Portable detection system for early diagnosis of orthopedic joint diseases

Abstract

The invention discloses a portable detection system for early diagnosis of orthopaedic joint diseases, and relates to the technical field of orthopaedic joint disease diagnosis, which comprises the following steps: acquiring acoustic and motion feature vectors of joint activities, calculating a quality score based on an acoustic signal-to-noise ratio and a motion goodness of fit, performing weighted splicing on features by taking the quality score as a weight after normalization, generating a fusion feature vector; thirdly, reconstructing features by using a reverse decoding model, comparing the same degree with the original features, weighting to obtain fusion same degree, and iteratively adjusting the weight to optimize the fusion effect; and performing clustering analysis on the candidate feature vectors, replacing the candidate feature vectors if the candidate feature vectors deviate from a clustering center, and finally inputting a health model to output a diagnosis report. Through dynamic weighting of quality scores, high-fidelity features are processed preferentially, the information density and discrimination capability are improved, and low-quality data interference is reduced; and self-checking and dynamic correction are realized through reconstruction and similarity comparison, so that the fidelity and consistency of the characteristics are ensured, and the robustness of the system is enhanced.

Description

Technical Field

[0001] This invention relates to the field of orthopedic joint disease diagnostic technology, and more specifically, this application relates to a portable detection system for the early diagnosis of orthopedic joint diseases. Background Technology

[0002] With the development of precision medicine and mobile health technologies, the early screening and diagnosis of orthopedic joint diseases is gradually shifting from a hospital-centered model that relies on large imaging equipment to a portable, community-based, and home-based model. To improve the comprehensiveness and accuracy of diagnosis, existing portable testing systems have begun to integrate multimodal data, such as simultaneously acquiring acoustic signals (e.g., friction sounds, popping sounds) and kinematic signals (e.g., joint angles, angular velocities) generated during joint movement.

[0003] These existing technologies typically employ a simple "acquisition-fusion-diagnosis" pipeline. The common practice is to acquire acoustic and motion data separately using microphone arrays and inertial measurement units (IMUs), extract features, directly concatenate them, or fuse them using fixed weights. The fused feature vector is then input into a pre-trained classification model (such as a support vector machine or neural network) to directly output the diagnostic result. The advantage of this type of technology is that it initially utilizes the complementarity of multimodal information, theoretically providing richer diagnostic criteria compared to single-modal systems.

[0004] However, this traditional multimodal fusion method lacks dynamic evaluation of the data quality itself and a closed-loop verification mechanism for the reliability of the fusion results. Specifically, after feature fusion, there is a lack of reverse verification of feature validity; even when reverse verification is performed, it is often based on isolated single detection data, lacking an "experience base" built using historical health data to perform secondary verification of the current detection results. Therefore, a portable detection system for the early diagnosis of orthopedic joint diseases is proposed to address this problem. Summary of the Invention

[0005] To address the aforementioned technical problems, a portable detection system for the early diagnosis of orthopedic joint diseases is provided. This technical solution resolves the issues raised in the background section.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] This application discloses a portable detection system for the early diagnosis of orthopedic joint diseases, including: The data acquisition module is used to acquire acoustic feature vectors and motion feature vectors during joint movement. The vector acquisition module is used to take the signal-to-noise ratio of the acoustic feature vector as the acoustic quality score and the degree of fit between the motion feature vector and the preset standard motion feature vector as the motion quality score. After normalizing the acoustic quality score and motion quality score, weight coefficients are obtained. Based on the weight coefficients, the acoustic feature vector and motion feature vector are weighted and concatenated to obtain the fused feature vector. The vector reconstruction module is used to input the fused feature vector into the inverse decoding model to reconstruct the acoustic feature vector and motion feature vector; The reconstructed acoustic feature vector and motion feature vector are compared with the acoustic feature vector and motion feature vector before weighted splicing to obtain the acoustic similarity and motion similarity. Based on the weighting coefficients, acoustic similarity and motion similarity are weighted and fused to obtain fusion similarity. If the fusion similarity is less than the preset similarity threshold, the weight coefficients are adjusted and the fusion similarity is recalculated until the fusion similarity is greater than the similarity threshold or the number of re-splicing reaches the preset maximum number of iterations. The processed fusion feature vector is then used as a candidate feature vector. The effective judgment module is used to divide the candidate feature vector and historical candidate feature vector into clusters, calculate the cluster distance from the candidate feature vector to the cluster center and the distance threshold obtained from the historical candidate feature vector within the cluster. If the cluster distance is greater than the distance threshold, the historical candidate feature vector at the cluster center is used to replace the current candidate feature vector as the final feature vector. The risk diagnosis module is used to input the final feature vector into the health model to obtain a joint diagnosis report.

[0008] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This application uses a quality score based on the signal-to-noise ratio of acoustic feature vectors and the degree of agreement between motion feature vectors to weight and concatenate acoustic and motion feature vectors to generate a fused feature vector. This solves the problem of blind feature weight allocation and inability to distinguish data reliability when simply fusing multimodal data in the prior art. It achieves dynamic priority weighting of high-fidelity features, improves the information density and discriminative ability of the fused feature vector, and avoids interference of low-quality data on the diagnostic model. 2. This application reconstructs acoustic and motion feature vectors through a reverse decoding model and compares them with the original features. The fusion similarity is obtained by weighting based on quality scores, and an iterative adjustment mechanism is set up to overcome the shortcomings of existing systems in lacking feature fusion effectiveness verification and closed-loop optimization. This enables self-verification and dynamic correction of the feature fusion process, ensures the fidelity and consistency of feature information, and enhances the robustness of the system to noise and abnormal data. Attached Figure Description

[0009] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. Wherein: Figure 1 This is a structural block diagram of the portable detection system for early diagnosis of orthopedic joint diseases proposed in this invention; Figure 2 This is a flowchart illustrating the standard motion feature vector acquisition process in this invention. Figure 3 This is a flowchart of the candidate feature vector acquisition process in this invention; Figure 4 This is a flowchart of the method for clustering effectiveness analysis of reassembled fused feature vectors in this invention. Detailed Implementation

[0010] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0011] While existing portable detection systems for orthopedic joint diseases have incorporated multimodal data fusion, they generally employ a simple linear "acquisition-fusion-diagnosis" process, lacking dynamic evaluation of data quality and a closed-loop verification mechanism for the reliability of fusion results. Specifically, feature fusion lacks reverse validation of its effectiveness and relies solely on single-test data, failing to utilize historical health data for secondary verification, thus limiting the accuracy and robustness of diagnostic results.

