Inplanatable analysis method and system for burn prediction result

Abstract

The invention provides an interpretability analysis method and system for a burn prediction result, and belongs to the technical field of burn intelligent analysis. According to the method, a target result is set, a biomarker corresponding to the burn prediction result is continuously adjusted, an improved biomarker is obtained, and the burn prediction result is obtained. And predicting the improved biomarkers by using a random forest model to obtain an improved result, and continuously adjusting the biomarkers corresponding to the burn prediction result until the improvement target is achieved, so that the burn prediction result is interpreted and analyzed according to the improvement target, and the burn prediction accuracy is improved. The technical problem that in the prior art, a machine learning mode cannot be used for carrying out interpretable analysis on a grading and typing prediction result is solved, a clinically understandable decision basis is formed, and systematic calculation support is provided for burn diagnosis and treatment.

Description

Technical Field

[0001] This invention belongs to the field of intelligent burn analysis technology, and relates to an analysis method for burn prediction results using counterfactual analysis technology, specifically an interpretability analysis method and system for burn prediction results. Background Technology

[0002] In medicine, burns refer to tissue damage caused by heat (including flames, hot liquids, steam, or high-temperature objects), electric current, chemical substances (strong acids, strong alkalis), radioactive materials, and certain light waves (such as intense ultraviolet radiation). These burns primarily occur when these causative agents act on the skin and deeper tissues, leading to protein denaturation, cell dysfunction, and even necrosis. Current techniques mainly employ a "three-degree four-classification" or "four-degree five-classification" system to classify the severity of burns. However, these classification methods suffer from drawbacks such as high subjectivity, vague quantitative standards, and difficulties in dynamic assessment.

[0003] See patent application CN2025114554222, which discloses a method for predicting burn grading and classification using machine learning (random forest model). The predicted burn grading and classification results, as shown in the attached figures, are determined by biomarkers BG1, IL-1β, EGF, and BG2. These biomarkers reflect the severity of burns, providing an objective and quantitative predictive reference. However, this machine learning-based grading and classification prediction is a black-box approach, lacking interpretability analysis of the results and thus failing to translate them into clinically understandable decision-making criteria. Summary of the Invention

[0004] In view of the above-described background technology, the existing technology of using machine learning for graded classification prediction is a black box prediction, which has the technical problem that the graded classification prediction results cannot be interpreted. In order to address this technical problem, the present invention proposes an interpretable analysis method and system for burn prediction results.

[0005] This invention sets a target result. By continuously adjusting the biomarkers corresponding to the burn prediction results, improved biomarkers can be obtained. This will improve biomarkers Improved results predicted using the random forest model By continuously adjusting the biomarkers corresponding to the burn prediction results, until... This approach aims to improve burn prediction results and interpret them accordingly, addressing the technical challenge of interpreting graded and classified burn predictions using existing machine learning methods. It provides a clinically understandable basis for decision-making and offers systematic computational support for burn diagnosis and treatment.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A method for interpretability analysis of burn prediction results includes the following steps: S1: Obtain the target result of the burn. The set of biomarkers corresponding to burn prediction results Where i is the biomarker index and n is the number of biomarkers. For the i-th biomarker; S2: Determine the adjustment increment Using adjustment increment For the i-th biomarker Adjustments were made to obtain improved biomarkers. This will improve biomarkers As a clinical burn diagnostic factor, prediction was performed using a random forest model, resulting in improved outcomes. ; S3: Improve the results With the target result Compare; like Then the improvement goal will be achieved; like Then modify and adjust the increment. Repeat step S2 until the improvement target is achieved; S4: Use the improvement targets to interpret and analyze the burn prediction results.

[0007] Further specifying, S1 specifically includes: S1.1: Obtain the target result of the burn. The set of biomarkers corresponding to burn prediction results ; S1.2: Determine the overall weight of each biomarker, and determine the biomarker set based on the overall weight of each biomarker. Adjustment priority for each biomarker; S1.3: Execute S2 according to the adjustment priority of each biomarker.

[0008] Further specifying, the formula for calculating the overall weight of each biomarker is as follows: In the formula, Let be the overall weight of the i-th biomarker, which is dimensionless; SHAP contribution value of the i-th biomarker after standardization, dimensionless; The importance weight of the contribution value to SHAP, dimensionless; The degree centrality of the i-th biomarker after standardization is dimensionless; Weights for degree centrality; The betweenness centrality of the i-th biomarker after standardization is dimensionless; The betweenness centrality weight is dimensionless. To represent the compact centrality of the i-th biomarker after standardization, dimensionless; The centrality weights are dimensionless.

