A plantar pressure full cycle evaluation and orthotic recommendation generation system fusing static and dynamic data

Abstract

This invention discloses a system for full-cycle assessment and orthotic suggestion generation of plantar pressure, integrating static and dynamic data, relating to the field of medical devices. It includes: a data acquisition module that collects basic data, static plantar pressure data, and dynamic data, and preprocesses this data to form a dataset through a database module; an assessment module that trains an assessment model using the dataset and outputs assessment results for the individuals being assessed; and an orthotic suggestion module that trains a suggestion generation model using intermediate features from the dataset and the assessment model and outputs orthotic suggestions for the individuals being assessed. This invention achieves efficient integration of the entire process from data acquisition to suggestion output, adapts to the actual needs of clinical rehabilitation guidance, and possesses strong practicality and promotional value.

Description

Technical Field

[0001] This invention relates to the field of medical device technology, and more specifically to a system for full-cycle assessment of plantar pressure and generation of orthopedic recommendations that integrates static and dynamic data. Background Technology

[0002] Flat feet, high arches, inversion / eversion of the foot, as well as metatarsalgia, heel pain, plantar fasciitis, and other common foot and ankle disorders not only cause foot pain and limited walking function, but can also trigger a chain of musculoskeletal disorders such as knee varus and compensatory scoliosis, seriously affecting patients' quality of life and athletic ability. As the core support structure for weight-bearing and movement, the pressure distribution characteristics of the sole of the foot are closely related to foot and ankle diseases; therefore, plantar pressure assessment is a key basis for the diagnosis of foot and ankle diseases and the development of orthopedic treatment plans.

[0003] However, despite some progress in plantar pressure assessment, several limitations remain. Specifically, existing systems often collect static or dynamic plantar pressure data separately, neglecting the correlation between the two. Data collection is limited in scope, resulting in incomplete assessments. Furthermore, current plantar pressure assessments rely solely on physician experience, leading to insufficient accuracy. Additionally, orthotic recommendations are also based on physician assessments, depending on standardized orthotic protocols and clinical experience. This results in orthotic insole parameters being significantly influenced by subjectivity, insufficiently targeted rehabilitation training recommendations, and unsatisfactory orthotic outcomes.

[0004] Therefore, how to conduct plantar pressure assessment based on multidimensional data to improve the accuracy of the assessment and generate targeted orthopedic recommendations is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, the present invention provides a system for full-cycle assessment of plantar pressure and generation of orthopedic recommendations that integrates static and dynamic data. By collecting and integrating multi-dimensional data, personalized recommendations are generated, thereby solving the problems of incomplete assessment and inaccurate recommendations in existing systems.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: This invention discloses a system for full-cycle assessment of plantar pressure and generation of orthopedic recommendations that integrates static and dynamic data, comprising: a data acquisition module, a database module, an assessment module, and an orthopedic recommendation module; The data acquisition module collects basic data, static plantar pressure data, and dynamic data from several volunteers, as well as foot and ankle disease types and historical orthopedic recommendations for volunteers with abnormal plantar pressure; it also collects basic data, static plantar pressure data, and dynamic data from individuals to be evaluated. The database module preprocesses the volunteers' basic data, static plantar pressure data, and dynamic data as sample features, and uses the corresponding foot and ankle disease type and degree as sample labels to obtain the first type of sample data and construct the first dataset; it also preprocesses the basic data, static plantar pressure data, and dynamic data of the personnel to be evaluated to obtain the predicted input data of the personnel to be evaluated. The evaluation module trains an evaluation model using the first dataset; it then inputs the predicted input data into the trained evaluation model to obtain the evaluation results for the person to be evaluated. The database module further preprocesses the historical orthopedic suggestion data as a sample output and uses the intermediate features of the evaluation model as a sample input to obtain a second type of sample data and construct a second dataset. The orthopedic suggestion module trains a suggestion generation model using the second dataset, and inputs the intermediate features extracted by the evaluation model from the predicted input data into the trained suggestion generation model to obtain orthopedic suggestions for the person to be evaluated.

[0007] Furthermore, the data acquisition module includes a static data acquisition unit, a dynamic data acquisition unit, and an information acquisition unit; The static data acquisition unit collects plantar pressure data of volunteers and individuals under evaluation in a static posture with both feet evenly bearing weight and eyes looking straight ahead using a plantar pressure plate measuring instrument. The unit also collects the height of volunteers and individuals under evaluation using a medical height ruler and their weight using a medical electronic scale. The dynamic data acquisition unit collects plantar pressure data of volunteers and those to be evaluated during natural walking and running through a gait track laid with several plantar pressure plate measuring instruments, and uses this data as the dynamic data. The information collection unit collects information on the age, gender, foot length, foot width, arch height, and leg length of volunteers and those to be evaluated through electronic medical records, doctor's consultation records, or questionnaires; it also collects the type and severity of the volunteers' foot and ankle diseases, as well as corresponding historical orthopedic recommendations. The foot and ankle disorders include one or more of the following: flat feet, high arches, inversion of the foot, eversion of the foot, metatarsalgia, heel pain, and plantar fasciitis; The historical orthotic recommendation data includes: the three-dimensional height of the orthotic insole, as well as the type, frequency, and duration of rehabilitation training; The basic data includes: height, weight, age, gender, foot length, foot width, arch height, and leg length.

