A method for predicting depression based on gender and season-specific neural activity

Abstract

The present application belongs to the technical field of biomedical signal processing, and specifically relates to a depression prediction method based on gender and season-specific neural activity, aiming to improve the accuracy of the prediction model. The method includes obtaining electroencephalogram signals of multiple depression patients, and recording the gender attribute, data collection season, and depression severity scale score of each depression patient. The electroencephalogram signals are preprocessed, and based on the preprocessed electroencephalogram signals, a multi-dimensional feature set of multiple brain regions in multiple sub-frequency bands is extracted; the multi-dimensional feature set includes frequency domain features, nonlinear dynamic features, and inter-brain region information flow features. With gender attribute and data collection season as influencing factors, statistical analysis is performed on the multi-dimensional feature set, and features with statistical significance are selected to form a target feature subset. The target feature subset is used to train a machine learning model, and the trained prediction model is used to output the depression severity quantification index prediction result of the depression patient.

Description

Technical Field

[0001] This invention relates to the field of biomedical signal processing technology, and in particular to a method for predicting depression based on gender- and season-specific neural activity. Background Technology

[0002] Depression is a complex and prevalent mental disorder that has become a major global public health problem. Currently, the clinical assessment of the severity of depression mainly relies on physician interviews and patient self-rating scales, such as the Beck Depression Self-Rating Scale. Although this method is widely used, it is highly dependent on the physician's subjective experience and the patient's immediate state and cooperation. It is difficult to avoid assessment distortions caused by the patient's avoidance personality or social expectation bias, and it suffers from problems such as strong subjectivity and poor timeliness.

[0003] In the search for objective biological markers, electroencephalograms (EEGs) have attracted attention because they can directly reflect the brain's neurophysiological activity. Existing technologies utilize EEG signal characteristics, such as power spectral density and nonlinear dynamics, to predict the severity of depression using machine learning models. However, most of these methods treat patients with depression as a homogeneous group, constructing general predictive models. In reality, numerous studies have shown that gender factors (such as differences in hormone levels and psychosocial stressors) and seasonal factors (such as daylight hours and climate changes) significantly modulate the expression of depressive symptoms and brain neural activity patterns. Existing general models ignore these key regulatory variables, resulting in features that lack group specificity and limiting predictive accuracy and robustness. Summary of the Invention

[0004] The purpose of this invention is to provide a method for predicting depression based on gender and season-specific neural activity, aiming to improve the accuracy of the prediction model.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] This invention provides a method for predicting depression based on gender and season-specific neural activity, comprising: S1: acquiring EEG signals from multiple patients with depression and recording the gender attribute, data collection season, and depression severity scale score for each patient; wherein the EEG signals are multi-channel EEG data acquired through a standard electrode system, and the depression severity scale score is the Beck Depression Scale score; S2: preprocessing the EEG signals and extracting multidimensional feature sets of multiple brain regions in multiple sub-frequency bands based on the preprocessed EEG signals; the multidimensional feature sets include frequency domain features, nonlinear dynamic features, and inter-brain information flow features; S3: statistically analyzing the multidimensional feature sets using gender attribute and data collection season as influencing factors, and selecting statistically significant features to form a target feature subset; S4: training a machine learning model using the target feature subset, and using the trained prediction model to output the prediction results of the quantitative index of depression severity for patients.

[0007] The preprocessing of the EEG signal includes the following steps: channel relocation and redundant electrode removal; filtering and notch filtering of the EEG signal; segmenting the filtered EEG signal and removing abnormal amplitude signal segments; rereference processing of the abnormal-removed EEG signal; dividing the EEG signal into sub-frequency bands and delineating the target brain region according to the preset brain region range.

[0008] The frequency domain features are power spectral density features, the nonlinear dynamic features are fuzzy entropy features, and the brain inter-brain information flow features are transfer entropy features.

[0009] The power spectral density features were extracted using the power spectral analysis method based on fast Fourier transform to calculate the average power spectral density of each channel's EEG signal in each sub-frequency band. The fuzzy entropy features were extracted using the fuzzy entropy algorithm based on phase space reconstruction to calculate the average fuzzy entropy of each channel's EEG signal in each sub-frequency band. The transfer entropy features were extracted using a pre-set transfer entropy analysis toolbox to calculate the average transfer entropy of each channel's EEG signal in each sub-frequency band.

[0010] Based on the preprocessed EEG signals, multidimensional feature sets of multiple brain regions in multiple sub-frequency bands are extracted, including: dividing the EEG signals into signals belonging to different brain regions, and extracting multidimensional features of each brain region signal in each sub-frequency band.