[0012] To address the aforementioned issues, this application calculates and normalizes the quality scores of acoustic and motion features into weight coefficients, then weights and concatenates the feature vectors to obtain a fused feature vector. Next, a reverse decoding model is used to reconstruct the feature vector, and reverse verification is performed by comparing the similarity between the reconstructed and original vectors. The weight coefficients are iteratively adjusted until a similarity threshold or the maximum number of iterations is met, outputting candidate feature vectors. Subsequently, based on historical candidate feature vector clustering analysis, cluster distance and thresholds are calculated, and abnormal candidate feature vectors are replaced to obtain the final feature vector. Finally, the vector is input into a health model to generate a joint diagnosis report. This application achieves closed-loop optimization and reliability improvement of multimodal fusion through dynamic quality assessment, reconstruction verification, and historical data clustering validation, enhancing the accuracy and adaptability of early diagnosis.

[0013] like Figure 1 As shown, a portable detection system for the early diagnosis of orthopedic joint diseases is introduced, including: The data acquisition module 100 is used to acquire acoustic feature vectors and motion feature vectors during joint movement. The vector acquisition module 200 is used to take the signal-to-noise ratio of the acoustic feature vector as the acoustic quality score and the degree of fit between the motion feature vector and the preset standard motion feature vector as the motion quality score; after normalizing the acoustic quality score and the motion quality score, weight coefficients are obtained; based on the weight coefficients, the acoustic feature vector and the motion feature vector are weighted and concatenated to obtain the fused feature vector. The vector reconstruction module 300 is used to input the fused feature vector into the inverse decoding model to reconstruct the acoustic feature vector and motion feature vector; compare the reconstructed acoustic feature vector and motion feature vector with the acoustic feature vector and motion feature vector before weighted splicing to obtain the acoustic similarity and motion similarity; perform weighted fusion of acoustic similarity and motion similarity based on weight coefficients to obtain fusion similarity; if the fusion similarity is less than the preset similarity threshold, adjust the weight coefficients and recalculate the fusion similarity until the fusion similarity is greater than the similarity threshold or the number of re-sponsoring reaches the preset maximum number of iterations, and use the processed fused feature vector as the candidate feature vector; The effective judgment module 400 is used to divide the candidate feature vector and historical candidate feature vector into clusters, calculate the cluster distance from the candidate feature vector to the cluster center and the distance threshold obtained from the historical candidate feature vector within the cluster. If the cluster distance is greater than the distance threshold, the historical candidate feature vector at the cluster center is used to replace the current candidate feature vector as the final feature vector. The risk diagnosis module 500 is used to input the final feature vector into the health model to obtain a joint diagnosis report.

[0014] Regarding step 100: Portable sensors (such as microphones and inertial measurement units) are worn at the joints to simultaneously acquire triboacoustic data representing internal joint friction and vibration, as well as motion data representing changes in the spatial position of the joint during the target joint's movement cycle. The motion data includes the maximum range of motion, average angular velocity, and acceleration indicators representing the smoothness of the movement.

[0015] The process of acquiring triboacoustic data includes: Acquire the raw acoustic signal sequence generated by joint movement from the microphone array. .

[0016] For the original acoustic signal sequence An adaptive filtering algorithm is used to remove environmental noise from the original sound waveform data based on reference noise samples, thereby obtaining clean acoustic data after noise reduction.

[0017] Through noise reduction processing technology, this application achieves dual purification of the original triboacoustic data: by using an adaptive filtering algorithm to remove environmental noise based on reference noise samples, it solves the problem that triboacoustic data in existing portable detection systems is easily affected by environmental noise interference, effectively improving the purity and signal-to-noise ratio of the acoustic signal, providing a high-fidelity data foundation for subsequent short-time Fourier transform and feature extraction, enhancing the system's ability to sensitively capture the subtle friction sound features of early joint lesions, and avoiding the risk of missed diagnosis caused by noise masking.

[0018] The process of acquiring motion data includes: The joint angle data of the target joint acquired by the inertial measurement unit in a time sequence over one complete motion cycle is denoted as follows: ,in This refers to the sampling time point.

[0019] Calculation of motion amplitude index: based on joint angle time series Calculate its maximum value and minimum value The difference, as the maximum angle of activity :

[0020] Calculation of motion velocity index: time series of joint angles The angular velocity sequence is obtained by performing first-order differentiation. And calculate the average of the absolute values ​​of the angular velocity sequence as the average angular velocity. :

[0021] Motion smoothness index calculation: for angular velocity sequences The angular acceleration sequence is obtained by performing first-order differentiation. And then The acceleration sequence is obtained by performing first-order differentiation. Calculate the root mean square value of the acceleration sequence as an acceleration index. :

[0022] Through the aforementioned data acquisition technology, this application achieves simultaneous acquisition and quantitative characterization of acoustic and motion multimodal data within the target joint's motion cycle. By coordinating the microphone array and inertial measurement unit, the acoustic signal sequence and joint angle time series generated by joint activity are simultaneously captured, providing a raw data foundation for subsequent analysis. By calculating the maximum range of motion, average angular velocity, and acceleration indices, multidimensional quantification of joint motion amplitude, velocity, and smoothness is achieved, solving the problem of incomplete assessment using traditional single motion parameters. The introduction of the acceleration index can sensitively reflect subtle abnormal changes in motion smoothness, providing crucial quantitative evidence for early joint functional degeneration. This data acquisition scheme provides highly reliable, multidimensional data support for subsequent feature fusion and disease risk diagnosis.