[0009] Further specifying, the SHAP contribution value of the i-th biomarker The calculation formula is: In the formula, The SHAP contribution value for the i-th biomarker is dimensionless. This is a dimensionless set of biomarkers corresponding to burn prediction results. The prediction result obtained by using the random forest model after adding the i-th biomarker is dimensionless. A collection of biomarkers The prediction results obtained using the random forest model are dimensionless; n is the number of biomarkers, in units of 1, and i is the serial number of the biomarker, in units of 1. The SHAP contribution value of the i-th biomarker After standardization, the SHAP contribution value of the i-th biomarker is obtained. .

[0010] Further limiting the degree centrality of the i-th biomarker after standardization Betweenness centrality of the i-th biomarker after standardization and the tight centrality of the i-th biomarker after standardization It is obtained by processing the i-th biomarker through graph and network analysis and then standardizing the processing results.

[0011] Further specifying the degree centrality of the i-th biomarker The calculation formula is: In the formula, In the graph network model, the degree centrality of the i-th biomarker is dimensionless; The adjustment increment for the i-th biomarker is dimensionless; n is the number of biomarkers. Betweenness centrality of the i-th biomarker The calculation formula is: In the formula, In the graph network model, the betweenness centrality of the i-th biomarker is dimensionless; Both j and j are path points in the graph network model; For the i-th biomarker Number of shortest pathological pathways between path point j and path point j, in units of: number; In the graph network model, Number of shortest pathological pathways between path point j and path point j, in units of: number; tight centrality of the i-th biomarker The calculation formula is: In the formula, In the graph network model, represents the compact centrality of the i-th biomarker, dimensionless; j represents the path point in the graph network model, dimensionless; and n represents the number of biomarkers, in units of 1; The distance correlation between the i-th biomarker and the j-th path point is dimensionless; Degree centrality of the i-th biomarker After standardization, the degree centrality of the i-th biomarker is obtained. ; the betweenness centrality of the i-th biomarker After standardization, the betweenness centrality of the i-th biomarker is obtained. ; the tight centrality of the i-th biomarker After standardization, the compact centrality of the i-th biomarker was obtained. .

[0012] Further specifying, the adjustment increment Includes a set of reduction coefficients and increase the set of coefficients If the SHAP contribution value of the i-th biomarker If the value is greater than τ, then the coefficient set is reduced. For the i-th biomarker Make adjustments; if If <τ, then use the increased coefficient set. For the i-th biomarker Adjustments will be made; if the requirements are not met... >τ and If the condition is <τ, then iterate through the set of reduction coefficients. and increase the set of coefficients For the i-th biomarker Adjustments are made to determine the i-th biomarker. The minimum adjustment amount.

[0013] Further specifying, the method of utilizing adjustment increments For the i-th biomarker Adjustments were made to obtain improved biomarkers. Specifically: In the formula, Let i be the i-th biomarker, which is dimensionless; This is the adjustment increment for the i-th biomarker, dimensionless; Let be the i-th improved biomarker, which is dimensionless.

[0014] Further specifying, S4 specifically includes: using the improvement target to determine the expected burn grade after adjustment, the adjusted biomarkers, and the adjustment increment. Based on burn prediction results, adjusted expected burn severity, adjusted biomarkers, and adjustment increments. The results of burn prediction were interpreted and analyzed.

[0015] An interpretability analysis system for burn prediction results, used to implement the aforementioned method for interpretable analysis of burn prediction results, includes: Acquisition module: Used to acquire the target results of burns. The set of biomarkers corresponding to burn prediction results Where i is the biomarker index and n is the number of biomarkers. For the i-th biomarker; Improvement module: Used to determine the adjustment increment. Using adjustment increment For the i-th biomarker Adjustments were made to obtain improved biomarkers. This will improve biomarkers As a clinical burn diagnostic factor, prediction was performed using a random forest model, resulting in improved outcomes. ; Judgment module: used to evaluate the improvement results With the target result Compare; if If so, the improvement goal is achieved; if Then modify and adjust the increment. Repeat the process of improving the module until the improvement goal is achieved; And the interpretation and analysis module: used to interpret and analyze the burn prediction results using the improvement targets.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention provides an interpretability analysis method for burn prediction results, which involves setting target results. By continuously adjusting the biomarkers corresponding to the burn prediction results, improved biomarkers can be obtained. This will improve biomarkers Improved results predicted using the random forest model By continuously adjusting the biomarkers corresponding to the burn prediction results, until... This approach aims to improve burn prediction results and interpret them accordingly, addressing the technical challenge of interpreting graded and classified burn predictions using existing machine learning methods. It provides a clinically understandable basis for decision-making and offers systematic computational support for burn diagnosis and treatment.