[0008] Furthermore, the preprocessing of the volunteers' basic data, static plantar pressure data, and dynamic data, as well as the preprocessing of the basic data, static plantar pressure data, and dynamic data of the individuals to be evaluated, specifically includes: Basic data preprocessing: Gender is coded; height, weight, age, foot length, foot width, arch height and leg length are cleaned, outliers are removed and filled, and then the data is standardized and sorted in order to obtain the basic data vector; Static data preprocessing: Based on the pressure values ​​of each sensor in the static plantar pressure data, the corresponding grayscale images are converted and noise filtering and smoothing are performed to obtain static feature maps; Dynamic data preprocessing: Taking the first heel strike in the dynamic data of plantar pressure as the starting point of a cycle and the second heel strike on the same side as the ending point of a cycle, dynamic data of 5 consecutive complete gait cycles of natural walking and 5 consecutive complete gait cycles of running are obtained. Based on the pressure values ​​of each sensor in each frame of the dynamic data, grayscale images of each frame are obtained, and noise filtering and smoothing are performed to finally obtain the dynamic feature map sequence. The preprocessing of the historical orthodontic recommendation data specifically includes: Based on the three-dimensional height of the orthotic insole, a coordinate system is established with the foot length direction as the X-axis, the foot width direction as the Y-axis, and the foot height direction as the Z-axis to obtain the three-dimensional height matrix of the orthotic insole; the type of rehabilitation training is one-hot encoded, and the frequency and duration are standardized to obtain the rehabilitation training vector.

[0009] Furthermore, the evaluation model includes a CNN network component, a CNN-LSTM network component, and a multilayer perceptron component; The CNN network takes the static feature map as input and processes it through multiple layers of convolution, pooling, and activation functions to obtain the first feature matrix; The CNN-LSTM network takes the dynamic feature map sequence as input and processes it through convolution, gating mechanism and attention mechanism to obtain the second feature matrix; The multilayer perceptron takes the first feature matrix, the second feature matrix, and the basic data vector as input, and obtains the predicted probability of foot and ankle disease type and severity through feature fusion and mapping.

[0010] Furthermore, the CNN network portion sequentially includes: a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a third pooling layer, a fourth convolutional layer, and a fourth pooling layer; the number of convolutional kernels in the first to fourth convolutional layers are 16, 32, 64, and 128, respectively, the kernel size of the first to fourth convolutional layers is 3, and the activation function of the convolutional kernels is the ReLU activation function; the first to fourth pooling layers all use max pooling.

[0011] Furthermore, the structure of the CNN-LSTM network is as follows: CNN layer, LSTM layer, and attention layer; The CNN layers sequentially include: a fifth convolutional layer, a fifth pooling layer, a sixth convolutional layer, and a sixth pooling layer; the number of convolutional kernels in the fifth and sixth convolutional layers are 32 and 64 respectively, the kernel size of the fifth and sixth convolutional layers is 3×3, and the activation function is ReLU activation function; the fifth and sixth pooling layers all use max pooling; the CNN layers extract features from the feature maps of each frame in the dynamic feature map sequence, and output a feature map sequence with uniform dimensions; The LSTM layer consists of two stacked LSTM units, with each LSTM unit having 128 hidden units. The LSTM layer takes the feature map sequence output by the CNN layer as input and extracts the temporal feature tensor from the feature map sequence through the gating mechanism of the LSTM unit. The attention layer sequentially comprises a tensor multiplication layer, a fully connected layer, a flattening layer, a softmax layer, a reshape layer, and a lambda layer. The tensor multiplication layer receives the temporal feature tensor and performs self-attention interaction computation. The fully connected layer performs a fully connected mapping on the output of the tensor multiplication layer to generate attention weight associated features. The flattening layer compresses the attention weight associated features from multidimensional features into a one-dimensional vector to obtain an initial attention weight vector. The softmax layer normalizes the initial attention weight vector so that the sum of the weights is 1. The reshape layer reshapes the normalized weight vector into a result tensor that matches the dimension of the temporal feature tensor output by the LSTM layer. The lambda layer performs element-wise multiplication between the result tensor and the temporal feature tensor, and sums and aggregates all weighted feature vectors along the temporal length direction to finally output the second feature matrix.

