A parkinson's disease early identification system and method based on multi-task learning

Abstract

The application provides a Parkinson disease early identification system and method based on multi-task learning, and belongs to the technical field of machine learning. First, a Parkinson disease discrimination data set covering action, language and visual modalities is constructed to provide comprehensive symptom data and identity for subsequent analysis. Then, in combination with medical prior rules, a symptom recognition module integrating multiple traditional machine learning models is constructed to preliminarily discriminate the action and language data. A short-term recognition model is constructed to enhance the multi-modal disease perception ability and improve the discrimination effect of short-term Parkinson disease recognition. Finally, based on the recognition result sequence of multiple time periods, mathematical modeling analysis is carried out to fuse statistical trends and group differences, and to realize stable comprehensive judgment of Parkinson disease. Through the collaborative training and knowledge sharing among tasks, the model can enhance the comprehensive discrimination ability of multi-modal pathological signals of Parkinson disease, so as to realize more accurate and sensitive early identification.

Description

Technical Field

[0001] This invention belongs to the field of machine learning technology, and in particular relates to an early identification system and method for Parkinson's disease based on multi-task learning. Background Technology

[0002] In the field of Parkinson's disease research, existing identification methods typically focus on a single modality or a single symptom dimension, resulting in a lack of comprehensive understanding of the complex and diverse early manifestations of Parkinson's disease in practical applications. Early clinical manifestations of Parkinson's disease involve multiple sensory channels, including speech, movement, and vision, and there are potential connections between these channels. This places higher demands on the feature fusion and task coordination capabilities of identification models. However, traditional identification methods suffer from task fragmentation and insufficient feature fusion when processing multimodal data, making it difficult to achieve a holistic perception and accurate identification of the disease manifestations. Simultaneously, single-task models struggle to adapt to the dynamic evolutionary relationships between Parkinson's symptoms and fail to capture the deep, underlying interactive information behind multi-source pathological data, leading to low early identification accuracy and poor generalization ability. With the rapid development of artificial intelligence technology, multi-task learning frameworks, with their advantages in multimodal information collaborative modeling, knowledge transfer, and improved generalization ability, have brought new opportunities for the early identification of Parkinson's disease. Multi-task learning can unify the modeling of multiple tasks such as speech analysis, action recognition, and visual perception, uncovering potential connections between symptom manifestations and improving the model's ability to analyze and understand complex pathological patterns.

[0003] Existing methods for commercializing scientific research results include the following:

[0004] (1) Early identification method of Parkinson's disease based on clinical expert rules: Clinical neurologists formulate symptom identification rules for Parkinson's disease based on years of clinical experience, such as typical manifestations like voice tremor, bradygait, and muscle rigidity. These rules are then manually assessed through questionnaires, physical sign scales, and simplified motor function assessments to assist in the initial identification and classification of Parkinson's disease. This method relies on expert knowledge, and the identification process is characterized by strong human intervention and poor adaptability.

[0005] (2) Early identification method of Parkinson's disease based on traditional machine learning: By collecting sensor data such as voice, movement, and electromyography, manually designed features such as MFCC, frequency domain jitter, and movement frequency are extracted, and classification models such as support vector machine, random forest, and K-nearest neighbors are constructed to identify the patient's state. This method depends on the quality of feature design, makes it difficult to fully explore the potential nonlinear correlations between symptoms, and has limited model generalization ability.

[0006] (3) Early identification method of Parkinson's disease based on deep neural networks: Construct an end-to-end deep learning architecture, such as convolutional neural networks, recurrent neural networks or Transformer models, to learn feature representations directly from the original speech or video sequences, and combine symptom labels for supervised training to achieve automated feature extraction and symptom identification. This method can mine deep pathological features in multimodal signals, but there is still room for improvement in terms of inter-task collaborative modeling and data efficiency.

[0007] Current methods for early identification of Parkinson's disease still have many limitations in terms of accuracy and practicality. Methods based on clinical expert rules rely heavily on human experience, lack objective quantitative indicators, and struggle to adapt to diverse individual symptom presentations. Traditional machine learning methods depend on manually constructed features, making it difficult to capture complex nonlinear pathological features and exhibiting poor generalization ability across different scenarios. While deep neural network methods possess end-to-end modeling capabilities, their training typically relies on large amounts of labeled data, and they lack effective collaborative modeling mechanisms between different tasks, making it difficult to fully utilize multi-source information. Overall, current methods still need further improvement in robustness, generalization, and task fusion capabilities. Summary of the Invention

[0008] To address the aforementioned problems, the first aspect of this invention provides an early identification system for Parkinson's disease based on multi-task learning, characterized in that it includes a data acquisition module, a symptom identification module, a short-term identification model for Parkinson's disease, and a comprehensive discrimination module for Parkinson's disease.

[0009] The data acquisition module is used to acquire user voice expression audio data and limb movement sequence data. Specifically, the limb movement sequence data is the limb inertial movement sequence data collected by the accelerometer, which is used as data representing Parkinson's disease symptoms; at the same time, motion image data is acquired.

[0010] The symptom recognition module is based on the user's voice expression audio data and limb movement sequence data collected by the data acquisition module. It constructs a medical prior knowledge random forest model and a medical constraint embedded gradient boosting tree model, and combines them with an adaptive weighted logistic regression model to output the Parkinson's disease symptom recognition results.

[0011] The Parkinson's disease short-term identification model first extracts convolutional features from each frame of the acquired motion image data using a frame-level visual perception unit, and enhances the response of facial muscles and key limb regions by combining an attention mechanism. Second, it models the temporal features of continuous video frames through a cross-frame symptom fusion unit to obtain cross-frame dynamic features. Finally, through a multimodal symptom fusion unit, it uses a multilayer perception network to perform feature calculation processing on the Parkinson's disease symptom prediction results obtained from the symptom identification module, and fuses them with the cross-frame dynamic features to guide the network to focus on the semantic regions of the image related to medical symptoms. Finally, the Parkinson's disease identification head outputs the early identification results of Parkinson's disease.

[0012] The Parkinson's disease comprehensive discrimination module performs multiple disease probability detections using a short-term identification model for Parkinson's disease under long content frames to obtain the Parkinson's disease factor sequence of the subject; then, it performs multi-step mathematical analysis on the factor sequences of normal people and Parkinson's patients to finally obtain the standard identification result of Parkinson's disease.

[0013] Preferably, the process of constructing the dataset used to train the symptom recognition module includes:

[0014] Audio data of speech expression collected using a microphone Limb motion sequence data acquired by deploying accelerometers on the limbs And use it as data for the symptom characteristics of Parkinson's disease. Input data ;

[0015] When processing audio data During data collection, non-invasive electromyography (EMG) sensors are deployed to measure the tension of the laryngeal muscles during vocalization. Surface strain gauge sensors were deployed to collect the force of the throat muscles during vocalization. Deploy an electromagnetic vocalization meter to detect the amplitude of tongue movement. Deploy a small inertial measurement unit to measure the mandibular opening angle. ;

[0016] In processing limb movement sequence data During data acquisition, electromyography (EMG) sensors were deployed to collect the amplitude of EMG signals from the forearm muscle groups. Interval of electromyographic signal bursts Flexible pressure sensors were deployed to collect the peak force of the fingers' gripping operations. Deploy a plantar pressure sensor array to collect peak heel force. ;

[0017] The data collected during the above process Combined as output data of the symptom recognition module .

