Parkinson's disease intelligent diagnosis and analysis system based on human body postures

Abstract

The invention relates to the technical field of auxiliary medical instruments, and particularly discloses a Parkinson's disease intelligent diagnosis and analysis system based on human body postures, the Parkinson's disease intelligent diagnosis and analysis system comprises an acquisition end, a processing end, an evaluation end and an optimization end, the acquisition end comprises a shooting unit and a data transmission unit, the processing end comprises a data receiving unit and a data processing unit, and the evaluation end is connected with the data receiving unit. The evaluation end comprises an evaluation and scoring unit and a report generation unit, the optimization end comprises a system optimization unit, the model construction and training module comprises a model training stage and a fine tuning stage, the model training stage constructs a positive sample and a negative sample, and the fine tuning stage is used for optimizing a classification task result; through model pre-training and fine tuning, the system can learn feature representation with higher robustness and high discrimination degree, so that more accurate diagnosis and evaluation results are provided in motion evaluation of Parkinson's disease, and the performance and practicability of the model are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of auxiliary medical devices, and particularly relates to an intelligent diagnosis and analysis system for Parkinson's disease based on human postures. Background Art

[0002] Parkinson's disease is a common neurodegenerative disease in the middle-aged and elderly. The lesion site is mainly located in the substantia nigra striatum. It is a disease caused by the degeneration and loss of substantia nigra-dopaminergic neurons and the decrease of dopamine content in the striatum. The typical clinical manifestations are resting tremor, muscle rigidity, bradykinesia, abnormal posture, etc. In addition to extrapyramidal system lesions, Parkinson's disease is mainly manifested by sleep disorders, mental and behavioral abnormalities, depression, and various sensory and autonomic nerve abnormalities that exist before the appearance of typical motor symptoms; in addition to the adverse effects of motor symptoms on the quality of life of PD patients, current clinical research has gradually expanded to non-motor symptoms, which also affect patients in severe cases and greatly affect their quality of life.

[0003] In the Chinese patent with the publication number CN114694830A, a human posture estimation method for the auxiliary diagnosis of Parkinson's disease is mentioned, including the following steps: 1) According to the motor function examination part of the international authoritative Parkinson's disease scale UPDRS, determine the actions that need to be automatically analyzed, and quantify each parameter under the guidance of experts so that the computer can process it; 2) Collect and label the posture data of Parkinson's disease patients and healthy people; 3) Determine the operation process and backbone network of the posture estimation algorithm based on the deep convolutional neural network; 4) Use the collected data to train the neural network; 5) Use the video taken by the patient according to the specified requirements, apply the trained deep convolutional neural network to obtain the key point coordinates, and then calculate the required physiological indicators and the UPDRS action level of this action to achieve human posture estimation. The present invention can accurately identify the postures of Parkinson's disease patients and be used for doctor-assisted diagnosis.

[0004] However, although the above technical solution can estimate the postures of patients, thus facilitating the doctor to diagnose the condition, this method uses the human posture estimation as the diagnosis basis, and the overall operation process is relatively complex, with many estimation processes and requiring multi-project processing, and there are still large errors in the estimation results. Therefore, it needs to be improved by the staff. Summary of the Invention

[0005] The purpose of the present invention is to provide an intelligent diagnosis and analysis system for Parkinson's disease based on human postures to solve the problems raised in the above background art.

