Attention deficit hyperactivity disorder symptom score prediction method based on bipartite hypothesis framework

By adding a symptom score prediction subnetwork to the autoencoder network and combining it with the binary hypothesis framework, the problem of low accuracy in predicting ADHD symptom scores was solved, achieving high accuracy and fast symptom assessment.

CN118942681APending Publication Date: 2024-11-12HOHAI UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411061309.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-02
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

Existing methods for predicting ADHD symptom scores have low accuracy and are greatly affected by human factors, making it difficult to achieve rapid and accurate symptom assessment.

Method used

A symptom score prediction subnetwork is added to the feature extraction stage of the autoencoder network. Symptom scores are predicted using a binary hypothesis framework. The combination of the autoencoder network and the symptom score prediction subnetwork enables objective prediction of symptom scores.

Benefits of technology

It achieves highly accurate prediction of ADHD symptom scores, reduces the influence of human factors, and improves the speed and accuracy of symptom assessment.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118942681A_ABST
    Figure CN118942681A_ABST
Patent Text Reader

Abstract

The invention discloses an attention deficit hyperactivity disorder symptom score prediction method based on a bipartite hypothesis framework, and the method comprises the steps: carrying out different types of label hypotheses for a test sample, and obtaining typical features under different label hypotheses; performing feature extraction on typical features of the test sample and the training sample through a hypothesis network; inputting the test sample high-level features into the symptom sub-prediction sub-network to obtain training sample high-level features with the best clustering effect, taking a test sample label hypothesis corresponding to the test sample high-level features as correct hypothesis estimation of a test sample label, and taking the hypothesis label as an estimation label of a test sample and outputting the hypothesis label; and inputting the candidate symptom scores under different label hypotheses of the test sample and the correct hypothesis estimation of the test sample label into a biological identification detection module together, selecting the label hypothesis of the test sample which is the same as the correct hypothesis estimation, and taking the corresponding candidate symptom score under the hypothesis as a final biological identification and outputting the final biological identification.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a method for predicting symptom scores of attention deficit hyperactivity disorder based on a binary hypothesis framework, and belongs to the field of intelligent auxiliary classification diagnosis of mental illness. Background Art

[0002] Intelligent auxiliary diagnosis methods based on brain imaging data (a variety of biological data generated by magnetic resonance imaging) are an auxiliary diagnosis method based on artificial intelligence and are of great significance in rapid diagnosis. Based on the public brain imaging database of ADHD-200, many intelligent auxiliary diagnosis methods for ADHD have been proposed in recent years. Among them, the auxiliary diagnosis classification method based on the binary hypothesis framework and autoencoding network (Tang, Y., et al, ADHD classification using auto-encoding neural network and binary hypothesis testing, 2022) has received much attention with a high classification accuracy (>95%). This method mainly consists of three parts: feature selection, feature extraction, and final classification, but this method does not achieve the assessment of the severity of ADHD. In actual clinical applications, the observation of ADHD symptoms is also crucial, and symptom scores are obtained through different scales to assess the severity of ADHD. In the implementation of this observation process, it is necessary to integrate the evaluation results of multiple aspects such as doctor observation, parent feedback, and child participation (such as audio-visual testing), which will inevitably be affected by human subjective factors.

[0003] Therefore, it is urgent to start from objective brain imaging data and build the relationship between brain imaging data and symptom scores, so as to achieve a more accurate prediction of symptom scores and achieve a faster and more accurate symptom score evaluation result. However, there are few studies and reports on existing ADHD symptom score prediction methods, such as the linear support vector regression (SVR) method (Song, S., et al, Predicting disease severity in children with combined attention deficit hyperactivity disorder using quantitative features from structural MRI of amygdaloid and hippocampal subfields, 2021), and the prediction accuracy of symptom scores is also low. Summary of the invention

[0004] The purpose of the present invention is to add a symptom score prediction subnetwork to the autoencoder network in the feature extraction stage based on the existing binary hypothesis framework. The symptom scores of the test samples under different label hypotheses are predicted, and in the symptom score judgment module in the symptom score identification stage, the symptom score under the corresponding hypothesis is selected according to the label prediction result of the test sample, and finally the score prediction performance is achieved that is better than the existing ADHD symptom score prediction method.

