An automatic assessment method for limb dysfunction in spinal cord injury based on clustering

Through automatic clustering-based assessment method, combined with the scale and physical examination data of spinal cord injury patients, semi-supervised K-means clustering model was trained and RefScore indicators were calculated, and the subjective problem of the evaluation of limb dysfunction in spinal cord injury was solved, and the subjective problem of spinal cord injury limb dysfunction was achieved. Accurate and reliable quantitative assessment of limb dysfunction, suitable for rehabilitation treatment and auxiliary equipment adaptation in patients with spinal cord injury.

CN114533041BActive Publication Date: 2024-12-03XI AN JIAOTONG UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210079367.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-01-24
Publication Date
2024-12-03
Estimated Expiration
2042-01-24

AI Technical Summary

Technical Problem

The existing method of limb dysfunction assessment of spinal cord injury is highly subjective, resulting in inaccurate and reliable evaluation. It is especially affected in areas with scarce medical resources and is difficult to conduct quantitative evaluation, which affects the formulation of rehabilitation treatment plans and the appropriate auxiliary equipment. match.

Method used

Using automatic clustering-based evaluation method, machine learning technology combined with limb dysfunction evaluation characteristics, digitized processing is performed using SCIM scale, Berg scale, AISA lesion index and other data, semi-supervised K-means clustering model, and RefScore is calculated Indicators were used to quantitatively evaluate the degree of limb dysfunction in patients with spinal cord injury.

Benefits of technology

It realizes an accurate and reliable quantitative assessment of the degree of limb dysfunction in patients with spinal cord injury, improves the accuracy and reliability of the assessment, is simple to operate and has anti-interference ability, and is suitable for rehabilitation treatment and auxiliary equipment adaptation for patients with spinal cord injury.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114533041B_ABST
    Figure CN114533041B_ABST
Patent Text Reader

Abstract

The present invention discloses a clustering-based automatic assessment method for limb dysfunction in spinal cord injury. First, the commonly used spinal cord injury limb function assessment scale and physical examination information are digitized; secondly, an improved semi-supervised K-means clustering method is proposed, and an assessment model is constructed based on it to achieve automatic differentiation of patients with different degrees of limb dysfunction; finally, a reference score index RefScore is proposed to determine the limb dysfunction level of each type of patient after clustering. This method proposes for the first time to use a clustering model to integrate the features of limb dysfunction assessment, and can make a quantitative assessment of the degree of limb dysfunction in patients with spinal cord injury. It has the advantages of high accuracy, strong reliability, simple operation, strong robustness, and strong anti-interference ability and easy promotion. At the same time, it provides an effective solution and approach for the dysfunction assessment of other types of diseases.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of intelligent assessment of limb dysfunction, and in particular to a clustering-based limb dysfunction assessment method, which is particularly suitable for limb dysfunction assessment of patients with spinal cord injury. Background Art

[0002] Spinal cord injury (SCI) is the most serious complication of spinal fracture. Due to the displacement of the vertebral body or the protrusion of bone fragments into the spinal canal, the spinal cord or cauda equina nerve will be damaged to varying degrees. The assessment of limb dysfunction in patients with spinal cord injury is of great significance in both the early treatment and the prognosis and rehabilitation stages. In the early treatment stage, an accurate assessment of the patient's limb dysfunction level can reflect the severity of the patient's spinal cord injury, thereby helping the clinician to formulate a correct treatment plan. In the prognosis stage, the patient's limb dysfunction assessment results can be used as a monitoring and evaluation indicator of the patient's rehabilitation process. Doctors can adjust the patient's rehabilitation treatment plan in real time based on this indicator to achieve the best rehabilitation effect.

