Method, device and system for assisting in identifying attention deficit hyperactivity disorder based on electrocardiogram signals

Abstract

The present invention discloses a method and device system for assisting in identifying attention deficit hyperactivity disorder (ADHD) based on electrocardiogram (ECG) signals. The method comprises: (1) collecting and processing ECG data of a subject whose risk of developing ADHD is to be assessed; (2) performing deep feature extraction on the processed ECG data of the subject using a one-dimensional convolutional neural network; generating a classification heat map from the feature map of the convolutional neural network using Score-CAM, and extracting time domain, frequency domain, and local statistical features from the generated classification heat map; and (3) inputting the time domain, frequency domain, and local statistical features into a machine learning classifier for classification, thereby obtaining an assessment result of the subject's risk of developing ADHD. The present invention combines deep learning with multiple machine learning methods, not only improving classification performance but also providing rich feature interpretability.

Description

Technical Field

[0001] The present invention relates to the field of medical technology, and in particular to a method, device and system for assisting in identifying attention deficit hyperactivity disorder based on electrocardiogram signals. Background Art

[0002] Attention Deficit Hyperactivity Disorder (ADHD) is a common neurodevelopmental disorder in children, characterized by inattention, impulsivity, and hyperactivity. ADHD affects approximately 5% to 7% of children and adolescents, and its etiology involves multiple factors including genetics, neurophysiology, and the environment. This disease not only significantly affects patients' daily lives, learning, and social interactions, but may also persist into adulthood, leading to reduced work efficiency and impaired social function. At the same time, ADHD is often accompanied by comorbidities such as anxiety, conduct disorder, and oppositional defiant disorder, further increasing the physical and mental burden on patients.

[0003] Currently, the diagnosis of ADHD mainly relies on clinical questionnaires (such as DSM-5 assessment criteria) and behavioral observation. However, these methods are highly subjective, and the diagnostic results are easily affected by the evaluator's experience and fluctuations in patient performance. In addition, traditional methods are time-consuming and costly, making it difficult to meet the clinical needs of large-scale screening and early intervention. Especially in areas with limited medical resources, diagnostic delays or misdiagnosis often occur. Therefore, there is an urgent need to develop an efficient and objective auxiliary diagnostic tool to improve diagnostic accuracy and shorten diagnostic time.

[0004] In recent years, electrocardiogram (ECG) signals have become a hot topic for studying the physiological characteristics of ADHD due to their close correlation with autonomic nervous system activity. Previous studies have shown that ECG signals in patients with ADHD exhibit distinct abnormalities, such as an overall decrease in heart rate variability (HRV) and irregular signal rhythms. However, most existing studies rely primarily on linear time-domain and frequency-domain analysis methods, which struggle to capture the complex nonlinear characteristics of the signal. Furthermore, these methods have limited classification accuracy and lack robustness when used with large-scale data.

[0005] Furthermore, while the application of machine learning and deep learning technologies in medical diagnosis has been steadily increasing in recent years, ADHD research remains in its early stages of exploration. For example, some studies have attempted to utilize simple statistical features combined with traditional machine learning algorithms for classification, but insufficient feature extraction has limited classification performance. For example, Chinese patent publication CN118983076A discloses a method and system for assisting the diagnosis of attention deficit hyperactivity disorder (ADHD) in children. This method includes the following steps: developing cognitive ability tests based on the cognitive characteristics and developmental status of a specific child population; collecting cognitive ability test data from subjects in the ADHD diagnosis group and a control group; preprocessing the collected cognitive ability test data; constructing a support vector machine (SVM) model, and utilizing the trained SVM model to assist in the diagnosis of ADHD in children. Furthermore, while deep learning technology can extract deep-level signal features, its "black box" nature and lack of model interpretability have limited its widespread clinical application.

[0006] Therefore, how to combine deep learning and machine learning technologies to improve the model classification performance while enhancing its clinical interpretability has become an important technical challenge in current ADHD research. Summary of the Invention

[0007] The present invention provides a device for assisting in identifying attention deficit hyperactivity disorder based on electrocardiogram signals. It combines deep learning with multiple machine learning methods, which not only improves the classification performance but also provides rich feature interpretability.