[0017] 2. This invention utilizes counterfactual analysis to reduce the coefficient set. and / or increase the set of coefficients Identify the i-th biomarker The adjustment amount is used to determine the i-th biomarker. The minimum adjustment amount, such that the i-th biomarker The adjustment range is minimized, and then the optimal improvement target is determined, so that the i-th improved biomarker is obtained after adjustment. The predicted improvement target is to improve the current burn severity to a more ideal level.

[0018] 3. This invention determines the biomarker set based on the comprehensive weight of each biomarker. The adjustment priority of each biomarker is determined, and the biomarkers are adjusted according to the determined adjustment priority. This provides important guidance for the search strategy in counterfactual analysis and realizes a multi-dimensional comprehensive evaluation from local predictive importance to global structural importance. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the method for interpretability analysis of burn prediction results in this invention; Figure 2 This is a schematic diagram of the burn prediction result interpretability analysis system of the present invention; Figure 3A diagram showing the importance ranking of the SHAP contribution values ​​of the four biomarkers BG1, IL-1β, EGF, and BG2; Figure 4 This is a diagram of the network topology formed by four biomarkers: BG1, IL-1β, EGF, and BG2. Figure 5 A radar chart showing the centrality analysis of four biomarkers: BG1, IL-1β, EGF, and BG2. Figure 6 A heatmap showing the characteristic associations formed by four biomarkers: BG1, IL-1β, EGF, and BG2. Figure 7 This is a logical block diagram of a hierarchical decision computing framework. Detailed Implementation

[0020] The technical solution of the present invention will be further explained and described below with reference to the accompanying drawings and embodiments, but the present invention is not limited to the embodiments described below.

[0021] See Figure 1 This invention proposes a method for interpretability analysis of burn prediction results, comprising the following steps: S1: Obtain the target result of the burn. The set of biomarkers corresponding to burn prediction results Where i is the biomarker index and n is the number of biomarkers. Let be the i-th biomarker; in this invention, BG1, IL-1β, EGF, and BG2 are used as the biomarkers corresponding to the burn prediction results, that is, in this invention, =BG1、 =IL-1β、 =EGF、 =BG2.

[0022] Specifically, S1 includes: S1.1: Obtain the target result of the burn. The set of biomarkers corresponding to burn prediction results ; S1.2: Determine the overall weight of each biomarker, and determine the biomarker set based on the overall weight of each biomarker. Adjustment priority for each biomarker; The formula for calculating the overall weight of each biomarker is as follows: In the formula, Let be the overall weight of the i-th biomarker, which is dimensionless; SHAP contribution value of the i-th biomarker after standardization, dimensionless; The importance weight of the contribution value to SHAP, dimensionless; The degree centrality of the i-th biomarker after standardization is dimensionless; Weights for degree centrality; The betweenness centrality of the i-th biomarker after standardization is dimensionless; The betweenness centrality weight is dimensionless. To represent the compact centrality of the i-th biomarker after standardization, dimensionless; The centrality weights are dimensionless.

[0023] S1.3: Execute S2 according to the adjustment priority of each biomarker.

[0024] S2: Determine the adjustment increment Using adjustment increment For the i-th biomarker Adjustments were made to obtain improved biomarkers. This will improve biomarkers As a clinical burn diagnostic factor, prediction was performed using a random forest model, resulting in improved outcomes. Among them, adjusting the increment Includes a set of reduction coefficients and increase the set of coefficients If the SHAP contribution value of the i-th biomarker If the value is greater than τ, then the coefficient set is reduced. For the i-th biomarker Make adjustments; if If <τ, then use the increased coefficient set. For the i-th biomarker Adjustments will be made; if the requirements are not met... >τ and If the condition is <τ, then iterate through the set of reduction coefficients. and increase the set of coefficients For the i-th biomarker Make adjustments. Utilize adjustment increments. For the i-th biomarker Adjustments were made to obtain improved biomarkers. Specifically: In the formula, Let i be the i-th biomarker, which is dimensionless; This is the adjustment increment for the i-th biomarker, dimensionless; Let be the i-th improved biomarker, which is dimensionless.