[0003] At present, the limb dysfunction of patients with spinal cord injury is mainly evaluated clinically based on scales, physical examinations, and interviews. Commonly used scales include the American Spinal Injury Association (ASIA) Injury Scale, the Spinal Cord Injury Independence Measure (SCIM), and the Berg scale, which can evaluate the motor function, degree of injury, and ability to perform daily activities in patients with spinal cord injury. Common clinical physical examination data include muscle strength, muscle tension, and joint angles, which reflect the degree of muscle activation and muscle spasm of the patient, and are of great significance for judging the degree of functional impairment of the patient. Commonly used image information includes MRI and CT images, which can intuitively reflect the location and severity of spinal cord injury. The above examination information provides a basis for the evaluation of limb dysfunction in patients with spinal cord injury from different aspects, but the final evaluation result requires doctors to rely on their own experience to conduct a comprehensive analysis of the above multi-source information. Different doctors may draw different conclusions based on the same effective information, which is highly subjective. This is bound to seriously affect the accuracy and reliability of the evaluation of the degree of limb dysfunction in patients with spinal cord injury, especially in remote areas and rural areas with scarce medical resources, where the impact is more worthy of attention.

[0004] In addition, when doctors make a diagnosis, it is difficult to make a quantitative assessment of the degree of limb dysfunction of patients due to the lack of a unified assessment standard, which is not conducive to the formulation of rehabilitation treatment plans and the adaptation of assistive devices. Therefore, studying new limb dysfunction assessment methods is one of the research topics of the applicant. Summary of the invention

[0005] In view of the defects of the existing spinal cord injury limb dysfunction assessment methods such as strong subjectivity, the purpose of the present invention is to provide a clustering-based automatic limb dysfunction assessment method, which uses machine learning technology combined with limb dysfunction assessment characteristics to complete a quantitative automatic assessment of the degree of dysfunction of the upper and lower limbs of patients with spinal cord injury.

[0006] In order to achieve the above tasks, the present invention is implemented through the following technical solutions:

[0007] A clustering-based automatic assessment method for spinal cord injury limb dysfunction, characterized in that the specific implementation steps are as follows:

[0008] Step 1: Collect patient scale and physical examination data, including: SCIM scale, Berg scale, AISA impairment index, spinal cord injury cross section, and muscle strength and tension of eight joints including shoulder, elbow, wrist, hand, hip, knee, ankle and foot;

[0009] Step 2: Digitize the non-digital language descriptive information in the data to form a feature set;

[0010] Step 3: Normalize all features;

[0011] Step 4: According to the correlation between the features and the upper and lower limbs, the feature set is divided into upper limb related parts and lower limb related parts, which are used to evaluate the upper limbs and lower limbs respectively;

[0012] Step 5: Set the number of levels to be divided and calculate the similarity between two patients. Train the semi-supervised K-means clustering model based on the above feature set and patient similarity relationship.

[0013] Step 6: Calculate the reference score index RefScore of each category of limb dysfunction after clustering based on the clustering results of the feature set, which is used to judge the severity of limb dysfunction of each category of patients.

[0014] In the above step 2, the method for digitizing the non-digital information in the data is as follows:

[0015] ① The AISA impairment index is divided into five levels: A, B, C, D, and E, indicating that the severity of spinal cord injury decreases in sequence. In digital processing, the AISA impairment index A, B, C, D, and E are represented by 1, 2, 3, 4, and 5 respectively.

[0016] ② The cross-sectional position of the spinal cord injury was digitized according to the spinal segmentation rules. From top to bottom, the 7 cervical vertebrae were represented as 1-7, the 12 thoracic vertebrae were represented as 8-19, the 5 lumbar vertebrae were represented as 20-24, and the 5 sacral vertebrae were represented as 25-29.

[0017] In the above step 3, the normalization calculation formula of the feature is:

[0018]

[0019] Where: F_Score ij is the score of the i-th reference feature of the j-th patient, F_Score_Standard i is the highest score that the i-th feature can achieve.

[0020] The specific method of step 4 above is: check the scale and physical examination data one by one, include the parts related to the upper limbs into the upper limb feature set, and the parts related to the lower limbs into the lower limb feature set; the division results are:

[0021] The upper limb feature set included: upper limb strength (shoulder, elbow, wrist, and hand), upper limb tone (shoulder, elbow, wrist, and hand), SCIM scale upper limb part, spinal cord injury level, and AISA impairment index;

[0022] The lower extremity feature set includes: lower extremity strength (hip, knee, ankle, and foot), lower extremity tone (hip, knee, ankle, and foot), SCIM lower extremity component, Berg scale, spinal cord injury cross-section, and AISA impairment index.

