A knee osteoarthritis multi-biosignal monitoring device and classification method

Abstract

The application provides a knee osteoarthritis multi-biological signal monitoring device and a classification method, aims to realize real-time dynamic monitoring of knee osteoarthritis, and improves the accuracy of multi-classification of knee motion signals by using multi-biological knee motion signals, and has important significance for observation of knee degenerative changes and prevention of osteoarthritis in specific applications.

Description

Technical Field

[0001] This invention relates to the field of medical testing equipment and biological signal analysis and classification technology, specifically to a monitoring device and classification method for multiple biological signals in knee osteoarthritis. Background Technology

[0002] The knee joint is a typical rotating joint and an important weight-bearing part of the human body. Long-term flexion and extension movements or incorrect movement patterns can lead to osteoarthritis of the knee, and this degenerative change is irreversible. Clinically, the Kellgren-Lawrence (KL) classification is commonly used to measure the severity of knee osteoarthritis, and it is divided into grades from mild to severe: Grade 0 (normal knee joint), Grade I, Grade II, Grade III, and Grade IV (severe osteoarthritis of the knee joint).

[0003] Currently, clinical examinations of the knee joint are divided into invasive and non-invasive methods. Invasive examinations often use arthroscopy, which requires minimally invasive surgery to achieve diagnosis, and the process is painful. Non-invasive examinations mainly include X-ray imaging, computed tomography (CT), and magnetic resonance imaging (MRI). However, X-ray imaging and CT scans involve ionizing radiation, and MRI equipment is too expensive, making them unsuitable for frequent examinations. In addition, some doctors use palpation to examine the condition of the knee joint, but all of these examinations require diagnosis by a specialist and cannot meet the needs of ordinary people for real-time dynamic monitoring of knee osteoarthritis at home. In particular, using knee joint motion information to characterize the health of the knee joint is of great significance for the study of knee osteoarthritis. However, current related studies collect knee joint motion signals in a single way, which cannot completely and accurately express the knee joint motion state information, and thus cannot classify knee joint motion signals more accurately. Moreover, knee joint motion signal classification is a multi-classification problem according to the Kellgren-Lawrence (KL) standard. Knee osteoarthritis is a degenerative change. Existing knee joint motion signal classification methods do not consider the progressive changes between different levels of signals, resulting in low recognition of early and milder knee joint motion signals.

[0004] With the development of sensor technology, wearable technology and deep learning, inspired by palpation, it has become possible to collect multi-biometric information to monitor knee osteoarthritis and classify multi-biometric knee joint motion signals. However, the knee joint structure is relatively small, and the selection of the type and number of related acquisition sensors, their placement, and the overall structure of the wearable device all affect the monitoring effect of multi-biometric signals, requiring further research. Summary of the Invention

[0005] To address the aforementioned problems, this invention proposes a monitoring device and classification method for multiple biological signals in knee osteoarthritis. The aim is to achieve real-time dynamic monitoring of knee osteoarthritis and improve the accuracy of multi-classification of knee joint motion signals by utilizing multiple biological knee joint motion signals. In practical applications, this has significant implications for observing knee degenerative changes and preventing osteoarthritis.

[0006] The technical solution of this invention is as follows:

[0007] A monitoring device for multiple biological signals in knee osteoarthritis, characterized in that it includes a data acquisition component, a main control board hardware module, a wireless communication module, a data processing module, a KL standard classification module, and a wearable structural component.

[0008] The acquisition component collects vibration signal data, acoustic emission signal data, electromyographic signal data, and knee flexion and extension angle change data of the knee joint during the human body's standing-squatting-standing process, forming a multi-dimensional knee joint motion signal;

[0009] The main control board hardware module collects multi-dimensional knee joint motion signals through a wireless communication module, and summarizes and preprocesses the data.

[0010] The data processing module preprocesses and decomposes the multi-dimensional knee joint motion signals received by the main control board hardware module to obtain signal decomposition coefficients, which are used to construct a dataset.

[0011] The KL standard classification module constructs a classification model dataset from the signal decomposition coefficients obtained by the data processing module; a multi-volume integral class model with a central ordinal loss function is selected for training to obtain a target classification model, and the target classification model is used to perform KL standard classification of motion signals of knee osteoarthritis.

[0012] The wearable structural component is worn on the human knee joint to accommodate the fixed installation of various bio-information sensors and the placement of wireless communication modules.

[0013] Furthermore, the acquisition components include a triaxial accelerometer, a contact acoustic sensor, a surface electromyography (EMG) sensor, a miniature inertial sensor, and a wireless communication module. The triaxial accelerometer is used to acquire vibration signals from the patellar position of the knee joint during the standing-squatting-standing process. The contact acoustic sensor is used to acquire acoustic emission signals from the medial and lateral compartments of the knee joint during this process. The EMG sensor is used to acquire surface EMG signals from the quadriceps femoris muscle during the standing-squatting-standing process. The miniature inertial sensor is used to acquire data on changes in knee flexion and extension angles during the standing-squatting-standing process. The wireless communication module is used for data transmission with the main control board hardware.

