Parkinson's disease risk prediction method based on deep learning, electronic equipment and medium

Abstract

The invention provides a Parkinson's disease risk prediction method based on deep learning, an electronic device and a medium, and the method comprises the steps: integrating medical data from different sources, extracting visual features from a hand-drawn spiral line image through a CNN, capturing the dynamic changes in motion and voice from motion sensing data and voice data through an LSTM, and carrying out the prediction of the Parkinson's disease risk. Various data sources can be comprehensively analyzed; the importance of different source features is weighted through an attention mechanism, so that more dynamic and adaptive feature fusion is realized; according to the method, the complex relationship and interaction among different data types can be captured, subtle and complex pathological characteristics of the Parkinson's disease are identified, and the accuracy and efficiency of risk prediction of the Parkinson's disease are improved.

Description

Technical Field

[0001] The present invention belongs to the field of medical information processing and statistical analysis, and in particular relates to a Parkinson's disease risk prediction method based on deep learning, an electronic device and a medium. Background Art

[0002] Parkinson's disease (PD) is a gradually progressive neurodegenerative disease. With the aging of the population, the number of Parkinson's disease patients is on the rise, posing a major challenge to both society and the medical system.

[0003] The main symptoms of Parkinson's disease include resting tremor, muscle rigidity, bradykinesia and postural gait disorders. These symptoms are mainly caused by the decrease of dopamine levels in the brain and pathological changes of substantia nigra cells. Although scientific research has not yet clarified the exact cause of Parkinson's disease, the hypothesis that it is related to genetic, environmental and other factors is widely accepted. In addition, although Parkinson's disease cannot be cured at present, early prediction and timely medical intervention can significantly improve the quality of life of patients.

[0004] In the prediction, handwriting and handwriting tests are considered to be very effective auxiliary risk prediction tools in the early stage of Parkinson's disease. These tests are simple, low-cost, and can quickly reflect the patient's motor dysfunction. The handwriting movements of Parkinson's patients are usually slower than those of healthy people, and their handwriting is trembling and their motor control ability is weakened, which can be identified by simple handwriting analysis.

[0005] However, although handwriting analysis provides a feasible method for risk prediction of Parkinson's disease, there are still many challenges. For example, the clinical manifestations of Parkinson's disease in the early stages may be confused with other movement disorders such as essential tremor (ET), which makes accurate prediction complicated. In addition, traditional handwriting analysis requires doctors or trained professionals to interpret the results, which is not only inefficient but also may be affected by subjective factors.

[0006] The paper A Lightweight Parkinson's Disease Diagnostic Method Based on Hand-Drawn Spiral Lines (Wang L, Liu H, Chen X, et al. A Lightweight Parkinson's Disease Diagnostic Method Based on Hand-Drawn Spiral Lines [C] / / Proceedings of the 2024 4th International Conference on Bioinformatics and Intelligent Computing. New York, NY, USA: Association for Computing Machinery, 2024: 217-220.) only used a single data source in a single scenario. The model has the problem of weak generalization ability, which limits the depth and accuracy of Parkinson's disease risk prediction. Summary of the invention

[0007] The purpose of the present invention is to address the deficiencies in the existing technology and provide a Parkinson's disease prediction method, electronic device and medium based on deep learning, which improves the accuracy and efficiency of Parkinson's disease risk prediction by fusing and analyzing multi-source medical data.

[0008] In order to achieve the above object, the technical solution adopted by the present invention is:

[0009] A Parkinson's disease risk prediction method based on deep learning, comprising the following steps:

[0010] S1: Acquire multi-source data of a Parkinson's disease patient, wherein the multi-source data includes a hand-drawn spiral image, motion sensor data, and voice data;

[0011] S2: preprocessing the multi-source data;

[0012] S3: using the multi-source data processed by S2 as the input of a deep learning network model to train the deep learning network model, wherein the deep learning network model includes a CNN network, an LSTM network, an attention mechanism layer, and a fully connected layer, wherein the CNN network and the LSTM network are respectively connected to the attention mechanism layer, and the attention mechanism layer is connected to the fully connected layer;