[0006] To achieve the above purpose, the present invention provides the following technical solutions:

[0007] An intelligent diagnosis and analysis system for Parkinson's disease based on human body postures, comprising:

[0008] An acquisition end, a processing end, an evaluation end and an optimization end;

[0009] The acquisition end includes a shooting unit and a data transmission unit, the processing end includes a data receiving unit and a data processing unit, the evaluation end includes an evaluation and scoring unit and a report generation unit, and the optimization end includes a system optimization unit;

[0010] The data processing unit includes a video processing module and a model construction and training module. The video processing module is used for pose estimation to extract joint key points and construct an original skeleton sequence, and obtains positive and negative samples of the original SkeletonSequence through image enhancement;

[0011] The model construction and training module includes a model training stage and a fine-tuning stage. In the model training stage, a positive sample and a negative sample are constructed. The positive sample is obtained by data augmentation from the original sample and is similar to the original sample in characteristics. The negative sample has different temporal and spatial characteristics from the original sample respectively. The constructed samples and similarity metrics are used to train the model to enhance spatio-temporal discrimination ability;

[0012] The fine-tuning stage is used to optimize the classification task result. The model parameters obtained in the pre-training stage are used to initialize the model parameters in the fine-tuning stage, and then the model is enhanced through a fully connected layer and a softmax function to perform supervised classification based on the fused joint bone features, further optimizing the performance of the model, especially in specific classification tasks.

[0013] Preferably, the shooting unit is used for the patient to use a shooting device to self-shoot the motor function of Parkinson's disease in a home scene, and the data transmission unit is used to transmit the patient video recorded by the shooting unit to the processing end, and the transmission method is one of wireless transmission and wired transmission.

[0014] Preferably, the data receiving unit is used to receive the video transmitted in the data transmission unit, classify and sort the data according to time and actions after receiving the data, and store the video at the position of the data processing unit.

[0015] Preferably, the evaluation and scoring unit is based on the trained model, predicts and evaluates the patient data, determines whether the patient has Parkinson's disease, and gives corresponding clinical suggestions according to the evaluation results.

[0016] Preferably, the report generation unit is used to generate a comprehensive evaluation report after completing data analysis and scoring. The report not only includes the quantitative score of motor function, but also provides an overall Parkinson's disease risk assessment. The report is designed to provide clear and easy-to-understand information for users to quickly understand their health status.

[0017] Preferably, the system optimization unit is used to continuously optimize and improve the AI detection system, improve its accuracy, efficiency and user-friendliness, and make corresponding adjustments and updates according to the feedback and requirements in actual applications.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0019] (1) Through model pre-training and fine-tuning, the system can learn more robust and discriminative feature representations, thus providing more accurate diagnosis and evaluation results in the motor assessment of Parkinson's disease, and improving the performance and practicality of the model.

[0020] (2) After data analysis and scoring, a comprehensive evaluation report will be generated. The report not only includes the quantitative score of motor function, but also provides an overall Parkinson's disease risk assessment. The report is designed to provide clear and easy-to-understand information for users so that they can quickly understand their health status. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 is the system flow chart of the present invention;

[0022] Figure 2 is the skeleton sequence diagram of the present invention;

[0023] Figure 3 is the model training flow chart of the present invention;

[0024] Figure 4 is the overall framework diagram of the graph convolutional network model of the present invention;

[0025] In the figure: 1. Shooting unit; 2. Data transmission unit; 3. Data receiving unit; 4. Data processing unit; 5. Evaluation and scoring unit; 6. Report generation unit; 7. System optimization unit. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0026] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0027] Embodiment 1:

[0028] Please refer to Figures 1 to 4 shown in the figure, a smart diagnosis and analysis system for Parkinson's disease based on human body posture, comprising:

[0029] An acquisition end, a processing end, an evaluation end and an optimization end;

[0030] The acquisition end includes a shooting unit 1 and a data transmission unit 2, the processing end includes a data receiving unit 3 and a data processing unit 4, the evaluation end includes an evaluation and scoring unit 5 and a report generation unit 6, and the optimization end includes a system optimization unit 7;

[0031] The shooting unit 1 is used to allow the patient to use a shooting device to self-shoot the motor function of Parkinson's disease in a home scenario;

[0032] To ensure the video quality, the user should select an environment suitable for testing:

[0033] 1) Lighting: The lighting should be uniform and sufficient, avoiding overly bright or dark areas. The user can use natural light or indoor lighting, but direct light or shadows should be avoided;

[0034] 2) Background: The background should be simple and non-interfering to reduce the visual noise that may appear in the video. It is recommended to choose a clean wall or use a background cloth;

[0035] 3) Noise: The environment should be relatively quiet to avoid background noise interfering with the video recording. The user can choose to conduct the test during a relatively quiet period of the day or use noise-canceling headphones.

