Multi-model fused lung squamous cell carcinoma survival probability prediction system

Abstract

The invention discloses a lung squamous cell carcinoma survival probability prediction system fusing multiple models, and relates to the technical field of medical data analysis. Comprising a data acquisition module used for acquiring a lung squamous carcinoma clinical data set and an inspection index of a patient; the data preprocessing and dynamic feature table construction module is used for preprocessing the lung squamous cell carcinoma clinical data set in existing data processing and dynamically updating a physiological feature information table; the model training and predicting module is used for carrying out training and online prediction on various survival analysis models; and the SHAP interpretation and weight calculation module is used for calling a corresponding SHAP algorithm to obtain an average absolute SHAP value of each input feature. According to the method, model interpretation conflict measurement, invalid variable elimination and pseudo high risk verification are provided, and automatic arbitration or artificial recheck is realized through weighted scoring, so that the reasonability and safety of prediction and intervention suggestions are ensured, the decision risk is reduced, and the clinical trust is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of medical data analysis, and in particular to a multi-model fusion system for predicting the survival probability of squamous cell lung cancer. Background Art

[0002] Prognostic assessment of squamous cell lung carcinoma is crucial in clinical diagnosis and treatment, as well as in scientific research. In clinical practice, doctors need to conduct a comprehensive assessment based on a patient's clinical indicators (such as TNM staging and PS score), imaging features (CT / PET-CT), pathological data (differentiation level, PD-L1 expression), and molecular markers (such as gene mutations). However, the optimal modeling methods for different data types vary. For example, structured clinical data is well-suited to traditional statistical models, imaging data relies on deep learning, and genomic data requires complex feature screening algorithms. To ensure the accuracy of squamous cell lung carcinoma survival probability prediction, a multi-model-integrated squamous cell lung carcinoma survival probability prediction system needs to be designed.

[0003] After searching, the Chinese invention patent application with publication number "CN119008022A" proposed "a method, device and storage medium for constructing a prediction model for liver cancer precancerous lesions". Through a biomarker, namely cfDNA, combined with whether there is a peak of cfDNA in a specific interval of the peak graph, liver cancer precancerous lesions can be accurately predicted. Compared with the existing technology, this solution is simple. It only needs to detect the cfDNA peak graph to diagnose HCC precancerous lesions. There is no need for puncture and complex instruments and equipment and experimental plans. It is simple and fast to operate, low cost, short detection cycle and strong universality.

[0004] The Chinese invention patent application with publication number "CN115620903A" proposes "a method and system for predicting the risk of bone metastasis of malignant tumors in all cancer types". By constructing a cancer classification system based on the prevalence of bone metastasis, and based on the cancer patient's cancer type, age, gender, insurance status, tumor histological stage, TNM stage, and liver, lung, and brain metastasis, a pan-cancer bone metastasis risk prediction model is provided to determine the current probability of bone metastasis in tumor patients and accurately predict the probability of bone metastasis in patients with all malignant tumors. The model is easy to use and has high prediction accuracy.

[0005] However, in actual use, the above-mentioned disclosed method systems and the method systems in the prior art are difficult to quantify and capture because the interaction effects between different data modalities are difficult to capture, and different models give opposite prediction tendencies or importance assessments for certain features. There is a lack of systematic conflict measurement and arbitration mechanism, resulting in a lack of reliable interpretation basis for the fusion results, making it difficult to gain clinical trust. Summary of the Invention

[0006] The purpose of the present invention is to provide a multi-model fusion system for predicting the survival probability of lung squamous cell carcinoma to solve the problems raised in the above background technology.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a multi-model fusion lung squamous cell carcinoma survival probability prediction system, comprising:

[0008] Data acquisition module, used to obtain clinical data sets and test indicators of patients with lung squamous cell carcinoma;

[0009] The data preprocessing and dynamic feature table construction module is used to preprocess the existing data of the lung squamous cell carcinoma clinical dataset and dynamically update the physiological feature information table;

[0010] Model training and prediction module, used for training and online prediction of various survival analysis models;

