Parkinson's disease gait analysis and diagnosis method based on deep learning

Abstract

The invention discloses a Parkinson's disease gait analysis and diagnosis method based on deep learning, and belongs to the technical field of medical signal processing. The method comprises the following steps: acquiring multi-modal time sequence gait data of Parkinson's disease, preprocessing the multi-modal time sequence gait data, and dividing a data set to obtain preprocessed multi-modal time sequence gait data; the multi-modal time sequence gait data at least comprises foot pressure data, lower limb motion acceleration and lower limb swing angular velocity; designing a feature extraction network structure of a multi-channel convolutional neural, extracting features from the multi-modal time sequence gait data, and obtaining useful features from local to global; establishing a Transform feature fusion module, and introducing a multi-head self-attention mechanism to process all input useful feature data in parallel; a softmax classification diagnosis module is provided, Dropout and regularization technologies are adopted to avoid overfitting of data, an Adam optimizer is utilized to optimize model parameters, and a Parkinson's disease gait analysis diagnosis result is obtained.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of medical signal processing, and particularly relates to a Parkinson's disease gait analysis diagnosis method based on deep learning. BACKGROUND

[0002] Traditional Parkinson's disease diagnosis methods mainly rely on clinicians and diagnosis scales, and such diagnosis methods have strong subjectivity and are difficult to realize objective quantitative evaluation of gait movement disorders of Parkinson's disease patients. Therefore, Parkinson's disease gait analysis is a key technology for solving the problems of low efficiency and strong subjectivity of traditional Parkinson's disease diagnosis methods.

[0003] In recent years, with the development of deep learning technology, more and more researches begin to explore the use of gait data to assist in the diagnosis of Parkinson's disease. Gait information has significant characteristics such as periodicity, stability, rhythm and space-time, and is not only cost-effective, but also easy to obtain. In order to solve the gait feature extraction task, the existing manual feature extraction method has the problems of high dimension, large amount of calculation and poor robustness. Deep learning can automatically extract and learn high-level features from raw time series data, avoiding complex manual feature extraction, and providing a new idea for early diagnosis and disease monitoring of Parkinson's disease. In addition, considering the spatial distribution between multiple sensors, different size features and data correlation are extracted, and feature extraction and correlation analysis are performed from different angles. Deep learning technology shows great potential in Parkinson's disease gait analysis diagnosis. Therefore, it is crucial to find a more suitable deep learning network to solve the problems of automatic feature extraction, multi-dimensional data fusion and computational complexity. SUMMARY

[0004] The application aims to solve the problems of low efficiency and strong subjectivity of traditional Parkinson's disease diagnosis methods, and provides a Parkinson's disease gait analysis diagnosis method based on deep learning to improve the accuracy and objectivity of Parkinson's disease diagnosis.

[0005] The application provides a Parkinson's disease gait analysis diagnosis method based on deep learning, and the specific steps of the analysis diagnosis method are as follows:

[0006] Obtain multi-modal time series gait data of Parkinson's disease, preprocess the multi-modal time series gait data and divide the data set to obtain preprocessed multi-modal time series gait data; the multi-modal time series gait data at least includes foot pressure data, lower limb movement acceleration and lower limb swing angular velocity;

[0007] Design a multi-channel convolutional neural feature extraction network structure to extract features from the multi-modal time series gait data and obtain useful features from local to global;

[0008] A Transformer feature fusion module is established, and a multi-head self-attention mechanism is introduced to process all input useful feature data in parallel to effectively capture and depict the spatial relationship of the features.

[0009] A softmax classification diagnosis module is proposed to classify and diagnose Parkinson's disease gait based on the fusion features.

[0010] Preferably, the multi-modal time series gait data of Parkinson's disease is obtained, preprocessed and divided into data sets to obtain preprocessed multi-modal time series gait data, and the specific steps are as follows:

[0011] Parkinson's gait data is obtained through a wearable device, wherein the wearable device includes an accelerometer, a gyroscope and a pressure sensor, the pressure sensor is arranged on the shoe sole for collecting foot pressure data, the accelerometer is arranged on the lower limb for collecting lower limb movement acceleration, and the gyroscope is arranged on the lower limb for collecting lower limb swing angular velocity.

