Parkinson's disease auxiliary diagnosis and treatment system based on multi-task continuous learning algorithm

Abstract

The invention relates to a Parkinson's disease condition auxiliary diagnosis and treatment system based on a multi-task continuous learning technology, and the system comprises the steps: collecting the voice data of a patient through a bone conduction microphone, carrying out the preprocessing and feature extraction, and carrying out the disease evaluation through a multi-task learning algorithm; the model of each patient not only considers common characteristics, but also can adapt to differences among individuals, so that more accurate personalized prediction is provided; according to the multi-task continuous learning algorithm, a high-accuracy model can be trained only by providing a small amount of data for each patient, and the difficulty and cost of data collection are reduced; the system can feed back personalized Parkinson's disease assessment results only by simple pronunciation of a user, so that the Parkinson's disease management accuracy and efficiency are improved, and better user experience is brought to patients.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and more specifically, to a Parkinson's disease assisted diagnosis and treatment system based on a multi-task continuous learning algorithm. Background Art

[0002] With the rapid progress in the field of Artificial Intelligence (AI), researchers are exploring its potential applications in medical diagnosis. AI can assist doctors or directly serve patients, providing rapid and accurate medical assessments and diagnostic recommendations. This technology not only improves the accuracy and efficiency of diagnosis but also significantly enhances the patient experience.

[0003] Parkinson's Disease (PD) is a common neurodegenerative disease, mainly characterized by bradykinesia, muscle rigidity, resting tremor, and postural gait abnormalities. Dysarthria, as a common early symptom, appears in approximately 90% of patients. Although existing speech technologies are widely used in multiple fields, there are still two major challenges in Parkinson's disease diagnosis: one is that traditional recordings are vulnerable to environmental noise interference, affecting data quality, and removing noise also incurs more costs; the other is that most algorithms lack personalized design, using general machine learning models and failing to fully consider individual differences. These problems limit the application effect of existing technologies in the early detection and condition monitoring of Parkinson's disease. Therefore, it is of great significance to develop a speech diagnosis technology more suitable for Parkinson's disease patients.

[0004] Bone Conduction (BC) technology is a technology that transmits sound through bones, which can significantly reduce the interference of environmental noise and improve the quality of speech signals. In the diagnosis of Parkinson's disease, the application of bone conduction technology can more accurately capture the speech characteristics of patients and provide reliable data support for diagnosis. Given the limitations of traditional recording technology in a noisy environment, bone conduction technology provides an effective solution to this problem.

[0005] Multi-task continuous learning technology can learn the common and individual characteristics among patients during the training process, thereby providing more accurate personalized diagnoses. Utilizing the correlation between tasks, the multi-task continuous learning algorithm can not only improve the prediction accuracy in the case of limited data volume but also fully consider the individual differences of patients and continuously learn to improve the prediction accuracy as the number of clinical diagnoses of patients increases.

[0006] Currently, most Parkinson's disease assisted diagnosis and treatment systems are based on a unified data model, unable to reflect the individual differences of patients. If a separate model is established for each patient, there is a lack of sufficient diagnosis and treatment data, severely limiting the actual application effect of the system. Summary of the Invention

[0007] In view of the above deficiencies in the prior art, the present invention provides a Parkinson's disease assisted diagnosis and treatment system based on a continuous multi-task learning algorithm, which can realize early prediction of Parkinson's disease and assessment of disease progression, has a simple structure, and high prediction accuracy.

[0008] The technical solution adopted by the present invention is as follows:

[0009] A Parkinson's disease assisted diagnosis and treatment system based on bone conduction recording and multi-task continuous learning algorithm, including a bone conduction microphone, a data preprocessing module, a prediction set management module, a feature extraction module, a training set management module, a continuous multi-task learning algorithm, and a user interface module;

[0010] The bone conduction microphone is used to collect the voice data of the patient;

[0011] The data preprocessing module performs preprocessing steps on the collected voice signals, including but not limited to noise reduction, denoising, removing background noise, and accurately extracting useful voice segments;

[0012] The feature extraction module extracts a series of acoustic features from the processed voice data to form an acoustic feature data set;

[0013] The prediction set management module is used to store the patient's self-monitoring of the disease and the recording feature data of the patient who has not been diagnosed and scored by the doctor according to experience and medical tests;

[0014] The training set management module is used to store the patient's recording feature data and the total score (total_UPDRS) made by the doctor for the patient's condition according to experience and medical test results; input the data set with the total score (total_UPDRS) label into the multi-task learning algorithm for training to generate a prediction model;

[0015] Take the voice feature data to be recognized as input and apply it to the personal model trained in advance using the multi-task learning algorithm;

[0016] The multi-task learning model outputs the predicted total score (total_UPDRS) of the patient;

[0017] The multi-task continuous learning algorithm utilizes the correlation between multiple patient data during the training process, can adapt to the voice differences of individual patients, trains a personalized model for each patient, and can update the model as the number of user diagnoses increases.

