Adolescent depression prediction system

Abstract

The present application relates to the technical field of health care informatics, in particular to a system for predicting depression of teenagers, comprising an acoustic characteristic representation module, including a co-integrated ultrasonic transducer array and electroencephalogram electrode, an electrode substrate embedded with an acoustic impedance sensor, and a temperature-sensitive acoustic coupling adhesive layer on the surface of the electrode substrate; a multi-layer medium acoustic modeling module: used for establishing an accurate head tissue acoustic transmission model according to individual anatomical structure characteristics, for quantifying the frequency dispersion effect generated when the electroencephalogram signal propagates in different tissue layers, etc.; solving the problem that in the prior art, due to the change of the myelin sheath thickness of teenagers, the tissue-specific dispersion has a systematic influence on the functional connection network, resulting in that the depression-related network abnormalities and the development-related transmission changes cannot be distinguished.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of healthcare informatics, in particular to a youth depression prediction system. BACKGROUND

[0002] Early prediction and intervention of youth depression is crucial for preventing the disease from worsening; existing depression prediction methods mainly rely on clinical assessment, questionnaire survey or behavior data analysis based on machine learning, but these methods are highly subjective and susceptible to individual reporting bias, and lack objective physiological indicators. Research has found that the early stage of youth depression is related to electroencephalogram signals, and by obtaining electroencephalogram signals as the basis for data analysis of youth depression prediction, there is higher objectivity, however, the dispersion effect and phase distortion of electroencephalogram signals occur when penetrating different tissue layers. Especially in the prefrontal region, due to the uneven thickness of the skull and the influence of cerebrospinal fluid flow, frequency-dependent phase delay occurs in the transmission process of electroencephalogram signals, which can lead to false estimation of brain region synchronization in functional connectivity analysis. This dispersion effect has an individualized spatial distribution pattern and is closely related to the development state of the brain. The youth period is a critical stage of white matter myelination in the brain, and changes in myelin thickness will further change the signal transmission characteristics. Existing technologies treat electroencephalogram signals as propagating in a homogeneous medium, completely ignoring the systematic impact of this tissue-specific dispersion on functional connectivity networks, resulting in the inability to distinguish between depression-related network abnormalities and development-related transmission changes, thus leading to biased prediction results. SUMMARY

[0003] The technical problem to be solved by the present application is how to improve the youth depression prediction system to solve the problem of biased prediction results caused by the systematic impact of tissue-specific dispersion of myelin thickness changes on functional connectivity networks, which makes it impossible to distinguish between depression-related network abnormalities and development-related transmission changes.

[0004] To solve the above technical problems, the technical scheme adopted by the present application is:

[0005] The youth depression prediction system based on brain-computer interface comprises:

[0006] The acoustic characteristic representation module comprises a co-integrated ultrasonic transducer array and electroencephalogram electrode, an acoustic impedance sensor is embedded in the electrode substrate, and a temperature-sensitive acoustic coupling adhesive layer is provided on the surface of the electrode substrate;

[0007] The multi-layer medium acoustic modeling module is used to establish an accurate head tissue acoustic transmission model according to individual anatomical structure characteristics, and to quantify the dispersion effect of electroencephalogram signals when propagating in different tissue layers;

[0008] A phase compensation filtering module is configured to design and implement phase compensation according to the acoustic modeling result, so as to correct phase distortion of the electroencephalogram signal generated in the tissue transmission process;

[0009] A frequency dispersion invariant functional connectivity calculation module is configured to calculate a frequency dispersion invariant functional connectivity index according to the compensated signal output by the adaptive phase compensation algorithm, so as to ensure that the functional connectivity network analysis is immune to residual phase distortion through introduction of a frequency dispersion stability index and a dynamic weight mechanism;

[0010] A Bayesian network module is configured to construct a dynamic Bayesian network model according to the result of the frequency dispersion invariant functional connectivity calculation, so as to perform time series prediction and uncertainty quantification of depression risk; and a more robust prediction framework is established through integration of the frequency dispersion stability index and the acoustic feature;

[0011] A depression intervention decision module is configured to construct a multi-level and adaptive decision system for accurately triggering and adjusting a depression intervention strategy according to the output of the Bayesian network modeling, in combination with the frequency dispersion stability index and the acoustic feature.

[0012] Further, in the use of the multi-layer medium acoustic modeling module in the above-mentioned adolescent depression prediction system, the following is specifically included: obtaining the skull thickness distribution of an individual based on CT image data; measuring the cerebrospinal fluid pulsation velocity through Doppler ultrasound; and calculating the acoustic transmission matrix of each tissue layer:

[0013] Ti =

[0014] wherein, is the complex wave number of the i-th layer (rad / m); ;

[0015] is the frequency;

[0016] is the frequency-dependent phase velocity;

[0017] is the frequency-dependent attenuation coefficient (Np / m);

[0018] is the thickness of the i-th layer; obtained through CT image segmentation, with an accuracy of ±0.1 mm;

[0019] is the acoustic impedance of the i-th layer (Rayl);

[0020] ;

[0021] is the medium density (kg / m³);

[0022] the acoustic wave propagation speed for the i-th layer of medium;

[0023] the imaginary unit;

[0024] Derive the acoustic transfer function of the system from the total transmission matrix:

[0025]

[0026] where, is the complex transfer function, containing amplitude and phase information; is the reference impedance;

[0027]

[0028] Further, in the use of the phase compensation filter module in the above-mentioned adolescent depression prediction system, specifically includes:

[0029] The acoustic transfer function obtained based on the multi-layer medium acoustic modeling , calculate the relative phase difference between channels i and j:

[0030]

[0031] where, is the acoustic transfer function of channel i, from the acoustic modeling output, is a complex function, containing amplitude and phase information;

[0032] is the acoustic transfer function of channel j, also from the acoustic modeling output;

[0033] is the operator extracting the phase angle of the complex number, the unit is radian (rad), the output range is [-π, π];

[0034] Indicates the phase distortion of channel i relative to channel j, which varies with frequency;

[0035] Construct a frequency domain compensation filter for correcting phase distortion and dispersion effect:

[0036]

[0037] where, is the imaginary unit;

[0038] is the dispersion compensation coefficient, the unit is second (s), used to adjust the strength of group delay compensation, determined by maximum likelihood estimation;

[0039] To correct static phase offset, ensure that the inter-channel phase difference after compensation is zero;

[0040] To correct the dispersion effect caused by group delay variation, reduce the frequency-dependent phase distortion.

[0041] Further, in the use of the dispersion-invariant functional connectivity calculation module in the above-mentioned adolescent depression prediction system, specifically comprising:

[0042] Perform empirical mode decomposition on the compensated signal;

[0043] Extract the instantaneous phase of each IMF component;

[0044] Calculate the dispersion stability index:

[0045] Wherein: is the standard deviation (rad) of the phase difference sequence

[0046] is a normalization factor corresponding to the maximum possible phase difference change range;

[0047] Calculate the dispersion-weighted phase lag index:

[0048]

[0049] Wherein: is the attenuation coefficient, determined by cross-validation;

[0050] is the dispersion stability index, the value range [0, 1];

[0051] Construct a functional connectivity network considering dispersion stability;

[0052] Connection weight:

[0053] Wherein: is the directed connection weight of channel i to j;

[0054] is the total number of channels.

[0055] Further, in the use of the Bayesian network module in the above-mentioned adolescent depression prediction system, specifically comprising:

[0056] State space model definition:

[0057] ;

[0058] Wherein: is the dispersion adjustment coefficient, dimension ​×1, estimated via variational Bayesian method;

[0059] The average dispersion stability index is a scalar:

[0060]

[0061] This is the acoustic characteristic vector, which includes the acoustic impedance and transmission loss of each channel;

[0062] This is the dispersion stability adjustment coefficient;

[0063] This is the acoustic feature adjustment matrix;

[0064] Q is the process noise covariance matrix, with dimensions DS×DS, assumed to be a diagonal matrix. ;

[0065] Observation model:

[0066]

[0067] in: It is a nonlinear mapping function, implemented through a three-layer feedforward neural network;

[0068]

[0069] b1 is the hidden layer bias vector; b2 is the output layer bias vector; b3 is the output adjustment bias term;

[0070] The observation noise adjustment vector;

[0071] To observe noise;

[0072] Let be the dispersion sensitivity coefficient, a dimensionless positive real number, which controls the sensitivity of observation noise to dispersion effects. It determines the sensitivity of observation noise to dispersion effects when... When the deviation from the ideal value of 1 is observed, the rate of noise amplification is measured.

[0073] Initialize the model prior based on data from healthy populations:

[0074] State Priors:

[0075] in:

[0076] The average functional connectivity vector of 100 healthy adolescents;

[0077] Covariance matrix of functional connectivity for healthy population;

[0078] Compute posterior distribution based on variational Bayesian inference; Compute depression risk probability:

[0079]

[0080] wherein: , , are logistic regression coefficients, trained from historical data;

[0081] is the average clustering coefficient scalar.

[0082] Further, in the use of the depression intervention decision module in the above-mentioned adolescent depression prediction system, the use specifically comprises:

[0083] Construct a decision feature vector:

[0084] Feature vector composition:

[0085] wherein: is the depression risk probability, from Bayesian network modeling (0-1 range);

[0086] is the average frequency dispersion stability index (0-1 range);

[0087] is the acoustic feature vector;

[0088] : is the risk probability change rate;

[0089] is the global efficiency index scalar;

[0090] is the average clustering coefficient scalar;

[0091] is the modular index scalar;

[0092] Calculate the dynamic weight comprehensive intervention score:

[0093]

[0094] wherein, ;

[0095]

[0096] wherein: ;

[0097]

[0098] wherein, ;

[0099] is the L2 norm of the acoustic vector;

[0100] Adaptively adjusting the intervention parameter based on the comprehensive intervention score.

[0101] Further, in the use of the depression intervention decision module in the adolescent depression prediction system described above, the adaptively adjusting the intervention parameter based on the comprehensive intervention score specifically includes:

[0102] Adjusting the intensity of the neurofeedback:

[0103]

[0104] wherein, ;

[0105] ;

[0106] Adjusting the task difficulty coefficient:

[0107]

[0108] wherein,

[0109] Adjusting the duration of the continuous feedback:

[0110]

[0111] wherein, seconds; is the base duration;

[0112] is the change sensitivity coefficient.