[0012] Furthermore, the multilayer perceptron section sequentially includes: an input layer, a first fully connected layer, a second fully connected layer, and a multi-task output head; The input layer flattens the first feature matrix and the second feature matrix into a one-dimensional vector, and concatenates it with the basic data vector to obtain a fused feature vector. The first fully connected layer contains 520 neurons, and the second fully connected layer contains 256 neurons. Each neuron in the fully connected layer uses the ReLU activation function, and feature mapping and compression are performed through the two fully connected layers. The task output head includes a disease type prediction branch and a disease severity prediction branch, with the output of the second fully connected layer as input to both branches. The disease type prediction branch contains 7 output neurons, which use the Sigmoid activation function to output the probability of the existence of each disease. The disease severity prediction branch contains 21 output neurons, which use the Softmax activation function to output the probability distribution of each severity level corresponding to each disease.

[0013] Furthermore, the proposed generation model includes a first generative adversarial network and a second generative adversarial network; The first generative adversarial network takes the intermediate features as input and generates a three-dimensional height matrix of orthopedic insoles through a trained generator; the second generative adversarial network takes the intermediate features as input and generates rehabilitation training vectors through a trained generator. The intermediate features include the first feature matrix, the second feature matrix, and the basic data vector.

[0014] Furthermore, the first generative adversarial network includes a generator G1 and a discriminator D1; the generator G1 includes a static branch, a dynamic branch, a weighted fusion layer, and an output layer; The static branch includes a static intermediate feature map sub-branch and a static matrix sub-branch. The static intermediate feature map sub-branch sequentially includes two deconvolution layers and a batch normalization layer. The first feature matrix is ​​deconvolved by the deconvolution layers, and nonlinearity is introduced by the sigmoid activation function. The size is then adjusted by the batch normalization layer to obtain the static intermediate feature map. The static matrix sub-branch sequentially includes three deconvolution layers and a batch normalization layer. The basic data vector is deconvolved by the deconvolution layers, and the size is then adjusted by the batch normalization layer to obtain the static weight matrix. The static intermediate feature map and the static weight matrix are multiplied element-wise to obtain the static height matrix. The dynamic branch includes a dynamic intermediate feature map sub-branch and a dynamic matrix sub-branch. The dynamic intermediate feature map sub-branch sequentially includes two deconvolution layers and a batch normalization layer. The second feature matrix is ​​deconvolved by the deconvolution layers, and nonlinearity is introduced by the sigmoid activation function. The size is then adjusted by the batch normalization layer to obtain the dynamic intermediate feature map. The dynamic matrix sub-branch sequentially includes three deconvolution layers and a batch normalization layer. The base data vector is deconvolved by the deconvolution layers, and the size is then adjusted by the batch normalization layer to obtain the dynamic weight matrix. The dynamic intermediate feature map and the dynamic weight matrix are multiplied element-wise to obtain the dynamic height matrix. The weighted fusion layer performs a weighted fusion of the static height matrix and the dynamic height matrix, and outputs the result through the output layer. The discriminator D1 includes three convolutional layers and two fully connected layers. It takes the actual orthopedic insole 3D height matrix and the predicted 3D height matrix output by the generator G1 as inputs. It extracts features through the three convolutional layers and outputs the discrimination result through the two fully connected layers to determine the authenticity of the predicted 3D height matrix.

[0015] Furthermore, the second generative adversarial network includes a generator G2 and a discriminator D2; The generator G2 includes an input layer, three fully connected layers, and an output layer. The input layer flattens the first feature matrix and the second feature matrix into a one-dimensional vector and concatenates it with the basic data vector to obtain a fused feature vector. Then, it passes through three fully connected layers for nonlinear mapping, with each neuron in the fully connected layer using the ReLU activation function. Finally, the output layer outputs the training type probability distribution, frequency, and duration values ​​of the rehabilitation training vector. The neurons in the output layer corresponding to the rehabilitation training type use the Sigmoid activation function, while the neurons corresponding to the frequency and duration use the Linear activation function. The discriminator D2 takes the real rehabilitation training vector and the predicted rehabilitation training vector output by the generator G2 as input, and processes them through two fully connected layers in sequence to output the discrimination result and determine the authenticity of the rehabilitation training vector.