[0018] The user interface module is used to provide a disease analysis report.

[0019] Preferably, the acoustic features include at least, but are not limited to: Jitter(%), Jitter(Abs), Jitter:RAP, Jitter:PPQ5, Jitter:DDP, Shimmer, Shimmer(dB), Shimmer:APQ3, Shimmer:APQ5, Shimmer:APQ11, Shimmer:DDA, NHR, HNR, RPDE, DFA, and PPE; and each data record also contains the patient's total score (total_UPDRS) as a label.

[0020] A diagnosis and treatment method for a Parkinson's disease assisted diagnosis and treatment system based on bone conduction recording and multi-task continuous learning algorithm, including the following control processes:

[0021] Bone conduction voice acquisition: When the patient goes to the hospital for diagnosis or self-tests at home, the patient emits a specified sound through a general bone conduction microphone and records it 6 times at irregular intervals. The duration and sound intensity of each recording are different; the acquisition device records the patient's voice data and transmits it to the system for storage and processing through the interface module.

[0022] Data preprocessing: Extract useful voice segments by automatically removing parts with a volume lower than the set threshold; the processed data is stored in the database.

[0023] Feature extraction: Call the open-source Praat software to extract various acoustic features from the voice data processed by the preprocessing module.

[0024] According to the Unified Parkinson's Disease Rating Scale, give the total score (total_UPDRS) of the patient's Parkinson's disease at that time as a label for training the total_UPDRS prediction model.

[0025] The features extracted from the voice recorded during the examination and the doctor's diagnostic score are both stored in the training set, and the features extracted from the voice recorded by the patient during self-measurement are stored in the prediction set.

[0026] Data storage: Both the training set and the prediction set are stored in the MySQL database in the form of a relational table, waiting for further analysis and model training.

[0027] Model training: Import the diagnosis and treatment data of accumulated Parkinson's disease patients into the training set of the assisted diagnosis and treatment system in the form of an Excel table.

[0028] The model training module reads the stored training set and calls the multi-task continuous learning algorithm to learn a personalized prediction model for the patient's Parkinson's disease from the training set. After the model is trained, it is automatically saved for the patient's self-monitoring of the disease or the doctor's assisted diagnosis.

[0029] The model training module reads the newly added training data in the training set and calls the multi-task continuous learning algorithm to obtain a more accurate personalized prediction model for the Parkinson's disease condition of the patient. After the model is trained, it is automatically saved for use in the patient's self-monitoring of the condition or the doctor's auxiliary diagnosis.

[0030] Preferably, the specific working steps of the multi-task continuous learning algorithm are as follows:

[0031] (1) The multi-task continuous learning algorithm uses the single-task learning algorithm to learn k task models θ 1 ,..., θ k , and initializes the basis matrix L shared by multiple tasks as a column vector with θ 1 ,..., θ k = [θ 1 ,..., θ k , with a dimension of d×k, where d is the feature dimension and k is the number of columns of the basis matrix;

[0032] Assume that the feature dimension d = 16 and the rank of the basis matrix k = 5, then the dimension of the low-rank matrix L is 16×5;

[0033] The single-task learning algorithm uses the ridge regression calculation formula θ t = (X t X t T + λI) -1 X t Y t . Where θ t is the prediction model of the total_UPDRS score of the patient numbered t learned through the single-task learning algorithm, X t is the matrix composed of the voice features of the patient numbered t, the superscript T is the transpose operation of the matrix, and the superscript -1 is the inverse operation of the matrix. Y t is the total_UPDRS score obtained by the doctor's unified scoring of the patient numbered t according to the UPDRS 3.0 version, I is the identity matrix, and λ is the ridge coefficient;

[0034] Single-task incremental learning: Use the single-task incremental ridge regression learning algorithm θ t = (X t X t T + X ta X ta T + λI) -1 (X t Y t + X ta Y ta ) to learn. Where X tais a matrix composed of newly added voice features for the patient numbered t. The superscript T represents the transpose operation of the matrix, and the superscript -1 represents the inverse operation of the matrix. Y ta is the newly added total_UPDRS score obtained by the doctor's unified scoring of the patient numbered t according to the UPDRS version 3.0.