[0113] The beneficial effects of the present application are: the cooperation of the ultrasonic transducer and the acoustic impedance sensor: the real-time monitoring of the tissue characteristics is realized, and the transmission path estimation error is reduced; the cooperation of the acoustic modeling and the phase compensation: the phase distortion caused by the frequency dispersion is accurately corrected, and the reliability of the connection network is improved; the cooperation of the frequency dispersion stability index and the Bayesian model: the recognition ability of the model to the real neural connection is enhanced. By improving the adolescent depression prediction system, the problem that the depression-related network abnormalities and development-related transmission changes cannot be distinguished due to the systematic influence of the tissue-specific frequency dispersion of the change of the myelin sheath thickness of adolescents in the prior art is solved. BRIEF DESCRIPTION OF DRAWINGS

[0114] Figure 1 is the structure block diagram of the adolescent depression prediction system related to the specific embodiments of the present application. DETAILED DESCRIPTION

[0115] To describe the technical contents of the present application, the purposes and effects achieved, the following will be described in combination with the embodiments and the accompanying drawings.

[0116] Please refer to Figure 1 The embodiment of the present application relates to a youth depression prediction system, comprising:

[0117] An acoustic characteristic representation module, comprising a ultrasonic transducer array and an electroencephalogram electrode co-integrated, an electrode substrate embedded with an acoustic impedance sensor, and an electrode substrate surface provided with a temperature-sensitive acoustic coupling adhesive layer;

[0118] Specifically, the ultrasonic transducer array adopts PZT-5H piezoelectric material, with a size of 1mm×1mm×0.3mm, an 8×8 array produced by a photolithography process, a center frequency of 1MHz, and a bandwidth of 0.5-1.5MHz;

[0119] Specifically, the acoustic impedance sensor adopts a MEMS surface acoustic wave sensor, with a resonance frequency of 100MHz and a sensitivity of 0.001MRayl, and is connected with a signal processing circuit through gold wire bonding; the preparation of the temperature-sensitive acoustic coupling adhesive layer is that a polyurethane-silicone copolymer (mass ratio 60:40) is dissolved in dimethylbenzene, with a solid content of 25%, a film thickness of 200μm produced by spin coating, and an acoustic impedance of 1.55MRay at 35℃.

[0120] A multi-layer medium acoustic modeling module: used for establishing an accurate head tissue acoustic transmission model according to individual anatomical structure characteristics, for quantifying the frequency dispersion effect generated when the electroencephalogram signal propagates in different tissue layers;

[0121] Specifically, a 256-slice CT scanner is used to obtain head axial images, with scanning parameters of voltage 120kV, current 250mA, layer thickness 0.5mm, and pixel size 0.4mm×0.4mm; an adaptive threshold segmentation algorithm is adopted to extract the inner and outer boundaries of the skull, and the local thickness is calculated:

[0122] -

[0123] Wherein, is the skull thickness at the coordinate ; is the coordinate of the i-th point on the outer surface; is the coordinate of the corresponding point on the inner surface; is the number of sampling points, usually 25 points / mm 2 ;

[0124] A transcranial Doppler ultrasound instrument (probe frequency 2MHz) is used to measure the blood flow velocity of the middle cerebral artery, with a sampling frequency of 100Hz. The cerebrospinal fluid pulsation velocity is estimated by the following relationship:

[0125]

[0126] where, (t) is the cerebrospinal fluid flow rate at time t; is the blood flow velocity in middle cerebral artery; k is the coupling coefficient, with a value range of 0.3-0.5; τ is the phase delay, with a typical value of 0.1 s; is the basal flow rate, with a value of 2-3 cm / s;

[0127] Discretize the signal transmission path into multiple layers, and consider each layer as a homogeneous medium to calculate its acoustic characteristics;

[0128] For the i-th layer medium, its transmission matrix is represented as:

[0129] Ti =

[0130] where, is the complex wave number (rad / m) of the i-th layer medium; ;

[0131] is the frequency;

[0132] is the frequency-dependent phase velocity;

[0133] is the frequency-dependent attenuation coefficient (Np / m);

[0134] is the thickness of the i-th layer; obtained by CT image segmentation, with an accuracy of ±0.1 mm;

[0135] is the acoustic impedance (Rayl) of the i-th layer;

[0136]

[0137] is the medium density (kg / m³);

[0138] is the acoustic wave propagation speed of the i-th layer medium;

[0139] is the imaginary unit;

[0140] The values of the parameters of each tissue layer are as follows:

[0141] Skin layer:

[0142] Thickness: 1.5-2.5 mm (individual measurement);

[0143] Density: 1100 kg / m³;

[0144] Phase velocity: (m / s);

[0145] Attenuation coefficient: (dB / cm);

[0146] Skull layer:

[0147] Thickness: 3-8 mm (individual measurement);

[0148] Density: 1900 kg / m³;

[0149] Phase velocity: (m / s);

[0150] Attenuation coefficient: (dB / cm);

[0151] Cerebrospinal fluid layer:

[0152] Thickness: 3-8 mm (individual measurement);

[0153] Density: 1005 kg / m³;

[0154] Phase velocity: (m / s);

[0155] Attenuation coefficient: (dB / cm);

[0156] Brain tissue layer:

[0157] Thickness: remaining part of the transmission path;

[0158] Density: 1040 kg / m³;

[0159] Phase velocity: (m / s);

[0160] Attenuation coefficient: (dB / cm);

[0161] For a transmission path comprising N layers, the total transmission matrix is:

[0162]

[0163] The acoustic transfer function of the system is derived from the total transmission matrix:

[0164]

[0165] where, is the complex transmission function, containing amplitude and phase information;

[0166] The reference impedance is the water characteristic impedance (1.48 x 10 6 Rayl);

[0167]

[0168] A phase compensation filter module is used to design and implement phase compensation according to the results of acoustic modeling, to correct the phase distortion of the brain electrical signal in the process of tissue transmission;

[0169] It includes:

[0170] The acoustic transfer function obtained based on the acoustic modeling of the multi-layer medium The relative phase difference between channels i and j is calculated:

[0171]

[0172] Channel i and channel j represent different electrode positions on the brain electrical acquisition device. These channels are placed on the scalp surface according to the international 10-20 system or a denser layout, and each channel records the electrical activity of the brain area below it.

