SHAP interpretability-based lung squamous cell carcinoma survival prediction method and system

Abstract

The invention discloses a lung squamous cell carcinoma survival prediction method and system based on SHAP interpretability, and relates to the technical field of medical data analysis. Comprising the following steps: constructing a dynamically updated physiological feature information table based on a lung squamous cell carcinoma clinical data set of a target under multiple time nodes; and associating different survival analysis models with the physiological feature information table to predict the survival probability of the lung squamous cell carcinoma in the target body, the marginal contribution of each input feature to the output of each model is obtained by using the SHAP interpretation algorithm, and the feature uniqueness score is transparently displayed in the fusion process, so that the survival probability of the lung squamous cell carcinoma in the target body can be predicted in the actual use process. The method can visually understand how each feature affects the final prediction, and achieves automatic arbitration or manual recheck through providing a model to explain conflict measurement, invalid variable elimination and pseudo high risk verification and weighted scoring, thereby guaranteeing the reasonability and safety of prediction and intervention suggestions, reducing the decision risk, and enhancing the clinical trust.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical data analysis, in particular to a lung squamous carcinoma survival prediction method and system based on SHAP explainability. BACKGROUND

[0002] The prognosis evaluation of lung squamous carcinoma is crucial in clinical diagnosis and treatment and scientific research. In clinical practice, doctors need to comprehensively evaluate the patient's clinical indicators (such as TNM stage, PS score), imaging features (CT / PET-CT), pathological data (differentiation degree, PD-L1 expression), and molecular markers (such as gene mutations). Different data types have different optimal modeling methods - for example, structured clinical data is suitable for traditional statistical models, imaging data relies on deep learning, and genomic data requires complex feature selection algorithms. In order to ensure the accuracy of lung squamous carcinoma survival probability prediction, a lung squamous carcinoma survival prediction method and system based on SHAP explainability need to be designed.

[0003] According to the search, the Chinese invention patent application with publication number "CN117690591A" proposes a "chronic kidney disease progression risk prediction method, device, equipment and storage medium", which trains a progression prediction model based on the historical medical record information of chronic kidney disease patients; obtains sample data of the patient to be predicted, inputs the sample data into the progression prediction model for risk prediction, and obtains a risk score; calls a machine learning interpretability tool to evaluate the progression prediction model and the obtained risk score, and obtains the contribution value of each evaluation index in the sample data to the risk score; and generates a chronic kidney disease progression risk prediction result according to the risk score and the contribution value of each evaluation index to the risk score. This method can provide more accurate and effective prediction results for chronic kidney disease progression risk.

[0004] The Chinese invention patent application with publication number "CN115602324A" proposes an "osteoporosis risk prediction system based on an ultrasonic osteoporosis risk prediction model", which uses the currently recognized osteoporosis diagnostic standard as a control standard, uses the results of temporal window ultrasonic penetration and temporal window skull thickness obtained by ultrasonic examination technology, analyzes and predicts the relationship between osteoporosis and related factors, constructs different osteoporosis risk prediction models, draws nomograms, establishes a scoring table, uses the computer programming processing system function of the osteoporosis risk prediction system, respectively matches the prediction models suitable for different basic data tuples, and respectively calculates the scores of different basic data tuples and the predicted osteoporosis probability according to the scoring table. The highest score and the highest osteoporosis probability are output after comparison by the output program, and the prediction accuracy is good.

[0005] In actual use, the method system disclosed above and the method system in the prior art are difficult to quantitatively capture the interaction effect between different data modalities, and different models give opposite prediction tendencies or importance evaluations for some features, lack of systematic conflict measurement and arbitration mechanism, resulting in that the result after fusion lacks reliable explanation basis and is difficult to obtain clinical trust. SUMMARY

[0006] The purpose of the present application is to provide a lung squamous cell carcinoma survival prediction method and system based on SHAP explainability to solve the problems raised in the background art.