[0011] The SHAP interpretation and weight calculation module is used to call the corresponding SHAP algorithm to obtain the average absolute SHAP value of each input feature and perform weight enhancement on features that exceed the critical threshold;

[0012] The multi-model fusion and dynamic weight assignment module is used to use the meta-learner in the Stacking framework to perform nonlinear combination of the prediction results of each survival analysis model and use the feature uniqueness score obtained by SHAP value as the decision basis for ensemble weights to achieve dynamic weighting;

[0013] The arbitration logic module is used to calculate the comprehensive priority score of each intervention plan and trigger high-confidence execution, manual review or re-feature extraction, and output a priority list of intervention strategies with confidence labels;

[0014] The intervention strategy generation and management module is used to generate intervention plans based on core features and optimize the order and execution process of the plans according to arbitration logic;

[0015] A visual clinical decision-making support interface module, which displays basic patient information, model prediction curves, feature importance, intervention strategies, and arbitration labels in the form of charts and cards;

[0016] The system management and interface module is used to maintain user permissions, log records, model updates and retraining scheduling, and provides interfaces for integration with third-party systems.

[0017] As a further preferred embodiment of the present technical solution, the SHAP algorithm includes: TreeSHAP algorithm, linear SHAP algorithm, DeepSHAP algorithm and SHAP approximate calculation method based on permutation importance, and the SHAP algorithms are respectively adapted to the corresponding model structures to obtain the average absolute SHAP value.

[0018] As a further preferred embodiment of this technical solution, the TreeSHAP algorithm is applicable to the tree-based SHAP-interpretable gradient boosting model, the linear SHAP interpreter is adapted to the semi-parametric characteristics of the Cox proportional hazard model based on the survival analysis framework, the DeepSHAP framework is specially designed for the deep temporal attention network model of the deep neural network architecture, and the random forest survival regression model adopts the SHAP approximate calculation method of the feature permutation method.

[0019] As a further preferred embodiment of the present technical solution, in the arbitration logic module, the model interpretation conflict measurement index determines the conflict situation by comparing the SHAP value directions and absolute value trends of different models, the invalid variable elimination ratio index eliminates variables based on the proportion of variables not included in the physiological characteristic information table in the predicted output, and the rationality verification index detects the aggregation trend of the predicted values ​​near the upper limit of the training data through statistical analysis to mark pseudo-high risks.

[0020] As a further optimization of this technical solution, the model interpretation conflict measurement index, the invalid variable elimination ratio index and the rationality verification index are used to obtain the comprehensive priority score of each intervention strategy through weighted calculation. 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, the manual review mechanism needs to be triggered. When the score is ≤3 points, the feature weights are automatically re-extracted.

[0021] As a further preferred embodiment of the present technical solution, the survival analysis model includes a SHAP explanatory gradient boosting model, a Cox proportional hazard model based on a survival analysis framework, a deep temporal attention network model, and a random forest survival regression model.

[0022] As a further preferred embodiment of the present technical solution, the core characteristics include: PD-L1, TNM, white blood cell count, age and gender, and the intervention plans include: surgery + drug treatment, radical chemoradiotherapy ± immunotherapy, and myelosuppressive drug treatment.

[0023] As a further preferred embodiment of the present technical solution, the data acquisition module obtains the target patient's lung squamous cell carcinoma clinical data set and test indicators from the hospital's HIS / EMR system and laboratory information system. The target patient's lung squamous cell carcinoma clinical data set includes age, gender, TNM stage, white blood cell count, and PD-L1 expression level. The test indicators refer to a critical threshold table, which is determined by the treatment guideline.

[0024] As a further preferred embodiment of the present technical solution, the critical threshold is determined based on a critical threshold table, and weight enhancement is performed on features that exceed the critical threshold by implementing feature weight enhancement for variables whose values ​​in the feature importance distribution list exceed the feature items corresponding to 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.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] This multi-model fusion system for predicting survival probability for squamous cell lung cancer uses multi-model parallel training and stacking fusion, combined with SHAP value-driven dynamic weight allocation. It leverages the complementary advantages of each model in high-dimensional sparse features, staging and immune indicators, temporal dependencies, and interaction analysis, ensuring the accuracy and generalization of survival probability prediction.