[0173] The phase angle of the complex number is extracted for the operator, with the unit of radian (rad), and the output range is [-π, π];

[0174] The phase distortion of channel i relative to channel j varies with frequency.

[0175] Since the phase angle may be wrapped in the interval [-π, π], unwrapping is needed to ensure continuity:

[0176] For the frequency sequence , the phase difference increment is calculated:

[0177]

[0178] Frequency range: 0.5-45 Hz (covering the main frequency band of brain electrical signals), frequency resolution Δf = 0.1 Hz, and frequency point number K = 445;

[0179] If > π, then for plus or minus an integer multiple of 2π, so that ≤ π; repeat until all frequency points are processed to obtain continuous phase difference ;

[0180] The first derivative of the phase distortion with respect to frequency, i.e. the group delay difference, is calculated to quantify the frequency dispersion effect;

[0181]

[0182] where: = 2 (twice the frequency resolution).

[0183] For end points , use forward or backward difference, e.g. ;

[0184] denotes the relative group delay between channels i and j in seconds, reflecting the time delay difference of different frequency components of the signal.

[0185] Construct a frequency-domain compensation filter to correct the phase distortion and dispersion effect:

[0186]

[0187] where, is the imaginary unit;

[0188] is the dispersion compensation coefficient in seconds (s), used to adjust the strength of group delay compensation, determined by maximum likelihood estimation;

[0189] is to correct the static phase offset, ensuring that the phase difference between channels after compensation is zero;

[0190] is to correct the dispersion effect caused by the change of group delay, reducing the frequency-dependent phase distortion;

[0191] In practical applications, to maintain the signal amplitude unchanged, the filter amplitude can be close to 1 through normalization processing, but the adjustment effect of β is retained;

[0192] Calibrate the value of β using healthy adolescent EEG data to ensure optimal compensation effect:

[0193] Collect EEG data of M healthy adolescents (without depressive symptoms), e.g. M = 50; At the same time, collect their acoustic transfer functions , for each subject, calculate the phase difference after compensation for all channels (i, j):

[0194]

[0195] where, and are the frequency representations of the EEG signal, = is the conjugate filter;

[0196] Define the ideal phase difference (Assuming there is no inherent phase difference between channels in a healthy state);

[0197] Calculate the phase error:

[0198] | |

[0199] Assuming the error follows a Gaussian distribution, construct the likelihood function:

[0200]

[0201] in, The standard deviation of the error is initially set to 0.1 rad.

[0202] Solve :

[0203]

[0204] Using grid search in Solve within the specified range, with a step size of 0.001s, the typical optimized value is approximately... ;

[0205] The frequency-domain compensated filter is converted into a time-domain finite impulse response (FIR) filter for real-time processing: The time-domain impulse response is obtained by performing the inverse discrete Fourier transform (IDFT):

[0206]

[0207] Where: N=64: filter length, corresponding to a time length of 0.5 s (sampling rate 128 Hz);

[0208] n=0,1,……,N-1: Time-domain index;

[0209] To ensure For real sequences, Apply Hermitian symmetry constraints:

[0210]

[0211] Where * denotes complex conjugation;

[0212]

[0213] F s The sampling rate of the EEG signal (e.g., 128 Hz);

[0214] Applying Hamming windows to reduce Gibbs phenomenon:

[0215]

[0216] Final filter coefficients:

[0217] is the final FIR filter with real coefficients of length N that can be used for convolution with the EEG signal;

[0218] The compensation filter is applied in real-time computer signal processing, with EEG sampling rate of 128 Hz, frame length of 4 s (512 points), and 50% overlap. Fourier transform is performed on each frame of signal to obtain and ;

[0219] The compensated signal is:

[0220]

[0221]

[0222] where to ensure compensation consistency;

[0223] The theoretical form of is the discrete-time Fourier transform of

[0224] Fourier transform is performed on to obtain the compensated ;

[0225]

[0226]

[0227] This algorithm directly uses the transmission function output by acoustic modeling to calculate the phase distortion, ensuring that the compensation is based on individual tissue characteristics. Group delay calculation relies on the phase dispersion relationship provided by acoustic modeling, while beta estimation is calibrated through healthy data, allowing the compensation to adapt to population variations. The entire process forms a closed loop, real-time correcting signal transmission distortion and improving the reliability of functional connectivity analysis.