[0035] (3) According to the shared basis matrix L and the model θ obtained in (1) t , call the open-source sparse coding library to obtain S t ; S t is the coefficient of the disease prediction model for a specific patient;

[0036] (4) Dictionary learning, according to S obtained in (3) t call the open-source dictionary learning library to update the shared basis matrix L; the shared basis matrix L is composed of the single-task models of k patients as column vectors in (1), and is used as the shared basis matrix for the prediction models of multiple patients, and is fine-tuned here through dictionary learning;

[0037] (5) Store L, S, and θ: for continuous learning when a new training set is added next time, where S is the matrix composed of all S t and θ is the matrix composed of all θ t , t is the patient number; the personalized prediction model parameters of the patient numbered t are L×S t ; if the voice detection feature data of a patient forms a vector x, then x×L×S t is the disease prediction result of the patient.

[0038] The model obtained by the multi-task continuous learning algorithm includes the patient common shared basis matrix L and the patient personalized model coefficient S t , and the personalized disease severity prediction model parameters of the patient numbered t are: L×S t ;

[0039] Disease prediction:

[0040] After the voice data self-measured by the patient is extracted by the feature extraction module, it is input into the personalized prediction model trained and saved in the workflow (5), and the model gives the disease prediction result for the patient to self-evaluate the disease;

[0041] Result display:

[0042] The prediction result is displayed to the patient and the doctor through the user interface, including the disease trend and treatment suggestions.

[0043] The beneficial effects of the present invention compared with the prior art:

[0044] The Parkinson's disease assisted diagnosis and treatment system based on bone conduction recording collects the voice data of patients through a bone conduction microphone. After preprocessing and feature extraction, a multi-task learning algorithm is used for disease assessment. The model for each patient not only considers common features but also can adapt to the differences between individuals, thus providing more accurate personalized predictions. The multi-task continuous learning algorithm can train a highly accurate model with only a small amount of data from each patient, reducing the difficulty and cost of data collection. Users only need to make simple pronunciations, and the system can provide personalized Parkinson's disease assessment results, improving the accuracy and efficiency of Parkinson's disease management and bringing a better user experience to patients.

[0045] The present invention combines bone conduction technology and a multi-task learning model to achieve early prediction of Parkinson's disease and assessment of disease progression. The system has a simple structure and high prediction accuracy, and can be widely applied to primary medical institutions and remote areas to provide convenient disease monitoring and assisted diagnosis and treatment tools for patients.

[0046] The Parkinson's disease assisted diagnosis and treatment system based on the multi-task continuous learning algorithm is divided into three role modules: patients, doctors, and administrators. Each module has clear responsibilities and permissions to ensure data security and management convenience. By combining bone conduction technology and a multi-task learning model, the system can more accurately capture the voice features of patients and provide personalized disease assessment and management solutions. Brief Description of the Drawings

[0047] Figure 1 is the function diagram of the Parkinson's disease assisted diagnosis and treatment management system of the present invention;

[0048] Figure 2 is the schematic diagram of the patient's disease development curve. Detailed Embodiments

[0049] The present invention will be described in detail below with reference to the drawings and embodiments:

[0050] Appendix Figure 1 It can be known that a Parkinson's disease assisted diagnosis and treatment system based on bone conduction recording and a multi-task continuous learning algorithm includes a bone conduction microphone, a data preprocessing module, a prediction set management module, a feature extraction module, a training set management module, a continuous multi-task learning algorithm, and a user interface module;

[0051] The bone conduction microphone uses a non-invasive method to collect the voice data of patients;

[0052] Due to its unique characteristic of receiving sound through bone conduction, the external noise collected during bone conduction is very low, and the voice with sound waves can be directly used as a sample, significantly reducing the interference to patients and the processing cost.

[0053] Record the voice of patients pronouncing the vowel 'a'. Each patient records 6 times a day, with different durations and sound intensities for each recording. Eventually, all recorded voice samples will be handed over to the data preprocessing unit for processing.

[0054] The data preprocessing module aims to automatically improve data quality and perform preprocessing steps on the collected voice signals, including but not limited to noise reduction, denoising, removing background noise, and accurately extracting useful voice segments.

[0055] When the initial voice samples are obtained by bone conduction acquisition devices, since such devices themselves have a low noise level, the preprocessing module will automatically remove the parts with a volume lower than the set threshold, achieving the removal of invalid voice parts without manual intervention, ensuring that only the voice data with practical significance is retained. This automated processing flow not only improves the accuracy of subsequent analysis but also significantly enhances the reliability of the final output results.