[0014] Furthermore, the triaxial accelerometer is placed on the patella of the knee joint; the contact acoustic sensor is placed in the medial and lateral compartments of the knee joint, respectively; the surface electromyography sensor is placed in the quadriceps femoris muscle of the thigh; and the inertial sensor is placed above the femur of the thigh above the knee joint.

[0015] Furthermore, the wearable structure includes placement positions for all sensors and wireless communication modules, allowing different sensors and wireless communication modules to be inserted into corresponding elastic shells, with pre-reserved wire outlet holes; the wearable structure sensor placement includes one accelerometer sensor located at the patella, two contact acoustic sensors located in the medial and lateral compartments of the knee joint respectively, one surface electromyography sensor located in the quadriceps muscle, and one small inertial sensor located above the femur of the thigh.

[0016] Furthermore, the wearable structural components are custom-made by 3D printing from elastic materials.

[0017] Furthermore, the main control board hardware includes a wireless communication module 1, a data acquisition card, a central microprocessor, a microprocessor, a wireless communication module 2, an audio broadcasting module, a power supply module, a status monitoring module, a display module, and a storage module. The wireless communication module 1 receives multidimensional knee joint data sent by the acquisition components. The biosignal conditioning module amplifies and filters the acquired multidimensional knee joint data, eliminating baselines and interference in various bio-information data. The data acquisition card processes and integrates the data. The central microprocessor coordinates and controls each module. The microprocessor performs data preprocessing, signal decomposition, and KL standard classification of knee osteoarthritis motion signals. The wireless communication module 2 sends the classification results report to a designated device. The audio broadcasting module provides voice prompts for the monitoring and classification processes, as well as device warnings. The power supply module provides power to each module. The status monitoring module monitors the operating status of sensors and devices. The display module displays device information, the operating status of each module, and signal classification results. The storage module stores personal bio-information, acquired multidimensional knee joint motion data, and signal classification results.

[0018] The method for classifying multiple biological signals in knee osteoarthritis using the aforementioned monitoring device includes the following steps:

[0019] Step 1: Data processing, including data normalization, periodic partitioning, and dual-density dual-tree decomposition, specifically including the following sub-steps:

[0020] Step 1.1: The collected knee joint vibration data, acoustic emission data, surface electromyography data, and angle change data during the human standing-squatting-standing process are filtered by a low-pass filter to eliminate interference from human noise and signal drift.

[0021] Step 1.2: Normalize the knee joint motion vibration data, acoustic emission data, and surface electromyography data obtained in Step 1.1, using the following formulas:

[0022]

[0023] Where x is the one-dimensional signal to be processed, x′ is the obtained one-dimensional signal, and max(x) and min(x) are the maximum and minimum values ​​in the x signal, respectively;

[0024] Step 1.3: Calculate the inflection point of the knee joint flexion and extension angle data during the human body's standing-squatting-standing cycle. Using the calculated inflection point, manually segment the normalized knee joint motion vibration data, acoustic emission data, and surface electromyography data obtained in Step 2 into corresponding cycle motion segments.

[0025] Step 1.4: Perform dual-density dual-tree complex wavelet decomposition on the knee joint motion vibration periodic segment data, acoustic emission periodic segment data and surface electromyography periodic segment data obtained in Step 1.3 to obtain the wavelet coefficients and scaling coefficients corresponding to the signal subbands of different decomposition levels for each type of data segment;

[0026] Step 2: KL standard classification processing, including the construction of a multi-channel, multi-scale information coefficient matrix dataset and the construction of a multi-volume integral class model with a center ordinal loss function, specifically including the following sub-steps:

[0027] Step 2.1: Decompose the wavelet coefficients and scaling coefficients of the same period segment in the data processing module, and then splice them together to form the corresponding column vectors. Then arrange the column vectors corresponding to each decomposition layer to form the multi-scale coefficient matrix of the data segment, namely the multi-scale coefficient matrix of the knee joint motion vibration period segment, the multi-scale coefficient matrix of the acoustic emission period segment, and the multi-scale coefficient matrix of the surface electromyography period segment.

[0028] Step 2.2: The multi-sensor information multi-scale coefficient matrices obtained in Step 2.1 are mutually corresponding according to their periods, and the multi-scale coefficient matrices within the same period are stacked and fused to form a multi-channel multi-scale information coefficient matrix;

[0029] Step 2.3: Perform the operations of Step 2.1 and Step 2.2 on all the segmented periodic data to form a multi-channel, multi-scale information coefficient matrix dataset; the operator annotates the dataset according to the KL standard and divides the annotated dataset into training and test sets;

[0030] Step 2.4: Construct the KL standard classification model with the center ordinal loss function, including the following:

[0031] a. A multi-volume integral class model was chosen as the initial model for classification;

[0032] The central ordinal loss function for multi-volume integral models is defined as follows:

[0033] and

[0034] Among them, w ij Let w represent the penalty weight between the predicted class j and the true class i, where i,j∈{1,...,n}, and w... ij ∈W, where W represents the n×n central ordinal loss matrix, i.e., the penalty matrix between the predicted and true classes; n represents the classes of motion signals for different grades of knee osteoarthritis, n=5; and p represents the predicted probability output by the softmax function of the multi-volume integral class model. Each predicted class has a penalty weight of 1; the greater the difference between the predicted and true classes, the higher the corresponding penalty weight. m represents the size of each batch of samples, x... k d represents the sample characteristics in a batch of samples. i This represents the feature center of the i-th class corresponding to the sample in the batch, and its relationship with feature x. k The dimensions should be consistent, and the sum of the squares of the distances from each sample feature to the feature center should be as small as possible, resulting in more compact intra-class samples. λ controls the weights of the loss function.