[0228] The dispersion-invariant functional connectivity calculation module is used to calculate the functional connectivity index resistant to dispersion interference based on the compensated signal output by the adaptive phase compensation algorithm. By introducing a dispersion stability index and a dynamic weight mechanism, the functional connectivity network analysis is ensured to be immune to residual phase distortion;

[0229] Further processing is performed on the compensated signal output by the adaptive phase compensation algorithm:

[0230] Empirical mode decomposition (EMD) is used to decompose each channel signal into intrinsic mode functions:​​​

[0231]

[0232] wherein:

[0233] is the k-th intrinsic mode function, satisfying the narrowband condition;

[0234] is the residual component;

[0235] is the number;

[0236] for each component, a Hilbert transform is performed to obtain the analytic signal:

[0237]

[0238] wherein:

[0239] is the Hilbert transform operator;

[0240] is the imaginary unit;

[0241] the instantaneous phase is calculated as:

[0242]

[0243] wherein: is the instantaneous phase of the k-th IMF of channel i (rad); a phase unwrapping process is applied to ensure continuity;

[0244] the compensated inter-channel phase difference is calculated as:

[0245]

[0246] wherein: is the weight of the k-th IMF, calculated based on the energy proportion;

[0247] is the number of effective IMFs, excluding high-order components dominated by noise;

[0248]

[0249] wherein: is the standard deviation of the phase difference sequence (rad);

[0250] is the normalization factor, corresponding to the maximum possible phase difference variation range;

[0251] Anti-interference phase synchronization index based on compensated phase difference:

[0252] Traditional PLI calculation:

[0253]

[0254] where: is a sign function returning ±1;

[0255] is the average value within a time window T;

[0256] Time window T = 4 seconds, overlap 50%;

[0257] Compute weighted PLI:

[0258]

[0259] where: is a decay coefficient, determined by cross-validation, typical value λ = 1.5;

[0260] is a frequency dispersion stability index, value range [0, 1];

[0261] Construct a functional connectivity network considering frequency dispersion stability;

[0262] Connection weight:

[0263] where: is the directed connection weight of channel i to j;

[0264] is the total number of channels (32);

[0265] Retain statistically significant connections, threshold based on zero hypothesis test:

[0266]

[0267] where: is the mean of all is the standard deviation of all Extract key graph theory features from the functional connectivity network:

[0268] Calculate global efficiency:

[0269]

[0270]

[0271]

[0272] where:​​ Degree of node i (number of connections);

[0273] Modularity index Q is calculated using Louvain algorithm to partition the network into modules:

[0274]

[0275] Where: Total connection strength of the network; Module to which node i belongs; Kronecker delta function;

[0276] The above module directly uses the compensated signal output by the adaptive phase compensation algorithm , which has been corrected by the filter for the phase distortion caused by the organizational dispersion; the dispersion stability index DSI is calculated based on the variability of the compensated phase difference, reflecting the degree of residual distortion, and is integrated as a weight factor into the PLI calculation, forming a double protection mechanism;

[0277] Bayesian network module, for constructing a dynamic Bayesian network model based on the results of functional connectivity calculation of dispersion invariance, for time series prediction and uncertainty quantification of depression risk; by integrating the dispersion stability index and acoustic features, a more robust prediction framework is established;

[0278] A dynamic system containing functional connection state and observation variables is constructed:

[0279] Define state variables:

[0280]

[0281] Where: Functional connection matrix of dispersion invariance at time t (N x N); Matrix vectorization operation, converting N x N matrix to 1 vector; State vector, dimension (N=32 channels);

[0282] State evolution equation:

[0283] ;

[0284] Where: Dispersion adjustment coefficient, dimension 1, estimated by variational Bayesian estimation;

[0285] Average dispersion stability index, scalar:

[0286]

[0287] is the acoustic feature vector, including acoustic impedance and transmission loss of each channel;

[0288] is the dispersion stability adjustment coefficient;

[0289] is the acoustic feature adjustment matrix;

[0290] Q is the process noise covariance matrix, with dimensions DSxDS, and is assumed to be a diagonal matrix: Q = diag(q1, q2, …, qDS);

[0291]

[0292] is the equivalent acoustic impedance (MRayl) of channel i;

[0293] is the transmission loss (dB) at the reference frequency 10Hz;

[0294] is the acoustic dispersion stability index;

[0295] is the normalization coefficient (typical value 0.01);

[0296] The relationship between the state variable and the multi-modal observation data is established as follows:

[0297] Observation variable definition:

[0298]

[0299] wherein: is the compensated phase lag index matrix;

[0300] is the global efficiency index scalar;

[0301] is the average clustering coefficient scalar;

[0302] is the modularization index scalar;

[0303] The observation equation is:

[0304] wherein: is a nonlinear mapping function, realized by a three-layer feedforward neural network;

[0305]

[0306] b1 is a hidden layer bias vector; b2 is an output layer bias vector; b3 is an output adjustment bias term;

[0307] is an observation noise adjustment vector;

[0308] is an observation noise;

[0309] is a frequency dispersion sensitivity coefficient, is a positive real number without dimension, which controls the sensitivity of the observation noise to the frequency dispersion effect, which determines the rate of the observation noise amplification when deviates from the ideal value 1 ;

[0310] Initialize the model prior based on the healthy population data:

[0311] State prior:

[0312] where:

[0313] is the average functional connectivity vector of 100 healthy adolescents;

[0314] is the functional connectivity covariance matrix of the healthy population, which is regularized as a diagonal matrix;

[0315] Parameter prior:

[0316]

[0317]

[0318]

[0319] where: ; is the dimension of the observation vector;

[0320] Use the mean field variational inference to approximate the posterior distribution:

[0321] Variational distribution setting:

[0322]

[0323] where each variational distribution is a Gaussian distribution:

[0324]

[0325]

[0326] Lower bound of evidence:

[0327]

[0328] Coordinate ascent update principle:

[0329] State update:

[0330]

[0331]

[0332] where, is the neural network Jacobian matrix;

[0333] Parameter update:

[0334]

[0335] Compute depression risk probability: compute real-time depression risk score based on posterior distribution;

[0336] Compute health reference distance:

[0337] where is the Mahalanobis distance;

[0338] Compute risk probability:

[0339] where: , , is the logistic regression coefficient, trained on historical data;

[0340] Typical values: ; ; ;

[0341] is the average clustering coefficient scalar.