[0023] In step 5 above, the number of levels to be divided can be set according to the needs. In the introduction of this method, the three levels are taken as an example, namely mild, moderate and severe. The clustering method used in the training model is the improved semi-supervised K-means clustering. The specific calculation method is as follows:

[0024] ① Select some representative features from the feature set to form a reference feature vector, and calculate the absolute value distance between the reference feature vectors of each patient. The calculation formula is:

[0025]

[0026] In the formula, dist ij Represents the absolute value distance between the reference feature vector of the i-th patient and the j-th patient, feature pi is the value of the pth reference feature of the ith patient, feature pj is the value of the pth reference feature of the jth patient, and m means that m reference features are selected;

[0027] The reference characteristics selected for the upper limb assessment model include: AISA impairment index, spinal cord injury cross section, total score of the upper limb part of the SCIM scale, total upper limb muscle strength, and upper limb muscle tension;

[0028] The reference features selected for the lower limb assessment model include: AISA impairment index, spinal cord injury section location, Berg scale total score, SCIM scale lower limb part total score, lower limb muscle strength sum and lower limb muscle tension.

[0029] ②Set the distance threshold dist th , patient sample pairs with a distance less than the threshold are classified into the must_link class, and patient sample pairs with a distance greater than the threshold are classified into the cannot_link class; the distth calculation formula is:

[0030] dist th =m*percent

[0031] In the formula, m means that m reference features are selected, and percent can be optimally set between 0 and 1 according to the data situation. In this method, percent is set to 0.1.

[0032] Patient sample pairs with a distance less than the threshold will be classified into the must_link class, indicating that the similarity between the patients in this sample pair is large, and the two samples in this sample pair should be classified into the same class during clustering.

[0033] ③ Using the data of must_link class, the improved semi-supervised K-means algorithm is used to train the semi-supervised K-means clustering model.

[0034] The calculation method of the limb dysfunction level reference index RefScore described in the above step 6 is as follows:

[0035] Some representative features are selected as reference features, and the limb dysfunction level reference index RefScore is calculated through linear weighted fusion of the reference features. The limb dysfunction level of patients in each category can be determined based on this index. The larger the RefScore, the lower the degree of disability. The calculation method is as follows:

[0036]

[0037]

[0038] Among them: feature ij is the normalized value of the i-th reference feature of the j-th patient, k represents the category label after clustering, and N k is the total number of patients included in the kth category. flag i Indicates the type of feature, flag i =1, the higher the feature score, the lower the degree of obstacle. i = -1, higher feature scores indicate higher levels of impairment; ik is the normalized result of the mean score of the i-th feature of the k-th patient; m represents the m selected reference features;

[0039] The selected reference characteristics included: AISA impairment index, spinal cord injury cross section, Berg scale total score, SCIM scale total score, total muscle strength, and muscle tension;

[0040] RefScore k represents the reference score of limb dysfunction level for patients in category k;

[0041] After calculating the RefScore of each category after clustering, the limb dysfunction level of the category of patients with the largest RefScore value is mild, and the limb dysfunction level of the category of patients with the smallest RefScore value is severe.

[0042] The clustering-based limb dysfunction assessment method of the present invention proposes for the first time to use the clustering model to integrate the characteristics of limb dysfunction assessment, and can make a quantitative assessment of the degree of limb dysfunction in patients with spinal cord injury. It has the advantages of high accuracy, strong reliability, simple operation, strong robustness, and strong anti-interference ability and easy promotion. At the same time, it provides an effective solution and approach for the dysfunction assessment of other types of diseases. When using the improved semi-supervised K-means algorithm to determine the class to which a new sample belongs, not only the distance between the new sample and each cluster center is considered, but also the degree of similarity between the new sample and the sample pairs composed of each type of sample is considered. If the new sample and the sample pairs composed of a certain type of sample have only a small proportion of high similarity, the new sample will not be classified into this class. The proposed reference score index RefScore is used to determine the degree of limb dysfunction of each type of patient after clustering. The RefScore index plays an important guiding role in exploring the physiological significance of the training results based on unsupervised and semi-supervised machine learning models. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is the evaluation model training flowchart.

[0044] Figure 2 This is a comparison chart of the clustering effects of the improved semi-supervised K-means algorithm and the commonly used clustering algorithm applied to the assessment of upper limb dysfunction.

