Lower limb gait data acquisition and processing method and system for auxiliary diagnosis of Parkinson's disease

Abstract

The invention relates to the technical field of medical detection, in particular to a lower limb gait data collecting and processing method and system for auxiliary diagnosis of Parkinson's disease, and after obtained lower limb gait data of a patient is subjected to filtering, noise reduction, fusion, normalization and other preprocessing operations, a training data set is created. In the aspect of classification, an improved CNN network is combined with a converter, and correlation between gait data is classified. The CNN network is used for deep feature extraction, replaces an embedded layer of Transform as input, does not perform flattening processing, optimizes a masking process, and calculates the global correlation of data through a multi-head attention mechanism. And finally, performing classified output by using an MLP head. The network structure allows parallel data input, so that the calculation time is greatly reduced. The improved network structure not only realizes higher classification accuracy, but also reduces the network parameter volume, so that the model is possible to be implanted into mobile and embedded equipment.

Description

Technical Field

[0001] The present invention relates to the field of medical detection technology, and in particular to a method and system for collecting and processing lower limb gait data for assisting in diagnosing Parkinson's disease. Background Art

[0002] Clinically, the clinical diagnosis of Parkinson's disease often requires a lengthy and expensive pathological diagnosis. Patients often show unstable postures. Since Parkinson's disease is accompanied by non-motor symptoms, including autonomic dysfunction, depression, and hallucinations, the clinical diagnosis of Parkinson's disease often requires a lengthy and expensive pathological diagnosis, including neuroimaging, genetic diagnosis, biochemical tests, and drug experiments. The assessment of Parkinson's patients includes the Unified Parkinson's Disease Rating Scale (UPDRS) and the Hoehn & Yahr Rating Scale (H&Y). Although these methods can obtain relatively accurate clinical results for Parkinson's patients, their disadvantages are also obvious, namely, time-consuming, expensive, and complicated.

[0003] With the development of wearable sensors, and because gait-related data has the characteristics of periodicity, measurable range of motion, and spatiotemporal correlation, scholars continue to use various algorithms to analyze it to detect or predict Parkinson's disease. Some studies use supervised machine learning algorithms to train classification tasks for Parkinson's patients. However, most of these studies still focus on binary or multivariate classification problems that distinguish patients from healthy people. In addition, for data such as human gait, it is necessary to manually create dataset features for classification, which not only leads to high dimensionality and large data volume, but also lacks robustness. Summary of the invention

[0004] In view of this, the purpose of the present invention is to propose a method and system for collecting and processing lower limb gait data for auxiliary diagnosis of Parkinson's disease, so as to solve the problems of high dimension, large data volume and lack of robustness in the existing methods for diagnosing Parkinson's disease.

[0005] Based on the above purpose, the present invention provides a method for collecting and processing lower limb gait data for assisting in the diagnosis of Parkinson's disease, comprising the following steps:

[0006] S1. Obtain the time series data of the multimodal gait of the patient's lower limbs, use the improved CNN backbone network to extract the features of the multimodal data of the lower limbs from the time series data, and output the token matrix of the extracted features;

[0007] S2. Use the position encoder to position encode the token matrix of the extracted features output by the improved CNN backbone network to obtain the position encoding vector token;

[0008] S3, input the position encoding vector token into the Transformer encoder, use the multi-head attention mechanism to obtain the output vector of enhanced features;

[0009] S4. Input the output vector of the Transformer encoder into the multi-layer perceptron and output the classification result of the time series data of the multimodal gait of the patient's lower limbs.

[0010] Preferably, the improved CNN backbone network includes: a first 1D convolution layer with a convolution kernel size of 3, a second 1D convolution layer with a convolution kernel size of 4, a first maximum pooling layer, a third 1D convolution layer with a convolution kernel size of 4 and a second maximum pooling layer, and each of the first 1D convolution layer and the second 1D convolution layer is followed by a ReLU activation function and a batch normalization layer.

[0011] Preferably, the inactivation rate of the improved CNN backbone network is set to 0.2 to reduce network parameters and prevent overfitting.