[0005] The technical solution of the present invention is as follows:

[0006] A method for predicting symptom scores of attention deficit hyperactivity disorder based on a binary hypothesis framework comprises the following steps:

[0007] Step 1: Make different types of label assumptions for the test samples, namely, ADHD disease label assumption H0 and health label assumption H1; input the brain image data of the test samples with label assumptions into the feature selection module together with the brain image data of the training sample set to obtain the typical features (typical brain image data) of the test samples under the disease label assumption H0. And the typical characteristics of the test samples under the healthy label hypothesis H1 And the typical features of the training samples under different label assumptions are also and Among them, the training sample set has been annotated with diagnostic labels;

[0008] Step 2: Compare the typical features of the test samples and training samples By assuming that the network Net H0 Implement feature extraction; extract the typical features of test samples and training samples By assuming that the network Net H1 Implement feature extraction;

[0009] Each hypothesis network contains an autoencoder network (including an encoding subnetwork and a decoding subnetwork), a classification subnetwork and a symptom prediction subnetwork. First, the encoding subnetwork outputs the high-level features of the training samples, which correspond to different label hypotheses of the test samples, denoted as and Then the decoder subnetwork is input for reconstruction, and at the same time, the typical features and high-level features of the test samples under different label assumptions are combined. The symptom score prediction subnetworks under the corresponding hypotheses are input respectively to obtain the candidate symptom scores under the corresponding hypotheses, which are recorded as B0 and B1;

[0010] Step 3: Output the high-level features of the test samples under different hypothesis networks. and Input symptom prediction sub-network, respectively for high-level features and Perform clustering measure scoring calculation to obtain the corresponding clustering measure values, recorded as D0 and D1; compare the numerical values ​​of D0 and D1 through the comparison judgment module, identify the high-level features of the training sample with the best clustering effect, and take the test sample label hypothesis corresponding to the high-level features of the training sample as the correct hypothesis estimate of the test sample label, recorded as At the same time, the hypothesized label is used as the estimated label of the test sample and output, completing the classification process of the test sample;

[0011] Step 4: Estimation of the candidate symptom scores B0 and B1 of ADHD under different label assumptions of the test sample and the correct hypothesis of the test sample label Input into the biomarker detection module together; after biomarker judgment, select and estimate the correct hypothesis The label hypothesis of the same test sample is used, and the candidate symptom score of ADHD corresponding to the hypothesis is used as the final ADHD biomarker and output, thereby completing the ADHD biomarker detection process.

[0012] Furthermore, the symptom score prediction process for the test sample is as follows:

[0013] In the symptom score identification stage, the corresponding candidate symptom prediction scores obtained by the symptom score prediction subnetwork under the label hypothesis H0 and H1 of the test sample are output as B0 and B1; the candidate symptom prediction scores B0 and B1 are input into the symptom score judgment module of the symptom score identification stage; at the same time, using the existing binary hypothesis framework and the ADHD classification diagnosis method based on the autoencoder network, the correct estimation hypothesis corresponding to the test sample label is obtained through the ADHD label prediction stage, which is recorded as In the symptom score judgment module, select The candidate symptom prediction scores corresponding to the same label hypothesis are output as the final symptom prediction scores of the test samples.

[0014] Furthermore, in step 2, the input and output of the symptom score prediction subnetwork are processed as follows:

[0015] Step 21: The encoder subnetwork in the autoencoder network processes the data containing training samples and test samples to obtain high-level features of the samples. At the same time, the feature selection stage generates typical features of these samples. The feature extraction here is performed under a given label hypothesis, i.e., the ADHD disease label hypothesis H0 or the healthy label hypothesis H1;

[0016] Step 22: The high-level features and typical features obtained in step 21 are further screened to ensure that only those features that match the label hypothesis of the current test sample are retained. Whether it is high-level features or typical features, only those sample features that belong to the same category as the current hypothesis label will be used by the symptom score prediction subnetwork;

[0017] Step 23: The high-level features and typical features after screening and integration are sent to the symptom score prediction subnetwork, which outputs the candidate symptom prediction scores of samples including training samples and test samples that are the same as the current label hypothesis.

[0018] Furthermore, the training process of the symptom score prediction sub-network parameters is as follows:

[0019] The symptom score prediction sub-network parameters are only trained on training samples; the error between the candidate symptom prediction score of the training sample and its corresponding true symptom score is required to be minimized; when minimizing, only the symptom score prediction sub-network parameters are updated, and the autoencoder network parameters are not updated.

[0020] Beneficial effects of the present invention:

[0021] After the sample obtains the diagnosis prediction result (whether it is classified as an ADHD patient group or a healthy sample group), the present invention further performs symptom score prediction on the corresponding group, thereby achieving a symptom score prediction result with high prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is the overall framework of the method of the present invention;

[0023] Figure 2 It is the structural framework of the symptom score prediction network of the present invention. DETAILED DESCRIPTION

[0024] The method of the present invention is further described below in conjunction with the accompanying drawings.