[0013] The training process of the deep learning network model includes:

[0014] The CNN network is used to process the preprocessed hand-drawn spiral image to extract visual features;

[0015] The LSTM network is used to process the pre-processed motion sensor data and speech data to extract time series features;

[0016] Concatenate the visual features and time series features to obtain the concatenated features F concat ;

[0017] Through the attention mechanism, the concatenated feature F concat Perform weighted processing to obtain the weighted feature F attention =α·(F concat ·W v );

[0018] Use the fully connected weighted feature F attention Processing to obtain the final feature F fcn =W fcn ·F attention +b fcn ;

[0019] Iteratively optimize and train the deep learning network model using a cross entropy loss function to obtain a Parkinson's disease risk prediction model;

[0020] Among them, W v is the weight of the value vector, α is the attention weight, W fcn is the weight of the fully connected layer, b fcn is the bias of the fully connected layer.

[0021] The Parkinson's disease risk prediction model is used to predict Parkinson's disease.

[0022] The present invention integrates medical data from different sources, extracts visual features from hand-drawn spiral images through CNN, and captures dynamic changes in motion and speech from motion sensor data and speech data through LSTM, so as to comprehensively analyze multiple data sources; the importance of features from different sources is weighted through the attention mechanism, so as to achieve more dynamic and adaptive feature fusion; the present invention can capture the complex relationships and interactions between different data types, identify the subtle and complex pathological features of Parkinson's disease, and improve the accuracy and efficiency of Parkinson's disease risk prediction.

[0023] Furthermore, the implementation process of S2 includes:

[0024] The hand-painted spiral image is sequentially subjected to resolution adjustment, normalization processing, contrast enhancement and sharpening processing;

[0025] Performing filtering and standardization processing on the motion sensing data in sequence;

[0026] The speech data is analyzed using Mel frequency cepstral coefficients, speech features are extracted, and the speech features are filtered and standardized.

[0027] By preprocessing the hand-drawn spiral images, the image quality is improved and converted into a format for CNN network processing; by preprocessing the motion sensor data, the balance of the data in different dimensions is ensured, the sensitivity of the model to specific scales is reduced, and the availability and accuracy of the data are improved; the speech data is preprocessed to identify speech pattern changes associated with Parkinson's disease.

[0028] Furthermore, the CNN network includes a plurality of cascaded convolutional networks, and the convolutional network includes a convolutional layer and a pooling layer; the hand-drawn spiral image is processed by convolution and pooling to extract visual features; the expression of the convolution processing is as follows:

[0029] Among them, F ij represents the output feature value of the jth feature of the i-th layer, W i represents the weight of the i-th layer, X j is the jth feature, b i is the bias of the i-th layer, * represents the convolution operation, and ReLU is the activation function.

[0030] CNN can extract key visual features from hand-drawn spiral images.

[0031] Furthermore, the LSTM network includes an input gate, a forget gate, an output gate, and a candidate cell state unit;

[0032] The implementation process of using LSTM network to process pre-processed motion sensor data and voice data includes:

[0033] The preprocessed motion sensor data and speech data are used as the input of the forget gate, input gate, output gate and candidate cell state unit, respectively, and the activation vector f of the forget gate is obtained. t , the activation vector i of the input gate t , the activation vector o of the output gate t , candidate values ​​for the new cell state

[0034] f t =σ(W f ·[h t-1 , x t ]+b f )

[0035] i t =(σ(W k ·[h t-1 ,x t ]+b k )

[0036]

[0037] Update cell status C t , C t The expression is as follows:

[0038] Extract time series features h t =o t *tanh(C t );

[0039] Among them, C t , C t-1 are the cell states at the current moment and the previous moment, respectively, h t ,h t-1 are the hidden states of the current moment and the previous moment, σ is the sigmoid activation function, tanh is the hyperbolic tangent activation function, x t is the motion sensor data and voice data at the current moment, W f , W k , W o , W c are the weights of the forget gate, input gate, output gate and candidate cell state unit, b f 、b k 、b o 、b c They are the biases of the forget gate, input gate, output gate and candidate cell state unit respectively.

