A method and device for evaluating the severity of parkinson's disease by feature fusion

Abstract

The application discloses a Parkinson disease severity evaluation method and device based on feature fusion, and the method comprises the following steps: inputting a training set of a target patient into a bidirectional long short-term memory network trained by all patient data to obtain shared parameter features; selecting a Markov boundary of a unified Parkinson disease rating scale (UPDRS) of the target patient from the shared parameter features by using an incremental association Markov boundary algorithm; inputting the training set of the target patient into a bidirectional long short-term memory network trained by target patient data to obtain specific parameter features; fusing the Markov boundary of the UPDRS and the specific parameter features, and inputting the same into a full connection layer network to obtain the Parkinson disease severity of the target patient. The above technical scheme learns information from other tasks to make up for the insufficient amount of target patient data, and effectively improves the accuracy of Parkinson disease severity prediction.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and specifically to a method and apparatus for assessing the severity of Parkinson's disease by feature fusion. Background Technology

[0002] Remote monitoring has gained significant attention due to its ability to provide frequent, objective, and personalized assessments of Parkinson's disease severity. The challenge lies in the high heterogeneity among Parkinson's disease patients, who vary in demographics, genetic risk factors, comorbidities, stages, and treatment regimens. This heterogeneity necessitates the development of specific symptom assessment models for each patient. However, the limited data available for each patient often significantly impacts the predictive accuracy of individual symptom assessment models. Therefore, leveraging data from other Parkinson's patients to compensate for the limited data available for specific patients has become a pressing issue. Summary of the Invention

[0003] The purpose of this invention is to provide a method and apparatus for assessing the severity of Parkinson's disease by feature fusion. By learning information from other tasks, it makes up for the lack of data on the target patient and effectively improves the accuracy of predicting the severity of Parkinson's disease.

[0004] To achieve the above objectives, the following technical solution was adopted:

[0005] In a first aspect, the present invention provides a method for assessing the severity of Parkinson's disease based on feature fusion, comprising:

[0006] The training set of the target patient is input into a bidirectional long short-term memory network trained using data from all patients to obtain the shared parameter features;

[0007] The incremental correlation Markov boundary algorithm is used to select the Markov boundary of the Unified Parkinson's Disease Rating Scale (UPDRS) for the target patient from the shared parameter features;

[0008] The training set of the target patients is input into a bidirectional long short-term memory network trained using the target patient data to obtain the specific parameter features;

[0009] The Markov boundary of the UPDRS is fused with the specific parameter features, and then input into a fully connected layer network to obtain the severity of Parkinson's disease in the target patient.

[0010] Optionally, the training set of the target patient is input into a bidirectional long short-term memory network trained using data from all patients to obtain the shared parameter features, including:

[0011] The dataset D = {X, Y} contains data from n Parkinson's disease patients, where X = {X1, ..., X...} i , ..., X n Let} represent the feature data of n patients, Y = {Y1, ..., Y2}. i , ..., Y n} represents the severity score of Parkinson's disease symptoms for n patients. Specifically, the severity score of Parkinson's disease symptoms is quantitatively assessed using the Unified Parkinson's Disease Rating Scale (UPDRS); X i and Y i Let X represent the feature data and UPDRS score of the i-th patient, respectively; optionally, select the feature data X of the i-th patient from D = {X, Y}. i And UPDRS score Y i Data for the target patients;

[0012] Employing a hard-sharing mechanism for parameters in multi-task learning, patients with Parkinson's disease in X are considered as multiple related tasks. The data for each patient in X are divided into training, validation, and test sets. A pre-trained bidirectional long short-term memory (BSL-M) network is obtained by training the BSL-M network using the training set of all patients and fine-tuning its hyperparameters using the validation set of all patients. Multiple tasks share the hidden layers of this network. The BSL-M network has L layers, and the dimension of the hidden state in the k-th layer is V. k , k∈[1,L], the trained network parameters are represented as θ X ;

[0013] The training set of the target patients is input into a pre-trained bidirectional long short-term memory network, and the shared parameter features of the target patients are calculated according to the following formula:

