Depression assessment system based on resting heart and brain coupling analysis and implementation method

Abstract

The invention discloses a depression assessment system based on resting heart and brain coupling analysis and an implementation method. The system performs depression scale score prediction through a depression symptom assessment model; the implementation method comprises the steps that synchronous multi-lead electroencephalogram signals and electrocardiosignals with the same time duration are collected for a plurality of users; the collected electroencephalogram signals and electrocardiosignals are preprocessed; feature sequence extraction is carried out on the preprocessed electroencephalogram signals and electrocardiosignals, and traditional heart rate variability features are extracted; calculating causal nonlinear coupling strength between each frequency band feature sequence and the RR interval feature sequence of the electroencephalogram to obtain a heart and brain coupling feature set; based on the heart rate variability feature set and the heart and brain coupling feature set, establishing a depressive symptom evaluation model; according to the depression feature set obtained through calculation, the depression symptom evaluation model outputs the quantitative score of the depression degree. The method can avoid the limitation that a traditional scale depends on subjective evaluation, and can be used for early screening of depression.

Description

Technical Field

[0001] The present invention relates to the technical field of biomedical signal processing, and particularly relates to a depression assessment system and implementation method based on resting cardio-cerebral coupling analysis. Background Art

[0002] Depression is a common mental disorder, and its physiological mechanism involves abnormal functions of the central nervous system and the autonomic nervous system. In recent years, cardio-cerebral coupling has become an important direction for studying the neurophysiological characteristics of depression, aiming to reveal the interaction between brain activity and cardiac autonomic regulation. However, the cardio-cerebral interaction relationship is complex, showing non-linear and non-stationary characteristics, and traditional linear analysis methods are difficult to accurately describe its dynamic coupling pattern.

[0003] Existing studies have shown that the abnormal autonomic nerve function in depression patients is manifested as a decrease in heart rate variability, especially the reduction of high-frequency components, which reflects the impairment of vagus nerve regulation function. Heart rate variability can be used as an index of central-peripheral integration. In addition, cardio-cerebral coupling studies have shown that the cardio-cerebral interaction pattern in depression patients has changed, such as abnormal coupling between specific frequency bands of electroencephalogram signals and heart rate dynamics. However, most of the existing depression assessment methods rely on psychological questionnaires and subjective scales, lacking objective, physiological index-driven quantitative assessment means, and it is difficult to comprehensively reflect the complex physiological manifestations of depression by using heart rate variability or electroencephalogram characteristics alone. Summary of the Invention

[0004] Object of the Invention: The object of the present invention is to provide a depression assessment system and implementation method based on resting cardio-cerebral coupling analysis. By quantifying cardio-cerebral coupling characteristics and combining heart rate variability analysis, a depression degree prediction model is established, making the assessment of depression more objective and repeatable, and improving the ability to identify physiological abnormalities related to depression.

[0005] Technical Solution: A depression assessment system based on resting cardio-cerebral coupling analysis includes:

[0006] A data acquisition module, which uses a multi-channel acquisition device to synchronously acquire the original electrocardiogram signal and the original electroencephalogram signal of a user in a resting state for a certain period of time, and collects the clinical assessment scale of depression of the user;

[0007] A cardio-cerebral signal preprocessing module, which performs signal processing according to the acquired original electroencephalogram signal and original electrocardiogram signal;

[0008] A feature sequence extraction module, which calculates the characteristic sequences of each frequency band of the electroencephalogram according to the preprocessed electroencephalogram signal; calculates the RR interval characteristic sequence and the heart rate variability feature set according to the preprocessed electrocardiogram signal;

[0009] A coupling estimation module that calculates cardio-cerebral coupling features based on the extracted characteristic sequences of each frequency band of electroencephalogram (EEG) and the characteristic sequence of RR interval; the cardio-cerebral coupling features include two interaction directions, namely heart-to-brain and brain-to-heart, multi-channel of the whole brain and multiple EEG frequency bands.

[0010] A depression assessment model establishment module that constructs a depression symptom assessment model based on the cardio-cerebral coupling features, heart rate variability features, and data from the clinical depression assessment scale to quantitatively evaluate the depression level of an individual.

[0011] An output result module for outputting the depression level score of the user.

[0012] A method for realizing depression assessment based on resting cardio-cerebral coupling analysis. Based on the above-mentioned depression assessment system, by quantifying the cardio-cerebral coupling features and combining heart rate variability analysis, a depression symptom assessment model is established, and finally a quantitative score of the depression level is output. The steps are as follows:

[0013] S1, Collect synchronous multi-lead EEG signals and electrocardiogram (ECG) signals of the same time length from multiple users.

[0014] S2, Preprocess the collected EEG signals and ECG signals.

[0015] S3, Extract characteristic sequences from the preprocessed EEG signals and ECG signals, and extract traditional heart rate variability features.

[0016] S4, Calculate the causal nonlinear coupling strength between the characteristic sequences of each frequency band of EEG and the characteristic sequence of RR interval to obtain the nonlinear coupling features from heart to brain and the nonlinear coupling features from brain to heart. The two nonlinear coupling features together constitute the cardio-cerebral coupling feature set.

[0017] S5, Establish a depression symptom assessment model based on the heart rate variability feature set and the cardio-cerebral coupling feature set.

[0018] S6, According to the calculated depression feature set, output the quantitative score of the depression level by the depression symptom assessment model.