[0040] LSTM is used to process motion sensor data and speech samples with obvious time series characteristics. Through its internal gating mechanism, LSTM can capture the temporal dependencies and long-term patterns in the data and maintain the continuous transmission of information over time.

[0041] Based on the same inventive concept, the present invention also provides an electronic device, including:

[0042] one or more processors;

[0043] A memory having one or more programs stored thereon, which, when executed by the one or more processors, enables the one or more processors to implement the steps of a Parkinson's disease risk prediction method based on deep learning.

[0044] Based on the same inventive concept, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of a Parkinson's disease risk prediction method based on deep learning.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] The present invention integrates medical data from different sources, extracts visual features from hand-drawn spiral images through CNN, and captures dynamic changes in motion and speech from motion sensor data and speech data through LSTM, so as to comprehensively analyze multiple data sources; the importance of features from different sources is weighted through the attention mechanism, so as to achieve more dynamic and adaptive feature fusion; the present invention can capture the complex relationships and interactions between different data types, identify the subtle and complex pathological features of Parkinson's disease, and improve the accuracy and efficiency of Parkinson's disease risk prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 Schematic diagram of the Parkinson's disease risk prediction method based on deep learning of the present invention. DETAILED DESCRIPTION

[0048] The present invention will be described in detail below in conjunction with the embodiments. It should be noted that the embodiments and features of the embodiments of the present invention can be combined with each other without conflict. For the convenience of description, if the words "upper", "lower", "left" and "right" appear in the following, they only indicate that the upper, lower, left and right directions are consistent with the drawings themselves, and do not limit the structure.

[0049] Example

[0050] like Figure 1 This embodiment proposes a Parkinson's disease risk prediction system based on deep learning, which integrates a convolutional neural network (CNN) and a long short-term memory network (LSTM) to analyze multimodal data (including images, motion sensor data, and voice samples), including the following steps:

[0051] Step 1: Data Collection

[0052] The Parkinson's disease risk prediction system data in this embodiment mainly includes hand-drawn spiral image data, motion data, and voice samples.

[0053] A1: Image data

[0054] The hand-drawn spiral images were stored in a high-resolution format so that subsequent processing could analyze subtle features such as tremors and line breaks in the image. These data were used to analyze the correlation between image features and Parkinson's disease progression.

[0055] A2: Sports data

[0056] The motion data includes detailed records of limb movements, such as acceleration and gyroscope readings, which have been sampled at high frequency and signal processed. By using this data, time series analysis and pattern recognition can be performed to assess the extent of motor dysfunction.

[0057] A3: Voice Sample

[0058] The speech samples included recordings of Parkinson's disease patients speaking freely and reading specific texts in a controlled environment. These data have been transcribed and analyzed to provide key speech features, such as changes in pitch and speech rate, for monitoring changes in language expression.

[0059] Step 2: Data Preprocessing

[0060] Data preprocessing is a key step to ensure the accuracy and efficiency of subsequent analysis. In this embodiment, images, motion sensor data, and voice samples are all subjected to a special preprocessing process to ensure data quality and consistency and optimize the subsequent deep learning processing effect.

[0061] B1: Image data preprocessing

[0062] The main purpose of image data preprocessing is to improve image quality and convert it into a format suitable for convolutional neural network (CNN) processing.

[0063] Image resizing: All collected images are first resized to a uniform resolution, such as 256x256 pixels. This step is done using bilinear interpolation to ensure maximum retention of image details during the scaling process.

[0064] Normalization: The image pixel values ​​are normalized to the interval [0,1]. This operation can speed up the convergence of the neural network and improve the numerical stability during training. The normalization formula is

[0065]

[0066] Among them, pixel is the pixel value of the image, min and max are the minimum and maximum values ​​of all pixels in the image, respectively, and normalized is the normalized pixel.