[0012] Preferably, the calculation formula of the position code is:

[0013]

[0014] Among them, PE represents position encoding, and pos is the dimension d model The position index of the input vector, i∈[0,d model / 2] is the dimension of the current position vector encoding, d model Represents the dimension of the data after CNN layer conversion;

[0015] Select trigonometric function as position encoding, and use the following formula to calculate the relative position relationship between vectors separated by k vectors:

[0016] sin(α+β)=sinαcosβ+cosαsinβ

[0017] cos(α+β)=cosαcosβ-sinαsinβ

[0018] The vector code at position pos+k is determined by the vector codes at positions pos and k, and the calculation formula is: PE(pos+k,2i)=PE(pos,2i)×PE(k,2i+1)+PE(pos,2i+1)×PE(k,2i);

[0019] PE(pos+k,2i+1)=PE(pos,2i+1)×PE(k,2i+1)-PE(pos,2i)×PE(k,2i);

[0020] Among them, k represents the interval between two position encoding vectors, and α is β is

[0021] Preferably, in the Transformer encoder, the multi-head attention mechanism parallelizes the calculations of the quary, key and value in multiple dot product attentions, and concatenates the results of each parallelization and outputs them, specifically including:

[0022] Linearly project quary, key and value h times to d respectively q ,d k and d v Dimensionally;

[0023] Setting k =d v =d model / h correlation;

[0024] Perform parallel attention operations to obtain k Heads and connect them in series;

[0025] Multiply all heads by the matrix w o , the output is obtained, and the output dimension is consistent with the output of each layer of the residual connection encoder;

[0026] The formula for multi-head attention is:

[0027] MultiHead(Q,K,V)=Concat(Head 1 ,…,head h )W O #

[0028]

[0029] in,

[0030] are parameter matrices.

[0031] Preferably, the calculation formula of dot product attention is:

[0032]

[0033] in, Represents the scaling factor.

[0034] Preferably, the multilayer perceptron includes two fully connected layers, the middle hidden layer is 4 times the input dimension, and the output is the number of categories.

[0035] Preferably, the method further comprises:

[0036] S5. Use the optimizer for gradient update in the back propagation process in deep learning. The optimizer formula is:

[0037] m t =β 1 ·m t-1 +(1-β 1 )·g(w t )#

[0038] v t =β 2 ·v t-1 +(1-β 2 )·g 2 (w t )#

[0039]

[0040] Among them, β is set 1 = 0.9 and β 2 =0.999, α is the learning rate. m t and v t are the first-order momentum and the second-order momentum, g(w t ) is the parameter w t The loss gradient at time t, ∈, is set to a very small positive number that prevents the denominator from being zero.

[0041] The present invention also provides a lower limb gait data acquisition and processing system for assisting in the diagnosis of Parkinson's disease, comprising:

[0042] A data acquisition component, used for acquiring gait data of the patient's lower limbs;

[0043] The data processing module is used to process the patient's lower limb gait data collected by the data acquisition component using the patient's lower limb gait data processing method as described in any one of claims 1 to 8, and obtain the classification result of the patient's lower limb multimodal gait time series data.

[0044] Preferably, the data acquisition component includes:

[0045] An inertial measurement unit with orthogonally mounted triaxial MEMS gyroscopes and triaxial accelerometers installed at the patient's left and right ankle joints to measure the angular velocity and acceleration of the patient's lower limbs along three orthogonal axes;

[0046] Pressure sensors are installed on the front and back soles of the patient's feet to collect pressure data.

[0047] Beneficial effects of the present invention:

[0048] This system can be worn on the lower limbs of PD patients for a short period of time to collect and analyze data, and obtain results based on the H&Y table classification in a short period of time, helping doctors to quickly diagnose the patient's condition, thereby reducing the time-consuming and costly processes such as pharmacological analysis and clinical observation in the original diagnosis process.