[0016] As can be seen from the above technical solutions, compared with the prior art, the present invention provides a system for full-cycle assessment and orthopedic suggestion generation of plantar pressure that integrates static and dynamic data. Through multi-dimensional data fusion and intelligent model design, it achieves a comprehensive upgrade of plantar pressure assessment and orthopedic suggestion, with significant technical effects. Specifically, (1) the present invention solves the defect of single data acquisition dimension in the prior art. The system simultaneously collects plantar pressure data in static standing, dynamic walking and running states. Combined with basic information such as height and arch height, it constructs a three-dimensional acquisition system of static, dynamic and basic data, which fully captures the distribution characteristics and correlation rules of plantar pressure in different exercise states, avoiding the one-sidedness of single data assessment. (2) the present invention greatly improves the assessment accuracy. The assessment model integrates the spatial feature extraction capability of CNN, the temporal feature capture advantage of CNN-LSTM and the key information focusing function of attention mechanism. Through multi-task output head, it simultaneously predicts the probability and degree of foot and ankle diseases, which has high accuracy. (3) Based on the intermediate features of the evaluation model, this invention generates a three-dimensional height matrix of orthotic insoles and a rehabilitation training vector through a dual generative adversarial network. The parameters of the orthotic insoles are tailored to the individual's plantar pressure distribution, and the rehabilitation training is matched to the individual's actual situation, thus realizing the precise generation of personalized orthotic plans. This invention achieves efficient connection of the entire process from data collection to suggestion output, adapts to the actual needs of clinical rehabilitation guidance, and has strong practicality and promotional value. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the overall structure of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] This invention discloses a system for full-cycle plantar pressure assessment and orthopedic suggestion generation that integrates static and dynamic data, such as... Figure 1 As shown, it includes: a data acquisition module, a database module, an assessment module, and an orthopedic suggestion module; The data acquisition module collects basic data, static and dynamic plantar pressure data from several volunteers, as well as data on foot and ankle disease types and historical orthopedic recommendations for volunteers with abnormal plantar pressure; it also collects basic data, static and dynamic plantar pressure data from individuals to be evaluated. The database module preprocesses the volunteers' basic data, static plantar pressure data, and dynamic data as sample features, and uses the corresponding foot and ankle disease type and degree as sample labels to obtain the first type of sample data and construct the first dataset; it also preprocesses the basic data, static plantar pressure data, and dynamic data of the personnel to be evaluated to obtain the predicted input data for the personnel to be evaluated. The evaluation module trains an evaluation model using the first dataset; the predicted input data is then input into the trained evaluation model to obtain the evaluation results for the person to be evaluated. The database module also preprocesses historical orthopedic suggestion data as sample output and uses intermediate features of the evaluation model as sample input to obtain a second type of sample data and construct a second dataset. The orthopedic suggestion module uses the second dataset to train a suggestion generation model. The intermediate features extracted by the evaluation model from the predicted input data are then input into the trained suggestion generation model to obtain orthopedic suggestions for the person to be evaluated.

[0021] In one specific embodiment, the data acquisition module includes a static data acquisition unit, a dynamic data acquisition unit, and an information acquisition unit; The static data acquisition unit collects plantar pressure data of volunteers and those being evaluated in a static posture with both feet evenly bearing weight and eyes looking straight ahead using a plantar pressure plate measuring device. The unit also collects the height of volunteers and those being evaluated using a medical height measuring tape and their weight using a medical electronic scale. The dynamic data acquisition unit collects plantar pressure data of volunteers and those to be evaluated during natural walking and running through a gait track laid with several plantar pressure plate measuring instruments, and uses this data as dynamic data. The information collection unit collects information on the age, gender, foot length, foot width, arch height, and leg length of volunteers and those to be evaluated through electronic medical records, doctor's consultation records, or questionnaires; it also collects the type and severity of foot and ankle diseases of volunteers, as well as corresponding historical orthopedic recommendations. Foot and ankle disorders include one or more of the following: flat feet, high arches, inversion of the foot, eversion of the foot, metatarsalgia, heel pain, and plantar fasciitis; Historical orthopedic recommendations include: the three-dimensional height of the orthopedic insole, as well as the type, frequency, and duration of rehabilitation training; Basic data includes: height, weight, age, gender, foot length, foot width, arch height, and leg length.

[0022] In one specific embodiment, preprocessing is performed on the volunteer's baseline data, static plantar pressure data, and dynamic data, as well as on the baseline data, static plantar pressure data, and dynamic data of the person to be evaluated. Specifically, this includes: Basic data preprocessing: Gender is coded; height, weight, age, foot length, foot width, arch height and leg length are cleaned, outliers are removed and filled, and then the data is standardized and sorted in order to obtain the basic data vector; Static data preprocessing: Based on the pressure values ​​of each sensor in the static plantar pressure data, the corresponding grayscale images are converted and noise filtering and smoothing are performed to obtain static feature maps; Dynamic data preprocessing: Taking the first heel strike in the dynamic data of plantar pressure as the starting point of a cycle and the second heel strike on the same side as the ending point of a cycle, dynamic data of 5 consecutive complete gait cycles of natural walking and 5 consecutive complete gait cycles of running are obtained. Based on the pressure values ​​of each sensor in each frame of the dynamic data, grayscale images of each frame are obtained, and noise filtering and smoothing are performed to finally obtain the dynamic feature map sequence. Preprocessing of historical orthodontic recommendation data includes: Based on the three-dimensional height of the orthotic insole, a coordinate system is established with the foot length direction as the X-axis, the foot width direction as the Y-axis, and the foot height direction as the Z-axis to obtain the three-dimensional height matrix of the orthotic insole; the type of rehabilitation training is one-hot encoded, and the frequency and duration are standardized to obtain the rehabilitation training vector.