[0014]

[0015] Among them, X i Divided into training set, validation set and test set, f represents the training set of the target patients. X (·) indicates a bidirectional long short-term memory network pre-trained using training and validation sets from all patients. This refers to the shared parameter features obtained after inputting the training set of the target patient into a pre-trained bidirectional long short-term memory network; Where M represents the number of samples in the training set of the target patient, V Ldenoted as the dimension of the hidden state of the Lth layer of a bidirectional long short-term memory network pre-trained using the training and validation sets of all patients;

[0016] Optionally, an incremental correlation Markov boundary algorithm is used to select the Markov boundary of the Unified Parkinson's Disease Rating Scale (UPDRS) for the target patient from the shared parameter features, including:

[0017] In the forward phase of the incremental correlated Markov boundary algorithm, the shared parameter features of the target patient are calculated according to the following formula. The correlation strength z between the features in column p and the UPDRS score for Parkinson's disease symptom severity. p :

[0018]

[0019]

[0020] Where, p∈[1, 2V] L ], r p express The Pearson correlation coefficient between the p-th column feature and the UPDRS of the target patient. and Y j They represent The feature value in the j-th row and p-th column of the target patient and the UPDRS score of the j-th sample in the training set of the target patient. express The average value of the features in column p. This represents the average UPDRS score of all samples in the training set of the target patient; according to The correlation strength between all features and UPDRS is used to obtain the feature f′ with the highest correlation strength;

[0021] Calculate the P-value of the feature f′ with the strongest association. If the P-value of feature f′ is less than a set threshold β, then feature f′ is not independent of the UPDRS score of the target patient, and feature f′ is added to the Markov boundary set B of the Unified Parkinson's Disease Rating Scale (UPDRS) for the target patient. i In the middle; traversal calculation The strength of association between other features and the UPDRS of the target patient, and the corresponding p-value, until... If the P-value of a featureless target patient is less than the threshold β, the forward phase of the incremental association Markov boundary algorithm ends, and the resulting Markov boundary set of the target patient's UPDRS score is B. i ;

[0022] In the backward phase of the incremental associated Markov boundary algorithm, at B iUnder the condition of \f, calculate whether the P-value of each feature f is greater than the set threshold β, where f∈B i B i \f indicates that f will be removed from B. i The deleted feature set; if the p-value of feature f is greater than the set threshold β, then feature f is independent of the UPDRS score of the target patient and should be removed from B. i Delete feature f; if the p-value of feature f is not greater than the set threshold β, then feature f is not independent of the UPDRS score of the target patient, and feature f should be retained in B. i In the middle; calculate B in sequence according to the above steps. i The independence between each feature f and the UPDRS score of the target patient is determined, until all features B are traversed. i For each feature, the backward phase of the incremental association Markov boundary algorithm ends, and the updated B... i Markov boundaries for the Unified Parkinson's Disease Rating Scale (UPDRS) for the final target patients;

[0023] Optionally, the training set of the target patients is input into a bidirectional long short-term memory network trained using the target patient data to obtain the specific parameter features, including:

[0024] The training set of the target patient is input into a pre-trained bidirectional long short-term memory neural network according to the following formula to obtain the specific parameter feature F. i :

[0025]

[0026] Where f(·) represents a bidirectional long short-term memory network pre-trained using the training and validation sets of the target patient, the network having L′ layers and the hidden state dimension of the k′-th layer being V′. k′ , k′∈[1,L′], the trained network parameters are represented as θ, Among them, V′ L′ This represents the dimension of the hidden states in the L′ layer of f(·).

[0027] Optionally, the Markov boundary of the UPDRS is fused with the specific parameter features, and then input into a fully connected layer network to obtain the severity of Parkinson's disease in the target patient, including:

[0028] The Markov boundary B of the target patient's UPDRS is determined according to the following formula. i With the specific parameter feature F i To merge:

[0029]

[0030] in, Indicates the concatenation operation, F i ′ represents the Markov boundary B of the UPDRS of the target patient. i and the specific parameter feature F i The fused feature set;

[0031] The fused feature set F i Inputting l fully connected layers yields the severity of Parkinson's disease symptoms in the target patient, where the m-th fully connected layer contains l neurons. m , m∈[1,l].