[0027] In addition, the SHAP interpretation 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 during the fusion process. This allows for an intuitive understanding of how each feature affects the final prediction during actual use, improving clinical acceptance of the system output.

[0028] It should also be added that by combining the clinical critical threshold to strengthen the weight of the SHAP value, we can ensure that more attention is paid to key risk or protective factors, achieve responsiveness to the individual characteristics of patients, and avoid the shortcomings of static or empirical weight allocation;

[0029] It is worth noting that by proposing model interpretation conflict measurement, invalid variable elimination, and pseudo-high-risk verification, and realizing automated arbitration or manual review through weighted scoring, the rationality and safety of predictions and intervention recommendations are ensured, decision-making risks are reduced, and clinical trust is enhanced;

[0030] Finally, personalized intervention strategies are automatically generated through the ranking of core feature importance and arbitration scores, and the execution process, priority, and safety instructions are displayed in a visual interface in a card-like and graphical manner, helping clinicians to formulate and adjust treatment plans more efficiently. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 A diagram of the method steps when the system of the present invention is running;

[0032] Figure 2 It is the prediction graph of the XGBoost-SHAP model disclosed in the present invention;

[0033] Figure 3 is a prediction graph of the Cox proportional hazards model disclosed in the present invention;

[0034] Figure 4 is the prediction graph of the DeepTimeAttn ​​model disclosed in the present invention;

[0035] Figure 5 It is a prediction graph of the random forest survival regression model disclosed in the present invention;

[0036] Figure 6 The operation logic diagram of the proposed arbitration logic rules disclosed in the present invention;

[0037] Figure 7 This is a diagram of the clinical decision-making support interface for squamous cell lung cancer disclosed in the present invention;

[0038] Figure 8 This is a module composition diagram of the disclosed system of the present invention. DETAILED DESCRIPTION

[0039] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only 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 making creative efforts are within the scope of protection of the present invention.

[0040] Before understanding the technical solution proposed in this application, it should be clear that SHAP mentioned in this application is a model interpretation technology that assigns a uniqueness score to each input feature in the model to indicate the contribution of the input feature to the final predicted structure.

[0041] It should also be added that the method of fusing multiple models in this application is achieved through the weighted averaging method and stacking integration strategy in the existing technology. Specifically, the Stacking integration learning framework of the existing technology is first used to construct a multi-level prediction system, and the prediction results of different models are nonlinearly combined through the meta-learner.

[0042] It should be added that by using SHAP values ​​as the decision basis for model weight allocation, calculating the uniqueness score of each model, and dynamically adjusting the integration weights based on the uniqueness score, not only can different models form complementary advantages in capturing linear relationships, nonlinear interactions, and temporal dependencies, but feature-level prediction basis visualization is also achieved through SHAP's interpretable output.

[0043] Specifically, refer to Figure 1 It can be seen that the present invention provides a prediction method, including: steps S100 to S500.

[0044] Step S100: constructing a dynamically updated physiological feature information table based on a clinical dataset of lung squamous cell carcinoma of the target at multiple time nodes.

[0045] It should be noted that the physiological feature information table provides feature input for different survival analysis models. In addition, the content of the lung squamous cell carcinoma clinical data set in step S100 includes: age, gender, TNM stage (clinical tumor stage), white blood cell count and PD-L1 expression level.

[0046] It should be noted that age is included in the clinical dataset of squamous cell lung carcinoma mainly because the distribution characteristics of different age groups can significantly affect the progression and treatment effects of squamous cell lung carcinoma. The introduction of gender information is to capture potential biological differences and their impact on the treatment response of squamous cell lung carcinoma. TNM staging (clinical tumor staging) is a core indicator in clinical diagnosis, providing information on tumor size, lymph node involvement, and distant metastasis. White blood cell count reflects the patient's immune status. PD-L1 is a protein on the surface of tumor cells, namely programmed death receptor-1, which can "trick" the immune system into not attacking cancer cells.

