A method for assessing the risk of depression based on functional near-infrared spectroscopy

Abstract

The application discloses a depression risk assessment method based on functional near-infrared spectroscopy, comprising the following steps: S1, collecting functional near-infrared spectroscopy of depression patients and healthy personnel, and preprocessing the functional near-infrared spectroscopy to serve as a data set; S2, training a near-infrared spectroscopy extraction model through the data set; S3, extracting the functional near-infrared spectroscopy of a person to be risk assessed, preprocessing the functional near-infrared spectroscopy, inputting the near-infrared spectroscopy extraction model, extracting the blood oxygen dynamics characteristics of each region of the brain of the person to be risk assessed, and performing depression risk assessment.

Description

Technical Field

[0001] This application relates to the field of information analysis and processing, and in particular to a method for assessing the risk of depression based on functional near-infrared spectroscopy. Background Technology

[0002] In recent years, functional near-infrared spectroscopy (fNIRS) has gained widespread attention in neuroscience and biomedical engineering due to its non-invasiveness, portability, and relatively low cost. fNIRS assesses neural activity by measuring changes in blood oxygen concentration in the cerebral cortex, demonstrating unique advantages. Despite its enormous potential in assessing neural activity, existing methods still face numerous challenges. Traditional fNIRS analysis methods primarily rely on machine learning models, with typical examples including Support Vector Machines (SVM), Random Forests (RF), and k-Nearest Neighbors (KNN). These methods typically require manual feature extraction from fNIRS signals to generate low-dimensional data suitable for classification. Common feature extraction methods include time-domain and frequency-domain analysis, such as extracting average power, peak amplitude, and time-related changes in blood oxygen concentration. However, the feature extraction process heavily depends on expert experience and domain knowledge; different experimental designs and data preprocessing methods can significantly affect the extraction results. This not only increases the complexity of model construction but also raises the barrier to data analysis, limiting its widespread adoption in non-specialized application scenarios.

[0003] In existing research on fNIRS data processing, a significant trade-off exists between the number of channels and model performance. While a high number of channels can improve spatial resolution and thus enhance the model's classification ability, it also significantly increases the cost of hardware and computing resources, making it difficult to widely implement, especially in resource-constrained real-world applications. In contrast, low-channel-count fNIRS data often performs poorly in classification accuracy due to insufficient signal feature extraction, failing to meet the needs of in-depth analysis of complex brain functional activities. Therefore, how to achieve efficient and accurate brain function assessment under low-channel-count conditions has become a critical challenge that urgently needs to be addressed. The core issue is whether it is possible to bypass the complex feature extraction process and directly use the raw data for classification and diagnosis, while ensuring excellent system performance in terms of efficiency and robustness. Summary of the Invention

[0004] To address the aforementioned shortcomings in existing technologies, this invention provides a method for assessing the risk of depression based on functional near-infrared spectroscopy. This method designs an end-to-end architecture and spatiotemporal module statistical features, enabling direct processing of raw data and solving the problem of brain function assessment relying on a high number of channels.

[0005] To achieve the aforementioned objectives, the technical solution adopted by this invention is: a method for assessing the risk of depression based on functional near-infrared spectroscopy, comprising:

[0006] S1. Collect functional near-infrared spectra of patients with depression and healthy individuals, and preprocess them to obtain the dataset;

[0007] S2. Train the near-infrared spectroscopy extraction model using the dataset;

[0008] S3. Extract functional near-infrared spectra of individuals to be assessed for risk, preprocess them, and input them into a trained near-infrared spectroscopy extraction model to extract blood oxygen dynamics characteristics of different brain regions of the individuals to be assessed for depression risk.

[0009] Furthermore: S1 includes:

[0010] S11. Collect functional near-infrared spectra of patients with depression and healthy individuals during vocabulary fluency tests;

[0011] S12. Remove motion artifacts from the functional near-infrared spectrum and filter it to obtain the processed functional near-infrared spectrum.

