ADHD auxiliary diagnosis model construction method, control equipment and program product

Abstract

The invention discloses an ADHD auxiliary diagnosis model construction method, control equipment and a program product, relates to the field of medical artificial intelligence, and obtains a high-performance ADHD auxiliary diagnosis model through a small sample. The method comprises the following steps: preprocessing acquired sample data, screening out a predetermined number of indexes with the importance higher than that of a diagnosis result in the sample data to obtain a sample set, and dividing the sample set to obtain a training set and a test set; and constructing a logistic regression classifier model, training the logistic regression classifier model by using the training set, and testing the trained logistic regression classifier model by using the test set to obtain the ADHD auxiliary diagnosis model. According to the method, the problem that an ADHD prediction model depends on the sample size is solved, and the feasibility of clinical application is improved.

Description

Technical Field

[0001] The present invention relates to the field of medical artificial intelligence, and in particular to a method for constructing an auxiliary diagnosis model for ADHD (Attention Deficit Hyperactivity Disorder), a control device, and a program product. Background Art

[0002] There are two traditional ways to diagnose ADHD: 1) A diagnostic method that combines scale assessment, behavioral task testing, and physician clinical experience.

[0003] Currently, the diagnosis of ADHD is primarily based on criteria such as the DSM-5 (Diagnostic and Statistical Manual of Mental Disorders) or the ICD-10. Clinicians typically rely on interviews, scale scores, and behavioral observations to make diagnostic decisions.

[0004] 2) A diagnostic method that separates scale assessment from behavioral task test analysis.

[0005] In clinical practice and some research, doctors or researchers usually collect behavioral task test data or standardized scale assessment data for evaluation. The two types of evaluation processes and the results obtained are independent of each other.

[0006] The above traditional diagnostic methods have the following shortcomings: First, it relies heavily on the doctor's accumulated knowledge and experience, and the doctor makes subjective judgments based on experience. However, the diagnostic standards of different doctors are not uniform, which inevitably introduces subjective judgment factors, affecting the objectivity of the diagnostic results, and the consistency of the diagnostic results is poor. In addition, scale assessment data is usually based on long-term behavioral observations or guardians' experience of living with the subjects to quantify the relevant indicators of the subjects, while behavioral task test data reflects the subjects' immediate cognitive behavioral performance under objective experimental conditions. The two differ in terms of time dimension, observation perspective, data structure, etc. This data fragmentation phenomenon has caused the diagnosis and intervention guidance of ADHD to remain at the stage of "experience + data assistance", and it is impossible to fully realize the potential of data integration and analysis.

[0007] As artificial intelligence (AI) technology plays an increasingly prominent role across industries, AI-enabled reasoning models for predicting ADHD are also emerging. Currently, common ADHD prediction models rely heavily on high-dimensional, large-scale datasets such as neuroimaging (fMRI, DTI, EEG) and genomic data. Researchers often use deep learning (such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs)) or more complex machine learning algorithms to build ADHD prediction models.

[0008] Current AI-based ADHD prediction methods rely on learning mechanisms that rely on hundreds or even thousands of sample data points to ensure model generalization. However, in clinical settings, limited by factors such as patient numbers and recruitment difficulties, sample sizes often fall short of these requirements, hindering their clinical application. Furthermore, building and maintaining large-scale sample data sets requires significant human and material resources, resulting in high costs and a long timeline for clinical application. Summary of the Invention

[0009] The object of the present invention is to provide a method and control device for constructing an ADHD auxiliary diagnosis model to address all or part of the above-mentioned problems, so as to obtain a high-performance ADHD auxiliary diagnosis model with only a small number of samples.

[0010] The technical solution adopted in the present invention is as follows: A method for constructing an ADHD auxiliary diagnosis model, comprising: Preprocessing the acquired sample data; the sample data includes multi-dimensional indicators and diagnostic results for ADHD assessment; Screening out a predetermined number of indicators with the highest importance relative to the diagnosis result from the sample data to obtain a sample set, and dividing the sample set into a training set and a test set; A logistic regression classifier model is constructed, the logistic regression classifier model is trained using the training set, and the trained logistic regression classifier model is tested using the test set to obtain an ADHD auxiliary diagnosis model.

[0011] On the other hand, the present application also provides a control device, including a processor and a storage medium, wherein the storage medium stores computer instructions, and the processor runs the computer instructions in the storage medium to execute the above-mentioned ADHD auxiliary diagnosis model construction method.

[0012] In addition, the present application also provides an ADHD auxiliary diagnosis program product, which includes a computer program. When the computer program is executed by a processor, the ADHD auxiliary diagnosis model constructed by the above-mentioned ADHD auxiliary diagnosis model construction method is obtained.