[0018] Preferably, the symptom recognition module adopts a modular architecture design with separate prediction of base components and meta-components; in the base component, a random forest model incorporating prior medical knowledge is designed to predict Parkinson's disease symptoms and obtain the predicted Parkinson's disease symptoms. A medically constrained embedded gradient boosting tree model was designed to predict Parkinson's disease symptoms, and the prediction results were obtained. ;

[0019] The meta-component uses an adaptive weighted logistic regression model, and this model is used to evaluate the prediction results. and prediction results Scene-adaptive fusion processing is performed, and weights are dynamically assigned based on the predicted features. After the fusion processing is completed, the final prediction result of Parkinson's disease symptoms is obtained. .

[0020] Preferably, the medical prior knowledge random forest model is a multi-objective random forest model architecture, that is, it outputs data based on the symptom representation of Parkinson's disease. Each Parkinson's symptom in the dataset is matched with an independent set of regression trees;

[0021] A symptom coefficient correlation matrix was constructed based on existing medical literature. The random forest model consists of Q groups of random forest trees, each group being independent of the others, and the input to each group of random forest trees is data representing Parkinson's disease symptoms. Input data The output of each random forest tree is only responsible for prediction. One of the symptoms of Parkinson's disease, its predictive outcome is , ;

[0022] Each group of random forest trees contains 230 trees, with a maximum depth of 25 layers and a minimum number of leaf node samples of 6. During model training, the splitting criterion for each group of random forest trees is selected as the MSE (Mean Squared Error) loss function, and a medical prior knowledge splitting criterion function is introduced on this basis. The MSE loss values ​​calculated for Parkinson's disease symptoms of type O and type P were respectively and After that, and Perform summation and join Multiplying these results yields the splitting function calculations for the 0th and pth classes of Parkinson's disease symptoms. Finally, summing the splitting function calculations across all classes yields the medical prior knowledge splitting standard function. The calculation results.

[0023] Preferably, the medically constrained embedded gradient boosting tree model contains a total of Q independent gradient boosting tree sub-models, each corresponding to the output data of predicted Parkinson's disease symptom representations. One of the specific symptoms of Parkinson's disease;

[0024] Adding a symptom value constraint loss term to the model training loss function ensures that the predicted Parkinson's disease symptom values ​​preferentially fall within the medically permissible symptom range, thereby improving the clinical reliability of the model's predictions. Specifically:

[0025] Set each Parkinson's disease symptom The range of medical values, The model's predicted output is The overall training loss function of the model designed based on medical constraints The calculation method for this function is as follows: The sum of, among which This represents the function for retrieving the maximum value.

[0026] The medically constrained embedded gradient boosting tree model consists of Q independent LightGBM models, with each LightGBM model receiving Parkinson's disease symptom data as input. Input data Each gradient boosting tree model output is only responsible for predicting one Parkinson's disease symptom, and its prediction result is... , Each LightGBM model has 270 base regression trees, a learning rate of 0.09, and a maximum depth limit of 18 layers.

[0027] Preferably, the frame-level visual perception unit includes an upper processing pathway to fully perceive the local Parkinson's pathological features of the Parkinson's patient, and a lower processing pathway to fully perceive the whole-body Parkinson's pathological features.

[0028] In the above processing path, preliminary feature processing is first performed through two sequentially connected SFEnet modules. Then, the obtained data feature map is subjected to regional adaptive pooling operation using a regional pooling unit. The regional pooling unit first divides the output feature map of the SFEnet module into feature subspaces, then performs adaptive average pooling operation on each feature subspace, and finally concatenates the pooling operation results of each sub-region to obtain the global feature vector.

[0029] Simultaneously, a feature difference calculation layer is used to calculate the feature differences between the left and right parts of the body, thereby incorporating the asymmetric feature information of Parkinson's disease. The feature difference calculation layer first divides the input feature map into two symmetrical regions, left and right, along the vertical axis of the image center, and extracts the feature maps of the left and right sides respectively. Then, it performs element-by-element comparison on the feature points of corresponding positions on the left and right sides, calculates their absolute difference in each channel, and finally extracts the symmetric difference features that reflect the degree of inconsistency between the left and right sides of the body movements of Parkinson's patients.

[0030] Subsequently, residual connections are introduced to perform channel concatenation of the asymmetric feature information and the output feature information of two sequentially connected SFEnet modules. After further feature processing using the SFEnet module, the feature is obtained. ;

[0031] In the processing pathway, the SFEnet module is first introduced to perform preliminary feature processing operations to improve the ability to identify pathological features throughout the body. Then, the stage module of the OpenPose network is introduced to extract heatmaps of key human points and their spatial topology, capturing abnormal manifestations in limb stability and tremor control in Parkinson's patients. Based on the extracted skeleton graph structure embedding representation, a pose structure map is constructed through the graphsage layer, and a graph convolutional layer is introduced to model the spatial relationships between skeleton graph nodes, enhancing the global coordination representation between actions. Finally, a graph attention layer is introduced to dynamically adjust the information propagation intensity between key nodes, amplifying the core pathological pose signal to obtain features. ;

[0032] Finally, the features and characteristics The data is fed into a convolutional layer for final processing, and the ReLU activation function is used to complete the final activation to obtain the frame-level image output features. .

[0033] Preferably, the SFEnet module first includes two interconnected convolutional layers, a batch normalization layer, and a ReLU activation function for basic feature extraction; then, it uses spatially compressed convolutional layers, residual blocks, lightweight attention layers, and pooling layers for final processing to complete the overall construction of the SFEnet module.

[0034] Preferably, the multimodal symptom fusion unit is based on the obtained Parkinson's disease symptom prediction results. First, a Batch Normalization (BN) layer is used for feature normalization. Then, three interconnected fully connected layers are used for preliminary feature computation. Next, the Mish activation function is used for non-linear feature activation. Subsequently, residual connective blocks are introduced to further enhance the computational capability for numerical symptom features. Finally, after processing through two interconnected layers and a Simgoid activation function, the overall construction of the numerical Parkinson's disease symptom processing unit is completed, and the numerical fusion features of Parkinson's disease symptoms are calculated. .

[0035] Preferably, the cross-frame symptom fusion unit targets the input frame-level image feature sequence. By using an additive injection time-aware method, features are output to each frame-level image. Summation and addition of learnable temporal coding vectors , Temporal frame-level features are obtained. Subsequently, a gated attention mechanism was introduced to calculate the ill-conditioned feature weights corresponding to each temporal frame-level feature. Subsequently, the calculated ill-conditioned feature weights were used. and temporal frame-level features The dynamic features of Parkinson's disease across frames are obtained by calculating the corresponding cumulative summation. .