[0019] Furthermore, the collected EEG signals and ECG signals are preprocessed by the heart and brain signal preprocessing module. The detailed implementation steps are as follows:

[0020] S21, Apply a band-pass filter with a frequency range of 0.5 - 45 Hz and a 50 Hz notch filter to the collected ECG signals and EEG signals simultaneously for filtering and denoising, and resample them to 250 Hz.

[0021] S22, Apply the Pan-Tompking algorithm to detect the R peaks in the ECG signals and perform visual inspection and correction, and synchronously mark the EEG signals according to the R peaks.

[0022] S23. Manually remove bad segments and interpolate bad leads through visual inspection; reject contaminated heart and brain data segments due to head movement and drift, with each deleted data segment starting and ending with an R peak.

[0023] S24. Perform independent component analysis on the EEG signals and use single-lead ECG data as additional input; visually observe each component and manually reject components related to eye movement and cardiac artifacts.

[0024] S25. Apply common average referencing to the EEG signals.

[0025] Furthermore, the feature sequence extraction module intercepts data of the same length from the preprocessed ECG and EEG signals starting from the same R peak.

[0026] The implementation steps for extracting the feature sequence from the preprocessed EEG signals are as follows:

[0027] S311. Decompose the preprocessed EEG signals using multivariate empirical mode decomposition to extract a set of IMFs; for each user, all EEG channels obtain the same number of IMFs.

[0028] S312. Use a matching algorithm to select IMF components in the frequency band of interest among different individuals.

[0029] S313. Apply the Hilbert transform to the IMF components in the frequency band of interest to obtain the envelope curve of the IMF components, and downsample the envelope curve to obtain the EEG feature sequences X of each frequency band representing central nervous activity.

[0030] The implementation steps for extracting the RR interval feature sequence from the preprocessed ECG signals are as follows:

[0031] S321. Based on the marked R peaks in the preprocessed ECG signals, calculate all RR interval times of the ECG signals to construct the original RR interval sequence.

[0032] S322. Use cubic spline interpolation to uniformly interpolate the original RR interval sequence at a sampling rate of 250 Hz, and downsample the uniformly sampled RR interval sequence by 15 times to obtain the RR interval feature sequence Y representing autonomic nerve activity.

[0033] The implementation steps for extracting traditional heart rate variability features from the original and interpolated RR interval feature sequences are as follows:

[0034] S331. Calculate the heart rate HR, standard deviation of RR intervals SDNN, root mean square of successive differences RMSSD, and ratio pNN50 of the percentage of the number of changes in consecutive normal sinus intervals exceeding 50 milliseconds in the time domain of the original RR interval sequence.

[0035] S332. On the frequency domain of the RR interval feature sequence after interpolation, calculate the low-frequency component LF with the RR interval sequence frequency ranging from 0.04 to 0.15 Hz, the high-frequency component HF with the frequency ranging from 0.15 to 0.4 Hz, and the ratio LF / HF.

[0036] S333. Construct a traditional heart rate variability feature set including HR, SDNN, RMSSD, pNN50, LF, HF, and LF / HF.

[0037] Furthermore, through central frequency calculation and Hungarian algorithm matching, extract the IMFs in the frequency band of interest, including the following steps:

[0038] S3121. Set the target frequency band range of the frequency band of interest.

[0039] S3122. Each IMF component is a time series signal with a total number of points N. For the value of any point n, denoted as IMF(n), apply the Hilbert transform to IMF(n) to obtain H(IMF(n)), and construct the analytic signal z(n) with the following formula:

[0040]

[0041] z(n) = IMF(n) + jH(IMF(n))

[0042] Calculate the instantaneous phase φ(n) of the IMF component and derive the instantaneous frequency f(n) with the following formula:

[0043]

[0044] Perform time averaging on the instantaneous frequency f(n) to obtain the central frequency F of this IMF component with the following formula:

[0045]

[0046] where n ∈ [0, N], representing the data point index;

[0047] S3123. Calculate the central frequencies of all IMFs of each user respectively, screen out the users whose central frequencies of IMFs continuously fall within the entire target frequency band range, and use the IMF components of this user that fall within the target frequency band range as the reference set for matching the IMFs of interest.

[0048] S3124. For the IMF set of each user, calculate the cost matrix between the IMF set and the reference set.

[0049] S3125. Use the Hungarian algorithm to solve the optimal matching and find the minimum cost allocation scheme between the IMFs and the target frequency band to minimize the total allocation error.

[0050] After matching by the Hungarian algorithm, the IMF components of each user correspond one-to-one with the target frequency bands, and the specific IMF components of interest are obtained.

[0051] Further, the coupling estimation module uses convergent cross-mapping to calculate the causal non-linear coupling strength between the characteristic sequences of each EEG frequency band and the characteristic sequence of the RR interval. The specific steps are as follows:

[0052] S41, for the characteristic sequence X of each EEG frequency band and the characteristic sequence Y of the RR interval, perform phase space reconstruction respectively; for the value of the t-th data point on the characteristic sequence X of each EEG frequency band, denoted as X(t), construct a D-dimensional vector x(t), and the formula is as follows:

[0053] x(t) = <X(t), X(t - τ), X(t - 2τ), …, X(t - (D - 1)τ), t = 1 + (D - 1)τ, …, L

[0054] where τ is the constant time delay, D is the embedding dimension; L is the total number of points in the sequence length;

[0055] The set of these D-dimensional vectors x(t) constitutes the shadow manifold M of the phase space X ;

[0056] The same method is applied to the characteristic sequence Y of the RR interval, and a D-dimensional vector y(t) is constructed for the t-th data point. Finally, the set of all D-dimensional vectors obtains the shadow manifold M Y ;