[0049] The present invention creates a training data set by subjecting the acquired patient lower limb gait data to preprocessing operations such as filtering, denoising, fusion and normalization. In terms of classification, the improved CNN network is combined with the transformer to classify the correlation between gait data. The CNN network is used for deep feature extraction, replacing the embedding layer of the Transformer as input, without flattening, optimizing the masking process, and calculating the global correlation of the data through a multi-head attention mechanism. Finally, the MLP head is used for classification output. This network structure allows parallel input of data, greatly reducing the calculation time. The improved network structure not only achieves higher classification accuracy, but also reduces the network parameter volume, making it possible to implant the model into mobile and embedded devices. Therefore, the rapid diagnosis system for staging of multiple sclerosis patients proposed by us has good portability, can be remotely monitored and diagnosed, and can ensure the speed and accuracy of diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0051] Figure 1 A block diagram of a data processing method according to an embodiment of the present invention;

[0052] Figure 2 A data processing model block diagram of an embodiment of the present invention;

[0053] Figure 3 This is a structural diagram of an improved CNN backbone network according to an embodiment of the present invention;

[0054] Figure 4 Schematic diagram of the encoder structure of an embodiment of the present invention;

[0055] Figure 5 A schematic diagram of a scaled dot product attention mechanism according to an embodiment of the present invention;

[0056] Figure 6 Schematic diagram of a multi-head attention mechanism according to an embodiment of the present invention;

[0057] Figure 7 A schematic diagram of a confusion matrix evaluation index according to an embodiment of the present invention;

[0058] Figure 8 Schematic diagram of network loss and accuracy according to an embodiment of the present invention. DETAILED DESCRIPTION

[0059] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific embodiments.

[0060] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0061] The present invention provides a method for collecting and processing lower limb gait data to assist in the diagnosis of Parkinson's disease. The overall structure is as follows: Figure 1 As shown, the method comprises the following steps:

[0062] S1. Obtain the time series data of the multimodal gait of the patient's lower limbs, use the improved CNN backbone network to extract the features of the multimodal data of the lower limbs from the time series data, and output the token matrix of the extracted features. The improved CNN backbone network structure is as follows: Figure 4 As shown in Figure 2, the network mainly realizes the high-dimensional mapping of data features through the improved inverted residual block structure, and increases the number of time series inputs through a convolution kernel of size 3 to improve its feature dimension. c×l .

[0063] Among them, c is the number of input data channels, which is also the data width, and l is the time series window width. Then, the depthwise convolution in the depthwise separable convolution is used, and the convolution kernel size is 4. Finally, the feature matrix is ​​reduced in dimension using the convolution kernel of size 4. Except for the last convolution layer, each convolution operation is followed by a ReLU activation function and a batch normalization layer. To avoid overfitting, neurons are randomly inactivated after the last convolution layer. The inactivation rate is set to 0.2. Finally, without flattening, the output dimension is directly d model The token matrix, d model=c.

[0064] S2. Use the position encoder to position encode the token matrix of the extracted features output by the improved CNN backbone network to obtain the position encoding vector token;

[0065] Since the attention mechanism processes all input data in parallel and calculates the correlation between them, it does not pay attention to the positional relationship of the input data. However, for Parkinson's patients, the occurrence of abnormal gait is also closely related to their relative position before and after, so the positional relationship is used to calculate and reflect this relationship.

[0066]

[0067] Among them, pos is the dimension d model The position index of the input vector, i∈[0,d model / 2] is the dimension of the current position vector encoding, d model Represents the dimension of the data after CNN layer conversion, d model It is a custom parameter in the model, and this parameter runs through the Transformer model. In this model, CNN is used instead of the word embedding layer in the classic Transformer to convert the input data into a dimension of d. model The dimension of the data before and after passing through each layer in the Encoder remains d model .

[0068] sin(α+β)=sinαcosβ+cosαsinβ

[0069] cos(α+β)=cosαcosβ-sinαsinβ

[0070] These two formulas are sum and difference formulas of trigonometric functions, which are used to derive the positional relationship between the pos position and the pos+k position vector encoding in the following formula. α and β are respectively used in the derivation process. and Substitute in. The advantage of choosing trigonometric functions as position encoding is their periodicity, so the above formula can be used to calculate the relative position relationship between vectors separated by the k vector. The relative relationship between pos and pos+k can be obtained according to the above formula, where k represents the interval between the two position encoding vectors, which is used to explain how the position encoding reflects the relative position relationship below.