[0056] The feature extraction module extracts a series of acoustic features from the processed voice data to form an acoustic feature dataset.

[0057] The prediction set management module is used to store the voice recording feature data of patients' self-monitoring of their conditions and the voice recording features of patients who have not yet been diagnosed and scored by doctors based on experience and medical tests.

[0058] By connecting to the doctor scoring interface, doctors directly score these samples, thus continuously enriching the data volume of the training set and improving the accuracy of the model.

[0059] The training set management module is used to store the voice recording feature data of patients and the total score (total_UPDRS) given by doctors for the patients' conditions based on experience and medical test results. Input the dataset with the total score (total_UPDRS) label into the multi-task learning algorithm for training to generate a prediction model.

[0060] Use the voice feature data to be recognized as input and apply it to the personal model previously trained using the multi-task learning algorithm.

[0061] Output the predicted total score (total_UPDRS) of the patient through the multi-task learning model.

[0062] The multi-task continuous learning algorithm utilizes the correlation between multiple patients' data during the training process to improve the generalization ability of the model. At the same time, it can adapt to the voice differences of individual patients, train a personalized model for each patient, and update the model as the number of user diagnoses increases.

[0063] The user interface module supports multi-language output, is suitable for users in different regions of the world, and is used to provide a condition analysis report.

[0064] It includes the disease trend and treatment suggestions to help doctors and patients intuitively understand the disease progress. The system supports the login of three roles: patients, doctors, and administrators, and each role has corresponding permissions to ensure data security and convenient management, further enhancing the practicality and operability of the system. The specific functions of the patient role module include: voice collection, disease prediction, and diagnosis and treatment records, etc.; the specific functions of the doctor role module include: patient information query, auxiliary diagnosis and treatment, and disease diagnosis, etc.; the specific functions of the administrator role module include: user management, model training, statistical reports, etc.

[0065] Preferably, the acoustic features include at least but are not limited to: Jitter(%), Jitter(Abs), Jitter:RAP, Jitter:PPQ5, Jitter:DDP, Shimmer, Shimmer(dB), Shimmer:APQ3, Shimmer:APQ5, Shimmer:APQ11, Shimmer:DDA, NHR, HNR, RPDE, DFA, and PPE; and each data record also contains the patient's total score (total_UPDRS) as a label.

[0066] A Parkinson's disease auxiliary diagnosis and treatment system based on bone conduction recording and multi-task continuous learning algorithm includes the following control processes:

[0067] Bone conduction voice collection: When the patient goes to the hospital for diagnosis or self-tests at home, the patient emits a specified sound (vowel letter 'a') through a general bone conduction microphone, and records it 6 times at irregular intervals, and the duration and sound intensity of each recording are different; recording sound through the bone conduction microphone can avoid the influence of environmental background noise to the greatest extent. The collection device records the patient's voice data and transmits it to the system for storage and processing through the interface module;

[0068] Data preprocessing: Extract useful voice segments by automatically removing the part with a volume lower than the set threshold; the processed data is stored in the database;

[0069] Feature extraction: The feature extraction software module of the auxiliary diagnosis and treatment system calls the open-source Praat software to extract various acoustic features from the voice data processed by the preprocessing module;

[0070] When the patient is clinically examined by a doctor in the hospital, the doctor gives the total score (total_UPDRS) of the patient's Parkinson's disease at that time as a label according to the unified Parkinson's rating scale (version 3.0 of UPDRS) for training the total_UPDRS prediction model.

[0071] When a patient is diagnosed in the hospital, the features extracted from the voice recorded during the examination and the doctor's diagnosis score are both stored in the training set, and the features extracted from the voice recorded by the patient's self-measurement are stored in the prediction set;

[0072] Data storage: Both the training set and the prediction set are stored in the MySQL database in the form of a relational table, waiting for further analysis and model training.

[0073] The system will automatically and continuously accumulate data on the patient's self-detection and the doctor's scoring of the patient's condition diagnosis.