[0036] Preferably, the Parkinson's disease identification head is based on the Parkinson's cross-frame dynamic features calculated by the cross-frame symptom fusion unit. And the numerical fusion features of Parkinson's symptoms calculated by the multimodal symptom fusion unit. First, the two features are normalized using a channel normalization layer. After normalization, channel concatenation is performed. The concatenated features are then processed through two interconnected convolutional layers. Next, a linear transformation layer with Tanh activation is introduced to enhance non-linear expressiveness. Finally, a double fully connected layer with residual connections is used for further processing. The activation function process yields the final identification result of Parkinson's disease symptoms. ,in This indicates that the person is not ill. This indicates an illness with a severity level of 1. This indicates that the illness is of severity level 2. This indicates an illness with a severity level of 3. This indicates an illness with a severity level of 4. This indicates an illness with a severity level of 5.

[0037] Preferably, the Parkinson's disease comprehensive discrimination module, based on the early identification model of Parkinson's disease, continuously identifies Parkinson's disease in the subject within multiple time periods, and integrates the identification results obtained from different time periods to form a Parkinson's disease identification result sequence. ;

[0038] Based on the calculated Parkinson's disease identification result sequence Spdr, the trend slope of this sequence is calculated by constructing a first-order linear regression fit. If this A positive value indicates that the Parkinson's identification result shows an increasing trend over time, which is consistent with the characteristics of Parkinson's disease development;

[0039] For sequence Modeling the sequence perturbation, i.e., solving the sequence perturbation. Standard deviation The larger this value, the lower the stability of the Parkinson's disease symptoms in the test subject; the smaller this value, the more stable the Parkinson's disease symptoms in the test subject, that is, the higher the diagnosis rate of Parkinson's disease.

[0040] The Parkinson's disease identification result sequences for the normal population were calculated using the early identification model for Parkinson's disease. And the Parkinson's disease identification result sequence of the confirmed cases. Subsequently, the Spdr sequence for identifying Parkinson's disease in the test subjects and the Spdr sequence for identifying the normal population were calculated respectively. Normal group Mahalanobis distance and the sequence of confirmed case identification results The confirmed case, Ma, is far from ;

[0041] Subsequently, based on the Mahalanobis distance of the normal group Distance from confirmed case Ma The distance anomaly index is obtained by calculating the ratio. The higher this index, the closer the subject is to the Parkinson's disease group;

[0042] Based on the calculated distance anomaly index The calculated trend slope and standard deviation , and Summation and Union Calculate the discriminant factor by constructing the final discriminant for Parkinson's disease through difference calculation. ;

[0043] Set judgment threshold When the discriminant factor Greater than the judgment threshold If the test subject is diagnosed with Parkinson's disease, the severity level is the rounded result of the mean of the severity levels of Parkinson's disease in the Parkinson's disease identification result sequence Spdr.

[0044] A second aspect of the present invention provides a method for early identification of Parkinson's disease based on multi-task learning, which deploys the early identification system for Parkinson's disease as described in the first aspect in a hospital workstation system, and includes the following process:

[0045] S1, acquire audio data of the user's voice expression and limb movement sequence data. The limb movement sequence data is specifically the limb inertial movement sequence data collected by the accelerometer, which is used as data to characterize Parkinson's disease symptoms; at the same time, acquire motion image data.

[0046] S2, based on the user's voice expression audio data and limb movement sequence data collected by the data acquisition module, constructs a medical prior knowledge random forest model and a medical constraint embedded gradient boosting tree model, and combines them with an adaptive weighted logistic regression model to output the Parkinson's disease symptom recognition results.

[0047] S3 firstly, based on the acquired motion image data, the frame-level visual perception unit performs convolutional feature extraction on each frame of the image, and combines an attention mechanism to enhance the response of facial muscles and key limb regions. Secondly, the cross-frame symptom fusion unit models the temporal features of continuous video frames to obtain cross-frame dynamic features. Finally, the multimodal symptom fusion unit uses a multilayer perception network to perform feature calculation processing on the Parkinson's disease symptom prediction results obtained by the symptom recognition module, and fuses them with the cross-frame dynamic features to guide the network to focus on the image semantic regions related to medical symptoms. Finally, the Parkinson's disease recognition head outputs the early recognition results of Parkinson's disease.

[0048] S4. Under long content frames, the Parkinson's disease short-term identification model is used to perform multiple disease probability detections to obtain the Parkinson's disease factor sequence of the subject; then, the sequence is compared with the factor sequences of normal people and Parkinson's patients through multi-step mathematical analysis to finally obtain the standard identification result of Parkinson's disease.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] (1) Construction of a subjective intelligent identification mechanism: This invention effectively breaks through the reliance on the experience of neurology experts and manual consultation and assessment in traditional methods. It uses multimodal objective data to drive the model for automatic identification, avoids human subjective misjudgment, and significantly improves the objectivity, consistency and scalability of early identification of Parkinson's disease.

[0051] (2) Deep symptom coupling modeling mechanism for pathological heterogeneity: In view of the characteristics of multi-source heterogeneity and complex expression of Parkinson's disease symptoms, this invention automatically models the nonlinear coupling relationship between multimodal features through neural network structure, effectively capturing early weak pathological signals such as actions, language, and expressions, and improving the sensitivity and recognition accuracy of early symptom manifestations.

[0052] (3) High data efficiency multi-task fusion modeling mechanism: Considering the practical problems of high difficulty in collecting Parkinson's disease data and high labeling cost, this invention adopts a multi-task joint training strategy, makes full use of the synergistic association between symptom prediction and symptom classification, enhances the robustness of the model under low sample conditions, and improves the generalization performance in complex clinical scenarios. Attached Figure Description

[0053] Figure 1 This is a flowchart illustrating the overall technical route of the present invention.

[0054] Figure 2 This is a structural diagram of the Parkinson's disease short-term identification model of the present invention.

[0055] Figure 3 This is a structural diagram of the frame-level visual perception unit of the present invention.

[0056] Figure 4 This is a comparison chart showing the accuracy of Parkinson's disease identification in this invention.

[0057] Figure 5 This is a comparison chart of ablation experiments using the algorithm of this invention. Detailed Implementation

[0058] This invention provides a system and method for early identification of Parkinson's disease based on multi-task learning, the overall technical approach of which is as follows: Figure 1 As shown. The overall process includes four core stages:

[0059] S1, Construction of Parkinson's Disease Discrimination Dataset: Early-stage Parkinson's disease patients exhibit abnormalities in body movement, speech expression, and facial expressions. Therefore, this invention constructs a complete Parkinson's disease symptom representation dataset and an early Parkinson's disease identification dataset, encompassing audio data of speech expression and limb movement sequence data to characterize early Parkinson's disease symptoms; it also includes motion image data.

[0060] S2, Parkinson's disease symptom recognition module: Based on the limb movement time-series data and user voice expression audio data collected in S1, a medical prior knowledge random forest model and a medical constraint embedded gradient boosting tree model are constructed, and combined with an adaptive weighted logistic regression model to obtain the Parkinson's disease symptom recognition results.