[0007] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0008] In a first aspect, a lung squamous cell carcinoma survival prediction method based on SHAP explainability is provided, comprising:

[0009] Based on the lung squamous cell carcinoma clinical data set at multiple time nodes, a dynamically updated physiological feature information table is constructed;

[0010] Different survival analysis models are associated with the physiological feature information table to predict the survival probability of lung squamous cell carcinoma in the target body, and the SHAP method is used to determine the proportional weight value of each input variable;

[0011] Based on the weight ordering of the input variables, the feature variables affecting the survival probability of lung squamous cell carcinoma are screened out, and a plurality of intervention strategy schemes corresponding to the feature variables are constructed according to the weight gradient;

[0012] The arbitration logic rules are constructed according to the treatment guidelines, and the arbitration logic rules are used to analyze the execution order and content of the plurality of intervention strategy schemes;

[0013] Based on the arbitration logic rules, the intervention strategy scheme is adjusted, the execution process and content configuration are optimized, and dynamic presentation and output are realized through a visual clinical auxiliary decision-making interface.

[0014] As a further preferred embodiment of the present technical solution, the method for determining the proportional weight value of each input variable using the SHAP method comprises:

[0015] For the prediction results of each survival analysis model, a matching SHAP explanation algorithm is applied to obtain the SHAP value of the input variable;

[0016] The obtained SHAP values are statistically analyzed, the average absolute SHAP value of each input feature is first calculated to quantify its global contribution, then normalization processing is implemented to ensure that the sum of all feature contribution weights is 1, and finally a feature importance distribution list with normalized weights is generated;

[0017] Based on the critical threshold table of variables in the lung squamous carcinoma clinical dataset, a weighted optimization adjustment is performed on the feature importance distribution list.

[0018] As a further preferred embodiment of the technical solution, the operation logic of the proposed arbitration logic rule comprises:

[0019] According to the model explanation conflict metric, it is analyzed whether the prediction basis of different models is consistent;

[0020] According to the invalid variable elimination proportion index established according to the invalid variables in the physiological characteristic information table, the intervention strategy scheme corresponding to the prediction value output by the corresponding survival analysis model is analyzed and eliminated;

[0021] According to the prediction output rationality verification index, the pseudo-high risk trap of the survival analysis model training data is avoided;

[0022] Three types of different indicators are fused to construct the proposed arbitration logic rule.

[0023] As a further preferred embodiment of the technical solution, the model explanation conflict metric is whether the SHAP value distribution generated by different models has opposite trends;

[0024] The invalid variable elimination proportion index is that the variables in the collected clinical data set cannot be completely and one-to-one corresponding to the contents in the physiological characteristic information table;

[0025] The rationality verification index is determined according to the time range of the training data set.

[0026] As a further preferred embodiment of the technical solution, the model explanation conflict metric, the invalid variable elimination proportion index, and the rationality verification index are used to calculate the comprehensive priority score of each intervention strategy by weighting, the score interval is 0-10 points, when the score is greater than or equal to 7, it is determined as a high confidence scheme, when the score is in the interval of 4-6, an artificial review mechanism is triggered, and when the score is less than or equal to 3, the step S200 is returned to extract the feature weight.

[0027] As a further preferred embodiment of the technical solution, the model explanation conflict metric is determined by comparing the SHAP value direction and absolute value trend of different models, the invalid variable elimination proportion index is eliminated based on the proportion of variables not included in the physiological characteristic information table in the prediction output, and the rationality verification index detects the aggregation trend of the prediction value near the upper limit of the training data to mark the pseudo-high risk.

[0028] As the further preferred technical solution, the feature importance distribution list implements the method of weighted optimization adjustment, the critical threshold is determined by acquiring the critical threshold table, and the weight of the feature exceeding the critical threshold is strengthened by the variable of the feature item in the critical threshold table in the feature importance distribution list, the feature weight is strengthened, and the original weight configuration is maintained for the variable not exceeding the threshold, and the weight strengthening proportion is determined by calculating the percentage of the absolute difference between the feature importance distribution value and the critical threshold.

[0029] The second scheme is to further improve the lung squamous carcinoma survival prediction method based on SHAP explainability, and a lung squamous carcinoma survival prediction system based on SHAP explainability is also proposed, which uses the lung squamous carcinoma survival prediction method based on SHAP explainability disclosed above.