[0074] Model training

[0075] When the auxiliary diagnosis and treatment system is put into use in the hospital, after the administrator enters the account number and password to log in to the auxiliary diagnosis and treatment system, click the "Import Initial Training Data" button on the administrator interface, and import the accumulated diagnosis and treatment data of Parkinson's disease patients into the training set of the auxiliary diagnosis and treatment system in the form of an Excel table;

[0076] After the administrator imports the initial training data, they can click the "Model Training" button. After the administrator clicks the "Initial Training" button, the model training module of the auxiliary diagnosis and treatment system reads the stored training set, calls the multi-task continuous learning algorithm to learn from the training set to obtain a personalized prediction model for the patient's Parkinson's disease condition. After the model is trained, it is automatically saved for use in the patient's self-monitoring of the condition or the doctor's auxiliary diagnosis;

[0077] After the auxiliary diagnosis and treatment system is put into operation, as the number of times the patient goes to the hospital for doctor diagnosis increases and new patients are added, new diagnosis and treatment data will be automatically accumulated and added to the training set.

[0078] The administrator can log in to the auxiliary diagnosis and treatment system at regular intervals and click the "Retrain" button on the administrator interface. The model training module of the auxiliary diagnosis and treatment system reads the newly added training data in the stored training set, calls the multi-task continuous learning algorithm to learn to obtain a more accurate personalized prediction model for the patient's Parkinson's disease condition. After the model is trained, it is automatically saved for use in the patient's self-monitoring of the condition or the doctor's auxiliary diagnosis.

[0079] The multi-task continuous learning algorithm can learn the common and individual characteristics among patients during the training process, improving the generalization ability of the model and the accuracy of personalized diagnosis.

[0080] Preferably, the specific working steps of the multi-task continuous learning algorithm are as follows:

[0081] (1) The multi-task continuous learning algorithm uses a single-task learning algorithm to learn and obtain k task models θ 1 ,..., θ k , and θ 1 ,..., θ kInitialize the basis matrix \(L = [\theta 1 , \cdots, \theta k \) shared by multiple tasks, with dimension \(d\times k\), where \(d\) is the feature dimension and \(k\) is the number of columns of the basis matrix;

[0082] Assume the feature dimension \(d = 16\) and the rank of the basis matrix \(k = 5\), then the dimension of the low-rank matrix \(L\) is \(16\times5\);

[0083] The single-task learning algorithm uses the ridge regression calculation formula \(\theta t =(X t X t T +\lambda I) -1 X t Y t . Where \(\theta t \) is the prediction model of the total UPDRS score of the patient numbered \(t\) learned by the single-task learning algorithm, \(X t \) is the matrix composed of the voice features of the patient numbered \(t\), the superscript \(T\) is the transpose operation of the matrix, and the superscript \(-1\) is the inverse operation of the matrix. \(Y t \) is the total UPDRS score obtained by the doctor's unified scoring of the patient numbered \(t\) according to the UPDRS 3.0 version, \(I\) is the identity matrix, and \(\lambda\) is the ridge coefficient;

[0084] Single-task incremental learning: Use the single-task incremental ridge regression learning algorithm \(\theta t =(X t X t T +X ta X ta T +\lambda I) -1 (X t Y t +X ta Y ta ) for learning. Where \(X ta \) is the matrix composed of the newly added voice features of the patient numbered \(t\), the superscript \(T\) is the transpose operation of the matrix, and the superscript \(-1\) is the inverse operation of the matrix. \(Y ta \) is the newly added total UPDRS score obtained by the doctor's unified scoring of the patient numbered \(t\) according to the UPDRS 3.0 version.

[0085] (3) According to the shared basis matrix \(L\) and the model \(\theta t \) obtained in (1), call the open-source sparse coding library to obtain \(S t \); \(S t \) is the coefficient of the disease prediction model for a specific patient;

[0086] (4) Dictionary learning, according to \(S tCall the open source dictionary learning library to update the shared basis matrix L; the shared basis matrix L is composed of the single-task models of k patients as column vectors in (1), and is used as the shared basis matrix of the prediction model for multiple patients. It is fine-tuned through dictionary learning here;

[0087] (5) Store L, S, and θ: for continued learning when a new training set is added next time, where S is the sum of all S t The matrix composed of θ is all θ t The matrix is ​​composed of t, where t is the patient number; the parameters of the personalized prediction model for the patient numbered t are L×S t ; The patient's speech detection feature data constitutes a vector x, then x×L×S t That is the predicted result of the patient's condition.

[0088] The model learned by the multi-task continuous learning algorithm includes the patient common shared basis matrix L, the patient personalized model coefficient S t , the parameters of the personalized disease severity prediction model for patient number t are: L×S t ;

[0089] Disease prediction:

[0090] The patient's self-measured speech data is extracted by the feature extraction module and then input into the personalized prediction model trained and saved in the workflow (5). The model gives the disease prediction result for the patient to self-evaluate the disease;

[0091] The voice data measured during the diagnosis and examination of the patient in the hospital is input into the trained model after the features are extracted by the feature extraction module. The model gives the prediction results of the disease for the doctor's diagnosis reference.