[0061] S3, Construction of a short-term identification model for Parkinson's disease: Based on the Parkinson's disease symptoms obtained from the analysis of limb movement sequence data and language temporal data in S2, and combined with the motion image data collected in S1, this invention constructs a short-term identification model for Parkinson's disease that includes a frame-level visual perception unit and a symptom fusion unit to obtain the Parkinson's disease discrimination results under the collected video.

[0062] S4, the Parkinson's Disease Comprehensive Discrimination Module: Based on the short-term Parkinson's disease identification results obtained in S3, the probability of the test subject having the disease can be determined to a certain extent. However, short-term test results have a large degree of individual randomness, and the short-term symptoms of some diseases are similar to those of Parkinson's disease, making misjudgment easy to occur based on short-term analysis alone. To improve the identification accuracy, this invention uses the short-term Parkinson's disease identification model in S3 to perform multiple disease probability detections under long content frames to obtain the Parkinson's disease factor sequence of the test subject; then, this sequence is compared with the factor sequences of normal individuals and Parkinson's patients through multi-step mathematical analysis to finally obtain the standard Parkinson's disease identification result.

[0063] S5, Parkinson's Disease Recognition Model Deployment and Operation: The trained Parkinson's disease symptom recognition module, short-term Parkinson's disease recognition model, and comprehensive Parkinson's disease discrimination module are integrated into the hospital workstation system. The model receives real-time input action, speech, and image data streams, and sequentially performs inference through the Parkinson's disease symptom recognition module, the short-term Parkinson's disease recognition model, and the comprehensive Parkinson's disease discrimination algorithm. The final output Parkinson's disease discrimination result specifically includes information on the Parkinson's disease status and severity level. This recognition result provides timely reference for the clinical diagnosis of Parkinson's disease.

[0064] The invention will be further described below with reference to specific embodiments.

[0065] I. Dataset Construction

[0066] In the stage of constructing the Parkinson's disease symptom representation dataset, this invention conducts data collection operations under medical laboratory conditions on normal individuals and Parkinson's disease patients, specifically including the acquisition of speech expression audio data using a microphone. Limb motion sequence data collected by accelerometers deployed on the limbs And use it as data for the symptom characteristics of Parkinson's disease. Input data ;

[0067] Characterizing input data calibration: Simultaneously, the audio data during the above process... During data collection, non-invasive electromyography (EMG) sensors are deployed to measure the tension of the laryngeal muscles during vocalization. Surface strain gauge sensors were deployed to collect the force of the throat muscles during vocalization. Deploy an electromagnetic vocalization meter to detect the amplitude of tongue movement. Deploy a small inertial measurement unit to measure the mandibular opening angle. ;

[0068] The limb movement sequence data during the above process During data acquisition, electromyography (EMG) sensors were deployed to collect the amplitude of EMG signals from the forearm muscle groups. Interval of electromyographic signal bursts Flexible pressure sensors were deployed to collect the peak force of the fingers' gripping operations. Deploy a plantar pressure sensor array to collect peak heel force. ;

[0069] The data collected during the above process Combined as data for Parkinson's disease symptom representation Output data This allows for the specific symptom labeling of the input data. The input data is then then... With output data By combining these data, a complete set of Parkinson's disease symptom characteristics can be obtained. , .

[0070] Following the data construction method described above, a total of N Parkinson's disease symptom data were collected. The obtained N data points on Parkinson's disease symptoms Combined to complete the Parkinson's disease symptom representation dataset The construction process.

[0071] Motion Image Acquisition and Early Identification Dataset Construction for Parkinson's Disease: This invention aims to achieve early identification of Parkinson's disease by referencing Parkinson's disease symptom data. Output data in In addition, high-definition cameras are used to collect motion image data. This will provide more refined and accurate body and facial movement features for subsequent Parkinson's disease identification;

[0072] Construction of input data for early identification of Parkinson's disease: S1-1 Output data Parkinson's disease motion imaging data acquired in S1-2 Together as early identification data for Parkinson's disease One input data in , ;

[0073] Construction of an early identification dataset for Parkinson's disease: Based on the constructed early input data for Parkinson's disease. The data collection involved professional medical personnel analyzing the Parkinson's disease status of participants (both healthy individuals and Parkinson's patients). Severity of Parkinson's disease Perform judgment and data labeling processing;

[0074] Specifically, illness refers to the condition of the illness. A label of 0 indicates the presence of the disease, while a label of 1 indicates no disease. Additionally, the severity of Parkinson's disease is also considered. The classification includes five levels (1, 2, 3, 4, 5), with higher levels indicating more severe Parkinson's disease. and Data is stitched together to obtain Parkinson's disease diagnostic data, which can then be used as early identification data for Parkinson's disease. One output data in .

[0075] A total of N Parkinson's disease early identification data were collected according to the described data collection method. The obtained N pieces of early identification data for Parkinson's disease Combined to complete the Parkinson's disease early identification dataset The construction process. It should be noted that the number of data entries in the early identification dataset for Parkinson's disease and the number of data entries in the Parkinson's disease symptom representation dataset... The data entries are consistent.

[0076] II. Construction of Parkinson's Disease Symptom Recognition Module

[0077] The Parkinson's disease symptom recognition module designed in this invention adopts an improved machine learning model component and an integrated learning overall architecture design, which can further improve the accuracy and robustness of Parkinson's disease symptom recognition, provide a direct symptom feature basis for subsequent Parkinson's disease recognition, and ensure the stability of the final Parkinson's disease recognition.

[0078] 1. Parkinson's Disease Symptom Recognition Module Design: The Parkinson's disease symptom recognition module designed in this invention adopts a modular architecture design with separate prediction of base components and meta-components. In the base component, a random forest model incorporating prior medical knowledge is designed to predict Parkinson's disease symptoms and obtain the predicted results. A medically constrained embedded gradient boosting tree model was designed to predict Parkinson's disease symptoms, and the prediction results were obtained. .

[0079] The component uses an adaptive weighted logistic regression model, and the prediction results are analyzed using this model. and prediction results Scene-adaptive fusion processing is performed, and weights are dynamically assigned based on the predicted features. After the fusion processing is completed, the final prediction result of Parkinson's disease symptoms is obtained. .

[0080] 2. Design of a Random Forest Model Based on Medical Prior Knowledge: The random forest model designed in this invention is a multi-objective random forest model architecture, that is, it outputs data based on the symptom representation of Parkinson's disease. Each Parkinson's symptom in the dataset is matched with an independent set of regression trees.

[0081] Based on existing medical prior knowledge There are correlations among the eight categories of Parkinson's disease symptoms included in the study. Based on clinical experience, some Parkinson's disease symptoms show strong correlations, indicating a high probability of co-occurrence, while others show weak correlations, indicating that these symptoms do not affect each other.

[0082] This invention constructs a symptom coefficient correlation matrix based on existing medical literature:

[0083]

[0084] in, , , Indicates the output data Of the eight types of Parkinson's disease symptoms, and The stronger the correlation between the two types of Parkinson's disease symptoms, the closer this value is to 1.

[0085] The medical prior knowledge random forest model designed in this invention includes a total of 8 groups of random forest trees. Each group of random forest trees is independent of the others, and the input to each group of random forest trees is Parkinson's disease symptom representation data. Input data The output of each random forest tree is only responsible for prediction. One of the symptoms of Parkinson's disease, its predictive outcome is , .