[0045] Figure 3 This is a comparison chart of the clustering effects of the improved semi-supervised K-means algorithm and the commonly used clustering algorithm applied to the assessment of lower limb dysfunction.

[0046] The present invention is further described in detail below with reference to the accompanying drawings and embodiments. DETAILED DESCRIPTION

[0047] See also Figure 1 This embodiment provides a clustering-based automatic assessment method for spinal cord injury limb dysfunction, and the specific implementation steps are as follows:

[0048] Step 1: Collect patient scale and physical examination data, including: SCIM scale, Berg scale, AISA impairment index, spinal cord injury cross section, shoulder, elbow, wrist, hand, hip, knee, ankle and foot joint strength and tension;

[0049] Step 2: Digitize the non-digital information in the data to form a feature set. The digital processing method is as follows:

[0050] ① The AISA impairment index is divided into five levels: A, B, C, D, and E, indicating that the severity of spinal cord injury decreases in sequence. In digital processing, the AISA impairment index A, B, C, D, and E are represented by 1, 2, 3, 4, and 5 respectively.

[0051] ② The cross-sectional position of the spinal cord injury was digitized according to the spinal segmentation rules. From top to bottom, the 7 cervical vertebrae were represented as 1-7, the 12 thoracic vertebrae were represented as 8-19, the 5 lumbar vertebrae were represented as 20-24, and the 5 sacral vertebrae were represented as 25-29.

[0052] Step 3: Normalize the features. The normalization calculation formula is as follows:

[0053]

[0054] Where: F_Score ij is the score of the i-th reference feature of the j-th patient, F_Score_Standard i is the highest score that the i-th feature can achieve.

[0055] Step 4: Check the scale and physical examination data one by one, and include the parts related to the upper limbs into the upper limb feature set, and the parts related to the lower limbs into the lower limb feature set.

[0056] The upper limb feature set includes: upper limb strength (shoulder, elbow, wrist, hand), upper limb muscle tone (shoulder, elbow, wrist, hand), upper limb part of the SCIM scale, spinal cord injury level, and AISA impairment index;

[0057] The lower limb feature set includes: lower limb strength (hip, knee, ankle, foot), lower limb tone (hip, knee, ankle, foot), SCIM lower limb part, Berg scale, spinal cord injury cross-section, and AISA impairment index.

[0058] Step 5: Set the number of levels to be divided, and calculate the similarity between two patients based on the distance of the feature vector; use the clustering idea in machine learning to build an automatic evaluation model. The model uses an improved semi-supervised K-means clustering algorithm to train the clustering model based on the above feature set and patient similarity relationship. The number of levels to be divided can be set according to needs.

[0059] In this embodiment, the classification into three levels is taken as an example, and the three levels are mild, moderate and severe. The specific calculation method of the model is as follows:

[0060] ① Select some representative features to form reference feature vectors, and calculate the absolute value distance between the reference feature vectors of each patient. The calculation formula is:

[0061]

[0062] Among them, dist ij Represents the absolute value distance between the reference feature vector of the i-th patient and the j-th patient, feature pi is the value of the pth reference feature of the ith patient, feature pj is the value of the pth reference feature of the jth patient, and m means that m reference features are selected.

[0063] In this embodiment, the reference features selected by the upper limb assessment model include: AISA impairment index, spinal cord injury cross section, total score of the upper limb part of the SCIM scale, total upper limb muscle strength and upper limb muscle tension;

[0064] The reference characteristics selected for the lower limb assessment model include: AISA impairment index, spinal cord injury cross section, Berg scale total score, SCIM scale lower limb part total score, lower limb muscle strength sum and lower limb muscle tension.

[0065] ②Set the distance threshold dist th , patient sample pairs with a distance less than the threshold are classified into the must_link class, and patient sample pairs with a distance greater than the threshold are classified into the cannot_link class. dist th The calculation formula is:

[0066] dist th =m*percent

[0067] Here, m means that m reference features are selected, and percent can be optimally set between 0 and 1 according to the data situation. In this embodiment, percent is set to 0.1.

[0068] Patient sample pairs with a distance less than the threshold will be classified into the must_link class, indicating that the similarity between the patients in this sample pair is large, and the two samples in this sample pair should be classified into the same class during clustering.