[0013] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: This application sorts and filters the indicator features in the sample data by importance, so that the classifier model can only focus on the key indicators of the sample data to optimize the model parameters, thereby achieving the effect of rapid convergence while ensuring the accuracy of the classification effect. Through this design, the model training can be completed using only a small sample, and the superiority of the model performance is guaranteed. It is generally applicable to usage scenarios of various sizes and improves the feasibility of clinical application. Compared with the method of improving the attention of the model through the attention mechanism, this application does not have additional sample data consumption and processes for adjusting attention, and the required amount of sample data can be further reduced. This application does not require the construction of a complex model architecture, is easy for users to operate, and has a low threshold for use. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The present invention will now be described by way of example with reference to the accompanying drawings, in which: Figure 1 This is a flow chart of the method for constructing an ADHD auxiliary diagnosis model. DETAILED DESCRIPTION

[0015] All features disclosed in this specification, or all steps in the disclosed methods or processes, except mutually exclusive features and / or steps, can be combined in any manner.

[0016] Any feature disclosed in this specification (including any appended claims and abstract), unless otherwise stated, may be replaced by other equivalent or similar features. In other words, unless otherwise stated, each feature is only an example of a series of equivalent or similar features.

[0017] The problem of ADHD diagnosis by doctors based solely on scale assessment data and behavioral task test data is that there are serious subjective factors and poor consistency. Traditional ADHD diagnosis through artificial intelligence modeling requires reliance on large amounts of data for model training, which makes it difficult to apply to clinical application scenarios. In the embodiments of the present application, a method for constructing an ADHD auxiliary diagnosis model, a control device and a program product are provided, which aim to overcome the existing method's reliance on big data and improve its feasibility in clinical applications.

[0018] A method for constructing an ADHD auxiliary diagnosis model, such as Figure 1 As shown, it includes: S1. Preprocess the acquired sample data. The sample data includes multi-dimensional indicators and diagnostic results for ADHD assessment.

[0019] As an optional implementation, the acquired sample data includes scale assessment data and behavioral task test data.

[0020] Sample assessment data refers to the indicators obtained by completing a standard scale, usually in the form of a score. Scale assessment data typically includes indicators for at least one of the SNAP-IV, Conners, and CBCL scales, but may also include indicators for more scales.

[0021] Behavioral task test data is indicator data obtained by having the subject perform a specific task. Unlike scale assessments, behavioral task test data is usually interactive, meaning the subject is given a task to perform, and the indicator data involved in the subject's performance of the task is collected. For example, behavioral task test data includes test data for at least one of the Go / No-Go task, N-Back task, and Stroop task. The so-called test data includes at least one of the following indicators: reaction time (RT), accuracy (ACC), reaction time coefficient of variation (CVReaction), and d-prime. Depending on the task, it may also include integrated efficiency score (IES) and working memory span (WM Span).

[0022] Each sample data item contains the aforementioned scale assessment data and behavioral task test data, as well as the patient / volunteer's diagnosis result. Typically, the diagnosis result is marked as 1 or 0, with 1 indicating ADHD and 0 indicating no ADHD.

[0023] The acquired sample data needs to be preprocessed to improve the data quality. As an optional implementation, the method for preprocessing the sample data includes: 1) Outlier processing.

[0024] Indicators collected in the form of scales or behavioral tasks may not be accurate, and there may be outliers. Processing outliers can avoid or reduce the deviations introduced in the subsequent application of sample data.

[0025] In some feasible implementations, outliers of each indicator in the sample data are processed by deletion or correction.

[0026] Taking the Z-Score outlier detection method as an example, the Z-Score exceeds 3 times the standard deviation. The indicators are trimmed or corrected.

[0027] The Z-Score of the indicator is calculated as follows: , Where, Respectively represent the i-th index in the sample data X Z-Score, the value of the ith index, the mean of the ith index (i.e. the set of the ith index of all sample data), and the standard deviation of the ith index.

[0028] The index whose Z-Score exceeds is regarded as an outlier, which is clipped or corrected. This preprocessing operation can be implemented using the sklearn.preprocessing.StandardScaler class of the scikit-learn library.

[0029] 2) Data standardization.

[0030] The sample data after outlier processing is standardized to unify the feature distribution.

[0031] In some possible embodiments, the sample data can be standardized to have a mean of 0 and a standard deviation of 1 using methods such as StandardScaler of the scikit-learn library.

[0032] 3) Feature merging.

[0033] As mentioned above, the scale assessment data and the behavioral task test data are collected in two different ways. Traditionally, doctors diagnose by making empirical judgments based on these two kinds of data, which results in the fragmentation of data.