[0035] When j = i, q = 1 - p, meaning the loss is smaller when the predicted class matches the true class. The loss function can be reduced to:

[0036] and

[0037] c. Using the dataset obtained in step 2.3 and the center ordinal loss function defined in b, train the initial multi-volume integral class model in a. When the adjustable loss function reaches its minimum, the KL standard classification target model is obtained.

[0038] Step 2.5: Based on the KL standard classification target model and the multi-channel multi-scale information coefficient matrix of knee joint motion obtained in Step 2.4, identify the categories of multi-biological signals of knee osteoarthritis motion and output the classification results.

[0039] Beneficial effects

[0040] The specific beneficial effects of this invention are as follows:

[0041] 1. This invention provides a monitoring device and classification method for multiple biosignals in knee osteoarthritis, including an acquisition component for dynamically acquiring knee joint vibration signals, acoustic emission signals, electromyographic signals, and flexion-extension angle signals; a motherboard hardware module for data aggregation and other module scheduling and control; a data processing module for multi-sensor data normalization, periodic segmentation, and dual-density dual-tree decomposition; a KL standard classification module for constructing a multi-channel, multi-scale information coefficient matrix dataset and a central ordinal loss function classification model; and a wearable structural component for fixed installation of multiple sensors and placement of a wireless communication module. This invention is easy for ordinary people to wear at home for real-time dynamic monitoring of knee joint status information and KL standard classification of multiple biosignals of knee joint movement, which is beneficial for observing degenerative changes in knee osteoarthritis and preventing early osteoarthritis.

[0042] 2. In this invention, multiple biosignal sensors are used to collect information on the periodic movements of the knee joint, resulting in more complete and accurate information on the knee joint's motion state, which is beneficial for improving the accuracy of knee joint motion signal classification. Simultaneously, each sensor operates independently, ensuring that the collected data does not interfere with each other. The main control board hardware module receives data via a wireless communication module, facilitating transmission and ensuring signal accuracy.

[0043] 3. In this invention, dual-density dual-tree complex wavelet transform is used to decompose periodic segments of knee joint vibration data, acoustic emission data, and electromyographic signal data to obtain wavelet coefficients and scaling coefficients at each level for different data types. Then, the wavelet coefficients and scaling coefficients at each level of the same segment decomposition are concatenated to form corresponding column vectors. These column vectors are then arranged column-wise to form a multi-scale coefficient matrix, resulting in multi-scale coefficient matrices for different data types. Finally, the multi-scale coefficient matrices of different types within the same period are stacked and fused to form a multi-channel multi-scale information coefficient matrix. This multi-channel multi-scale information coefficient matrix integrates the finer-grained frequency characteristics of different scales from multiple sensor data, obtaining multi-dimensional and complete knee joint flexion-extension biological information. This achieves a comprehensive and accurate characterization of knee joint flexion-extension features, thereby improving the accuracy of knee osteoarthritis classification.

[0044] 4. In this invention, the KL standard classification module uses the center ordinal loss function to train a multivolume integral classification model. In deep learning, cross-entropy loss is the default loss for all classification categories, without considering the proximity between different categories. However, the categories of motion signals for different levels of knee osteoarthritis are progressively advanced. The center ordinal loss function uses different penalty weights for the predicted and true categories of multidimensional signals from different levels of osteoarthritis. The greater the distance between the two categories, the higher the penalty weight, thus widening the distance between different categories. Simultaneously, it optimizes the distance between each sample feature and the corresponding category center, enhancing intra-class compactness and improving the accuracy of multi-biosignal classification for early-stage knee osteoarthritis.

[0045] 5. In this invention, the wearable structure includes a fixing device for multiple sensors and a placement device for the wireless communication module, both 3D printed from elastic material, possessing good ductility and fixing effect, and the sensors and modules are easy to disassemble and install. The invention will be further described below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0046] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0047] Figure 1 Block diagram of the acquisition component and motherboard hardware module structure of this invention;

[0048] Figure 2 This document outlines the flowcharts for the data processing module and the KL standard classification module.

[0049] Figure 3 This method provides a schematic diagram of knee joint motion data period segmentation.

[0050] Figure 4 A schematic diagram illustrating the calculation of the central ordinal loss function in this invention;

[0051] Figure 5 Schematic diagram of the wearable structure of the present invention; (a) outer side view, (b) inner side view.