[0036] With the assistance of AI technology, the user only needs to shoot through the camera (an ordinary smartphone can meet the requirements) to realize the daily assessment of the motor function of Parkinson's disease. Doctors can complete the diagnosis process within 3 minutes, and the diagnosis speed is increased by 10 times. Through AI technology, let the patient use an ordinary smartphone to self-shoot and complete the daily evaluation of the motor function of Parkinson's disease in a home scenario, saving a large amount of time for patients and doctors to visit or follow up;

[0037] The data processing unit 4 includes a video processing module and a model construction and training module. The video processing module is used for pose estimation to extract joint key points and construct an original skeleton sequence, and obtain positive and negative samples of the original SkeletonSequence through image enhancement;

[0038] In the SmartPD Evaluator, pose detection is achieved by using Mediapipe's Hand Detector and Body Pose Detector. Mediapipe is an open-source machine learning framework that provides a variety of pre-trained models, including models for hand and body pose detection.

[0039] Hand Detector: Used to detect and track the key points of the hand. Each hand has 21 key points, including the tips of the fingers, knuckles, palm center, etc. The positions of these key points are crucial for analyzing hand movements and tremors.

[0040] Body Pose Detector: Used to detect and track the key points of the body. The human pose is represented by 33 key points, which cover the main joints and bone connection points of the body, such as shoulders, elbows, wrists, knees, and ankles. The positions of these key points are crucial for analyzing body movements and postural control. The specific positions are as Figure 2 shown.

[0041] The key point coordinates extracted from multiple video frames form the skeleton sequence of the corresponding human body. This sequence is a time series data that records the changes in the positions of key points over time when the user performs a motion task. The skeleton sequence is the key to understanding the user's motion pattern because it captures the dynamic characteristics of the user's actions.

[0042] After the data is uploaded to the server or processed on the local device, the application will use a Graph Convolutional Network (GCN) to analyze the video data. GCN is a deep learning model specifically designed to process graph-structured data, and it can effectively capture and analyze the complex patterns of human movements.

[0043] In the feature extraction stage, GCN will perform the following key steps:

[0044] 1) Pose estimation: The system first uses pose estimation techniques, such as Mediapipe's Hand Detector and Body Pose Detector, to identify the human key points in the video. These key points include 21 key points (Hand LandMarks) of the hand and 33 key points (Pose LandMarks) of the body;

[0045] 2) Skeleton sequence construction: The key point coordinates extracted from multiple video frames are used to construct the skeleton sequence (Skeleton Sequence) of the human body. This sequence represents the motion pattern of the human body over time.

[0046] 3) Image Enhancement: To improve the generalization ability of the model, the system generates positive and negative samples through image enhancement techniques. For example, the first or last frame of the original Skeleton Sequence is translated left or right or mirrored to create negative samples.

[0047] 4) Model Classification: After the skeleton sequence construction and image enhancement, the system uses a Graph Convolutional Network (GCN) to process this data. GCN is a neural network specifically designed to process graph-structured data, and it can effectively analyze and learn the complex relationships in the skeleton sequence. Through the input skeleton sequence, GCN classifies the actions performed by the current user into 5 levels, which may represent different stages from normal to severe Parkinson's disease symptoms. By learning the features extracted from the data, GCN can identify the movement patterns related to Parkinson's disease, such as the location, type, and severity of tremors.

[0048] Through these steps, the system can extract features related to Parkinson's disease motor symptoms, such as the location, type, and severity of tremors, from video data. These features are then used to evaluate the user's motor function and give corresponding scores.