[0012] S13. The processed functional near-infrared spectrum is converted into hemoglobin concentration to obtain low-frequency blood oxygenation change information, which is then used as a dataset.

[0013] Furthermore, in S2, the near-infrared spectroscopy extraction model processes the input data using the following methods:

[0014] S21. Input the input data into 16 parallel time feature modules to capture key time series features at different frequencies;

[0015] S22. Input key temporal features of different frequencies into the spatial feature module to identify the blood oxygen dynamics features of different brain regions;

[0016] S24. The blood oxygen dynamics characteristics of each brain region are averaged and then input into the statistical feature module to obtain the original depression risk assessment information.

[0017] S25. Expand the original depression risk assessment information into a one-dimensional feature vector, and process it through a fully connected layer and the Softmax activation function to obtain the depression risk assessment result.

[0018] Furthermore, the temporal feature module and the spatial feature module together constitute the spatiotemporal feature module, which includes a 2D convolutional layer, a first batch of normalized layers, a deep 2D convolutional layer, a second batch of normalized layers, and an ELU activation function layer connected in sequence.

[0019] Furthermore, in S24, the statistical feature module's methods for processing the blood oxygen dynamics characteristics of different brain regions include:

[0020] S241. Calculate the minimum, maximum, and weighted average values ​​of the blood oxygen dynamics characteristics of each brain region over time t.

[0021] S242. The minimum, maximum and weighted average values ​​of the blood oxygen dynamics characteristics of each brain region at time t are summed element by element to obtain statistical characteristics.

[0022] S243. The blood oxygen dynamics characteristics of each brain region are sequentially batch standardized, processed by activation function, and 2D average pooling to obtain the processed blood oxygen dynamics characteristics.

[0023] S244. The statistical characteristics and the processed blood oxygen kinetic characteristics are added element by element to obtain the original depression risk assessment information.

[0024] Furthermore: In S241, the expressions for calculating the minimum, maximum, and weighted average values ​​of the blood oxygenation dynamics characteristics of various brain regions over time t are as follows:

[0025]

[0026] Among them, M min (x i M represents the minimum value of the blood oxygen dynamic characteristics of different brain regions over time t. max (x i () represents the maximum value of the blood oxygen dynamic characteristics of each brain region over time t. denoted as the weighted average of the blood oxygen dynamics characteristics of different brain regions over time t, where T represents the time window, sup represents the upper bound of the time window T, inf represents the lower bound of the time window T, and w(t) represents the weighting function.

[0027] The beneficial effects of this invention are as follows:

[0028] An end-to-end architecture and spatiotemporal module statistical features were designed to enable direct processing of raw data and improve the model's ability to capture the spatiotemporal features of functional near-infrared spectral signals. This allows for excellent performance even with low-channel input, thereby improving the accuracy of depression risk assessment. Attached Figure Description

[0029] Figure 1 This is a flowchart of a method for assessing the risk of depression based on functional near-infrared spectroscopy.

[0030] Figure 2 This is a schematic diagram of a functional near-infrared spectral collection method.

[0031] Figure 3 This is a schematic diagram of the near-infrared spectroscopy extraction model.

[0032] Figure 4 This is a schematic diagram of the statistical feature model structure.

[0033] Figure 5 This is a comparison of performance metrics and the magnitude of metric changes for the Depression Dataset under full-channel and low-channel conditions.

[0034] Figure 6 This chart compares the performance metrics of the Stroke Dataset under full-channel and low-channel conditions, and shows the magnitude of metric changes.

[0035] Figure 7 The graph shows the performance of different models on the Depression Dataset dataset.

[0036] Figure 8 This chart compares the performance metrics of traditional machine learning models with data that has undergone feature extraction as input, and deep learning models with raw data that has not undergone feature extraction as input. Detailed Implementation

[0037] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0038] like Figure 1 As shown, in one embodiment of the present invention, a method for assessing the risk of depression based on functional near-infrared spectroscopy is provided, comprising:

[0039] S1. Collect functional near-infrared spectra of patients with depression and healthy individuals, and preprocess them to obtain the dataset.