[0052] The components include: 1. a triaxial accelerometer; 2. a first contact acoustic emission sensor; 3. a second contact acoustic emission sensor; 4. a surface electromyography sensor; 5. a small inertial sensor; and 6. a wireless communication module. Detailed Implementation

[0053] The embodiments of the present invention are described in detail below. These embodiments are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0054] This embodiment proposes a monitoring device and classification method for multiple biological signals in knee osteoarthritis. The device structure is as follows: Figure 1 As shown, the purpose is to monitor knee osteoarthritis in real time and classify the multi-biological signals of knee osteoarthritis movement. The obtained parameters can be further used to observe early knee degenerative changes and prevent knee osteoarthritis.

[0055] This embodiment collects multi-dimensional biosignals during the knee joint's standing-squatting-standing process, preprocesses and decomposes the multi-biosignal data to obtain a dataset, and uses the dataset to train a multi-volume integral classification model with a central ordinal loss function to obtain a target classification model. Based on the target classification model, KL standard classification can be performed on the multi-biosignals of knee osteoarthritis, realizing real-time monitoring of knee osteoarthritis and classification of multi-biosignals of knee osteoarthritis in ordinary people at home, which is of great significance for the prevention of osteoarthritis.

[0056] First, the acquisition component collects multi-biosignal data of the knee joint during the standing-squatting-standing process. Then, the data is transmitted to the central microprocessor 1 via the wireless communication module for aggregation and processing. The microprocessor 2 preprocesses and decomposes the aggregated data and uses it for training the KL standard classification target model. The resulting target classification model is then used for multi-biosignal classification of knee osteoarthritis.

[0057] Specifically, in the application of multi-biosignal monitoring and classification for knee osteoarthritis, the device includes: an acquisition component for collecting knee joint vibration signals, first acoustic emission signals, second acoustic emission signals, electromyographic signals, and flexion-extension angle signals during the knee joint standing-squatting-standing process; a data processing module for normalizing, periodically segmenting, and decomposing the multi-biosignals; a KL standard classification module for constructing a multi-channel, multi-scale information coefficient matrix dataset, training and classifying the multi-biosignal target model for knee osteoarthritis; and an integrated wearable structural component for mounting and fixing the multi-sensor and placing the wireless communication module.

[0058] Specifically, the monitoring device for multiple biosignals in knee osteoarthritis includes: a data acquisition component, a main control board hardware module, a data processing module, a KL standard classification module, and wearable structural components.

[0059] The acquisition component collects vibration signal data, acoustic emission signal data, electromyographic signal data, and knee flexion and extension angle change data of the knee joint during the human body's standing-squatting-standing process.

[0060] The main control board hardware module collects multi-dimensional knee joint motion signals through a wireless communication module, and summarizes and processes the data.

[0061] The data processing module preprocesses and decomposes the multi-dimensional knee joint motion signals received by the main control board hardware module to obtain signal decomposition coefficients, which facilitates the construction of the dataset.

[0062] The KL standard classification module constructs a classification model dataset from the signal decomposition coefficients obtained by the data processing module; a multi-volume integral class model with a central ordinal loss function is selected for training to obtain a target classification model, and the target classification model is used to perform KL standard classification of motion signals of knee osteoarthritis.

[0063] The wearable structural component is designed to be worn on the human knee joint and accommodates the fixed installation of various bio-information sensors and the placement of wireless communication modules.

[0064] like Figure 1 As shown, the main control board hardware includes a first wireless communication module for receiving multi-sensor data, a biosignal conditioning module for amplifying and filtering the acquired multi-dimensional biosignals, a data acquisition card for processing and integrating data, a first central microprocessor for coordinating and controlling various modules, a second microprocessor for centralized data processing and KL standard classification modules, a second wireless communication module for sending osteoarthritis multi-biosignal classification reports to designated devices (mobile phones), a voice broadcast module for providing voice prompts for the knee osteoarthritis multi-biosignal monitoring and classification process and device warnings, a status monitoring module for monitoring the operating status of sensors and devices, a display module for displaying device information, the working status of each module, and signal classification results, a storage module for saving personal bio-information, multi-dimensional knee joint motion data, and signal classification results, and a power supply module for powering each module.

[0065] The wearable structure is entirely custom-made by 3D printing from elastic material, containing all the placement positions for sensors and wireless communication modules. Different sensors and wireless communication modules can be inserted into their respective elastic shells, with pre-reserved wire outlet holes. The wearable structure's sensor placement includes one accelerometer sensor located at the patella, two contact acoustic sensors located in the medial and lateral compartments of the knee joint respectively, one surface electromyography sensor located in the quadriceps muscle, and one small inertial sensor located above the femur in the thigh.

[0066] The data acquisition components include: one triaxial accelerometer for acquiring vibration signals of the patellar position of the knee joint during the standing-squatting-standing process; two contact acoustic sensors for acquiring acoustic emission signals of the medial compartment (composed of the medial tibial plateau and the medial femoral condyle) and the lateral compartment (composed of the lateral tibial plateau and the lateral femoral condyle) of the knee joint during the standing-squatting-standing process; one surface electromyography (EMG) sensor for acquiring surface EMG signals of the quadriceps femoris muscle during the standing-squatting-standing process; one small inertial sensor for acquiring data on knee flexion and extension angle changes during the standing-squatting-standing process; and a wireless communication module for data transmission with the main control board hardware.