[0008] The technical solutions of the present invention are as follows:

[0009] A method for assisting in identifying attention deficit hyperactivity disorder based on electrocardiogram signals, comprising:

[0010] (1) Collecting and processing electrocardiogram data of subjects whose risk of attention deficit hyperactivity disorder needs to be assessed;

[0011] (2) A one-dimensional convolutional neural network is used to extract deep features from the processed ECG data of the subjects; a classification heat map is generated from the feature map of the convolutional neural network using Score-CAM, and time domain, frequency domain and local statistical features are extracted from the generated classification heat map;

[0012] (3) The time domain, frequency domain and local statistical features are input into the machine learning classifier for classification to obtain the assessment results of the subject's risk of developing attention deficit hyperactivity disorder.

[0013] In step (1), processing the electrocardiogram data includes: capturing a signal segment from the 3rd second to the 9th second of lead II in the electrocardiogram data as processed electrocardiogram data.

[0014] In step (2), the one-dimensional convolutional neural network includes four convolutional layers and fully connected layers connected in sequence, and each convolution layer is followed by batch normalization (BatchNorm), ReLU activation function and maximum pooling layer.

[0015] The first convolutional layer uses a convolution kernel of 3 filters with a kernel size of 27 to capture the basic local features of the ECG signal;

[0016] The second convolutional layer uses a convolution kernel of 10 filters with a kernel size of 15 to extract the time domain related features of the ECG signal;

[0017] The third convolutional layer uses a convolution kernel of 10 filters with a kernel size of 3 to further extract the mid-order features of the signal;

[0018] The fourth convolutional layer uses a convolution kernel of 10 filters with a kernel size of 3, which is responsible for capturing the high-order nonlinear characteristics of the ECG signal;

[0019] Each convolution layer is followed by batch normalization (BatchNorm) and ReLU activation function, and the feature dimension is reduced by the maximum pooling layer. The final feature is flattened by the fully connected layer to form a deep feature.

[0020] The training process of the one-dimensional convolutional neural network includes: collecting and processing the electrocardiogram data of several healthy control children (HC) and children with attention deficit hyperactivity disorder (ADHD) to construct a training data set; and using the training data set to train the one-dimensional convolutional neural network.

[0021] The construction of the training dataset includes: collecting 12-lead ECG signal data from healthy control children and children with attention deficit hyperactivity disorder aged 6 to 12 years old in a real clinical environment, excluding drug intervention and other disease interference factors; intercepting the signal fragment of a specific time period (3rd to 9th seconds) of lead II to construct the training dataset.

[0022] Furthermore, the one-dimensional convolutional neural network was trained using the Adam optimizer with a learning rate of 0.0002.

[0023] Furthermore, L2 regularization and Dropout techniques are combined in the one-dimensional convolutional neural network training process to improve the generalization performance of the model.

[0024] In step (2), Score-CAM generates a classification heat map by weighting the activation maps of each convolutional layer of the one-dimensional convolutional neural network.

[0025] The time domain features include the global activation mean, activation value standard deviation, and maximum activation value.

[0026] The global mean activation value reflects the average level of activation across the heatmap and is used to measure the overall activity of the signal. The mean activation value for the ADHD group was generally higher than that for the healthy control group, indicating greater signal fluctuation. The standard deviation of the activation value indicates the distribution of activation values ​​in the heatmap. The standard deviation for the ADHD group was significantly higher than that for the healthy control group, revealing the instability of their signal characteristics. The maximum activation value, the location of the strongest activation value in the heatmap, reflects the area of ​​greatest interest to the model and is typically concentrated near the QRS complex and T wave.

[0027] The frequency domain features include Fourier transform mean, spectrum peak, and power peak.

[0028] The Fourier transform mean was used to analyze the spectral distribution of the heat map signal. The ADHD group generally exhibited a higher spectral mean, indicating the complexity of the signal's frequency components. The spectral peak, which captures the most significant frequency component in the signal, was found to fluctuate more significantly in the ADHD group than in the HC group. The power peak, which indicates the concentration of signal energy in the frequency domain, was higher in the ADHD group than in the HC group, reflecting more dramatic energy fluctuations.