[0025] S3: Improve the results With the target result Compare; like Then the improvement goal will be achieved; like Then modify and adjust the increment. Repeat step S2 until the improvement target is achieved; S4: Interpretation and analysis of burn prediction results using improvement targets; S4 specifically includes: using improvement targets to determine the adjusted expected burn grade, adjusted biomarkers, and adjustment increments. Based on burn prediction results, adjusted expected burn severity, adjusted biomarkers, and adjustment increments. The results of burn prediction were interpreted and analyzed.

[0026] See Figure 3 The SHAP contribution value of the i-th biomarker mentioned above The calculation formula is: In the formula, The SHAP contribution value for the i-th biomarker is dimensionless. This is a dimensionless set of biomarkers corresponding to burn prediction results. The prediction result obtained by using the random forest model after adding the i-th biomarker is dimensionless. A collection of biomarkers The prediction results obtained using the random forest model are dimensionless; n is the number of biomarkers (unit: number), and i is the biomarker index (unit: number); where... Figure 3 The SHAP value in this context refers to the SHAP contribution value.

[0027] The SHAP contribution value of the i-th biomarker After standardization, the SHAP contribution value of the i-th biomarker is obtained. Among them, the SHAP contribution value of the i-th biomarker after standardization. The calculation formula is: In the formula, The SHAP contribution value for the i-th biomarker is dimensionless. The SHAP contribution value of the m-th biomarker, dimensionless; The highest SHAP contribution value among all biomarkers. Here is the numerical stability constant, and its value is... The purpose of setting this value is to prevent division by zero.

[0028] In this invention, the degree centrality of the i-th biomarker after standardization is... Betweenness centrality of the i-th biomarker after standardization and the tight centrality of the i-th biomarker after standardization It is obtained by processing the i-th biomarker through graph and network analysis and then standardizing the processing results.

[0029] Degree centrality of the i-th biomarker mentioned above The calculation formula is: In the formula, In the graph network model, the degree centrality of the i-th biomarker is dimensionless; The adjustment increment for the i-th biomarker is dimensionless; n is the number of biomarkers. Betweenness centrality of the i-th biomarker mentioned above The calculation formula is: In the formula, In the graph network model, the betweenness centrality of the i-th biomarker is dimensionless; Both j and j are path points in the graph network model; For the i-th biomarker Number of shortest pathological pathways between path point j and path point j, in units of: number; In the graph network model, Number of shortest pathological pathways between path point j and path point j, in units of: number; The tight centrality of the i-th biomarker mentioned above The calculation formula is: In the formula, In the graph network model, represents the compact centrality of the i-th biomarker, dimensionless; j represents the path point in the graph network model, dimensionless; and n represents the number of biomarkers, in units of 1; The distance correlation between the i-th biomarker and the j-th path point is dimensionless; See Figure 4 , Figure 5 and Figure 6The graph network analysis is built upon a network of biomarker interactions. The network topology uses nodes to represent biomarkers, with node size reflecting the importance of features derived from SHAP analysis. Edge weights are calculated based on eigenvalue correlation (threshold > 0.3). Four nodes are color-coded for easy identification of different pathological factors. The centrality analysis radar chart comprehensively displays three indicators: degree centrality, betweenness centrality, and tight centrality. Degree centrality reflects the number of direct connections between features, betweenness centrality identifies bridge nodes in the network, and tight centrality measures information transmission efficiency. The feature association heatmap precisely displays the correlation strength between biomarkers in matrix form, with numerical labels providing specific quantitative information on the association. Figure 5 In this paper, polar coordinates are used to visually compare three network centrality measures of nodes (degree centrality, betweenness centrality, and compact centrality). Degree centrality (red line) reflects the number of direct connections between nodes; betweenness centrality (blue line) reflects the control a node has over the information flow path; and compact centrality (green line) reflects the average "distance" efficiency from a node to other points in the network. Figure 6 The original correlation coefficients between biomarkers are presented in the form of a numerical matrix, with the color intensity indicating the strength of the correlation. The darker the color, the stronger the correlation.

