Method and system for explaining output result of depth model, and depression screening equipment
The data processing system calculates the representation value of the depth model output results of the known feature matrix, which solves the problem that the depth model output results is difficult to explain, and realizes the explanation and calibration of the depth model output results.
Patent Information
- Application Number
- CN202510174812.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-18
- Publication Date
- 2025-06-06
AI Technical Summary
Because the output results of the depth model are based on complex features, it is difficult to directly explain the physical practical significance behind it, which makes it impossible to review and calibrate its output results manually or other means.
Through the data processing system, the processor is used to obtain the output results after data processing, including preprocessing models, depth models and human-computer interactive machines. The system calculates the representation value of the known feature matrix for unknown feature matrix or output results, and obtains the interpretation dimension and interpretation values corresponding to the output results.
The explanation of the output results of the depth model is realized. By obtaining the interpretability of known features to the output results or unknown feature matrix, the projection of abstract features in the depth model in the real world is formed, explaining the physical significance of the depth model.
Smart Images
Figure CN120105059A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of physiological electrical signal processing, and specifically relates to a method, system, and depression screening device for interpreting output results of a deep model. Background Art
[0002] Since the deep model obtains various parameters based on a larger data set, its built-in neural network structure can extract more complex features, so that its output results usually get better results than traditional machine learning models. However, this also makes it impossible to directly analyze the output results of the deep model from a numerical perspective based on the features that make a huge contribution, resulting in very abstract output results. It is difficult to explain the physical reality behind them, especially in classification, regression, clustering and other issues. In other words, it is difficult to explain which known features the output results are based on, resulting in the inability to review and calibrate the output results manually or by other means. Summary of the invention
[0003] The present invention provides a method, system, and depression screening device for interpreting the output results of a deep model, so as to obtain the explanatory dimension corresponding to the output results through the interpretability of known features to an unknown feature matrix.
[0004] In order to solve the above technical problems, the present invention provides a data processing system, which uses a processor to obtain the output results after data processing, including: a processor, on which a preprocessing model, a deep model, and a human-computer interaction machine are arranged; the preprocessing model obtains preprocessing data and its corresponding known feature matrix based on the original data; the deep model obtains the output results and its corresponding unknown feature matrix based on the preprocessing data; a function area is arranged on the operation interface of the human-computer interaction machine; a feature visualization option and a report generation option are arranged on the function area; wherein the feature visualization option is used to calculate the characterization value of the known feature matrix for the output result or the unknown feature matrix, so as to obtain the explanation result corresponding to the output result, that is, the explanation dimension and the explanation value; the report generation option is used to display the output result and the corresponding explanation result on the report display interface.
[0005] Furthermore, obtaining the explanation result corresponding to the output result includes: selecting any explanation method to obtain candidate values of any dimension in the unknown feature matrix and the output result or each dimension in the known eigenvalue matrix; judging whether the candidate value meets the threshold; if so, selecting the optimal candidate value as the representation value, recording the representation value as the explanation value, and recording the known feature corresponding to the representation value as the explanation dimension; if not, changing other explanation methods to obtain candidate values; when all explanation methods fail to obtain the representation value, adjusting the threshold of the candidate value so that the candidate value meets the threshold.
[0006] Furthermore, obtaining candidate values of any dimension in the unknown feature matrix and each dimension in the known eigenvalue matrix includes: any one of the first interpretation method, the second interpretation method, and the third interpretation method; wherein the first interpretation method is configured to calculate the correlation between any dimension in the unknown feature matrix and each dimension in the known feature matrix, and use the correlation coefficient as the candidate value; the second interpretation method is configured to calculate the correlation between any dimension and each dimension in the known feature matrix after performing principal component analysis on the unknown feature matrix, and use the correlation coefficient as the candidate value; the third interpretation method is configured to regress any dimension in the unknown feature matrix with each dimension in the known feature matrix to obtain a loss value, and use the loss value as the candidate value.
[0007] Further, obtaining candidate values of any dimension in the unknown feature matrix and the output result includes: any one of the fourth interpretation method and the fifth interpretation method; wherein the fourth interpretation method is configured to obtain a reference template of the known feature matrix based on historical data with the same output result, calculate the correlation between the reference template and each dimension in the known feature matrix of the online data, and use the correlation coefficient as the candidate value; the fifth interpretation method is configured to obtain a reference template of the known feature matrix based on historical data with different output results, calculate the separability of the reference template and each dimension in the known feature matrix of the online data, and use the separability coefficient as the candidate value.