[0023] In one specific embodiment, the evaluation model includes a CNN network part, a CNN-LSTM network part, and a multilayer perceptron part; The CNN network takes a static feature map as input and processes it through multiple layers of convolution, pooling, and activation functions to obtain the first feature matrix; The CNN-LSTM network takes a sequence of dynamic feature maps as input and processes them through convolution, gating, and attention mechanisms to obtain the second feature matrix. The multilayer perceptron part takes the first feature matrix, the second feature matrix and the basic data vector as input, and obtains the predicted probability of foot and ankle disease type and severity through feature fusion and mapping.

[0024] In one specific embodiment, the CNN network part sequentially includes: a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a third pooling layer, a fourth convolutional layer, and a fourth pooling layer; the number of convolutional kernels in the first to fourth convolutional layers are 16, 32, 64, and 128, respectively, the kernel size of the first to fourth convolutional layers is 3, and the activation function of the convolutional kernels is the ReLU activation function; the first to fourth pooling layers all use max pooling.

[0025] In a specific embodiment, the structure of the CNN-LSTM network is as follows: CNN layer, LSTM layer, attention layer; The CNN layers consist of: a fifth convolutional layer, a fifth pooling layer, a sixth convolutional layer, and a sixth pooling layer. The number of kernels in the fifth and sixth convolutional layers are 32 and 64, respectively, and the kernel size is 3×3. The activation function for all six convolutional layers is ReLU. Max pooling is used in all six pooling layers. The CNN layers extract features from each frame of the dynamic feature map sequence, outputting a feature map sequence with uniform dimensions. The LSTM layer consists of two stacked LSTM units, with each LSTM unit having 128 hidden units. The LSTM layer takes the feature map sequence output by the CNN layer as input and extracts the temporal feature tensor from the feature map sequence through the gating mechanism of the LSTM unit. The attention layer consists of a tensor multiplication layer, a fully connected layer, a flattening layer, a softmax layer, a reshape layer, and a lambda layer. The tensor multiplication layer receives the temporal feature tensor and performs self-attention interaction computation. The fully connected layer maps the output of the tensor multiplication layer to generate attention weight associated features. The flattening layer compresses the attention weight associated features from multidimensional features into a one-dimensional vector to obtain the initial attention weight vector. The softmax layer normalizes the initial attention weight vector so that the sum of the weights is 1. The reshape layer reshapes the normalized weight vector into a result tensor that matches the dimension of the temporal feature tensor output by the LSTM layer. The lambda layer performs element-wise multiplication between the result tensor and the temporal feature tensor, and sums and aggregates all weighted feature vectors along the temporal length direction to finally output the second feature matrix.

[0026] In one specific embodiment, the multilayer perceptron section sequentially includes: an input layer, a first fully connected layer, a second fully connected layer, and a multi-task output head; The input layer flattens the first and second feature matrices into one-dimensional vectors and concatenates them with the basic data vectors to obtain the fused feature vector. The first fully connected layer contains 520 neurons, and the second fully connected layer contains 256 neurons. Each neuron in the fully connected layer uses the ReLU activation function, and feature mapping and compression are performed through the two fully connected layers. The task output head includes a disease type prediction branch and a disease severity prediction branch. The input to both branches is the output of the second fully connected layer. The disease type prediction branch contains 7 output neurons, which use the Sigmoid activation function to output the probability of the existence of each disease. The disease severity prediction branch contains 21 output neurons (3 neurons for each disease, corresponding to mild, moderate and severe, respectively), which use the Softmax activation function to output the probability distribution of each severity level for each disease.

[0027] In one specific embodiment, it is suggested that the generative model include a first generative adversarial network and a second generative adversarial network; The first generative adversarial network takes the intermediate features as input and generates a three-dimensional height matrix of orthopedic insoles through a trained generator; the second generative adversarial network takes the intermediate features as input and generates rehabilitation training vectors through a trained generator. The intermediate features include the first feature matrix, the second feature matrix, and the basic data vector.