[0067] There is at least one acquisition device; the triaxial accelerometer is placed on the patella of the knee joint; the contact acoustic sensors are placed in the medial and lateral compartments of the knee joint respectively; the surface electromyography sensor is placed in the quadriceps femoris muscle of the thigh; and the inertial sensor is placed above the femur of the thigh above the knee joint.

[0068] The data acquisition components include one triaxial accelerometer for acquiring vibration signals of the patella position during the standing-squatting-standing process, two contact acoustic sensors for acquiring acoustic emission signals of the medial and lateral compartments of the knee joint during the standing-squatting-standing process, one surface electromyography (EMG) sensor for acquiring electromyographic signals of the quadriceps muscle during the standing-squatting-standing process, one small inertial sensor for acquiring data on changes in the flexion and extension angles of the knee joint during the standing-squatting-standing process, and one wireless communication module for data transmission with the main control board hardware.

[0069] The method for classifying multiple biological signals in knee osteoarthritis using the aforementioned monitoring equipment includes the following theoretical content:

[0070] The data processing module includes data normalization, periodic segmentation, and dual-density dual-tree decomposition. Data operations include the following steps:

[0071] Step 1: Filter the knee joint vibration data, acoustic emission data, surface electromyography data, and angle change data collected by the main control board hardware module during the human standing-squatting-standing process using a low-pass filter to eliminate interference from human noise and signal drift.

[0072] Step 2: Normalize the knee joint motion vibration data, acoustic emission data, and surface electromyography data obtained in Step 1, using the following formulas:

[0073]

[0074] Where x is the one-dimensional signal to be processed, x′ is the obtained one-dimensional signal, and max(x) and min(x) are the maximum and minimum values ​​in the x signal, respectively;

[0075] Step 3: Calculate the inflection point of the knee joint flexion and extension angle data during the human body's standing-squatting-standing cycle. Using the calculated inflection point, manually segment the normalized knee joint motion vibration data, acoustic emission data, and surface electromyography data obtained in Step 2 into corresponding cycle motion segments to reduce redundant non-cycle motion data.

[0076] Step 4: Perform dual-density dual-tree complex wavelet decomposition on the knee joint motion vibration periodic segment data, acoustic emission periodic segment data and surface electromyography periodic segment data obtained in Step 3 to obtain the wavelet coefficients and scaling coefficients corresponding to the signal subbands of different decomposition levels for each type of data segment.

[0077] Preferably, the process of the KL standard classification module includes the following steps:

[0078] Step 1: Decompose the wavelet coefficients and scaling coefficients of the same period segment in the data processing module and splice them together to form the corresponding column vectors. Then arrange the column vectors corresponding to each decomposition layer to form the multi-scale coefficient matrix of the data segment, namely the multi-scale coefficient matrix of the knee joint motion vibration period segment, the multi-scale coefficient matrix of the acoustic emission period segment, and the multi-scale coefficient matrix of the surface electromyography period segment.

[0079] Step 2: The multi-scale coefficient matrices of the multi-sensor information obtained in Step 1 are mutually corresponding according to their periods, and the multi-scale coefficient matrices within the same period are stacked and fused to form a multi-channel multi-scale information coefficient matrix.

[0080] Step 3: Perform the operations of Step 1 and Step 2 on all the segmented periodic data to form a multi-channel, multi-scale information coefficient matrix dataset; specialist doctors annotate the dataset according to the KL standard, and divide the annotated dataset into training and test sets;

[0081] Step 4: Since knee osteoarthritis is a degenerative process, a KL standard signal classification model with a central ordinal loss function is constructed, including:

[0082] a. A multi-volume integral class model is chosen as the initial model for the target classification model;

[0083] The central ordinal loss function for multi-volume integral models is defined as follows:

[0084] and

[0085] Among them, w ij Let w represent the penalty weight between the predicted class j and the true class i, where i,j∈{1,...,n}, and w... ij ∈W, where W represents the n×n central ordinal loss matrix, i.e., the penalty matrix between the predicted and true classes; n represents the classes of motion signals for different grades of knee osteoarthritis, n=5; and p represents the predicted probability output by the softmax function of the multi-volume integral class model. Each predicted class has a penalty weight of 1; the greater the difference between the predicted and true classes, the higher the corresponding penalty weight. m represents the size of each batch of samples, x... k d represents the sample characteristics in a batch of samples. i This represents the feature center of the i-th class corresponding to the sample in the batch, and its relationship with feature x. k The dimensions are consistent, and it is desirable to minimize the sum of the squares of the distances from each sample feature to the feature center, resulting in more compact intra-class samples. λ controls the weights of the loss function.