[0071] PE(pos+k,2i)=PE(pos,2i)×PE(k,2i+1)+PE(pos,2i+1)×PE(k,2i)

[0072] PE(pos+k,2i+1)=PE(pos,2i+1)×PE(k,2i+1)-PE(pos,2i)×PE(k,2i)

[0073] The vector encoding of the position pos+k can be determined by the vector encoding of the position pos and k. This means that the relative position relationship between the vectors is also included in the k position encoding.

[0074] S3, input the position encoding vector token into the Transformer encoder, use the multi-head attention mechanism to obtain the output vector of enhanced features;

[0075] In a conventional Transformer, there are two parts, the encoder and the decoder, which are used for feature extraction and output prediction respectively. In this embodiment, only the encoder part is used. Compared with the classic encoder, the input of gait data has a neat shape. Unlike the translation task, sentences of different lengths need to be masked to align the shapes, while in the encoder, the masking operation is cancelled. Figure 4 As shown in Figure 1, the encoder consists of multiple blocks, each of which contains two sub-layers, namely multi-head attention and MLP, in which residual connections and normalization operations are added. Compared with the classic encoder, the input of gait data has a neat shape.

[0076] The attention mechanism in Transformer is scaled dot product attention. Figure 6 As shown in Figure 2, the attention value between different tags is obtained by performing a dot product operation on the quary, key, and value vectors. The attention is calculated according to the following formula to obtain a dimension d model To avoid information loss due to the value after the dot product being too large or too small, the author introduces As a scaling factor, keep values ​​within a reasonable range.

[0077]

[0078] In order to extract different dimensions and data correlations, the Transformer in this embodiment adopts a multi-head attention mechanism. Figure 6 As shown in , quary, key, and value are linearly projected h times to d respectively. q ,d k and d v In order to maintain a similar computational complexity to self-attention, we set d k =d v =d model / h correlation. Then perform parallel attention operation to obtain k heads and connect them in series. Finally, right-multiply all heads by matrix wo , and get the output. The output dimension is consistent with the output of each layer of the residual connection encoder.

[0079] MultiHead(Q,K,V)=Concat(head 1 ,…,head h )W O #

[0080]

[0081] Among them, the parameter matrix

[0082] To prevent overfitting, a deactivation operation with a deactivation rate of 0.2 was added before normalization. After 6 encoder structures, the dimension was d model Output.

[0083] S4. Input the output vector of the Transformer encoder into the multi-layer perceptron and output the classification result of the time series data of the multimodal gait of the patient's lower limbs.

[0084] After CNN and Transformer processing, the final output size is batcsize×w×d model The expected output size of the four categories is batchsize×4, and two fully connected layers are set. The middle hidden layer is 4 times the input dimension, and the output is the number of categories. Random dropout is added between the two linear fully connected layers to prevent overfitting.

[0085] As an embodiment, the method further includes:

[0086] S5. Use the optimizer for gradient update in the back propagation process in deep learning. The optimizer formula is:

[0087] m t =β 1 ·m t-1 +(1-β 1 )·g(w t )#

[0088] v t =β 2 ·v t-1 +(1-β 2 )·g 2 (w t )#

[0089]

[0090] Among them, β is set 1 = 0.9 and β 2= 0.999, α is the learning rate. m and v are the first-order momentum and the second-order momentum respectively, g(w) is the loss gradient of parameter w at time t, and ∈ is set to a very small positive number to prevent the denominator from being zero.

[0091] The learning rate determines the speed of each parameter update. If the speed is too fast or too slow, it will hinder the model from converging to the optimal solution. In order to update the network more efficiently, the encoder uses a warm-up variable learning rate method. First, the learning rate rises linearly from a very small value to the set maximum value, and then decays exponentially. Set the step size num=3000.

[0092] The embodiment of this specification also provides a lower limb gait data acquisition and processing system for assisting in the diagnosis of Parkinson's disease, comprising:

[0093] A data acquisition component, used for acquiring gait data of the patient's lower limbs;

[0094] The data processing module is used to process the patient's lower limb gait data collected by the data acquisition component using the patient's lower limb gait data processing method as described in any one of claims 1 to 8, and obtain the classification result of the patient's lower limb multimodal gait time series data.