[0057] S42, use the uniform multivariate average mutual information method and the false nearest neighbor method to estimate the time delay and optimal dimension of the EEG IMF envelope sequence;

[0058] S43, for the t-th data point, use the k-nearest neighbor method to find the D + 1 nearest neighbor nodes in the shadow manifold M X denoted as t1,..., t D+1 ;

[0059] S44, calculate the weight ω of every m neighbor nodes according to the distance between x(t) and the neighbor nodes m , and the formula is as follows:

[0060]

[0061] where, d[x(t), x(t m )] is the Euclidean distance between two vectors;

[0062] S45, map t1,..., t D+1 to Y, and obtain the cross-mapping estimation value from X to Y for the t-th data point through weighted summation The formula is as follows:

[0063]

[0064] Y(t m ) represents the value obtained by mapping the coordinates of the m-th neighbor node obtained from X to Y;

[0065] S46. Calculate the absolute value of the Pearson correlation coefficient between the cross-mapping estimate from the EEG frequency-band feature sequence X to the RR interval feature sequence Y and the true value on the RR interval feature sequence Y as the heart-to-brain non-linear coupling coefficient CCM Y→X , and the formula is as follows:

[0066]

[0067] wherein, corr(·) represents the Pearson correlation operation, represents the corresponding set of cross-mapping estimates from X to Y;

[0068] Similarly, according to steps S43 to S46, calculate the cross-mapping estimates from the RR interval feature sequence Y to the EEG frequency-band feature sequences X, and calculate the absolute value of the correlation coefficient with the true values on the EEG frequency-band feature sequences X as the brain-to-heart non-linear coupling coefficient CCM X→Y ;

[0069] If the CCM value is close to 0, it indicates a weak non-linear coupling, and if it is close to 1, it indicates a strong coupling.

[0070] Furthermore, the steps for establishing a depression symptom assessment model based on the heart rate variability feature set and the heart-brain coupling feature set are as follows:

[0071] S51. Obtain the scale scores of multiple users based on the clinical assessment scale for depression, and divide the users into a healthy group and a depression group according to the scale scores; through statistical analysis, obtain the heart rate variability feature set and the heart-brain coupling feature set with significant inter-group differences; form a depression feature set from the heart rate variability feature sets and the heart-brain coupling feature sets with inter-group differences of multiple users, and divide it into a training set and a test set according to the ratio of 8:2;

[0072] S52. Use a machine learning method to train the depression symptom assessment model with the training set;

[0073] S53. Use K-fold cross-validation to verify the depression symptom assessment model, and adjust the hyperparameters to save the depression symptom assessment model that meets the conditions.

[0074] Compared with the prior art, the significant effects of the present invention are as follows:

[0075] 1. In the implementation method of the present invention, through the synchronous acquisition and refined processing of electrocardiogram (ECG) and electroencephalogram (EEG) signals, it can ensure the accurate matching of the ECG RR interval and EEG data, and fully exploit the signal synchronization information. The multivariate empirical mode decomposition is used to adaptively extract different frequency band components of the brain, and the Hungarian algorithm is used to solve the problem of component matching among individuals, which fully preserves the nonlinear structure of the signals and provides a basis for the quantification of nonlinear heart-brain coupling. From the perspective of nonlinear dynamics, the coupling relationship between the heart and the brain is quantified, and combined with heart rate variability analysis, it can more comprehensively reflect the autonomic nerve function abnormality of patients with depression, providing a scientific basis for the objective quantitative assessment of depression;

[0076] 2. In the implementation method of the present invention, not only the traditional heart rate variability characteristics are concerned, but also the convergent cross-mapping method is applied to the synchronous heart and brain characteristic time series to construct a multi-channel and multi-frequency band bidirectional nonlinear causal heart-brain coupling feature set. By analyzing the differences between the healthy group and the depressive group, the features highly correlated with depression are found, and combined with machine learning to construct a depressive symptom assessment model, and the depressive scale score is predicted through the depressive symptom assessment model, which can avoid the limitation of the traditional scale relying on subjective assessment and can be used for the early screening of depression;

[0077] 3. The system of the present invention only relies on resting-state EEG and ECG signals, without external task stimulation, reducing the requirements for user cooperation. At the same time, the detection process is non-invasive and convenient, suitable for large-scale clinical screening and remote monitoring, improving the clinical practicability. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1 is a schematic structural diagram of the depression assessment system of the present invention;

[0079] Figure 2 is a flowchart of the implementation method of the present invention; Figure 3 is a predicted depression scale score graph obtained by predicting the score of the Hamilton 24-item scale using the test set. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0080] The present invention will be further described in detail below with reference to the accompanying drawings of the specification and the specific embodiments.

[0081] As Figure 1 shown, this embodiment discloses a depression assessment system based on heart-brain coupling analysis, including the following modules:

[0082] A data acquisition module 1, which collects the clinical assessment scale of the user's depression and synchronously acquires the original ECG signal and the original EEG signal of the user in a certain period of time in the resting state by using a multi-channel acquisition device;

[0083] The heart and brain signal preprocessing module 2 processes the collected original EEG signals and original ECG signals;

[0084] The feature sequence extraction module 3 calculates the EEG feature sequences in each frequency band according to the preprocessed EEG signals, and is also used to calculate the RR interval feature sequence (the time interval sequence of adjacent R peaks of the ECG signal) and the heart rate variability feature set according to the preprocessed ECG signals;

[0085] The coupling estimation module 4 calculates the heart-brain coupling features according to the extracted EEG feature sequences in each frequency band and the RR interval feature sequence. The heart-brain coupling features involve two interaction directions from heart to brain and from brain to heart, the whole brain channels and multiple EEG frequency bands;

[0086] The depression assessment model establishment module 5 constructs a depression symptom assessment model based on the heart-brain coupling features, heart rate variability features and the data of the depression clinical assessment scale to quantitatively evaluate the depression degree of an individual;

[0087] The output result module 6 evaluates and outputs the depression degree of the user based on the depression symptom assessment model.