[0029] Local statistical features include sliding window mean, sliding window standard deviation, and local maximum.

[0030] The sliding window mean, calculated as the average level of local activation values ​​over a sliding window, showed greater fluctuations in the sliding mean for the ADHD group, reflecting the robust signal variability across time periods. The sliding window standard deviation, used to quantify the volatility of local activation values, was significantly higher in the ADHD group than in the HC group, indicating signal instability on small timescales. The local maximum, the maximum activation value within a sliding window, identifies key regions of the local signal, and the ADHD group typically exhibited higher local maxima.

[0031] Statistical analysis of heat map features revealed significant differences in signal characteristics between the ADHD and HC groups. Time domain analysis revealed higher overall activation levels and volatility in the ADHD group's signals, reflecting the disordered activity of their autonomic nervous system. Frequency domain analysis revealed the diversity and complexity of the spectral components in the ADHD group, indicating greater irregularity in their signals. Local statistical features further demonstrated that the dynamic changes in the ECG signals of ADHD patients were more pronounced in the temporal dimension.

[0032] In step (3), the machine learning classifier is support vector machine (SVM), random forest (RF), logistic regression (LR), K-nearest neighbor algorithm (KNN), decision tree (DT) or XGBoost.

[0033] Training a machine learning classifier involves:

[0034] Collect and process the electrocardiogram data of several healthy control children and children with attention deficit hyperactivity disorder, and divide the processed electrocardiogram data into training and test sets;

[0035] The training set is input into the trained one-dimensional convolutional neural network. Score-CAM is used to generate a classification heat map for the convolutional neural network feature map. Time domain, frequency domain and local statistical features are extracted from the generated classification heat map.

[0036] The extracted time domain, frequency domain and local statistical features are input into the machine learning classifier to train the machine learning classifier.

[0037] The present invention also provides a device for assisting in identifying attention deficit hyperactivity disorder based on electrocardiogram signals, comprising:

[0038] A data acquisition and processing module collects and processes the electrocardiogram data of subjects whose risk of developing attention deficit hyperactivity disorder needs to be assessed;

[0039] The heatmap feature extraction module uses a one-dimensional convolutional neural network to perform deep feature extraction on the processed ECG data of the subjects. Score-CAM is used to generate a classification heatmap from the convolutional neural network feature map, and time domain, frequency domain and local statistical features are extracted from the generated classification heatmap.

[0040] The classification judgment module inputs the time domain, frequency domain and local statistical features into the machine learning classifier for classification, and obtains the assessment results of the risk of the subject suffering from attention deficit hyperactivity disorder.

[0041] The present invention also provides a system for assisting in identifying attention deficit hyperactivity disorder based on electrocardiogram signals, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the method for assisting in identifying attention deficit hyperactivity disorder based on electrocardiogram signals are implemented.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] (1) High precision: Experimental results show that the method of the present invention achieves a classification accuracy of over 95% on clinical datasets.

[0044] (2) Interpretability: The heat map generated by Score-CAM technology improves the interpretability of the model, helping doctors understand the decision-making basis of the model and making the model's diagnostic results more transparent and credible.

[0045] (3) Versatility: Multiple machine learning classifiers are used to verify the extracted features, which improves the robustness and reliability of the model.

[0046] (4) Low cost and non-invasive: Compared with traditional imaging and behavioral assessment methods, ECG signal-based methods have the advantages of low cost and convenience for clinical outpatient screening. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a flow chart of the present invention for assisting in identifying attention deficit hyperactivity disorder based on electrocardiogram signals;

[0048] Figure 2 Schematic diagram of the 1D-CNN model structure;

[0049] Figure 3 The heat map feature distribution histograms of the ADHD and HC groups; (a) is the average activation value distribution histogram; (b) is the maximum activation value distribution histogram; (c) is the activation value standard deviation distribution histogram; (d) is the Fourier transform mean distribution histogram; (e) is the Fourier transform standard deviation distribution histogram; (f) is the maximum Fourier transform value distribution histogram; (g) is the power mean distribution histogram; (h) is the power maximum distribution histogram; (i) is the local mean distribution histogram; (j) is the local standard deviation distribution histogram; (k) is the local maximum distribution histogram;