[0092] The disease prediction result is a total score (total_UPDRS) value reflecting the severity of the patient's Parkinson's disease, which is displayed to the patient and the doctor through the user interface. The doctor can give the actual score of the patient's disease based on clinical experience and other medical testing methods.

[0093] Results:

[0094] The prediction results are presented to patients and doctors through the user interface, including disease trends and treatment recommendations.

[0095] The system provides visual charts to help doctors and patients intuitively understand the progression of the disease.

[0096] User Role:

[0097] 1. Patient:

[0098] Speech was recorded via a bone conduction sound acquisition device.

[0099] View one's own medical condition report and prediction curve.

[0100] One can view a detailed medical condition analysis report on the user interface, including the trend of the medical condition and treatment suggestions.

[0101] 2. Doctor:

[0102] View the medical condition prediction results of the patient, including the development trend of the medical condition and treatment suggestions.

[0103] Diagnose the patient, score the current degree of the patient's medical condition, and adjust the patient's treatment plan.

[0104] One can view the prediction results of the patient in the "assisted diagnosis and treatment" interface for further diagnosis and treatment.

[0105] 3. Administrator:

[0106] Responsible for data management, model training, and system configuration.

[0107] Can perform user management, data import, and model training.

[0108] Regularly check the data quality of the training set and update the model using the multi-task continuous learning algorithm when necessary to improve the prediction accuracy of the model.

[0109] User interface module

[0110] 1. Log in to the system:

[0111] Users of various roles open the system, select the corresponding identities of "patient", "doctor", "administrator", enter the username and password, and click the "Log in" button.

[0112] If the username or password is incorrect, the system will display an error message "Incorrect username / password", and the patient needs to re-enter the correct username and password.

[0113] 2. Voice collection:

[0114] The patient collects voice through the provided bone conduction microphone. The system will record the voice characteristics of the patient, which will be used for subsequent medical condition prediction.

[0115] The collected voice data is transmitted to the server through the system. The server will perform preliminary processing on these data, extract key features, and store them in the prediction set of the MySQL database in the form of a table.

[0116] 3. Medical condition prediction:

[0117] If a specialized model for the patient already exists in the system, the system will use this model to predict the patient's voice data and obtain the total score (total_UPDRS) value that reflects the severity of the patient's Parkinson's disease.

[0118] The prediction results will be displayed in the patient's "My Condition Report" interface, where the patient can view the trend of their condition.

[0119] If there is no specialized model for the patient in the system, the system will notify the doctor for manual evaluation.

[0120] The doctor will view the patient's basic information and voice data in the system and score the patient based on clinical experience. These scores will be part of the training data.

[0121] The data scored by the doctor will be added to the system's training set. Each time the doctor scores a patient, the system will automatically record this data.

[0122] 4. Condition Monitoring and Diagnosis Records

[0123] Patients can view the information of their previous self-condition monitoring and diagnosis records from hospital visits.

[0124] 5. Model Training:

[0125] The administrator clicks the "Import Initial Data" button to import the patient diagnosis and treatment data previously accumulated in this hospital into the training set at one time.

[0126] After successful import, the administrator clicks the "Model Training" button, and the model training module reads the training set to start training the initial model.

[0127] After training is completed, the system will pop up a "Training Successful" prompt box and draw relevant data graphs of the model.

[0128] When the system goes online and runs, it will automatically and continuously accumulate the patient's self-detection data and the data scored by the doctor for the patient's condition diagnosis.

[0129] When the data volume of the training set increases, the administrator can restart the model training. The model training module reads the updated training set and calls the multi-task continuous learning algorithm to train and obtain a personalized prediction model.

[0130] The trained model will be saved in the system for condition prediction during the patient's self-condition check and for auxiliary diagnosis prediction during the doctor's in-person consultation.

[0131] 6. User Management:

[0132] The administrator can add, delete, modify, and query the role information of administrators, doctors, and patients through the user management module.

[0133] 7. Statistical Report:

[0134] Administrators can use the statistical report module to count the information on the diagnosis and treatment of patients by doctors in the hospital, as well as the self - detection and in - hospital consultation information of patients.

[0135] 8. Patient Information Query:

[0136] Doctors can use the patient information query module to query the previous self - monitoring and in - hospital diagnosis and treatment information of the patients they have consulted.