[0030] Degree centrality of the i-th biomarker After standardization, the degree centrality of the i-th biomarker is obtained. ; the betweenness centrality of the i-th biomarker After standardization, the betweenness centrality of the i-th biomarker is obtained. ; the tight centrality of the i-th biomarker After standardization, the compact centrality of the i-th biomarker was obtained. .

[0031] The degree centrality of the i-th biomarker after standardization is obtained. The calculation formula is: In the formula, To obtain the degree centrality of the i-th biomarker after standardization, which is dimensionless; Let be the degree centrality of the i-th biomarker, which is dimensionless; Let m be the degree centrality of the m-th biomarker, dimensionless. It represents the maximum degree centrality among all biomarkers. Here is the numerical stability constant, and its value is... The purpose of setting this value is to prevent division by zero.

[0032] The betweenness centrality of the i-th biomarker after standardization is obtained. The calculation formula is: In the formula, The betweenness centrality of the i-th biomarker after standardization is dimensionless. represents the betweenness centrality of the i-th biomarker, which is dimensionless; The betweenness centrality of the m-th biomarker is dimensionless; This represents the maximum betweenness centrality among all biomarkers; Here is the numerical stability constant, and its value is... The purpose of setting this value is to prevent division by zero.

[0033] After standardization, the compact centrality of the i-th biomarker was obtained. The calculation formula is: In the formula, To obtain the compact centrality of the i-th biomarker after standardization, which is dimensionless; The compact centrality of the i-th biomarker is dimensionless; The compact centrality of the m-th biomarker is dimensionless; This represents the maximum value among all biomarkers in terms of tight centrality. Here is the numerical stability constant, and its value is... The purpose of setting this value is to prevent division by zero.

[0034] This invention combines counterfactual analysis with a target priority decision-making mechanism, providing clinically logical search direction guidance for counterfactual analysis. Based on clinical knowledge of burn severity, a burn severity function s:{0,1,...,5} is defined. s(c) = Where 0 represents normal tissue, a larger value indicates a more severe burn, c is the burn severity index, and s(c) is the burn severity level. For the prediction results... All possible target categories are sorted by four levels of priority: The first priority P1 = {s(c) = s(c0) - 1} contains direct improvement targets with a severity level one lower than the current one, where c0 is the serial number of the initially predicted burn severity; The second priority P2 = {s(c) < s(c0) - 1} contains multi-level improvement targets with a severity level two or more lower than the current one; The third priority P3 = {0} is the ideal target for restoring normal tissue; The fourth priority P4 = C\({c0} ∪ P1 ∪ P2 ∪ P3) contains all other categories. This hierarchical structure reflects the progressive principle of clinical treatment: First, try the most likely direct improvement, then consider greater improvement, and finally consider the ultimate goal of restoring normal tissue. Based on the combination of this counterfactual analysis method and the target priority decision-making mechanism, the above-mentioned burn prediction results, adjusted expected burn grades, adjusted biomarkers, and adjustment increments can be obtained. Interpret and analyze the burn prediction results. For example, the adjusted expected burn grade is third-degree burn c t = 3, which is the same as the burn prediction result. The adjusted biomarker is IL-1β, thus forming a clinically understandable decision basis.

[0035] See Figure 7 , for the above-mentioned counterfactual analysis method and target priority decision-making mechanism, the decision basis is the constructed hierarchical decision calculation framework. The hierarchical decision calculation framework includes four logical layers, namely the input parsing layer, the analysis and weighting layer, the feature weight fusion module, and the target-driven decision-making module. The input parsing layer receives the biomarker set corresponding to the burn prediction result , that is, it receives [BG1, EGF, IL-1β, BG2], and calls the random forest model to predict the burn grade to obtain the burn prediction result; Then, use the analysis and weighting layer to conduct multi-dimensional analysis on [BG1, EGF, IL-1β, BG2], and calculate the SHAP contribution value of the biomarker to quantify the contribution direction and intensity of each biomarker to the burn prediction result, and construct a network topology graph by calculating the degree centrality, betweenness centrality, and closeness centrality of each biomarker; Then use the feature weight fusion module based on the comprehensive weight of the biomarker The corresponding comprehensive weights are calculated. Finally, using the goal-driven decision-making module, based on the burn prediction results, strategies for transitioning to adjacent milder grades (e.g., from grade 5 to grade 4) are prioritized. If no feasible path exists, the possibility of transitioning to multiple milder grades or even normal tissue (grade 0) is explored. Finally, other grade changes are considered. This layer systematically traverses each biomarker according to priority, combining the direction indicated by the biomarker's SHAP contribution value to generate counterfactual samples scaled on the original values ​​with a series of adjustment increments (e.g., 0.1, 0.3, 1.1, 1.3, etc.). The modified samples are then re-input into the random forest model to evaluate whether the improvement results reach the target grade and to calculate the prediction confidence. Finally, it ranks all feasible intervention options according to the formula "efficiency = confidence × comprehensive weights" and outputs a structured list of intervention recommendations. Therefore, the "structure" of this system is a non-parametric intelligent decision-making process that tightly couples interpretability analysis, network analysis, clinical priority rules, and iterative search strategies.

