A depression detection system

Abstract

The present invention discloses a depression detection system, including: an exhaled gas collection module, a waveform signal acquisition module and a detection module; the gas channel of the exhaled gas collection module gradually narrows along the flow direction of the exhaled gas, and is used to discharge the original gas in the gas channel and collect the exhaled gas by creating a flow rate difference at the air inlet and the air outlet of the gas channel; the waveform signal acquisition module is connected to the exhaled gas collection module, and is used to collect the waveform signal of the exhaled gas; the detection module is connected to the waveform signal acquisition module, and is used to detect depression according to the waveform signal. The present invention sets the gas channel of the exhaled gas collection module to a shape that gradually narrows along the gas flow direction, the gas channel has a large volume and a small flow rate at the air inlet; the gas channel has a small volume and a large flow rate at the air outlet, so that the original gas in the gas channel is discharged more thoroughly, reducing the interference of the original gas, thereby making the depression detection result more accurate.

Description

Technical Field

[0001] The present invention relates to the technical field of medical data processing, and in particular to a depression detection system. Background Art

[0002] In the field of depression detection, with the deepening of biomarker research, breath analysis technology based on the concept of electronic nose is gradually showing its potential. This technology detects volatile organic compounds (VOCs) in the patient's exhaled breath through a high-precision sensor array. The content of these compounds is closely related to the physiological and psychological state of the human body, providing a new perspective for non-invasive detection of mental illnesses such as depression.

[0003] Current respiratory analysis equipment is divided into air bag respiratory analysis equipment and air bagless respiratory analysis equipment. Among them, the air bagless respiratory analysis equipment has an intermediate cabin for storing gas inside, and depression is detected by blowing air directly into the intermediate cabin. The air bagless respiratory analysis equipment needs to discharge the interfering gas in the intermediate cabin during use. The traditional solution is to let the test subject blow as much air as possible to fill the intermediate cabin with the gas to be tested and discharge the interfering gas out of the intermediate cabin.

[0004] However, the above scheme places high demands on the test subjects. Long-term and large-volume blowing increases the burden on patients and the threshold for using the equipment. It is difficult to ensure that the interfering gas is discharged from the intermediate cabin instead of mixing with the gas to be tested, which in turn affects the accuracy of depression detection. Summary of the invention

[0005] In order to solve the above problems, the present invention discloses a depression detection system.

[0006] The present invention discloses a depression detection system, comprising: an exhaled gas collection module, a waveform signal acquisition module and a detection module;

[0007] The gas channel of the exhaled gas collection module gradually narrows along the flow direction of the exhaled gas, and is used to discharge the original gas in the gas channel and collect the exhaled gas by creating a flow rate difference at the gas inlet and the gas outlet of the gas channel;

[0008] The waveform signal acquisition module is connected to the exhaled gas collection module and is used to collect the waveform signal of the exhaled gas;

[0009] The detection module is connected to the waveform signal acquisition module and is used to perform depression detection according to the waveform signal.

[0010] Preferably, the central axis of the air inlet and the central axis of the air outlet of the gas channel intersect at a preset point.

[0011] Preferably, the waveform signal acquisition module includes a solid electrolyte carbon dioxide sensor, a sensor sensitive to combustible gases, and a sensor sensitive to malodorous gases.

[0012] Preferably, the detection module includes a feature extraction unit, a feature selection unit and a depression detection unit;

[0013] The feature extraction unit is connected to the waveform signal acquisition module and is used to extract features from the waveform signal to obtain a feature set;

[0014] The feature selection unit is connected to the feature extraction unit and is used to select the feature set to obtain an optimal feature subset;

[0015] The depression detection unit is connected to the feature selection unit and is used to perform depression detection according to the optimal feature subset.

[0016] Preferably, the feature set is selected to obtain an optimal feature subset, specifically:

[0017] Based on the feature selection algorithm integrating Inf-FSs and SVM-RFE, the feature set is selected to obtain the optimal feature subset.

[0018] Preferably, based on the feature selection algorithm combining Inf-FSs and SVM-RFE, the feature set is selected to obtain the optimal feature subset, specifically:

[0019] According to the Inf-FSs algorithm and the feature set, a neighboring order matrix is ​​obtained;

[0020] According to the SVM-RFE algorithm and the feature set, a weight matrix is ​​obtained;

[0021] The feature set is selected according to the neighboring order matrix and the weight matrix to obtain an optimal feature subset.