[0086] When j = i, q = 1 - p, meaning the loss is smaller when the predicted class matches the true class. The loss function can be reduced to:

[0087] and

[0088] c uses the dataset obtained in step 3 and the center ordinal loss function defined in b to train the initial multi-volume integral class model in a. When the adjustable loss function reaches its minimum, the target model for KL standard classification is obtained.

[0089] Step 5: Based on the KL standard classification target model and the multi-channel multi-scale information coefficient matrix of knee joint periodic motion obtained in Step 4, determine the category of multiple biological signals of knee osteoarthritis and output the classification results.

[0090] Data processing module and KL standard classification module, such as Figure 2 As shown.

[0091] The data processing module includes data normalization, periodic partitioning, and dual-density dual-tree decomposition. The specific steps are as follows:

[0092] Step 1: The knee joint vibration data, acoustic emission 1 data, acoustic emission 2 data, surface electromyography data and angle change data collected by the main control board hardware module during the human body standing-squatting-standing process are filtered by a fourth-order Butterworth low-pass filter to eliminate human body noise and signal drift interference.

[0093] Step 2: Normalize the knee joint motion vibration data, acoustic emission 1 data, acoustic emission 2 data, and surface electromyography data obtained in Step 1, respectively, using the following formula:

[0094]

[0095] Where x is the one-dimensional signal to be processed, x′ is the obtained one-dimensional signal, and max(x) and min(x) are the maximum and minimum values ​​in the x signal, respectively;

[0096] Step 3: Calculate the inflection point of the knee flexion-extension angle data during the human standing-squatting-standing cycle. Using the calculated inflection point, manually segment the normalized knee joint motion vibration data, acoustic emission data, and surface electromyography data obtained in Step 2 into corresponding periodic motion segments to reduce redundant non-periodic motion data. Figure 3 As shown;

[0097] Step 4: Perform dual-density dual-tree complex wavelet decomposition on the knee joint motion vibration periodic segment data, acoustic emission 1 periodic segment data, acoustic emission 2 periodic segment data and surface electromyography periodic segment data obtained in Step 3 to obtain the decomposition coefficients corresponding to the signal subbands of different data at multiple scales.

[0098] Dual-density dual-tree complex wavelet transform includes two different scaling functions. and four different wavelet functions {ψ h,i ,ψ g,i}, i = 1, 2. Two wavelet functions in the same group deviate from each other by 0.5 units. Wavelet functions in different groups approximate Hilbert transform pairs, i.e.

[0099] ψ g,1 (t)≈H{ψ h,1 (t)},ψ g,2 (t)≈H{ψ h,2 (t)}

[0100] Two scaling functions and four wavelet functions constitute one complex scaling function and two complex wavelet functions, i.e.

[0101] ψ i (t)=ψ h,i (t)+jψ g,i (t) i=1,2

[0102] in, ψ are the scaling functions of low-pass filters h0 and g0, respectively; h,1 , ψ h,2 , ψ g,1 , ψ g,2 These are the wavelet coefficients of the high-pass filters h1, h2, g1, and g2, respectively. The specific decomposition formula is as follows:

[0103]

[0104] In the formula, x' represents periodic segment data; j represents the number of decomposition levels; r represents the coefficients of the wavelet function ψ during compression; A j and B r Let r represent the scaling coefficients of the j-th level decomposition and the wavelet coefficients of each compression coefficient r, where r = 1, 2, ..., j; k represents the coefficients of the scaling function and the wavelet function when shifted. Represents a series of scaling functions; ψ r,k This represents a series of wavelet functions. In this embodiment, the decomposition level is set to 3 levels, resulting in the following: the first level decomposition scale coefficient A1 and wavelet coefficient B1; the second level decomposition scale coefficient A2 and wavelet coefficients B1 and B2; and the third level decomposition scale coefficient A3 and wavelet coefficients B1, B2, and B3.

[0105] The KL standard classification module includes the construction of a multi-channel, multi-scale information coefficient matrix dataset and the construction of a multi-volume integral class model using the central ordinal loss function. The process is as follows:

[0106] Step 1: Decompose the wavelet coefficients and scaling coefficients of each decomposition layer of the periodic data segments from multiple sensors in the data processing module, and construct a multi-scale coefficient matrix using the coefficients of each layer. In this embodiment, the scaling coefficients A1 and wavelet coefficients B1 of the first layer are concatenated to form a one-dimensional column vector d1. Similarly, the decomposition coefficients of the second and third layers form one-dimensional column vectors d2 and d3, respectively. Thus, the multi-scale coefficient matrix D = [d1, d2, d3]. Similarly, the multi-scale coefficient matrix of the knee joint motion vibration periodic segment is V1 = [v1, v2, v3], the multi-scale coefficient matrix of the acoustic emission 1 data periodic segment is S1 = [s1, s2, s3], the multi-scale coefficient matrix of the acoustic emission 2 data periodic segment is S2 = [s'1, s'2, s'3], and the multi-scale coefficient matrix of the surface electromyography periodic segment is E2 = [e1, e2, e3].

[0107] Step 2: The multi-scale coefficient matrices of multiple sensor data obtained in Step 1 are correlated with each other according to the time period, and the multi-scale coefficient matrices within the same time period are stacked and fused to form a multi-channel multi-scale information coefficient matrix.

