Insomnia feature extraction and identification method and system based on deep learning

Abstract

The invention relates to the technical field of insomnia recognition, in particular to an insomnia feature extraction and recognition method and system based on deep learning, and the method comprises the steps: collecting the brain function changes of a subject under different task conditions through designing a task experiment, obtaining the fMRI data of the subject in a resting state and the task experiment through functional magnetic resonance imaging, and obtaining the fMRI data of the subject; obtaining resting state fMRI data and task state fMRI data; collecting personalized data of a subject, wherein the personalized data comprises sleep behavior data, psychological assessment data and gene data; performing dynamic segmentation on the task state fMRI data and the resting state fMRI data through a dynamic function connection analysis model, and calculating dynamic function connection matrixes at different time points; constructing a multi-level insomnia feature extraction model to carry out insomnia feature extraction; and constructing a multi-task insomnia recognition model, carrying out automatic insomnia recognition and classification based on the insomnia characteristics and the dynamic function connection matrix, and predicting the insomnia severity.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of insomnia recognition, in particular to an insomnia feature extraction and recognition method and system based on deep learning. BACKGROUND

[0002] Insomnia is a common sleep disorder that seriously affects people's physical and mental health, and is closely related to depression, anxiety, Alzheimer's disease and other neuropsychiatric diseases. Timely and accurate identification of insomnia patients and analysis of features are crucial for personalized treatment and intervention. Traditional insomnia assessment relies on polysomnography (PSG) to exclude other sleep disorders that may cause insomnia symptoms, but polysomnography (PSG) detection process is complex, costly, and greatly affected by laboratory environment, making it difficult to achieve large-scale screening. Therefore, exploring more efficient and convenient insomnia recognition methods has become an important direction of current research.

[0003] In recent years, deep learning technology has made significant progress in medical signal processing, especially in sleep monitoring and sleep disorder recognition. By building a neural network-based classification model, physiological signals such as electroencephalogram (EEG), electrocardiogram (ECG), heart rate variability (HRV) can be automatically extracted and pattern recognized, realizing intelligent screening of insomnia patients. Existing research shows that deep learning models such as convolutional neural network (CNN) and long short-term memory network (LSTM) can effectively mine the spatiotemporal features of sleep-related signals, thereby improving the accuracy of insomnia detection. In addition, combining attention mechanism, adaptive feature extraction and multi-modal data fusion technologies can further improve the performance of deep learning models in insomnia recognition tasks.

[0004] However, existing deep learning-based insomnia recognition technology still faces many challenges. First, traditional structural MRI mainly reflects static anatomical features, making it difficult to capture dynamic neural functional changes related to insomnia, resulting in limited adaptability of the model to individual differences. Second, insomnia is often accompanied by comorbidities such as anxiety and depression, and existing models have difficulty effectively distinguishing specific brain function features of insomnia during feature extraction, which may mistakenly include comorbid factors into the classification criteria, thereby reducing the classification accuracy of the model. Therefore, a new method combining deep learning technology is urgently needed to more accurately capture the dynamic neural patterns of insomnia while excluding the influence of comorbid factors on insomnia classification accuracy, improving the accuracy and clinical applicability of insomnia recognition.

[0005] Therefore, an insomnia feature extraction and recognition method and system based on deep learning are proposed. SUMMARY

[0006] The application aims to provide a deep learning-based insomnia feature extraction and recognition method and system to more accurately capture the dynamic neural patterns of insomnia while excluding the influence of comorbid factors on insomnia classification accuracy, thereby improving the accuracy and clinical applicability of insomnia recognition. The method comprises: collecting the brain function changes of subjects under different task conditions through a designed task experiment, and obtaining resting-state fMRI data and task-state fMRI data by acquiring the fMRI data of subjects under resting state and the task experiment using functional magnetic resonance imaging; collecting personalized data of subjects, including sleep behavior data, psychological assessment data, and genetic data; performing dynamic segmentation on the task-state fMRI data and the resting-state fMRI data through a dynamic functional connectivity analysis model, and calculating the dynamic functional connectivity matrix at different time points; constructing a multi-level insomnia feature extraction model to extract insomnia features; constructing a multi-task insomnia recognition model to automatically recognize and classify insomnia based on the insomnia features and the dynamic functional connectivity matrix, and predicting the severity of insomnia.

[0007] To achieve the above-mentioned purpose, the application provides the following technical solutions:

[0008] A deep learning-based insomnia feature extraction and recognition method comprises:

[0009] Collecting the brain function changes of subjects under different task conditions through a designed task experiment, and obtaining resting-state fMRI data and task-state fMRI data by acquiring the fMRI data of subjects under resting state and the task experiment using functional magnetic resonance imaging;

[0010] Collecting personalized data of subjects, including sleep behavior data, psychological assessment data, and genetic data;

[0011] Performing dynamic segmentation on the task-state fMRI data and the resting-state fMRI data through a dynamic functional connectivity analysis model, and calculating the dynamic functional connectivity matrix at different time points;

[0012] Constructing a multi-level insomnia feature extraction model to extract insomnia features; the multi-level insomnia feature extraction model comprises: a first layer for extracting spatial features from the resting-state fMRI data; a second layer for extracting temporal dynamic features from the task-state fMRI data; and a third layer for fusing the personalized data, the spatial features, and the temporal dynamic features to obtain the insomnia features;

[0013] Constructing a multi-task insomnia recognition model to automatically recognize and classify insomnia based on the insomnia features and the dynamic functional connectivity matrix, and predicting the severity of insomnia.

[0014] Preferably, the task-based fMRI experiment includes: cognitive tasks, emotion regulation tasks, and sleep induction tasks.

[0015] The cognitive tasks include: memory test tasks, attention tasks, and executive function tasks.