[0342] Learn model hyperparameters via EM algorithm:

[0343] E-step: compute posterior expectation using the above variational inference;

[0344] M-step: update parameters by maximizing the complete-data likelihood:

[0345]

[0346]

[0347] Depression intervention decision module: used to construct a multi-level, adaptive decision system based on the output of the Bayesian network modeling, combined with the frequency dispersion stability index and acoustic features, for accurate triggering and adjusting of depression intervention strategies;

[0348] Feature vector composition:

[0349] Where: is the depression risk probability, from Bayesian network modeling (0-1 range);

[0350] is the mean dispersion stability index (0-1 range);

[0351] is the acoustic feature vector;

[0352] : is the rate of change of risk probability;

[0353] is the global efficiency index scalar;

[0354] is the mean clustering coefficient scalar;

[0355] is the modularity index scalar;

[0356] Feature normalization:

[0357]

[0358] Where: is the healthy population feature mean; is the healthy population feature standard deviation;

[0359] A dual-threshold decision system based on signal quality and depression risk is constructed;

[0360] Comprises: First layer: signal quality assessment:

[0361] If ; the signal quality is excellent;

[0362] If < 0.85, ≥ 0.70, the signal quality is good;

[0363] If < 0.70, ≥ 0.60, the signal quality is medium;

[0364] If < 0.60, the signal quality is poor;

[0365] Second layer: depression risk classification:

[0366] If < 0.30, the risk level is low;

[0367] If <0.50, ≥ 0.30, then the risk level is low;

[0368] If < 0.70, ≥ 0.50, then the risk level is medium-high;

[0369] If ≥ 0.70, then the risk level is high;

[0370] Third layer: fusion decision matrix

[0371] Referring to Table 1:

[0372] Table 1

[0373]

[0374] Calculate the comprehensive intervention score to quantify the decision strength:

[0375] Basic score:

[0376] Wherein: ;

[0377] Preferably,

[0378] Network feature correction score:

[0379] Wherein,

[0380] Preferably,

[0381] Comprehensive intervention score:

[0382] Wherein, ;

[0383] Preferably,

[0384] The L2 norm of the acoustic vector;

[0385] Adjust the intervention strength parameter based on the score:

[0386] Adjust the strength of the neural feedback:

[0387]

[0388] Wherein: ; (range strength);

[0389] ; (score range);

[0390] Adjustment task difficulty coefficient:

[0391]

[0392] wherein:

[0393] Adjustment duration of feedback:

[0394]

[0395] wherein: seconds; base duration;

[0396] change sensitivity coefficient;

[0397] When signal quality is poor, prioritize acquisition optimization, including:

[0398] Optimization effect evaluation:

[0399]

[0400] Optimization success criteria:

[0401] Emergency response is initiated for high-risk situations:

[0402] Emergency condition detection:

[0403]

[0404] exponential function;

[0405] Emergency intervention protocol:

[0406] Immediately initiate high-intensity neurofeedback ( ); send an alarm to the monitoring system in parallel; extend the duration of a single intervention to 300 seconds; start continuous monitoring mode (sampling interval shortened to 1 second);

[0407] To avoid frequent fluctuations in decisions and maintain the continuity of interventions, set an exponentially weighted decision:

[0408]

[0409] wherein: current decision weight; smoothed decision;

[0410] Maintain a history of the last N=10 decisions, and only when there are M=5 consecutive decisions in agreement, make significant strategy changes;

[0411] As a case study, the following real-time data was recorded during the brain-computer interface-based depression prediction and intervention process for a 16-year-old female adolescent:

[0412] Monitoring period: 20:00-20:30 on March 15, 2024;

[0413] Features derived from Bayesian network modeling: (normalized)

[0414] The value is 0.73; the signal quality is good.

[0415] (32 channels) Details are as follows:

[0416] Forehead area:

[0417] Channel 1: FP1: 1.5523 (equivalent impedance 1.58 MRayl, transmission loss -11.2dB, DSI=0.88);

[0418] Channel 2: FP2: 1.5481 (equivalent impedance 1.57 MRayl, transmission loss -11.1dB, DSI=0.89);

[0419] Channel 3: Fpz: 1.5498 (equivalent impedance 1.575 MRayl, transmission loss -11.0dB, DSI=0.87);

[0420] Channel 4: AF3: 1.5321 (equivalent impedance 1.56 MRayl, transmission loss -10.9dB, DSI=0.86);

[0421] Channel 5: AF4: 1.5308 (equivalent impedance 1.55 MRayl, transmission loss -10.8dB, DSI=0.87);

[0422] Channel 6: F7: 1.4983 (equivalent impedance 1.52 MRayl, transmission loss -10.5dB, DSI=0.85);

[0423] Channel 7: F8: 1.5012 (equivalent impedance 1.53 MRayl, transmission loss -10.6dB, DSI=0.84);

[0424] Channel 8: Fz: 1.5215 (equivalent impedance 1.545 MRayl, transmission loss -10.7dB, DSI=0.86);

[0425] Forehead area:

[0426] Channel 9: F3: 1.4789 (equivalent impedance 1.50 MRayl, transmission loss -10.4 dB, DSI = 0.83);

[0427] Channel 10: F4: 1.4817 (equivalent impedance 1.51 MRayl, transmission loss -10.3 dB, DSI = 0.84);