[0012] The time series gait data of each modality is intercepted, and the time series gait data in the preset starting time period and the preset ending time period is removed, and the jitter of the time series gait data is reduced through median filtering;

[0013] The multi-modal time series gait data is normalized and scaled to the same range to avoid data imbalance;

[0014] The normalized multi-modal time series gait data is intercepted by sliding window to obtain a fixed length sample segment and is divided into three parts: training set, validation set and test set.

[0015] Preferably, the feature extraction network structure of the multi-channel convolutional neural network is designed to extract features from the multi-modal time series gait data, obtain useful features from local to global, and make the following adjustments:

[0016] The windowed multi-channel time series data i∈R C×L is taken as input, where C is the input data channel number and feature dimension, L is the time series window width, and R is the data value type, i.e. all data points are real numbers;

[0017] The multi-channel convolutional neural network includes three CNN heads with different convolution kernels; each CNN head includes three convolution layers and two max pooling layers; the kernel sizes of 3, 6 and 9 are selected in the convolution layers of the three convolution heads; the first convolution layer is used to reduce the input dimension over time while increasing the feature dimension, the stride of all convolution layers is set to 1, Dropout and regularization techniques are used after each convolution layer to limit the complexity of the model, reduce overfitting and improve the generalization ability of the model;

[0018] And after the second convolutional layer and the third convolutional layer, there is a ReLU activation function and a batch normalization layer, the ReLU activation function is used to increase the ability of the model to deal with nonlinear problems, the input value is y, and the calculation formula of the activation function f is as follows:

[0019]

[0020] Preferably, the Transformer feature fusion module is established, the multi-head self-attention mechanism is introduced to process all input useful feature data in parallel, and the specific steps are as follows:

[0021] The position encoding is introduced, the position encoding is calculated by applying different frequencies of sine and cosine functions to the positions of the input sequence, and the formula is as follows:

[0022]

[0023] In the formula, PE is the position vector, d model represents the dimension size of the input vector, pos is the position index of the input vector with dimension d, and i is the index of the current position vector encoding;

[0024] The PE of the pos+k position is:

[0025] PE(pos+k,2i)=PE(pos,2i)×PE(k,2i+1)+PE(pos,2i+1)×PE(k,2i)

[0026] PE(pos+k,2i+1)=PE(pos,2i+1)×PE(k,2i+1)-PE(pos,2i)×PE(k,2i)

[0027] Then, the attention-based encoder structure is constructed to discover feature and data correlation through the spatial distribution between multiple sensors.

[0028] The multi-head self-attention mechanism module is constructed to extract feature and data correlation of different sizes.

[0029] Preferably, the softmax classification diagnosis module is a multi-class decision module composed of a full connection layer and a softmax function.

[0030] Preferably, the Adam optimizer is used to optimize the parameters of the feature extraction network, the Transformer feature fusion module and the softmax classification diagnosis module, which specifically includes:

[0031] The Adam optimizer module is constructed and adjusted as follows:

[0032] First, the exponential moving average of the first and second moments of the gradient is calculated:

[0033] m t = β1·m t-1 + (1-β1)·g

[0034] v t = β2·v t-1 + (1-β2)·g 2

[0035] In the formula, m t and v t represent the first and second moments of the gradient, g represents the current gradient, t represents the current iteration number, and β1 and β2 are adjustable exponential decay rates;

[0036] Bias correction is introduced to compensate for the initialization bias of the first and second moments, and the calculation formula is as follows:

[0037]

[0038] Then, according to the first moment estimate variable and the second moment estimate variable after bias correction, the model parameters are updated:

[0039]

[0040] In the formula, ω represents the parameters of the model, α is the learning rate, and ε is a very small number to avoid division by zero error.

[0041] Compared with the existing Parkinson's disease diagnosis technology, the present application has the following beneficial effects:

[0042] The present application proposes a Parkinson's disease gait analysis and diagnosis method based on deep learning, which solves the problems of relying on subjective evaluation and lacking quantitative gait features in traditional diagnosis methods. By designing an advanced neural network architecture, the present application can automatically extract key features from gait data and use the powerful capabilities of deep learning for feature learning and pattern recognition. In addition, the present application introduces multi-modal data fusion technology, integrates gait information from different sensors, and enhances the accuracy and robustness of diagnosis. Further improve the recognition ability of the model to the gait features of Parkinson's disease, thereby improving the accuracy and reliability of the diagnosis. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 is a flowchart of the Parkinson's disease gait analysis and diagnosis method based on deep learning of the present application;