[0016] The emotion regulation tasks include: emotional picture and / or video presentation tasks and emotion reappraisal tasks.

[0017] The sleep induction tasks include: environmental control and relaxation training.

[0018] Preferably, the sleep behavior data includes: objective sleep data, subjective sleep data, and behavioral habit data.

[0019] The psychological assessment data includes: depression and anxiety assessment, stress assessment, and emotion regulation ability assessment.

[0020] The genetic data includes: genes related to circadian rhythm, genes related to neurotransmitters, and genes related to sleep disorders.

[0021] Preferably, the dynamic functional connectivity analysis model includes: a data preprocessing unit, a dynamic window segmentation unit, a high-order mutual information calculation unit, a nonlinear Granger causality analysis unit, and a functional connectivity calculation unit.

[0022] The data preprocessing unit preprocesses the resting-state fMRI data and the task-state fMRI data, including: denoising, low-frequency filtering, head motion and physiological artifact correction, and brain region division.

[0023] The dynamic window segmentation unit uses a sliding time window method to adaptively adjust the window size based on different task conditions, and divides the preprocessed task-state fMRI data and resting-state fMRI data into multiple time windows.

[0024] The high-order mutual information calculation unit calculates the high-order mutual information between each brain region in the time window of M brain regions of interest, generating a nonlinear functional connectivity matrix.

[0025] The nonlinear Granger causality analysis unit estimates the nonlinear causal relationship between each brain region in the time window of M brain regions of interest through a deep learning model, generating a directional functional connectivity matrix.

[0026] The functional connectivity calculation unit fuses the nonlinear functional connectivity matrix and the directional functional connectivity matrix to obtain the dynamic functional connectivity matrix at different time points.

[0027] Preferably, the multi-level insomnia feature extraction model includes: the first layer, the second layer, and the third layer.

[0028] The first layer processes the resting-state fMRI data through a graph neural network to extract the spatial features; the spatial features include functional connectivity patterns of each brain region, resting-state network topological structure, and activation intensity distribution between each brain region;

[0029] The second layer processes the task-state fMRI data through an attention mechanism to analyze the time sequence variation characteristics of the brain under different task conditions, and obtains the time dynamic characteristics;

[0030] The third layer fuses the personalized data, the spatial features, and the time dynamic characteristics through a feature-level fusion method to obtain the insomnia features.

[0031] Preferably, the multi-task insomnia recognition model is constructed based on the insomnia features and the dynamic functional connectivity matrix for automatic insomnia recognition and classification, and prediction of insomnia severity; the specific process is as follows:

[0032] The multi-task insomnia recognition model is constructed by designing a multi-task learning framework, and the multi-task insomnia recognition model includes an insomnia classification unit and an insomnia severity prediction unit;

[0033] The insomnia features and the dynamic functional connectivity matrix are input into the insomnia classification unit and the insomnia severity prediction unit, respectively;

[0034] The insomnia classification unit extracts spatiotemporal patterns through deep learning, and classifies the insomnia state through a fully connected layer and a Softmax classifier;

[0035] The insomnia severity prediction unit calculates the insomnia severity score based on the attention mechanism, and compares it with the clinical standard.

[0036] Preferably, a deep learning-based insomnia feature extraction and recognition system includes:

[0037] An fMRI data acquisition module is used to collect the brain function changes of the subject under different task conditions by designing a task experiment, and to obtain the fMRI data of the subject under the resting state and the task experiment using functional magnetic resonance imaging, thereby obtaining the resting-state fMRI data and the task-state fMRI data;

[0038] A personalized data acquisition module is used to collect the personalized data of the subject, including sleep behavior data, psychological assessment data, and genetic data;

[0039] A dynamic functional connectivity analysis module is used to perform dynamic segmentation on the task-state fMRI data and the resting-state fMRI data through a dynamic functional connectivity analysis model, and to calculate the dynamic functional connectivity matrix at different time points.

[0040] an insomnia feature extraction module, configured to construct a multi-level insomnia feature extraction model to perform insomnia feature extraction; the multi-level insomnia feature extraction model comprises: a first layer configured to extract spatial features from the resting-state fMRI data; a second layer configured to extract time dynamic features from the task-state fMRI data; and a third layer configured to fuse the personalized data, the spatial features and the time dynamic features to obtain the insomnia features;

[0041] an insomnia recognition and classification module, configured to construct a multi-task insomnia recognition model to perform automatic insomnia recognition and classification based on the insomnia features and the dynamic functional connectivity matrix, and predict the severity of insomnia.

[0042] Compared with the prior art, the insomnia feature extraction recognition method based on deep learning provided by the present application has the following beneficial effects:

[0043] 1. The method collects the brain function changes of the subject under different task conditions by designing a task experiment, and obtains the brain function data in the resting state and the task state by using functional magnetic resonance imaging, so as to comprehensively evaluate the brain function activity of the individual. At the same time, combined with personalized data, including sleep behavior data, psychological evaluation data and genetic data, the sleep condition, psychological state and genetic characteristics of the individual can be comprehensively considered, the influence of comorbidity factors on the insomnia classification accuracy is excluded, and the accuracy and individualization level of subsequent insomnia recognition are improved.

[0044] 2. The present application proposes a dynamic functional connectivity analysis model to dynamically segment the task-state fMRI data and the resting-state fMRI data, calculate the dynamic functional connectivity matrix at different time points, and can capture the dynamic neural function changes of the brain function network. Compared with the traditional static functional connectivity analysis, this method can more accurately reflect the abnormal dynamic changes of the brain network of insomnia patients, and improve the understanding and recognition ability of the mechanism of insomnia.