[0067] Image enhancement: including contrast enhancement and sharpening to highlight key features in the image, such as edges and textures. Contrast enhancement is achieved by adjusting the histogram of the image, while sharpening enhances the high-frequency part of the image by applying a high-pass filter.

[0068] B2: Motion sensor data preprocessing

[0069] The preprocessing of motion sensor data aims to improve the availability and accuracy of data and provide high-quality input for LSTM network analysis.

[0070] Filtering: Apply a low-pass filter, such as a Butterworth filter, to remove high-frequency noise. The filtering formula can be expressed as:

[0071]

[0072] Where y(t) is the output of the filter at time t, and w s is the cutoff frequency and k is the filter order.

[0073] Data standardization: All sports data are standardized according to Standardization is performed, where x1 is the original data, μ and σ1 are the mean and standard deviation of the data respectively. This step ensures the balance of the data in different dimensions and reduces the sensitivity of the model to a specific scale.

[0074] B3: Voice Data Preprocessing

[0075] Feature extraction: In this embodiment, the primary task of speech data preprocessing is to extract key acoustic features useful for Parkinson's disease risk prediction. To this end, Mel frequency cepstral coefficients (MFCC) are used to capture the envelope of the short-time power spectrum of the speech signal, because MFCC features can effectively represent the basic properties of speech signals and are widely used in speech recognition and related health status analysis.

[0076] MFCC calculation process:

[0077] 1. Pre-emphasis: Pre-emphasize the original speech signal to enhance the high-frequency part. The formula is

[0078] s'(n1)=s(n1)-α1×s(n1-1)

[0079] Among them, s'(n1) is the processed signal value of n1 at the current moment, s(n1) is the original signal value of n1 at the current moment, α1 is the smoothing factor, which ranges from 0.9 to 1, and s(n1-1) is the signal value of n1-1 at the previous moment.

[0080] 2. Framing and window function: The speech signal is divided into short frames, each frame is about 2040 milliseconds, and there is a certain overlap between frames. The Hanning window is applied in each frame to reduce the boundary effect. The formula of the Hanning window is as follows:

[0081]

[0082] Where w(n) is the value of the Hanning window at the nth point, 0≤n≤N-1, and N is the length of the window function.

[0083] 3. Fast Fourier Transform (FFT): Perform FFT on each frame of data and convert it to the frequency domain.

[0084] 4. Mel filter bank: The result of FFT is processed by the Mel filter bank, which imitates the sensitivity of the human ear to sounds of different frequencies and emphasizes important frequency components.

[0085] 5. Logarithmic Energy: Calculate the logarithm of the energy of each filter output.

[0086] 6. Discrete Cosine Transform (DCT): Perform DCT on the logarithmic energy to reduce the correlation between features and obtain the final MFCC.

[0087] After the MFCC features are extracted, further signal cleaning and encoding are required to ensure that the data input into the neural network is high-quality and standardized.

[0088] Signal cleaning: Noise reduction is performed on the MFCC features, especially removing background noise and other interference. This is achieved through the spectral thresholding method, which sets an energy threshold to identify and reduce noise components below the threshold.

[0089] Feature encoding: The cleaned MFCC features are standardized. Specifically, the mean and standard deviation of the MFCC features are calculated to process each feature so that the mean of the data in each feature dimension is 0 and the standard deviation is 1. This makes it have a unified scale, which is convenient for neural network processing.

[0090] Step 3: Feature Extraction

[0091] In this embodiment, feature extraction is achieved through a convolutional neural network (CNN) and a long short-term memory network (LSTM), which are designed for different types of data to ensure that the most useful information is extracted from each type of data.

[0092] C1: Convolutional Neural Networks (CNN) applied to image data

[0093] The application of CNN in the field of image processing has been widely proven to be effective, especially in feature extraction. In this embodiment, CNN is used to automatically extract key visual features, such as shape and texture, from pre-processed images of hand-drawn spiral lines of Parkinson's disease patients.