[0025] like Figure 1 As shown, a classification diagnosis method for ADHD based on a binary hypothesis framework includes the following steps:

[0026] Step 1: Make different types of label assumptions for the test samples, i.e., label the ADHD disease label assumption H0 and the healthy label assumption H1; input the brain imaging data of the test samples with label assumptions and the brain imaging data of the training sample set (multiple training samples with diagnostic label annotations) into the feature selection module to obtain the typical features (typical brain imaging data) of the test samples under the disease label assumption H0. And the typical characteristics of the test samples under the healthy label hypothesis H1 And the typical features of the training samples under different label assumptions are also and

[0027] Step 2: Compare the typical features of the test samples and training samples By assuming that the network Net H0 Implement feature extraction; extract the typical features of test samples and training samples By assuming that the network Implement feature extraction;

[0028] Each hypothesis network contains an autoencoder network (including an encoding subnetwork and a decoding subnetwork), a classification subnetwork and a symptom prediction subnetwork. First, the encoding subnetwork outputs the high-level features of the training samples, which correspond to different label hypotheses of the test samples, denoted as and Then the decoder subnetwork is input for reconstruction, and at the same time, the typical features and high-level features of the test samples under different label assumptions are combined. The symptom score prediction subnetworks under the corresponding hypotheses are input respectively to obtain the candidate symptom scores under the corresponding hypotheses, which are recorded as B0 and B1;

[0029] Step 3: The high-level features of the training samples output by the test samples under different hypothesis networks are and Input symptom prediction sub-network, respectively for high-level features and Perform clustering measure scoring calculations to obtain the corresponding clustering measure values, recorded as D0 and D1. Through the comparison and judgment module, compare the numerical values ​​of D0 and D1, identify the high-level features of the training sample with the best clustering effect, and take the test sample label hypothesis corresponding to the high-level features of the training sample as the correct hypothesis estimate of the test sample label, recorded as At the same time, the hypothesized label is used as the estimated label of the test sample and output, completing the classification process of the test sample;

[0030] Step 4: Estimation of the candidate symptom scores B0 and B1 of ADHD under different label assumptions of the test sample and the correct hypothesis of the test sample label, Input into the biomarker detection module together; after biomarker judgment, select and estimate the correct hypothesis The label hypothesis of the same test sample is used, and the candidate symptom score of ADHD corresponding to the hypothesis is used as the final ADHD biomarker and output, thereby completing the ADHD biomarker detection process.

[0031] The structure of the symptom prediction subnetwork, such as Figure 2 As shown in the figure, the symptom score prediction sub-network consists of a concatenation layer, three fully connected layers and corresponding activation functions. The symptom score prediction sub-network is a multi-input single-output network.

[0032] The symptom score prediction sub-network parameters are only trained on training samples; the error between the candidate symptom prediction score of the training sample and its corresponding true symptom score is required to be minimized; when minimizing, only the symptom score prediction sub-network parameters are updated, and the autoencoder network parameters are not updated.

[0033] Embodiment 1:

[0034] 1) Experimental conditions

[0035] In this experiment, the resting fMRI blood oxygen concentration level (BOLD) data provided by the ADHD-200 public database was used and matched to the brain template (AAL-116 template) to generate BOLD signals based on brain regions. Subsequently, the Pearson correlation operation was performed on the BOLD signals of the brain regions to obtain the functional connection data between brain regions, and the functional connection data between brain regions were used as brain imaging data to predict ADHD symptom scores based on a bipartite network.

[0036] The various network parameters of the present invention are shown in Table 1:

[0037] Table 1 Input and output size of each network module used in the present invention

[0038] Network Name enter Output Support Vector Machine Recursive Feature Elimination 58×1 35×1 Encoder 35×1 20×1 Decoder 20×1 35×1 Classifier 20×1 2×1 Symptom score prediction subnetwork (35+20)×1 1×1

[0039] The symptom score prediction subnetwork of the present invention inputs typical features and high-level features. Typical features have a size of (35×1), while high-level features have a size of (20×1). These two features are first concatenated together to form a new feature layer with a dimension of (55×1). Next, this new feature layer is operated through a series of fully connected layers and activation functions, each of which has a different output dimension: the first layer outputs (32×1) and the second layer outputs (16×1). Finally, the results of all these operations are passed to the last layer, which is a fully connected layer with an output of (1×1), which is the predicted ADHD symptom score. This design allows the model to comprehensively consider typical features and high-level features in order to more accurately predict ADHD symptom scores.

[0040] The symptom score prediction subnetwork parameters of the present invention are only trained on training samples; the error between the candidate symptom prediction score of the training sample and its corresponding true symptom score is required to be minimized; when minimizing, only the symptom score prediction subnetwork parameters are updated, and the autoencoder network parameters are not updated.

[0041] 2) Experimental content According to the specific experimental parameter settings in Table 1, experiments were conducted on multiple site data of the ADHD-200 database for implementation example 1 to predict symptom scores, and the results are shown in Table 2. The BSP-Net column shows the results of this method.

[0042] Table 2: Comparative analysis of different results

[0043]

[0044] From the results column, we can see that BSP-Net has high accuracy on all datasets, with MSE values ​​below 16 and correlation coefficients between true scores and predicted scores exceeding 0.96. Compared with other methods, BSP-Net effectively utilizes high-level features of training subjects and effectively improves the prediction of symptom scores under the learning true label hypothesis.