[0045] 3. The present application adopts a multi-level insomnia feature extraction model to extract spatial features from the resting-state fMRI data, extract time dynamic features from the task-state fMRI data, and fuse them with personalized data to form high-dimensional insomnia feature expression. On this basis, a multi-task insomnia recognition model is constructed, combined with the insomnia features and the dynamic functional connectivity matrix, to perform automatic insomnia recognition and classification, and predict the severity of insomnia. This method can improve the recognition ability and prediction accuracy of insomnia, and provide more effective support for individualized intervention and clinical diagnosis. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 a flowchart of an insomnia feature extraction recognition method based on deep learning provided by an embodiment of the present application;

[0047] Figure 2A structural diagram of an insomnia feature extraction and recognition system based on deep learning is provided for an embodiment of the present application.

[0048] Figure 3 A working principle diagram of a multi-level insomnia feature extraction model is provided for an embodiment of the present application.

[0049] Figure 4 A working principle diagram of a multi-task insomnia recognition model is provided for an embodiment of the present application. DETAILED DESCRIPTION

[0050] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0051] Insomnia is a common sleep disorder that seriously affects people's physical and mental health and is closely related to depression, anxiety, Alzheimer's disease and other neuropsychiatric diseases. Timely and accurate identification of insomnia patients and analysis of features are crucial for personalized treatment and intervention. Traditional insomnia evaluation relies on polysomnography (PSG) to exclude other sleep disorders that may cause insomnia symptoms, but the polysomnography (PSG) detection process is complex, costly and greatly affected by laboratory environment, making it difficult to achieve large-scale screening. Therefore, exploring more efficient and convenient insomnia recognition methods has become an important direction of current research.

[0052] The present application proposes an insomnia feature extraction and recognition method and system based on deep learning, which realizes more accurate capture of dynamic neural patterns of insomnia, while excluding the influence of comorbid factors on insomnia classification accuracy, thereby improving the accuracy and clinical applicability of insomnia recognition. In order to illustrate that the method of the present application can improve the accuracy and clinical applicability of insomnia recognition, the effectiveness of the present application will be explained in two embodiments below.

[0053] Embodiment one

[0054] In the embodiments of the present application, the method proposed in the present application is used to more accurately capture the dynamic neural patterns of insomnia while excluding the influence of comorbid factors on insomnia classification accuracy, thereby improving the accuracy and clinical applicability of insomnia recognition. The embodiments of the present application are directed to insomnia recognition of 150 subjects of different age groups in hospital A. The insomnia diagnosis process of the subjects in hospital A will be described in detail below according to the content; among them, Figure 1 Figure 1 ​A specific flowchart of the method proposed by the present application includes: collecting the brain function changes of the subjects under different task conditions by designing task experiments, and obtaining resting state fMRI data and task state fMRI data by using functional magnetic resonance imaging to acquire the fMRI data of the subjects under resting state and the task experiments; collecting personalized data of the subjects, including: sleep behavior data, psychological assessment data and genetic data; performing dynamic segmentation on the task state fMRI data and the resting state fMRI data by using a dynamic functional connectivity analysis model, and calculating the dynamic functional connectivity matrix at different time points; constructing a multi-level insomnia feature extraction model to extract insomnia features; constructing a multi-task insomnia recognition model to automatically recognize and classify insomnia based on the insomnia features and the dynamic functional connectivity matrix, and predicting the severity of insomnia. Figure 2 The structure diagram of the system proposed by the present application is shown in Figure 1 The following description is made in combination with the contents in

[0055] A deep learning-based insomnia feature extraction and recognition method includes:

[0056] Collect the brain function changes of the subjects under different task conditions by designing task experiments, and obtain resting state fMRI data and task state fMRI data by using functional magnetic resonance imaging to acquire the fMRI data of the subjects under resting state and the task experiments;

[0057] The task experiment includes: cognitive tasks, emotion regulation tasks and sleep induction tasks;

[0058] The cognitive tasks include: memory test tasks, attention tasks and executive function tasks; the emotion regulation tasks include: emotion picture and / or video presentation tasks and emotion reappraisal tasks; the sleep induction tasks include: environmental control and relaxation training.

[0059] Specifically, the memory test task requires the subjects to remember and recall words, images and / or events to evaluate the brain activity in the memory encoding and extraction process;

[0060] The attention task studies the neural mechanisms related to attention control by asking the subjects to focus on specific stimuli and / or maintain attention under interference;

[0061] The executive function task requires the subjects to make decisions under conflicting information to evaluate the executive control function of the prefrontal cortex;

[0062] The emotion picture and / or video presentation task shows the subjects pictures and / or videos with specific emotional content to observe the brain region activity related to emotional response;

[0063] The mood re-evaluation task requires the subject to re-interpret and / or evaluate emotional stimuli to study the neural mechanism in the process of emotional regulation;

[0064] The environmental control needs to control the environment to be quiet and dim enough to relax the subject and induce the subject to enter a sleep state;

[0065] The relaxation training promotes the occurrence of sleep by guiding the subject to perform relaxation exercises such as deep breathing and / or meditation.

[0066] The embodiments of the present application design task experiments, including cognitive tasks, emotional regulation tasks and sleep induction tasks, combined with functional magnetic resonance imaging technology, to systematically collect the brain function changes of the subjects under different task conditions, so as to comprehensively reveal the brain neural mechanism of insomnia patients. The cognitive task can accurately measure the cognitive function state of the subject, especially the memory ability, attention control and executive function closely related to insomnia, so as to identify the possible impact of insomnia on high-level cognitive processes. The emotional regulation task effectively evaluates the response of the subject to emotional stimuli and the emotional regulation ability, helping to identify abnormal neural activity patterns in the emotional regulation process of insomnia individuals, which is of great value to understanding the relationship between insomnia and emotional disorders such as anxiety and depression. The sleep induction task artificially creates conditions conducive to sleep during fMRI scanning, so as to study the brain activity characteristics of insomnia patients during sleep induction, providing objective data support for exploring the physiological mechanism of insomnia. The design of task experiments not only enriches the dimension of brain function data, but also makes the analysis of the neural mechanism of insomnia more comprehensive and in-depth, and can combine the dynamic neural changes of individuals to improve the accuracy of subsequent insomnia identification and classification, providing scientific basis for clinical diagnosis and individualized intervention.