[0032] Secondly, the present invention provides a feature-fusion-based Parkinson's disease severity assessment device, applied to the above-mentioned method, comprising:

[0033] The first acquisition module is configured to input the training set of the target patient into a bidirectional long short-term memory network trained using all patient data to acquire the shared parameter features.

[0034] The selection module is configured to use the incremental correlated Markov boundary algorithm to select the Markov boundary of the Unified Parkinson's Disease Rating Scale (UPDRS) for the target patient from the shared parameter features.

[0035] The second acquisition module is configured to input the training set of the target patient into a bidirectional long short-term memory network trained using the target patient data to acquire the specific parameter features;

[0036] The fusion module is configured to fuse the Markov boundary of the UPDRS with the specific parameter features, input the fusion into a fully connected layer network, and obtain the severity of Parkinson's disease in the target patient.

[0037] Due to the adoption of the above technical solution, the beneficial effects achieved by this invention are as follows:

[0038] This invention provides a feature fusion-based method for assessing the severity of Parkinson's disease. By treating all Parkinson's patients as multiple related tasks, it allows each target patient to learn complementary information from data from other tasks, thus alleviating the problem of limited data for the target patient. First, a bidirectional long short-term memory (LSTM) network is established for the target patient to obtain unique information about their condition. Then, the LSM network is trained using the training and validation sets of all Parkinson's patients in the dataset, fully exploring the intrinsic correlations among multiple patients. The learned intrinsic correlations are represented by shared parameter features. Next, an incremental associative Markov boundary algorithm is used to select the Markov boundary for the severity score of the target patient's condition from the shared parameter features. Finally, based on the unique information about the target patient's condition, the Markov boundary information obtained from other patient data is fused to assess the severity of the target patient's Parkinson's disease. By learning information from other tasks, the limited data for the target patient is compensated for, effectively improving the accuracy of predicting the severity of Parkinson's disease. Attached Figure Description

[0039] Figure 1 A flowchart illustrating a method for assessing the severity of Parkinson's disease based on feature fusion according to an embodiment of the present disclosure is shown.

[0040] Figure 2 A structural block diagram of a feature-fusion-based Parkinson's disease severity assessment device according to an embodiment of the present disclosure is shown. Detailed Implementation

[0041] In the following, exemplary embodiments of the present disclosure will be described in detail with reference to the accompanying drawings to enable those skilled in the art to readily implement them. Furthermore, for clarity, portions unrelated to the description of the exemplary embodiments have been omitted from the drawings.

[0042] In this disclosure, it should be understood that terms such as “comprising” or “having” are intended to indicate the presence of features, figures, steps, behaviors, components, parts or combinations thereof disclosed in this specification, and are not intended to exclude the possibility of the presence or addition of one or more other features, figures, steps, behaviors, components, parts or combinations thereof.

[0043] It should also be noted that, unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other. This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0044] Remote monitoring has gained significant attention due to its ability to provide frequent, objective, and personalized assessments of Parkinson's disease severity. The challenge lies in the high heterogeneity among Parkinson's disease patients, who vary in demographics, genetic risk factors, comorbidities, stages, and treatment regimens. This heterogeneity necessitates the development of specific symptom assessment models for each patient. However, the limited data available for each patient often significantly impacts the predictive accuracy of individual symptom assessment models. Therefore, leveraging data from other Parkinson's patients to compensate for the limited data available for specific patients has become a pressing issue.

[0045] This disclosure is made to at least partially solve the problems in the prior art that the inventors have discovered.

[0046] Figure 1 A flowchart illustrating a feature fusion-based method for assessing the severity of Parkinson's disease according to an embodiment of this disclosure is shown, such as... Figure 1 As shown, the method for assessing the severity of Parkinson's disease based on feature fusion may include the following steps S101-S104:

[0047] In step S101, the training set of the target patient is input into a bidirectional long short-term memory network trained using all patient data to obtain the shared parameter features;

[0048] In step S102, the incremental correlation Markov boundary algorithm is used to select the Markov boundary of the Unified Parkinson's Disease Rating Scale (UPDRS) for the target patient from the shared parameter features.