[0088] The data acquisition module 1 synchronously acquires ECG signals and EEG signals through non-invasive sensors. It is preferred to use an electrode cap with 32 or more lead numbers and a layout that conforms to the 10-20 system electrode placement method calibrated by the International EEG Society for collecting EEG signals. Two additional leads are extended on the EEG amplifier, and heart electrode patches are used to closely adhere to the strongest pulse fluctuation point on the user's wrist and the inner side of the ankle to keep the EEG data and ECG data collected simultaneously. Before signal acquisition, the user is required to sit on a comfortable seat and sit still for 8-10 minutes. After the body and mind are relaxed, the user keeps eyes closed and rests but remains awake, and collects EEG data and ECG data for more than 5 minutes. During the signal acquisition process, the user is required to maintain a sitting posture, with both hands perpendicular to both sides, relax as much as possible, avoid facial muscle activities such as deep breathing, frowning, and swallowing, and avoid body movement. The online signal sampling frequency is set to 1000 Hz. In addition, before collecting EEG and ECG signals, the depression clinical scale scores of the user are collected. The preferred scale is the Hamilton 24-item Depression Scale (HAMD-24), and the Montgomery Depression Scale (MADRS), Beck Depression Inventory (BDI), etc. can also be selected.

[0089] The time length of the preprocessed ECG and EEG signals intercepted by the feature sequence extraction module 3 is preferably not less than 3 minutes. The target frequency bands of the IMF of interest for the preprocessed EEG signals are delta (1-4Hz), theta (4-8Hz), alpha (8-13Hz), beta (13-30Hz), gamma (30-45Hz), or delta (1-4Hz), theta (4-8Hz), alpha (8-13Hz).

[0090] Depression assessment model building module 5 selects SVM regression (SVR) to train the depression feature set, takes the Hamilton 24-item depression rating scale as the target variable, and trains the model to quantitatively predict the degree of depression. The prediction accuracy, mean square error (MSE), R 2 The depression assessment model is evaluated by using indicators such as values ​​to ensure that the depression assessment model can accurately quantify the individual's depression level.

[0091] like Figure 2 FIG. 1 is a flowchart of a method for implementing depression assessment based on resting heart-brain coupling analysis, including the following steps:

[0092] Step 1: collect the clinical assessment scale for depression and collect the original EEG signals and original ECG signals;

[0093] The data acquisition module 1 collects synchronous multi-lead EEG signals and ECG signals of the same length of time from multiple users, and also includes a clinical assessment scale for depression, to obtain depression scale scores for multiple users.

[0094] Step 2: preprocessing the collected original EEG signals and original ECG signals;

[0095] The steps of preprocessing the original EEG signal and the original ECG signal by the heart and brain signal preprocessing module 2 are as follows:

[0096] Step 21, applying a 0.5-45 Hz bandpass filter and a 50 Hz notch filter to the original ECG signal and the original EEG signal for filtering and denoising, and resampling to 250 Hz;

[0097] Step 22, applying the Pan-Tompking algorithm to detect the R peak in the ECG signal and perform visual inspection and correction, and synchronously marking the EEG signal according to the R peak;

[0098] Step 23, through visual inspection, manually remove bad segments and interpolate bad guides; remove heart and brain data segments with multiple channels simultaneously affected by head motion and drift pollution, and each deleted data segment starts and ends with an R peak, that is, the length of the deleted data segment is an integer multiple of the heart cycle;

[0099] Step 24: Perform independent component analysis on the EEG signals. Select the runica algorithm, set the number of PCA components to the total number of electrode derivatives minus the interpolated bad derivatives, and use the single-lead ECG data as an additional input to enhance the recognition of cardiac artifacts. Visually observe each component and manually remove the components related to eye movements and cardiac artifacts; the number of components removed for each individual user is controlled within 4;

[0100] Step 25: Perform whole-brain average re-referencing on the EEG signals.

[0101] Step 3: Extract feature sequences from the preprocessed EEG and ECG signals, and extract traditional heart rate variability features;

[0102] The feature sequence extraction module 3 takes the same R peak as the starting point for the preprocessed ECG and EEG signals, and intercepts data not shorter than 3 minutes. The implementation of calculating the feature sequences of each frequency band of EEG, the RR interval feature sequence, and the traditional heart rate variability features is as follows:

[0103] (31) Extract feature sequences from the preprocessed EEG signals;

[0104] The main steps are as follows:

[0105] Step 311: Decompose the preprocessed EEG signals using multivariate empirical mode decomposition to adaptively extract a set of intrinsic mode functions (IMFs). These IMF components can capture the intrinsic frequency characteristics of brain activities and retain the non-linear characteristics of the original signals. For each user, all EEG channels obtain the same number of IMFs, but the number of IMFs may be different among different individuals;