[0094] Network architecture: CNN consists of a combination of multiple convolutional layers and pooling layers, each of which uses multiple filters to automatically capture local features in the image. As the network level deepens, the extracted features change from simple to complex, and a high-level understanding of the image content can be gradually formed.

[0095] Convolutional layer: In each convolutional layer, the image is convolved with a set of learned kernels (filters) to produce a set of feature maps. This process can be expressed as:

[0096] Among them, F ij represents the output feature value of the jth feature of the i-th layer, W i represents the weight matrix of the i-th layer, X jis the jth feature, b i is the bias term of the i-th layer, * represents the convolution operation, and ReLU is the activation function.

[0097] Pooling layer: The pooling layer follows the convolutional layer to reduce the spatial size of the feature map and improve the invariance of the features. The most common pooling operations are maximum pooling and average pooling.

[0098] C2: Long Short-Term Memory (LSTM) applied to motion and speech data

[0099] LSTM networks are particularly suitable for processing data with time series characteristics, such as motion sensor data and speech data, because they are able to capture temporal dependencies and long-term patterns in the data.

[0100] Core structure: The LSTM unit consists of three gate structures - input gate, forget gate and output gate, which together control the inflow, storage and outflow of information.

[0101] The preprocessed motion sensing data and speech data are passed as input to three gates (forget gate, input gate, output gate) and the calculation module of candidate cell states.

[0102] The forget gate controls the cell state C at the previous moment t-1 Which information needs to be retained to the current cell state C t The formula is:

[0103] f t =σ(W f ·[h t-1 , x t ]+b f )

[0104] The input gate controls which information in the current input information needs to be written into the cell state C at the current moment t The formula is:

[0105] i t =(σ(W k ·]h t-1 ,x t ]+b k )

[0106] The candidate cell state provides new information, and the input gate controls the extent to which it is written into the cell state. The candidate value of the new cell state is:

[0107]

[0108] The output gate controls which information in the current cell state needs to be output to the hidden state at the current moment. The formula is:

[0109] The cell status is updated as follows:

[0110]

[0111] Output:

[0112] h t =o t *tanh(C t )

[0113] Among them, C t , C t-1 are the cell states at the current moment and the previous moment, respectively, h t ,h t-1 are the hidden states of the current moment and the previous moment, σ is the sigmoid activation function, tanh is the hyperbolic tangent activation function, and x t is the motion sensor data and voice data at the current moment, W f , W k , W o , W c are the weights of the forget gate, input gate, output gate and candidate cell state unit, b f 、b k 、b o 、b c They are the biases of the forget gate, input gate, output gate and candidate cell state unit respectively.

[0114] Step 4: Feature fusion and analysis

[0115] In the Parkinson's disease risk prediction system, feature fusion and analysis are key links to ensure that the information extracted from multi-source data can be fully utilized to make accurate risk predictions. Traditional feature fusion methods such as simple concatenation may not be sufficient to capture the complex relationships between different data sources, so an improved fusion strategy is proposed, combined with the design and implementation of a deep learning model.

[0116] D1: Improved data fusion strategy

[0117] In order to more effectively fuse the features extracted by CNN and LSTM, this embodiment adopts a multimodal fusion layer, which not only considers the connection of features, but also introduces an attention mechanism to weight the importance of different source features, thereby achieving more dynamic and adaptive feature fusion.

[0118] Multimodal fusion layer: This layer first simply concatenates the outputs of CNN and LSTM, then re-weights the importance of each feature through the self-attention module, and finally uses the fully connected layer to process the fused features. The self-attention mechanism dynamically adjusts the weight of each feature by calculating the interdependence between features, so that the model can focus on information with greater risk prediction value. The specific process is as follows:

[0119] h cnn is the output feature vector of the last layer of CNN, h lstm It is the last layer output feature vector of LSTM, after feature concatenation:

[0120] F concat =[h cnn ;h lstm ]

[0121] F concat It is the concatenated feature representation.