[0022] Preferably, the feature set is selected according to the neighboring order matrix and the weight matrix to obtain an optimal feature subset, specifically:

[0023] Obtaining a first feature importance score and a second feature importance score according to the neighboring order matrix and the weight matrix;

[0024] Normalizing and summing the first feature importance score and the second feature importance score to obtain a final feature importance score;

[0025] The feature set is selected according to the final feature importance score to obtain an optimal feature subset.

[0026] Preferably, feature extraction is performed on the waveform signal, specifically:

[0027] At least one of transient feature extraction, frequency domain feature extraction and decomposition signal feature extraction is performed on the waveform signal.

[0028] Preferably, depression detection is performed according to the optimal feature subset, specifically:

[0029] Inputting the optimal feature subset into a classification model to obtain a classification accuracy rate corresponding to the optimal feature subset;

[0030] Depression detection is performed according to the classification accuracy.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] (1) The present invention sets the gas channel of the exhaled gas collection module to a shape that gradually narrows along the flow direction of the exhaled gas. The part of the exhaled gas entering the gas channel has a large volume and a low flow rate; the part leaving the gas channel has a small volume and a high flow rate, so that the original gas in the gas channel is discharged more thoroughly; further, the flow rate of the air inlet part is relatively slow, which, on the one hand, allows the tester to complete the blowing in a more relaxed state without excessive force or long-term continuous blowing; on the other hand, it helps to distribute the exhaled gas more evenly in the gas channel, reduces the turbulence and eddy current phenomenon in the gas flow, reduces the disturbance and damage to the gas sample, and is conducive to maintaining the original characteristics and composition of the gas sample, facilitating subsequent accurate analysis;

[0033] (2) The gas outlet of the gas channel of the present invention is narrow. Due to the limitation of space and the increase of flow rate, the collision and diffusion between gas molecules are strengthened, which helps to push the interfering gas remaining in the gas channel further to the gas outlet and discharge it, thereby improving the accuracy of depression detection;

[0034] (3) The waveform signal acquisition module of the present invention is equipped with multiple sensors, which can collect waveform signals of different exhaled gases and provide multi-dimensional reference data for the detection of depression;

[0035] (4) The system of the present invention analyzes waveform signals by using transient features, frequency domain features, and signal decomposition features simultaneously, which helps to detect depression more comprehensively and accurately;

[0036] (5) The present invention combines the use of neighboring matrix and weight matrix to help discover the potential relationships and interactions between features and select a more refined and efficient optimal feature subset, thereby improving the accuracy of depression detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1is a schematic diagram of the structure of a depression detection system of the present invention;

[0038] Figure 2 is a schematic diagram of the structure of a gas channel according to an embodiment of the present invention;

[0039] Figure 3 is the average response curve of sensor TGS2610-D00 for healthy controls and patients with depression;

[0040] Figure 4 is the average response curve of sensor TGS2603 for healthy controls and patients with depression;

[0041] Figure 5 is the average response curve of sensor TGS826 for healthy controls and patients with depression;

[0042] Figure 6 is the average response curve of sensor SP3S-AQ2 for healthy controls and patients with depression;

[0043] Figure 7 is the average response curve of sensor TGS822 for healthy controls and patients with depression;

[0044] Figure 8 is the average response curve of sensor TGS2600-temperature modulation for healthy controls and patients with depression;

[0045] Fig. 9 is the average response curve of sensor TGS2620 for healthy controls and patients with depression;

[0046] Fig.10 is the relationship between the peak response of each sensor when the multiple sensors are used in the depression detection system of the present invention and the breath analysis device of the prior art for different blowing times (hydrogen, 50.6 ppm);

[0047] Fig.11 is the relationship between the peak response of each sensor when the multiple sensors are used in the depression detection system of the present invention and the breath analysis device of the prior art for different blowing times (ammonia, 10 ppm);

[0048] Fig.12 is a comparison diagram of the responses of the depression detection system proposed by the present invention (left) and the air bag device of the prior art (right);

[0049] Fig.13 It is a typical sample display of waveform signal;

[0050] In the figure, 1. air inlet; 2. air outlet; 3. exhaled gas; 4. outside; 5. first two-way valve; 6. gas channel; 7. first micro vacuum pump; 8. second two-way valve; 9. sensor array; 10. second micro vacuum pump. DETAILED DESCRIPTION

[0051] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present invention. However, it should be clear to those skilled in the art that the present invention may be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present invention.