[0106] Step 312: Since the number of IMFs may be different for the EEG signals of different users after multivariate empirical mode decomposition, and the corresponding frequency components are not exactly the same, a matching algorithm needs to be used to select the IMF components in the frequency band of interest for different individuals. The steps of extracting the IMFs in the frequency band of interest through central frequency calculation and Hungarian algorithm matching are as follows:

[0107] Step 3121: Set the target frequency band range of the frequency band of interest;

[0108] Step 3122: Each IMF component is a time series signal with a total length (i.e., the total number of points) of N. For the value of any point n, denoted as IMF(n), where n ∈ [0, N], representing the data point index. Apply the Hilbert transform to IMF(n) to obtain H(IMF(n)), and construct the analytic signal z(n), the formula is as follows:

[0109]

[0110] z(n) = IMF(n) + jH(IMF(n)) (2)

[0111] Calculate the instantaneous phase φ(n) of the IMF component and derive the instantaneous frequency f(n). The formulas are as follows:

[0112]

[0113] Perform a time average on the instantaneous frequency f(n) to obtain the central frequency F of this IMF component. The formula is as follows:

[0114]

[0115] Where N represents the total number of points of the IMF component.

[0116] Step 3123: Calculate the central frequency of all IMF components of each user respectively, and screen out the users whose central frequencies of the IMF continuously fall within the entire target frequency band range. Use the IMF components of this user that fall within the target frequency band range as the reference set for matching the interested IMF components;

[0117] Step 3124: For the IMF set of each user, calculate the cost matrix between the IMF set and the reference set. The elements of the cost matrix are obtained by subtracting the Pearson correlation coefficient between the amplitudes of two IMFs from 1. The lower the cost, the more similar the two IMFs are. The formula is as follows:

[0118]

[0119] Where c i,j represents the cost, represents the i-th IMF component of the reference set, represents the j-th IMF component in the IMF set other than the reference set, and corr(·) represents the Pearson correlation operation;

[0120] At the same time, introduce cost matrix optimization measures. Since the IMFs obtained by multivariate empirical mode decomposition are usually sorted from high to low in frequency, when matching, the cost of IMF pairs that are far apart is set to infinity to avoid unreasonable matching;

[0121] Step 3125: Use the Hungarian algorithm to solve the optimal matching and find the minimum-cost allocation scheme between the IMF and the target frequency band to minimize the total allocation error;

[0122] Step 3126: After matching by the Hungarian algorithm, the IMF components of each user correspond one by one to the target frequency band, and the specific interested IMF components are obtained.

[0123] Step 313: Apply the Hilbert transform to the IMF components in the frequency band of interest to obtain the envelope curve of the IMF components, and downsample the envelope curve by 15 times to reduce the data dimension and improve the calculation efficiency. Finally, the characteristic sequences X of each EEG frequency band representing central nervous activity are obtained. For the value of any point n' in X, denoted as X(n'), the formula is as follows:

[0124]

[0125] where Downsample(·) represents the downsampling operation, r represents the downsampling ratio, which is 15 here; n' represents the data index after downsampling, n'∈[0,L], represents the floor operation.

[0126] (32) Extract the RR interval characteristic sequence from the preprocessed electrocardiogram signal;

[0127] The main steps are as follows:

[0128] Step 321: Based on the R-peak events marked in the preprocessed electrocardiogram signal, calculate the time interval between adjacent R-peaks of the electrocardiogram signal, and construct the original RR interval sequence;

[0129] RR K =t K -t K-1 (8)

[0130] where RR K represents the time interval between the Kth adjacent R-peaks, t K represents the time point when the Kth R-peak appears, t K-1 represents the time point when the (K - 1)th R-peak appears;

[0131] Step 322: Since the heart rate is unevenly sampled, use cubic spline interpolation to uniformly interpolate the original RR interval sequence at a sampling rate of 250 Hz, and downsample the uniformly sampled RR interval sequence by 15 times. Finally, the RR interval characteristic sequence Y representing autonomic nerve activity is obtained.

[0132] (33) Extract traditional heart rate variability characteristics from the original and interpolated RR interval characteristic sequences;

[0133] The main steps are as follows:

[0134] Step 331: Calculate the heart rate (HR), standard deviation of RR intervals (SDNN), root mean square of successive differences (RMSSD), and ratio of the percentage of the number of changes in consecutive normal sinus intervals exceeding 50 milliseconds (pNN50) in the time domain of the original RR interval sequence;

[0135] Step 332: Calculate the low-frequency component LF with a frequency range of 0.04 - 0.15 Hz, the high-frequency component HF with a frequency range of 0.15 - 0.4 Hz, and their ratio LF / HF on the frequency domain of the interpolated RR interval feature sequence.

[0136] Step 333: Construct a traditional heart rate variability feature set including HR, SDNN, RMSSD, pNN50, LF, HF, and LF / HF.

[0137] Step Four: Calculate the causal nonlinear coupling strength between the EEG feature sequences of each frequency band and the RR interval feature sequence to obtain the nonlinear coupling features from heart to brain and from brain to heart.