[0008] Furthermore, the report display interface also displays the explanation dimension and explanation value corresponding to the output result.
[0009] Furthermore, a positioning data option is also provided on the operation interface to call the original data or pre-processed data corresponding to any output result.
[0010] In the second aspect, the present invention provides a depression screening device configured with a data processing system, which uses the subject's EEG data as original data. When the output end of the deep model is configured as a classification model, the output result is a binary classification result of depression; when the output end of the deep model is configured as a regression model, the output result is the degree of depression; when the output end of the deep model is configured as a clustering model, the output result is a depression subtype.
[0011] In the third aspect, the present invention provides a method for explaining the output results of a deep model, including: obtaining preprocessed data and its corresponding known features based on the original data; obtaining the output results and its corresponding unknown feature matrix based on the preprocessed data; calculating the characterization value of the known feature matrix to the unknown feature matrix or the output result; screening the dimensions in the known feature matrix as the explanation dimensions according to the characterization values, and using the characterization values as the explanation values.
[0012] The beneficial effect of the present invention is that the data processing system of the present invention uses a processor to obtain the output results after data processing, and obtains the interpretability of the known features to the output results or the unknown feature matrix by means of the correlation or consistency between the intermediate variables of the deep model and the interpretable traditional features, and obtains the explanatory dimension corresponding to the output results, forming the projection of these abstract features in the deep model in the real world, and then explaining the physical meaning of the entire deep model. This method no longer attempts to explain the actual meaning of these features extracted by the deep model itself or the process of extracting features by the deep model, nor does it require strict acquisition of the real meaning completely corresponding to each feature extracted by the deep model, but rather obtains as much as possible the parts of the features that contribute greatly to the model results that can be intuitively explained based on the current data, and then uses intuitive documents, numerical values and charts to output, which can support both specific analysis of single subject results and automated processing of batch subject results.
[0013] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description and the drawings.
[0014] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS
[0015] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0016] Figure 1 It is the workflow diagram of the data processing system.
[0017] Figure 2 It is a schematic diagram of the operating interface of the human-computer interaction machine.
[0018] Figure 3 yes Figure 2 A partial enlarged view of point A in the middle.
[0019] Figure 4 Figure 2 is a schematic diagram of reporting the average energy in the alpha band as an explanatory dimension.
[0020] Figure 5 It is a schematic diagram of the comparison between the topographic map of the spatial distribution of features and the template of healthy subjects as a reference range.
[0021] Figure 6 It is a histogram of the specific characteristic values of a subject relative to the reference range. DETAILED DESCRIPTION
[0022] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0023] Example 1
[0024] See Figure 1-Figure 6 , this embodiment 1 provides a data processing system, which uses a processor to obtain the output result after data processing, including: a processor, on which a preprocessing model, a deep model, and a human-computer interaction machine are arranged; the preprocessing model obtains preprocessing data and its corresponding known features based on the original data; the deep model obtains the output result and its corresponding unknown feature matrix based on the preprocessing data; a function area is arranged on the operation interface of the human-computer interaction machine; a feature visualization option and a report generation option are arranged on the function area; wherein the feature visualization option is used to calculate the characterization value of the known feature matrix for the output result or the unknown feature matrix, so as to obtain the explanation result corresponding to the output result, that is, the explanation dimension and the explanation value; the report generation option is used to display the output result and the corresponding explanation result on the report display interface.
[0025] Optionally, the preprocessing model includes, but is not limited to, power frequency notching, bandpass filtering, blink artifact detection, electromyography artifact detection, motion artifact detection, bad conductor detection, bad conductor removal, bad conductor interpolation, amplitude normalization, etc.
[0026] Optionally, the known features, such as but not limited to interpretable traditional features such as alpha lateralization and beta band energy, are usually written based on the calculation steps or calculation formulas of a certain known feature to write corresponding function codes, wherein the input of the function is a preprocessed data matrix, and the output is the interpretable traditional eigenvalue matrix, that is, the known feature matrix.