[0067] Preferably, the personalized data of the subject is collected, including sleep behavior data, psychological assessment data and genetic data;

[0068] The sleep behavior data includes objective sleep data, subjective sleep data and behavior habit data;

[0069] The psychological assessment data includes depression and anxiety assessment, stress assessment and emotional regulation ability assessment;

[0070] The genetic data includes genes related to circadian rhythm, genes related to neurotransmitters and genes related to sleep disorders.

[0071] Specifically, the objective sleep data includes total sleep time, sleep latency, deep sleep and light sleep ratio, wake-up times and rapid eye movement sleep duration;

[0072] The subjective sleep data includes Pittsburgh Sleep Quality Index, Insomnia Severity Index and sleep diary record;

[0073] The behavior habit data includes: taking sleep-aiding drugs, work-rest time, and the influence of diet on sleep;

[0074] The depression and anxiety assessment is assessed by Beck Depression Inventory, Hamilton Depression Scale, and Generalized Anxiety Scale;

[0075] The stress assessment is assessed by cortisol level and perceived stress scale;

[0076] The emotional regulation ability assessment is assessed by Cognitive Emotion Regulation Scale;

[0077] The genes related to circadian rhythm include: CLOCK gene, PER gene family, and BMAL1 gene;

[0078] The genes related to neurotransmitters include: 5-HTTLPR, COMT, and GABA receptor genes;

[0079] The genes related to sleep disorders include: HCRTR2 and DAOA;

[0080] The gene data is extracted from blood samples and / or saliva samples, and analyzed using gene sequencing and / or SNP genotyping technology.

[0081] The embodiments of the present application comprehensively collect personalized data of the subjects, including sleep behavior data, psychological assessment data, and gene data, to provide more comprehensive and accurate support for subsequent insomnia feature extraction and recognition. In terms of sleep behavior data, objective sleep parameters and subjective sleep assessment are combined to effectively quantify the sleep quality of individuals, and behavior habit data is combined to in-depth analyze external factors leading to insomnia. In terms of psychological assessment data, the depression, anxiety, and stress levels of individuals are comprehensively measured, and their emotional regulation ability is assessed, to provide objective and subjective combined psychological state analysis through various scales and biological indicators, thereby revealing the relevance between psychological factors and insomnia. In terms of gene data, genes related to circadian rhythm regulation, neurotransmitter function, and sleep disorders are closely related, to provide biological basis for exploring the genetic susceptibility of individuals. Through multi-level and multi-dimensional data collection, precise characterization of insomnia individuals is realized, to provide more scientific support for subsequent insomnia recognition, classification, and severity prediction, and also to provide important reference for personalized intervention and precision medicine, as shown in Table 1.

[0082] Table 1 shows the influence of task experiment and personalized data on insomnia recognition.

[0083] Table 1 Influence of task experiment and personalized data on insomnia recognition

[0084]

[0085] Preferably, the task-state fMRI data and the resting-state fMRI data are dynamically segmented by a dynamic functional connectivity analysis model to calculate dynamic functional connectivity matrices at different time points; refer to Figure 3 ;

[0086] The dynamic functional connectivity analysis model comprises a data preprocessing unit, a dynamic window segmentation unit, a high-order mutual information calculation unit, a nonlinear Granger causality analysis unit, and a functional connectivity calculation unit.

[0087] The data preprocessing unit preprocesses the resting-state fMRI data and the task-state fMRI data, including denoising, low-frequency filtering, head motion and physiological artifact correction, and brain region division.

[0088] The dynamic window segmentation unit uses a sliding time window method to adaptively adjust the window size based on different task conditions, and divides the preprocessed task-state fMRI data and resting-state fMRI data into multiple time windows.

[0089] The high-order mutual information calculation unit calculates the high-order mutual information between each brain region in the time window of M brain regions of interest, generating a nonlinear functional connectivity matrix.

[0090] The nonlinear Granger causality analysis unit estimates the nonlinear causal relationship between each brain region in the time window of M brain regions of interest through a deep learning model to generate a directional functional connectivity matrix.

[0091] The functional connectivity calculation unit fuses the nonlinear functional connectivity matrix and the directional functional connectivity matrix to obtain the dynamic functional connectivity matrix at different time points.

[0092] Specifically, the denoising uses independent component analysis to remove physiological noise and instrument noise.

[0093] Since low-frequency fluctuations below 0.1 Hz are considered to be related to neural activity, the low-frequency filtering uses a Butterworth filter to remove high-frequency signals.

[0094] The head motion and physiological artifact correction uses the Friston-24 parameter regression method for head motion correction, and the RETROICOR method to remove heart rate and respiratory artifacts.

[0095] The brain region division divides the brain into M regions of interest using a standard brain atlas.