[0050] Figure 4 The cross-validation performance comparison chart of 1D-CNN and classifier; (a) is the accuracy comparison line chart of ten-fold cross-validation; (b) is the precision comparison line chart of ten-fold cross-validation; (c) is the recall comparison line chart of ten-fold cross-validation; (d) is the F1 score comparison line chart of ten-fold cross-validation;

[0051] Figure 5 Examples of 1D-Score-CAM heat maps for ADHD samples (A) and HC samples (B). DETAILED DESCRIPTION

[0052] The present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be noted that the following examples are intended to facilitate understanding of the present invention and do not have any limiting effect on the present invention.

[0053] The present invention designs an automatic detection technology for attention deficit hyperactivity disorder (ADHD) based on ECG signals by deeply analyzing children's electrocardiogram (ECG) signals, combining one-dimensional convolutional neural network (1D-CNN) and multiple machine learning classifiers (the flowchart is shown in the figure). Figure 1This technology, centered around the heatmap generated by Score-CAM, constructs a feature extraction method that combines time-domain, frequency-domain, and local statistical features. This classifier accurately classifies the ECG signatures of individuals with ADHD versus those without. By combining 1D-CNN deep feature extraction with interpretability techniques, this invention significantly improves the transparency and clinical applicability of the model's classification performance, ultimately forming an efficient and interpretable ADHD detection process that enables clinicians to rapidly screen and make preliminary diagnoses without relying on behavioral assessments.

[0054] The specific technical solutions of the present invention are further illustrated by the following examples.

[0055] Devices for assisting in identifying attention deficit hyperactivity disorder based on electrocardiogram signals include:

[0056] (1) ECG data acquisition and processing module

[0057] In this example, ECG data from 2,368 children diagnosed with ADHD and 2,368 healthy controls (HC) were collected from a real-world clinical setting without medication. All participants were matched 1:1 by age, and patients with cardiovascular disease, which could affect ECG characteristics, were excluded to ensure data authenticity and representativeness.

[0058] The ECG signals collected were recorded using 12 leads. To improve signal quality and consistency, only lead II was used for research. The sampling frequency was 300 Hz, and the signal segments from 3 to 9 seconds were captured, resulting in analysis samples of 3,000 sampling points per segment. No filtering or noise reduction was performed on the signals, preserving all characteristic information from the original data.

[0059] (2) Deep learning feature extraction module

[0060] A one-dimensional convolutional neural network (1D-CNN) was constructed to gradually extract signal features through four layers of convolution and pooling (the model structure is as follows Figure 2 The specific process is as follows:

[0061] The first convolutional layer uses a 3-filter convolution kernel with a kernel size of 27 to capture the basic local features of the ECG signal. Batch normalization (BatchNorm) and the ReLU activation function enhance the model's stability and nonlinear expression capabilities.

[0062] The second convolutional layer uses a 10-filter convolution kernel with a kernel size of 15 to extract time-domain features of the ECG signal. The pooling layer uses a 2×2 maximum pooling operation to reduce feature dimensionality and retain key information.

[0063] The third convolutional layer uses a convolution kernel with 10 filters and a kernel size of 3 to further extract the mid-order pattern of the signal and reduce the dimensionality through pooling operations.

[0064] The fourth convolutional layer uses a 10-filter convolution kernel with a kernel size of 3 to capture the high-order nonlinear features of the ECG signal. It combines batch normalization with pooling to improve feature compactness and expressiveness.

[0065] Each convolution layer is followed by batch normalization (BatchNorm) and ReLU activation functions, and a max pooling layer is used to reduce feature dimensionality and mitigate overfitting risks. The final features are flattened through a fully connected layer to form deep features.

[0066] The one-dimensional convolutional neural network is trained using the Adam optimizer with a learning rate of 0.0002. During the training process, L2 regularization and Dropout techniques are combined to improve the generalization performance of the model.