[0047] It is worth noting that based on existing treatment guidelines (such as NCCN guidelines, IARC data, ELCC consensus, and medical safety consensus), the critical thresholds of each variable in the obtained lung squamous cell carcinoma clinical dataset can be seen in the following table: Critical thresholds of variables in lung squamous cell carcinoma clinical dataset.

[0048]

[0049] It is worth noting that when the critical thresholds of these variables obtained through the critical threshold table of variables in the clinical data set of squamous cell lung cancer are used to construct a dynamically updated physiological characteristic information table in step S100, data screening needs to be combined with the above-mentioned critical thresholds. In addition, the bone marrow suppressive drugs in the critical threshold table of variables in the clinical data set of squamous cell lung cancer are common clinical prescription drugs for squamous cell lung cancer in the prior art.

[0050] For example: gemcitabine + platinum (such as carboplatin) or paclitaxel + carboplatin. Due to the specificity of the physiological characteristics of patients with squamous cell lung cancer, the applicant does not make any specific restrictions on the type of bone marrow suppressive drugs.

[0051] As a preferred embodiment, this embodiment is mainly used to supplement the specific content of the physiological characteristic information table in step S100. It is worth noting that in this embodiment, the input feature items in the physiological characteristic information table match the variable item contents in the critical threshold table of variables in the clinical data set of lung squamous cell carcinoma, specifically age, gender, TNM stage, white blood cell count and PD-L1 expression level.

[0052] Step S200: Associating different survival analysis models with the physiological characteristic information table to predict the survival probability of lung squamous cell carcinoma in the target body, and using the SHAP method to determine the proportional weight value of each input variable.

[0053] It should be noted that in step S200, the different survival analysis models are the SHAP interpretable gradient boosting (XGBoost-SHAP) model, the Cox proportional hazards model based on the survival analysis framework, the deep temporal attention network (DeepTimeAttn) model, and the random forest survival regression model.

[0054] It is worth noting that the SHAP explanatory gradient boosting model implements adaptive feature selection for the high-dimensional sparse features of PD-L1 expression levels. Figure 2 It can be seen that the SHAP interpretive gradient boosting model is used to predict the trend of survival rate within 30 days with different age groups (50 years old, 65 years old, 75 years old) and PD-L1 expression grade combinations as input. It is worth noting that Figure 2 The middle curve shows that even in the elderly group (such as 75 years old), even if PD-L1 expression is high (≥50%), the initial value of the survival probability output by the model is significantly lower than that of the other age groups (50 and 65 years old), indicating that the SHAP explanatory gradient boosting model is significantly sensitive to age and immune expression levels.

[0055] In addition, it should be added that the Cox proportional hazard model based on the survival analysis framework in this application uses TNM stage and white blood cell count as variable inputs. It is worth noting that TNM stage is the main variable. The later the stage, the higher the hazard ratio, which has a stronger explanatory effect on the increased risk of death. White blood cell count is involved in model fitting as an immune response indicator. When the white blood cell count is too low, it indicates immune system suppression, which affects the prognosis of treatment. Figure 3 It can be seen that as the white blood cell count gradually decreases, the survival probability curve shows a steeper downward trend, which indicates that the state of the immune system plays a key role in the patient's prognosis. At the same time, combined with the analysis results of TNM staging, it can be found that even if the white blood cell count of early-stage patients is low, their 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 to the prediction model.

[0056] It is worth noting that the deep temporal attention network model in this application uses age and PD-L1 expression level as input variables, refer to Figure 4It can be seen that the survival probability curve shows an obvious stratification phenomenon. Especially under the combination of different age groups and PD-L1 expression levels, the model can capture more detailed time-dependent characteristics. For example, when PD-L1 is highly expressed (≥50%), the survival probability curve of young patients (less than 60 years old) has a relatively gentle downward trend, while elderly patients (such as 75 years old) show a more significant survival rate attenuation, which further illustrates the impact of the complex interaction between age and immune markers on prognosis.