[0023] The process of obtaining acoustic feature vectors includes: The processed original acoustic signal sequence Perform a short-time Fourier transform to obtain its spectral representation. ; Based on spectrum Calculate the Mel-frequency cepstral coefficients of each frame of the signal to construct the MFCC feature matrix. ,in For filter bank indexes on the Mel scale; For the MFCC characteristic matrix Along the time axis Statistical features are extracted, and the mean, standard deviation, skewness, and kurtosis are calculated and concatenated to form an acoustic feature vector. .

[0024] The process of standardizing motion data to obtain motion feature vectors includes:

[0025] Feature standardization sub-step: Calculate the maximum active angle Mean angular velocity and acceleration index Standardize each quantity separately to eliminate the influence of dimensions. The standardization formula is:

[0026] in, , , These are the means of each indicator obtained in advance from the dataset. , , This represents the corresponding standard deviation.

[0027] Standardized indicators , , Combined in a predetermined order, they form a one-dimensional kinematic feature vector. For subsequent analysis:

[0028] The acoustic feature vector and motion feature vector are normalized using the min-max method.

[0029] Through the above technical solutions, this application achieves refined construction and standardized processing of acoustic and motion feature vectors. By performing short-time Fourier transform on the acoustic signal and extracting the statistical features of the Mel frequency cepstral coefficients, an acoustic feature vector is constructed, solving the problem of incomplete capture of weak acoustic features such as joint friction sounds by traditional methods, and realizing multi-dimensional quantitative representation of the time-frequency characteristics of acoustic signals. By standardizing the motion indicators based on a standard dataset, the dimensional differences between the maximum angle of motion, average angular velocity, and acceleration indicators are eliminated, solving the problem of inconsistent time scales when directly fusing multi-source motion parameters. Furthermore, the min-max method is used to normalize the feature vectors, unifying acoustic and motion features with different physical meanings and magnitudes into the same numerical range, laying a comparability foundation for subsequent weighted fusion based on quality scores, and improving the coordination of multimodal feature fusion and the accuracy of system decision-making.

[0030] Regarding step 200:

[0031] The process of calculating the acoustic quality score includes:

[0032] Based on acoustic feature vectors Calculate acoustic quality score The calculation formula is as follows:

[0033] in, For feature vector-based The calculated signal-to-noise ratio is used to quantify the energy ratio of the effective signal component to the noise component; These are normalized weighting coefficients used to map the signal-to-noise ratio to scores within a preset scoring range.

[0034] Furthermore, a correlation analysis was conducted between the signal-to-noise ratio of the acoustic feature vectors and the clinical diagnostic results of joint friction sounds. The frequency band weights in the signal-to-noise ratio calculation are dynamically adjusted based on the results of correlation analysis, prioritizing the retention of frequency bands with high correlation to pathological acoustic features.

[0035] This application also proposes to verify the validity of the acoustic quality score after obtaining it, including: Obtain the historical acoustic feature vector that is the same as the current acoustic feature vector and motion feature vector, the historical acoustic quality score corresponding to the historical motion feature vector, and the current acoustic quality score; The average historical acoustic quality score is calculated, and the difference between this average and the current acoustic quality score is used to calculate the first deviation value. Calculate the second deviation value between the historical acoustic quality score and the mean score; Historical acoustic quality scores with a second deviation value less than the first deviation value were selected, and their mean was calculated. Replace the current acoustic quality score with the mean value.

[0036] Through the above technical solution, this application achieves real-time validity verification and dynamic correction of acoustic quality scores. By performing deviation analysis and mean substitution between the current acoustic quality score and historical score data, it effectively identifies and smooths score distortion caused by transient interference or data anomalies, improving the stability and reliability of the quality scoring system. This adaptive correction mechanism can optimize the current score using historical experience data without introducing additional detection procedures, ensuring both real-time response and enhanced robustness to noise data. Furthermore, by establishing a dynamic verification process based on historical benchmarks, a closed-loop management technology for data quality is formed, further improving the accuracy of early diagnosis of joint diseases and the confidence of the overall system decision-making.

[0037] The process of calculating the performance quality score includes: Acquire raw motion data generated by joint movement, including maximum range of motion. Mean angular velocity and acceleration indicators And constitute the original motion feature vector. .

[0038] For the original motion feature vector The standardization process is performed to obtain the standardized motion feature vector. The processing formula is:

[0039] in, This is the vector of mean values ​​for each calculated athletic indicator. This is the corresponding standard deviation vector.

[0040] Based on standardized motion feature vectors Calculate the quality of motion score By calculating its feature vector compared to the standard motion trajectory This is achieved through the degree of matching, and its calculation formula is:

[0041] in, for and The difference measurement function between them is preferably Euclidean distance or Mahalanobis distance.

[0042] like Figure 2 The process of obtaining standard motion feature vectors is illustrated below: Simultaneously collect joint angle data of the target joint during the motion cycle from healthy individuals; The joint angle data is processed to obtain motion data, which includes the maximum range of motion, average angular velocity, and acceleration indices. The extracted motion data is then statistically analyzed and normalized, and the mean value of the motion data is calculated as the standard motion feature vector.

[0043] This application also proposes updating the standard motion feature vector after obtaining it, including: Regularly collect new sample data from healthy individuals to obtain updated standard motion feature vectors; Calculate the Mahalanobis distance between the updated motion feature vector and the existing standard motion feature vector; When the Mahalanobis distance exceeds the preset Mahalanobis distance threshold (e.g., 0.1), the existing standard motion feature vector is replaced with the updated standard motion feature vector.

[0044] Furthermore, hierarchical standard motion feature vector groups can be established based on age and gender factors to improve the group adaptability of motion quality scores.