[0028] In one specific embodiment, the first generative adversarial network includes a generator G1 and a discriminator D1; the generator G1 includes a static branch, a dynamic branch, a weighted fusion layer, and an output layer. The static branch includes a static intermediate feature map sub-branch and a static matrix sub-branch. The static intermediate feature map sub-branch consists of two deconvolutional layers and a batch normalization layer. The first feature matrix is ​​deconvolved by the deconvolutional layers, and non-linearity is introduced through the Sigmoid activation function. The size is then adjusted by the batch normalization layer to obtain the static intermediate feature map. The static matrix sub-branch consists of three deconvolutional layers and a batch normalization layer. The basic data vector is deconvolved by the deconvolutional layers, and the size is then adjusted by the batch normalization layer to obtain the static weight matrix. The static intermediate feature map and the static weight matrix are multiplied element-wise to obtain the static height matrix. The dynamic branches include a dynamic intermediate feature map sub-branch and a dynamic matrix sub-branch. The dynamic intermediate feature map sub-branch consists of two deconvolutional layers and a batch normalization layer. The second feature matrix is ​​deconvolved by the deconvolutional layers and nonlinearity is introduced through the sigmoid activation function. The size is then adjusted by the batch normalization layer to obtain the dynamic intermediate feature map. The dynamic matrix sub-branch consists of three deconvolutional layers and a batch normalization layer. The base data vector is deconvolved by the deconvolutional layers and the size is then adjusted by the batch normalization layer to obtain the dynamic weight matrix. The dynamic intermediate feature map and the dynamic weight matrix are multiplied element-wise to obtain the dynamic height matrix. The weighted fusion layer performs a weighted fusion of the static height matrix and the dynamic height matrix, and outputs the result through the output layer. The discriminator D1 consists of three convolutional layers and two fully connected layers. It takes the real orthopedic insole 3D height matrix and the predicted 3D height matrix output by the generator G1 as input. It extracts features through the three convolutional layers and outputs the discrimination result through the two fully connected layers to determine the authenticity of the predicted 3D height matrix.

[0029] In one specific embodiment, the second generative adversarial network includes a generator G2 and a discriminator D2; The generator G2 consists of an input layer, three fully connected layers, and an output layer. The input layer flattens the first and second feature matrices into a one-dimensional vector and concatenates it with the basic data vector to obtain a fused feature vector. This vector is then passed through three fully connected layers for non-linear mapping, with each neuron in the fully connected layers using the ReLU activation function. Finally, the output layer outputs the training type probability distribution, frequency, and duration values ​​of the rehabilitation training vector. The neurons in the output layer corresponding to the rehabilitation training type use the Sigmoid activation function, while the neurons corresponding to the frequency and duration use the Linear activation function. The discriminator D2 takes the real rehabilitation training vector and the predicted rehabilitation training vector output by the generator G2 as input, and processes them through two fully connected layers in sequence to output the discrimination result and determine the authenticity of the rehabilitation training vector.

[0030] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0031] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A system for full-cycle assessment of plantar pressure and generation of orthopedic recommendations integrating static and dynamic data, characterized in that, include: Data acquisition module, database module, assessment module, and orthodontic suggestion module; The data acquisition module collects basic data, static plantar pressure data, and dynamic data from several volunteers, as well as foot and ankle disease types and historical orthopedic recommendations for volunteers with abnormal plantar pressure; it also collects basic data, static plantar pressure data, and dynamic data from individuals to be evaluated. The database module preprocesses the volunteers' basic data, static plantar pressure data, and dynamic data as sample features, and uses the corresponding foot and ankle disease type and degree as sample labels to obtain the first type of sample data and construct the first dataset; it also preprocesses the basic data, static plantar pressure data, and dynamic data of the personnel to be evaluated to obtain the predicted input data of the personnel to be evaluated. The evaluation module trains an evaluation model using the first dataset; it then inputs the predicted input data into the trained evaluation model to obtain the evaluation results for the person to be evaluated. The database module further preprocesses the historical orthopedic suggestion data as a sample output and uses the intermediate features of the evaluation model as a sample input to obtain a second type of sample data and construct a second dataset. The orthopedic suggestion module trains a suggestion generation model using the second dataset, and inputs the intermediate features extracted by the evaluation model from the predicted input data into the trained suggestion generation model to obtain orthopedic suggestions for the person to be evaluated.

2. The system for full-cycle assessment and orthopedic suggestion generation of plantar pressure integrating static and dynamic data as described in claim 1, characterized in that, The data acquisition module includes a static data acquisition unit, a dynamic data acquisition unit, and an information acquisition unit; The static data acquisition unit collects plantar pressure data of volunteers and those to be evaluated in a static posture with their feet evenly distributed and their eyes looking straight ahead, using a plantar pressure plate measuring instrument. This data is used as the static data. The height of volunteers and those to be evaluated was collected using a medical height chart, and their weight was collected using a medical electronic scale. The dynamic data acquisition unit collects plantar pressure data of volunteers and those to be evaluated during natural walking and running through a gait track laid with several plantar pressure plate measuring instruments, and uses this data as the dynamic data. The information collection unit collects information on the age, gender, foot length, foot width, arch height, and leg length of volunteers and those to be evaluated through electronic medical records, doctor's consultation records, or questionnaires. The volunteers' foot and ankle disease types and severity, as well as corresponding historical orthopedic recommendations, were collected. The foot and ankle disorders include one or more of the following: flat feet, high arches, inversion of the foot, eversion of the foot, metatarsalgia, heel pain, and plantar fasciitis; The historical orthotic recommendation data includes: the three-dimensional height of the orthotic insole, as well as the type, frequency, and duration of rehabilitation training; The basic data includes: height, weight, age, gender, foot length, foot width, arch height, and leg length.