[0067] (3) Heat map feature extraction and interpretability module

[0068] To enhance the clinical interpretability of the model, this paper uses Score-CAM technology to generate classification heatmaps from convolutional neural network feature maps. Score-CAM calculates the contribution of each feature region to the classification result, visually presenting the decision-making basis of the deep learning model in the form of a heatmap, helping to understand the signal regions that the model focuses on.

[0069] Extract time domain, frequency domain and local statistical features from the generated heat map (feature histogram such as Figure 3 ), as follows:

[0070] Time domain features: 1) Global activation mean: reflects the average level of the overall activation value of the heat map, which is used to measure the overall activity of the signal. The activation mean of the ADHD group is generally higher than that of the healthy control group (HC), indicating that their signal fluctuations are greater. 2) Activation value standard deviation: represents the distribution range of the heat map activation value. The standard deviation of the ADHD group is significantly higher than that of the HC group, revealing the instability of its signal characteristics. 3) Maximum activation value: The location of the strongest activation value in the heat map reflects the area of ​​greatest concern to the model, which is usually concentrated near the QRS complex wave and T wave.

[0071] Frequency Domain Features: 1) Fourier Transform Mean: This is used to analyze the spectral distribution of the signal corresponding to the heat map. The ADHD group generally exhibits a higher spectral mean, indicating the complexity of the signal's frequency components. 2) Spectral Peak: This captures the most significant frequency components in the signal. The ADHD group exhibits greater fluctuations in this feature compared to the HC group. 3) Power Peak: This indicates the concentration of signal energy in the frequency domain. The ADHD group has a higher power peak than the HC group, reflecting more dramatic energy fluctuations in their signal.

[0072] Local statistical features: 1) Sliding window mean: This is the average level of local activation values ​​calculated using a sliding window. The ADHD group had a wider range of sliding mean fluctuations, reflecting the strong changes in the signal at different time periods. 2) Sliding window standard deviation: This is used to quantify the volatility of local activation values. The ADHD group had a significantly higher standard deviation than the HC group, indicating that their signal is unstable on small time scales. 3) Local maximum: The maximum activation value within the sliding window identifies the key area of ​​the local signal. The ADHD group generally has higher local maximum values.

[0073] Statistical analysis of heat map features revealed significant differences in signal characteristics between the ADHD and HC groups. Time domain analysis revealed higher overall activation levels and volatility in the ADHD group's signals, reflecting the disordered activity of their autonomic nervous system. Frequency domain analysis revealed the diversity and complexity of the spectral components in the ADHD group, indicating greater irregularity in their signals. Local statistical features further demonstrated that the dynamic changes in the ECG signals of ADHD patients were more pronounced in the temporal dimension.

[0074] The statistical results of the heat map features significantly improved the accuracy and robustness of ADHD classification, laying a solid foundation for the clinical promotion of the model.

[0075] (4) ADHD classification and performance verification module

[0076] The extracted heatmap features were input into the following six machine learning classifiers for ADHD and HC classification: support vector machine (SVM); random forest (RF); logistic regression (LR); K-nearest neighbor (KNN); decision tree (DT); and XGBoost. Classifier performance was evaluated using a 10-fold cross-validation method, which involves splitting the dataset and repeating training and testing to comprehensively assess the stability and robustness of the model.

[0077] The results show that the combination of 1D-CNN and random forest classifier performs best (the difference in classifier performance is shown in Figure 2). Figure 4 The performance indicators are as follows: classification accuracy: 95.97%; precision: 0.96; recall: 0.96; F1 score: 0.96.

[0078] To further verify the reliability of the classification results and the interpretability of the model, a sample of one ADHD subject and one HC subject was randomly selected for heat map visualization analysis (the sample electrocardiogram heat maps of ADHD and HC are shown in Figure 2). Figure 5The results show that in the ADHD group, the heatmap activation pattern is more dispersed, with the model primarily focusing on the QRS complex and T wave regions. This indicates that ADHD patients exhibit significant abnormal fluctuations in their ECG signals in these regions. In the HC group, activation values ​​are concentrated near the R wave, and the signal distribution is relatively stable, reflecting the regularity and consistency of ECG signals in healthy individuals.