[0122] Next is the implementation of the self-attention mechanism:

[0123] 1. Calculate the self-attention weight of feature representation:

[0124] α=softmax(F concat ·W q ·(F concat ·W k ) T )

[0125] Among them, W q and W k are the weight matrices of the query vector and key vector respectively.

[0126] 2. Weighting feature representations:

[0127] F attention =α·(F concat ·W v )

[0128] Among them, E attention is the feature representation after weighting by the attention module, W v is the weight matrix of the value vector and α is the attention weight.

[0129] Finally, the fused features are processed using a fully connected layer for final risk prediction:

[0130] F fcn ×W fcn ·F attention +b fcn Among them, F fcn is the output feature of the fully connected layer, Wfcn is the weight of the fully connected layer, b fcn is the bias of the fully connected layer.

[0131] D2: Design and implementation of deep learning models

[0132] Based on the improved fusion strategy, a deep learning model suitable for multimodal data analysis is designed to ensure the effective utilization of features extracted from image, motion, and speech data sources.

[0133] Model Architecture: The basic architecture of the model includes initial convolutional and LSTM layers for extracting visual and sequence features, followed by multimodal fusion layers for further analysis of the features and final classification of pathological status.

[0134] Training process:

[0135] Using an end-to-end training strategy, the entire model is trained from input to output to ensure that all components are optimized in a coordinated manner.

[0136] Cross entropy is used as the loss function to optimize the classification accuracy, while regularization strategy and dropout technology are used to prevent overfitting.

[0137] The Adam optimizer is used because its adaptive learning rate can speed up convergence and improve training efficiency.

[0138] Step 5: Generate prediction results

[0139] Data input: The patient's image, motion, and speech data are directly input into the model after preprocessing.

[0140] Feature analysis and fusion: The system uses a pre-trained deep learning model that combines the output features of a convolutional neural network (CNN) and a long short-term memory network (LSTM) for analysis.

[0141] Output layer: The model output layer uses the softmax function to process the output F of the fully connected layer. fcn , calculate the probability of each category, the formula is as follows:

[0142] Y = softmax(F fcn )

[0143] Among them, Y is the probability of each category, that is, the risk prediction output.

[0144] Incremental learning: The system implements an incremental learning strategy, continuously integrating new data and feedback to update the model, and adjusting the model parameters using the following formula:

[0145]

[0146] Among them, θnew represents the updated parameters of the model, θ old is the current parameter of the model, η is the learning rate, is the gradient of the parameter θ, L is the loss function, and x, y are the new data and their labels.

[0147] This embodiment achieves comprehensive recognition and analysis of Parkinson's disease symptoms by comprehensively using images, motion sensor data and voice samples. This multimodal analysis method overcomes the limitations of traditional reliance on a single data source and improves the comprehensiveness and accuracy of lung nodule risk prediction.

[0148] This embodiment uses convolutional neural networks (CNN) and long short-term memory networks (LSTM) to automatically extract key features and learn the complex relationships between these features. This not only improves the efficiency of feature extraction, but also enables the model to capture deeper pathological information, thereby ensuring the accuracy and reliability of lung nodule risk prediction.

[0149] This embodiment provides an electronic device, including:

[0150] one or more processors;

[0151] A memory having one or more programs stored thereon, which, when executed by one or more processors, enables the one or more processors to implement the steps of a Parkinson's disease risk prediction method based on deep learning.

[0152] In some implementations, the memory may be a high-speed random access memory (RAM), and may also include a non-volatile memory, such as at least one disk memory.

[0153] In some other implementations, the processor may be a central processing unit (CPU), a digital signal processor (DSP), or other general-purpose processors of various types, which are not limited herein.

[0154] This embodiment provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of a method for predicting Parkinson's disease risk based on deep learning.