[0086] Each random forest tree contains 230 trees, with a maximum depth of 25 layers and a minimum of 6 leaf node samples per tree. During model training, the splitting criterion for each random forest tree is the Mean Squared Error (MSE) loss function, and a medical prior knowledge splitting criterion function is designed based on this. Specifically, the MSE loss values ​​calculated for Parkinson's disease symptoms of type O and type P are respectively and After that, and Perform summation and join Multiplying these results yields the splitting function calculations for the 0th and pth classes of Parkinson's disease symptoms. Finally, summing the splitting function calculations across all classes yields the medical prior knowledge splitting standard function. The calculation results.

[0087] Therefore, the splitting function By incorporating the prior medical knowledge of the correlation between Parkinson's disease symptoms into the model process without affecting the training process of the single-group random forest tree, the predictive effect of related symptoms is further improved, thereby enhancing the overall prediction accuracy.

[0088] 3. Medical Constraint Embedded Gradient Boosting Tree Model Design: The gradient boosting tree model designed in this invention contains a total of 8 independent gradient boosting tree sub-models, each corresponding to the output data of predicted Parkinson's disease symptom representations. A specific symptom of Parkinson's disease.

[0089] To incorporate the known limits of Parkinson's disease symptoms in the medical field into the gradient boosting tree model constructed in this invention, a symptom value constraint loss term is added to the model training loss function to ensure that the predicted Parkinson's disease symptom values ​​preferentially fall within the medically permissible symptom value range, thereby improving the clinical reliability of the model's predictions. Specifically:

[0090] Set each Parkinson's disease symptom The range of medical values, The model's predicted output is The overall training loss function of the model designed based on medical constraints The calculation method for this function is as follows: The sum of, among which This represents the function for retrieving the maximum value.

[0091] Furthermore, the medically constrained embedded gradient boosting tree model includes eight independent LightGBM models, each of which is input to Parkinson's disease symptom representation data. Input data Each gradient boosting tree model output is only responsible for predicting one Parkinson's disease symptom, and its prediction result is... , Each LightGBM model has 270 base regression trees, a learning rate of 0.09, and a maximum depth limit of 18 layers. The overall training check function constructed above is used as the training criterion during model training.

[0092] 4. Adaptive Weighted Logistic Regression Model Design: Based on the prediction results of the obtained random forest model and the prediction results obtained from the gradient boosting tree model An adaptive weighted logistic regression model was used to adaptively fuse the results to obtain the final prediction results for Parkinson's disease symptoms. Furthermore, the weights are derived from the logistic regression network based on the input data. Dynamically generated.

[0093] III. Construction of a Short-Term Identification Model for Parkinson's Disease

[0094] This invention proposes a short-term recognition neural model architecture for Parkinson's disease. The model takes numerical data of Parkinson's symptoms and motion videos of the subject as inputs and consists of three main modules.

[0095] First, since the input images are video frame images, a frame-level visual perception unit is designed to extract convolutional features from each frame of the video, combined with an attention mechanism to enhance the response of facial muscles and key limb regions. Second, a cross-frame symptom fusion unit is used to model the temporal features of consecutive video frames, extracting cross-frame dynamic features. Finally, a multimodal symptom fusion unit is constructed, and a multilayer perceptual network is used to predict Parkinson's disease symptoms obtained from the S2 process. Feature computation is performed and fused with cross-frame dynamic features to guide the network to focus on image semantic regions related to medical symptoms, ultimately outputting early Parkinson's disease identification results. The overall block diagram of the short-term Parkinson's disease identification model is shown below. Figure 2 As shown.

[0096] 1. Frame-level visual perception unit architecture design

[0097] The structural block diagram of the frame-level visual perception unit constructed in this invention is as follows: Figure 3 As shown, actual Parkinson's patients exhibit differences in overall and localized motor function compared to healthy individuals. Therefore, while analyzing the overall and localized motor function of Parkinson's patients using image data is computationally intensive, the rich image features processed can further improve the accuracy of Parkinson's disease identification.

[0098] The frame-level visual perception unit includes an upper data processing pathway to fully perceive the local Parkinsonian pathological features of the Parkinsonian subject, and a lower processing pathway to fully perceive the systemic Parkinsonian pathological features.

[0099] In the above processing path, preliminary feature processing is first performed through two sequentially connected SFEnet modules. Then, a designed regional pooling unit is used to perform adaptive pooling operations on the obtained data feature map in different regions. This helps the model further improve its ability to perceive key local ill-conditioned features and reduce interference from redundant features. The regional pooling unit first divides the output feature map of the SFEnet module into several feature subspaces, then performs adaptive average pooling operations on each feature subspace, and finally concatenates the pooling results of each sub-region to obtain the global feature vector.

[0100] Due to the inherent symmetrical body structure of Parkinson's patients, this invention designs a feature difference calculation layer to calculate the feature differences between the left and right parts of the body, thereby further incorporating the asymmetrical feature information of Parkinson's disease. The feature difference calculation layer first divides the input feature map into two symmetrical regions, left and right, along the vertical axis of the image center, extracting the left and right feature maps respectively. Then, it performs element-by-element comparison of feature points at corresponding positions on the left and right sides, calculating their absolute differences in each channel. Finally, it extracts symmetrical difference features reflecting the degree of inconsistency in the left-right body movements of Parkinson's patients.

[0101] Subsequently, residual connections are introduced to perform channel concatenation of the asymmetric feature information and the output feature information of two sequentially connected SFEnet modules. Further feature processing is then performed using the SFEnet module to complete the construction of the data processing path and obtain the features. .

[0102] In the lower processing pathway designed in this invention, the SFEnet module is first introduced to complete preliminary feature processing operations, thereby improving the ability to identify systemic pathological features. Subsequently, the stage module of the OpenPose network is introduced to extract heatmaps of key human points and their spatial topology, thereby capturing the abnormal performance of Parkinson's patients in limb stability and tremor control. Based on the extracted skeleton graph structure embedding representation, a graphsage layer is designed to construct a posture structure map, and a graph convolutional layer is further introduced to model the spatial relationships between skeleton graph nodes, enhancing the global coordination representation between actions. Finally, a graph attention layer is introduced to dynamically adjust the information propagation intensity between key nodes, further amplifying the core pathological posture signals, thereby ensuring that the model can smoothly perceive the systemic symptom structure of Parkinson's disease and obtain features. .

[0103] Finally, the features and characteristics The data is fed into a convolutional layer for final processing, and the ReLU activation function is used to complete the final activation to obtain the frame-level image output features. .

[0104] The SFEnet module used in this module first includes two interconnected convolutional layers, a batch normalization layer, and a ReLU activation function for basic feature extraction. Then, spatially compressed convolutional layers, residual blocks, lightweight attention layers, and pooling layers are used for final processing to complete the overall construction of the SFEnet module.