[0011] The brain regions include at least the frontal lobe, central lobe, and parieto-occipital lobe; the sub-bands include Delta, Theta, Alpha, Beta, and Gamma bands.

[0012] In step S3, the statistical analysis is a two-way ANOVA, and statistical significance means that the main effect of gender or the interaction effect between gender and season meets the preset significance level.

[0013] The machine learning model is a gradient boosting decision tree model.

[0014] The training process of the gradient boosting tree model includes: dividing the dataset corresponding to the target feature into a training set and a test set according to a preset ratio; optimizing the parameters of the gradient boosting tree model using a hyperparameter optimization tool; and training the model using the optimized parameters and the training set to obtain a depression severity prediction model.

[0015] Step S5 follows step S4: using the SHapley additive interpretation tool to perform interpretability analysis on the output of the prediction model and identify the key features that contribute the most to the prediction results.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0017] 1. The depression prediction method based on gender- and season-specific neural activity provided in this application incorporates consideration of the impact of gender and season on the severity of depression throughout the entire process. This allows the target feature subset to accurately capture the unique neural activity patterns of patients with depression in different genders and seasons, overcoming the limitations of the one-size-fits-all adaptability of general models and significantly improving the prediction accuracy for specific groups. By constructing a multi-dimensional feature system including frequency domain features (power spectral density), nonlinear dynamic features (fuzzy entropy), and brain inter-regional information flow features (transmission entropy), and simultaneously covering the frontal lobe, central, parietal-occipital brain regions, and Delta, Theta, Alpha, Beta, and Gamma sub-bands, a full-dimensional characterization of EEG signals is achieved.

[0018] 2. The method provided in this application introduces the SHapley additive interpretation tool to decompose the output of the prediction model and identify the key features that contribute the most to the prediction (such as the transfer entropy of a specific brain region and the fuzzy entropy of a specific frequency band). This breaks the black box barrier of traditional machine learning models, enabling clinicians to clearly grasp the core basis of the prediction and providing concrete characteristic evidence for the study of the neural mechanisms of depression. Attached Figure Description

[0019] Figure 1 This is a flowchart of a method for predicting depression based on gender and season-specific neural activity provided in an embodiment of this application;

[0020] Figure 2 This is a schematic diagram illustrating the performance verification of a depression severity prediction model provided in an embodiment of this application;

[0021] Figure 3 This is a predictive model interpretability analysis diagram provided in an embodiment of this application. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0023] For example, such as Figure 1 As shown in the embodiments of this application, a method for predicting depression based on gender- and season-specific neural activity is provided, including:

[0024] S1: Acquire EEG signals from multiple patients with depression and record the gender attribute, data collection season, and depression severity scale score for each patient; the EEG signals are multi-channel EEG data acquired through a standard electrode system, and the depression severity scale score is the Beck Depression Scale (BDI) score.

[0025] Exemplary experimental data is from the publicly accessible TDBRAIN EEG dataset. Data recording was performed according to standardized methods developed by BrainResource Ltd. 26-channel signals were recorded using the Compumedics Quickcap or ANT-Neuro Waveguard Cap with sintered Ag / AgCl electrodes based on the 10-10 electrode international system, acquired at a sampling rate of 500 Hz. Reference to the raw EEG data was converted to the mean values ​​of bilateral mastoid processes (A1 and A2 electrodes) during offline analysis, with the AF2 electrode as the ground point. Longitudinal and transverse eye movements were also recorded. Two minutes of data were collected from each subject during both open and resting states. The dataset includes 132 patients with major depressive disorder (aged 18.66–78.00 years), including 65 males (15 in spring, 15 in summer, 16 in fall, and 19 in winter) and 67 females (21 in spring, 12 in summer, 20 in fall, and 14 in winter). All participants completed the Beck Depression Rating Scale, which was used as a quantitative indicator of the severity of depression in this application.

[0026] S2: Preprocess the EEG signal and extract multidimensional feature sets of multiple brain regions in multiple sub-frequency bands based on the preprocessed EEG signal. The multidimensional feature sets include frequency domain features, nonlinear dynamic features, and inter-brain information flow features. Among them, the frequency domain feature is the power spectral density (PSD) feature, the nonlinear dynamic feature is the fuzzy entropy (FuzzyEn) feature, and the inter-brain information flow feature is the transfer entropy (TransEn) feature.