[0030] Compared with the prior art, the beneficial effects of the present application are:

[0031] The lung squamous carcinoma survival prediction method and system based on SHAP explainability fully utilize the complementary advantages of each model in high-dimensional sparse features, staging and immune indicators, time series dependence and interaction analysis through multi-model parallel training and Stacking fusion combined with SHAP value driven dynamic weight distribution, which guarantees the accuracy and generalization ability of survival probability prediction;

[0032] In addition, the SHAP explanation algorithm is used to obtain the marginal contribution of each input feature to the output of each model, and the feature uniqueness score is transparently displayed in the fusion process, so that in actual use, the influence of each feature on the final prediction can be intuitively understood, and the acceptance of the system output by clinicians is improved;

[0033] It should be added that by combining the clinical critical threshold with the SHAP value, the weight is strengthened to ensure that more attention is paid to key risk or protective factors, to realize the response ability to individual characteristics of patients, and to avoid the shortcomings caused by static or experience weight distribution;

[0034] It is worth noting that by proposing model explanation conflict measurement, invalid variable elimination and pseudo high risk verification, and by weighted scoring to realize automatic arbitration or manual review, the rationality and safety of prediction and intervention suggestions are guaranteed, the decision risk is reduced, and the clinical trust is enhanced;

[0035] Finally, the core feature importance ranking and arbitration score are used to automatically generate individualized intervention strategies, and the execution process, priority and safety instructions are displayed in a card and chart form in the visualization interface, which helps clinicians to more efficiently formulate and adjust treatment plans. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 The method steps of the system running of the present application are shown in the figure.

[0037] Figure 2 a prediction graph of the XGBoost-SHAP model disclosed by the present application;

[0038] Figure 3 a prediction graph of the Cox proportional hazards model disclosed by the present application;

[0039] Figure 4 a prediction graph of the DeepTimeAttn model disclosed by the present application;

[0040] Figure 5 a prediction graph of the random forest survival regression model disclosed by the present application;

[0041] Figure 6 a running logic diagram of the proposed arbitration logic rule disclosed by the present application;

[0042] Figure 7 a lung squamous carcinoma clinical auxiliary decision interface graph disclosed by the present application. DETAILED DESCRIPTION

[0043] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0044] Before understanding the technical solutions proposed in the present application, it should be clear that SHAP mentioned in the present application is a model explanation technology, which assigns a unique score to each input feature in the model to represent the contribution of the input feature to the final prediction structure.

[0045] In addition, it should be supplemented that the method of fusing multiple models in the present application is realized by the weighting average method and the stacking integration strategy in the prior art. Specifically, first, a multi-level prediction system is constructed by using the existing Stacking ensemble learning framework, and the prediction results of different models are nonlinearly combined by a meta-learner.

[0046] It should be supplemented that the SHAP value is taken as the decision basis for model weight allocation, and the unique score of each model is calculated, and the unique score is used to dynamically adjust the integrated weight. Not only does this make different models complementary in capturing linear relationships, nonlinear interactions and time-dependent properties, but it also realizes feature-level prediction basis visualization through the explainability output of SHAP.

[0047] Specifically, reference is made to Figure 1It can be known that the application provides a prediction method, comprising: step S100-step S500.

[0048] Step S100: based on the lung squamous carcinoma clinical data set at multiple time nodes, a dynamically updated physiological characteristic information table is constructed.

[0049] It should be noted that the physiological characteristic information table provides feature input for different survival analysis models, and in addition, the content of the lung squamous carcinoma clinical data set in step S100 includes age, gender, TNM stage (tumor clinical stage), white blood cell count and PD-L1 expression level.

[0050] It should be noted that the age as an item of content in the lung squamous carcinoma clinical data set, mainly due to the distribution characteristics of different age groups will have significant difference on lung squamous carcinoma progression and treatment effect, the introduction of gender information is to capture the potential biological differences and their influence on lung squamous carcinoma treatment response, TNM stage (tumor clinical stage) as the core index in clinical diagnosis, provides the information of tumor size, lymph node involvement and distant metastasis, white blood cell count reflects the immune status of patients, PD-L1 is a protein on the surface of tumor cells, that is, programmed death receptor-1, which can "deceive" the immune system and make it not attack cancer cells.

[0051] It is worth noting that according to the existing treatment guidelines (such as NCCN guidelines, IARC data, ELCC consensus and medical safety consensus), the critical threshold of each variable in the acquired lung squamous carcinoma clinical data set is as follows lung squamous carcinoma clinical data set variable critical threshold table.