[0027] Optionally, the deep model generally includes a convolution layer, an embedding layer, and a fully connected layer; the fully connected layer is usually modified to adapt to different tasks, such as a deep model applied to a classification task, a deep model applied to a regression task, and a deep model applied to a clustering task, that is, the classification model, regression model, and clustering model in this case. Since the unknown feature matrix is a matrix composed of numerical values formed before the fully connected layer after convolution and embedding when a deep model processes preprocessed data to obtain output results, this type of numerical matrix is usually obtained based on the multi-layer neural network structure in the deep model. There is no intuitive formula with actual physical meaning, and it is difficult to directly explain its meaning.
[0028] As an optional setting method of the operation interface.
[0029] For details, see Figure 2 The operation interface is provided with a function area A, a waveform display area B, and a positioning data area C from top to bottom.
[0030] See Figure 3 , function area A is set with various operation options, such as file options, preprocessing options, classification model options, regression model options, feature visualization options, report generation options, etc., so as to gradually implement the overall process when selecting each option and obtain the interpretation dimension of the output results. Each operation option is an independent step, and its specific functions are as follows:
[0031] (1) File option: This function is to select the original data file. Select the corresponding menu item in the drop-down menu to read the original data files in different formats.
[0032] (2) Preprocessing option: After selecting this function, the preprocessing model will be used to perform operations from "raw data" to "preprocessing" to obtain preprocessed data.
[0033] (3) Classification model option: This is a type of deep model. This function obtains the classification task results. Multiple specific classification models can be selected in the drop-down menu. Since the deep model and the preprocessing model are bound together, after selecting this function, the operation from "raw data" to "classification result" will be performed, and the preprocessed data will be saved as the preprocessed data of a classification model for the classification task, so that the known feature matrix can be calculated based on the preprocessed data later.
[0034] (4) Regression model option: This is another type of deep model. This function obtains the regression task results and allows you to select from a variety of specific regression models in the drop-down menu. Since the deep model and the preprocessing model are bound together, after selecting this function, the operation from "raw data" to "classification results" will be performed, and the preprocessed data will be saved as the preprocessed data of a regression model for the regression task, so that the known feature matrix can be calculated based on the preprocessed data later.
[0035] (5) Feature visualization option: This function will use the preprocessed data to obtain different known features and their corresponding eigenvalue matrices; then, the explanatory dimension corresponding to the output result is obtained through the interpretability of the known features to the unknown feature matrix, that is, the known features that can best explain the output result are found, and then displayed in a visual way. The visualization method includes but is not limited to tables, spatial distribution of features or brain topography, distribution of features over time, such as line graphs, bar graphs, heat maps, and distribution of features relative to a reference range, such as histograms and box plots. In addition, the feature visualization option also sets a threshold adjustment item to adjust the threshold of any characterization value.
[0036] Table 1: Table visualization
[0037]
[0038] (6) Generate report option: After clicking and confirming through the preview report pop-up window, you can finally save the report in a document, table or PDF format. The report display interface also displays the interpretation method and representation value corresponding to each interpretation dimension. Figure 4 As shown in the figure, through data slicing, we can see the traditional features that are most relevant to the full-time output results, that is, 02-alpha means that the explanatory dimension of the output result is the average energy of the alpha band of the O2 channel, and the correlation coefficient of 0.4 means that the average energy of the alpha band is used as the representation value of the known features and the position feature matrix, that is, the explanatory value.