[0096] Different task conditions (resting state, cognitive task, emotional task, sleep induction) affect the temporal characteristics of brain activity, so the window size needs to be adaptively adjusted according to different task conditions, including:

[0097] Resting state: the brain activity is stable, and a larger window such as 40-60s is suitable;

[0098] Cognitive task: the information processing speed is fast, and a smaller window such as 20-40s is suitable;

[0099] Emotional regulation task: the emotional response has a medium time scale, and a medium window such as 30-50s is suitable;

[0100] Sleep induction task: the brain state changes slowly, and a larger window such as 50-70s is suitable;

[0101] The adaptive adjustment formula of the window size is:

[0102] W=W0·λ T ;

[0103] Wherein, W is the window size after adaptive adjustment; W0 is the initial window size; λ T is the weight of different task conditions; in this embodiment, the resting state is 1.2, the cognitive task is 0.8, the emotional regulation task is 1.0, and the sleep induction task is 1.1;

[0104] The division formula of the time window is:

[0105] X k ={X(t)|t∈[kS,kS+W]},k=1,2,...,K;

[0106] Wherein, X k is the data of the kth time window; t is time; S is the step; K is the total number of windows;

[0107] The calculation formula of the high-order mutual information is:

[0108]

[0109] Wherein, is the high-order mutual information; and are the resting state and task state fMRI data of brain region i and brain region j in the time window k; P(x i ) and P(x j ) are the marginal probability distributions of brain region i and brain region j;

[0110] The nonlinear functional connection matrix is represented as:

[0111]

[0112] where H k (i,j) is the nonlinear functional connectivity matrix;

[0113] The formula of the nonlinear Granger causality analysis is:

[0114]

[0115] where f() is the nonlinear function fitting the dynamic linearity of the brain region j; g() is the nonlinear function modeling the influence of the brain region i on the brain region j; is the resting-state and task-state fMRI data sequence of the brain region j in the past p time steps; is the resting-state and task-state fMRI data sequence of the brain region i in the past p time steps; and ε is the error term;

[0116] The directional functional connectivity matrix is represented as:

[0117]

[0118] where G k (i,j) is the directional functional connectivity matrix; and NGC() is the nonlinear Granger causality analysis function;

[0119] The dynamic functional connectivity matrix is represented as:

[0120] F k = α·H k (i,j) + (1-α)·G k (i,j);

[0121] where F k is the dynamic functional connectivity matrix; and α is a weight parameter optimized by cross-validation;

[0122] The method of the embodiment of the application realizes fine processing and analysis of resting state and task state fMRI data by constructing a dynamic functional connectivity analysis model, thereby improving the accuracy and reliability of insomnia feature extraction. First, physiological noise and instrument noise are removed by a data preprocessing unit, thereby enhancing the signal-to-noise ratio of the fMRI data. At the same time, head motion correction is performed using the Friston-24 parameter regression method, and the RETROICOR method is used to remove heart rate and respiration artifacts, further reducing the interference of motion artifacts and physiological noise, and ensuring the stability and reliability of the data. In the dynamic window segmentation process, the sliding time window method is used to adaptively adjust the window size under the task condition, so that it can match the dynamic change characteristics of the brain under different task states. This optimization strategy of window size can effectively balance the time resolution and functional connectivity stability, and improve the accuracy of dynamic feature extraction. In terms of functional connectivity calculation, first, the non-linear correlation between brain regions is calculated by high-order mutual information to construct a non-linear functional connectivity matrix, which can overcome the limitations of traditional linear analysis methods and capture more complex neural signal correlations. Further, a deep learning model is used for non-linear Granger causality analysis to estimate the directional information flow between brain regions and generate a directional functional connectivity matrix, thereby revealing the causal interaction relationship between different brain regions. Finally, the non-linear functional connectivity matrix and the directional functional connectivity matrix are fused by a functional connectivity calculation unit to obtain the dynamic functional connectivity matrix at different time points, realizing the dynamic tracking of the brain functional network under different states. This method not only can capture key brain network changes related to insomnia, but also can distinguish different insomnia subtypes, improve the accuracy of insomnia recognition and classification, and provide more accurate neuroimaging basis for individualized intervention strategies. Referring to Table 2, Table 2 gives the comparison of insomnia recognition accuracy of dynamic functional connectivity analysis and traditional static functional connectivity analysis.

[0123] Table 2 Comparison of insomnia recognition accuracy of dynamic functional connectivity analysis and traditional static functional connectivity analysis

[0124] Number Static functional connectivity matrix Dynamic functional connectivity matrix Insomnia identification accuracy improvement 1 84.0％ 91.0％ +7％ 2 85.5％ 92.3％ +6.8％ 3 83.0％ 90.2％ +7.1％ 4 82.5％ 91.3％ +7.4％ 5 84.2％ 90.8％ +6.6％

[0125] Preferably, a multi-level insomnia feature extraction model is constructed for insomnia feature extraction; the multi-level insomnia feature extraction model comprises: a first layer for extracting spatial features from the resting state fMRI data; a second layer for extracting time dynamic features from the task state fMRI data; and a third layer for fusing the individualized data, the spatial features and the time dynamic features to obtain the insomnia features; see Figure 4 ;

[0126] The multi-level insomnia feature extraction model comprises the first layer, the second layer and the third layer;

[0127] The first layer processes the resting-state fMRI data through a graph neural network to extract the spatial features; the spatial features include: functional connectivity patterns of each brain region, resting-state network topology, and activation intensity distribution between each brain region;

[0128] The second layer processes the task-state fMRI data through an attention mechanism to analyze the temporal variation characteristics of the brain under different task conditions, and obtains the time dynamic characteristics;

[0129] The third layer fuses the personalized data, the spatial features and the time dynamic characteristics through a feature-level fusion method to obtain the insomnia characteristics.

[0130] Specifically, the process of processing the resting-state fMRI data through a graph neural network is: calculating the Pearson correlation coefficient between the time series of each brain region according to the M brain regions of interest, and constructing a resting-state functional connectivity matrix; converting the resting-state functional connectivity matrix into a graph structure, and calculating the spatial features using a graph convolution network; wherein the functional connectivity pattern of each brain region extracts the functional connectivity features between different brain regions through graph embedding learning; the resting-state network topology reflects the network organization mode between brain regions by calculating network centrality, modularity and other topological indicators; the activation intensity distribution between each brain region extracts the activity intensity features of different brain regions by analyzing the power spectrum of the time series of each brain region.