[0045] Through the aforementioned technical solution for acquiring standard motion trajectories, this application achieves objective quantification and standardized modeling of healthy joint movement patterns. By collecting standard motion data from healthy individuals and extracting key motion indicators, it overcomes the limitations of traditional assessments that rely on subjective experience or single reference values. Through statistical analysis and normalization of multi-dimensional motion indicators, a standard motion trajectory vector representing the movement characteristics of healthy joints is established. By periodically collecting new sample data from healthy individuals and determining the update timing based on Mahalanobis distance thresholds, the timeliness and representativeness of the standard motion feature vector are ensured, effectively avoiding benchmark drift caused by changes in population characteristics or equipment aging. By establishing stratified standard motion feature vector groups based on age and gender, the applicability of a single standard to groups with different physiological characteristics is addressed, improving the accuracy and group adaptability of motion quality scoring. This solution provides a scientific and reliable benchmark for subsequent motion quality scoring, effectively improving the system's sensitivity and specificity in identifying early joint dysfunction, and enhancing the objectivity and comparability of diagnostic results.

[0046] The process of weighting and concatenating feature vectors based on quality scores to obtain a fused feature vector includes: Obtaining acoustic feature vectors Motion feature vectors and the corresponding acoustic quality score. and exercise quality score .

[0047] Based on acoustic quality score and exercise quality score Calculate the weights of the acoustic eigenvectors respectively. and the weights of the motion feature vectors The weight calculation formula is:

[0048] Using the calculated weights and For acoustic feature vectors and motion feature vectors Perform weighted fusion to generate a fused feature vector. The fusion formula is:

[0049] in, This represents a vector concatenation operation, which involves first concatenating the weighted feature vectors... and Connect the two ends together to form a higher-dimensional fusion feature vector.

[0050] Output fused feature vector This is used for subsequent reverse decoding and health status analysis.

[0051] Through the above technical solutions, this application achieves standardized processing, quality score quantification, and dynamic weighted fusion of acoustic and motion feature vectors. By standardizing motion features based on the mean and standard deviation of a standard dataset, the dimensional differences of multiple indicators are eliminated, providing a unified benchmark for subsequent fusion analysis. By calculating the signal-to-noise ratio of acoustic feature vectors as a quality score, and using the fit between standardized motion features and standard trajectories to calculate the motion quality score, an independent quantitative assessment of the reliability of multimodal data is achieved. By dynamically assigning weights based on the quality score and performing weighted concatenation to generate fused feature vectors, the problem of fixed weight allocation in traditional multimodal fusion is solved, achieving the dominant fusion of high-reliability features, improving the discriminative information quality of the fused feature vectors and the reliability of system decisions, and providing a high-quality data foundation for subsequent reverse verification and disease risk diagnosis.

[0052] Regarding step 300: like Figure 3The flowchart shown illustrates the process of obtaining candidate feature vectors.

[0053] The process of reconstructing acoustic and motion feature vectors by inputting the fused feature vectors into a pre-trained inverse decoding model includes: Obtain the fused feature vector and load the pre-trained inverse decoding model. .

[0054] fuse feature vectors As input, it is fed into the pre-trained inverse decoding model. The model performs inverse decoding, mapping the fused features back to the original acoustic and motion feature spaces, thereby reconstructing the acoustic feature vectors in parallel. and motion feature vectors This reconstruction process is achieved through the model's parameterized mapping function, which is mathematically expressed as:

[0055] in, Indicates the reverse decoding model The model is trained using a pre-trained set of parameters to minimize the difference between the reconstructed feature vector and the original input feature vector.

[0056] Output the reconstructed acoustic feature vector and motion feature vectors The dimensions of the reconstructed vectors are respectively the same as those of the acoustic feature vectors before weighted concatenation. and motion feature vectors The dimensions correspond to each other.

[0057] The process of constructing the reverse decoding model includes: Obtain the training dataset, which contains multiple sets of synchronously acquired, preprocessed standard acoustic feature vectors. and standard motion feature vector ,in For sample index, The total number of samples.

[0058] Forward fusion sub-step: Based on a pre-defined quality score or derived from the data. and The standard feature vectors in each group of samples are weighted and fused to generate a fused feature vector. As training input:

[0059] Model building and training sub-steps: Building an inverse decoding model The model contains at least one shared encoder network and one decoder network with two parallel output branches; it fuses feature vectors. The input model and its computation process are represented as follows:

[0060] in, and These are the acoustic and motion feature vectors reconstructed from the model, respectively. This is the set of parameters to be optimized in the model. The loss function is then optimized. For model parameters During training, the loss function aims to minimize the difference between the reconstructed feature vector and the original standard feature vector, and is defined as follows:

[0061] in, and The hyperparameters used to balance the magnitudes of the two reconstruction losses This represents the L2 norm of a vector. The training process iteratively optimizes the parameters using the backpropagation algorithm. This continues until the loss function converges.

[0062] Save the final model parameters obtained after training. Thus, a pre-trained inverse decoding model is obtained. , used to complete the refactoring task.

[0063] Furthermore, the output layer can use the Sigmoid activation function to restrict the reconstructed feature vector to the range [0, 1].

[0064] The average decoding confidence of the inverse decoding model is periodically obtained within the calibration time window. If the average decoding confidence is consistently lower than the preset decoding confidence threshold (e.g., 0.85), the model is updated or retrained. The calibration time window is set according to actual needs.

[0065] The process of comparing the reconstructed feature vector with the feature vector before weighted fusion to obtain acoustic and motion similarity includes: Obtain the original acoustic feature vector before weighted fusion Compared with the original motion feature vector And the acoustic feature vector reconstructed by the inverse decoding model and motion feature vectors .