[0049] In step S103, the training set of the target patient is input into the bidirectional long short-term memory network trained using the target patient data to obtain the specific parameter features;

[0050] In step S104, the Markov boundary of the UPDRS is fused with the specific parameter features and input into a fully connected layer network to obtain the severity of Parkinson's disease in the target patient.

[0051] This disclosure provides a feature fusion-based method for assessing the severity of Parkinson's disease. By treating all Parkinson's patients as multiple related tasks, it allows each target patient to learn complementary information from data from other tasks, thus alleviating the problem of limited data for the target patient. First, a bidirectional long short-term memory (LSTM) network is established for the target patient to obtain unique information about their condition. Then, the LSM network is trained using the training and validation sets of all Parkinson's patients in the dataset, fully exploring the intrinsic correlations among multiple patients. The learned intrinsic correlations are represented by shared parameter features. Next, an incremental associative Markov boundary algorithm is used to select the Markov boundary for the severity score of the target patient's condition from the shared parameter features. Finally, based on the unique information about the target patient's condition, the Markov boundary information obtained from other patient data is fused to assess the severity of the target patient's Parkinson's disease. By learning information from other tasks, the limited data for the target patient is compensated for, effectively improving the accuracy of predicting the severity of Parkinson's disease.

[0052] According to embodiments of this disclosure, step S101, which involves inputting the training set of the target patient into a bidirectional long short-term memory network trained using data from all patients, to obtain the shared parameter features, includes:

[0053] A1: The dataset D = {X, Y} contains data from n Parkinson's disease patients, where X = {X1, ..., X...} i , ..., X n Let} represent the feature data of n patients, Y = {Y1, ..., Y2}. i , ..., Y n} represents the severity score of Parkinson's disease symptoms for n patients. Specifically, the severity score of Parkinson's disease symptoms is quantitatively assessed using the Unified Parkinson's Disease Rating Scale (UPDRS); X i and Y i Let X represent the feature data and UPDRS score of the i-th patient, respectively; select the feature data X of the i-th patient from D = {X, Y}. i And UPDRS score F i The data refers to the target patient; for example, n can take the value 42.

[0054] A2: Employing a hard-sharing mechanism for parameters in multi-task learning, patients with Parkinson's disease in X are considered as multiple related tasks. The data for each patient in X is divided into training, validation, and test sets. A pre-trained bidirectional long short-term memory (BSL-MS) network is obtained by training the BSL-MS network using the training set of all patients and tuning its hyperparameters using the validation set of all patients. Multiple tasks share the hidden layers of this network. The BSL-MS network has L layers, and the hidden state dimension of the k-th layer is V. k , k∈[1,L], the trained network parameters are represented as θ X For example, if L can take the value 3, then k∈[1,3], and V1, V2 and V3 can take the values ​​128, 64 and 32 respectively.

[0055] A3: Input the training set of the target patient into a pre-trained bidirectional long short-term memory network, and calculate the shared parameter features of the target patient according to the following formula:

[0056]

[0057] Among them, X i Divided into training set, verification set, and test set. f represents the training set of the target patients. X (·) indicates a bidirectional long short-term memory network pre-trained using training and validation sets from all patients. This refers to the shared parameter features obtained after inputting the training set of the target patient into a pre-trained bidirectional long short-term memory network; Where M represents the number of samples in the training set of the target patient, V L Let F3 represent the hidden state dimension of the Lth layer of a bidirectional long short-term memory network pre-trained using the training and validation sets of all patients; for example, L can be 3, F3 can be 32, and M can be 100.