[0039] Optional, see Figure 2 and Figure 3 There are also waveform data display value adjustment items between function area A and waveform display area B, such as Sens (μv / ch) option, TC (s) option, HF (Hz) option, Notch option, DispLength (s) option, Montage option, channel selection (elimination) option, AR option, page turning option, etc., to adjust the display results of waveform data of each channel. Among them, Sens (μv / ch) option: Sens is the abbreviation of Sensitivity, which corresponds to the scaling parameters commonly used in this field; the unit μv / ch represents microvolts per channel, that is, in the drawing area assigned to each channel, the maximum vertical value corresponds to how many microvolts of EEG data; different Sens values can be used to display the same data to change the displayed waveform size. TC (s) option: TC is the abbreviation of time constant Time Constant, which can be calculated by the formula Converted to the high-pass cutoff frequency of the filter, the unit s represents seconds. HF(Hz) option: HF is the abbreviation of High Frequency, that is, the low-pass cutoff frequency of the filter, the unit Hz represents Hertz. Notch option: Notch is the abbreviation of Power Frequency Notch. If this checkbox is checked, it means that the data will be notched at the power frequency, where the power frequency is the preset default value; the power frequency is generally 50Hz or 60Hz, depending on the AC power frequency in different countries or regions, and is set to 50Hz here. DispLength(s) option: DispLength is the abbreviation of display length Display Length, corresponding to the duration parameter commonly used in this field; the unit s represents seconds, that is, a waveform of several seconds in length is drawn in the drawing area displayed on each screen. Montage option: Montage is a reasonable and orderly arrangement of EEG derivatives or channels. These channels are created to display the activity of the entire head and provide lateralization and localization information. It can be intuitively understood as the reference method of the electrode, that is, the drop-down box is responsible for realizing the "re-reference" function. Specifically, AR Montage uses the average value of a certain number of electrodes as a reference, and LE Montage uses a lead adapter to connect the left and right ears to provide a more stable reference point. Select (eliminate) channel options: This function is to select channels. You can select the channels to be eliminated or the channels to be displayed. Specifically, after clicking this button, a pop-up window will appear, and all channels of the current input data will be displayed in the list box. Select the channels that need to be displayed on the current interface, or the channels that need to be eliminated, and then click the "Selection Completed" button. The channels displayed in the waveform display area of the main interface will be updated accordingly. This means selecting or eliminating channels. AR option: This function is noise reduction. Click this button to refresh the waveform display area of the current main interface and display the waveform data after the fixed preprocessing process. The "fixed preprocessing process" mentioned here usually includes bandpass filtering, power frequency notching, re-reference, channel selection, and removal of specific artifacts based on ICA or other methods. Specific artifacts may include "electromyographic artifacts", "electrooculographic artifacts", "motion artifacts", etc. It can also refer to the "threshold detection-based preprocessing" process that first obtains or specifies a threshold based on the current data or data set, then performs abnormal detection based on the threshold, and finally interpolates the abnormal channels and discards the abnormal time period. Page turning option: Contains symbols such as "<<", "<", ">", and ">>": This function controls the page turning of the waveform display area of the main interface; "<<" means going back one page, "<" means the back part, ">" means the forward part, and ">>" means going forward one page. Here, the "forward part" or "backward part" is set to P% of the length of a page, P∈(0,100).
[0040] Optional, see Figure 1 and Figure 2, the positioning data area C is the time domain feature / long-term result display area, which is initially empty. You can click on any position of the data in the positioning data area C to call and display the original data or pre-processed data corresponding to any interpretation channel in the waveform display area B to realize the positioning data supply. Through the displayed visualization results, users can intuitively locate the specific moment of the abnormal value (or outlier) on the entire data. Through interactive methods such as clicking, the waveform display area of the operation interface will refresh the currently displayed time range to the moment corresponding to the abnormal value, so that users can confirm whether the current feature value is reliable. For example, when implementing the classification task of depression diagnosis, the user (or doctor) observes that the deep model result of a subject's data is "depression". It is necessary to further view the traditional features used to explain the current results, such as alpha lateralization, beta band energy, etc. When it is observed that the distribution of a certain traditional feature dimension, or several specific feature values in a traditional feature is relatively outlier (or different from the range of healthy indicators), if the user has doubts about the current feature calculation process, the "Locate Data" function can be used to quickly trace back to the frequency domain, channel, or time period used to calculate the current feature, and then confirm that the current feature value is reliable, that is, confirm the reliability of the output results of the deep model explained using the current feature.
[0041] Furthermore, obtaining the explanation result corresponding to the output result includes: selecting any explanation method to obtain candidate values of any dimension in the unknown feature matrix and the output result or each dimension in the known eigenvalue matrix; judging whether the candidate value meets the threshold; if so, selecting the optimal candidate value as the representation value, recording the representation value as the explanation value, and recording the known feature corresponding to the representation value as the explanation dimension; if not, changing other explanation methods to obtain candidate values; when all explanation methods fail to obtain the representation value, adjusting the threshold of the candidate value so that the candidate value meets the threshold.
[0042] As an optional implementation method of calculating the characterization value of the known feature matrix to the unknown feature matrix.