[0044] Figure 2 is a logic diagram of the Parkinson's disease gait analysis and diagnosis method based on deep learning of the present application;

[0045] Figure 3 Structure diagram of the multi-channel convolutional neural network feature extraction module of the Parkinson's disease gait analysis diagnosis method based on deep learning of the present application;

[0046] Figure 4 Structure diagram of the attention-based encoder module of the Parkinson's disease gait analysis diagnosis method based on deep learning of the present application;

[0047] Figure 5 Structure diagram of the multi-head self-attention module of the Parkinson's disease gait analysis diagnosis method based on deep learning of the present application;

[0048] Figure 6 Accuracy and loss curve diagram of the training data set and test data set of the Parkinson's disease gait analysis diagnosis method based on deep learning of the present application. DETAILED DESCRIPTION

[0049] In order to make the person in the art better understand the present application, the present application will be further explained below in combination with specific embodiments, which do not limit the present application in any form. Moreover, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by the person of ordinary skill in the art without making creative efforts belong to the scope of protection of the present application.

[0050] Please refer to the accompanying drawings, the specific implementation steps of the Parkinson's disease gait analysis diagnosis method based on deep learning are as follows:

[0051] 1. Obtain multi-modal time series gait data of Parkinson's disease, pre-process the multi-modal time series gait data and divide the data set to obtain pre-processed multi-modal time series gait data; the multi-modal time series gait data at least includes foot pressure data, lower limb motion acceleration and lower limb swing angular velocity.

[0052] The specific steps are as follows:

[0053] Obtain Parkinson's gait data through a wearable device, wherein the wearable device includes an accelerometer, a gyroscope and a pressure sensor, the pressure sensor is arranged on the shoe sole for collecting foot pressure data; the accelerometer is arranged on the lower limb for collecting lower limb motion acceleration; the gyroscope is arranged on the lower limb for collecting lower limb swing angular velocity;

[0054] Intercept the time series gait data of each modality, remove the time series gait data in the preset starting time period and the preset ending time period, and reduce the jitter of the time series gait data through median filtering; for example, remove the data of the first 10 seconds and the last 10 seconds;

[0055] The multi-modal time sequence gait data is normalized and scaled to the same range to avoid data imbalance.

[0056] The normalized multi-modal time sequence gait data is windowed to obtain a fixed-length sample segment and is divided into three parts: a training set, a validation set and a test set.

[0057] Specifically, the continuous multi-modal time sequence gait data is divided into a fixed-length sample segment by windowing, and the window size is 240 and the step length is 110.

[0058] 2. A multi-channel convolutional neural network feature extraction network structure is designed to extract features from the multi-modal time sequence gait data and obtain useful features from local to global.

[0059] The specific steps are as follows:

[0060] The windowed multi-channel time series data i∈R C×L is taken as input, where C is the input data channel number and feature dimension, L is the time series window width, and R is the data numerical type, i.e. all data points are real numbers;

[0061] The multi-channel convolutional neural network includes three CNN heads with different convolution kernels, and the convolution kernels are 3, 6 and 9; preferably, the convolution kernel with a convolution kernel of 3 is used to extract local features of foot pressure data; the convolution kernel with a convolution kernel of 6 is used to extract local features of lower limb movement acceleration and lower limb swing angular velocity; and the convolution kernel with a convolution kernel of 9 is used to extract global features of foot pressure data, lower limb movement acceleration and lower limb swing angular velocity.

[0062] Each CNN head includes three convolution layers and two max-pooling layers; wherein the first convolution layer is used to reduce the input dimension over time while increasing the feature dimension, the stride of all convolution layers is set to 1, and Dropout and regularization techniques are used after each convolution layer to limit the complexity of the model, reduce overfitting and improve the generalization ability of the model; wherein the Dropout rate range is set to 0.2 to 0.4. In addition to the last convolution layer, there is a ReLU activation function and a batch normalization layer after each convolution layer, the ReLU activation function is used to increase the ability of the model to handle nonlinear problems, the input value is y, and the calculation formula of the activation function f is as follows:

[0063]

[0064] 3. A Transformer feature fusion module is established, a multi-head self-attention mechanism is introduced to process all input useful feature data in parallel, and the spatial relationship of the features is effectively captured and described.