[0052]

[0053] It is worth noting that the critical threshold of these variables obtained through the lung squamous carcinoma clinical data set variable critical threshold table is used to construct the dynamically updated physiological characteristic information table in step S100, and the data screening needs to be combined with the above critical threshold, in addition, the myelosuppressive drugs in the lung squamous carcinoma clinical data set variable critical threshold table are common lung squamous carcinoma clinical prescription drugs in the prior art.

[0054] For example: one of gemcitabine + platinum (such as carboplatin) or paclitaxel + carboplatin, due to the specificity of the physiological characteristic information of lung squamous carcinoma patients, therefore the applicant does not make specific limitation on the type of myelosuppressive drugs.

[0055] As a preferred embodiment, the present embodiment is mainly used to supplement the specific content of the physiological characteristic information table in step S100. It is worth noting that in the present embodiment, the input characteristic items in the physiological characteristic information table and the variable item content in the critical threshold table of the variables in the lung squamous carcinoma clinical data set are matched, specifically age, gender, TNM stage, white blood cell count, and PD-L1 expression level.

[0056] Step S200: different survival analysis models are associated with the physiological characteristic information table to predict the survival probability of lung squamous carcinoma in the target body, and the SHAP method is used to determine the proportional weight value of each input variable.

[0057] It should be noted that in step S200, the different survival analysis models are SHAP-explained gradient boosting (XGBoost-SHAP) model, Cox proportional hazards model based on survival analysis framework, DeepTimeAttn model, and random forest survival regression model.

[0058] It is worth noting that the SHAP-explained gradient boosting model implements adaptive feature selection for high-dimensional sparse characteristics of PD-L1 expression level. In the actual prediction process, reference is made to Figure 2 It can be seen that the SHAP-explained gradient boosting model is used to predict the survival rate change trend within 30 days with different age groups (50 years old, 65 years old, 75 years old) and PD-L1 expression grading combinations as input. It is worth noting that Figure 2 The curve shows that the survival probability of the high age group (e.g. 75 years old) is significantly lower than that of the remaining age groups (50 years old and 65 years old) even if the PD-L1 expression is high (≥ 50%), thus showing the significant sensitivity of the SHAP-explained gradient boosting model to age and immune expression level.

[0059] In addition, it should be supplemented that the Cox proportional hazards model based on the survival analysis framework in the present application takes TNM stage and white blood cell count as variable input. It is worth noting that TNM stage is the main variable, and the risk ratio is higher for later stages, and the explanation of the increased risk of death is strong. White blood cell count as an immune response indicator participates in model fitting. When white blood cells are too low, it indicates that the immune system is suppressed, affecting treatment prognosis. Through Figure 3 It can be seen that as the white blood cell count decreases, the survival probability curve shows a more steep downward trend, which shows that the state of the immune system plays a key role in patient prognosis. At the same time, combined with the analysis results of TNM stage, it can be found that even if the white blood cell count of early-stage patients is low, the survival probability is still significantly higher than that of late-stage patients, which further verifies the importance of the interaction between staging and immune indicators on the prediction model.

[0060] It is worth noting that the deep time sequence attention network model in this application is inputted with age and PD-L1 expression level as variables, and the reference Figure 4 It can be seen that the survival probability curve presents a clear hierarchical phenomenon, especially under the combination of different age groups and PD-L1 expression levels, the model can capture more fine time-dependent features, for example, in the case of high PD-L1 expression (≥50%), the survival probability curve of young patients (less than 60 years old) declines gently, while the survival rate of elderly patients (such as 75 years old) shows a more significant decline, which further illustrates the influence of the complex interaction between age and immune markers on prognosis.

[0061] Finally, it should be added that the random forest survival regression model in this application inputs gender and TNM stage as variables, and the reference Figure 5 It can be seen that the survival probability curve shows that gender and TNM stage have a significant impact on the survival prediction of lung squamous cell carcinoma patients, specifically, the survival probability curve of male patients is lower than that of female patients in each stage, and with the progression of TNM stage, the survival probability curve presents a clear downward trend.