[0108] Step 3: Perform the operations of Step 1 and Step 2 on all the segmented periodic data to form a multi-channel, multi-scale information coefficient matrix dataset; specialist doctors annotate the dataset according to the KL classification standard, and divide the annotated dataset into training and test sets;

[0109] Step 4: Since knee osteoarthritis is a degenerative process, construct a KL standard classification model with a central ordinal loss function, including:

[0110] a. Choose a multi-volume integral class model as the initial model for classification, such as the ResNet series models, VGG series models, and other deep learning classification models;

[0111] The central ordinal loss function for multi-volume integral models is defined as follows:

[0112] and

[0113] Among them, w ij Let w represent the penalty weight between the predicted class j and the true class i, where i,j∈{1,...,n}, and w... ij∈W, where W represents the n×n central ordinal loss matrix, i.e., the penalty matrix between the predicted and true classes; n represents the classes of motion signals for different grades of knee osteoarthritis, n=5; and p represents the predicted probability output by the softmax function of the multi-volume integral class model. Each predicted class has a penalty weight of 1; the greater the difference between the predicted and true classes, the higher the corresponding penalty weight. m represents the size of each batch of samples, x... k d represents the sample characteristics in a batch of samples. i This represents the feature center of the i-th class corresponding to the sample in the batch, and its relationship with feature x. k The dimensions are consistent, and it is desirable to minimize the sum of the squares of the distances from each sample feature to the feature center, resulting in more compact intra-class samples. λ controls the weights of the loss function.

[0114] When j = i, q = 1 - p, meaning the loss is smaller when the predicted class matches the true class. The loss function can be reduced to:

[0115] and

[0116] c uses the dataset obtained in step 3 and the center ordinal loss function defined in b to train the initial multi-volume integral class model in a. When the adjustable loss function reaches its minimum, the KL standard classification target model is obtained.

[0117] Step 5: Based on the KL standard classification target model and the multi-channel, multi-scale information coefficient matrix of knee joint motion obtained in Step 4, identify the categories of multi-biological signals of knee osteoarthritis motion and output the classification results.

[0118] External and internal views of the wearable structure are as follows Figure 5 As shown, the device includes the placement of sensors and wireless communication modules. The sensor locations include one accelerometer sensor at the patella, two contact acoustic sensors located in the medial and lateral compartments of the knee joint, one surface electromyography (EMG) sensor in the quadriceps femoris muscle, and one small inertial sensor above the femur in the thigh. The wearable structure is entirely custom-made from elastic material using 3D printing, allowing the sensors and modules to be inserted into corresponding elastic shells. It has wire exit holes on the outside and fits snugly against the skin on the inside, making it simple and easy to disassemble.

[0119] The above is only one of the preferred embodiments of the present invention, but it is not limited to the present invention. Any substitutions and changes made to the design scheme and concept of the present invention should be included within the scope of the present invention.

[0120] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention without departing from the principles and spirit of the present invention.

Claims

1. A knee osteoarthritis multi-biosignal classification method, characterized by: Comprising the following steps: Step 1: data processing, including normalization, period segmentation and double-density double-tree complex wavelet decomposition of data, specifically comprising the following sub-steps: Step 1.1: low-pass filter the collected knee joint vibration data, acoustic emission data, surface electromyography data and angle change data during the human body standing-squatting-standing process to eliminate the interference of human body noise and signal drift; Step 1.2: normalize the knee joint motion vibration data, acoustic emission data and surface electromyography data obtained in step 1.1, the formula is as follows: wherein, is the one-dimensional signal to be processed, is the one-dimensional signal obtained, and are respectively the maximum and minimum values in the signal. Step 1.3: calculate the inflection point of the knee joint flexion and extension angle data during the human body standing-squatting-standing cycle, and manually segment the normalized knee joint motion vibration data, acoustic emission data and surface electromyography data obtained in step 1.2 into corresponding cycle motion segments using the obtained inflection point; Step 1.4: perform double-density double-tree complex wavelet decomposition on the knee joint motion vibration cycle segment data, acoustic emission cycle segment data and surface electromyography cycle segment data segmented in step 1.3 to obtain wavelet coefficients and scale coefficients corresponding to different decomposition layers of each type of data segment; Step 2: K-L standard classification processing, including multi-channel multi-scale information coefficient matrix data set construction and center ordinal loss function multi-convolution classification model construction, specifically comprising the following sub-steps: Step 2.1: splice the wavelet coefficients and scale coefficients obtained in the data processing module for each layer to form corresponding column vectors, and then arrange the column vectors corresponding to each layer of decomposition in columns to form a multi-scale coefficient matrix for the data segment, i.e., a knee joint motion vibration cycle segment multi-scale coefficient matrix, an acoustic emission cycle segment multi-scale coefficient matrix, and a surface electromyography cycle segment multi-scale coefficient matrix; Step 2.2: stack the multi-scale coefficient matrices in the same cycle according to the corresponding periods to form a multi-channel multi-scale information coefficient matrix; Step 2.3: perform steps 2.1 and 2.2 on all segmented cycle segment data to form a multi-channel multi-scale information coefficient matrix data set; the operator labels the data set according to the K-L standard, and divides the labeled data set into a training set and a test set; Step 2.4: construct a K-L standard classification model with a center ordinal loss function, including the following contents: a Select a multi-convolution classification model as the initial classification model; b The center ordinal loss function of the multi-convolution classification model is defined as follows: and , wherein, denotes the penalty weight between the predicted class and the true class , while , denotes the center ordinal loss matrix, i.e., the penalty matrix between the predicted class and the true class, denotes the class of the different levels of knee osteoarthritis movement signals, = 5, denotes the predicted probability output by the softmax function of the multi-convolution classification model; the penalty weight of each predicted class to itself is 1, and the farther the difference between the predicted class and the true class, the higher the corresponding penalty weight; denotes the size of each batch of samples, denotes the sample feature in the batch sample, denotes the feature center corresponding to the th class of the sample in the batch sample, which has the same dimension as the feature , and requires that the smaller the sum of squares of the distance of each sample feature from the feature center, the better, and the more compact the intra-class samples; controls the weight of . Let Time, That is, the loss is smaller when the predicted class is consistent with the true class, and the loss function can be reduced to: and c Train the multi-convolution classification initial model in a using the data set obtained in step 2.3 and the center ordinal loss function defined in b, and obtain the K-L standard classification target model when the adjustable loss function reaches the minimum; Step 2.5: Based on the K-L standard classification target model obtained in step 2.4 and the knee joint motion multi-channel multi-scale information coefficient matrix, identify the class of the knee joint osteoarthritis motion multi-biological signal, and output the classification result.