[0131] The process of processing the task-state fMRI data through an attention mechanism is: inputting the task-state fMRI data represented as a time series into a Transformer or LSTM network, identifying the neural activity changes at key time points through a self-attention mechanism, and obtaining the time dynamic characteristics.

[0132] Table 3 shows the comprehensive performance of the multi-level insomnia feature extraction model in classification and severity prediction.

[0133] Table 3 Comprehensive performance of multi-level insomnia feature extraction model in classification and severity prediction

[0134] Performance index Traditional feature extraction Multi-level insomnia feature extraction model Insomnia identification accuracy (%) 84.1 93.5 F1-score 0.81 0.90 Severity prediction RMSE 3.5 1.1

[0135] The embodiments of the present application construct a multi-level insomnia feature extraction model, comprehensively utilize resting state and task state fMRI data and personalized data, and realize fine description of insomnia-related neural mechanisms from multiple dimensions. The functional connection mode, network topology structure and activation intensity distribution of each brain region are extracted from the resting state fMRI data by using the graph neural network, which can accurately capture the basic functional organization characteristics of the brain in the resting state. By converting the resting state functional connection matrix into a graph structure and using the graph convolution network for graph embedding learning and topology index calculation, the deep representation of the resting state network characteristics is realized. At the same time, the task state fMRI data is represented as a time sequence, and the self-attention mechanism in the Transformer or LSTM network is used to extract the neural activity changes at the key time points, and the features reflecting the temporal dynamic changes of the brain under different task conditions are obtained. The spatial features and temporal dynamic features and the personalized data are fused at the feature level to form a high-dimensional and comprehensive insomnia feature expression, which not only significantly improves the accuracy and robustness of insomnia recognition and classification, but also provides a strong scientific basis for in-depth understanding of the neural physiological mechanism of insomnia, early diagnosis and development of personalized intervention strategies.

[0136] Preferably, a multi-task insomnia recognition model is constructed based on the insomnia features and the dynamic functional connection matrix for automatic insomnia recognition and classification, and prediction of insomnia severity; the specific process is as follows:

[0137] The multi-task insomnia recognition model is constructed by designing a multi-task learning framework, and the multi-task insomnia recognition model includes an insomnia classification unit and an insomnia severity prediction unit;

[0138] The insomnia features and the dynamic functional connection matrix are input into the insomnia classification unit and the insomnia severity prediction unit, respectively;

[0139] The insomnia classification unit extracts spatio-temporal patterns by deep learning, and classifies insomnia states by a fully connected layer and a Softmax classifier;

[0140] The insomnia severity prediction unit calculates the insomnia severity score based on the attention mechanism, and compares it with the clinical standard.

[0141] Specifically, the insomnia classification unit uses a deep learning model, such as a convolutional neural network (CNN) or a recurrent neural network (RNN), to extract spatio-temporal features from the input insomnia features and dynamic functional connection matrix; the extracted spatio-temporal features are further processed by a fully connected layer; finally, a Softmax classifier is used for insomnia state classification, outputting the probability distribution of insomnia and non-insomnia classes, and selecting the class with a probability higher than a preset threshold for final classification prediction.

[0142] The insomnia severity prediction unit adopts a deep learning model based on an attention mechanism, automatically assigns different weights to each input insomnia feature and dynamic functional connectivity matrix, and performs weighted summation on the input features according to the weights to obtain weighted features; the weighted features are input into a fully connected layer to obtain the insomnia severity score; the severity score is expressed as a score interval of 0 to 10, and by comparing with the clinical standard, the grades of mild, moderate and severe insomnia can be divided.

[0143] The embodiments of the present application can simultaneously realize automatic identification and classification of insomnia and prediction of the severity of insomnia by constructing a multi-task insomnia recognition model, which provides a more accurate and efficient way for clinical diagnosis of insomnia. The model can extract spatiotemporal patterns from insomnia features and dynamic functional connectivity matrices through deep learning technology, providing strong support for the classification of insomnia state. At the same time, combined with the attention mechanism, the model can automatically assign weights to different input features, thereby improving the prediction ability of the severity of insomnia, making the prediction results more in line with individual differences. After processing by the fully connected layer, the model can score the severity of insomnia according to the clinical standard, helping clinicians to accurately assess the severity of insomnia. In addition, the loss function of the joint optimization of insomnia classification and severity prediction tasks can ensure that the performance of both tasks is optimized, enhancing the robustness and accuracy of the model. The multi-task insomnia recognition model improves the efficiency and accuracy of insomnia recognition, providing strong support for early intervention and personalized treatment.

[0144] The embodiment of the application realizes comprehensive evaluation of brain function activities by designing task experiments, collecting brain function data of subjects under resting state and task state conditions, and combining sleep behavior, psychological evaluation and genetic personalized data, can comprehensively consider sleep condition, psychological state and genetic characteristics of individuals, exclude the influence of comorbidity factors on insomnia classification accuracy, and thus improve the accuracy and individualization level of subsequent insomnia recognition. With the support of the dynamic functional connectivity analysis model, the task state and resting state fMRI data are dynamically segmented and analyzed, so as to calculate the dynamic functional connectivity matrix reflecting the neural network state at different time points, which not only captures the dynamic changes of the brain network under different task conditions, but also accurately reflects the abnormal patterns of the brain network of insomnia patients, accurately reflects the abnormal dynamic changes of the brain network of insomnia patients, and improves the understanding and recognition ability of insomnia mechanism. Meanwhile, the multi-level insomnia feature extraction model extracts spatial features under the resting state and time dynamic features under the task state respectively, and fuses these features with personalized data to form high-dimensional insomnia feature expression, which provides more rich and accurate neuroimaging and behavioral biomarkers for subsequent insomnia recognition. The multi-task insomnia recognition model constructed based on these insomnia features and dynamic functional connectivity matrix realizes automatic recognition and classification of insomnia state by using a deep learning method, and can accurately predict the severity of insomnia, thereby improving the accuracy and individualization level of recognition, and providing a strong scientific basis for clinical diagnosis and precise intervention.