[0062] It is emphasized that in step S200, by nonlinearly combining the prediction results of the SHAP interpretation gradient boosting model, the Cox proportional hazards model based on the survival analysis framework, the deep time sequence attention network model and the random forest survival regression model, the multidimensional characteristics of lung squamous cell carcinoma patients and their interactions can be more comprehensively captured, specifically, the SHAP interpretation gradient boosting model is good at processing high-dimensional sparse features, the Cox proportional hazards model based on the survival analysis framework is more sensitive to staging and immune indicators, the deep time sequence attention network model can mine time-dependent features, and the random forest survival regression model performs well in the interaction analysis of gender and staging. This multi-model fusion method not only improves the prediction accuracy, but also enhances the robustness and generalization ability of the model.

[0063] As a preferred embodiment, the method for determining the proportion weight value of each input variable by using the SHAP method in step S200 mainly includes steps S201-S203.

[0064] Step S201: For the prediction results of each survival analysis model, apply the matching SHAP interpretation algorithm to obtain the SHAP value of the input variable.

[0065] It should be noted that the SHAP value is used to represent the marginal contribution of the input feature to the prediction result, and specifically, the calculation of the SHAP value follows the following technical specifications: TreeSHAP algorithm is used for SHAP interpretive gradient boosting model, linear SHAP interpreter is used based on the survival analysis framework of Cox proportional hazards model, DeepSHAP interpretation framework is applied to deep time sequence attention network model, and SHAP approximate calculation method based on permutation importance is used for random forest survival regression model.

[0066] Step S202: statistical analysis is performed on the obtained SHAP value, the average absolute SHAP value of each input feature is first calculated to quantify its global contribution, then normalization processing is implemented to ensure that the sum of the contribution weights of all features is 1, and finally a feature importance distribution list with normalized weights is generated.

[0067] It should be noted that the SHAP explanation algorithm applied in step S201 is based on existing technology: TreeSHAP algorithm is suitable for tree-based SHAP interpretive gradient boosting model, linear SHAP interpreter is adapted to the semi-parametric characteristics of Cox proportional hazards model based on survival analysis framework, DeepSHAP framework is designed for deep time sequence attention network model of deep neural network architecture, and SHAP approximate calculation method of feature permutation is used for random forest survival regression model.

[0068] Step S203: based on the critical threshold table of variables in the lung squamous cell carcinoma clinical data set, weighted optimization adjustment is performed on the feature importance distribution list.

[0069] Specifically, for the variables in the feature importance distribution list whose values exceed the corresponding feature items in the critical threshold table, feature weight strengthening is implemented, and for the variables that do not exceed the threshold, the original weight configuration is maintained. It should be noted that the weight strengthening proportion is determined by calculating the percentage of the absolute difference between the feature importance distribution value and the critical threshold value.

[0070] Step S300: based on the weight ordering of the input variables, the feature variables affecting the survival probability of lung squamous cell carcinoma are screened out, and a plurality of intervention strategy schemes corresponding to the feature variables are constructed according to the weight gradient.

[0071] It should be noted that in step S300, the core features that have a significant contribution to the prediction of lung squamous carcinoma survival probability are screened out by combining the ranking results of SHAP values, and targeted intervention measures are formulated according to the weight proportion of the core features, which include but are not limited to PD-L1 expression level, TNM stage, white blood cell count and age, and the importance ranking reflects the actual influence in the multi-dimensional model fusion analysis. Further, personalized treatment and monitoring programs are designed according to the importance distribution of different features, for example, for patients with high PD-L1 expression, immunotherapy strategies are preferred, and for patients with low white blood cell count, infection prevention and immune support measures need to be strengthened. In addition, in the age dimension, the low survival probability of elderly patients indicates that more conservative drug selection and dose adjustment mechanisms need to be included in the treatment plan.

[0072] Step S400: Constructing a recommendation arbitration logic rule according to the treatment guidelines, which is used to analyze the execution order and content of multiple intervention strategy schemes.

[0073] It should be noted that in step S400, according to the medical safety consensus in the treatment guidelines, there will be conflicts between different treatment schemes in the actual treatment process, for example: the interaction between intensive immunotherapy and chemotherapy will affect the treatment effect or increase the risk of adverse reactions, therefore, the construction of the recommendation arbitration logic rule in the present application is mainly used to weigh the priority of different intervention strategies, and the feasibility and safety of each scheme are quantified through the confidence score mechanism.