[0155] The contents explained in the above embodiments should be understood as these embodiments are only used to more clearly illustrate the present invention, and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

Claims

1. A Parkinson's disease risk prediction method based on deep learning, characterized in that: The following steps are involved: S1: Acquire multi-source data of a Parkinson's disease patient, wherein the multi-source data includes a hand-drawn spiral image, motion sensor data, and voice data; S2: preprocessing the multi-source data; S3: using the multi-source data processed by S2 as the input of a deep learning network model to train the deep learning network model, wherein the deep learning network model includes a CNN network, an LSTM network, an attention mechanism layer, and a fully connected layer, wherein the CNN network and the LSTM network are respectively connected to the attention mechanism layer, and the attention mechanism layer is connected to the fully connected layer; The training process of the deep learning network model includes: The CNN network is used to process the preprocessed hand-drawn spiral image to extract visual features; The LSTM network is used to process the pre-processed motion sensor data and speech data to extract time series features; Concatenate the visual features and time series features to obtain the concatenated features F concat ; Through the attention mechanism, the concatenated feature F concat Perform weighted processing to obtain the weighted feature F attention =α·(F concat ·W v ); Use the fully connected weighted feature F attetntion Processing to obtain the final feature F fcn =W fcn ·F attention +b fcn ; Iteratively optimize and train the deep learning network model using a cross entropy loss function to obtain a Parkinson's disease risk prediction model; Among them, W v is the weight of the value vector, α is the attention weight, W fcn is the weight of the fully connected layer, b fcn is the bias of the fully connected layer.

2. The Parkinson's disease risk prediction method based on deep learning according to claim 1, characterized in that: The implementation process of S2 includes: The hand-painted spiral image is sequentially subjected to resolution adjustment, normalization processing, contrast enhancement and sharpening processing; Performing filtering and standardization processing on the motion sensing data in sequence; The speech data is analyzed using Mel frequency cepstral coefficients, speech features are extracted, and the speech features are filtered and standardized.

3. The Parkinson's disease risk prediction method based on deep learning according to claim 1, characterized in that: The CNN network includes multiple cascaded convolutional networks, and the convolutional network includes a convolutional layer and a pooling layer. The hand-drawn spiral image is processed by convolution and pooling to extract visual features. The expression of the convolution processing is as follows: F ij =ReLU(W i *X j +b i ) Among them, F ij represents the output feature value of the jth feature of the i-th layer, W i represents the weight of the i-th layer, X j is the jth feature, b i is the bias of the i-th layer, * represents the convolution operation, and ReLU is the activation function.

4. The Parkinson's disease risk prediction method based on deep learning according to claim 1, characterized in that: The LSTM network includes an input gate, a forget gate, an output gate, and a candidate cell state unit; The implementation process of using LSTM network to process pre-processed motion sensor data and voice data includes: The preprocessed motion sensor data and speech data are used as the input of the forget gate, input gate, output gate and candidate cell state unit, respectively, and the activation vector f of the forget gate is obtained. t , the activation vector i of the input gate t , the activation vector o of the output gate t , candidate values ​​for the new cell state f t =σ(W f ·[h t-1 ,x t ]+b f ) i t =(σ(W k ·[h t-1 ,x t ]+b k ) the t =σ(W o ·[h t-1 ,x t ]+b o ) Update cell status C t , C t The expression is as follows: Extract time series features h t =o t *tanh(C t ); Among them, C t , C t-1 are the cell states at the current moment and the previous moment, respectively, h t ,h t-1 are the hidden states of the current moment and the previous moment, σ is the sigmoid activation function, tanh is the hyperbolic tangent activation function, x t is the motion sensor data and voice data at the current moment, W f , W k , W o , W c are the weights of the forget gate, input gate, output gate and candidate cell state unit, b f , b k , b o , b c They are the biases of the forget gate, input gate, output gate and candidate cell state unit respectively.

5. An electronic device, characterized in that: include: one or more processors; A memory having one or more programs stored thereon, which, when the one or more programs are executed by the one or more processors, enables the one or more processors to implement the steps of the method according to any one of claims 1 to 4.

6. A computer-readable storage medium, characterized in that: The computer program is stored therein, and when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 4 are implemented.

Images