[0034] In the embodiments of the present application, the indices of the scale assessment data and the behavioral task test data after standardization are spliced to form a comprehensive index matrix, which is used as the comprehensive feature matrix of the subjects.

[0035] S2, select a predetermined number of indices with higher importance relative to the diagnostic result from the sample data to obtain a sample set, and divide the sample set to obtain a training set and a test set.

[0036] The various indices collected by the sample data may not all be helpful or have greater help for the auxiliary diagnosis of ADHD. In the present application, by selecting indices with higher relative importance for the diagnostic result, limited sample data is concentrated to optimize the parameters of the classifier model, so that the classifier model can more efficiently learn the features of important indices with higher correlation with the diagnostic result, accelerate the model convergence speed, reduce the dependence on the amount of sample data (if all indices are learned, more sample data is needed to make the model converge, and the convergence speed is also slower), so that the present application scheme only needs small sample clinical data to obtain a high-performance (such as high classification accuracy) classifier model, improving the feasibility of clinical use.

[0037] As an optional implementation, a predetermined number of indicators in the sample data are screened out according to their importance relative to the diagnostic result, including the following processes: 1) Assess the importance of each indicator in the sample data relative to the diagnostic result respectively.

[0038] For the assessment of the importance of indicators, in some feasible implementations, the following methods can be used: A) Configure the kernel function, kernel function coefficient gamma, regularization parameter C (also known as penalty parameter), and sensitivity parameter of the support vector regression model SVR.

[0039] The objective function of the SVR model is: , The constraint condition is: , In the above formula, n represents the number of input sample data, respectively represent the weight vector, bias vector, and sensitivity parameter, and respectively are the i-th input vector and output vector of the SVR model, are all the slack variables corresponding to the i-th input vector, represents the L2 norm of w.

[0040] In some specific implementations, the parameters configured for the SVR model are shown in Table 1: Table 1 SVR parameter configuration table

[0041] The configuration code is as follows: from sklearn.svm import SVR, svr_model = SVR(kernel="linear", C=1.0, epsilon=0.1). B) Fit the SVR model with sample data.

[0042] Input the sample data into the SVR model to minimize the objective function for fitting.

[0043] C) Extract the coef coefficient of the SVR model, and sort the importance of the corresponding indicators according to the size of the coef coefficient.

[0044] According to the size of the coef coefficient of the SVR model, the importance of each input feature can be determined, and the importance of the corresponding indicators can be sorted.

[0045] 2) Filter out a predetermined number of indicators that are of high importance.

[0046] By selecting a predetermined proportion or number of indicators that rank high from the various indicators ranked by importance, the key indicators can be screened out.

[0047] The filtering code is as follows: from sklearn.feature_selection import SelectFromModel, feature_selector = SelectFromModel(svr_model, threshold="mean"), feature_selector.fit(X_train, y_train), X_train_selected = feature_selector.transform(X_train), X_test_selected = feature_selector.transform(X_test). For example, there are 50 indicators input into the SVR model, and the final output screened indicators are 10-15 indicators. These 10-15 indicators are the key indicators.

[0048] These selected indicators are then used to train the classifier model. Key indicators are screened from the preprocessed sample data, and the corresponding diagnostic results are retained to form a new sample set. This new sample set is then divided into training and test sets based on a ratio (e.g., 7:3 or 8:2). Both sets are constructed by randomly selecting sample data from the sample set.

[0049] S3. Construct a logistic regression classifier model, train the logistic regression classifier model using the training set, and test the trained logistic regression classifier model using the test set to obtain an ADHD auxiliary diagnosis model.

[0050] 1) The logistic regression classifier model is designed as follows: , In the formula, w is the weight vector, X is the input feature vector, y is the inference result, b is the bias term, It represents the probability that the inference result y is 1 given the input feature vector X.

[0051] 2) The objective function of the logistic regression classifier model is designed as: , Where i is the serial number of the sample data in the training set, which corresponds to the i-th input sample; N is the total amount of sample data in the training set; Indicates the diagnosis result of the i-th sample data, with a value of 1 indicating ADHD and a value of 0 indicating non-ADHD; ; is the regularization coefficient.

[0052] 3) Configure regularization coefficient , for example, configured as L1 regularization or L2 regularization.

[0053] 4) Configure the optimizer and optimization method for the logistic regression classifier model. For example, the optimizer can be lbfgs or liblinear, and the optimization method can be K-Fold cross-validation.

[0054] After building and configuring the logistic regression classifier model, the training samples in the training set are used to tune the parameters of the logistic regression classifier model. After completion, the performance of the logistic regression classifier model is tested using the test samples in the test set.