[0052] like Figure 1 As shown, the present invention discloses a depression detection system, including: an exhaled gas collection module, a waveform signal acquisition module and a detection module;

[0053] The exhaled gas collection module comprises a first two-way valve 5, a gas channel 6, a first micro vacuum pump 7 and a second two-way valve 8;

[0054] The gas channel 6 of the exhaled gas collection module gradually narrows along the flow direction of the exhaled gas 3, and is used to discharge the original gas in the gas channel 6 and collect the exhaled gas 3 by creating a flow rate difference at the gas inlet 1 and the gas outlet 2 of the gas channel 6;

[0055] Preferably, the central axis of the air inlet and the central axis of the air outlet of the gas channel 6 intersect at a preset point.

[0056] like Figure 2 As shown, in one embodiment, the gas channel 6 adopts a structure in which the volume of the side where the air inlet 1 is located is large and the volume of the side where the air outlet 2 is located is small. According to the theory of fluid dynamics, the volume of the side where the air inlet 1 is located is large, and the flow rate of the inhaled gas is small; the volume of the side where the air outlet 2 is located is small, and the flow rate is large. Due to the limitation of the space at the air outlet 2 and the increase in the flow rate, the collision and diffusion effect between gas molecules are strengthened, which helps to further discharge the interfering gas remaining in the gas channel 6; further, the flow rate on the side where the air inlet 1 is located is slower, on the one hand, it allows the tester to complete the blowing in a more relaxed state without excessive force or long-term continuous blowing; on the other hand, it helps to distribute the exhaled gas 3 more evenly in the gas channel 6, reduces the turbulence and eddy current phenomenon in the gas flow, reduces the disturbance and damage to the gas sample, and is conducive to maintaining the original characteristics and composition of the exhaled gas 3, facilitating subsequent accurate analysis;

[0057] In this embodiment, the air inlet 1 is arranged above the gas channel 6, and the air outlet 2 is arranged on the side wall of the gas channel 6, so that the exhaled air flow direction is not parallel to the outlet, reducing convection and preventing the exhaled air components from being directly lost from the air outlet 2.

[0058] The waveform signal acquisition module is connected to the exhaled gas collection module and is used to collect the waveform signal of the exhaled gas 3;

[0059] like Figure 1 As shown, the waveform signal acquisition module includes a sensor array 9 (arranged inside the air chamber), a signal processing circuit and a data acquisition card connected in sequence, and further includes a second micro vacuum pump 10 connected to the air chamber, and the second micro vacuum pump 10 is used to discharge the exhaled gas 3 after the reaction to the outside 4;

[0060] The signal processing circuit is used to send a temperature modulation signal to the sensor array 9; and to pre-process the sensor signal (voltage signal) acquired by the sensor array 9 to obtain a pre-processed sensor signal;

[0061] The data acquisition card is used to receive the pre-processed sensor signal (waveform signal) and send the temperature step signal to the signal processing circuit.

[0062] Preferably, the sensor array 9 includes a solid electrolyte carbon dioxide sensor, a sensor sensitive to combustible gases, and a sensor sensitive to malodorous gases.

[0063] Specifically, in Example 1, the sensor array 9 used is a solid electrolyte carbon dioxide sensor TGS4161, a sensor TGS2610-D00 sensitive to butane and carbon monoxide, and a sensor TGS2603 sensitive to ammonia and hydrogen sulfide. Figure 3 , Figure 4 Shown are the average response curves of TGS2610-D00 and TGS2603 for healthy controls and patients with depression during the test in Example 1.

[0064] The sensor array 9 is one of the core components of the depression detection system of the present invention. The design of the sensor array 9 directly determines the sensing range and sensing accuracy of the device.

[0065] In Example 2, in order to improve the performance of the depression detection system and reduce the interference of irrelevant information caused by the use of redundant sensors, the products of multiple sensor manufacturers were tested and compared. Based on the existing pathological manifestations of depression, 110 healthy volunteers and 160 volunteers with depression were additionally recruited for comparative testing experiments. After comprehensively considering factors such as sensor performance, optimal operating temperature, sensor stability, cross-response problems, and the range of perceptible target gases, the sensor array 9 of the present invention consisting of 8 gas sensors and 1 humidity and temperature sensor was finally selected. The specific configuration is shown in Table 1.