[0105] 2. Overall Design of Cross-Frame Disease Fusion Unit

[0106] Based on the obtained frame-level image output features , The present invention further designs a cross-frame symptom fusion unit to complete cross-frame processing and fusion operations between features of different video frames in order to extract cross-frame dynamic features of Parkinson's disease. .

[0107] In the cross-frame symptom fusion unit, the input frame-level image feature sequence is... By using an additive injection time-aware method, features are output to each frame-level image. Summation and addition of learnable temporal coding vectors , Temporal frame-level features are obtained. Subsequently, a gated attention mechanism was introduced to calculate the ill-conditioned feature weights corresponding to each temporal frame-level feature. Then, the calculated ill-conditioned feature weights are used. and temporal frame-level features The dynamic features of Parkinson's disease across frames are obtained by calculating the corresponding cumulative summation. .

[0108] 3. Multimodal symptom fusion unit

[0109] Based on the obtained Parkinson's disease symptom prediction results First, a Batch Normalization (BN) layer is used for feature normalization. Then, three interconnected fully connected layers are used for preliminary feature computation. Next, the Mish activation function is used for non-linear feature activation. Subsequently, residual connective blocks are introduced to further enhance the computational capability for numerical symptom features. Finally, after processing through two interconnected layers and a Simgoid activation function, the overall construction of the numerical Parkinson's disease symptom processing unit is completed, and the numerical fusion features of Parkinson's disease symptoms are calculated. .

[0110] 4. Construction of a Head Framework for Parkinson's Disease Recognition

[0111] Based on the calculated cross-frame dynamic features of Parkinson's disease and the calculated numerical fusion features of Parkinson's disease symptoms First, the two features mentioned above are normalized using a channel normalization layer. After normalization, channel concatenation is performed. The concatenated features are then further processed through two interconnected convolutional layers. Next, a linear transformation layer with Tanh activation is introduced to further enhance non-linear expressive power. Based on this, a double fully connected layer structure with residual connections is introduced for further processing. Finally, [the following steps are described]. The activation function process yields the final identification result of Parkinson's disease symptoms. ,in This indicates that the person is not ill. This indicates an illness with a severity level of 1. This indicates an illness with a severity level of 2. This indicates an illness with a severity level of 3. This indicates an illness with a severity level of 4. This indicates an illness with a severity level of 5.

[0112] IV. Implementation of the Parkinson's Disease Comprehensive Diagnosis Module

[0113] Time-dependent multi-recognition process implementation: Based on the Parkinson's disease symptom recognition results obtained from the final recognition. Because the symptoms of some other diseases may be similar to those of Parkinson's disease in a short period of time, using this result directly as the final criterion for Parkinson's disease may lead to a high misdiagnosis rate.

[0114] To improve the accuracy of comprehensive Parkinson's disease identification and effectively reduce the risk of misdiagnosis, this invention proposes a multi-time period identification result sequence analysis method. Based on a constructed early Parkinson's disease identification model, this method continuously identifies Parkinson's disease in the subject across multiple time periods and integrates the identification results from different time periods to form a Parkinson's disease identification result sequence. .

[0115] Sequence Self-Change Feature Analysis Unit Design: Parkinson's disease, as a neurodegenerative disease, typically exhibits a gradual worsening of symptoms over time. Based on the calculated Parkinson's disease identification result sequence Spdr, a first-order linear regression fitting is constructed to calculate the trend slope of this sequence. If this A positive value indicates that the Parkinson's identification result shows an increasing trend over time, which is consistent with the characteristics of Parkinson's disease development.

[0116] In addition to calculating the trend slope, this unit further analyzes the sequence. Modeling the sequence perturbation, i.e., solving the sequence perturbation. Standard deviation The larger this value, the lower the stability of the Parkinson's disease symptoms in the test subject; the smaller this value, the more stable the Parkinson's disease symptoms in the test subject, that is, the higher the diagnosis rate of Parkinson's disease.

[0117] Temporal distribution anomaly detection based on multi-population comparison: To avoid the risk of misjudgment caused by single-person sequence encouragement, a group temporal comparison mechanism is further introduced. The described Parkinson's disease identification model is used to calculate the Parkinson's disease identification result sequences for the normal population. And the Parkinson's disease identification result sequence of the confirmed cases. Subsequently, the Spdr sequence for identifying Parkinson's disease in the subjects and the Spdr sequence for identifying the normal population were calculated separately. Normal group Mahalanobis distance and the sequence of confirmed case identification results The confirmed case, Ma, is far from .

[0118] Subsequently, based on the Mahalanobis distance of the normal group Distance from confirmed case Ma The distance anomaly index is obtained by calculating the ratio. The higher this index, the closer the subject is to the Parkinson's disease group.

[0119] Construction of the final diagnostic mechanism for Parkinson's disease: based on the calculated distance anomaly index The calculated trend slope and standard deviation , and Summation and Union Calculate the discriminant factor by constructing the final discriminant for Parkinson's disease through difference calculation. .

[0120] In addition, a judgment threshold is set. When the discriminant factor Greater than the judgment threshold If the test subject is diagnosed with Parkinson's disease, the severity level is the rounded result of the mean of the severity levels of Parkinson's disease in the Parkinson's disease identification result sequence Spdr.

[0121] V. Model Deployment

[0122] Multimodal data acquisition is achieved by deploying an acquisition system consisting of three types of sensing devices in the Parkinson's disease screening scenario: a patient wearing an inertial sensor to acquire actual limb movement data, a patient using a voice microphone to acquire continuous speech and word pronunciation data, and a patient using a video camera to acquire image sequences containing facial expressions and limb movements.

[0123] Data preprocessing and synchronization mechanisms are implemented as follows: The acquired raw multimodal data undergoes time alignment, noise filtering, and format conversion. Filtering algorithms are used to remove outliers from motion sequence data and improve speech signal quality, while image data is normalized and standardized. All data is synchronized to a unified time-series index for easy model alignment analysis.

[0124] Deployment and Operation of the Parkinson's Disease Recognition Model: The trained Parkinson's disease symptom recognition module, short-term Parkinson's disease recognition model, and comprehensive Parkinson's disease discrimination module are integrated into the hospital workstation system. The model receives real-time input data streams of action, speech, and images, and sequentially performs inference through the Parkinson's disease symptom recognition module, the short-term Parkinson's disease recognition model, and the comprehensive Parkinson's disease discrimination algorithm. The final output Parkinson's disease discrimination result specifically includes information on the Parkinson's disease status and severity level. This recognition result provides timely reference for the clinical diagnosis of Parkinson's disease.

[0125] VI. Explanation of Experimental Results

[0126] To further demonstrate that the Parkinson's disease identification method and process proposed in this invention have good detection accuracy, a performance comparison experiment was conducted between this invention and existing Parkinson's disease identification methods. The comparative experimental methods included an SVM-based Parkinson's disease identification algorithm and a YOLOv5-based Parkinson's disease identification algorithm. Performance comparison tests were conducted for four different types of Parkinson's disease, and the identification accuracy was defined as: the number of correctly identified cases divided by the total number of identified cases.