[0428] Channel 11: Fl: 1.4923 (equivalent impedance 1.52 MRayl, transmission loss -10.5 dB, DSI = 0.82);

[0429] Channel 12: F2: 1.4956 (equivalent impedance 1.525 MRayl, transmission loss -10.4 dB, DSI = 0.83);

[0430] Channel 13: F5: 1.4682 (equivalent impedance 1.49 MRayl, transmission loss -10.2 dB, DSI = 0.81);

[0431] Channel 14: F6: 1.4719 (equivalent impedance 1.495 MRayl, transmission loss -10.1 dB, DSI = 0.82);

[0432] Channel 15: FC1: 1.4583 (equivalent impedance 1.48 MRayl, transmission loss -9.8 dB, DSI = 0.80);

[0433] Channel 16: FC2: 1.4621 (equivalent impedance 1.485 MRayl, transmission loss -9.7 dB, DSI = 0.81);

[0434] Central zone:

[0435] Channel 17: C3: 1.4238 (equivalent impedance 1.45 MRayl, transmission loss -9.5 dB, DSI = 0.78);

[0436] Channel 18: C4: 1.4285 (equivalent impedance 1.455 MRayl, transmission loss -9.4 dB, DSI = 0.79);

[0437] Channel 19: Cz: 1.4356 (equivalent impedance 1.46 MRayl, transmission loss -9.3 dB, DSI = 0.77);

[0438] Channel 20: Cl: 1.4189 (equivalent impedance 1.44 MRayl, transmission loss -9.6 dB, DSI = 0.76);

[0439] Channel 21 : C2: 1.4217 (equivalent impedance 1.445 MRayl, transmission loss -9.5 dB, DSI = 0.77);

[0440] Channel 22: C5: 1.4083 (equivalent impedance 1.43 MRayl, transmission loss -9.7 dB, DSI = 0.75);

[0441] Parietal region:

[0442] Channel 23: P3: 1.3956 (equivalent impedance 1.42 MRayl, transmission loss -9.8 dB, DSI = 0.74);

[0443] Channel 24: P4: 1.3989 (equivalent impedance 1.425 MRayl, transmission loss -9.7 dB, DSI = 0.75);

[0444] Channel 25: Pz: 1.4023 (equivalent impedance 1.43 MRayl, transmission loss -9.6 dB, DSI = 0.73);

[0445] Channel 26: P1 : 1.3887 (equivalent impedance 1.41 MRayl, transmission loss -9.9 dB, DSI = 0.72);

[0446] Channel 27: P2: 1.3921 (equivalent impedance 1.415 MRayl, transmission loss -9.8 dB, DSI = 0.73);

[0447] Channel 28: P7: 1.3785 (equivalent impedance 1.40 MRayl, transmission loss -10.1 dB, DSI = 0.71);

[0448] Occipital region:

[0449] Channel 29: O1 : 1.3658 (equivalent impedance 1.39 MRayl, transmission loss -10.2 dB, DSI = 0.70);

[0450] Channel 30: O2: 1.3689 (equivalent impedance 1.395 MRayl, transmission loss -10.1 dB, DSI = 0.71);

[0451] Channel 31 : z: 1.3723 (equivalent impedance 1.40 MRayl, transmission loss -10.0 dB, DSI = 0.69);

[0452] Channel 32: POz: 1.3856 (equivalent impedance 1.41 MRayl, transmission loss -9.9 dB, DSI=0.72);

[0453] ;

[0454] 0.68; risk level is medium-high;

[0455] 0.45;

[0456] 0.38;

[0457] 0.29;

[0458] (signal quality = "good", risk level = "medium-high") -> decision = "standard neurofeedback", priority = "high";

[0459] = 0.408 + 0.0675 + 0.018 = 0.4935;

[0460] = 0.22 + 0.114 + 0.087 = 0.421;

[0461] = 0.2468 + 0.1263 + 0.126 = 0.4991;

[0462] = 0.3 + 0.7x0.499 = 0.649;

[0463] = 1 / (1 + exp(-(-1.5 + 3.0x0.68 + 2.0x0.73))) = 0.881;

[0464] = 180 x (1 + 2.0x0.12) = 180x1.24 = 223;

[0465] Perform intervention plan:

[0466] Neurofeedback task: Prefrontal alpha wave enhancement training; Goal: Increase prefrontal alpha wave power to 130% of baseline value; Difficulty level: 0.881 (high difficulty); Feedback intensity: 0.649 (moderate to strong); Duration: 223 seconds;

[0467] Intervention effect evaluation:

[0468] 60 seconds after the start of the intervention: Prefrontal alpha wave power: increased from baseline to 125%; Engagement score: 0.72 (good); Signal quality maintenance: DSI = 0.71;

[0469] 5 minutes after the end of the intervention: ; showed a downward trend.

[0470] The above is only an embodiment of the present application, and does not limit the patent scope of the present application, any equivalent transformation or direct or indirect application in related technical fields using the content of the present application specification and drawings are also included in the patent protection scope of the present application.