[0069] ③ Using the data of the must_link class, the improved semi-supervised K-means clustering algorithm is used to train the semi-supervised K-means clustering model. The process of the improved semi-supervised K-means clustering algorithm is shown in the following table:

[0070] Table 1: Improved semi-supervised K-means clustering algorithm process

[0071]

[0072] The improved semi-supervised K-means clustering algorithm provided in this embodiment not only considers the distance between the new sample and each cluster center when determining the class to which the new sample belongs, but also considers the degree to which the sample pairs consisting of the new sample and samples of each class satisfy the must_link class. If only a small proportion of the sample pairs consisting of the new sample and samples of a certain class belong to the must_link class, the new sample will not be classified into that class.

[0073] Compared with traditional clustering methods such as K-means, spectral clustering, and density clustering, the improved semi-supervised K-means clustering algorithm has a significantly improved clustering effect. The silhouette coefficient is used as the evaluation index of the clustering result, and its calculation formula is as follows:

[0074]

[0075] Where: a(i) represents the dissimilarity within the cluster, which is the distance from sample i to the same cluster C. i The average distance of other samples, the smaller a(i), the more likely the sample should be clustered into this cluster. The mean of a(i) of all samples in cluster C is called the cluster dissimilarity of cluster C. b(i) represents the inter-cluster dissimilarity, which is the distance from sample i to other clusters C. j The average distance of all samples. The larger the b(i), the less sample i belongs to other clusters. The s(i) value is between [-1, 1]. The closer it is to 1, the more reasonable the clustering of sample i is; the closer it is to -1, the more sample i should be classified into other clusters; the closer it is to 0, the sample should be on the boundary. The mean s(i) of all samples is called the silhouette coefficient of the clustering result.

[0076] See also Figure 2 and Figure 3In order to verify the accuracy of the clustering-based automatic assessment method for limb dysfunction in spinal cord injury given in this embodiment, the inventor used the clustering model trained by the scale and physical examination data of 38 patients for verification. The results showed that compared with traditional clustering algorithms such as K-means and Gaussian mixture model, the clustering performance of the improved semi-supervised K-means clustering algorithm proposed in this embodiment is greatly improved. It has great application potential in clustering applications.

[0077] Step 6: Select some representative features as reference features, and calculate a limb dysfunction level reference value RefScore by integrating the reference features of samples in each cluster. It is used to determine the limb dysfunction level of samples in each category after clustering. The larger the RefScore, the lower the degree of disability. The calculation method is as follows:

[0078]

[0079]

[0080] Among them: feature ij is the normalized value of the i-th reference feature of the j-th patient, N k is the total number of patients included in the kth category. flag i Indicates the type of feature, flag i =1, the higher the feature score, the lower the degree of obstacle. i = -1, higher feature scores indicate higher levels of impairment. i k is the normalized result of the mean score of the i-th feature of the k-th patient. m means that m reference features are selected.

[0081] The reference features selected in this embodiment include: AISA impairment index, spinal cord injury cross section, Berg scale total score, SCIM scale total score, muscle strength sum and muscle tension. k Represents the reference score of limb dysfunction level for patients in category k.

[0082] After clustering, the RefScore of each category was calculated. The limb dysfunction level of the category with the largest RefScore value was mild, and the limb dysfunction level of the category with the smallest RefScore value was severe.