[0043] There are three ways to obtain candidate values for any dimension in the unknown feature matrix and each dimension in the known eigenvalue matrix, namely the first interpretation method, the second interpretation method, and the third interpretation method; wherein the first interpretation method is configured to calculate the correlation between any dimension in the unknown feature matrix and each dimension in the known feature matrix, take the correlation coefficient as the candidate value, and take the known feature with the largest correlation coefficient and higher than the first threshold as the explanatory dimension; the second interpretation method is configured to calculate the correlation between any dimension and each dimension in the known feature matrix after performing principal component analysis on the unknown feature matrix, take the correlation coefficient as the candidate value, and take the known feature with the largest correlation coefficient and higher than the second threshold as the explanatory dimension; the third interpretation method is configured to regress any dimension in the unknown feature matrix with each dimension in the known feature matrix to obtain a loss value, take the loss value as the candidate value, and take the known feature with the smallest loss value and lower than the third threshold as the explanatory dimension.
[0044] Optionally, the "correlation" refers not only to the correlation indicator "calculating the Pearson correlation coefficient", but also includes the results obtained by other types of correlation analysis methods, such as: Spearman rank correlation coefficient (or rank correlation coefficient), Kendall tau correlation coefficient, consistency correlation coefficient (Concordance Correlation Coefficient, CCC), multiple correlation coefficient, etc. In the PCA and correlation calculation steps, the operation interface will provide adjustable thresholds, and also allow users to select the number of principal components to be explained and allow direct specification of traditional feature libraries that can be used for interpretation.
[0045] Optionally, the variable matrix obtained by principal component analysis of the unknown feature matrix can be intuitively understood as a dimensionality reduction operation on the numerical matrix. Usually, there is a high correlation between the features extracted by the deep model, that is, the features are usually redundant. Therefore, generally speaking, the overall efficiency can be significantly improved through principal component analysis, avoiding repeated calculations of homogeneous features extracted by the deep model.
[0046] As an optional implementation method of calculating the characterization value of the output result of the known feature matrix.
[0047] like Figure 6 As shown in the figure, the dark color is the actual distribution range of the specific value of the feature of the subject. It can be seen that some of the features of the subject do not fall within the reference distribution range, indicating that the subject may be abnormal. Figure 5As shown, the reference of the healthy subject represents the interpretable feature reference value template calculated based on the healthy subject group; the difference between the two explains the possible reason why the final result of the subject is not healthy. Therefore, the interpretation result of the online signal can also be obtained with the help of the reference template of the historical data. Wherein obtaining the candidate value of any dimension and the output result in the unknown feature matrix may include: any one of the fourth interpretation method and the fifth interpretation method; wherein the fourth interpretation method is configured to obtain the reference template of the known feature matrix according to the historical data with the same output result, calculate the correlation between the reference template and each dimension in the known feature matrix of the online data, use the correlation coefficient as the candidate value, and use the known feature with the largest correlation coefficient and higher than the fourth threshold as the interpretation dimension. The fifth interpretation method is configured to obtain the reference template of the known feature matrix according to the historical data with different output results, calculate the separability of each dimension in the reference template and the known feature matrix of the online data, use the separable coefficient as the candidate value, and use the known feature with the largest separable coefficient and higher than the fifth threshold as the interpretation dimension.
[0048] Example 2
[0049] On the basis of Example 1, this Example 2 provides a depression screening device equipped with the data processing system, and uses the EEG data of the subject as the original data.
[0050] When the output end of the deep model is configured as a classification model, the output result is a binary classification result of depression, that is, yes or no, to screen whether the subject is depressed.
[0051] When the output end of the deep model is configured as a regression model, the output result is the depression level, the predicted scale value, or the predicted depression level of the patient, or the predicted improvement effect that the patient may achieve after a certain treatment.
[0052] When the output end of the deep model is configured as a clustering model, the output result is a depression subtype, such as outputting the predicted cause of the patient's depression, or outputting the predicted treatment plan that may be more suitable for the patient.
[0053] Example 3
[0054] Based on Example 1, this Example 3 provides a method for explaining the output results of a deep model, including: obtaining preprocessed data and its corresponding known features based on the original data; obtaining the output results and their corresponding unknown feature matrices based on the preprocessed data; calculating the characterization value of the known feature matrix to the unknown feature matrix or the output result; screening the dimensions in the known feature matrix as the interpretation dimensions according to the characterization value, and using the characterization value as the interpretation value. The process or steps of the method are, for example but not limited to, configured as a corresponding computer program, stored in a computer device, a computer-readable storage medium, or a computer program product, so as to be executed by a processor.