[0055] The training process of the logistic regression classifier model, in some feasible implementations, includes: 1) Configure the K value of the K-Fold cross-validation method. For example, if you configure K=5, 5-fold cross-validation will be used.

[0056] 2) Randomly shuffle the sample data in the training set.

[0057] The K-Fold cross-validation method divides the training set into K groups based on the configured K value. Each time, one of the groups is selected as the test subset, while the other groups are used as the training subsets to train and test the logistic regression classifier model. This cycle repeats K times. Therefore, randomly shuffling the sample data in the training set can improve the training performance of the logistic regression classifier model.

[0058] 3) Using the training set, the K-Fold cross-validation method was used to optimize the model parameters of the logistic regression classifier model, and the leave-one-out method was used to cross-validate the performance of the logistic regression classifier model.

[0059] The logistic regression classifier model optimizes the model parameters by learning the features of the training samples in the training set, and performs a performance test on the trained logistic regression classifier model.

[0060] The code for optimizing the model parameters of the logistic regression classifier model is as follows: cv = StratifiedKFold(n_splits=5, shuffle=True, random_state=42) grid_search = GridSearchCV( model, param_grid=param_grid, cv=cv, scoring='roc_auc', n_jobs=-1. ) The code for evaluating the performance (such as generalization ability) of the logistic regression classifier model is as follows: loo = LeaveOneOut(), y_pred_loo = [], y_true_loo = [], y_proba_loo = [], print(f"Evaluate {name} model using leave-one-out..."), for train_idx, test_idx in loo.split(X_train): X_loo_train, X_loo_test = X_train.iloc[train_idx], X_train.iloc[test_idx], y_loo_train, y_loo_test = y_train.iloc[train_idx], y_train.iloc[test_idx], if len(np.unique(y_loo_train))>1:# must have positive and negative classes X_loo_train_resampled, y_loo_train_resampled = smote.fit_resample(X_loo_train, y_loo_train), model_clone = clone(best_model), model_clone.fit(X_loo_train_resampled, y_loo_train_resampled), y_pred_loo.append(model_clone.predict(X_loo_test)[0]), y_true_loo.append(y_loo_test.iloc[0]), y_proba_loo.append(model_clone.predict_proba(X_loo_test)[0, 1]). After the above process, the ADHD auxiliary diagnosis model trained is obtained. Using the ADHD auxiliary diagnosis model, the scale evaluation data and behavior task test data of the to-be-tested person are input, and the inference result (probability of belonging to ADHD) is output by the ADHD auxiliary diagnosis model.

[0061] In addition, as an optional implementation, the ADHD auxiliary diagnosis model also outputs model explanation information, such as outputting importance ranking information of each index, to increase the explainability of the inference result.

[0062] The ADHD auxiliary diagnosis model constructed above is also verified in the embodiments of the present application: In the data set with a total sample data amount of 139, 90 cases (accounting for 64.7%) in the normal group (i.e., the diagnosis result is 0) and 49 cases (accounting for 35.3%) in the ADHD group (i.e., the diagnosis result is 1). Based on the logistic regression classifier model, modeling analysis is performed, and leave-one-out cross-validation is used to evaluate the performance of the model. The best parameter configuration of the model is: C = 0.1, class_weight (class weight) = 'balanced', penalty (regularization coefficient) = 'l2', and the optimization algorithm is lbfgs.

[0063] The evaluation result of leave-one-out cross-validation shows that the model performs well on the training set: accuracy (Accuracy): 82.88%, precision (Precision): 73.81%, recall (Recall): 79.49%, F1 value (F1-score): 76.54%, and area under ROC curve (AUC): 0.9124.

[0064] In addition, the evaluation result on the independent test set is more excellent, and the specific performance is as follows: accuracy (Accuracy): 92.86%, precision (Precision): 90.00%, recall (Recall): 90.00%, F1 value (F1-score): 90.00%, and area under ROC curve (AUC): 0.9833.

[0065] Overall, the logistic regression classifier model demonstrated high discriminative power and robustness on the current dataset, particularly with an AUC of 0.9833 on the test set, demonstrating excellent discrimination between the ADHD and normal groups. This demonstrates that the proposed solution can generate a high-performance ADHD auxiliary diagnosis model using a small sample size, eliminating the need for excessive reliance on large clinical samples and making it equally applicable to small-scale use cases (where only a small number of samples are available), thus enhancing the feasibility of clinical application.

[0066] Based on the concept of this application, an embodiment of this application also provides a control device, including a processor and a storage medium, in which computer instructions are stored. The processor runs the computer instructions in the storage medium to execute the above-mentioned ADHD auxiliary diagnostic model construction method.