[0058] According to embodiments of this disclosure, step S102, which involves using the incremental correlated Markov boundary algorithm to select the Markov boundary of the Unified Parkinson's Disease Rating Scale (UPDRS) for the target patient from the shared parameter features, includes:

[0059] B1: In the forward phase of the incremental association Markov boundary algorithm, the shared parameter features of the target patient are calculated according to the following formula. The correlation strength z between the features in column p and the UPDRS score for Parkinson's disease symptom severity. p :

[0060]

[0061]

[0062] Where, p∈[1, 2V] L ], r p express The Pearson correlation coefficient between the p-th column feature and the UPDRS of the target patient. and Y j They represent The feature value in the j-th row and p-th column of the target patient and the UPDRS score of the j-th sample in the training set of the target patient. express The average value of the features in column p. This represents the average UPDRS score of all samples in the training set of the target patient; according to The correlation strength between all features and UPDRS is used to obtain the feature f′ with the highest correlation strength; for example, L can take the value 3, V3 can take the value 32, then p∈[1, 64]; assuming The association strengths z1, z2, ..., z of the 64 features in the middle. 64 The values ​​are -1.3, 0.3, ..., 1.2, respectively, with the maximum correlation strength being z. 29 =1.4, then f′ is Features of column 29 in the middle.

[0063] B2: Calculate the P-value of the feature f′ with the strongest association. If the P-value of feature f′ is less than a set threshold β, then feature f′ is not independent of the UPDRS score of the target patient, and feature f′ is added to the Markov boundary set B of the Unified Parkinson's Disease Rating Scale (UPDRS) for the target patient. i In the middle; traversal calculation The strength of association between other features and the UPDRS of the target patient, and the corresponding p-value, until... If the P-value of a featureless target patient is less than the threshold β, the forward phase of the incremental association Markov boundary algorithm ends, and the resulting Markov boundary set of the target patient's UPDRS score is B. i For example, the P-value of feature f′ can be 0.001, and the threshold β can be 0.05. Because the P-value of feature f′ is less than the set threshold β, feature f′ is added to B. i In the middle; assuming Complete the traversal according to steps B1 and B2. After the 64 columns of features, the forward stage of the incremental association Markov boundary algorithm obtains B. i If B can contain 30 columns of features, then B i ∈R 100×30.

[0064] B3: In the backward phase of the incremental associated Markov boundary algorithm, in B i Under the condition of \f, calculate whether the P-value of each feature f is greater than the set threshold β, where f∈B i B i \f indicates that f will be removed from B. i The deleted feature set; if the p-value of feature f is greater than the set threshold β, then feature f is independent of the UPDRS score of the target patient and should be removed from B. i Delete feature f; if the p-value of feature f is not greater than the set threshold β, then feature f is not independent of the UPDRS score of the target patient, and feature f should be retained in B. i In the middle; calculate B in sequence according to the above steps. i The independence between each feature f and the UPDRS score of the target patient is determined, until all features B are traversed. i For each feature, the backward phase of the incremental association Markov boundary algorithm ends, and the updated B... i Markov boundaries for the Unified Parkinson's Disease Rating Scale (UPDRS) for the final target patients; for example, assuming B i ∈R 100×30 The threshold β is 0.05. If B i The feature f in column 3 of B i Given the condition that P is 0.06, then B i The feature f in column 3 will be from B i Delete B at this time i ∈R 100×29 And so on, traversing B in step B3 until all B is completed. i After all features are identified, the backward phase of the incremental associated Markov boundary algorithm yields B. i The data can contain 11 columns of features, then the final B i ∈R 100×11 Markov boundaries for the Unified Parkinson's Disease Rating Scale (UPDRS) for the target patients.

[0065] According to embodiments of this disclosure, step S103, which involves inputting the training set of the target patient into a bidirectional long short-term memory network trained using the target patient data to obtain the specific parameter features, includes:

[0066] The training set of the target patient is input into a pre-trained bidirectional long short-term memory neural network according to the following formula to obtain the specific parameter feature F. i :

[0067]

[0068] Where f(·) represents a bidirectional long short-term memory network pre-trained using the training and validation sets of the target patient, the network having L′ layers and the hidden state dimension of the k′-th layer being V′. k′ , k′∈[1,L′], the trained network parameters are represented as θ, Among them, V′ L′ Let F represent the dimension of the hidden state at layer L′ in f(·); for example, L′ can be 2, then k′∈[1,2], V1 and V2 can be 32 and 16 respectively, and M can be 100. Then the specific parameter feature F... i ∈R 100×32 .