[0066] Table 1 Sensor array settings and brief parameters

[0067]

[0068] In Table 1, volatile organic compounds are organic volatiles including acetone, etc., that is, a general term for volatile organic compounds. For sensors, sensor No. 1 is a fixed electrolyte carbon dioxide sensor, which is used to indicate the actual amount of exhaled air received, sensors No. 2 to No. 8 are metal oxide semiconductor sensors, which are used to detect the concentration of different target gases, and sensor No. 9 is a wet temperature sensor, which is used to compensate the data for humidity in the preprocessing step. According to the content introduced above, the components such as nitric oxide, hydrogen sulfide and carbon monoxide in the exhaled gas 3 of patients with depression will change, among which the concentrations of nitric oxide and hydrogen sulfide will decrease relative to the healthy control group, and the concentration of carbon monoxide will increase relative to the healthy control group. Corresponding to the design of the present invention, sensor No. 4, TGS2610-D00 is sensitive to combustible gases, including carbon monoxide; sensor No. 7, TGS2603 is sensitive to sulfides, including hydrogen sulfide, and objects with foul odors.

[0069] Although the sensor array 9 designed in Example 1 has a significant difference in response between healthy controls and patients with depression, it is still necessary to add other sensors to enhance the performance of the sensor array 9. Because metal oxide semiconductor sensors cannot sense only one type of gas, they are often more sensitive to a certain type of gas and less sensitive to other gases. At the same time, the correspondence between the sensor and the target gas is not one-to-one, and there are multiple sensors that can respond to a certain gas at the same time. Therefore, in Example 2, a small number of sensor groups that have high specificity for a certain task are combined with some sensor groups that have weaker specificity for this task, which is beneficial to enhancing the performance of the depression detection system. There are two principles for enhancing the performance of the depression detection system in Example 2. First, the other sensors used in the present invention also have certain response differences between healthy controls and patients with depression in the test. Therefore, they are not irrelevant sensors, but an effective information supplement. Detailed information can be seen in Figure 5 — Fig. 9 ; Second, in actual use, the subjects of detection are often not completely healthy, or do not have any other diseases except depression. The presence of other diseases may also cause changes in the concentrations of gases such as carbon monoxide, hydrogen sulfide, and nitric oxide. In order to eliminate the influence of other diseases that may exist in the subjects of detection, multiple sensors with different target gases are needed to more comprehensively characterize the existence pattern of the gas samples.

[0070] The detection module is connected to the waveform signal acquisition module and is used for detecting depression according to the waveform signal.

[0071] In one embodiment, the detection module is a micro single-board computer connected to a data acquisition card, and is used to receive waveform signals and perform depression detection based on the waveform signals.

[0072] Furthermore, the present invention also explores the influence of the blowing time of the test subject on the response loss of the airbag-free depression detection system, because the degree of elimination of environmental factors interference ultimately still depends on the total amount of blowing of the test subject. Fig.10 , Fig.11 As shown, it is found through experiments that when the blowing time is about 35 seconds, the response amplitude of the existing airbag-free device with an intermediate cabin is basically close to the response amplitude of the airbag device. If it is less than 30 seconds, the signal attenuation is more serious, resulting in more serious perception response loss. When the present invention is used, the required blowing time is further shortened, and only about 30 seconds are needed to approach the response amplitude of the airbag device.

[0073] Further, such as Fig.12As shown, the detection results of the depression detection system proposed by the present invention are closer to those of the breathing analysis device with airbags. In other words, the depression detection system proposed by the present invention can effectively alleviate the perception loss of the device without airbags, making its results close to the perception results of the device with airbags.

[0074] In another embodiment, the detection module includes a feature extraction unit, a feature selection unit, and a depression detection unit;

[0075] The feature extraction unit is connected to the waveform signal acquisition module and is used to extract features from the waveform signal to obtain a feature set;

[0076] Preferably, feature extraction is performed on the waveform signal to obtain a feature set, specifically:

[0077] At least one of transient feature extraction, frequency domain feature extraction and decomposition signal feature extraction is performed on the waveform signal, such as Fig.13 The following is a typical sample of the waveform signal.

[0078] Different from the steady-state response characteristics of the odor signal, its transient response characteristics usually carry additional usable information because they are usually related to the changes of the sample during the injection and purge stages. In the present invention, the waveform signal is subjected to transient feature extraction, including extraction of amplitude, differential, first-order derivative, second-order derivative, integral, time constant and phase characteristics, a total of 665 dimensions of transient features. Preferably, in the present invention, the transient feature extraction method is to extract features for each channel separately, and finally splice them to form a transient feature set of the sample.

[0079] The exhaled gas 3 collected at a certain time is selected as a sample, and each transient feature is extracted, specifically:

[0080] (1) Maximum amplitude feature

[0081] The amplitude is Fig.13 The value of the y-axis has the physical meaning of the voltage of the sensor. Therefore, the amplitude curve can be understood as the voltage change curve of the sensor.