2. A knee osteoarthritis multi-biosignal monitoring device implementing the method of claim 1, characterized by: It comprises a collection component, a main control board hardware module and a wearing structure component; The collection component collects vibration signal data, acoustic emission signal data, muscle electrical signal data and knee joint flexion angle change data during the standing-squatting-standing process of the human body, forming multi-dimensional knee joint motion signals; The main control board hardware includes a first wireless communication module, a biological signal conditioning module, a data acquisition card, a central microprocessor, a microprocessor, a second wireless communication module, a voice broadcast module, a power supply module, a state monitoring module, a display module and a storage module; the first wireless communication module is used for receiving multi-dimensional knee joint data sent by the collection component; the biological signal conditioning module is used for amplifying and filtering the collected multi-dimensional knee joint data, and eliminating the baseline and interference of each biological information data; the data acquisition card is used for processing and integrating data; the central microprocessor is used for coordinating and controlling each module; The microprocessor is used for data preprocessing, signal decomposition and knee osteoarthritis motion signal K-L standard classification; the second wireless communication module sends the classification result reported by the system to a designated device; the voice broadcast module is used for voice prompts of the monitoring process and the classification process, and device early warning; the power supply module is used for power supply of each module; The state monitoring module is used for monitoring the running state of the sensor and the device; the display module is used for displaying device information, the working state of each module and the signal classification result; The storage module is used for saving personal biological information, collected multi-dimensional knee joint motion data and signal classification results; The wearing structure component is used for wearing on the knee joint part of the human body, and meets the fixed installation of various biological information sensors and the placement of the wireless communication module.

3. The monitoring device of claim 2, wherein: The collection component includes a three-axis acceleration sensor, a contact acoustic sensor, a surface electromyography signal sensor, a small inertial sensor and a third wireless communication module; the three-axis acceleration sensor is used for collecting vibration signals of the patella position of the knee joint during the standing-squatting-standing process of the human body; the contact acoustic sensor is used for collecting acoustic emission signals of the medial compartment and the lateral compartment of the knee joint during the standing-squatting-standing process of the human body; the surface electromyography signal sensor is used for collecting surface electromyography signals of the quadriceps femoris during the standing-squatting-standing process of the human body; The small inertial sensor is used for collecting knee joint flexion angle change data during the standing-squatting-standing process of the human body; The third wireless communication module is used for data transmission with the main control board hardware.

4. The monitoring device of claim 3, wherein: The three-axis acceleration sensor is placed on the patella of the knee joint; the contact acoustic sensor is placed in the medial compartment and the lateral compartment of the knee joint respectively; the surface electromyography signal sensor is placed on the quadriceps femoris of the thigh; and the small inertial sensor is placed above the femur of the thigh.

5. The monitoring device of claim 2, wherein: The wearing structure component contains the placement positions of all sensors and wireless communication modules, can respectively insert different sensors and wireless communication modules into corresponding elastic shells, and has reserved wire holes; the sensors include one three-axis acceleration sensor located on the patella, two contact acoustic sensors respectively located in the medial compartment and the lateral compartment of the knee joint, one surface electromyography sensor located on the quadriceps femoris, and one small inertial sensor located above the femur of the thigh.

6. The monitoring device of claim 4, wherein: The wearing structure component is customized by 3D printing of an elastic material.

Images