[0145] Embodiment two

[0146] In embodiment one, the method proposed by the application successfully realizes more accurate capture of the dynamic neural patterns of insomnia, eliminates the influence of comorbidity factors on insomnia classification accuracy, and improves the accuracy and clinical applicability of insomnia recognition. To further verify the effectiveness of the application, another 150 subjects in hospital B were used to identify insomnia in the embodiment of the application.

[0147] A deep learning-based insomnia feature extraction and recognition system, comprising:

[0148] An fMRI data acquisition module is configured to collect brain function changes of subjects under different task conditions by designing task experiments, and acquire fMRI data of subjects under resting state and the task experiments by using functional magnetic resonance imaging, to obtain resting state fMRI data and task state fMRI data.

[0149] The task experiment includes a cognitive task, an emotion regulation task and a sleep induction task.

[0150] The cognitive task includes memory test task, attention task and executive function task; the emotion regulation task includes emotional picture and / or video presentation task and emotion reappraisal task; and the sleep induction task includes environment control and relaxation training.

[0151] Preferably, the personalized data acquisition module is used to acquire personalized data of the subject, including sleep behavior data, psychological assessment data and genetic data.

[0152] The sleep behavior data includes objective sleep data, subjective sleep data and behavior habit data.

[0153] The psychological assessment data includes depression and anxiety assessment, stress assessment and emotion regulation ability assessment.

[0154] The genetic data includes genes related to circadian rhythm, genes related to neurotransmitters and genes related to sleep disorders.

[0155] Preferably, the dynamic functional connectivity analysis module is used to perform dynamic segmentation on the task-state fMRI data and the resting-state fMRI data through a dynamic functional connectivity analysis model, and calculate dynamic functional connectivity matrices at different time points.

[0156] The dynamic functional connectivity analysis model includes a data preprocessing unit, a dynamic window segmentation unit, a high-order mutual information calculation unit, a nonlinear Granger causality analysis unit and a functional connectivity calculation unit.

[0157] The data preprocessing unit pre-processes the resting-state fMRI data and the task-state fMRI data, including denoising, low-frequency filtering, head motion and physiological artifact correction, and brain region division.

[0158] The dynamic window segmentation unit uses a sliding time window method to adaptively adjust the window size based on different task conditions, and divides the pre-processed task-state fMRI data and resting-state fMRI data into multiple time windows.

[0159] The high-order mutual information calculation unit calculates the high-order mutual information between each brain region in the time window of M brain regions of interest, and generates a nonlinear functional connectivity matrix.

[0160] The nonlinear Granger causality analysis unit estimates the nonlinear causal relationship between each brain region in the time window of M brain regions of interest through a deep learning model, and generates a directional functional connectivity matrix.

[0161] The functional connectivity calculation unit fuses the nonlinear functional connectivity matrix and the directional functional connectivity matrix to obtain the dynamic functional connectivity matrix at different time points.

[0162] Preferably, the insomnia feature extraction module is configured to construct a multi-level insomnia feature extraction model to extract insomnia features; the multi-level insomnia feature extraction model comprises: a first layer configured to extract spatial features from the resting-state fMRI data; a second layer configured to extract time dynamic features from the task-state fMRI data; and a third layer configured to fuse the personalized data, the spatial features and the time dynamic features to obtain the insomnia features.

[0163] The multi-level insomnia feature extraction model comprises the first layer, the second layer and the third layer.

[0164] The first layer is configured to process the resting-state fMRI data by a graph neural network to extract the spatial features; the spatial features comprise functional connection patterns of brain regions, resting-state network topological structures and activation intensity distribution between brain regions.

[0165] The second layer is configured to process the task-state fMRI data by an attention mechanism to analyze time sequence variation features of the brain under different task conditions to obtain the time dynamic features.

[0166] The third layer is configured to fuse the personalized data, the spatial features and the time dynamic features by a feature-level fusion method to obtain the insomnia features.

[0167] Preferably, the insomnia recognition and classification module is configured to construct a multi-task insomnia recognition model to automatically recognize and classify insomnia based on the insomnia features and the dynamic functional connection matrix, and to predict the severity of insomnia; the specific process is as follows:

[0168] The multi-task insomnia recognition model is constructed by designing a multi-task learning framework; the multi-task insomnia recognition model comprises an insomnia classification unit and an insomnia severity prediction unit.

[0169] The insomnia features and the dynamic functional connection matrix are input into the insomnia classification unit and the insomnia severity prediction unit, respectively.

[0170] The insomnia classification unit extracts spatio-temporal patterns by deep learning, and classifies insomnia states by a fully connected layer and a Softmax classifier.

[0171] The insomnia severity prediction unit calculates an insomnia severity score based on an attention mechanism, and compares the insomnia severity score with a clinical standard.

[0172] Table 4 shows insomnia recognition, classification and severity prediction results of the multi-task insomnia recognition model.

[0173] Table 4 Insomnia recognition, classification, and severity prediction results of the multitask insomnia recognition model

[0174] Subject number Insomnia identification result Classification result Severity score Severity level 1 Insomnia Mild insomnia 4.2 Mild 2 Insomnia Severe insomnia 8.1 Severe 3 Non-insomnia No 0.3 No 4 Insomnia Moderate insomnia 6.5 Moderate 5 Non-insomnia No 1.4 No

[0175] While embodiments of the present application have been shown and described with reference to particular embodiments thereof, it will be understood by those of ordinary skill in the art that various changes in form and details can be made therein without departing from the spirit and scope of the application as defined by the following claims and their equivalents.