[0049] Since the loss function of the behavior detection model uses Triplet Loss, which requires the original skeleton sequence, a positive sample, and a negative sample to form a triplet. The negative sample is obtained by image enhancement of the Anchor, such as translating the first or last frame of the extracted Skeleton Sequence left or right, or mirroring it left or right.

[0050] The model construction and training module includes a model training stage and a fine-tuning stage. In the model training stage, a positive sample and a negative sample are constructed. The positive sample is obtained by data augmentation from the original sample and is similar to the original sample in features. The negative sample has different temporal and spatial features from the original sample respectively. The constructed samples and similarity metrics are used to train the model to enhance spatio-temporal discrimination.

[0051] (1) For positive sample construction, a data augmentation scheme is designed to preserve the original motion semantics through the augmentation process. Augmented samples are created by applying random similarity transformations to each original video frame.

[0052] (2) The generation of spatially disordered negative samples aims to provide samples with the same temporal context but different spatial contexts as the anchor sample, which helps the model learn to distinguish differences in spatial features. The following are the steps to generate spatially disordered negative samples:

[0053] 1) Spatial global disorder: Apply global spatial transformations to the anchor samples, such as random translation, rotation, or scaling, while keeping the temporal order of the video unchanged. These transformations simulate the situation of performing the same action at different spatial positions.

[0054] 2) Local spatial transformation: Apply local spatial transformations to specific regions in the video frames, such as rotating or scaling only a part of the body while keeping the other parts unchanged. This helps the model learn to distinguish the differences in local spatial features.

[0055] (3) Consistent with the spatial global disorder scheme, use the temporal global disorder scheme to generate temporal negative samples with normal spatial context (but random temporal context) compared to the anchor samples, which helps the model learn to distinguish the differences in temporal features. The following are the steps to generate temporally disordered negative samples:

[0056] 1) Temporal global disorder: Apply global temporal transformations to the anchor samples, such as randomly rearranging the order of the video frames while keeping the spatial positions unchanged, simulating the situation of performing actions in different orders at the same spatial position.

[0057] 2) Local temporal transformation: Apply local temporal transformations to specific time periods in the video frames, such as stretching or compressing only a part of the video temporally while keeping the other parts in normal temporal order, which helps the model learn to distinguish the differences in local temporal features.

[0058] Graph Convolutional Neural Network

[0059] In the model, the human joints and bone structures are represented by a graph structure, and the graph structure is represented in the form of an adjacency matrix. If nodes i and j in the adjacency matrix are connected, the corresponding adjacency matrix element is set as: When there are action changes, different parameter matrices are specified for each convolutional layer of the grid by introducing a mask matrix; this mask matrix can mine and add action-related joint connections based on the natural connections of the human body, enabling effective aggregation of the spatial features of the human skeleton.

[0060] During the model pre-training stage, based on the self-supervised graph neural network framework, it aims to enhance the discriminability of spatio-temporal fine-grained features. The graph neural network can better represent the structures of joints and bones, enhancing the similarity of the output features of the joint flow and bone flow.

[0061] The fine-tuning stage is used to optimize the results of the classification task. The model parameters obtained in the pre-training stage are used to initialize the model parameters in the fine-tuning stage, and then the model is enhanced through fully connected layers and the softmax function to perform supervised classification based on the fused joint-bone features, further optimizing the performance of the model, especially in specific classification tasks.

[0062] Through model pre-training and fine-tuning, the system can learn more robust and discriminative feature representations, thus providing more accurate diagnosis and evaluation results in the motor assessment of Parkinson's disease. These steps are crucial for improving the performance and practicality of the model.

[0063] The data transmission unit 2 is used to transmit the patient video recorded by the shooting unit 1 to the processing end. The transmission method is one of wireless transmission and wired transmission. The data receiving unit 3 is used to receive the video transmitted in the data transmission unit 2. After receiving the data, the data is sorted according to time and action, and the video is stored at the position of the data processing unit 4.