3. The system for full-cycle assessment and orthopedic suggestion generation of plantar pressure integrating static and dynamic data as described in claim 2, characterized in that, The preprocessing of basic data, static plantar pressure data, and dynamic data of volunteers, as well as the preprocessing of basic data, static plantar pressure data, and dynamic data of individuals to be evaluated, specifically includes: Basic data preprocessing: Gender is coded; height, weight, age, foot length, foot width, arch height and leg length are cleaned, outliers are removed and filled, and then the data is standardized and sorted in order to obtain the basic data vector; Static data preprocessing: Based on the pressure values ​​of each sensor in the static plantar pressure data, the corresponding grayscale images are converted and noise filtering and smoothing are performed to obtain static feature maps; Dynamic data preprocessing: Taking the first heel strike in the dynamic data of plantar pressure as the starting point of a cycle and the second heel strike on the same side as the ending point of a cycle, dynamic data of 5 consecutive complete gait cycles of natural walking and 5 consecutive complete gait cycles of running are obtained. Based on the pressure values ​​of each sensor in each frame of the dynamic data, grayscale images of each frame are obtained, and noise filtering and smoothing are performed to finally obtain the dynamic feature map sequence. The preprocessing of the historical orthodontic recommendation data specifically includes: Based on the three-dimensional height of the orthotic insole, a coordinate system is established with the foot length direction as the X-axis, the foot width direction as the Y-axis, and the foot height direction as the Z-axis to obtain the three-dimensional height matrix of the orthotic insole; the type of rehabilitation training is one-hot encoded, and the frequency and duration are standardized to obtain the rehabilitation training vector.

4. The system for full-cycle assessment and orthopedic suggestion generation of plantar pressure integrating static and dynamic data as described in claim 3, characterized in that, The evaluation model includes a CNN network component, a CNN-LSTM network component, and a multilayer perceptron component. The CNN network takes the static feature map as input and processes it through multiple layers of convolution, pooling, and activation functions to obtain the first feature matrix; The CNN-LSTM network takes the dynamic feature map sequence as input and processes it through convolution, gating mechanism and attention mechanism to obtain the second feature matrix; The multilayer perceptron takes the first feature matrix, the second feature matrix, and the basic data vector as input, and obtains the predicted probability of foot and ankle disease type and severity through feature fusion and mapping.

5. The system for full-cycle assessment and orthopedic suggestion generation of plantar pressure integrating static and dynamic data according to claim 4, characterized in that, The CNN network consists of the following layers in sequence: a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a third pooling layer, a fourth convolutional layer, and a fourth pooling layer. The number of convolutional kernels in the first to fourth convolutional layers are 16, 32, 64, and 128, respectively. The kernel size in all of the first to fourth convolutional layers is 3, and the activation function of the convolutional kernels is ReLU. Max pooling is used in all of the first to fourth pooling layers.

6. The system for full-cycle assessment and orthopedic suggestion generation of plantar pressure integrating static and dynamic data according to claim 4, characterized in that, The structure of the CNN-LSTM network is as follows: CNN layer, LSTM layer, attention layer; The CNN layers sequentially include: a fifth convolutional layer, a fifth pooling layer, a sixth convolutional layer, and a sixth pooling layer; the number of convolutional kernels in the fifth and sixth convolutional layers are 32 and 64 respectively, the kernel size of the fifth and sixth convolutional layers is 3×3, and the activation function is ReLU activation function; the fifth and sixth pooling layers all use max pooling; the CNN layers extract features from the feature maps of each frame in the dynamic feature map sequence, and output a feature map sequence with uniform dimensions; The LSTM layer consists of two stacked LSTM units, with each LSTM unit having 128 hidden units. The LSTM layer takes the feature map sequence output by the CNN layer as input and extracts the temporal feature tensor from the feature map sequence through the gating mechanism of the LSTM unit. The attention layer sequentially comprises a tensor multiplication layer, a fully connected layer, a flattening layer, a softmax layer, a reshape layer, and a lambda layer. The tensor multiplication layer receives the temporal feature tensor and performs self-attention interaction computation. The fully connected layer performs a fully connected mapping on the output of the tensor multiplication layer to generate attention weight associated features. The flattening layer compresses the attention weight associated features from multidimensional features into a one-dimensional vector to obtain an initial attention weight vector. The softmax layer normalizes the initial attention weight vector so that the sum of the weights is 1. The reshape layer reshapes the normalized weight vector into a result tensor that matches the dimension of the temporal feature tensor output by the LSTM layer. The lambda layer performs element-wise multiplication between the result tensor and the temporal feature tensor, and sums and aggregates all weighted feature vectors along the temporal length direction to finally output the second feature matrix.