[0079] The embodiments described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for assisting in identifying attention deficit hyperactivity disorder based on electrocardiogram signals, characterized in that: include: (1) Collecting and processing electrocardiogram data of subjects whose risk of attention deficit hyperactivity disorder needs to be assessed; (2) A one-dimensional convolutional neural network is used to extract deep features from the processed ECG data of the subjects; the activation maps of each convolutional layer of the one-dimensional convolutional neural network are weighted by Score-CAM to generate a classification heat map, and the time domain, frequency domain and local statistical features are extracted from the generated classification heat map; The time domain features include global activation mean, activation value standard deviation, and maximum activation value; The frequency domain features include Fourier transform mean, spectrum peak, and power peak; Local statistical features include sliding window mean, sliding window standard deviation, and local maximum; (3) Inputting the time domain, frequency domain and local statistical features into the machine learning classifier for classification, and obtaining the assessment results of the subject's risk of developing attention deficit hyperactivity disorder; The machine learning classifier is a support vector machine, random forest, logistic regression, K-nearest neighbor algorithm, decision tree or XGBoost.

2. The method for assisting in identifying attention deficit hyperactivity disorder based on electrocardiogram signals according to claim 1, characterized in that: In step (1), processing the electrocardiogram data includes: capturing a signal segment from the 3rd second to the 9th second of lead II in the electrocardiogram data as processed electrocardiogram data.

3. The method for assisting in identifying attention deficit hyperactivity disorder based on electrocardiogram signals according to claim 1, characterized in that: In step (2), the one-dimensional convolutional neural network includes four convolutional layers and fully connected layers connected in sequence, and each convolution layer is followed by batch normalization, ReLU activation function and maximum pooling layer.

4. The method for assisting in identifying attention deficit hyperactivity disorder based on electrocardiogram signals according to claim 1, characterized in that: The training process of the one-dimensional convolutional neural network includes: collecting and processing electrocardiogram data of several healthy control children and children with attention deficit hyperactivity disorder to construct a training data set; and using the training data set to train the one-dimensional convolutional neural network.

5. The method for assisting in identifying attention deficit hyperactivity disorder based on electrocardiogram signals according to claim 1, characterized in that: Training a machine learning classifier involves: Collect and process the electrocardiogram data of several healthy control children and children with attention deficit hyperactivity disorder, and divide the processed electrocardiogram data into training and test sets; The training set is input into the trained one-dimensional convolutional neural network. Score-CAM is used to generate a classification heat map for the convolutional neural network feature map. Time domain, frequency domain and local statistical features are extracted from the generated classification heat map. The extracted time domain, frequency domain and local statistical features are input into the machine learning classifier to train the machine learning classifier.

6. A device for assisting in identifying attention deficit hyperactivity disorder based on electrocardiogram signals, comprising: A data acquisition and processing module collects and processes the electrocardiogram data of subjects whose risk of developing attention deficit hyperactivity disorder needs to be assessed; The heat map feature extraction module uses a one-dimensional convolutional neural network to perform deep feature extraction on the processed electrocardiogram data of the subject; the activation map of each convolution layer of the one-dimensional convolutional neural network is weighted by Score-CAM to generate a classification heat map, and time domain, frequency domain and local statistical features are extracted from the generated classification heat map; the time domain features include global activation mean, activation value standard deviation, and maximum activation value; the frequency domain features include Fourier transform mean, spectrum peak, and power peak; the local statistical features include sliding window mean, sliding window standard deviation, and local maximum; The classification judgment module inputs the time domain, frequency domain and local statistical features into a machine learning classifier for classification to obtain an assessment result of the risk of the subject suffering from attention deficit hyperactivity disorder; the machine learning classifier is a support vector machine, random forest, logistic regression, K nearest neighbor algorithm, decision tree or XGBoost.

7. A system for assisting in identifying attention deficit hyperactivity disorder based on electrocardiogram signals, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method for assisting in identifying attention deficit hyperactivity disorder based on electrocardiogram signals as described in any one of claims 1 to 5 are implemented.

Images