[0040] S2. Train the near-infrared spectroscopy extraction model using the dataset;

[0041] S3. Extract functional near-infrared spectra of individuals to be assessed for risk, preprocess them, and input them into a trained near-infrared spectroscopy extraction model to extract blood oxygen dynamics characteristics of different brain regions of the individuals to be assessed for depression risk.

[0042] In classification tasks, when the sample size is relatively small and the number of features is large, the phenomenon of the Curse of Dimensionality is extremely likely to occur, directly affecting the generalization ability of the model and resulting in a significant decrease in accuracy. To effectively alleviate this problem, we adopted a feature selection strategy of an interpretive model and combined it with a validation set evaluation mechanism to screen out the optimal channels from the input channels. This method not only improves the generalization performance of the model but also reduces the interference of redundant features on the classification results, ensuring that the model can still maintain a high discriminative ability in small-sample scenarios.

[0043] Specifically, S1 includes:

[0044] S11. Collect functional near-infrared spectroscopy (FNIRS) of depression patients and healthy individuals during a word fluency test.

[0045] As Figure 2 shown, during the word fluency test, the subjects were required to form words using Chinese radicals "家 (jiā)", "北 (běi)", and "花 (huā)" to stimulate brain activities related to speech generation.

[0046] In this application, a total of 48 probes were used in the FNIRS device. These probes were precisely arranged in the frontal, parietal, and temporal regions of the subjects, covering brain areas related to functions such as emotional regulation, language processing, and cognitive control. Each probe generated 3-channel data (i.e., HbO, HbR, and HbT), so a total of 144-channel time-series features were extracted, and the sampling frequency was 11 Hz. This frequency is relatively common in FNIRS research, can effectively balance data precision and volume, and is suitable for long-term experimental recordings.

[0047] S12. Remove motion artifacts from the functional near-infrared spectroscopy and perform filtering to obtain the processed functional near-infrared spectroscopy.

[0048] S13. Convert the processed functional near-infrared spectroscopy into hemoglobin concentration to obtain low-frequency blood oxygen change information and use it as a dataset.

[0049] Specifically, in S2, the processing method of the near-infrared spectroscopy extraction model for the input data includes:

[0050] S21. Input the input data into 16 parallel time feature modules respectively to capture key timing features of different frequencies.

[0051] S22. Input the key timing features of different frequencies into the spatial feature module to identify the blood oxygen dynamics features of each brain region.

[0052] S24. Input the blood oxygen dynamics features of each brain region into the statistical feature module after average pooling to obtain the original depression risk assessment information.

[0053] S25. Expand the original depression risk assessment information into a one-dimensional feature vector, and process it through a fully connected layer and the Softmax activation function to obtain the depression risk assessment result.

[0054] The near-infrared spectroscopy extraction model is a ResNet-like CNN architecture based on a brain-computer interface (BCI) using fNIRS. This model possesses the following key characteristics: (1) it can be flexibly applied to various BCI paradigms, demonstrating strong generalization ability; (2) it requires only a very small number of channels to complete efficient training, significantly reducing its dependence on high-dimensional data; (3) even with extremely limited fNIRS data, the near-infrared spectroscopy extraction model can still maintain high training performance and classification accuracy. A schematic diagram of the near-infrared spectroscopy extraction model is shown below. Figure 3 As shown, this architecture not only solves the modeling challenges under low-channel conditions but also optimizes for the sparsity of fNIRS signals, enabling the model to achieve robust performance even on small datasets. The near-infrared spectroscopy extraction model, while ensuring model compactness and computational efficiency, can efficiently capture key features in fNIRS data, making it particularly suitable for applications with a limited number of channels but requiring accurate brain activity identification. The model parameters of the near-infrared spectroscopy extraction model are shown in Table 1.

[0055] Table 1. Near-infrared spectroscopy extraction model parameters (when the training rounds are 400)

[0056] Training time 34.73 seconds Number of trainable parameters 5762 Number of untrainable parameters 80 Total number of parameters 5842

[0057] Specifically, the temporal feature module and the spatial feature module together form the spatiotemporal feature module, which includes a 2D convolutional layer, a first batch of normalized layers, a deep 2D convolutional layer, a second batch of normalized layers, and an ELU activation function layer connected in sequence.