[0027] For example, the preprocessing of EEG signals using MATLAB version R2020a and the EEGLAB toolbox version 2024.2 includes the following steps: channel relocation and redundant electrode removal of the EEG signals; filtering and notch filtering of the EEG signals; segmenting the filtered EEG signals and removing abnormal amplitude signal segments; rereference processing of the abnormal-removed EEG signals; dividing the EEG signals into sub-frequency bands and delineating target brain regions according to preset brain region ranges.

[0028] During actual data acquisition, electrodes may deviate from their preset positions due to the subject's scalp shape and wearing errors, such as tightness or slight offset. This can lead to misalignment, where an electrode that should be acquiring signals from the frontal lobe actually acquires signals from the central brain region. Channel repositioning (e.g., based on scalp anatomical landmarks or digital positioning tools) can correct this deviation, ensuring that the signal from each channel truly originates from the target brain region and preventing subsequent mis-extraction of brain region features. EEG signals are weak bioelectrical signals, typically in the microvolt range. During acquisition, they are easily contaminated by low-frequency drift, high-frequency noise, and other broadband interferences. Furthermore, the neural activity characteristics related to depression exist only within a specific frequency range. High-pass filtering to truncate low-frequency signals below 1Hz can eliminate the influence of baseline drift on EEG signal amplitude. Low-pass filtering to truncate high-frequency signals above 40Hz can eliminate interference from noise such as electromyography (EMG) in the effective EEG frequency band, thus ensuring data accuracy.

[0029] For example, extracting multidimensional feature sets of multiple brain regions across multiple sub-frequency bands based on preprocessed EEG signals includes: dividing the EEG signals into signals belonging to different brain regions, and extracting multidimensional features of each brain region's signals across each sub-frequency band. The brain regions include at least the frontal lobe, central lobe, and parieto-occipital lobe. The sub-frequency bands include the Delta band, Theta band, Alpha band, Beta band, and Gamma band.

[0030] High-pass and low-pass digital filters were set to 1Hz and 40Hz respectively, with notch frequencies between 48Hz and 52Hz. The signal was divided into 2-second segments, and signal segments with amplitudes below -100μV and above 100μV were removed. Then, the signal was rereferenced using a whole-brain average reference. The following sub-bands were extracted: Delta (1-4Hz), Theta (4-8Hz), Beta (8-13Hz), Alpha (13-30Hz), and Gamma (30-45Hz). Finally, electrodes Fp1, Fp2, F3, F4, F7, F8, Fz, FC3, FC4, and FCz were designated as frontal lobe regions; electrodes C3, C4, Cz, CP3, CP4, and CPz were designated as central brain regions; and electrodes P3, P4, P7, P8, Pz, O1, O2, and Oz were designated as parietal-occipital brain regions for further feature extraction and analysis.

[0031] For example, the power spectral density feature is extracted by using the power spectral analysis method of fast Fourier transform. The average value of the PSD of the 26 channels of the EEG signal at NFFT=1024 is calculated in a segmented period of 2 seconds. The values ​​of the PSD calculated in different segments are then averaged.

[0032] For example, the extraction of fuzzy entropy features employs a fuzzy entropy algorithm based on phase space reconstruction to calculate the average fuzzy entropy of each channel's EEG signal in each sub-frequency band. The specific algorithm is as follows:

[0033] Set an N-point sampling sequence: {u(i): 1≤i≤N}. Reconstruct the phase space of u(i) in sequence order to obtain a set of m-dimensional vectors (m≤N-2). The reconstructed vectors can be written as:

[0034] (1)

[0035] Where i = 1,2,…,N−m+1, (i) represents the average value.

[0036] It is the distance between two vectors. and Defined as the maximum difference between corresponding elements of two vectors, i.e. There are two vectors and The similarity between them is defined as:

[0037] (2)

[0038] Where n and r are the gradient and width of the exponential function, respectively.

[0039] Define function :

[0040] (3)

[0041] By repeating steps (2) and (3) in the same way, a sequence function of a set of (m + 1) dimensional vectors can be reconstructed based on the order of the sequence. .

[0042] Define the fuzzy entropy of a given time series:

[0043] (4)

[0044] Where m and r represent the phase space dimension and similarity tolerance, respectively, m = 2, r = 0.5 (approximately 0.25 × SD, where SD is the standard deviation of the time series).

[0045] For example, the extraction of transfer entropy features involves calculating the average transfer entropy of each channel's EEG signal across each sub-frequency band using a pre-defined transfer entropy analysis toolbox. More specifically, the HERMES toolbox (version 2020-04-26) in MATLAB is used to calculate the average transfer entropy for each subject across all segments. For two time series, the TransEn (representation) from X to Y is defined by adjusting for past observations:

[0046] (5)

[0047] in, and These represent the historical states of Y and X, respectively.