[0064] Based on the trained model, the evaluation and scoring unit 5 predicts and evaluates the patient data, determines whether the patient has Parkinson's disease, and gives corresponding clinical suggestions according to the evaluation results.

[0065] By analyzing the video action data of the patient, the motor function of the patient will be quantitatively evaluated, including aspects such as walking, balance, and posture control; the trained model will be used to identify movement abnormalities and Parkinson's disease characteristics, as shown in the following table:

[0066]

[0067]

[0068] At the same time, each index is divided into 0-4 quantitative levels, as shown in the following table:

[0069] Level Definition Impact 0 Normal or unobstructed / 1 Minor impairment Does not affect daily functions 2 Mild impairment Has a certain impact on daily functions 3 Moderate impairment Significantly affects daily functions 4 Severe impairment Severe functional limitation

[0070] In addition, "UR" (Unable to Rate) may also be used to represent the situation that cannot be evaluated.

[0071] Users need to complete these tasks according to the instructions, and at the same time ensure that the camera can clearly capture their actions. When the user completes the corresponding actions, the application will detect in real time whether the actions are standard. If the actions seriously deviate from the action requirements, the application will give corresponding prompts.

[0072] Based on the above evaluation steps, a comprehensive Parkinson's disease evaluation report will be generated. The report will provide a prediction of the likelihood of Parkinson's disease in the patient and provide treatment suggestions and management strategies for individual patients for doctors.

[0073] The report generation unit 6 is used to generate a comprehensive evaluation report after completing data analysis and scoring. The report not only includes the quantitative score of motor function, but also provides an overall Parkinson's disease risk assessment. The design of the report provides clear and easy-to-understand information for users to quickly understand their health status.

[0074] The main content of the report includes:

[0075] 1) Quantitative scoring: List in detail the scores for each motor task and the comprehensive score obtained according to the MDS-UPDRS guidelines;

[0076] 2) Risk assessment: Based on the scoring results, provide an assessment of the risk of Parkinson's disease, which may include low, medium, and high risk levels;

[0077] 3) Chart display: Use charts and graphs to visually display the scoring results, such as bar charts, line charts, etc., to help users better understand their motor function status;

[0078] 4) Trend analysis: If the user has been evaluated before, the report will also include trend analysis to show changes in the user's health status;

[0079] 5) Personalized recommendations: Based on the assessment results, provide personalized health recommendations and suggestions for lifestyle adjustments.

[0080] The following aspects are designed to facilitate users' interpretation of the report:

[0081] 1) Read the abstract: First, read the abstract part of the report to understand the overall assessment results and risk levels.

[0082] 2) View the charts: View the charts and graphs to understand the scores of each motor task and how they affect the overall score.

[0083] 3) Understand the scores: Understand the meaning of each score level and how they reflect the severity of Parkinson's disease.

[0084] 4) Areas of concern: Identify the areas that need attention pointed out in the report, which are aspects where users may need further examinations or consultations with doctors.

[0085] 5) Consult a doctor: If the report recommends further medical consultations, users should discuss the assessment results with a doctor and take actions according to the doctor's advice.

[0086] The system optimization unit 7 is used to continuously optimize and improve the AI detection system, enhance its accuracy, efficiency, and user-friendliness, and make corresponding adjustments and updates based on the feedback and requirements in actual applications.

[0087] Collect more diverse data of Parkinson's disease patients and continuously expand the training dataset to improve the generalization ability of the model, enabling it to better adapt to the characteristics and manifestations of different patients, including:

[0088] 1) Patient data collection: Collaborate with more medical institutions to collect data of Parkinson's disease patients of different ages, genders, races, and disease stages;

[0089] 2) Data diversity: Ensure that the dataset includes cases of various movement disorders, including patients with mild to severe symptoms, and the performance of different movement tasks;

[0090] 3) Data quality control: Implement strict data quality control processes to ensure that the collected data is accurate, reliable, and complies with ethical standards.