[0095] As an implementation method, the data collection component includes:

[0096] An inertial measurement unit with orthogonally mounted triaxial MEMS gyroscopes and triaxial accelerometers installed at the patient's left and right ankle joints to measure the angular velocity and acceleration of the patient's lower limbs along three orthogonal axes;

[0097] Pressure sensors are installed on the front and back soles of the patient's feet to collect pressure data.

[0098] Specifically, the inertial measurement unit can use the IMU18B sensor, which has a digital control circuit inside. The controller compensates for errors such as temperature and installation error angle, and finally outputs information such as angular velocity, acceleration and temperature to the outside through the RS422 communication protocol.

[0099] The pressure sensor can use an FSR pressure sensor, which is a piezoresistor installed on the front and back sides of the patient's foot to collect pressure data. The main control chip uses the STM32 processor of STMicroelectronics. The functions implemented by the peripheral circuit include RS422 to TTL conversion, pressure analog signal conversion, and sending data to the cloud.

[0100] Experiments were conducted on multi-classification and binary classification problems using the improved CNN transformer model. The experiments were validated using time series data from four categories, namely healthy, stage one to stage three. In most cases, the "healthy" and "stage one" categories performed best, with an accuracy of 99% on the test set, while the "stage two" category also achieved an accuracy of more than 96%. In binary classification, the results are shown in Table 1, which can completely distinguish between the stage two and stage three experimental groups. In most cases, the stage one and stage three categories performed best, with an accuracy of 100% on the test set, while the other two categories also achieved an accuracy of more than 96%.

[0101] Table 1. Binary classification accuracy

[0102]

[0103] Next, the evaluation indicators such as accuracy, sensitivity, specificity, precision and f1 score were calculated based on the confusion matrix, as shown in Table 2 and Figure 7 shown.

[0104] The results show that the detection accuracy of normal people is 96.70%, and the binary classification problem of Parkinson's disease patients and healthy people can also achieve a high recognition rate. In the multi-classification problem of patients in different stages (stages one to three), the accuracy of stage one patients is the lowest, only 97.83%. The misclassification rate of stage two patients is only 1.04%.

[0105] Table 2. Evaluation indicators of this method

[0106]

[0107] In order to verify the performance of the model proposed in this method (hereinafter referred to as the improved CNN-Transformer), a network model corresponding to the self-built dataset was established based on the multi-head convolutional neural network proposed by Luigi Borzì et al. Borzì, Sigcha, Rodríguez-Martín and Olmo (2023) and the LSTM neural network designed by E et al. Balaji et al. (2021). Their feature channels were modified to the feature channels of the self-built dataset, and the results obtained after training on the same dataset are as follows Figure 8 As shown in the figure, the multi-head CNN has the highest accuracy on the self-built dataset, reaching 83.3%, while the LSTM has an accuracy of 91.6%. The multi-head CNN and the improved CNN-Transformer model are comparable in running speed and can both run on the CPU, but compared with the LSTM, the improved CNN-Transformer model reduces the number of parameters by half while ensuring accuracy, greatly improving the running speed.

[0108] Those skilled in the art should understand that the discussion of any of the above embodiments is only exemplary and is not intended to imply that the scope of the present invention is limited to these examples; under the concept of the present invention, the technical features in the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other changes in different aspects of the present invention as described above, which are not provided in detail for the sake of simplicity. Any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for collecting and processing lower limb gait data for assisting diagnosis of Parkinson's disease, characterized in that: The method comprises the following steps: S1. Obtain the time series data of the multimodal gait of the patient's lower limbs, use the improved CNN backbone network to extract the features of the multimodal data of the lower limbs from the time series data, and output the token matrix of the extracted features; S2. Use the position encoder to position encode the token matrix of the extracted features output by the improved CNN backbone network to obtain the position encoding vector token; S3, input the position encoding vector token into the Transformer encoder, use the multi-head attention mechanism to obtain the output vector of enhanced features; S4. Input the output vector of the Transformer encoder into the multi-layer perceptron and output the classification result of the time series data of the multimodal gait of the patient's lower limbs.