[0127] The comparison results are as follows Figure 4 As shown, the algorithm proposed in this invention exhibits the best detection accuracy across four different severity levels of Parkinson's disease. This is because the algorithm performs joint analysis and processing of image data and numerical symptom data, and also incorporates a mathematical analysis algorithm to further reduce interference from suspected cases.

[0128] To demonstrate that the Parkinson's disease comprehensive judgment module proposed in this invention has a positive effect on improving the accuracy of Parkinson's disease identification, a performance comparison experiment was conducted under two different conditions: freezing the algorithm and using the algorithm normally.

[0129] like Figure 5 As shown, in addition to comparing the model performance of the frozen judgment analysis algorithm and the normal use judgment analysis algorithm, this invention also applies the designed judgment analysis algorithm to the SVM Parkinson's disease recognition algorithm. The results show that this judgment analysis algorithm can improve the recognition accuracy. This is because the algorithm based on mathematical analysis can further reduce the interference caused by suspected symptoms and accidental factors, and by observing over a longer time scale, it ensures the stability and accuracy of the recognition.

[0130] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0131] While the specific embodiments of the present invention have been described above, they are not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A Parkinson's disease early identification system based on multi-task learning, characterized in that, It includes a data acquisition module, a symptom recognition module, a short-term Parkinson's disease identification model, and a comprehensive Parkinson's disease discrimination module; The data acquisition module is used to acquire user voice audio data and limb movement sequence data, and use them as data representing symptoms of Parkinson's disease; at the same time, it acquires motion image data. The symptom recognition module is based on the user's voice expression audio data and limb movement sequence data collected by the data acquisition module. It constructs a medical prior knowledge random forest model and a medical constraint embedded gradient boosting tree model, and combines them with an adaptive weighted logistic regression model to output the Parkinson's disease symptom recognition results. The medical prior knowledge random forest model is a multi-objective random forest model architecture, that is, it outputs data based on the symptom representation of Parkinson's disease. Each Parkinson's symptom was matched with an independent set of regression trees; a symptom coefficient correlation matrix was constructed based on existing medical literature. The random forest model consists of Q groups of random forest trees, each group being independent of the others, and the input to each group of random forest trees is data representing Parkinson's disease symptoms. Input data The output of each random forest tree is only responsible for prediction. One of the symptoms of Parkinson's disease, its predictive outcome is , ; The described medical prior knowledge random forest model contains 230 trees per group of random forest trees, with a maximum depth of 25 layers and a minimum number of leaf node samples of 6. During model training, the splitting criterion for each group of random forest trees is selected as the MSE (Mean Squared Error) loss function, and a medical prior knowledge splitting criterion function is introduced on this basis. ; The MSE loss values ​​calculated for Parkinson's disease symptoms of type O and type P were respectively and After that, and Perform summation and join Multiplying these results yields the splitting function calculations for the 0th and pth classes of Parkinson's disease symptoms. Finally, summing the splitting function calculations across all classes yields the medical prior knowledge splitting standard function. The calculation results; The medically constrained embedded gradient boosting tree model contains a total of Q independent gradient boosting tree sub-models, each corresponding to the output data for predicting Parkinson's disease symptom representations. A specific symptom of Parkinson's disease; define each Parkinson's disease symptom The range of medical values, The model's predicted output is The overall training loss function of the model designed based on medical constraints The calculation method for this function is as follows: The sum of, among which This represents the function for retrieving the maximum value. The adaptive weighted logistic regression model, based on the prediction results of the obtained random forest model, and the prediction results obtained from the gradient boosting tree model An adaptive weighted logistic regression model was used to adaptively fuse the results to obtain the final Parkinson's disease symptom identification result. ; The Parkinson's disease short-term identification model first extracts convolutional features from each frame of the acquired motion image data using a frame-level visual perception unit, and enhances the response of facial muscles and key limb regions by combining an attention mechanism. Second, it models the temporal features of continuous video frames through a cross-frame symptom fusion unit to obtain cross-frame dynamic features. Finally, through a multimodal symptom fusion unit, it uses a multilayer perception network to perform feature calculation processing on the Parkinson's disease symptom identification results obtained by the symptom identification module, and fuses them with the cross-frame dynamic features to guide the network to focus on the semantic regions of the image related to medical symptoms. Finally, the Parkinson's disease identification head outputs the early identification results of Parkinson's disease. The Parkinson's disease comprehensive discrimination module performs multiple disease probability detections using a short-term Parkinson's disease identification model under long content frames to obtain the Parkinson's disease factor sequence of the test subject. The sequence was then subjected to multi-step mathematical analysis with factor sequences from normal individuals and Parkinson's patients to obtain the standard identification results for Parkinson's disease.

2. The Parkinson's disease early identification system based on multi-task learning as described in claim 1, characterized in that: The process of building the dataset used to train the symptom recognition module includes: Audio data of speech expression collected using a microphone The limb motion data collected by accelerometers deployed on the limbs are specifically limb motion sequence data. And use it as data for the symptom characteristics of Parkinson's disease. Input data ; When processing audio data During data collection, non-invasive electromyography (EMG) sensors are deployed to measure the tension of the laryngeal muscles during vocalization. Surface strain gauge sensors were deployed to collect the force of the throat muscles during vocalization. Deploy an electromagnetic vocalization meter to detect the amplitude of tongue movement. Deploy a small inertial measurement unit to measure the mandibular opening angle. ; In collecting limb inertial movement sequence data During data acquisition, electromyography (EMG) sensors were deployed to collect the amplitude of EMG signals from the forearm muscle groups. Interval of electromyographic signal bursts Flexible pressure sensors were deployed to collect the peak force of the fingers' gripping operations. Deploy a plantar pressure sensor array to collect peak heel force. ; The data collected during the above process Combined as output data of the symptom recognition module .

3. The Parkinson's disease early identification system based on multi-task learning as described in claim 1, characterized in that: The symptom recognition module adopts a modular architecture design with separate prediction of base components and meta-components. In the base component, a random forest model incorporating prior medical knowledge is designed to predict Parkinson's disease symptoms and obtain the predicted symptoms. A medically constrained embedded gradient boosting tree model was designed to predict Parkinson's disease symptoms, and the prediction results were obtained. ; The meta-component uses an adaptive weighted logistic regression model, and this model is used to evaluate the prediction results. and prediction results Scene-adaptive fusion processing is performed, and weights are dynamically assigned based on the predicted features. After the fusion processing is completed, the final prediction result of Parkinson's disease symptoms is obtained. .

4. The Parkinson's disease early identification system based on multi-task learning as described in claim 1, characterized in that: The medically constrained embedded gradient boosting tree model comprises Q independent LightGBM models, with each LightGBM model receiving Parkinson's disease symptom data as input. Input data Each gradient boosting tree model output is only responsible for predicting one Parkinson's disease symptom, and its prediction result is... , Each LightGBM model has 270 base regression trees, a learning rate of 0.09, and a maximum depth limit of 18 layers.