[0057] Finally, it should be added that the random forest survival regression model in this application uses sex and TNM stage as input variables, refer to 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 patients with lung squamous cell carcinoma. Specifically, the survival probability curve of male patients is lower than that of female patients in all stages, and with the advancement of TNM stage, the survival probability curve shows a clear downward trend.

[0058] It should be emphasized that in step S200, by nonlinearly combining the prediction results of the SHAP explanatory gradient boosting model, the Cox proportional hazard model based on the survival analysis framework, the deep temporal attention network model, and the random forest survival regression model, the multidimensional characteristics of patients with lung squamous cell carcinoma and their interactions can be more comprehensively captured. Specifically, the SHAP explanatory gradient boosting model is good at processing high-dimensional sparse features, the Cox proportional hazard model based on the survival analysis framework is highly sensitive to stage and immune indicators, the deep temporal attention network model can mine time-dependent features, and the random forest survival regression model performs well in the interaction analysis of gender and stage. This multi-model fusion approach not only improves the accuracy of prediction, but also enhances the robustness and generalization ability of the model.

[0059] As a preferred implementation scheme, this implementation scheme is mainly used for determining the proportional weight value of each input variable using the SHAP method in step S200, including: steps S201 to S203.

[0060] Step S201: For the prediction results of each survival analysis model, a matching SHAP interpretation algorithm is applied to obtain the SHAP value of the input variable.

[0061] It should be noted that the SHAP value is used to represent the marginal contribution of the input feature to the prediction result. Specifically, the calculation of the SHAP value follows the following technical specifications: the TreeSHAP algorithm is used for the SHAP interpretive gradient boosting model, the linear SHAP interpreter is used for the survival analysis framework based on the Cox proportional hazard model, the DeepSHAP interpretation framework is applied to the deep temporal attention network model, and the SHAP approximation calculation method based on permutation importance is used for the random forest survival regression model.

[0062] Step S202: Statistical analysis is performed on the obtained SHAP values. 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.

[0063] It should be noted that the SHAP interpretation algorithms used in step S201 are all implemented based on existing technologies: the TreeSHAP algorithm is suitable for the tree-based SHAP interpretation type gradient boosting model, the linear SHAP interpreter is adapted to the semi-parametric characteristics of the Cox proportional hazards model based on the survival analysis framework, the DeepSHAP framework is specially designed for the deep temporal attention network model of the deep neural network architecture, and the random forest survival regression model adopts the SHAP approximate calculation method of the feature permutation method.

[0064] Step S203: Based on the critical threshold table of variables in the lung squamous cell carcinoma clinical dataset, perform weighted optimization adjustment on the feature importance distribution list.

[0065] Specifically, for variables whose values ​​in the feature importance distribution list exceed the corresponding feature items in the critical threshold table, feature weight enhancement is implemented. For variables that do not exceed the threshold, the original weight configuration is maintained. It should be noted that the weight enhancement ratio is determined by calculating the percentage of the absolute difference between the feature importance distribution value and the critical threshold.

[0066] Step S300: Based on the weight ranking of the 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.

[0067] It should be noted that in step S300, by combining the SHAP value ranking results, core features that significantly contribute to the prediction of lung squamous cell carcinoma survival probability are screened, and targeted intervention measures are formulated based on their weight ratios. These core features include but are not limited to PD-L1 expression level, TNM stage, white blood cell count, and age. Their importance ranking reflects their actual influence in the multidimensional model fusion analysis. Furthermore, personalized treatment and monitoring plans are designed based on the importance distribution of different features. For example, for patients with high PD-L1 expression, immunotherapy strategies are prioritized, while for patients with low white blood cell counts, infection prevention and immune support measures are required. In addition, in terms of age, the lower risk of survival in elderly patients suggests the need to incorporate more conservative drug selection and dose adjustment mechanisms into the treatment plan.