[0074] Specifically, referring to Figure 6 It can be seen that the running logic of the recommendation arbitration logic rule includes steps S401-S404.

[0075] Step S401: According to the model explanation conflict measurement index, analyze whether the prediction basis of different models is consistent.

[0076] It should be noted that since different models use different input features and structures, such as SHAP explanation gradient boosting model preferring PD-L1 high-dimensional features, Cox proportional risk model focusing on TNM staging, and deep time series attention network model extracting time series features, the prediction results will have directional conflicts.

[0077] Therefore, in the actual execution of step S401, the content of the model explanation conflict measurement index is "whether the SHAP value distribution generated by different models (such as the direction and absolute value of the SHAP value of PD-L1 and TNM stage) has opposite trends (such as one model considers that high expression of PD-L1 enhances survival probability, and another model considers that survival probability decreases)".

[0078] Step S402: According to the invalid variable in the physiological characteristic information table, an invalid variable elimination ratio index is established, and the intervention strategy scheme corresponding to the prediction value output by the corresponding survival analysis model is analyzed.

[0079] It should be noted that, due to the individual differences of lung squamous cell carcinoma patients, the variables in the collected clinical data set cannot completely correspond to the contents in the physiological characteristic information table. It is particularly emphasized that the variable items disclosed in step S100 cannot cover all clinical data items of lung squamous cell carcinoma patients. Therefore, in the actual prediction process, the prediction results generated by using the invalid variables in the physiological characteristic information table as characteristic inputs and the intervention strategy schemes generated therefrom should be excluded from the consideration based on data analysis.

[0080] Step S403: According to the prediction output rationality checking index, the pseudo-high-risk trap of the survival analysis model training data is avoided.

[0081] It should be particularly noted that, due to the limitation of the prediction output of the survival analysis model to the time range of the training data set (such as the longest observation survival period of X days), it will lead to misjudgment of high-risk cases. Therefore, the core function of step S403 is to improve the clinical rationality of the prediction results of the survival analysis model through algorithm optimization.

[0082] Specifically, by performing statistical analysis on the survival time prediction value output by the survival analysis model, when it is detected that the prediction value is significantly clustered near the upper limit threshold of the training data (for example, more than 90% of the samples converge to X, the upper limit), it is automatically marked as a pseudo-high-risk label. For the prediction results with pseudo-high-risk label, the pathological characteristics of lung squamous cell carcinoma and the actual physiological parameters of the patient are comprehensively considered, and the intervention strategy scheme generated by the corresponding survival analysis model is dynamically adjusted. It should be noted that the time range associated with the training data set (such as the longest observation survival period of X days) indicates that the upper limit threshold of the training data is the Xth day.

[0083] Step S404: Fusion of three different types of indexes to construct a suggestion arbitration logic rule.

[0084] It should be noted that in step S404, the three different types of indexes are the indexes disclosed in steps S401 to S403.

[0085] Specifically, the model explanation conflict metric (weight 0.5), the invalid variable elimination ratio metric (weight 0.3), and the rationality verification metric (weight 0.2) are combined by weighted calculation to obtain the comprehensive priority score of each intervention strategy, the score interval is 0-10 points, when the score is ≥7, it is determined as a high confidence scheme, when the score is in the interval of 4-6 points, an artificial review mechanism is triggered, and when the score is ≤3 points, it is automatically returned to step S200 to extract the feature weight again.

[0086] Step S500: adjusting the intervention strategy scheme based on the arbitration logic rule, optimizing the execution process and content configuration, and realizing dynamic presentation and output through a visual clinical auxiliary decision interface.

[0087] It should be noted that the purpose of step S500 is to structurally integrate the feature variable weight, model prediction output, intervention strategy suggestion, and arbitration confidence score obtained in the foregoing steps S100 to S404, and finally form a visual survival prediction and intervention suggestion interface for doctors to refer to, so as to improve the clinical auxiliary decision efficiency and interpretability in the treatment process of lung squamous cell carcinoma.