[0082] The maximum amplitude feature refers to the maximum response value of each sensor in the sample, indicating the maximum response degree of each sensor to the sample. The present invention adopts the sensor array 9 of the embodiment 2 mentioned above, using a total of 7 sensors, namely sensors 2 to 8 for actual analysis, thus generating a total of 7-dimensional features.

[0083] (2) Amplitude downsampling

[0084] Amplitude downsampling means dividing each sensor in the sample into 21 segments. For the first 20 segments, 55 sampling points are taken in each segment. Since each sensor saves a total of 1153 sampling points in one sampling, 53 sampling points are taken in the last segment. The average amplitude is calculated in each segment to represent the segment, and a total of 147-dimensional features are generated here.

[0085] (3) Amplitude downsampling after normalization

[0086] Amplitude downsampling after normalization means that each sensor in the sample is divided by its maximum amplitude to achieve normalization, and then each sensor in the sample is divided into 21 segments. For the first 20 segments, 55 sampling points are taken in each segment. Since each sensor saves a total of 1153 sampling points in one sampling, 53 sampling points are taken in the last segment. For each segment, the average amplitude is calculated to represent the small segment, and a total of 147-dimensional features are generated here.

[0087] (4) Amplitude difference

[0088] Amplitude difference refers to the calculation of the amplitude difference of each sensor in the sample, taking once in the injection phase, twice in the reaction phase, and once in the sweep phase. In the design of the present invention, the sampling time points of the segmentation are [0, 64, 288, 512, 736, 960], and a total of 35-dimensional features are generated here.

[0089] (5) Maximum first-order derivative and minimum first-order derivative

[0090] The maximum first-order derivative and the minimum first-order derivative refer to calculating the maximum first-order derivative value and the minimum first-order derivative value of each sensor for the sample, where a total of 14-dimensional features are generated.

[0091] (6) First-order derivative downsampling

[0092] The sensors in the sample are divided into 21 segments. For the first 20 segments, 55 sampling points are taken in each segment. Since each sensor saves a total of 1153 sampling points in one sampling, 53 sampling points are taken in the last segment. The average derivative value is calculated in each segment to represent the segment, and a total of 147-dimensional features are generated here.

[0093] (7) Maximum and minimum second-order derivatives during sample injection and purge phases

[0094] The maximum second-order derivative and the minimum second-order derivative in the sample injection phase and the purge phase refer to: calculating the maximum second-order derivative value and the minimum second-order derivative value of each sensor of the sample in the injection phase, that is, the 8th to 64th sampling points, and in the sweeping phase, that is, the 512th to 1153rd sampling points, where a total of 28-dimensional features are generated.

[0095] (8) Segmental integration

[0096] Segmented integration means that for each sensor in the sample, one sampling is performed in the injection phase, two samplings are performed in the reaction phase, and one sampling is performed in the cleaning phase to calculate the integral value of the segment. In the design of the present invention, the segmented sampling time points are [0, 64, 288,512, 736, 960], and a total of 35-dimensional features are generated here.

[0097] (9) Time constant-amplitude

[0098] Time constant-amplitude refers to calculating the sampling time points at which the amplitude of each sensor reaches 100%, 90%, 60%, and 30% of the maximum amplitude for each sensor in the sample, generating a total of 28-dimensional features.

[0099] (10) Time constant - first-order derivative

[0100] The time constant-first-order derivative refers to the sampling time points when the first-order derivative of each sensor in the sample reaches the maximum value and the minimum value, and a total of 14-dimensional features are generated here.

[0101] (11) Time constant-second derivative

[0102] Time constant-second-order derivative refers to the sampling time points at which the second-order derivative of each sensor in the injection phase, i.e., the 8th to 64th sampling points, and the second-order derivative of each sensor in the sweep phase, i.e., the 512th to 1153th sampling points, reaches the maximum and minimum values. A total of 28-dimensional features are generated here.

[0103] (12) Phase characteristics

[0104] Phase feature means that for each sensor in the sample, the phase space is constructed based on the amplitude of each sensor and its first-order derivative, that is, the signal curve is redrawn with the amplitude as the x-axis and the first-order derivative as the y-axis. The signal curve in the phase space is taken once in the injection phase, twice in the reaction phase, and once in the sweep phase, and the integral value of the segment is calculated. In the design of the present invention, the segmented sampling time points are [0, 64, 288, 512, 736, 960], and a total of 35-dimensional features are generated here.