[0137] 9. Auxiliary Diagnosis and Treatment:

[0138] Doctors can request patients to collect voice data on - site and, based on the voice examination data of the patients on the same day, call the prediction model to predict the condition for auxiliary diagnosis and treatment.

[0139] 10. Disease Diagnosis:

[0140] Doctors can, based on their clinical diagnosis experience, the voice examination data of the patients on the same day, and other medical test results, and referring to the prediction results of the auxiliary diagnosis and treatment, give the total score (total_UPDRS) of the disease diagnosis for the patient. The diagnosis and treatment situation of this patient will be recorded in the training set.

[0141] The Parkinson's disease auxiliary diagnosis and treatment system based on bone - conduction recording collects the voice data of patients through a bone - conduction microphone. After pre - processing and feature extraction, a multi - task learning algorithm is used for disease assessment; the model for each patient not only considers common features but also can adapt to the differences between individuals, thus providing more accurate personalized predictions; the multi - task continuous learning algorithm can train a high - accuracy model with only a small amount of data provided by each patient, reducing the difficulty and cost of data collection; users only need to make a simple pronunciation, and the system can provide personalized Parkinson's disease assessment results, improving the accuracy and efficiency of Parkinson's disease management and bringing a better user experience to patients.

[0142] The present invention combines bone - conduction technology and a multi - task learning model to achieve early prediction of Parkinson's disease and assessment of disease progression; the system has a simple structure and high prediction accuracy, and can be popularized and applied to primary medical institutions and remote areas, providing patients with convenient disease monitoring and auxiliary diagnosis and treatment tools.

[0143] The Parkinson's disease auxiliary diagnosis and treatment system based on the multi - task continuous learning algorithm is divided into three role modules: patients, doctors, and administrators. Each module has clear responsibilities and authorities, ensuring data security and convenient management. By combining bone - conduction technology and a multi - task learning model, the system can more accurately capture the voice characteristics of patients and provide personalized disease assessment and management plans.

[0144] The above are only the preferred embodiments of the present invention, and do not impose any formal restrictions on the structure of the present invention. Any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention all fall within the scope of the technical solution of the present invention.

Claims

1. A Parkinson's disease auxiliary diagnosis and treatment system based on bone conduction recording and multi-task continuous learning algorithm, characterized by : It includes bone conduction microphone, data preprocessing module, prediction set management module, feature extraction module, training set management module, continuous multi-task learning algorithm, and user interface module; Bone conduction microphone, used to collect patient voice data; The data preprocessing module performs preprocessing steps on the collected speech signals, including but not limited to noise reduction and denoising, removing background noise and accurately extracting speech segments whose sound signal strength exceeds a threshold; The feature extraction module extracts a series of acoustic features from the processed speech data to form an acoustic feature data set; A prediction set management module, used to store patient self-monitoring condition and patient recording feature data that has not yet been diagnosed and scored by a doctor based on experience and medical tests; The training set management module is used to store the patient's recording feature data and the total score (total_UPDRS) made by the doctor based on the patient's clinical diagnosis based on experience and medical test results; Input the dataset with the total score (total_UPDRS) label into the multi-task continuous learning algorithm for training to generate a prediction model; The speech feature data to be recognized is used as input and applied to the personal model that has been pre-trained using the multi-task continuous learning algorithm; The multi-task learning model outputs the predicted total patient score (total_UPDRS); The multi-task continuous learning algorithm utilizes the correlation between multiple patient data during the training process, and can adapt to the voice differences of individual patients, train a personalized model for each patient, and update the model as the number of user diagnoses increases; The user interface module is used to provide a disease analysis report.

2. The Parkinson's disease auxiliary diagnosis and treatment system based on bone conduction recording and multi-task continuous learning algorithm according to claim 1 is characterized in that : The acoustic features include at least but are not limited to: Jitter (%), Jitter (Abs), Jitter: RAP, Jitter: PPQ5, Jitter: DDP, Shimmer, Shimmer (dB), Shimmer: APQ3, Shimmer: APQ5, Shimmer: APQ11, Shimmer: DDA, NHR, HNR, RPDE, DFA and PPE; and each data record also contains the patient's total score (total_UPDRS) as a label.