5. The Parkinson's disease early identification system based on multi-task learning as described in claim 1, characterized in that: The frame-level visual perception unit includes an upper processing pathway to fully perceive the local Parkinson's pathological features of the Parkinson's patient, and a lower processing pathway to fully perceive the whole-body Parkinson's pathological features. In the above processing path, the initial feature processing operation is first performed through two sequentially connected SFEnet modules. Then, the regional pooling unit performs regional adaptive pooling operation on the obtained data feature map. The regional pooling unit first divides the output feature map of the SFEnet module into multiple feature subspaces, then performs adaptive average pooling operation on each feature subspace, and finally concatenates the pooling operation results of each sub-region to obtain the global feature vector. At the same time, a feature difference calculation layer is used to calculate the feature differences between the left and right parts of the body, thereby incorporating the asymmetric feature information of Parkinson's disease; The feature difference calculation layer first divides the input feature map into two symmetrical regions, left and right, along the vertical axis of the image center, and extracts the feature map of the left and right sides respectively. Then, it performs element-by-element comparison on the feature points at corresponding positions on the left and right sides, calculates their absolute difference in each channel, and finally extracts the symmetrical difference features that reflect the degree of inconsistency between the left and right sides of the body movements of Parkinson's patients. Subsequently, residual connections are introduced to perform channel concatenation of the asymmetric feature information and the output feature information of two sequentially connected SFEnet modules. After further feature processing using the SFEnet module, the feature is obtained. ; In the processing pathway, the SFEnet module is first introduced to perform preliminary feature processing operations to improve the ability to identify pathological features throughout the body. Then, the stage module of the OpenPose network is introduced to extract heatmaps of key human points and their spatial topology, capturing abnormal manifestations in limb stability and tremor control in Parkinson's patients. Based on the extracted skeleton graph structure embedding representation, a pose structure map is constructed through the graphsage layer, and a graph convolutional layer is introduced to model the spatial relationships between skeleton graph nodes, enhancing the global coordination representation between actions. Finally, a graph attention layer is introduced to dynamically adjust the information propagation intensity between key nodes, amplifying the core pathological pose signal to obtain features. ; Finally, the features and characteristics The data is fed into a convolutional layer for final processing, and the ReLU activation function is used to complete the final activation to obtain the frame-level image output features. .

6. The Parkinson's disease early identification system based on multi-task learning as described in claim 5, characterized in that: The cross-frame symptom fusion unit targets the input frame-level image feature sequence. By using an additive injection time-aware method, features are output to each frame-level image. Summation and addition of learnable temporal coding vectors , Temporal frame-level features are obtained. Subsequently, a gated attention mechanism was introduced to calculate the ill-conditioned feature weights corresponding to each temporal frame-level feature. ; Then, the calculated ill-conditioned feature weights were used. and temporal frame-level features The dynamic features of Parkinson's disease across frames are obtained by calculating the corresponding cumulative summation. .

7. The Parkinson's disease early identification system based on multi-task learning as described in claim 5, characterized in that: The Parkinson's disease identification head is based on the Parkinson's disease cross-frame dynamic features calculated by the cross-frame symptom fusion unit. And the numerical fusion features of Parkinson's symptoms calculated by the multimodal symptom fusion unit. First, the two features are normalized using a channel normalization layer. After normalization, channel concatenation is performed. The concatenated features are then processed through two interconnected convolutional layers. Next, a linear transformation layer with Tanh activation is introduced to enhance non-linear expressiveness. Finally, a double fully connected layer with residual connections is used for further processing. The activation function process yields the final identification result of Parkinson's disease symptoms. ,in This indicates that the person is not ill. This indicates an illness with a severity level of 1. This indicates that the illness is of severity level 2. This indicates an illness with a severity level of 3. This indicates an illness with a severity level of 4. This indicates an illness with a severity level of 5.

8. The Parkinson's disease early identification system based on multi-task learning as described in claim 1, characterized in that: The Parkinson's disease comprehensive discrimination module, based on the early identification model of Parkinson's disease, continuously identifies Parkinson's disease in the test subject over multiple time periods, and integrates the identification results obtained from different time periods to form a Parkinson's disease identification result sequence. ; Based on the calculated Parkinson's disease identification result sequence Spdr, the trend slope of this sequence is calculated by constructing a first-order linear regression fit. If this A positive value indicates that the Parkinson's identification result shows an increasing trend over time, which is consistent with the characteristics of Parkinson's disease development; For sequence Modeling the sequence perturbation, i.e., solving the sequence perturbation. Standard deviation The larger this value, the lower the stability of the Parkinson's disease symptoms in the test subject; the smaller this value, the more stable the Parkinson's disease symptoms in the test subject, that is, the higher the diagnosis rate of Parkinson's disease. The Parkinson's disease identification result sequences for the normal population were calculated using the early identification model for Parkinson's disease. And the Parkinson's disease identification result sequence of the confirmed cases. Subsequently, the Spdr sequence for identifying Parkinson's disease in the test subjects and the Spdr sequence for identifying the normal population were calculated respectively. Normal group Mahalanobis distance and the sequence of confirmed case identification results The confirmed case, Ma, is far from ; Subsequently, based on the Mahalanobis distance of the normal group Distance from confirmed case Ma The distance anomaly index is obtained by calculating the ratio. The higher this index, the closer the subject is to the Parkinson's disease group; Based on the calculated distance anomaly index The calculated trend slope and standard deviation , and Summation and Union Calculate the discriminant factor by constructing the final discriminant for Parkinson's disease through difference calculation. ; Set judgment threshold When the discriminant factor Greater than the judgment threshold If the test subject is diagnosed with Parkinson's disease, the severity level is the rounded result of the mean of the severity levels of Parkinson's disease in the Parkinson's disease identification result sequence Spdr.

9. A method for early identification of Parkinson's disease based on the early identification system for Parkinson's disease as described in any one of claims 1 to 8, characterized in that: Deploying an early Parkinson's disease identification system within a hospital workstation system includes the following processes: S1, acquire audio data of the user's voice expression and limb movement sequence data, and use them as data representing symptoms of Parkinson's disease; at the same time, acquire motion image data; S2, based on the user's voice expression audio data and limb movement sequence data collected by the data acquisition module, constructs a medical prior knowledge random forest model and a medical constraint embedded gradient boosting tree model, and combines them with an adaptive weighted logistic regression model to output the Parkinson's disease symptom recognition results. S3 firstly, based on the acquired motion image data, the frame-level visual perception unit performs convolutional feature extraction on each frame of the image, and combines an attention mechanism to enhance the response of facial muscles and key limb regions. Secondly, the cross-frame symptom fusion unit models the temporal features of continuous video frames to obtain cross-frame dynamic features. Finally, the multimodal symptom fusion unit uses a multilayer perception network to perform feature calculation processing on the Parkinson's disease symptom recognition results obtained by the symptom recognition module, and fuses them with the cross-frame dynamic features to guide the network to focus on the image semantic regions related to medical symptoms. Finally, the Parkinson's disease recognition head outputs the early recognition results of Parkinson's disease. S4. Multiple disease probability detections are performed using a short-term identification model for Parkinson's disease under long content frames to obtain the Parkinson's disease factor sequence of the subject. The sequence was then subjected to multi-step mathematical analysis with factor sequences from normal individuals and Parkinson's patients to obtain the standard identification results for Parkinson's disease.

Images