[0105] In the present invention, the amplitude downsampling, normalized amplitude downsampling and first-order derivative downsampling features are divided into 21 segments according to time difference, with the purpose of reducing the original data (1153 dimensions) to 147-dimensional features, and summarizing the original change information of the response curve by calculating the amplitude average value in each small segment.

[0106] Furthermore, since the depression detection system perceives samples in three main stages: the inhalation stage, the reaction stage, and the exhalation stage. In the present invention, for the amplitude difference and the segmented integral features, the segmented sampling time points are [0, 64, 288, 512, 736, 960], where [0, 64] corresponds to the inhalation stage, [64, 512] corresponds to the reaction stage, and [512, 960] corresponds to the exhalation stage. Different features can be calculated at each stage of the perception process to improve the accuracy and effectiveness of the detection.

[0107] The exhaled gas 3 collected at a certain time is selected as a sample, and each frequency domain feature is extracted, specifically:

[0108] (1) Wavelet transform characteristics

[0109] By calculating the approximation coefficients of the wavelet transform results, the low-frequency components in the signal can be extracted, and the overall shape and main features of the signal can be retained. For the parameters used in the wavelet transform, the best parameter combination is wname = 'haar', level = 5 after experimental verification.

[0110] (2) FFT (Fast Fourier Transform) Features

[0111] FFT converts the time domain signal into the frequency domain signal, revealing the energy distribution of the signal in different frequency components. By extracting FFT features, the periodic pattern in the signal, the intensity and distribution of the frequency components can be understood. After the sensor response signal is converted into a frequency signal through FFT, the amplitude of the extracted result is used as the final feature.

[0112] (3) STFT (Short-time Fourier Transform) characteristics

[0113] STFT first divides the original signal into multiple short time periods (called "windows"), then performs Fourier transform on the signal in each window to obtain the spectrum information in the window, and finally extracts features for the frequency information in each window. After the present invention converts the sensor response signal into a frequency signal through STFT, the amplitude spectrum of the extraction result is used as the final feature. For the parameters used by STFT, after experimental verification, the optimal parameter combination is fs = 8.0, window = 'hann', nperseg = 516, noverlap = None, nfft = None.

[0114] The exhaled gas 3 collected at a certain time is selected as a sample, and the decomposition features of each signal are extracted, specifically:

[0115] (1) EEMD (Ensembled Empirical Mode Decomposition) Features

[0116] Integrated empirical mode decomposition is an adaptive decomposition method for signal analysis. It decomposes the original signal into multiple intrinsic mode functions (IMFs) to analyze signal characteristics at multiple scales. These IMFs represent different frequency components in the signal and have adaptability and fidelity of local signal characteristics. The present invention decomposes the sensor response signal into IMFs through EEMD, and extracts the mean, variance, standard deviation, kurtosis, skewness, energy and energy entropy of the results as the final features. For the parameters used by EEMD, after experimental verification, the best parameter combination is max_imf=4.

[0117] (2) VMD (Adaptive Signal Decomposition) Features

[0118] VMD is an adaptive signal decomposition method that can decompose complex signals into a set of a finite number of frequency modulation (FM) signals or intrinsic mode functions, each of which has specific frequency and bandwidth characteristics. By analyzing the frequency, energy and other characteristics of each IMF, the information of different components in the signal can be effectively captured, which helps to better understand the frequency changes and characteristic patterns of the signal. The present invention decomposes the sensor response signal into IMFs through VMD, and extracts the mean, variance, standard deviation, kurtosis, skewness, energy and energy entropy of the results as the final features. For the parameters used by VMD, after experimental verification, the optimal parameter combination is alpha = 3000, tau = 0, K = 4, DC = 0, init = 'random', tol = 1e-7.

[0119] The feature selection unit is connected to the feature extraction unit and is used to select the feature set to obtain the optimal feature subset. The essence of feature selection can be understood as selecting a part of the features from the feature set obtained after feature extraction to form a feature subset for the optimization of specific indicators. It usually selects a batch of the most valuable features from the original features to improve the model performance, and increases the calculation speed by reducing the dimension of the input data. It is an important means in the field of data analysis and an indispensable key step in the field of pattern recognition.

[0120] Preferably, the feature set is selected to obtain the optimal feature subset, specifically:

[0121] Based on the feature selection algorithm that combines Inf-FSs and SVM-RFE, the feature set is selected to obtain the optimal feature subset.