[0065] The specific steps are as follows:

[0066] The position encoding is introduced by applying different frequencies of sine and cosine functions to the position of the input sequence to calculate the position encoding;

[0067] The sine and cosine functions have periodicity, which can well represent the relative position of the data in the sequence, and the calculation formula is as follows:

[0068]

[0069] In the formula, PE is the position vector, d model represents the dimension size of the input vector, pos is the position index of the input vector with dimension d, and i is the index of the current position vector encoding;

[0070] sin(α+β)=sinαcosβ+cosαsinβ

[0071] cos(α+β)=cosαcosβ-sinαsinβ

[0072] The position encoding is composed of cosine functions with different frequencies, and the PE of pos+k position can be linearly represented by pos, which reflects the relative position relationship. Therefore, the PE of pos+k position is:

[0073] PE(pos+k,2i)=PE(pos,2i)×PE(k,2i+1)+PE(pos,2i+1)×PE(k,2i)

[0074] PE(pos+k,2i+1)=PE(pos,2i+1)×PE(k,2i+1)-PE(pos,2i)×PE(k,2i)

[0075] The position vector of pos+k can be represented as a linear combination of the 2i and 2i+1 dimensions of the position vectors of pos and k positions, and such linear combination means that the relative position information is contained in the position vector;

[0076] In order to enable the Transformer to effectively process sequence-to-sequence tasks, the encoder structure is constructed next.

[0077] The encoder is composed of multiple blocks, each block contains a multi-head self-attention and a feed-forward neural network, and a residual connection and normalization module are added in each sub-layer;

[0078] The residual connection and normalization module is composed of residual connection and normalization, and the calculation formula is as follows:

[0079] LayerNorm(X+Multihead(X))

[0080] LayerNorm(X+FeedForward(X))

[0081] Where X represents the input of the multi-head self-attention module or feedforward neural network module;

[0082] The feedforward neural network layer is a two-layer neural network that first linearly transforms the input Z into a higher dimension, then filters it through a nonlinear ReLU activation function, and finally linearly transforms it back to the original dimension.

[0083] In order to extract features of different sizes and data correlation, a multi-head self-attention mechanism module is constructed next.

[0084] The key part of the multi-head self-attention module is the scaled dot product attention. In each scaled dot product attention head, the embedding is mapped to three new tensors, namely query (Q), key (K), and value (V);

[0085] The attention score is calculated according to the standard model, and the output dimension is d. In order to avoid the information loss caused by the dot product being too large or too small, we introduce The scaling factor controls the value within a reasonable range and is normalized by the softmax activation function;

[0086] Each set of attention is used to map the input to a different sub-representation space, which allows the model to focus on different positions in different sub-representation spaces. The entire calculation process can be expressed as:

[0087] [Q,K,V]=F×[W q ,W k ,W v ]

[0088]

[0089] Multihead(Q,K,V)=Concat(head1,L,head h )W O

[0090] head i =Attention(Q i ,K i ,V i )

[0091] Where, and is the trainable parameter matrix, d k is the dimension of the key vector.

[0092] The multi-head attention module divides Q, K and V into h parts, where h represents the number of heads, and performs self-attention calculation in parallel within each head;

[0093] Subsequently, the outputs of the h heads are connected to the input of the feedforward neural network, and data of the same size is outputted;

[0094] Ensure that the output is consistent with the input in dimension, quickly focus on key features, and perform feature extraction and correlation analysis from different angles.

[0095] 4. A softmax classification diagnosis module is proposed, which is based on fusion features for classification and diagnosis of Parkinson's disease gait.

[0096] The specific steps are as follows:

[0097] The classifier consists of a fully connected layer and a softmax function to form a multi-class decision module. Adding a softmax layer on the fully connected layer is equivalent to a range constraint;

[0098] Given the severity class C, the normalization term Ensure that the sum of all probabilities of the function is 1, and the probability of each class sample is calculated by the softmax function σ, whose calculation formula is as follows:

[0099]

[0100] The performance indicators of the classifier, such as accuracy, precision, recall and F1-score, are used to evaluate the classification accuracy of the task.