Claims

1. Adolescent Depression Prediction System, characterized in that, include: The phase compensation filtering module is used to design and implement phase compensation based on acoustic modeling results to correct phase distortion of EEG signals during tissue transmission. The dispersion-invariant functional connection calculation module is used to calculate the functional connection index against dispersion interference based on the compensated signal output by the adaptive phase compensation algorithm. By introducing a dispersion stability index and a dynamic weighting mechanism, it ensures that the functional connection network analysis is not affected by residual phase distortion. The Bayesian network module is used to construct a dynamic Bayesian network model based on the results of dispersion-invariant functional connection calculations, for time-series prediction and uncertainty quantification of depression risk; a prediction box is established by integrating dispersion stability indices and acoustic features. Depression Intervention Decision Module: This module is used to construct a decision system based on the output of Bayesian network modeling, combined with dispersion stability indicators and acoustic features, to trigger and adjust depression intervention strategies. The acoustic characterization module includes an integrated ultrasonic transducer array and EEG electrodes, with an acoustic impedance sensor embedded in the electrode substrate and a temperature-sensitive acoustic coupling adhesive layer on the surface of the electrode substrate. Multilayer media acoustic modeling module: used to establish an accurate acoustic transmission model of cranial tissue based on individual anatomical structural characteristics, and to quantify the dispersion effect generated when EEG signals propagate in different tissue layers; The uses of the depression intervention decision-making module specifically include: Constructing decision feature vectors: Feature vector composition: in: The probability of depression is derived from Bayesian network modeling; The average dispersion stability index; For acoustic feature vectors; : represents the rate of change of risk probability; A scalar for global efficiency metrics; The average clustering coefficient is a scalar. Modular scalar indicators; Calculate the dynamic weighted comprehensive intervention score: in, ; in: ; in, ; Let L2 be the norm of the acoustic vector; Intervention parameters are adaptively adjusted based on comprehensive intervention scores; The uses of the depression intervention decision-making module, specifically the adaptive adjustment of intervention parameters based on comprehensive intervention scores, include: Adjusting the intensity of neurofeedback: in: ; ； Adjust the task difficulty level: in: Adjust the continuous feedback time: in: Seconds; based on duration; This is the sensitivity coefficient to change.

2. The adolescent depression prediction system according to claim 1, characterized in that, The applications of the multilayer media acoustic modeling module include: obtaining individual skull thickness distribution based on CT image data; measuring cerebrospinal fluid pulsation velocity via Doppler ultrasound; and calculating the acoustic transmission matrix for each tissue layer. If = in, denoted as the complex wave number (rad / m) of the i-th layer of medium. ; For frequency; The phase velocity is frequency-dependent; The attenuation coefficient is frequency-dependent. The thickness of the i-th layer is obtained through CT image segmentation. Let be the acoustic impedance of the i-th layer; The density of the medium; Let be the sound wave propagation speed of the i-th layer of medium; The imaginary unit; Derive the acoustic transfer function of the system from the total transfer matrix: in, It is a complex transfer function that includes amplitude and phase information; Reference impedance; 3. The adolescent depression prediction system according to claim 1, characterized in that, The phase compensation filtering module has the following specific applications: Acoustic transfer function obtained from multi-layer media acoustic modeling Calculate the relative phase difference between channels i and j: in, Let be the acoustic transfer function of channel i, derived from the acoustic modeling output, and be a complex function containing amplitude and phase information; The acoustic transfer function for channel j is also derived from the acoustic modeling output; Extract the phase angle of the complex number for the operator, in radians, with an output range of [-π, π]. This represents the phase distortion of channel i relative to channel j, which varies with frequency. Construct a frequency domain compensation filter to correct phase distortion and dispersion effects: in, The imaginary unit; , is the dispersion compensation coefficient, in seconds, used to adjust the strength of group delay compensation, and is determined by maximum likelihood estimation; To correct static phase shift and ensure that the phase difference between channels is zero after compensation; To correct the dispersion effect caused by group delay variation and reduce frequency-dependent phase distortion.

4. The adolescent depression prediction system according to claim 1, characterized in that, The uses of the dispersion-invariant function connection calculation module specifically include: Empirical mode decomposition is performed on the compensated signal; Extract the instantaneous phase of each IMF component; Calculate the dispersion stability index: in: Phase difference sequence Standard deviation; This is the normalization factor, corresponding to the maximum possible range of phase difference variation; Calculate the dispersion-weighted phase lag exponent: in: The attenuation coefficient was determined through cross-validation. This is a dispersion stability index, with a value range of [0,1]. Construct a functional connectivity network that considers dispersion stability; Connection weights: in: Let be the directed connection weight from channel i to j; This represents the total number of channels.

5. The adolescent depression prediction system according to claim 1, characterized in that, The uses of Bayesian network modules include: State-space model definition: ； in: For dispersion adjustment coefficient, dimension ×1, estimated via variational Bayesian method; The average dispersion stability index is a scalar: This is the acoustic characteristic vector, which includes the acoustic impedance and transmission loss of each channel; This is the acoustic feature adjustment matrix; Q is the process noise covariance matrix; Observation model: in: It is a nonlinear mapping function, implemented through a three-layer feedforward neural network; b1 is the hidden layer bias vector; b2 is the output layer bias vector; b3 is the output adjustment bias term; The observation noise adjustment vector; To observe noise; Let be the dispersion sensitivity coefficient, a dimensionless positive real number, which controls the sensitivity of observation noise to dispersion effects. It determines the sensitivity of observation noise to dispersion effects when... When the deviation from the ideal value of 1 is observed, the rate of noise amplification is measured. Initialize the model prior based on data from healthy populations: State Priors: in: The average functional connectivity vector of 100 healthy adolescents; For the functional connectivity covariance matrix of the healthy population; posterior distribution is calculated based on variational Bayesian inference; probability of depression risk is then calculated. in: , , These are the logistic regression coefficients, obtained through training with historical data; This is a scalar value representing the average clustering coefficient.

Images