[0088] Specifically, referring to Figure 7 It can be seen that the content displayed in the visual clinical auxiliary decision interface includes: basic clinical variable information (age, gender, TNM stage (tumor clinical stage), white blood cell count, and PD-L1 expression level) of the target patient, intervention strategy scheme and arbitration score (including triggered arbitration logic labels such as “model explanation conflict”, “variable invalidity elimination”, and “pseudo-high risk correction”), and intervention strategy recommendation card for displaying the content of the intervention strategy scheme after optimization of the execution process and content configuration.

[0089] In addition, it should be further supplemented that in the actual use process, the visual clinical auxiliary decision interface can export the individualized prediction result and intervention suggestion as a standardized structured electronic file (such as HL7 / CDA format), and through embedding into the existing hospital HIS system, the doctor can review it, or as patient follow-up data, it is imported into the existing hospital EMR system, so that the doctor can more efficiently obtain information when formulating a treatment scheme, and at the same time, the smoothness of cross-department cooperation is improved.

[0090] As a preferred embodiment, the prediction system of the present application can be realized on hardware by a general-purpose computing device (such as a server or a workstation) and a corresponding storage unit. The lung squamous cell carcinoma survival prediction system based on SHAP interpretability is divided into several functional modules, which work cooperatively, and the specific implementation is as follows:

[0091] A data collection module is configured to acquire a patient lung squamous carcinoma clinical data set and test indexes from a hospital HIS / EMR system and a laboratory information system (LIS) data source, the lung squamous carcinoma clinical data set including age, gender, TNM stage, white blood cell count, and PD-L1 expression level, and then output the lung squamous carcinoma clinical data set, wherein the test indexes refer to a critical threshold table.

[0092] A data preprocessing and dynamic feature table construction module is configured to perform preprocessing on the lung squamous carcinoma clinical data set in existing data processing, and screen and label the cleaned data in combination with the "variable critical threshold table", and then construct a physiological feature information table dynamically updated with time nodes through the screened and labeled data.

[0093] A model training and prediction module is configured to train and online predict a plurality of survival analysis models, and output survival probability prediction values of each model for a target patient within a specified time window (such as 30 days or 90 days).

[0094] A SHAP explanation and weight calculation module is configured to call a corresponding SHAP algorithm (TreeSHAP, linear SHAP, DeepSHAP, permutation importance SHAP) to obtain SHAP values of each input feature, statistically and normalize the average absolute SHAP values of each feature, generate a global feature importance distribution, perform weight reinforcement on features exceeding a critical threshold, and output a feature weight list optimized by normalization and threshold.

[0095] A multi-model fusion and dynamic weight distribution module is configured to use a meta-learner in a Stacking framework to nonlinearly combine prediction results of each model, use "feature uniqueness scores" derived by SHAP as a decision basis for integrated weights, realize dynamic weighting, and output a final survival probability prediction after weighting and fusion and a corresponding model contribution degree.

[0096] An arbitration logic module is configured to calculate a comprehensive priority score of each intervention scheme according to a (model explanation conflict metric, invalid variable elimination ratio metric, rationality verification metric) in a predetermined weight (0.5:0.3:0.2), trigger high-confidence execution, manual review, or re-feature extraction, and output an intervention strategy priority list with a confidence label.

[0097] An intervention strategy generation and management module is configured to generate intervention schemes (such as immunotherapy, chemotherapy adjustment, and infection prevention) for core features (PD-L1, TNM, white blood cell count, and age), and optimize and sort the schemes according to arbitration logic and design an execution process, finally recommend an intervention strategy card and its execution steps.

[0098] A visual clinical decision support interface module to present patient basic information, model prediction curves, feature importance, intervention strategies and their arbitration labels (e.g. "model conflict" "pseudo high risk correction") in charts and cards, and to output interactive web interface and structured reports.

[0099] A system management and interface module to maintain user permissions, logging, model update and retraining schedule, and to provide RESTful API interface for deep integration with third party systems (HIS, LIS, EMR).