3. A Parkinson's disease auxiliary diagnosis and treatment system based on bone conduction recording and multi-task continuous learning algorithm according to claim 1 or 2, characterized in that: The control process includes the following: Bone conduction voice collection: When patients go to the hospital for diagnosis or self-test at home, they use a universal bone conduction microphone to make a specified sound. The sound is recorded 6 times at irregular intervals, and the duration and sound intensity of each recording vary. The collection device records the patient's voice data and transmits it to the system for storage and processing through the interface module. Data preprocessing: extracting valid speech segments by automatically removing the parts with volume below a set threshold; the processed data is stored in a database; If no valid voice segment can be extracted, the system will prompt you to re-record; Feature extraction: calling the open source Praat software to extract multiple acoustic features from the speech data processed by the preprocessing module; According to the Unified Parkinson's Rating Scale, the patient's Parkinson's disease score (total_UPDRS) is given as a label to train the total score prediction model; The features extracted from the recorded speech during the examination and the physician’s diagnostic scores are stored in the training set, and the features extracted from the patient’s self-measured recorded speech are stored in the prediction set; Data storage, both the training set and the prediction set are stored in the MySQL database in the form of relational tables, waiting for further analysis and model training; Model training: import the accumulated diagnosis and treatment data of Parkinson's disease patients into the training set of the auxiliary diagnosis and treatment system in the form of Excel tables; The model training module reads the stored training set and calls the multi-task continuous learning algorithm to learn from the training set to obtain a personalized prediction model for the patient's Parkinson's disease. After the model is trained, it is automatically saved for the patient's self-monitoring of the disease or for the doctor's auxiliary diagnosis. The model training module reads the newly added training data in the stored training set, and calls the multi-task continuous learning algorithm to learn a more accurate personalized prediction model for the patient's Parkinson's disease. After the model is trained, it is automatically saved for the patient's self-monitoring of the condition or the doctor's auxiliary diagnosis.

4. The Parkinson's disease auxiliary diagnosis and treatment system based on bone conduction recording and multi-task continuous learning algorithm according to claim 3 is characterized in that: The specific working steps of the multi-task continuous learning algorithm are as follows: (1) The multi-task continuous learning algorithm uses the single-task learning algorithm to learn k task models θ1, ..., θ k , let θ1,...,θ k Initialize the basis matrix L shared by multiple tasks as a column vector L = [θ1, ..., θ k ], the dimension is d×k, where d is the feature dimension and k is the number of base matrix columns; Assuming the feature dimension d = 16 and the basis matrix rank k = 5, the dimension of the shared basis matrix L is 16 × 5; The single-task learning algorithm uses the ridge regression calculation formula θ t =(X t X t T +λI) -1 X t Y t ; where θ t To learn the total score prediction model of the patient numbered t through the single-task learning algorithm, X t is the matrix composed of the speech features of the patient numbered t, the superscript T is the transpose operation of the matrix, and the superscript -1 is the inverse operation of the matrix; Y t is the total score obtained by the doctor according to the UPDRS version 3.0 for the patient numbered t, I is the unit matrix, and λ is the ridge coefficient; Single-task incremental learning: Use the single-task incremental ridge regression learning algorithm θ t =(X t X t T +X ta X ta T +λI) -1 (X t Y t +X ta Y ta ) study; where X ta is the matrix composed of the newly added speech features of the patient numbered t, the superscript T is the transpose operation of the matrix, and the superscript -1 is the inverse operation of the matrix; Y ta It is the newly added total score obtained by the doctor's unified scoring of the patient numbered t according to the UPDRS version 3.0; (3) Based on the shared basis matrix L and model θ obtained in (1) t , call the open source sparse coding library to find S t ; S t Predict model coefficients for a specific patient's condition; (4) Dictionary learning: Based on S obtained in (3) t Call the open source dictionary learning library to update the shared basis matrix L; the shared basis matrix L is composed of the single-task models of k patients in (1) as column vectors, and is used as the shared basis matrix of the prediction model for multiple patients. It is fine-tuned through dictionary learning here; (5) Store L, S, and θ: for continued learning when a new training set is added next time, where S is the sum of all S t The matrix composed of θ is all θ t The matrix is ​​composed of t, where t is the patient number; the parameters of the personalized prediction model for the patient numbered t are L×S t ; The patient's speech detection feature data constitutes a vector x, then x×L×S t That is, the predicted outcome of the patient's condition; The model learned by the multi-task continuous learning algorithm includes the patient common shared basis matrix L, the patient personalized model coefficient S t , the parameters of the personalized disease severity prediction model for patient number t are: L×S t ; Disease prediction: The patient's self-measured speech data is extracted by the feature extraction module and then input into the personalized prediction model trained and saved in the workflow (5). The model gives the disease prediction result for the patient to self-evaluate the disease; Results: The prediction results are presented to patients and doctors through the user interface, including disease trends and treatment recommendations.

Images