[0069] According to embodiments of this disclosure, step S104, which involves fusing the Markov boundary of the UPDRS with the specific parameter features and inputting it into a fully connected layer network to obtain the severity of Parkinson's disease in the target patient, includes:

[0070] D1: The Markov boundary B of the target patient's UPDRS is determined according to the following formula. i With the specific parameter feature F i To merge:

[0071]

[0072] in, Indicates the concatenation operation, F i ′ represents the Markov boundary B of the UPDRS of the target patient. i and the specific parameter feature F i The fused feature set; for example, assuming the Markov boundary B of the UPDRS. i ∈R 100×11 The specific parameter feature F i ∈R 100×32 Then B i and F i F after fusion i ′∈R 100×43 .

[0073] D2: The fused feature set F i Inputting l fully connected layers yields the severity of Parkinson's disease symptoms in the target patient, where the m-th fully connected layer contains l neurons. m m∈[1, l]; for example, l can take the value 3, then m∈[1, 3], l1, l2 and l3 can take the values ​​32, 16 and 1 respectively; assuming the fused feature set F i ′∈R 100×43By inputting three fully connected layers, we can obtain the UPDRS scores for the severity of Parkinson's disease symptoms for these 100 samples.

[0074] Figure 2 A structural block diagram of a feature-fusion-based Parkinson's disease severity assessment device according to an embodiment of the present disclosure is shown. This device can be implemented as part or all of an electronic device through software, hardware, or a combination of both. Figure 2 As shown, the Parkinson's disease severity assessment device 200 with feature fusion includes a first acquisition module 201, a selection module 202, a second acquisition module 203, and a fusion module 204.

[0075] The first acquisition module 201 is configured to input the training set of the target patient into a bidirectional long short-term memory network trained using all patient data to acquire the shared parameter features;

[0076] The selection module 202 is configured to use the incremental correlation Markov boundary algorithm to select the Markov boundary of the Unified Parkinson's Disease Rating Scale (UPDRS) for the target patient from the shared parameter features.

[0077] The second acquisition module 203 is configured to input the training set of the target patient into a bidirectional long short-term memory network trained using the target patient data to acquire the specific parameter features;

[0078] The fusion module 204 is configured to fuse the Markov boundary of the UPDRS with the specific parameter features, and input it into a fully connected layer network to obtain the severity of Parkinson's disease in the target patient.

[0079] This disclosure discloses a feature fusion-based method for assessing the severity of Parkinson's disease. By treating all Parkinson's patients as multiple related tasks, it allows each target patient to learn complementary information from data from other tasks, thus alleviating the problem of limited data for the target patient. First, a bidirectional long short-term memory (LSTM) network is established for the target patient to obtain unique information about their condition. Then, the LSM network is trained using the training and validation sets of all Parkinson's patients in the dataset to fully explore the intrinsic correlations among multiple patients. The learned intrinsic correlations are represented by shared parameter features. Next, an incremental associative Markov boundary algorithm is used to select the Markov boundary for the severity score of the target patient's condition from the shared parameter features. Finally, based on the unique information about the target patient's condition, the Markov boundary information obtained from other patient data is fused to assess the severity of the target patient's Parkinson's disease. By learning information from other tasks, the limited data for the target patient is compensated for, effectively improving the accuracy of predicting the severity of Parkinson's disease.

[0080] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure.

[0081] The above description is merely a preferred embodiment of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features disclosed in this disclosure that have similar functions.