[0036] See Figure 2 The present invention also proposes an interpretability analysis system for burn prediction results, which is used to implement the above-mentioned interpretability analysis method for burn prediction results, including an acquisition module, an improvement module, a judgment module and an interpretation analysis module; Acquisition module: Used to acquire the target results of burns. The set of biomarkers corresponding to burn prediction results Where i is the biomarker index and n is the number of biomarkers. For the i-th biomarker; Improvement module: Used to determine the adjustment increment. Using adjustment increment For the i-th biomarker Adjustments were made to obtain improved biomarkers. This will improve biomarkers As a clinical burn diagnostic factor, prediction was performed using a random forest model, resulting in improved outcomes. ; Judgment module: used to evaluate the improvement results With the target result Compare; if If so, the improvement goal is achieved; if Then modify and adjust the increment. Repeat the process of improving the module until the improvement goal is achieved; Interpretation and Analysis Module: Used to interpret and analyze burn prediction results using improvement targets.

[0037] The present invention provides an interpretability analysis system for burn prediction results that is completely consistent with the aforementioned method for interpretability analysis of burn prediction results. For details regarding the acquisition module, improvement module, judgment module, and interpretation analysis module that are not fully disclosed in the burn prediction result interpretability analysis system, please refer to the description in the section on interpretability analysis method for burn prediction results above. The present invention will not repeat these details here.

[0038] The above description is only used to illustrate the technical solutions of the present invention, and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing, those skilled in the art should understand that modifications can still be made to the technical solutions described above, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the present invention.

Claims

1. A method for interpretability analysis of burn prediction results, characterized in that, Includes the following steps: S1: Obtain the target result of the burn. The set of biomarkers corresponding to burn prediction results Where i is the biomarker index and n is the number of biomarkers. For the i-th biomarker; S2: Determine the adjustment increment Using adjustment increment For the i-th biomarker Adjustments were made to obtain improved biomarkers. This will improve biomarkers As a clinical burn diagnostic factor, prediction was performed using a random forest model, resulting in improved outcomes. ; S3: Improve the results With the target result Compare; like Then the improvement goal will be achieved; like Then modify and adjust the increment. Repeat step S2 until the improvement target is achieved; S4: Use the improvement targets to interpret and analyze the burn prediction results.

2. The method for interpretability analysis of burn prediction results according to claim 1, characterized in that, S1 specifically includes: S1.1: Obtain the target result of the burn. The set of biomarkers corresponding to burn prediction results ; S1.2: Determine the overall weight of each biomarker, and determine the biomarker set based on the overall weight of each biomarker. Adjustment priority for each biomarker; S1.3: Execute S2 according to the adjustment priority of each biomarker.

3. The method for interpretability analysis of burn prediction results according to claim 2, characterized in that, The formula for calculating the overall weight of each biomarker is as follows: In the formula, Let be the overall weight of the i-th biomarker, which is dimensionless; SHAP contribution value of the i-th biomarker after standardization, dimensionless; The importance weight of the contribution value to SHAP, dimensionless; The degree centrality of the i-th biomarker after standardization is dimensionless; Weights for degree centrality; The betweenness centrality of the i-th biomarker after standardization is dimensionless; The betweenness centrality weight is dimensionless. To represent the compact centrality of the i-th biomarker after standardization, dimensionless; The centrality weights are dimensionless.