[0138] Use the coupling estimation module 4 to calculate the causal nonlinear coupling strength between the EEG feature sequences X of each frequency band and the RR interval feature sequence Y using convergent cross mapping. The specific steps are as follows:

[0139] Step 41: For the EEG feature sequences X of each frequency band and the RR interval feature sequence Y, perform phase space reconstruction respectively. For the value of the t-th data point on the EEG feature sequences X of each frequency band, denoted as X(t), construct a D-dimensional vector x(t) using the following formula:

[0140] x(t) = <X(t), X(t - τ), X(t - 2τ), …, X(t - (D - 1)τ)>, where t = 1 + (D - 1)τ, …, L (9)

[0141] where τ is the constant time delay, D is the embedding dimension, which are two key technical parameters for constructing the phase space. L is the total number of data points in the sequence length. The set of all D-dimensional vectors constitutes the shadow manifold M X 。

[0142] Using the same method as formula (8), apply it to the RR interval feature sequence Y. For the t-th data point, construct a D-dimensional vector y(t). Finally, the set of all D-dimensional vectors obtains the shadow manifold M Y 。

[0143] Step 42: Considering that the EEG is a multi-dimensional time series, use the uniform multi-average mutual information method and the false nearest neighbor method extended to multi-variate time series to estimate the time delay and optimal dimension of the EEG IMF envelope sequence. The RR interval feature sequence Y uses the same phase space parameters.

[0144] Step 43: For the t-th data point, use the k-nearest neighbor method to find the D + 1 neighbor nodes in the shadow manifold M X that are closest to x(t), denoted as t1,..., t D+1 。

[0145] Step 44: Calculate the weight ω of the m-th neighbor node according to the distance between x(t) and the neighbor nodes. m , and the formula is as follows:

[0146]

[0147] where d[x(t), x(t m )] is the Euclidean distance between two vectors, and e is the natural constant.

[0148] Step 45: Map t1,..., t D+1 to Y, and obtain the cross-map estimate value of the t-th data point from X to Y through weighted summation. The formula is as follows:

[0149]

[0150] where Y(t m ) represents the value obtained by mapping the coordinates of the m-th neighbor node obtained from X to Y.

[0151] Step 46: Calculate the absolute value of the Pearson correlation coefficient between the cross-map estimate value from the EEG frequency-band feature sequence X to the RR interval feature sequence Y and the true value on the RR interval feature sequence Y, and use it as the cardio-cerebral nonlinear coupling coefficient CCM. Y→X , and the formula is as follows:

[0152]

[0153] where corr(·) represents the Pearson correlation operation, represents the corresponding set of cross-map estimate values from X to Y. Since X contains both its own information and the features from Y, the predicted value of Y can be obtained through the shadow manifold M X of X, that is The strength of this causal relationship is quantified through the Pearson correlation.

[0154] Similarly, according to formulas (10) - (13), the cross-map estimate value from the RR interval feature sequence Y to the EEG frequency-band feature sequence X can also be calculated, and the absolute value of the correlation coefficient is calculated with the true value on the EEG frequency-band feature sequence X, as the cerebro-cardiac nonlinear coupling coefficient. The specific steps are as follows:

[0155] Step 47: For the t-th data point, use the k-nearest neighbor method to find the D + 1 neighbor nodes in the shadow manifold M Y that are closest to y(t);

[0156] Step 48: Calculate the weight ω of the m-th neighbor node according to the distance between y(t) and each neighbor node.m ;

[0157] Step 49: Map the D + 1 neighbor nodes closest to y(t) onto X, and obtain the cross - mapping estimate value of the t - th data point from Y to X through weighted summation.

[0158] Step 410: Calculate the absolute value of the Pearson correlation coefficient between the cross - mapping estimate value from the RR - interval feature sequence Y to the EEG feature sequences X in each frequency band and the true value on the EEG feature sequences X in each frequency band, and use it as the brain - to - heart non - linear coupling coefficient CCM. X→Y ;

[0159] Among them, steps 43 to 46 can be executed before steps 47 to 410, can be executed after steps 47 to 410, or can be executed simultaneously with steps 47 to 410.

[0160] If the CCM value is close to 0, it indicates a weak non - linear coupling; if it is close to 1, it indicates a strong coupling. The non - linear coupling characteristics from heart to brain and from brain to heart involve all - brain channels and multiple frequency bands of interest, jointly constituting the heart - brain coupling feature set.

[0161] Step Five: Establish a depression symptom assessment model based on the heart rate variability feature set and the heart - brain coupling feature set;

[0162] The steps for the depression assessment model establishment module 5 to establish a depression symptom assessment model based on the above - mentioned heart rate variability feature set and heart - brain coupling feature set are as follows:

[0163] Step 51: Screen depression - related features based on statistical analysis. Obtain the scale scores of multiple users based on the clinical assessment scale for depression, and divide the users into a healthy group and a depression group according to the scale scores. Through statistical analysis, obtain the heart rate variability feature set and heart - brain coupling feature set with significant inter - group differences. The heart rate variability feature set and heart - brain coupling feature set with inter - group differences of multiple users constitute the depression feature set, which is divided into a training set and a test set according to an 8:2 ratio;

[0164] Step 52: Use traditional machine learning methods to train the depression symptom assessment model with the training set. The input of the depression symptom assessment model is the depression feature set, and the output is the score of the clinical assessment scale for depression, which is used to evaluate the depression level of an individual.

[0165] Step 53: To avoid overfitting, K - fold cross - validation can be used to verify the depression symptom assessment model, and adjust the hyperparameters to ensure the generalization ability of the depression symptom assessment model, and save the depression symptom assessment model that meets the conditions.

[0166] Step 6: According to the calculated depression feature set, the depression symptom assessment model outputs a quantitative score of the depression level;

[0167] Furthermore, the output result module 6, according to the depression feature set calculated in Steps 1 to 4, as the input of the depression symptom assessment model, outputs its quantitative score of the depression level.