Claims

1. A feature fusion-based method for assessing the severity of Parkinson's disease, characterized in that, include: The training set of the target patients is input into a bidirectional long short-term memory network trained using data from all patients to obtain shared parameter features; The incremental correlation Markov boundary algorithm is used to select the Markov boundary of the Unified Parkinson's Disease Rating Scale (UPDRS) for the target patient from the shared parameter features; The step of employing the incremental correlated Markov boundary algorithm to select the Markov boundary of the Unified Parkinson's Disease Rating Scale (UPDRS) for the target patient from the shared parameter features includes: In the forward phase of the incremental correlated Markov boundary algorithm, the shared parameter features of the target patient are calculated according to the following formula. The first in The strength of the association between column features and the UPDRS score for the severity of Parkinson's disease symptoms : in, , express The first in The Pearson correlation coefficient between the column features and the UPDRS of the target patients. and They represent The Middle Line number The eigenvalues ​​of the column and the first eigenvalue in the target patient training set UPDRS score for each sample express The Middle The average value of column features, This represents the average UPDRS score of all samples in the training set of the target patient; according to The feature with the strongest correlation strength is obtained by analyzing the correlation strength between all features and UPDRS. ; Calculate the feature with the strongest correlation. The p-value when the feature The P-value is less than the set threshold. Then the feature The characteristics are not independent of the target patient's UPDRS score. Markov boundary set of the Unified Parkinson's Disease Rating Scale (UPDRS) for the target patients In the middle; traversal calculation The strength of association between other features and the UPDRS of the target patient, and the corresponding p-value, until... P-values ​​with no features are less than the threshold. Then the forward phase of the incremental correlation Markov boundary algorithm ends, and the resulting Markov boundary set of the target patient's UPDRS score is: ; In the backward phase of the incremental correlated Markov boundary algorithm, Under the condition of, calculate each feature Is the P-value greater than the set threshold? ,in, , Indicates will from The feature set after deletion; if features The P-value is greater than the set threshold. Then the feature Independent of the UPDRS score of the target patient, it should be from Deleting features If features The P-value is not greater than the set threshold. Then the feature The characteristics are not independent of the target patient's UPDRS score. Should be retained In the middle; calculate in sequence according to the above steps. Each feature Independence between the target patient's UPDRS score and the target patient's score, until all patients have been traversed. For each feature, the backward phase of the incremental association Markov boundary algorithm ends, and the updated... Markov boundaries for the Unified Parkinson's Disease Rating Scale (UPDRS) for the final target patients; The training set of the target patients is input into a bidirectional long short-term memory network trained using the target patient data to obtain specific parameter features; The Markov boundary of the UPDRS is fused with the specific parameter features, and then input into a fully connected layer network to obtain the severity of Parkinson's disease in the target patient.

2. The method for assessing the severity of Parkinson's disease based on feature fusion according to claim 1, characterized in that, The step of inputting the training set of the target patient into a bidirectional long short-term memory network trained using data from all patients to obtain the shared parameter features includes: Dataset Includes Data from a number of Parkinson's disease patients, among whom... express Individual patient characteristic data express The severity scores of Parkinson's disease symptoms for each patient were specifically assessed using the Unified Parkinson's Disease Rating Scale (UPDRS). and They represent the first Characteristic data and UPDRS scores of each patient; Select The Middle Characteristic data of each patient and UPDRS score Data for the target patients; Employing a hard-sharing mechanism for parameters in multi-task learning, Patients with Parkinson's disease were considered as having multiple related tasks; The data for each patient in the study were divided into training, validation, and test sets. The bidirectional long short-term memory (BLS) network was trained using the training set from all patients, and its hyperparameters were tuned using the validation set from all patients, resulting in a pre-trained BLS network. Multiple tasks shared the hidden layers of this network. The number of layers in the BLS network was [number missing]. , No. The dimension of the hidden state of the layer is , The trained network parameters are represented as ; The training set of the target patients is input into a pre-trained bidirectional long short-term memory network, and the shared parameter features of the target patients are calculated according to the following formula: Among them, Divided into training set, verification set, and test set. This represents the training set of the target patients. This indicates a bidirectional long short-term memory network pre-trained using training and validation sets from all patients. This refers to the shared parameter features obtained after inputting the training set of the target patient into a pre-trained bidirectional long short-term memory network; ,in, This indicates the number of samples in the training set for the target patient. This represents the first bidirectional long short-term memory network pre-trained using training and validation sets from all patients. The dimension of the hidden state of the layer.