[0101] It can be understood that, in order to avoid overfitting of data, the embodiment also uses the Adam optimizer to optimize the parameters of the feature extraction network, the Transformer feature fusion module and the softmax classification diagnosis module, specifically including:

[0102] First, construct the Adam optimizer module:

[0103] Calculate the exponential moving average of the first and second moments of the gradient:

[0104] m t =β1·m t-1 +(1-β1)·g

[0105] v t =β2·v t-1 +(1-β2)·g 2

[0106] In the formula, m t and v trespectively, g represents the current gradient, t represents the current iteration number, and β1 and β2 are adjustable exponential decay rates, which are set as β1 = 0.9 and β2 = 0.999;

[0107] the first moment of the gradient m t and the second moment v t Initialization is started from 0, which leads to underestimation in the initial stage, and a bias correction is introduced to compensate for the initialization bias, and the calculation formula is as follows:

[0108]

[0109] Then, the first moment estimation variable and the second moment estimation variable Update the model parameters:

[0110]

[0111] In the formula, ω represents the parameters of the model, α is the learning rate, and ε is a minimum value to avoid division by zero error. The Adam algorithm adaptively adjusts the learning rate, and adjusts the learning rate according to the size and stability of the gradient during the training process.

[0112] In some embodiments, the Parkinson's diagnosis result is a binary classification result. For example, the Parkinson's diagnosis result indicates normal or suffering from Parkinson's disease. In some embodiments, the Parkinson's diagnosis result is a multi-classification result, for example, the Parkinson's diagnosis result indicates normal or Parkinson's disease severity.

[0113] In specific use, according to the data of the collected sample and the classification model, it is determined whether it is a binary classification result or a multi-classification result. The embodiment selected is a multi-classification result. According to the Unified Parkinson's Disease Rating Scale (UPDRS), the severity is defined as 0 to 4.

[0114] 0 indicates normal; 1 indicates slow walking, small step or tendency to drag; 2 indicates difficulty walking, but still needs help, has some degree of panic or small steps; 3 indicates difficulty walking, severe abnormal gait, needs help; 4 indicates that walking cannot be assisted, so 4 is not classified.

[0115] The following describes an embodiment of the application. In this embodiment, data of an accelerometer, a gyroscope and a pressure sensor are collected, 10-meter turn-around walk and timed up-and-go tests are designed to collect necessary gait data, and gait analysis is performed. Among them, 10-meter turn-around walk refers to walking 10 meters straight, and then returning; the timed up-and-go test refers to standing up from a chair, walking 3 meters, turning back and sitting down.

[0116] Different parameters are selected to train the model respectively, the exponential decay rate is set to β1=0.9 and β2=0.999, and the Dropout rate range is set to 0.2 to 0.4. By comparing the diagnostic performance of different parameter models, the model with the optimal performance is selected as the diagnostic model. Among them, the performance evaluation relies on performance evaluation indexes such as confusion matrix, accuracy, precision, loss function and ROC curve.

[0117] In order to verify the diagnostic performance of the method of the application, the model proposed in the application is compared with the benchmark model on the data set, and the diagnostic result is obtained.

[0118] Method Accuracy (%) Precision (%) Recall (%) F1-score (%) CNN 93.96 93.97 93.91 93.94 Transformer 94.83 94.83 94.63 94.78 The model proposed by the present application 97.85 95.94 95.61 95.69

[0119] As can be seen from the data in the table, the accuracy of the model proposed in the application on the data set is increased by 3.89% and 3.02% compared with the two benchmark models respectively, in addition, the precision, recall and F1-score of the method of the application are also higher than those of other models, which shows the effectiveness and advancement of the model proposed in the application. In summary, the deep learning-based Parkinson's disease gait analysis and diagnosis method proposed in the application can effectively improve the accuracy and reliability of Parkinson's diagnosis.