[0100] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are only by way of example and that changes, modifications, substitutions and alterations can be made thereto without departing from the spirit and scope of the application as defined in the following claims, in which:

Claims

1. A survival prediction method for squamous cell lung cancer based on SHAP interpretability, characterized by: include: Based on the clinical dataset of lung squamous cell carcinoma at multiple time points, a dynamically updated physiological feature information table is constructed; Different survival analysis models were associated with the physiological characteristic information table to predict the survival probability of lung squamous cell carcinoma in the target body, and the proportional weight value of each input variable was determined using the SHAP method; Based on the weight ranking of input variables, characteristic variables that affect the survival probability of lung squamous cell carcinoma are screened out, and multiple intervention strategy plans corresponding to the characteristic variables are constructed according to the weight gradient; Constructing arbitration logic rules based on treatment guidelines to analyze the execution order and content of multiple intervention strategies; Adjust intervention strategies based on arbitration logic rules, optimize execution processes and content configuration, and achieve dynamic presentation and output through a visual clinical decision-making support interface.

2. The method for predicting survival of squamous cell lung cancer based on SHAP interpretability according to claim 1, characterized in that: Methods for determining the proportional weight values ​​of each input variable using the SHAP method include: For the prediction results of each survival analysis model, the matching SHAP interpretation algorithm is applied to obtain the SHAP value of the input variable; The obtained SHAP values ​​are statistically analyzed. First, the average absolute SHAP value of each input feature is calculated to quantify its global contribution. Then, normalization is performed to ensure that the sum of all feature contribution weights is 1. Finally, a feature importance distribution list with normalized weights is generated. Based on the critical threshold table of variables in the clinical dataset of lung squamous cell carcinoma, weighted optimization and adjustment of the feature importance distribution list were performed.

3. The method for predicting survival of squamous cell lung cancer based on SHAP interpretability according to claim 1, characterized in that: The recommended operating logic of arbitration logic rules includes: Based on the model interpretation conflict metrics, analyze whether the prediction basis between different models is consistent; Establish invalid variable elimination ratio indicators based on invalid variables in the physiological characteristic information table, and perform elimination analysis on the intervention strategy plans corresponding to the predicted values ​​output by the corresponding survival analysis model; Avoid the false high-risk trap of survival analysis model training data based on the prediction output rationality verification index; Integrate three different types of indicators to construct logical rules for recommended arbitration.

4. The method for predicting survival of squamous cell lung cancer based on SHAP interpretability according to claim 3, characterized in that: The model interpretation conflict metric is to compare whether the distributions of SHAP values ​​generated by different models have opposite trends; Invalid variables, the invalid variable elimination ratio index was established because the variables in the collected clinical data set could not completely correspond one-to-one with the contents in the physiological characteristic information table; The rationality check index is determined based on the time range of the training data set.

5. The method for predicting survival of squamous cell lung cancer based on SHAP interpretability according to claim 3, characterized in that: The model interpretation conflict measurement index, invalid variable elimination ratio index and rationality verification index are weightedly calculated to obtain a comprehensive priority score for each intervention strategy. The score range is 0-10 points. When the score is ≥7, it is judged as a high-confidence solution. When the score is in the range of 4-6 points, a manual review mechanism needs to be triggered. When the score is ≤3 points, it automatically returns to step S200 to re-extract feature weights.

6. The method for predicting survival of squamous cell lung cancer based on SHAP interpretability according to claim 3, characterized in that: The model interpretation conflict metric determines the conflict situation by comparing the direction and absolute value trend of the SHAP values ​​of different models. The invalid variable elimination ratio indicator eliminates variables based on the proportion of variables not included in the physiological characteristic information table in the predicted output. The rationality verification indicator detects the clustering trend of the predicted values ​​near the upper limit of the training data through statistical analysis to mark false high risks.

7. The method for predicting survival of squamous cell lung cancer based on SHAP interpretability according to claim 2, characterized in that: A method for implementing weighted optimization adjustment on the feature importance distribution list, determining the critical threshold by obtaining the critical threshold table, and performing weight enhancement on features that exceed the critical threshold. The method is to implement feature weight enhancement by the variables in the feature importance distribution list whose values ​​exceed the corresponding feature items in the critical threshold table. For variables that do not exceed the threshold, the original weight configuration is maintained, and the weight enhancement ratio is determined by calculating the percentage of the absolute difference between the feature importance distribution value and the critical threshold.

8. A SHAP-interpretable survival prediction system for squamous cell lung cancer, characterized by: The method for predicting survival of squamous cell lung cancer based on SHAP interpretability described in any one of claims 1 to 7 is used.

Images