[0058] The time feature module is used to extract information features of fNIRS data in the time dimension. Considering the lag and slowness of fNIRS data, we set the length of the time feature module to be as short as possible in order to effectively capture the low-frequency blood oxygenation changes of subjects when performing the VFT task.

[0059] In addition, given the importance of low-frequency information in fNIRS data, we set up 16 "time feature modules" to more accurately capture key time-series features at different frequencies.

[0060] The spatial feature module is used to identify the blood oxygen dynamics characteristics of different brain regions, enhance the model's sensitivity to neural activity in specific brain regions, and thus extract changes in blood flow and blood oxygen levels in different brain regions more precisely, enabling the capture and analysis of regional neural activity.

[0061] like Figure 4 As shown in S24, the statistical feature module's methods for processing the blood oxygen dynamics characteristics of different brain regions include:

[0062] S241. Calculate the minimum, maximum, and weighted average values ​​of the blood oxygen dynamics characteristics of each brain region over time t.

[0063] The expressions for calculating the minimum, maximum, and weighted average of the blood oxygen dynamics characteristics of different brain regions over time t are as follows:

[0064]

[0065] Among them, M min (x i M represents the minimum value of the blood oxygen dynamic characteristics of different brain regions over time t. max (x i () represents the maximum value of the blood oxygen dynamic characteristics of each brain region over time t. denoted as the weighted average of the blood oxygen dynamics characteristics of different brain regions over time t, where T represents the time window, sup represents the upper bound of the time window T, inf represents the lower bound of the time window T, and w(t) represents the weighting function.

[0066] S242. The minimum, maximum and weighted average values ​​of the blood oxygen dynamics characteristics of each brain region at time t are summed element by element to obtain statistical characteristics.

[0067] S243. The blood oxygen dynamics characteristics of each brain region are sequentially batch standardized, processed by activation function, and 2D average pooling to obtain the processed blood oxygen dynamics characteristics.

[0068] S244. The statistical characteristics and the processed blood oxygen kinetic characteristics are added element by element to obtain the original depression risk assessment information.

[0069] The statistical feature module not only relies on simple statistics, but also incorporates functional analysis, integral theory and optimization theory to ensure that the feature extraction of each channel can accurately describe the extreme behavior of the signal and its global trend.

[0070] In one embodiment of the present invention, the DepressionDataset dataset collected using a lexical fluency test is shown in Table 2:

[0071] Table 2 DepressionDataset dataset

[0072] male female total Depression patients 2 20 22 healthy people 8 9 17 total 10 29 39

[0073] The DepressionDataset dataset contains records from 39 participants, 22 of whom were clinically diagnosed with depression based on standardized psychiatric scales and physician assessments, as shown in the table. The remaining 17 participants were from a healthy control group. All participants were matched as closely as possible in terms of age, gender, and education level to minimize the influence of confounding variables on the results.

[0074] To verify the robustness of the proposed model, we used a publicly available fNIRS dataset from Qingdao University of Science and Technology as a comparative dataset. This dataset contains fNIRS signals from healthy individuals and stroke patients. Data acquisition was performed using the NIRSIT portable device, covering 204 probes, each recording data from the HbO and HbR hemoglobin channels. Compared to our main dataset, this comparative dataset not only uses different equipment but also has a significantly increased number of probes, yet it only contains two types of blood oxygenation signals. This diversity makes it an ideal candidate for testing the adaptability and generalization ability of near-infrared spectroscopy extraction models. In this application, this dataset is named the StrokeDataset dataset.

[0075] In this embodiment, we used the NirSpark software package to perform systematic preprocessing on the fNIRS signal. During fNIRS scanning, individuals often exhibit involuntary head movements, and these motion artifacts have a significant impact on functional connectivity analysis.