[0120] The above-described embodiments are only used to illustrate the technical solutions of the present application, but not limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1.A deep learning-based diagnosis method for Parkinson's disease gait analysis, characterized in that, Specifically comprising the following steps: Obtaining multi-modal time sequence gait data of Parkinson's disease, preprocessing the multi-modal time sequence gait data and dividing the data set to obtain preprocessed multi-modal time sequence gait data; the multi-modal time sequence gait data at least includes foot pressure data, lower limb motion acceleration and lower limb swing angular velocity; Designing a multi-channel convolutional neural feature extraction network structure, extracting features from the multi-modal time sequence gait data, and obtaining useful features from local to global; Establishing a Transformer feature fusion module, introducing a multi-head self-attention mechanism to parallelly process all input useful feature data to effectively capture and depict the spatial relationship of the features; Proposing a softmax classification diagnosis module based on the fusion features to classify and diagnose the Parkinson's disease gait. 2.The deep learning-based Parkinson's disease gait analysis diagnosis method of claim 1, wherein, The specific steps of obtaining the multi-modal time sequence gait data of Parkinson's disease, preprocessing the multi-modal time sequence gait data and dividing the data set to obtain the preprocessed multi-modal time sequence gait data are as follows: Obtaining Parkinson's gait data through a wearable device, wherein the wearable device includes an accelerometer, a gyroscope and a pressure sensor, the pressure sensor is arranged on the shoe sole for collecting foot pressure data; the accelerometer is arranged on the lower limb for collecting lower limb motion acceleration; and the gyroscope is arranged on the lower limb for collecting lower limb swing angular velocity; Intercepting the time sequence gait data of each mode, removing the time sequence gait data in the preset starting time period and the preset ending time period, and reducing the jitter of the time sequence gait data through median filtering; Normalizing the multi-modal time sequence gait data and scaling to the same range to avoid data imbalance; Sliding window interception is performed on the normalized multi-modal time sequence gait data to obtain fixed-length sample segments and divide them into three parts: training set, validation set and test set. 3.The deep learning-based Parkinson's disease gait analysis diagnosis method according to claim 1 or 2, characterized in that, The multi-channel convolutional neural feature extraction network structure is designed to extract features from the multi-modal time sequence gait data and obtain useful features from local to global, and the following adjustments are made: Windowed multi-channel time series data i e R C×L as input, where C is the number of input data channels and feature dimensions, L is the time series window width, and R is the numerical type of data, i.e., all data points are real numbers; The multi-channel convolutional neural network includes three CNN heads with different convolution kernels; each CNN head includes three convolution layers and two maximum pooling layers; the first convolution layer is used to reduce the input dimension and increase the feature dimension over time, the stride of all convolution layers is set to 1, Dropout and regularization techniques are used after each convolution layer to limit the complexity of the model, reduce overfitting and improve the generalization ability of the model; And a ReLU activation function and a batch normalization layer are arranged after the second convolution layer and the third convolution layer, the ReLU activation function is used to increase the ability of the model to handle nonlinear problems, the input value is y, and the calculation formula of the activation function f is as follows: 4.The deep learning-based Parkinson's disease gait analysis diagnosis method of claim 3, wherein, The Transformer feature fusion module is established, and the multi-head self-attention mechanism is introduced to parallelly process all input useful feature data, and the specific steps are as follows: Position encoding is introduced, which is calculated by applying different frequencies of sine and cosine functions to the positions of the input sequence, and the calculation formula is as follows: In the formula, PE is a position vector, d model represents the dimension size of the input vector, pos is the position index of the input vector with dimension d, and i is the index of the current position vector encoding. The PE of pos+k position 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) For the position vector of pos+k, it can be expressed as a linear combination of the 2i and 2i+1 dimensions of the position vector of pos position and k position, and such linear combination means that the relative position information is contained in the position vector; Then, an attention-based encoder structure is constructed to discover feature and data correlation through the spatial distribution between multiple sensors; A multi-head self-attention mechanism module is constructed to extract feature and data correlation of different sizes. 5.The deep learning-based Parkinson's disease gait analysis diagnosis method of claim 4, wherein, The softmax classification diagnosis module is a multi-class decision module composed of a full connection layer and a softmax function. 6.The deep learning-based diagnosis method of Parkinson's disease gait analysis according to claim 5, wherein, The parameters of the feature extraction network, the Transformer feature fusion module and the softmax classification diagnosis module are optimized by using the Adam optimizer, which specifically includes: An Adam optimizer module is constructed and adjusted as follows: First, calculate the exponential moving average of the first and second moments of the gradient: m t = β1· m t-1 + (1 - β1)· g v t = β2· v t-1 + (1 - β2)· g 2 In the formula, m t and v t respectively represent the first and second moments of the gradient, g represents the current gradient, t represents the current iteration number, and β1 and β2 are adjustable exponential decay rates. Bias correction is introduced to compensate for the initialization bias of the first and second moments, and the calculation formula is as follows: Then, the first moment estimation variable after bias correction and the second moment estimation variable Update model parameters: In the formula, ω represents the parameters of the model, α is the learning rate, and ε is a very small value to avoid division by zero error.

Images