Claims

1. A clustering-based automatic assessment method for limb dysfunction in spinal cord injury, characterized in that: The specific implementation steps are as follows: Step 1: Collect patient scale and physical examination data, including SCIM scale, Berg scale, AISA impairment index, spinal cord injury cross section, shoulder, elbow, wrist, hand, hip, knee, ankle and foot joint strength and tension; Step 2: Digitize the non-digital information in the data to form a feature set; Step 3: Normalize the features; Step 4: Classify the features according to their relevance to the upper and lower limbs, and divide the feature set into upper limb-related parts and lower limb-related parts, which are used to evaluate the upper limbs and lower limbs respectively; Step 5: Set the number of levels to be divided and calculate the similarity between patients. Train the semi-supervised K-means clustering model based on the above feature set and patient similarity relationship. Step 6: Calculate the reference score of limb dysfunction level of each type of patient after clustering based on the clustering results of the feature set RefScore , used to judge the severity of limb dysfunction in various types of patients; The method for digitizing the non-digital information in the data in step 2 is as follows: ① The AISA impairment index is divided into five levels: A, B, C, D, and E, which represent the decreasing severity of spinal cord injury. In the digital processing, the AISA impairment index A, B, C, D, and E are represented by 1, 2, 3, 4, and 5 respectively; ② The spinal cord injury cross-section was digitized according to the spinal segmentation rules. From top to bottom, the 7 cervical vertebrae were represented as 1-7, the 12 thoracic vertebrae were represented as 8-19, the 5 lumbar vertebrae were represented as 20-24, and the 5 sacral vertebrae were represented as 25-29; The calculation formula for feature normalization in step 3 is: In the formula, For the jth patient i The score of the reference features, For the i The highest score a characteristic can achieve; The specific implementation method of step 4 is: Check the scale and physical examination data one by one, and include the parts related to the upper limbs into the upper limb feature set, and the parts related to the lower limbs into the lower limb feature set; the division results are: The upper limb feature set includes upper limb strength, upper limb tension, SCIM upper limb part, spinal cord injury cross section, and AISA impairment index; the upper limb refers to the shoulder, elbow, wrist, and hand; The lower limb feature set includes lower limb strength, lower limb tension, SCIM lower limb part, Berg scale, spinal cord injury cross section and AISA impairment index; the lower limb refers to the hip, knee, ankle and foot; Reference score index for limb dysfunction level described in step 6 RefScore The calculation method is as follows: Select some representative features as reference features, and calculate the reference index of limb dysfunction level through linear weighted fusion of reference features RefScore According to this index, the degree of limb dysfunction of patients in each category can be determined. RefScore The larger the value, the lower the obstacle level. The calculation method is as follows: in: For the jth patient i The normalized value of the reference feature, For the k The total number of patients included in the class; Indicates the type of feature, Higher scores on the characteristics indicate lower levels of impairment. Higher scores on the characteristics indicate higher levels of impairment; For the k Type 1 patients i Normalized result of the mean of class feature scores; m Indicates that the selected m Reference features; The selected reference characteristics included: AISA impairment index, spinal cord injury cross section, Berg scale total score, SCIM scale total score, total muscle strength, and muscle tension; represents the reference score of limb dysfunction level for patients in category k; Calculate the clustering of each category RefScore back, RefScore The patients with the largest value had mild limb dysfunction. RefScore The level of limb dysfunction in the group with the smallest value is severe.

2. The method according to claim 1, characterized in that The number of levels required for the setting described in step 5 can be set according to needs, namely mild, moderate and severe.

3. The method according to claim 1, characterized in that The specific calculation method of the improved semi-supervised K-means clustering is as follows: ① Select some representative features to form reference feature vectors, and calculate the absolute value distance between the reference feature vectors of each patient. The calculation formula is: In the formula, Indicates i Patients and j The absolute value distance of the reference feature vector of the patients, For the i Patient No. p The value of the reference feature, For the j Patient No. p The value of the reference feature, m Indicates that the selected m Reference features; The reference characteristics selected for the upper limb assessment model include: AISA impairment index, spinal cord injury cross section, total score of the upper limb part of the SCIM scale, total upper limb muscle strength, and upper limb muscle tension; The reference characteristics selected for the lower limb assessment model include: AISA impairment index, spinal cord injury section location, Berg scale total score, SCIM scale lower limb part total score, lower limb muscle strength and lower limb muscle tension; ②Set distance threshold , patient sample pairs with distances less than the threshold are classified must_link Class, patient sample pairs with distance greater than the threshold are classified cannot_link kind; The calculation formula is: in, m Indicates that the selected m Reference features, percent The optimal setting can be made between 0 and 1 according to the data situation; Patient sample pairs with distances less than the threshold will be classified must_link class, indicating that the similarity between the patients in this sample pair is large, and the two samples in this sample pair should be classified into the same class during clustering; ③Use must_link For the data of the class, the improved semi-supervised K-means algorithm is used to train the semi-supervised K-means clustering model.