[0066] Calculate the original acoustic eigenvectors With reconstructed acoustic feature vectors Acoustic similarity between This similarity is quantified by calculating the cosine similarity between the two vectors, using the following formula:

[0067] in, The dimension of the acoustic feature vector. and Represent the first and second digits of the original vector and the reconstructed vector, respectively. One portion, Represents the vector dot product. This represents the L2 norm of a vector.

[0068] Calculate the original motion feature vector With reconstructing motion feature vectors similarity of motion between This similarity is also quantified using cosine similarity, and its calculation formula is:

[0069] in, The dimension of the motion feature vector. and Represent the first and second digits of the original vector and the reconstructed vector, respectively. Each component.

[0070] Output the calculated acoustic similarity Similarity to movement The similarity is a scalar value between -1 and 1. The closer the value is to 1, the higher the similarity between the reconstructed vector and the original vector.

[0071] Furthermore, the scalar values ​​are normalized to ensure that the output values ​​fall within the [0,1] interval: Obtain the original acoustic feature vector before weighted fusion Compared with the original motion feature vector And the acoustic feature vector reconstructed by the inverse decoding model and motion feature vectors .

[0072] Calculate the original acoustic eigenvectors With reconstructed acoustic feature vectors Acoustic similarity between This similarity is quantified by calculating the normalized cosine similarity between the two vectors to ensure that its value range is [0,1]. The calculation formula is as follows:

[0073] in, Let be the dimension of the acoustic feature vector. This formula linearly maps the range of cosine similarity from [-1, 1] to [0, 1].

[0074] Calculate the original motion feature vector With reconstructing motion feature vectors similarity of motion between Using the same normalization process, its calculation formula is:

[0075] in, is the dimension of the motion feature vector.

[0076] Output the calculated acoustic similarity Similarity to movement The similarity is a scalar value between 0 and 1. The closer the value is to 1, the higher the similarity between the reconstructed vector and the original vector.

[0077] The process of weighted fusion of similarity scores based on weight coefficients to obtain the fused similarity score includes: Obtaining acoustic similarity Similarity to movement and acoustic quality rating and exercise quality score .

[0078] Based on acoustic quality score and exercise quality score Calculate the fusion weights for acoustic similarity respectively. Fusion weights based on motion similarity The weight calculation formula is:

[0079] Using the calculated weights and Acoustic similarity Similarity to movement Perform weighted fusion to generate fusion similarity. The fusion formula is:

[0080] This linear weighted fusion ensures that the similarity contribution of the acoustic and motion dimensions is proportional to their corresponding data quality.

[0081] Output the calculated fusion similarity This value is a scalar between 0 and 1, used for subsequent quality control judgments.

[0082] The process of iteratively adjusting the weighting strategy based on fusion similarity includes: Obtain the initial fusion similarity Preset similarity threshold Maximum number of iterations and initial acoustic quality score and exercise quality score .

[0083] Set the iteration counter Determine the current degree of similarity in fusion. Is it less than the same degree threshold? And the number of iterations Is it less than the maximum number of iterations? If the conditions are met, proceed to the weight adjustment sub-step.

[0084] Based on the current iteration number The fusion weights for acoustic similarity are adjusted based on the initial quality score. Fusion weights based on motion similarity The adjustment strategy is to gradually increase the weight of features with lower quality scores to balance the contributions of both. The adjustment formula is as follows:

[0085]

[0086] in, The rate coefficient is used to control the magnitude of weight adjustment. After adjustment, the weights need to be normalized.

[0087]

[0088] Use the adjusted weights and Recalculate the fusion similarity :

[0089] Update iterative counter Then return the iterative judgment sub-step.

[0090] when or When the iteration terminates, the final adjusted fusion similarity is output. and the corresponding final weights and This is used for subsequent joint health status analysis. The degree of fusion similarity at present.

[0091] Through the above technical solutions, this application achieves real-time verification and dynamic optimization of the feature fusion process. By reconstructing acoustic and motion feature vectors through an inverse decoding model and comparing them with the original features using normalized cosine similarity, the problem of lacking closed-loop verification of feature fusion effectiveness in existing systems is solved, ensuring the information fidelity and consistency of the fused features. By weighting the similarity based on quality scores and introducing an iterative adjustment mechanism to dynamically optimize the weight allocation, the system automatically suppresses the contribution of low-quality data and strengthens the dominant role of high-confidence features, improving its adaptability to noise and abnormal data. Furthermore, by setting a similarity threshold and a maximum number of iterations as termination conditions, the controllability and reliability of the fusion results are ensured, thereby enhancing the sensitivity and specificity of early diagnosis of joint diseases.

[0092] This application further proposes the implementation process of the gradient descent algorithm: Get the current weight vector ,in, The weights of the acoustic feature vectors. The weights are assigned to the motion feature vectors, with initial weights set based on acoustic quality scores and motion quality scores; and the current fusion similarity is obtained. ;

[0093] Calculate fusion similarity For weight vector gradient The gradient characterizes the direction and extent of the influence of a small adjustment in weights on the fusion similarity, where the gradient components are... and This indicates the direction of change of the fusion similarity function at the current weight, used to guide the direction of weight updates. Based on the gradient and its inverse direction, the current acoustic and motion weights are fine-tuned according to a preset learning rate to make the fusion similarity evolve in the direction of increasing. The new weights of the weight vector are then updated using the gradient descent algorithm. The updated formula is:

[0094] in, As the initial weights, The learning rate is used to control the update step size. To merge the similarity function In weighting coefficients The gradient vector at that point; Output the updated weight vector And used to recalculate the fusion similarity; Furthermore, in the gradient descent algorithm, the learning rate is dynamically adjusted based on changes in fusion similarity. Specifically, if the improvement in fusion similarity in consecutive iterations is less than a preset improvement threshold (e.g., 0.05), the learning rate is reduced. To refine the search; if the similarity of the merged components fluctuates significantly, the learning rate is temporarily increased. To accelerate convergence.