Claims

1. A deep learning-based insomnia feature extraction and recognition method, characterized in that, The application relates to a method for automatic insomnia recognition and classification. The method comprises the following steps: Collecting brain function changes of a subject under different task conditions by designing a task experiment, and obtaining resting state fMRI data and task state fMRI data by acquiring fMRI data of the subject under a resting state and the task experiment through functional magnetic resonance imaging; Collecting personalized data of the subject, including sleep behavior data, psychological assessment data and genetic data; Performing dynamic segmentation on the task state fMRI data and the resting state fMRI data through a dynamic functional connection analysis model, and calculating a dynamic functional connection matrix at different time points; Constructing a multi-level insomnia feature extraction model to extract insomnia features; the multi-level insomnia feature extraction model comprises: a first layer for extracting spatial features from the resting state fMRI data; a second layer for extracting time dynamic features from the task state fMRI data; and a third layer for fusing the personalized data, the spatial features and the time dynamic features to obtain the insomnia features; 2. The insomnia feature extraction and identification method based on deep learning according to claim 1, characterized in that, Constructing a multi-task insomnia recognition model to automatically recognize and classify insomnia and predict the severity of insomnia based on the insomnia features and the dynamic functional connection matrix. The task experiment comprises cognitive tasks, emotion regulation tasks and sleep induction tasks; 3.The insomnia feature extraction and recognition method based on deep learning according to claim 1, characterized in that, The cognitive tasks comprise memory test tasks, attention tasks and executive function tasks; the emotion regulation tasks comprise emotion picture and / or video presentation tasks and emotion reappraisal tasks; and the sleep induction tasks comprise environment control and relaxation training. 4.The insomnia feature extraction and recognition method based on deep learning according to claim 1, characterized in that, The sleep behavior data comprises objective sleep data, subjective sleep data and behavior habit data; the psychological assessment data comprises depression and anxiety assessment, stress assessment and emotion regulation ability assessment; and the genetic data comprises genes related to circadian rhythm, genes related to neurotransmitters and genes related to sleep disorders. The dynamic functional connection analysis model comprises a data preprocessing unit, a dynamic window segmentation unit, a high-order mutual information calculation unit, a nonlinear Granger causality analysis unit and a functional connection calculation unit; The data preprocessing unit pre-processes the resting state fMRI data and the task state fMRI data, including denoising, low-frequency filtering, head motion and physiological artifact correction and brain region division; the dynamic window segmentation unit adaptively adjusts the window size based on different task conditions by using a sliding time window method, divides the pre-processed task state fMRI data and the resting state fMRI data into multiple time windows; the high-order mutual information calculation unit calculates the high-order mutual information between brain regions in the time window of M brain regions of interest, and generates a nonlinear functional connection matrix; the nonlinear Granger causality analysis unit estimates the nonlinear causal relationship between brain regions in the time window of M brain regions of interest through a deep learning model, and generates a directional functional connection matrix; and the functional connection calculation unit fuses the nonlinear functional connection matrix and the directional functional connection matrix to obtain the dynamic functional connection matrix at different time points.

5. The insomnia feature extraction and recognition method based on deep learning according to claim 1, characterized in that, The multi-level insomnia feature extraction model comprises the first layer, the second layer and the third layer. The first layer processes the resting-state fMRI data through a graph neural network to extract the spatial features; the spatial features comprise functional connection modes of each brain region, resting-state network topological structures and activation intensity distribution between each brain region; the second layer processes the task-state fMRI data through an attention mechanism to analyze time sequence variation characteristics of the brain under different task conditions to obtain the time dynamic features; and the third layer fuses the personalized data, the spatial features and the time dynamic features through a feature-level fusion method to obtain the insomnia features.

6. The insomnia feature extraction and recognition method based on deep learning according to claim 1, characterized in that, The multi-task insomnia recognition model is constructed based on the insomnia features and the dynamic functional connection matrix to automatically recognize and classify insomnia and predict insomnia severity; the specific process is as follows: The multi-task insomnia recognition model is constructed through a multi-task learning framework, and comprises an insomnia classification unit and an insomnia severity prediction unit; the insomnia features and the dynamic functional connection matrix are input into the insomnia classification unit and the insomnia severity prediction unit respectively; the insomnia classification unit extracts spatiotemporal patterns through deep learning and classifies insomnia states through a fully connected layer and a Softmax classifier; and the insomnia severity prediction unit calculates insomnia severity scores based on an attention mechanism and compares the scores with clinical standards.

7. A deep learning-based insomnia feature extraction and recognition system, characterized in that, The method comprises the following steps: An fMRI data acquisition module is configured to collect brain function changes of a subject under different task conditions through a designed task experiment, acquire fMRI data of the subject in a resting state and under the task experiment by using functional magnetic resonance imaging, and obtain resting-state fMRI data and task-state fMRI data; A personalized data acquisition module is configured to acquire personalized data of the subject, including sleep behavior data, psychological assessment data and genetic data; A dynamic functional connection analysis module is configured to perform dynamic segmentation on the task-state fMRI data and the resting-state fMRI data through a dynamic functional connection analysis model, and calculate dynamic functional connection matrices at different time points; An insomnia feature extraction module is configured to construct a multi-level insomnia feature extraction model to extract insomnia features; the multi-level insomnia feature extraction model comprises a first layer for extracting spatial features from the resting-state fMRI data, a second layer for extracting time dynamic features from the task-state fMRI data, and a third layer for fusing the personalized data, the spatial features and the time dynamic features to obtain the insomnia features; An insomnia recognition and classification module is configured to construct a multi-task insomnia recognition model based on the insomnia features and the dynamic functional connection matrix to automatically recognize and classify insomnia and predict insomnia severity.

Images