3. The method for assessing the severity of Parkinson's disease by feature fusion according to claim 1, characterized in that, The step of inputting the training set of the target patients into a bidirectional long short-term memory network trained using the target patient data to obtain the specific parameter features includes: The training set of the target patient is input into a pre-trained bidirectional long short-term memory neural network according to the following formula to obtain the specific parameter features. : in, This refers to a bidirectional long short-term memory network pre-trained using the training and validation sets of the target patients, the network having the following number of layers: , No. The dimension of the hidden state of the layer is , The trained network parameters are represented as , ,in, express The Middle The dimension of the hidden state of the layer.

4. The method for assessing the severity of Parkinson's disease by feature fusion according to claim 1, characterized in that, The process of fusing the Markov boundary of the UPDRS with the specific parameter features and inputting it into a fully connected layer network to obtain the severity of Parkinson's disease in the target patient includes: The Markov boundary of the target patient's UPDRS is determined according to the following formula. With the specific parameter features To merge: in, This indicates a splicing operation. This represents the Markov boundary of the UPDRS of the target patient. and the specific parameter features The fused feature set; The fused feature set enter A fully connected layer is used to obtain the severity of Parkinson's disease symptoms in the target patient, wherein the first... The number of neurons in a fully connected layer is , .

5. A feature-fusion-based device for assessing the severity of Parkinson's disease, characterized in that, include: The first acquisition module is configured to input the training set of the target patient into a bidirectional long short-term memory network trained using all patient data to acquire shared parameter features. The selection module is configured to use the incremental correlated Markov boundary algorithm to select the Markov boundary of the Unified Parkinson's Disease Rating Scale (UPDRS) for the target patient from the shared parameter features. The second acquisition module is configured to input the training set of the target patient into a bidirectional long short-term memory network trained using the target patient data to acquire specific parameter features; The fusion module is configured to fuse the Markov boundary of the UPDRS with the specific parameter features, and input it into a fully connected layer network to obtain the severity of Parkinson's disease in the target patient. The step of employing the incremental correlated Markov boundary algorithm to select the Markov boundary of the Unified Parkinson's Disease Rating Scale (UPDRS) for the target patient from the shared parameter features includes: In the forward phase of the incremental correlated Markov boundary algorithm, the shared parameter features of the target patient are calculated according to the following formula. The first in The strength of the association between column features and the UPDRS score for the severity of Parkinson's disease symptoms : in, , express The first in The Pearson correlation coefficient between the column features and the UPDRS of the target patients. and They represent The Middle Line 1 The eigenvalues ​​of the column and the first eigenvalue in the target patient training set UPDRS score for each sample express The Middle The average value of column features, This represents the average UPDRS score of all samples in the training set of the target patient; according to The feature with the strongest correlation strength is obtained by analyzing the correlation strength between all features and UPDRS. ; Calculate the feature with the strongest correlation. The p-value when the feature The P-value is less than the set threshold. Then the feature The characteristics are not independent of the target patient's UPDRS score. Markov boundary set of the Unified Parkinson's Disease Rating Scale (UPDRS) for the target patients In the middle; traversal calculation The strength of association between other features and the UPDRS of the target patient, and the corresponding p-value, until... P-values ​​with no features are less than the threshold. Then the forward phase of the incremental correlation Markov boundary algorithm ends, and the resulting Markov boundary set of the target patient's UPDRS score is: ; In the backward phase of the incremental correlated Markov boundary algorithm, Under the condition of, calculate each feature Is the P-value greater than the set threshold? ,in, , Indicates will from The feature set after deletion; if features The P-value is greater than the set threshold. Then the feature Independent of the UPDRS score of the target patient, it should be from Deleting features If features The P-value is not greater than the set threshold. Then the feature The characteristics are not independent of the target patient's UPDRS score. Should be retained In the middle; calculate in sequence according to the above steps. Each feature Independence between the target patient's UPDRS score and the target patient's score, until all patients have been traversed. For each feature, the backward phase of the incremental association Markov boundary algorithm ends, and the updated... Markov boundaries for the Unified Parkinson's Disease Rating Scale (UPDRS) for the final target patients.

Images