[0076] Following artifact correction, bandpass filtering was further performed, specifically within a frequency range of 0.01 to 0.1 Hz, to eliminate low-frequency noise caused by physiological activities such as pulse and respiration. This filtering step is crucial for improving the signal-to-noise ratio, further ensuring the accuracy and consistency of the data, thus laying a solid foundation for subsequent analysis.

[0077] After data preprocessing, each experimental sample is converted to a dimension of X∈R. T The matrix form is ×C, where T = 330 represents the number of sampling time points and C = 144 represents the number of channels. Specifically, each data point x i,j This indicates at time point t i Next, the Cth j The fNIRS signal intensity of each channel. Currently, the main methods for using fNIRS for judgment rely on manual judgment by medical experts, or feature extraction from existing data followed by classification using Support Vector Machines (SVM). This method depends to some extent on the experience and judgment of experts and may be affected by subjective factors. The raw data retains the complete information of the original signal, without further feature analysis, and directly reflects the spatiotemporal characteristics of the fNIRS signal.

[0078] Based on this, we designed the following experiment, and the experimental inputs are shown in Tables 3 and 4:

[0079] Table 3 Experimental settings for DepressionDataset

[0080] full channel low channel Feature extraction Near-infrared spectroscopy extraction model √ √ × EEGNet √ √ × KNN √ √ √ MLP √ √ √ RF √ √ √ SVM √ √ √

[0081] Table 4 StrokeDataset Experiment Settings

[0082] full channel low channel Feature extraction Near-infrared spectroscopy extraction model √ √ × EEGNet √ √ × KNN √ √ × MLP √ √ × RF √ √ × SVM √ √ ×

[0083] We chose the StrokeDataset for comparison not only to evaluate the model's performance on different data structures, but also because this dataset provides only two types of features (HbO and HbR), which differs from the feature composition of our main dataset. Furthermore, its numerous channels better demonstrate the adaptability of our near-infrared spectroscopy extraction model under low-channel conditions. Comparative experiments show that the near-infrared spectroscopy extraction model can effectively identify brain functional activity under different numbers of probes and feature types, indicating its robustness to noise, probe configuration, and signal channels. This further supports the broad applicability and reliability of the near-infrared spectroscopy extraction model in fNIRS brain function research.

[0084] Experimental Comparison Method: For traditional medical signal processing algorithms, such as Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Random Forest (RF), and Multilayer Perceptron (MLP), we first extract features from the original data. Each sample matrix X is transformed using time-frequency domain transformation and statistical feature analysis to generate a new feature matrix X′, which has the following form:

[0085] X′=f(X)

[0086] Here, f() represents a specific feature extraction method. These features will be used as input data to evaluate the classification performance of the above classic algorithms after feature extraction.

[0087] Evaluation experiments of various models using raw data from all channels as input: In order to evaluate the classification ability of various models for fNIRS signals without relying on feature engineering, the raw data without any feature processing was directly input into the near-infrared spectroscopy extraction model, EEGNet, SVM, KNN, RF and MLP models for training and evaluation. The results are shown in Table 5.

[0088] Table 5 shows the experimental results of each model evaluation using the raw data from all channels as input.

[0089]

[0090] Where Accuracy_D represents the accuracy of each model on the DepressionDataset, and Accuracy_S represents the accuracy of each model on the StrokeDataset.

[0091] Evaluation experiments of various models under low-channel raw data input: In practical applications, the number of channels in fNIRS devices is usually limited by hardware resources and experimental environment. Therefore, evaluating the algorithm performance under low-channel data is of great significance. In this experiment, we start from the original matrix X∈R T In ×C, select CL=8 channels to form a new low-channel dataset X. L ∈R T ×8. In the experimental setup, we directly input the low-channel data into the near-infrared spectroscopy extraction model, EEGNet, and SVM, KNN, RF, and MLP models for training and testing, maintaining the same hyperparameter settings and data preprocessing procedures as the full-channel experiment. The results are shown in Table 6:

[0092] Table 6 shows the experimental results of each model evaluation using the raw data from the low channel as input.