[0122] Preferably, based on the feature selection algorithm combining Inf-FSs and SVM-RFE, the feature set is selected to obtain the optimal feature subset, specifically:

[0123] S1. According to the Inf-FSs algorithm and feature set, the neighboring matrix is ​​obtained;

[0124] Inf-FSs refers to the supervised version of Infinite feature selection, that is, sample labels are required in the feature selection process.

[0125] S1 is specifically as follows: according to the Inf-FSs algorithm and feature set, a weighted fully connected graph is constructed, in which each node corresponds to a feature, and the edges between nodes are used to model the relationship between each pair of nodes. Finally, the importance of each feature is calculated by calculating the neighboring matrix of the graph. The present invention only retains the Fisher part of the relationship between nodes, and the final relationship modeling can be expressed as follows:

[0126]

[0127] In the formula, is the mean, is the standard deviation, is the sample of the category labeled 1 The average value of the features; is the sample of the category labeled 2 The average value of the features; is the sample of the category labeled 1 The standard deviation of the feature; is the sample of the category labeled 2 The standard deviation of a feature.

[0128] S2. Obtain a weight matrix according to the SVM-RFE algorithm and feature set;

[0129] SVM-RFE is a classic Wrapper feature selection method. The present invention inputs the feature set into the SVM classifier, optimizes the parameters of the SVM classifier, and obtains a weight matrix, which contains the contribution of each dimensional feature to the classification accuracy; according to the weight matrix, the feature with the lowest contribution is eliminated, and it is iterated to retrain a new SVM classifier and continue to eliminate until all features are eliminated. At this time, the order in which the features are eliminated is sorted from the latest to the earliest, which is the importance ranking of the feature set. At this time, according to the importance sorting, starting from the most important feature, the classifier is trained to check the classification accuracy. Each cycle adds a new feature to the feature subset in the order of importance. Finally, when the classification accuracy is no longer improved, the feature subset at this time is the optimal feature subset.

[0130] S3. According to the neighboring order matrix and the weight matrix, the feature set is selected to obtain the optimal feature subset.

[0131] Preferably, S3 is specifically:

[0132] S31, obtaining a first feature importance score and a second feature importance score according to the neighboring order matrix and the weight matrix;

[0133] S32, normalizing and summing the first feature importance score and the second feature importance score to obtain a final feature importance score;

[0134] S33. According to the final feature importance score, the feature set is selected to obtain the optimal feature subset.

[0135] According to the neighboring matrix and the weight matrix, two different feature importance scores, S1 and S2, can be obtained. S1 comes from solving the neighboring matrix of the graph, and S2 comes from the weight matrix of the SVM. Therefore, the first feature importance score and the second feature importance score are normalized and summed to obtain the final feature importance score through the following formula. By sorting the final feature importance scores from large to small, the features at the front are considered to be more important.

[0136]

[0137] Furthermore, the present invention uses ten-fold cross validation, when the parameter of the SVM classifier is set to kernel='linear', When the feature selection algorithm is used, the classification accuracy of depression and healthy samples is 98.90% when the features in the first 230 dimensions of the sorting results are tested. However, when the feature selection algorithm is not used, the classification accuracy can only be 91.08% when the features in the first 1792 dimensions of the sorting results are tested.

[0138] Traditional SVM-RFE only focuses on the correlation between features and labels, and lacks consideration of intra-feature redundancy. The Inf-FSs algorithm set in the present invention is an improved variant of the original algorithm, which separately retains the consideration of intra-feature redundancy. Therefore, the present invention can supplement the original model's consideration of feature redundancy by fusing the results of Inf-FSs with the results of SVM-RFE, thereby improving model performance.

[0139] The depression detection unit is connected to the feature selection unit and is used for performing depression detection according to the optimal feature subset.

[0140] Preferably, depression detection is performed based on the optimal feature subset, specifically:

[0141] Input the optimal feature subset into the classification model to obtain the classification accuracy corresponding to the optimal feature subset;

[0142] Depression detection based on classification accuracy.

[0143] The feature set or optimal feature subset can be combined with the commonly used classification models to achieve high-precision detection of depression. Table 2 shows the results of multiple classification experiments in which the optimal feature subset is input into the classification models such as SVM, KNN, random forest (RF), logistic regression (LR), and XGBoost.

[0144] Table 2 Classification accuracy results of healthy controls and depression samples on each model

[0145]

[0146] It can be seen from the above table that: based on the data sensed by the depression detection system of the present invention, healthy samples and depression samples can be effectively distinguished in most cases.