[0091] Model training and optimization

[0092] To further improve the performance of the system, model training and optimization are key steps:

[0093] 1) Model structure optimization: Explore different neural network architectures, such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), and graph convolutional networks (GCNs), to find the model structure most suitable for the Parkinson's disease assessment task;

[0094] Use GCN to extract the features of the nodes in the graph. GCN is composed of multiple stacked graph convolutional layers. Given the node features of the l-th layer, with the help of graph convolutional operations, the node features of the (l + 1)-th layer are:

[0095]

[0096] where: σ(·) is the activation function; I is the identity matrix, denotes the degree matrix of; X (l) is the node feature output by the l-th layer; W (l) is the parameter matrix of GCN.

[0097] The entire end-to-end graph convolutional network model is established by stacking multiple GCN layers and inserting SsrPool layers. The overall framework of the model is as Figure 4 shown. Specifically, the model regards a GCN SsrPool layer as a whole unit. By stacking GCN SsrPool layers, it continuously aggregates the node information of the graph until the representation of the graph is obtained. Each time a GCN layer and an SsrPool layer are passed, higher-level information aggregation is achieved on the basis of the original input, and the node feature information of the entire graph at different levels is retained through the readout module, so as to capture the local structure of the graph and the node feature information. Finally, the obtained graph-level representation is sent to a multi-layer perceptron classifier to predict the class label of the graph.

[0098] 2) Parameter adjustment: Through experiments and cross-validation, adjust the hyperparameters of the model, such as the learning rate, batch size, and optimizer type, to improve the training efficiency and prediction accuracy of the model.

[0099] 3) Improvement of learning algorithms: Research and implement new learning algorithms, such as transfer learning, meta-learning, and reinforcement learning, to improve the learning ability and adaptability of the model.

[0100] Most existing transfer learning techniques focus on minimizing the difference in marginal probability distributions of two-domain data to solve the domain shift problem between two-domain data. However, in actual application scenarios, the difference in conditional probability distributions is as common as the difference in marginal probability distributions, and the combined effect of the two will lead to differences in the distribution of positioning data collected at different times. To achieve more sufficient domain adaptation and thus minimize domain shift to the greatest extent, this paper considers minimizing the distribution differences of both probabilities while calculating the adaptation of the overall feature distribution of the data.

[0101] Marginal probability distribution adaptation

[0102] Since the probability distribution difference between two-domain data sets cannot be directly observed, for the convenience of calculation, the Maximum Mean Discrepancy (MMD) is introduced as the distance function for calculating the difference in marginal probability distributions. The expression for the difference in marginal probability distributions between two domains after mapping is:

[0103]

[0104] where, n s and n t represent the sample numbers of source domain data and target domain data respectively, M T represents the mapping matrix, represents the input data matrix composed of source domain and target domain data, H represents the reproducing Hilbert space, tr(·) represents solving the matrix, is the MMD matrix and can be constructed in the following way.

[0105]

[0106] Conditional probability distribution adaptation

[0107] Since the target domain lacks labeled samples, the difference in conditional probability distributions between two-domain data cannot be directly calculated. By training a base classifier with the source domain data X src , the target domain data X tar can be predicted, and the soft labels of X tar are obtained for subsequent calculations and multiple iterations are performed during the calculation process. Since the calculation of posterior probability is large, the sufficient statistic of the class-conditional probability distribution is used to approximate the conditional probability, and the specific conversion method is as follows:

[0108]

[0109] Then, modify the above MMD to measure the difference in class-conditional probability distributions between the two domains, so as to approximately represent the magnitude of the difference in conditional probability distributions.

[0110]

[0111] Among them, \(c\in\{1,2,\cdots,C\}\) represents the category of the label, and represents the samples of class \(c\) in the source domain data, represents the samples of class \(c\) in the target domain data, and respectively represent the numbers of samples of class \(c\) in the two domains, is the weight matrix, and the construction method is as follows:

[0112]

[0113] Among them, and respectively represent the prior class probabilities of class \(c\) on the source domain and the target domain.