[0048] S3: Using gender attribute and data collection season as influencing factors, perform statistical analysis on the multidimensional feature set, and select statistically significant features to form the target feature subset.

[0049] For example, in step S3, the statistical analysis is a two-way ANOVA, and statistical significance means that the main effect of gender or the interaction effect between gender and season meets a preset significance level. More specifically, in this embodiment of the application, for the two-way ANOVA, features with significant main effects of gender and gender-seasonal interaction effects, i.e., p < 0.05, are selected as the input features required to predict BDI scores. The above statistical analysis is performed using RStudio version 2024.09.0+375. The statistical results show that for the Delta band, all three EEG indicators have significant gender and seasonal interaction effects in almost the entire brain; for the Beta band, the PSD and FuzzyEn features have significant main effects of gender in the prefrontal cortex and significant main effects and interaction effects of gender in the parieto-occipital cortex; for the Gamma band, the PSD and FuzzyEn features have significant main effects of gender in the prefrontal cortex and on the Pz electrode, and additionally significant interaction effects on the Pz electrode; for the TransEn feature, there are significant main effects and interaction effects of gender within the prefrontal cortex and between the prefrontal cortex and the Pz electrode. Other sub-bands do not have significant main effect or interaction effect distributions.

[0050] S4: Train a machine learning model using a subset of target features, and use the trained prediction model to output the quantitative indicators of the severity of depression in patients.

[0051] For example, the machine learning model is a gradient boosting decision tree model. More specifically, this application embodiment uses the machine learning gradient boosting tree algorithm, namely the CatBoost algorithm, to train the model, and uses feature channels with significant gender main effects and interaction effects obtained through statistical analysis and screening in the feature statistical analysis module for model training and BDI regression prediction.

[0052] As one possible implementation, the training process of the gradient boosting tree model includes: dividing the dataset corresponding to the target feature into a training set and a test set according to a preset ratio; optimizing the parameters of the gradient boosting tree model using a hyperparameter optimization tool; and training the model using the optimized parameters and the training set to obtain a depression severity prediction model.

[0053] The dataset used in this embodiment is divided into training and testing sets at 80% and 20% respectively. This embodiment uses the catboost and hyperopt packages of Python 3.9.1 for hyperparameter optimization, including parameters such as bagging temperature, border count, depth, grow policy, iterations, L2 leaf regularization, learning rate, and random strength.

[0054] The method provided in this application embodiment further includes step S5 after step S4: using the SHApley Additive Interpretation Tool (SHAP) to perform interpretability analysis on the output of the prediction model, identifying the key features that contribute the most to the prediction results. To evaluate the prediction performance of the prediction model, this application embodiment calculates the Pearson correlation coefficient between the predicted BDI score and the actual BDI score; the higher the correlation coefficient, the better the prediction effect.

[0055] The SHAP tool quantifies the contribution of each feature to the prediction result of a single case (positive contributions increase the prediction score, while negative contributions decrease it), clearly presenting the correlation logic between key features and prediction results. This traceable and verifiable interpretation allows clinicians to correlate the prediction basis with known neural mechanisms of depression, significantly increasing their confidence in the model's prediction results and clearing a key obstacle for the technology to move from the laboratory to clinical application.

[0056] For example, refer to Figure 2 , Figure 2 The horizontal axis represents the true BDI score, which indicates the actual severity of depression as determined by clinical standard assessment (the higher the BDI score, the more severe the depression). Figure 2 The vertical axis represents the predicted BDI score, which is the quantitative result of the depression severity output by the gradient boosting decision tree model in this embodiment. Each scatter point in the figure represents the pairing of the actual value and the predicted value for one patient with depression. The closer the scatter point is to the "y=x diagonal", the smaller the deviation between the predicted value and the actual value.

[0057] Figure 2 The left and right plots represent males and females, respectively, and are divided into subplots according to season (spring, summer, autumn, and winter). Each seasonal subplot contains two sets of key statistical indicators: the Pearson correlation coefficient (r) and the significance level (p). The Pearson correlation coefficient (r) measures the strength of the linear correlation between the predicted and actual values; the closer r is to 1, the higher the prediction accuracy. The significance level (p) is used to verify the statistical reliability of the correlation; p < 0.05 indicates a significant correlation, meaning the prediction results are not random and the model performance is stable. The data in the figure shows that the r-values ​​are 0.60 (p=0.017) and 0.80 (p<0.001) in spring; 0.77 (p<0.001) and 0.86 (p<0.001) in summer; 0.72 (p=0.002) and 0.82 (p<0.001) in autumn; and 0.98 (p<0.001) and 0.96 (p<0.001) in winter. Furthermore, the p-values ​​for all groups are <0.05, indicating that the prediction model results are accurate and reliable.