[0067] In addition, an embodiment of the present application also provides an ADHD auxiliary diagnosis program product, which includes a computer program. When the computer program is executed by a processor, an ADHD auxiliary diagnosis model constructed by the above-mentioned ADHD auxiliary diagnosis model construction method is obtained.

[0068] The present invention is not limited to the aforementioned specific embodiments, but extends to any new features or any new combination disclosed in this specification, as well as any new method or process steps or any new combination disclosed.

Claims

1. A method for constructing an ADHD auxiliary diagnosis model, characterized in that: include: Preprocessing the acquired sample data; the sample data includes multi-dimensional indicators and diagnostic results for ADHD assessment; Screening out a predetermined number of indicators with the highest importance relative to the diagnosis result from the sample data to obtain a sample set, and dividing the sample set into a training set and a test set; A logistic regression classifier model is constructed, the logistic regression classifier model is trained using the training set, and the trained logistic regression classifier model is tested using the test set to obtain an ADHD auxiliary diagnosis model.

2. The method for constructing an ADHD auxiliary diagnosis model according to claim 1, wherein: The sample data includes scale assessment data and behavioral task test data; the scale test data includes at least an indicator for the assessment of at least one scale among the SNAP-IV scale, the Conners scale, and the CBCL scale; the behavioral task test data includes at least test data for at least one task among the Go / No-Go task, the N-Back task, and the Stroop task, and the test data includes at least one indicator among reaction time, accuracy, coefficient of variation of reaction time, and d-prime.

3. The method for constructing an ADHD auxiliary diagnosis model according to claim 2, wherein: Preprocess the acquired sample data, including: Deleting or correcting abnormal values ​​of each indicator in the sample data; Standardize each indicator; The indicators of the scale assessment data and the behavioral task test data are spliced ​​together to obtain a comprehensive indicator matrix.

4. The method for constructing an ADHD auxiliary diagnosis model according to claim 1, wherein: Screening out a predetermined number of indicators in the sample data that are most important relative to the diagnosis result includes: Evaluate the importance of each indicator in the sample data relative to the diagnostic result; Filter out a predetermined number of indicators that are of top importance.

5. The method for constructing an ADHD auxiliary diagnosis model according to claim 4, wherein: Evaluate the importance of each indicator in the sample data relative to the diagnostic result, including: Configure the kernel function, kernel function coefficient, regularization parameter, and sensitivity parameter of the support vector regression model; Fitting the support vector regression model using the sample data; The coef coefficient of the support vector regression model is extracted, and the importance of the corresponding indicators is sorted according to the size of the coef coefficient.

6. The method for constructing an ADHD auxiliary diagnosis model according to claim 5, wherein: The kernel function is configured as a linear kernel function, the kernel function coefficient is configured to be automatically generated, the regularization parameter is configured to be 1, and the sensitivity parameter is configured to be 0.

1.

7. The method for constructing an ADHD auxiliary diagnosis model according to claim 1, wherein: The construction of the logistic regression classifier model includes: Design the logistic regression classifier model as: ， Among them, w is the weight vector, X is the input feature vector, y is the inference result, and b is the bias term. Represents the probability that the inference result y is 1 given the input feature vector X; The objective function for designing the logistic regression classifier model is: ， Where i is the serial number of the sample data in the training set; N is the total amount of sample data in the training set; Indicates the diagnosis result of the i-th sample data, with a value of 1 indicating ADHD and a value of 0 indicating non-ADHD; ; is the regularization coefficient; Indicates the L2 norm of w; Configure regularization coefficients ; Configure the optimizer and optimization method for the logistic regression classifier model.

8. The method for constructing an ADHD auxiliary diagnosis model according to claim 7, wherein: Training the logistic regression classifier model using the training set includes: Configure the K value of the K-Fold cross-validation method; Randomly shuffling the sample data in the training set; The training set is used to optimize the model parameters of the logistic regression classifier model using the K-Fold cross-validation method, and the performance of the logistic regression classifier model is cross-checked using the leave-one-out method.

9. An ADHD auxiliary diagnosis program product, characterized in that: The method comprises a computer program, which, when executed by a processor, obtains an ADHD auxiliary diagnosis model constructed by the ADHD auxiliary diagnosis model construction method according to any one of claims 1 to 8.

10. A control device comprising a processor and a storage medium, wherein the storage medium stores computer instructions, characterized in that: The processor runs the computer instructions in the storage medium to execute the ADHD auxiliary diagnosis model construction method according to any one of claims 1-8.

Images