[0095] The above steps are executed iteratively until the fusion similarity is greater than the same threshold or the preset maximum number of iterations is reached. When the maximum number of iterations is reached but the fusion similarity still does not meet the same threshold, the system selects the weight combination corresponding to the highest fusion similarity value recorded during the iteration process as the final weight coefficient for output.

[0096] Through the above technical solution, this application achieves closed-loop adaptive optimization of feature fusion weights. By calculating the gradient of the fusion similarity with respect to the weights and iteratively adjusting in the opposite direction, it solves the problem that traditional fixed weights or simple rule adjustments cannot achieve fine optimization, enabling the weight allocation to automatically evolve towards maximizing the fusion similarity. The algorithm controls the update step size through a preset learning rate, ensuring the stability and convergence efficiency of the optimization process, thus quickly approaching the optimal weight configuration within a limited number of iterations. When the maximum number of iterations is reached and the fusion similarity still does not meet the threshold, the weight combination corresponding to the highest fusion similarity value recorded in historical iterations is selected as the final output, solving the problem that the optimization algorithm may fail to output usable results due to convergence difficulties. This intelligent optimization mechanism improves the accuracy of feature fusion and the system's adaptive capability, providing a reliable guarantee for the early and accurate diagnosis of joint diseases.

[0097] Regarding step 400: like Figure 4 As shown, the process of performing clustering effectiveness analysis on candidate feature vectors is introduced; The candidate feature vectors are merged with the historical candidate feature vectors to form a feature vector set; Based on clustering algorithms, the feature vectors in the feature vector set are divided into different clusters that represent the features of joint activity. Calculate the clustering distance from the current candidate feature vector to the center of its cluster, and compare this distance with the distance threshold calculated based on all historical candidate feature vectors within the cluster. If the cluster distance exceeds the distance threshold, the historical candidate feature vector of the center of its cluster is used as the replacement.

[0098] Specifically, this includes: obtaining candidate feature vectors and historical candidate feature vectors, and merging them to form a feature vector set. ,in Indicates the first 1 eigenvector The total number of vectors; Based on clustering algorithms (such as K-means algorithm), the feature vector set is divided into Divided into sum clusters In this context, each cluster represents a specific joint activity characteristic; For the current candidate feature vector Determine its cluster. , belong One of the items;

[0099] calculate To its cluster center Cluster distance The Euclidean distance formula is used:

[0100] Calculate clusters Distance threshold Based on all historical fusion feature vectors within this cluster to the center Distance statistics Specifically:

[0101] in, For clusters The average distance from the inner history vector to the center. Standard deviation This is a preset constant (usually taken as...). ).

[0102] Comparing cluster distances With threshold : like Then clustering is used. Central historical candidate feature vector Replace the current candidate feature vector ; Otherwise, keep These will be used as valid feature vectors for subsequent processing.

[0103] If the difference between the cluster distance and the distance threshold is greater than If the standard deviation is more than one-fold, it indicates that the current candidate feature vector is an outlier, possibly due to noise or abnormal activity. In this case, the historical candidate feature vector of the cluster center is used to replace the current candidate feature vector as the final feature vector. Otherwise, the current candidate feature vector is used directly as the final feature vector.

[0104] Through the aforementioned clustering effectiveness analysis technique, this application achieves anomaly identification and adaptive correction of candidate feature vectors. By performing cluster analysis on the current candidate feature vector and the historical candidate vector set and calculating its distance to the cluster center, the problem of traditional methods lacking an effective correction mechanism for anomaly feature vectors is solved. By dynamically calculating the distance threshold based on the statistical features of historical data and replacing anomaly vectors exceeding the threshold with cluster center vectors, outliers caused by data acquisition interference or fusion distortion are effectively filtered out, improving the quality and consistency of feature data input to the health model. This scheme enhances the system's fault tolerance to noisy data and the stability of output results, providing a guarantee for reliable diagnosis of joint disease risks.

[0105] Regarding step 500:

[0106] The process of generating joint indices and diagnostic reports based on the final feature vector includes: Obtain the final feature vector and load the health model .

[0107] The final feature vector Input health model The joint index, which characterizes the health of the joints, was calculated. The calculation process is expressed as follows:

[0108] in, This is the set of pre-trained parameters for a health model. The model is trained through supervised learning and can map fused features to a standardized health index scale.

[0109] The calculated joint index The risk level of joint disease is determined by comparing it with a preset risk threshold range. :

[0110] in, and This is a preset risk threshold.

[0111] Based on joint index and risk level Automatically generate a joint disease risk diagnosis report containing the following: Joint health index values ​​and corresponding health status descriptions; Health advice and warning information based on risk levels; Abstract of quantitative analysis of acoustic and motion characteristics.

[0112] Output joint index and structured joint disease risk diagnostic reports The report is presented in the form of an electronic document or a visual chart to assist in medical decision-making.

[0113] The process of building a health model includes:

[0114] Obtain the training dataset, which contains multiple sets of sample data, each set including: The fused feature vector generated by weighted concatenation and corresponding joint health status labels This label is quantified from clinical diagnostic results, among which For sample index, The total number of samples.

[0115] Building a health model The model consists of an input layer, at least one hidden layer, and an output layer, wherein: Input layer dimension and fused feature vector Dimensions Consistent; The output layer consists of a single neuron and is used to output the joint index. ; The hidden layers use non-linear activation functions, and the output layer uses the Sigmoid function to restrict the output to the range [0,1].

[0116] Use training data to adjust model parameters Optimization is performed by minimizing the loss function between the predicted joint index and the true health label. :

[0117] in, This is a regularization term used to prevent overfitting; training employs the backpropagation algorithm to iteratively optimize the parameters. This continues until the loss function converges.

[0118] Save the final model parameters obtained after training. To obtain a pre-trained healthy model It is used to assess the health of joints.

[0119] The average health confidence of the health model is periodically obtained within the data processing calibration window. If the average health confidence is consistently lower than the preset health confidence threshold (e.g., 0.85), the model is updated or retrained. The calibration window is set according to actual needs.