4. The method for interpretability analysis of burn prediction results according to claim 3, characterized in that, SHAP contribution value of the i-th biomarker The calculation formula is: In the formula, The SHAP contribution value for the i-th biomarker is dimensionless. This is a dimensionless set of biomarkers corresponding to burn prediction results. The prediction result obtained by using the random forest model after adding the i-th biomarker is dimensionless. A collection of biomarkers The prediction results obtained using the random forest model are dimensionless; n is the number of biomarkers, in units of 1, and i is the serial number of the biomarker, in units of 1. The SHAP contribution value of the i-th biomarker After standardization, the SHAP contribution value of the i-th biomarker is obtained. .

5. The method for interpretability analysis of burn prediction results according to claim 3, characterized in that, Degree centrality of the i-th biomarker after standardization Betweenness centrality of the i-th biomarker after standardization and the tight centrality of the i-th biomarker after standardization It is obtained by processing the i-th biomarker through graph and network analysis and then standardizing the processing results.

6. The method for interpretability analysis of burn prediction results according to claim 5, characterized in that, Degree centrality of the i-th biomarker The calculation formula is: In the formula, In the graph network model, the degree centrality of the i-th biomarker is dimensionless; The adjustment increment for the i-th biomarker is dimensionless; n is the number of biomarkers. Betweenness centrality of the i-th biomarker The calculation formula is: In the formula, In the graph network model, the betweenness centrality of the i-th biomarker is dimensionless; Both j and j are path points in the graph network model; For the i-th biomarker Number of shortest pathological pathways between path point j and path point j, in units of: number; In the graph network model, Number of shortest pathological pathways between path point j and path point j, in units of: number; tight centrality of the i-th biomarker The calculation formula is: In the formula, In the graph network model, represents the compact centrality of the i-th biomarker, dimensionless; j represents the path point in the graph network model, dimensionless; and n represents the number of biomarkers, in units of 1; The distance correlation between the i-th biomarker and the j-th path point is dimensionless; Degree centrality of the i-th biomarker After standardization, the degree centrality of the i-th biomarker is obtained. ; The betweenness centrality of the i-th biomarker After standardization, the betweenness centrality of the i-th biomarker is obtained. ; The close centrality of the i-th biomarker After standardization, the compact centrality of the i-th biomarker was obtained. .

7. The method for interpretability analysis of burn prediction results according to claim 4, characterized in that, The adjustment increment Includes a set of reduction coefficients and increase the set of coefficients If the SHAP contribution value of the i-th biomarker If the value is greater than τ, then the coefficient set is reduced. For the i-th biomarker Make adjustments; if If <τ, then use the increased coefficient set. For the i-th biomarker Adjustments will be made; if the requirements are not met... >τ and If the condition is <τ, then iterate through the set of reduction coefficients. and increase the set of coefficients For the i-th biomarker Adjustments are made to determine the i-th biomarker. The minimum adjustment amount.

8. The method for interpretability analysis of burn prediction results according to claim 1, characterized in that, The use of adjusting increments For the i-th biomarker Adjustments were made to obtain improved biomarkers. Specifically: In the formula, Let i be the i-th biomarker, which is dimensionless; This is the adjustment increment for the i-th biomarker, dimensionless; Let be the i-th improved biomarker, which is dimensionless.

9. The method for interpretability analysis of burn prediction results according to claim 1, characterized in that, Specifically, S4 includes: using improvement targets to determine the expected burn grade after adjustment, the adjusted biomarkers, and the adjustment increment. Based on burn prediction results, adjusted expected burn severity, adjusted biomarkers, and adjustment increments. The results of burn prediction were interpreted and analyzed.

10. A burn prediction result interpretability analysis system, used to implement the burn prediction result interpretability analysis method according to any one of claims 1-9, characterized in that, include: Acquisition module: Used to acquire the target results of burns. The set of biomarkers corresponding to burn prediction results Where i is the biomarker index and n is the number of biomarkers. For the i-th biomarker; Improvement module: Used to determine the adjustment increment. Using adjustment increment For the i-th biomarker Adjustments were made to obtain improved biomarkers. This will improve biomarkers As a clinical burn diagnostic factor, prediction was performed using a random forest model, resulting in improved outcomes. ; Judgment module: used to evaluate the improvement results With the target result Compare; if If so, the improvement goal is achieved; if Then modify and adjust the increment. Repeat the process of improving the module until the improvement goal is achieved; And the interpretation and analysis module: used to interpret and analyze the burn prediction results using the improvement targets.

Images