[0168] In this embodiment, the electroencephalogram, electrocardiogram data and the scores of the Hamilton 24-item scale of 70 users are collected. According to the scores of the Hamilton 24-item scale, they are divided into 32 healthy people and 38 depressed patients with different degrees. First, the cardio-cerebral coupling features and heart rate variability features related to depressive symptoms are extracted, so as to establish a depression feature set. The depression feature set is divided into a training set and a test set. The SVR model is trained with the training set, and the scores of the Hamilton 24-item scale are predicted with the test set. The predicted depression scale scores are as Figure 3 shown.

[0169] As Figure 3 can be seen, the method of the present invention can extract the cardio-cerebral coupling features and heart rate variability features related to depressive symptoms, establish a reliable depression symptom assessment model, and provide a more objective and accurate depression assessment tool for clinical practice.

Claims

1. A depression assessment system based on resting-state cardio-cerebral coupling analysis, characterized in that, include: The data acquisition module uses a multi-channel acquisition device to synchronously collect the user's original ECG signal and original EEG signal within a certain period of time in a resting state, and collects the user's clinical assessment scale for depression; The heart and brain signal preprocessing module performs signal processing based on the collected original EEG signals and original ECG signals; The feature sequence extraction module calculates the feature sequence of each frequency band of EEG based on the preprocessed EEG signal; According to the preprocessed ECG signal, the RR interval feature sequence and the heart rate variability feature set are calculated; The coupling estimation module calculates the heart-brain coupling characteristics based on the extracted EEG frequency band feature sequences and RR interval feature sequences; The heart-brain coupling features include two interaction directions: heart-to-brain and brain-to-heart, whole-brain multi-channels and multiple EEG frequency bands; The depression assessment model building module builds a depression symptom assessment model based on the characteristics of heart-brain coupling, heart rate variability, and data from the clinical assessment scale for depression, and quantitatively assesses the degree of depression of an individual. The output result module is used to output the user's depression level score.

2. A method for realizing depression assessment based on resting-state heart-brain coupling analysis, characterized in that, Based on the depression assessment system as claimed in claim 1, a depression symptom assessment model is established by quantifying the heart-brain coupling characteristics and combining heart rate variability analysis, and finally a quantitative score of depression degree is output; The steps include: S1, collect synchronous multi-lead EEG signals and ECG signals of the same length of time from multiple users; S2, preprocessing the collected EEG signals and ECG signals; S3, extracting feature sequences from the preprocessed EEG signals and ECG signals, and extracting traditional heart rate variability features; S4, calculate the causal nonlinear coupling strength between the characteristic sequences of each EEG frequency band and the characteristic sequence of RR intervals, and obtain the nonlinear coupling characteristics from heart to brain and from brain to heart. The two nonlinear coupling characteristics together constitute the heart-brain coupling feature set; S5, based on the heart rate variability feature set and the heart-brain coupling feature set, a depressive symptom assessment model was established; S6, based on the calculated depression feature set, the depression symptom assessment model outputs a quantitative score of the degree of depression.

3. The method for implementing depression assessment based on resting state cardio-cerebral coupling analysis according to claim 2, wherein The collected EEG and ECG signals are preprocessed by the heart and brain signal preprocessing module. The detailed implementation steps are as follows: S21, applying a 0.5-45 Hz bandpass filter and a 50 Hz notch filter to the collected ECG and EEG signals for filtering and denoising, and resampling to 250 Hz; S22, Pan-Tompking algorithm was applied to detect R peaks in ECG signals and visually inspected and corrected, and EEG signals were synchronously marked according to R peaks; S23, through visual inspection, manually remove bad segments and interpolate bad guides; remove heart and brain data segments contaminated by head movement and drift, and each deleted data segment starts and ends with an R peak; S24, perform independent component analysis on the EEG signal and use single-lead ECG data as additional input; visually observe each component and manually remove components related to eye movements and cardiac artifacts; S25, the EEG signal is averaged and re-referenced across the entire brain.

4. The method for implementing depression assessment based on resting heart-brain coupling analysis according to claim 2, wherein The feature sequence extraction module extracts data of the same length from the preprocessed ECG signal and EEG signal with the same R peak as the starting point; The implementation steps for extracting the feature sequence from the preprocessed EEG signals are as follows: S311, decompose the preprocessed EEG signals using multivariate empirical mode decomposition to extract a set of IMFs; for each user, the same number of IMFs are obtained for all EEG channels; S312, adopt a matching algorithm to select the IMF components in the frequency band of interest among different individuals; S313, apply the Hilbert transform to the IMF components in the frequency band of interest, obtain the envelope curve of the IMF components, and downsample the envelope curve to obtain the feature sequences X of each frequency band of EEG representing central nervous activity; The implementation steps for extracting the RR interval feature sequence from the preprocessed ECG signals are as follows: S321, mark the R peaks in the preprocessed ECG signals, calculate all RR interval times of the ECG signals, and construct the original RR interval sequence; S322, use cubic spline interpolation to uniformly interpolate the original RR interval sequence at a sampling rate of 250 Hz, and downsample the uniformly sampled RR interval sequence by 15 times to obtain the RR interval feature sequence Y representing autonomic nerve activity; The implementation steps for extracting traditional heart rate variability features from the original and interpolated RR interval feature sequences are as follows: S331, calculate the heart rate HR, the standard deviation SDNN of the RR intervals, the root mean square of successive differences RMSSD, and the ratio pNN50 of the percentage of the number of changes in consecutive normal sinus intervals exceeding 50 milliseconds in the time domain of the original RR interval sequence; S332, calculate the low-frequency component LF with a frequency of the RR interval sequence in the range of 0.04 - 0.15 Hz, the high-frequency component HF with a frequency in the range of 0.15 - 0.4 Hz, and the ratio LF / HF in the frequency domain of the interpolated RR interval feature sequence; S333, construct a traditional heart rate variability feature set including HR, SDNN, RMSSD, pNN50, LF, HF, and LF / HF.