[0093]

[0094] When both machine learning and deep learning models are fed with full-channel raw data: In our full-channel raw data evaluation experiment, we explored the performance of machine learning and deep learning models under ideal conditions with the same input. This experiment assumed ample equipment, computing resources, and time, aiming to compare the performance of the two types of models with high-dimensional raw data and a sufficient number of probes. Under these conditions, the performance of machine learning and deep learning models tended to be consistent, and even in some tasks, the machine learning model had a slight advantage. Experimental results show that although machine learning models have a slight advantage in some classification tasks when computation is sufficient, deep learning models exhibit broader adaptability without relying on complex preprocessing, providing important reference for practical applications.

[0095] When the input is raw data with 8 channels: The low-channel raw data evaluation experiment focuses on real-world application scenarios, simulating detection and evaluation under conditions of limited equipment, time, and computing resources. This experiment aims to simulate the actual conditions in which the general population cannot rely on high-end equipment and long-term testing during routine health checkups. Through low-channel fNIRS data input, this experiment explores how to still achieve effective detection and diagnosis while reducing equipment and time costs, thus providing a cost-effective solution for public health. In this scenario, traditional machine learning models show a significant performance decline due to limitations in data dimensionality and the number of channels (e.g., ...). Figure 5 (a) and Figure 6(a) shows that key metrics such as accuracy and recall have declined significantly (e.g., as shown in (a)). Figure 5 (b) and Figure 6 (b) is shown.

[0096] in, Figure 5 (a) A comparison of the performance metrics of different models under full-channel and low-channel conditions on the Depression Dataset dataset; Figure 5 (b) shows the variation of performance indicators for each model under full-channel and low-channel input; Figure 6 (a) A comparison of the performance metrics of different models under full-channel and low-channel conditions on the Stroke Dataset dataset; Figure 6 (b) shows the variation of performance indicators for each model under full-channel and low-channel input.

[0097] However, near-infrared spectroscopy extraction models can still maintain excellent classification performance by optimizing the input data dimensionality and channel selection, especially with low-channel input, such as... Figure 7 As shown, Figure 7 (a) is a visualization of the t-SNE results of the EEGNet model. Figure 7 (b) t-SNE visualization results of the near-infrared spectral extraction model; This shows that the near-infrared spectral extraction model exhibits strong adaptability and robustness when processing low-channel data, providing an economical and efficient solution for the promotion of fNIRS technology in practical applications.

[0098] We used t-SNE to visualize the features extracted by the near-infrared spectroscopy extraction model network, mapping the high-dimensional features to a two-dimensional space, and classifying and labeling the data points according to the labels. A clear boundary exists between the red and blue data points, indicating that different categories are effectively distinguished in the high-dimensional feature space, demonstrating the network's strong classification ability.

[0099] When the machine learning model is input with full-channel feature extraction data, and the deep learning model is input with 8-channel raw data, the experimental results are as follows: Figure 8 As shown, this experiment compared a traditional machine learning model (carefully extracted feature data from all channels) with a deep learning model (unprocessed raw data from low channels). Through this experiment, we clarified the optimal performance that traditional machine learning methods can achieve after meticulous preprocessing. Experimental results show that the near-infrared spectroscopy extraction model achieves better performance in core metrics such as accuracy and recall (e.g., Figure 8 As shown, all of them outperform traditional machine learning models, proving that deep learning models have significant advantages in handling real-world tasks and can reduce the complexity of manual feature selection.