[0120] Through the above technical solution, this application achieves quantitative assessment of joint health status and automated diagnostic report generation. By inputting the optimized fused feature vector into the health model, standardized joint indices are directly output, solving the problems of traditional diagnostic methods relying on subjective experience and being insensitive to early lesions, thus achieving an objective and quantitative representation of joint health status. By comparing the joint index with risk threshold ranges, low, medium, and high risk levels are automatically classified, and a structured diagnostic report containing a health status description, early warning information, and feature analysis summary is generated, solving the problems of unintuitive diagnostic results and lack of clinical guidance in existing portable systems. This solution combines multimodal feature fusion with intelligent decision-making, improving the automation level of the screening process and the reliability of diagnostic results, providing an efficient and reliable auxiliary tool for early intervention of joint diseases in community and home settings.

[0121] The above details the implementation steps of a portable detection system for early diagnosis of orthopedic joint diseases. The following section describes a specific application scenario using this embodiment: Scenario: A community health service center in a city plans to conduct a free knee joint health screening for residents aged 55 and above within its jurisdiction. The goal is to detect potential risks of osteoarthritis, osteoporosis, and other degenerative joint diseases early, so as to promptly recommend high-risk individuals to undergo in-depth examinations (such as MRI) at the hospital, achieving early diagnosis and early intervention.

[0122] Application flow of this system:

[0123] Step 1: Convenient Data Collection 1. The volunteer sits in a chair, and a staff member attaches a lightweight sensing device to both sides of their knee joint. The device integrates a high-sensitivity microphone array (for collecting internal friction and vibration acoustic signals during joint movement) and an inertial measurement unit (IMU, for recording the angle, angular velocity, and acceleration of joint movement).

[0124] 2. Staff members give standardized instructions via an app on a tablet: "Slowly straighten your knees completely, then slowly bend them back to the starting position. Repeat 5 times."

[0125] 3. Volunteers follow instructions, and the system synchronously collects the original acoustic signal sequence throughout the entire movement cycle. and joint angle time series .

[0126] Step 2: Intelligent Data Processing and Feature Fusion

[0127] 1. Feature extraction and quality assessment: Acoustic feature processing: The system processes the raw acoustic signal Adaptive filtering (based on a reference noise sample to remove ambient noise) is performed to obtain the clean acoustic data after noise reduction. Subsequently, MFCC features are extracted using short-time Fourier transform, and their statistical characteristics (mean, standard deviation, skewness, kurtosis) are calculated to form an acoustic feature vector. Simultaneously, acoustic quality scores are calculated based on the signal-to-noise ratio. .

[0128] Motion feature processing: The system uses joint angle data acquired by the IMU. Calculate the maximum active angle Mean angular velocity and acceleration index These indicators, after standardization, form motion feature vectors. Exercise quality score By comparing the current movement characteristics with the standard vectors in the pre-stored "Standard Movement Trajectory Database for Healthy Middle-aged and Elderly People", To calculate.

[0129] 2. Adaptive weighted fusion: The system scores acoustic quality. and exercise quality score Dynamically calculate weights and For acoustic feature vectors and motion feature vectors Perform weighted concatenation to generate a fused feature vector. .

[0130] 3. Closed-loop verification and optimization: Inverse Decoding: Fusing Feature Vectors Input inverse decoding model Reconstructing the acoustic feature vector and motion feature vectors .

[0131] Similarity comparison: The system calculates the normalized cosine similarity between the reconstructed vector and the original vector to obtain the acoustic similarity. Similarity to movement And based on the quality score, the data is weighted and merged into a fusion similarity score. .

[0132] Iterative adjustment: if If the values ​​are below a preset threshold (e.g., 0.8), the system dynamically adjusts the weights using a gradient descent algorithm. and Re-integrate and verify until... The process reaches the required number of iterations or the maximum number of iterations. This process ensures the fidelity of the fused features.

[0133] 4. Anomaly Correction: The system merges the current fused feature vector with the historical feature vectors and divides them into different joint activity feature clusters through cluster analysis (such as K-means). If the distance from the current vector to its cluster center exceeds a threshold calculated based on historical data, the cluster center vector is used instead to filter out abnormal data.

[0134] Step 3: Risk Diagnosis and Report Generation 1. Optimized fused feature vector Input health model Output a quantized joint index (For example, if it is 0.85, the range is 0-1, the higher the value, the healthier).

[0135] 2. The system automatically classifies risk levels based on preset risk thresholds. : Set it to 0.8. Set it to 0.5;

[0136] Low risk ( ); Medium risk ( ); High risk ( ); Joint Index It is 0.85, which is greater than The value is 0.8, therefore the risk level is low;

[0137] 3. The system automatically generates a structured joint disease risk diagnosis report, which includes:

[0138] Diagnostic conclusion: such as "good knee joint function, low risk" or "abnormal knee joint function, high risk, it is recommended to go to the orthopedic clinic as soon as possible".

[0139] Data Summary: Quantitative analysis of acoustic and kinematic characteristics, such as: "Joint movement smoothness (acceleration index) is below healthy levels; no obvious abnormalities were found in joint movement sounds."

[0140] Health advice: Specific recommendations are given for medium- and high-risk groups, such as "reducing stair climbing, avoiding lifting heavy objects, and engaging in low-impact exercises such as swimming and cycling."