[0147] Compared with the prior art, the present invention has the following beneficial effects:

[0148] (1) The present invention sets the exhaled gas collection module to a shape that gradually narrows along the flow direction of the exhaled gas. The part of the exhaled gas entering the exhaled gas collection module has a large volume and a low flow rate; the part of the exhaled gas leaving the exhaled gas collection module has a small volume and a high flow rate, so that the original gas in the exhaled gas collection module is discharged more thoroughly; further, the flow rate of the air inlet part of the exhaled gas collection module is relatively slow, which, on the one hand, allows the tester to complete the blowing in a more relaxed state without excessive force or long-term continuous blowing; on the other hand, it helps to distribute the exhaled gas more evenly in the exhaled gas collection module, reduces the turbulence and eddy current phenomena in the gas flow, reduces the disturbance and damage to the gas sample, and is conducive to maintaining the original characteristics and composition of the gas sample, facilitating subsequent precise analysis.

[0149] (2) The gas outlet of the exhaled gas collection module of the present invention is relatively narrow. Due to the limitation of space and the increase of flow rate, the collision and diffusion between gas molecules are enhanced, which helps to push the interfering gas remaining in the exhaled gas collection module further to the outlet and discharge it, thereby improving the accuracy of depression detection;

[0150] (3) The waveform signal acquisition module of the present invention is equipped with multiple sensors, which can collect waveform signals of different exhaled gases and provide multi-dimensional reference data for the detection of depression;

[0151] (4) The system of the present invention comprehensively analyzes waveform signals by using transient features, frequency domain features, and signal decomposition features simultaneously, which helps to detect depression more comprehensively and accurately;

[0152] (5) The system of the present invention uses a combination of neighbor-order matrices and weight matrices to help discover potential relationships and interactions between features and select a more refined and efficient optimal feature subset, thereby improving the accuracy of depression detection.

[0153] The above are only a few embodiments of the present application and do not constitute any form of limitation to the present application. Although the present application is disclosed as above with preferred embodiments, it is not intended to limit the present application. Any technician familiar with the profession, without departing from the scope of the technical solution of the present application, using the technical content disclosed above to make slight changes or modifications are equivalent to equivalent implementation cases and fall within the scope of the technical solution.

Claims

1. A depression detection system, characterized in that: include: Exhaled gas collection module, waveform signal acquisition module and detection module; The gas channel of the exhaled gas collection module gradually narrows along the flow direction of the exhaled gas, and is used to discharge the original gas in the exhaled gas collection module and collect the exhaled gas by creating a flow rate difference at the gas inlet and the gas outlet of the gas channel; The waveform signal acquisition module is connected to the exhaled gas collection module and is used to collect the waveform signal of the exhaled gas; The detection module is connected to the waveform signal acquisition module and is used to perform depression detection according to the waveform signal; The central axis of the air inlet and the central axis of the air outlet of the gas channel intersect at a preset point; The waveform signal acquisition module includes a solid electrolyte carbon dioxide sensor, a sensor sensitive to combustible gases, and a sensor sensitive to malodorous gases; The detection module includes a feature extraction unit, a feature selection unit and a depression detection unit; The feature extraction unit is connected to the waveform signal acquisition module and is used to extract features from the waveform signal to obtain a feature set; The feature selection unit is connected to the feature extraction unit and is used to select the feature set based on the feature selection algorithm integrating Inf-FSs and SVM-RFE to obtain the optimal feature subset, specifically: According to the Inf-FSs algorithm and the feature set, a neighboring order matrix is ​​obtained; The Inf-FSs mentioned here refers to the supervised version of Infinite feature selection; According to the SVM-RFE algorithm and the feature set, a weight matrix is ​​obtained; Obtaining a first feature importance score and a second feature importance score according to the neighboring order matrix and the weight matrix; Normalizing and summing the first feature importance score and the second feature importance score to obtain a final feature importance score; According to the final feature importance score, the feature set is selected to obtain an optimal feature subset; The depression detection unit is connected to the feature selection unit and is used to perform depression detection according to the optimal feature subset.

2. The depression detection system according to claim 1, characterized in that: The waveform signal is subjected to feature extraction, specifically: At least one of transient feature extraction, frequency domain feature extraction and decomposition signal feature extraction is performed on the waveform signal.

3. The depression detection system according to claim 1, characterized in that: Depression detection is performed according to the optimal feature subset, specifically: Inputting the optimal feature subset into a classification model to obtain a classification accuracy rate corresponding to the optimal feature subset; Depression detection is performed according to the classification accuracy.

Images