[0055] In the description of the embodiments of the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.
[0056] In the description of the present invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance.
[0057] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. The device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed.
[0058] With the above-mentioned ideal embodiments of the present invention as inspiration, through the above-mentioned description content, relevant personnel can make various changes and modifications without departing from the scope of the technical idea of the present invention.
Claims
1. A data processing system, using a processor to obtain output results after data processing, comprising: A processor, on which a preprocessing model, a depth model, and a human-computer interaction machine are arranged; The preprocessing model obtains preprocessing data and its corresponding known feature matrix based on original data; The deep model obtains output results and their corresponding unknown feature matrix based on preprocessed data; a function area is set on the operation interface of the human-computer interaction machine; it is characterized in that, The function area is provided with a feature visualization option and a report generation option; wherein The feature visualization option is used to calculate the representation value of the known feature matrix to the output result or the unknown feature matrix to obtain the explanation result corresponding to the output result, that is, the explanation dimension and the explanation value; The generate report option is used to display the output results and the corresponding interpretation results on the report display interface.
2. The data processing system according to claim 1, characterized in that: The explanation results corresponding to the output results include: Select any interpretation method to obtain the candidate values of any dimension in the unknown feature matrix and the output result or each dimension in the known eigenvalue matrix; Determine whether the candidate value meets the threshold; if so, select the best candidate value as the representation value, record the representation value as the explanation value, and record the known feature corresponding to the representation value as the explanation dimension; if not, replace other explanation methods to obtain the candidate value; When all interpretation methods fail to obtain the characterization value, the threshold of the candidate value is adjusted so that the candidate value meets the threshold.
3. The data processing system according to claim 2, characterized in that: Obtaining candidate values of any dimension in the unknown feature matrix and each dimension in the known eigenvalue matrix includes: any one of the first interpretation method, the second interpretation method, and the third interpretation method; wherein The first interpretation mode is configured to calculate the correlation between any dimension in the unknown feature matrix and each dimension in the known feature matrix, and use the correlation coefficient as a candidate value; The second interpretation method is configured to calculate the correlation between any dimension and each dimension in the known feature matrix after performing principal component analysis on the unknown feature matrix, and use the correlation coefficient as a candidate value; The third interpretation method is configured to regress any dimension in the unknown feature matrix with each dimension in the known feature matrix to obtain a loss value, and use the loss value as a candidate value.
4. The data processing system according to claim 2, characterized in that: Obtaining candidate values of any dimension and output result in the unknown feature matrix includes: any one of the fourth interpretation method and the fifth interpretation method; wherein The fourth interpretation method is configured to obtain a reference template of a known feature matrix based on historical data with the same output result, calculate the correlation between the reference template and each dimension in the known feature matrix of the online data, and use the correlation coefficient as a candidate value; The fifth interpretation method is configured to obtain a reference template of a known feature matrix based on historical data with different output results, calculate the separability of each dimension in the reference template and the known feature matrix of online data, and use the separable coefficient as a candidate value.
5. The data processing system according to claim 1, characterized in that: The report display interface also displays the explanation dimension and explanation value corresponding to the output result.
6. The data processing system according to claim 1, characterized in that: The operation interface is also provided with a positioning data option to call the original data or pre-processed data corresponding to any output result.
7. A depression screening device equipped with a data processing system as claimed in any one of claims 1 to 6, using the EEG data of the subject as raw data, characterized in that: When the output end of the deep model is configured as a classification model, the output result is a binary classification result of depression; When the output end of the deep model is configured as a regression model, the output result is the degree of depression; When the output end of the deep model is configured as a clustering model, the output result is a depression subtype.
8. A method for interpreting the output of a deep model, characterized in that: include: Obtain preprocessed data and its corresponding known features based on original data; Obtain output results and their corresponding unknown feature matrices based on preprocessed data; Calculate the characterization value of the known feature matrix to the unknown feature matrix or output result; The dimensions in the known feature matrix are selected as explanatory dimensions according to the representation values, and the representation values are used as explanatory values.