[0100] This demonstrates that our proposed near-infrared spectroscopy extraction model can still achieve efficient detection at a low cost even with low-channel raw data input, greatly improving its feasibility and universality in practical applications. Through this experiment, we propose a low-cost and widely applicable detection solution, promoting the adoption and application of portable fNIRS detection technology in the public health field, making the cost of brain dysfunction detection affordable for the general public, especially in the screening of mental illnesses such as depression. The low cost and high efficiency of the near-infrared spectroscopy extraction model meet the needs of large-scale applications, significantly lowering the technical threshold and enabling groups with limited economic means to enjoy such detection services, thereby expanding the coverage of mental health screening. With the increasing severity of global mental health issues, especially among adolescents and high-pressure working groups, early detection of depression has become particularly important. The near-infrared spectroscopy extraction model provides a practical solution for large-scale health screening, especially suitable for non-specialized medical environments such as schools and community health centers. By simplifying the detection process, the near-infrared spectroscopy extraction model is expected to become a convenient and efficient mental health screening tool, helping to detect potential problems early and intervene in a timely manner, thereby reducing the burden on the healthcare system and improving the mental health level of the population.

[0101] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for depression risk assessment based on functional near-infrared spectroscopy, characterized by, The method comprises the following steps: S1, collecting functional near-infrared spectroscopy of patients with depression and healthy personnel, and preprocessing to obtain a data set; S2, training a near-infrared spectroscopy extraction model based on the data set; S3, extracting the functional near-infrared spectroscopy of personnel to be risk evaluated, preprocessing, and inputting the trained near-infrared spectroscopy extraction model to extract the blood oxygen dynamics characteristics of each region of the brain of the personnel to be risk evaluated, and performing depression risk evaluation; In S2, the processing method of the near-infrared spectroscopy extraction model for input data comprises: S21, inputting the input data into 16 parallel time feature modules to capture key time sequence characteristics of different frequencies; S22, inputting the key time sequence characteristics of different frequencies into a spatial feature module to identify the blood oxygen dynamics characteristics of each region of the brain; S24, inputting the blood oxygen dynamics characteristics of each region of the brain into a statistical feature module after average pooling to obtain original depression risk evaluation information; S25, expanding the original depression risk evaluation information into a one-dimensional feature vector, and processing through a full connection layer and a Softmax activation function to obtain a depression risk evaluation result; In S24, the processing method of the statistical feature module for the blood oxygen dynamics characteristics of each region of the brain comprises: S241, calculate the minimum, maximum and weighted average of the blood oxygen dynamic characteristics of each region of the brain over time, respectively t ; S242. Transforming the blood oxygen dynamics characteristics of different brain regions over time. t The minimum, maximum, and weighted average values ​​of the channel signals are summed element by element to obtain statistical characteristics; S243, sequentially performing batch standardization, activation function processing and 2D average pooling on the blood oxygen dynamics characteristics of each region of the brain to obtain processed blood oxygen dynamics characteristics; S244, element-wise adding the statistical characteristics and the processed blood oxygen dynamics characteristics to obtain the original depression risk evaluation information. 2.The functional near-infrared spectroscopy-based depression risk assessment method according to claim 1, wherein, S1 comprises: S11, collecting functional near-infrared spectroscopy of patients with depression and healthy personnel during a word fluency test; S12, removing motion artifacts from the functional near-infrared spectroscopy and filtering to obtain processed functional near-infrared spectroscopy; S13, converting the processed functional near-infrared spectroscopy into hemoglobin concentration to obtain low-frequency blood oxygen change information, and taking the low-frequency blood oxygen change information as a data set. 3.The functional near-infrared spectroscopy-based depression risk assessment method according to claim 1, wherein, The time feature module and the spatial feature module jointly serve as a space-time feature module, and the space-time feature module comprises a 2D convolution layer, a first batch normalization layer, a deep 2D convolution layer, a second batch normalization layer and an ELU activation function layer connected in sequence. 4.The functional near-infrared spectroscopy-based depression risk assessment method according to claim 1, wherein, In S241, the minimum, maximum, and weighted average of the blood oxygenation dynamics of each region of the brain over time are calculated using the following expressions: t S241 min = min (S241) wherein represents the minimum value of the blood oxygenation dynamics of the respective region of the brain over time t , represents the maximum value of the blood oxygenation dynamics of the respective region of the brain over time t , represents the weighted average of the blood oxygenation dynamics of the respective region of the brain over time t , T represents the time window, sup represents the supremum of the time window T , T inf represents the infimum of the time window w , t represents the weight function.

Images