[0114] Through these technical improvements, the system will be able to capture the characteristics of Parkinson's disease more accurately and provide more reliable evaluation results.

[0115] The feedback from users is an important source for the continuous improvement of the system. To ensure that the system can meet the needs of users, the following measures are taken:

[0116] 1) User feedback mechanism: Establish a user feedback system that allows users to easily report problems, make suggestions, or share their usage experiences;

[0117] 2) Regular updates: According to user feedback and market trends, regularly update the functions and interfaces of the application to improve the user experience;

[0118] 3) Performance monitoring: Implement a performance monitoring system to track the running status of the application and user satisfaction in real time, so as to discover and solve problems in a timely manner.

[0119] Through the above technical methods and processes, provide more convenient and accurate diagnosis and evaluation tools for doctors and patients, reduce the need for manual operations, and improve the detection efficiency and accuracy.

[0120] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirits of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent diagnosis and analysis system for Parkinson's disease based on human body postures, characterized in that, Including: A collection end, a processing end, an evaluation end, and an optimization end; The collection end includes a shooting unit (1) and a data transmission unit (2), the processing end includes a data receiving unit (3) and a data processing unit (4), the evaluation end includes an evaluation and scoring unit (5) and a report generation unit (6), and the optimization end includes a system optimization unit (7); The data processing unit (4) includes a video processing module and a model construction and training module. The video processing module is used for pose estimation to extract joint key points and construct an original skeleton sequence, and obtains positive and negative samples of the original SkeletonSequence through image enhancement; The model construction and training module includes a model training stage and a fine-tuning stage. In the model training stage, a positive sample and a negative sample are constructed. The positive sample is obtained by data augmentation from the original sample and is similar to the original sample in characteristics. The negative samples have different temporal and spatial characteristics from the original sample respectively. The constructed samples and similarity metrics are used to train the model to enhance spatio-temporal discrimination; The fine-tuning stage is used to optimize the classification task result. The model parameters obtained in the pre-training stage are used to initialize the model parameters in the fine-tuning stage, and then the model is enhanced through a fully connected layer and a softmax function to perform supervised classification based on the fused joint bone features, further optimizing the performance of the model, especially in specific classification tasks.

2. The intelligent diagnosis and analysis system for Parkinson's disease based on human body postures according to claim 1, wherein: The shooting unit (1) is used for the patient to use a shooting device to self-shoot the motor function of Parkinson's disease in a home scenario. The data transmission unit (2) is used to transmit the patient video recorded by the shooting unit (1) to the processing end, and the transmission method is one of wireless transmission and wired transmission.

3. The intelligent diagnosis and analysis system for Parkinson's disease based on human body postures according to claim 1, wherein: The data receiving unit (3) is used to receive the video transmitted in the data transmission unit (2). After receiving the data, the data is sorted according to time and action, and the video is stored at the position of the data processing unit (4).

4. The intelligent Parkinson's disease diagnosis and analysis system based on human body postures according to claim 1, characterized in that: The evaluation and scoring unit (5) is based on the trained model to predict and evaluate the patient data, judge whether the patient has Parkinson's disease, and give corresponding clinical suggestions according to the evaluation results.

5. The intelligent diagnosis and analysis system for Parkinson's disease based on human body postures according to claim 1, wherein: The report generation unit (6) is used to generate an evaluation report after completing data analysis and scoring. The report includes a quantitative score of the motor function and a Parkinson's disease risk assessment.

6. The intelligent diagnosis and analysis system for Parkinson's disease based on human body postures according to claim 1, wherein: The system optimization unit (7) is used to continuously optimize and improve the AI detection system, improve its accuracy, efficiency, and user-friendliness, and make corresponding adjustments and updates according to the feedback and requirements in actual applications.

Images