vii.

34. F. Eckstein, F. Cicuttini, J- P Raynauld, J. C. Waterton and C. Peterfly, "Magnetic Resonance Imaging of articular cartilage in knee osteoarthritis (OA): morphological assessment". Osteoarthritis and Cartilage, 2006, pp. A46-A75.

All references are incorporated herein by reference.

Claims

We claim:

1. A system for assessing the health of a joint comprising: i) at least one device adapted to receive/capture a vibration signal from the joint; ϋ) data collection and processing means adapted to extract time-frequency infoπnation from the vibration signal and generate a feature vector; iii) correlation means configured to correlate the feature vector with anatomical data corresponding to the joint to identify latent variables; and iv) classification means configured to classify the vibration signal based on an input of said latent variables to produce an output relating to the health of the joint.

2. A system as defined in claim 1 , wherein the device is an accelerometer.

3. A system as defined in claim 1 , comprising a plurality of accelerometers.

4. A system as defined in claim 1, wherein the data collection and processing means is selected from the group consisting of Fourier transform (FT), the Fast Fourier transform (FFT), Wavelet Transform (WT), Discrete Wavelet Transform (DWT), Continuous Wavelet transform (CWT) and Wavelet Packets (WP).

5. A system as defined in claim 1, wherein the correlation means is a multivariate analysis method.

6. A system as defined in claim 5, wherein the multivariate analysis is selected from the group consisting of principal component analysis (PCA), SVD (singular value decomposition) and PLS (partial least squares).

7. A system as defined in claim 5, wherein the correlation means utilizes a dataset comprising an X-matrix including the feature vector and a Y-matrix including anatomical joint data.

8. A system as defined in claim 1, wherein the classification means is selected from the group consisting of a support vector machine (SVM), a radial basis function (RBF), recurrent neural networks, discriminant PLS and partial least square- discriminant analysis (PLS-DA).

9. A method for assessing the health of a joint comprising the steps of: i) capturing one or more vibration signals of the joint; ii) extracting time and frequency information from the captured vibration signal to generate a feature vector; iii) correlating the feature vector with anatomical data corresponding to the joint to identify latent variables; and iv) analyzing the latent variables of the feature vector as input to produce an output which indicates at least one classification of the joint.

10. A computer program product having computer readable code embodied therein for execution by a processor for configuring a computer to assess the health of a joint, said computer program product comprising instructions and data for configuring a processor of the computer system to: i) extract time and frequency information from a received vibration signal of the joint to generate a feature vector; ii) correlate the feature vector with anatomical data corresponding to the joint to identify latent variables; and ii) extracting time and frequency information from the captured vibration signal to generate a feature vector; iii) correlating the feature vector with anatomical data corresponding to the joint characteristic to identify latent variables; and iv) analyzing the latent variables of the feature vector as input to produce an output which quantifies the characteristic of the joint.

12. A method for tracking the effectiveness of a treatment on the transient state of health of a joint comprising the steps of: i) capturing one or more vibration signals of the joint at a first time period and at a second time period following said first time period; ii) extracting time and frequency information from the vibration signals to generate a feature vector for said first time period and a feature vector for said second time period; iii) computing the latent variables for said first and second time periods using the feature vectors; and iv) analyzing the latent variables of each feature vector to produce an output which quantifies the transient state of health of the joint.

13. A method to quantify a joint characteristic comprising the steps of:

i) introducing vibration signals to various parts of the joint by an external source of excitation; ii) extracting time and frequency information from the vibration signals to generate a feature vector; iii) correlating the feature vector with anatomical data of the joint characteristic

14. A method as defined in claim 14, wherein the joint characteristic is selected from the group consisting of cartilage thickness, osteophyte length, presence of a subchondral cyst and presence of meniscus tear.