[0141] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A portable detection system for early diagnosis of orthopedic joint diseases, characterized in that, Comprise: a data acquisition module for acquiring acoustic feature vectors and motion feature vectors during joint activity; a vector acquisition module for taking the signal-to-noise ratio of the acoustic feature vectors as an acoustic quality score and taking the degree of coincidence between the motion feature vectors and preset standard motion feature vectors as a motion quality score; further configured to normalize the acoustic quality score and the motion quality score to obtain a weight coefficient, based on the weight coefficient, to weight and splice the acoustic feature vectors and the motion feature vectors to obtain a fusion feature vector; a vector reconstruction module for inputting the fusion feature vector into an inverse decoding model to reconstruct the acoustic feature vectors and the motion feature vectors; further configured to compare the reconstructed acoustic feature vectors and motion feature vectors with the acoustic feature vectors and motion feature vectors before weighting and splicing to obtain acoustic similarity and motion similarity; further configured to weight and fuse the acoustic similarity and the motion similarity based on the weight coefficient to obtain a fusion similarity; further configured to adjust the weight coefficient and recalculate the fusion similarity if the fusion similarity is less than a preset similarity threshold, until the fusion similarity is greater than the similarity threshold or the number of re-splicing reaches a preset maximum iteration number, and to take the processed fusion feature vector as a candidate feature vector; an effective judgment module for clustering and dividing the candidate feature vector and historical candidate feature vectors, calculating a clustering distance of the candidate feature vector to a clustering center and a distance threshold obtained from historical candidate feature vectors in the clustering center, and replacing the current candidate feature vector with the historical candidate feature vector of the clustering center as a final feature vector if the clustering distance is greater than the distance threshold; a risk diagnosis module for inputting the final feature vector into a health model to obtain a joint diagnosis report.

2. The portable detection system for early diagnosis of orthopedic joint disease according to claim 1, characterized in that, The process of acquiring the acoustic feature vectors and the motion feature vectors comprises: synchronously acquiring friction acoustic data and motion data generated during the activity of the target joint in a motion cycle, wherein the motion data includes a maximum activity angle, an average angular velocity, and an acceleration index representing motion smoothness; processing the friction acoustic data using an adaptive filtering algorithm to obtain pure acoustic data free of noise interference; performing short-time Fourier transform on the pure acoustic data to calculate its mel-frequency cepstral coefficients, and performing statistical feature processing on the mel-frequency cepstral coefficients to obtain acoustic feature vectors; splicing the motion data after standardization processing to obtain motion feature vectors.

3. The portable detection system for early diagnosis of orthopedic joint disease according to claim 2, characterized in that, The process of acquiring the motion data comprises: acquiring joint angle data collected in time sequence by the target joint in a motion cycle; taking the difference between the maximum value and the minimum value of the joint angle data as the maximum activity angle; performing first-order differentiation on the joint angle data to obtain an angular velocity sequence, calculating the average value of the absolute value of the angular velocity sequence, and taking the average value as the average angular velocity; performing second-order differentiation on the angular velocity sequence to obtain an acceleration sequence, and calculating the root mean square value of the acceleration sequence as the acceleration index.

4. The portable detection system for early diagnosis of orthopedic joint disease according to claim 1, characterized in that, After acquiring the acoustic quality score, further perform validity verification, including: Acquire historical acoustic quality scores corresponding to the historical acoustic feature vectors and historical motion feature vectors same as the current acoustic feature vector and motion feature vector, and the current acoustic quality score; Calculate the score mean of the historical acoustic quality scores, and calculate the first deviation value by subtracting the current acoustic quality score; Calculate the second deviation value of the historical acoustic quality scores and the score mean; Screen out the historical acoustic quality scores with second deviation values less than the first deviation value, and calculate the mean value thereof; Replace the current acoustic quality score with the mean value.

5. The portable detection system for early diagnosis of orthopedic joint disease according to claim 1, characterized in that, The process of acquiring the standard motion feature vector includes: Synchronously acquire joint angle data generated by the target joint in a motion cycle from a healthy population; Process the joint angle data to obtain motion data, statistically analyze and normalize the extracted motion data, and calculate the mean value of the processed motion data as the standard motion feature vector.

6. The portable detection system for early diagnosis of orthopedic joint disease according to claim 5, characterized in that, After acquiring the standard motion feature vector, it is also updated, including: Periodically collect new sample data of the healthy population to acquire an updated standard motion feature vector; Calculate the Mahalanobis distance between the updated motion feature vector and the existing standard motion feature vector; When the Mahalanobis distance exceeds a preset Mahalanobis distance threshold, replace the existing standard motion feature vector with the updated standard motion feature vector.

7. The portable detection system for early diagnosis of orthopedic joint disease according to claim 1, wherein, The process of acquiring the fusion feature vector includes: acoustic feature vector , a motion feature vector , and corresponding acoustic quality scores and motion quality scores ; based on the acoustic quality score and the motion quality score , respectively, a weight of the acoustic feature vector and a weight of the motion feature vector , the weight calculation formula being: ； The calculated weights are used and The acoustic feature vector and the motion feature vector are weighted and spliced to obtain a fusion feature vector The fusion formula is: ； wherein represents a vector concatenation operation.

8. The portable detection system for early diagnosis of orthopedic joint disease according to claim 7, characterized in that, The process of acquiring the acoustic similarity and motion similarity includes: Acoustic feature vectors before weighted stitching are acquired Motion feature vectors , and reconstructed acoustic feature vectors and motion feature vectors ; Computing The cosine similarity between as acoustic similarity The formula is: ； wherein, is the dimension of the acoustic feature vector, denotes the dimension of denotes the dimension of denotes the L2 norm of a vector; Computing the same degree of motion between the same degree of motion between the same degree of motion between ； wherein is the dimension of the motion feature vector, denotes the first dimension of the motion feature vector, denotes the first dimension of the motion feature vector.

9. The portable detection system for early diagnosis of orthopedic joint disease according to claim 1, characterized in that, The process of adjusting the weight coefficient includes: Based on the gradient of the fusion similarity, use the gradient descent algorithm to iteratively update the weight until the fusion similarity is greater than the similarity threshold or the maximum iteration number is reached.

10. The portable detection system for early diagnosis of orthopedic joint disease according to claim 9, characterized in that, When the maximum iteration number is reached and the fusion similarity is still not greater than the similarity threshold, select the weight combination corresponding to the highest fusion similarity value recorded in the iteration process as the weight coefficient in the weighted splicing.

Images