2. The method for collecting and processing lower limb gait data for assisting diagnosis of Parkinson's disease according to claim 1, characterized in that: The improved CNN backbone network includes: a first 1D convolution layer with a convolution kernel size of 3, a second 1D convolution layer with a convolution kernel size of 4, a first maximum pooling layer, a third 1D convolution layer with a convolution kernel size of 4 and a second maximum pooling layer, and the first 1D convolution layer and the second 1D convolution layer are each followed by a ReLU activation function and a batch normalization layer.

3. The method for collecting and processing lower limb gait data for assisting diagnosis of Parkinson's disease according to claim 2, characterized in that: The inactivation rate of the improved CNN backbone network is set to 0.2 to reduce network parameters and prevent overfitting.

4. The method for collecting and processing lower limb gait data for assisting diagnosis of Parkinson's disease according to claim 1, characterized in that: The calculation formula of the position code is: Among them, PE represents position encoding, and pos is the dimension d model The position index of the input vector, i∈[0,d model / 2] is the dimension of the current position vector encoding, d model Represents the dimension of the data after CNN layer conversion; Select trigonometric function as position encoding, and use the following formula to calculate the relative position relationship between vectors separated by k vectors: sin(α+β)=sinαcosβ+cosαsinβ cos(α+β)=cosαcosβ-sinαsinβ The vector code at position pos+k is determined by the vector codes at positions pos and k, and the calculation formula is: PE(pos+k,2i)=PE(pos,2i)×PE(k,2i+1)+PE(pos,2i+1)×PE(k,2i); PE(pos+k,2i+1)= PE(pos,2i+1)×PE(k,2i+1)-PE(pos,2i)×PE(k,2i); Among them, k represents the interval between two position encoding vectors, and α is β is 5. The method for collecting and processing lower limb gait data for assisting diagnosis of Parkinson's disease according to claim 1, characterized in that: In the Transformer encoder, the multi-head attention mechanism parallelizes the calculations of the quary, key, and value in multiple dot product attentions, and concatenates the results of each parallelization and outputs them, specifically including: Linearly project quary, key and value h times to d respectively q ,d k and d v Dimensionally; Setting k =d v =d model / h correlation; Perform parallel attention operations to obtain k Heads and connect them in series; Multiply all heads by the matrix w o , the output is obtained, and the output dimension is consistent with the output of each layer of the residual connection encoder; The formula for multi-head attention is: in, are parameter matrices.

6. The method for collecting and processing lower limb gait data for assisting diagnosis of Parkinson's disease according to claim 5, characterized in that: The calculation formula of the dot product attention is: in, Represents the scaling factor.

7. The method for collecting and processing lower limb gait data for assisting diagnosis of Parkinson's disease according to claim 1, characterized in that: The multilayer perceptron includes two fully connected layers, the middle hidden layer is 4 times the input dimension, and the output is the number of categories.

8. The method for collecting and processing lower limb gait data for assisting diagnosis of Parkinson's disease according to claim 1, characterized in that: The method further comprises: S5. Use the optimizer for gradient update in the back propagation process in deep learning. The optimizer formula is: m t =β1·m t-1 +(1-β1)·g(w t )# v t =β2·v t-1 +(1-β2)·g 2 (w t )# Among them, β1=0.9 and β2=0.999 are set, and α is the learning rate. t and v t are the first-order momentum and the second-order momentum, g(w t ) is the parameter w t The loss gradient at time t, ∈, is set to a very small positive number that prevents the denominator from being zero.

9. A lower limb gait data acquisition and processing system for assisting in the diagnosis of Parkinson's disease, characterized in that: include: A data acquisition component, used for acquiring gait data of the patient's lower limbs; The data processing module is used to process the patient's lower limb gait data collected by the data acquisition component using the patient's lower limb gait data processing method as described in any one of claims 1 to 8, and obtain the classification result of the patient's lower limb multimodal gait time series data.

10. The lower limb gait data acquisition and processing system for assisting diagnosis of Parkinson's disease according to claim 9, characterized in that: The data acquisition component includes: An inertial measurement unit with orthogonally mounted triaxial MEMS gyroscopes and triaxial accelerometers installed at the patient's left and right ankle joints to measure the angular velocity and acceleration of the patient's lower limbs along three orthogonal axes; Pressure sensors are installed on the front and back soles of the patient's feet to collect pressure data.

Images