[0058] For example, refer to Figure 3 , Figure 3 It integrates four dimensions: feature type, frequency band, gender, and season. Feature type includes the three types of EEG data extracted in this application embodiment: PSD, FuzzyEn, and TransEn. Frequency bands include Delta band (1-4Hz), Beta band (13-30Hz), and Gamma band (30-45Hz). Figure 3 China uses highly visual methods to represent contributions. Figure 3 This visually demonstrates the correlation between a patient's BDI predicted score and characteristic type, frequency band, gender, and season. For example, a woman with a higher PSD in the Delta band during winter is strongly correlated with a higher BDI predicted score. This traceable contribution logic allows clinicians to link the predictive basis to the neural mechanisms of depression, such as the correlation between abnormal Beta band frequencies and mood disorders, addressing a pain point that traditional machine learning models cannot explain, and significantly improving clinical acceptance.

[0059] In the description of this specification, specific features, structures, materials, or characteristics may be combined in any suitable manner in one or more embodiments or examples.

[0060] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predicting depression based on gender- and season-specific neural activity, characterized in that, include: S1: Acquire EEG signals from multiple patients with depression and record the gender attribute, data collection season, and depression severity scale score for each patient; wherein, the EEG signals are multi-channel EEG data acquired through a standard electrode system, and the depression severity scale score is the Beck Depression Scale score; S2: Preprocess the EEG signals and extract multidimensional feature sets of multiple brain regions in multiple sub-frequency bands based on the preprocessed EEG signals; the multidimensional feature sets include frequency domain features, nonlinear dynamic features, and inter-brain information flow features; the frequency domain features are power spectral density features, the nonlinear dynamic features are fuzzy entropy features, and the inter-brain information flow features are transfer entropy features; the extraction of multidimensional feature sets of multiple brain regions in multiple sub-frequency bands based on the preprocessed EEG signals includes: classifying the EEG signals into categories belonging to different frequency bands. Signals from the same brain region are collected, and the multidimensional features of each brain region's signals in each sub-frequency band are extracted; the brain regions include at least the frontal lobe, central lobe, and parieto-occipital lobe; the sub-frequency bands include Delta, Theta, Alpha, Beta, and Gamma bands; S3: Using the gender attribute and the data collection season as influencing factors, statistical analysis is performed on the multidimensional feature set to select statistically significant features to form a target feature subset; the statistical analysis is a two-way ANOVA, and the statistical significance refers to the main effect of gender or the interaction effect of gender and season meeting a preset significance level; S4: A machine learning model is trained using the target feature subset, and the trained prediction model is used to output the prediction results of the quantitative index of the severity of depression patients; the machine learning model is a gradient boosting decision tree model.

2. The method for predicting depression based on gender- and season-specific neural activity according to claim 1, characterized in that, The preprocessing of the EEG signal includes the following steps: channel repositioning and redundant electrode removal of the EEG signal; filtering and notch filtering of the EEG signal; segmenting the filtered EEG signal and removing abnormal amplitude signal segments; rereference processing of the abnormal-free EEG signal; dividing the EEG signal into sub-frequency bands and delineating the target brain region according to a preset brain region range.

3. The method for predicting depression based on gender- and season-specific neural activity according to claim 1, characterized in that, The extraction of the power spectral density features employs a power spectral analysis method based on fast Fourier transform to calculate the average power spectral density of each channel's EEG signal in each sub-frequency band. The extraction of the fuzzy entropy features employs a fuzzy entropy algorithm based on phase space reconstruction to calculate the average fuzzy entropy of each channel's EEG signal in each sub-frequency band. The extraction of the transfer entropy features is achieved by using a pre-set transfer entropy analysis toolbox to calculate the average transfer entropy of each channel's EEG signal in each sub-frequency band.

4. The method for predicting depression based on gender- and season-specific neural activity according to claim 1, characterized in that, The training process of the gradient boosting tree model includes: dividing the dataset corresponding to the target feature into a training set and a test set according to a preset ratio; optimizing the parameters of the gradient boosting tree model using a hyperparameter optimization tool; and training the model using the optimized parameters and the training set to obtain a depression severity prediction model.

5. A method for predicting depression based on gender- and season-specific neural activity according to any one of claims 1-4, characterized in that, Step S5 follows step S4: using the SHapley additive interpretation tool to perform interpretability analysis on the output of the prediction model and identify the key features that contribute the most to the prediction results.

Images