5. The implementation method for depression assessment based on resting heart-brain coupling analysis according to claim 4, wherein, Extract the IMFs in the frequency band of interest through central frequency calculation and Hungarian algorithm matching, including the following steps: S3121, set the target frequency band range of the frequency band of interest; S3122, each IMF component is a time series signal with a total number of points N. For the value of any point n, denoted as IMF(n), apply the Hilbert transform to IMF(n) to obtain H(IMF(n)), and construct the analytic signal z(n), with the formula as follows: z(n) = IMF(n) + jH(IMF(n)) Calculate the instantaneous phase φ(n) of the IMF component, and derive the instantaneous frequency f(n), with the formula as follows: Perform time averaging on the instantaneous frequency f(n) to obtain the central frequency F of this IMF component, with the formula as follows: where n ∈ [0, N], representing the data point index; S3123, calculate the central frequencies of all IMFs of each user respectively, screen out the users whose central frequencies of IMFs continuously fall within the entire target frequency band range, and use the IMF components of this user that fall within the target frequency band range as the reference set for matching the interested IMF components; S3124, for the IMF set of each user, calculate the cost matrix between the IMF set and the reference set; S3125, the Hungarian algorithm is used to solve the optimal matching, find the minimum-cost allocation scheme between the IMF and the target frequency band, and minimize the total allocation error; S3126, after the matching by the Hungarian algorithm, the IMF components of each user correspond one-to-one with the target frequency band, and the specific IMF components of interest are obtained.

6. The method for realizing depression assessment based on resting state cardio-cerebral coupling analysis according to claim 4, wherein The coupling estimation module uses convergent cross mapping to calculate the causal nonlinear coupling strength between the characteristic sequences of each frequency band of EEG and the characteristic sequence of RR interval. The specific steps are as follows: S41, for the characteristic sequence X of each frequency band of EEG and the characteristic sequence Y of RR interval, phase space reconstruction is performed respectively; for the value of the t-th data point on the characteristic sequence X of each frequency band of EEG, denoted as X(t), a D-dimensional vector x(t) is constructed, and the formula is as follows: x(t) = <X(t), X(t - τ), X(t - 2τ), …, X(t - (D - 1)τ), t = 1 + (D - 1)τ, …, L where τ is the constant time delay, D is the embedding dimension; L is the total number of points in the sequence length; These sets of D-dimensional vectors x(t) form the shadow manifold M of the phase space X ; The same method is applied to the RR interval feature sequence Y, and a D-dimensional vector y(t) is constructed for the t-th data point. Finally, the set of all D-dimensional vectors yields the shadow manifold M Y ; S42, the uniform multivariate average mutual information method and the false nearest neighbor method are used to estimate the time delay and the optimal dimension of the EEG IMF envelope sequence; S43. For the t-th data point, use the k-nearest neighbor method to find the D + 1 neighbor nodes in the shadow manifold M that are closest to x(t), denoted as t1,..., t X ; D+1 ; S44. Calculate the weight ω of every m neighbor nodes according to the distance between x(t) and the neighbor nodes. The formula is as follows: m , as follows: where d[x(t), x(t m )] is the Euclidean distance between two vectors; S45, map t1,...,t D+1 onto Y, and obtain the cross - mapping estimation value of the t - th data point from X to Y through weighted summation The formula is as follows: Y(t m ) represents the value obtained by mapping the coordinates of the m-th neighbor node obtained from X onto Y; S46. Calculate the absolute value of the Pearson correlation coefficient between the cross-mapping estimate from the EEG frequency-band feature sequence X to the RR interval feature sequence Y and the true value on the RR interval feature sequence Y, and use it as the heart-to-brain nonlinear coupling coefficient CCM. Y→X , and the formula is as follows: where corr(·) represents the Pearson correlation operation, represents the corresponding set of cross - mapping estimation values from X to Y; Similarly, according to steps S43 to S46, calculate the cross - mapping estimate value from the RR - interval feature sequence Y to the EEG feature sequences X in each frequency band, and calculate the absolute value of the correlation coefficient with the true value on the EEG feature sequences X in each frequency band as the non - linear coupling coefficient CCM from brain to heart X→Y ; If the CCM value is close to 0, it indicates weak nonlinear coupling, and close to 1 indicates strong coupling.

7. The method for realizing depression assessment based on resting heart-brain coupling analysis according to claim 2, wherein, The steps to establish a depression symptom assessment model based on the heart rate variability feature set and the heart-brain coupling feature set are as follows: S51, based on the clinical assessment scale for depression, the scale scores of multiple users are obtained, and the users are divided into a healthy group and a depression group according to the scale scores; through statistical analysis, the heart rate variability feature set and the heart-brain coupling feature set with significant inter-group differences are obtained; the heart rate variability feature set and the heart-brain coupling feature set with inter-group differences of multiple users are combined to form a depression feature set, which is divided into a training set and a test set according to the ratio of 8:2; S52, the machine learning method is used to train the depression symptom assessment model with the training set; S53, the K-fold cross-validation is used to verify the depression symptom assessment model, and the hyperparameters are adjusted to save the depression symptom assessment model that meets the conditions.

Images