7. The system for full-cycle assessment and orthopedic suggestion generation of plantar pressure integrating static and dynamic data as described in claim 4, characterized in that, The multilayer perceptron section comprises, in sequence: an input layer, a first fully connected layer, a second fully connected layer, and a multi-task output head; The input layer flattens the first feature matrix and the second feature matrix into a one-dimensional vector, and concatenates it with the basic data vector to obtain a fused feature vector. The first fully connected layer contains 520 neurons, and the second fully connected layer contains 256 neurons. Each neuron in the fully connected layer uses the ReLU activation function, and feature mapping and compression are performed through the two fully connected layers. The task output head includes a disease type prediction branch and a disease severity prediction branch, with the output of the second fully connected layer as input to both branches. The disease type prediction branch contains 7 output neurons, which use the Sigmoid activation function to output the probability of the existence of each disease. The disease severity prediction branch contains 21 output neurons, which use the Softmax activation function to output the probability distribution of each severity level corresponding to each disease.

8. The system for full-cycle assessment and orthopedic suggestion generation of plantar pressure integrating static and dynamic data according to claim 4, characterized in that, The proposed generative model includes a first generative adversarial network and a second generative adversarial network; The first generative adversarial network takes the intermediate features as input and generates a three-dimensional height matrix of orthopedic insoles through a trained generator; the second generative adversarial network takes the intermediate features as input and generates rehabilitation training vectors through a trained generator. The intermediate features include the first feature matrix, the second feature matrix, and the basic data vector.

9. A system for full-cycle assessment and orthopedic suggestion generation of plantar pressure integrating static and dynamic data as described in claim 8, characterized in that, The first generative adversarial network includes a generator G1 and a discriminator D1; the generator G1 includes a static branch, a dynamic branch, a weighted fusion layer, and an output layer; The static branch includes a static intermediate feature map sub-branch and a static matrix sub-branch. The static intermediate feature map sub-branch sequentially includes two deconvolution layers and a batch normalization layer. The first feature matrix is ​​deconvolved by the deconvolution layers, and nonlinearity is introduced by the sigmoid activation function. The size is then adjusted by the batch normalization layer to obtain the static intermediate feature map. The static matrix sub-branch sequentially includes three deconvolution layers and a batch normalization layer. The basic data vector is deconvolved by the deconvolution layers, and the size is then adjusted by the batch normalization layer to obtain the static weight matrix. The static intermediate feature map and the static weight matrix are multiplied element-wise to obtain the static height matrix. The dynamic branch includes a dynamic intermediate feature map sub-branch and a dynamic matrix sub-branch. The dynamic intermediate feature map sub-branch sequentially includes two deconvolution layers and a batch normalization layer. The second feature matrix is ​​deconvolved by the deconvolution layers, and nonlinearity is introduced by the sigmoid activation function. The size is then adjusted by the batch normalization layer to obtain the dynamic intermediate feature map. The dynamic matrix sub-branch sequentially includes three deconvolution layers and a batch normalization layer. The base data vector is deconvolved by the deconvolution layers, and the size is then adjusted by the batch normalization layer to obtain the dynamic weight matrix. The dynamic intermediate feature map and the dynamic weight matrix are multiplied element-wise to obtain the dynamic height matrix. The weighted fusion layer performs a weighted fusion of the static height matrix and the dynamic height matrix, and outputs the result through the output layer. The discriminator D1 includes three convolutional layers and two fully connected layers. It takes the actual orthopedic insole 3D height matrix and the predicted 3D height matrix output by the generator G1 as inputs. It extracts features through the three convolutional layers and outputs the discrimination result through the two fully connected layers to determine the authenticity of the predicted 3D height matrix.

10. The system for full-cycle assessment and orthopedic suggestion generation of plantar pressure integrating static and dynamic data as described in claim 8, characterized in that, The second generative adversarial network includes a generator G2 and a discriminator D2; The generator G2 includes an input layer, three fully connected layers, and an output layer. The input layer flattens the first feature matrix and the second feature matrix into a one-dimensional vector and concatenates it with the basic data vector to obtain a fused feature vector. Then, it passes through three fully connected layers for nonlinear mapping, with each neuron in the fully connected layer using the ReLU activation function. Finally, the output layer outputs the training type probability distribution, frequency, and duration values ​​of the rehabilitation training vector. The neurons in the output layer corresponding to the rehabilitation training type use the Sigmoid activation function, while the neurons corresponding to the frequency and duration use the Linear activation function. The discriminator D2 takes the real rehabilitation training vector and the predicted rehabilitation training vector output by the generator G2 as input, and processes them through two fully connected layers in sequence to output the discrimination result and determine the authenticity of the rehabilitation training vector.

Images