[0045] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for predicting symptom scores of attention deficit hyperactivity disorder based on a binary hypothesis framework, characterized in that: The steps include: Step 1: Make different types of label assumptions for the test samples, namely, ADHD disease label assumption H0 and health label assumption H1; input the brain imaging data of the test samples with label assumptions into the feature selection module together with the brain imaging data of the training sample set to obtain the typical features of the test samples under the disease label assumption H0 And the typical characteristics of the test samples under the healthy label hypothesis H1 And the typical features of the training samples under different label assumptions are also and Among them, the training sample set has been annotated with diagnostic labels; Step 2: Compare the typical features of the test samples and training samples By assuming that the network Net H0 Implement feature extraction; extract the typical features of test samples and training samples By assuming that the network Net H1 Implement feature extraction; Each hypothesis network contains an autoencoder network, a classification subnetwork and a symptom prediction subnetwork, where the autoencoder network includes an encoding subnetwork and a decoding subnetwork. First, the encoding subnetwork outputs the high-level features of the training samples, which correspond to different label hypotheses of the test samples, denoted as and Then the decoder subnetwork is input for reconstruction, and at the same time the typical features and high-level features of the test samples under different label assumptions are The symptom score prediction subnetworks under the corresponding hypotheses are input respectively to obtain the candidate symptom scores under the corresponding hypotheses, which are recorded as B0 and B1; Step 3: Output the high-level features of the test samples under different hypothesis networks. and Input symptom prediction sub-network, respectively for high-level features and Perform clustering measure scoring calculation to obtain the corresponding clustering measure values, recorded as D0 and D1; compare the numerical values ​​of D0 and D1 through the comparison judgment module, identify the high-level features of the training sample with the best clustering effect, and take the test sample label hypothesis corresponding to the high-level features of the test sample as the correct hypothesis estimate of the test sample label, recorded as At the same time, the hypothesized label is used as the estimated label of the test sample and output, completing the classification process of the test sample; Step 4: Estimation of the candidate symptom scores B0 and B1 of ADHD under different label assumptions of the test sample and the correct hypothesis of the test sample label Input into the biomarker detection module together; after biomarker judgment, select and estimate the correct hypothesis The label of the same test sample is assumed, and the candidate symptom score of attention deficit hyperactivity disorder corresponding to the assumption is used as the final attention deficit hyperactivity disorder biomarker and output, thereby completing the attention deficit hyperactivity disorder biomarker detection process.

2. The method for predicting ADHD symptom scores based on a binary hypothesis framework according to claim 1, characterized in that: The symptom score prediction process for the test sample is as follows: In the symptom score identification stage, the corresponding candidate symptom prediction scores obtained by the symptom score prediction subnetwork under the label hypothesis H0 and H1 of the test sample are output as B0 and B1; the candidate symptom prediction scores B0 and B1 are input into the symptom score judgment module of the symptom score identification stage; at the same time, using the existing binary hypothesis framework and the ADHD classification diagnosis method based on the autoencoder network, the correct estimation hypothesis corresponding to the test sample label is obtained through the ADHD label prediction stage, which is recorded as In the symptom score judgment module, select The candidate symptom prediction scores corresponding to the same label hypothesis are output as the final symptom prediction scores of the test samples.

3. The method for predicting ADHD symptom scores based on a binary hypothesis framework according to claim 2, characterized in that: In step 2, for the input and output of the symptom score prediction subnetwork, the processing steps are as follows: Step 21: The encoder subnetwork in the autoencoder network processes the data containing training samples and test samples to obtain high-level features of the samples. At the same time, the feature selection stage generates typical features of these samples. The feature extraction here is performed under a given label hypothesis, i.e., the ADHD disease label hypothesis H0 or the healthy label hypothesis H1; Step 22: The high-level features and typical features obtained in step 21 are further screened to ensure that only those features that match the label hypothesis of the current test sample are retained. Whether it is high-level features or typical features, only those sample features that belong to the same category as the current hypothesis label will be used by the symptom score prediction subnetwork; Step 23: The high-level features and typical features after screening and integration are sent to the symptom score prediction subnetwork, which outputs the candidate symptom prediction scores of samples including training samples and test samples that are the same as the current label hypothesis.

4. The method for predicting ADHD symptom scores based on a binary hypothesis framework according to claim 3, characterized in that: The training process of the symptom score prediction subnetwork parameters is as follows: The symptom score prediction sub-network parameters are only trained on training samples; the error between the candidate symptom prediction score of the training sample and its corresponding true symptom score is required to be minimized; when minimizing, only the symptom score prediction sub-network parameters are updated, and the autoencoder network parameters are not updated.