[0068] Step S400: constructing a suggestion arbitration logic rule based on the treatment guideline, and the arbitration logic rule is used to analyze the execution order and content of multiple intervention strategy plans.

[0069] It should be noted that, in step S400, it can be seen through the medical safety consensus in the treatment guidelines that in the actual treatment process, there will be conflicts between different treatment plans. For example, there will be interactions between enhanced immunotherapy and chemotherapy, thereby affecting the treatment effect or increasing the risk of adverse reactions. Therefore, the construction of recommended arbitration logic rules is mainly used in this application to weigh the priorities of different intervention strategies and quantify the feasibility and safety of each plan through a confidence scoring mechanism.

[0070] Specifically, refer to Figure 6 It can be seen that the operation logic of the suggested arbitration logic rule includes: step S401 to step S404.

[0071] Step S401: Analyze whether the prediction bases of different models are consistent based on the model interpretation conflict metric.

[0072] It should be noted that since multiple models use different input features and structures, such as the SHAP explanatory gradient boosting model prefers the high-dimensional features of PD-L1, the Cox proportional hazard model focuses on TNM staging, and the deep temporal attention network model extracts temporal features, their prediction results will have directional conflicts.

[0073] Therefore, when step S401 is actually executed, the content of the model interpretation conflict metric is "comparing the SHAP value distributions (such as the direction and absolute value of the SHAP values ​​of PD-L1 and TNM stage) generated by different models to see whether there are opposite trends (such as one model believes that high expression of PD-L1 increases the probability of survival, while the other model believes that the probability of survival decreases)".

[0074] Step S402: Establish an invalid variable elimination ratio index based on the invalid variables in the physiological characteristic information table, and perform elimination analysis on the intervention strategy scheme corresponding to the predicted value output by the corresponding survival analysis model.

[0075] It should be pointed out that, given the individual differences among patients with squamous cell lung cancer, the variables in the collected clinical data set cannot completely correspond one-to-one with 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 patients with squamous cell lung cancer. Therefore, in the actual prediction process, the prediction results generated by using invalid variables in the physiological characteristic information table as feature inputs, as well as the intervention strategy plans generated thereby, should be excluded and not included in the consideration based on data analysis.

[0076] Step S403: Avoid the false high-risk trap of the survival analysis model training data based on the prediction output rationality verification index.

[0077] It should be noted that since the prediction output of the survival analysis model is limited by the time range of the training data set (e.g., the maximum observed survival period is X days), this may 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.

[0078] Specifically, by performing statistical analysis using existing technologies on the predicted survival time values ​​output by the survival analysis model, when it is detected that the predicted values ​​show a significant clustering trend near the upper threshold of the training data (for example, the predicted values ​​of more than 90% of the samples converge to the upper limit of X), they are automatically marked as pseudo-high-risk markers. For prediction results with pseudo-high-risk markers, the pathological characteristics of lung squamous cell carcinoma and the actual physiological parameters of the patient will be comprehensively considered to dynamically adjust the intervention strategy generated by the corresponding survival analysis model.

[0079] Step S404: Integrate three different types of indicators to construct a suggested arbitration logic rule.

[0080] It should be noted that, in step S404, the three different types of indicators are the indicators disclosed in steps S401 to S403.

[0081] Specifically, the model explains the conflict measurement index (weight 0.5), the invalid variable elimination ratio index (weight 0.3) and the rationality verification index (weight 0.2), and obtains the comprehensive priority score of each intervention strategy through weighted calculation. 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, the manual review mechanism needs to be triggered. When the score is ≤3 points, it automatically returns to step S200 to re-extract the feature weight.

[0082] Step S500: Adjust the intervention strategy plan based on the arbitration logic rules, optimize the execution process and content configuration, and realize dynamic presentation and output through a visual clinical decision-making support interface.

[0083] It should be noted that the purpose of step S500 is to structurally integrate the characteristic variable weights, model prediction outputs, intervention strategy recommendations and arbitration confidence scores obtained in the aforementioned steps S100 to S404, and ultimately form a visual survival prediction and intervention recommendation interface for doctors' reference, thereby improving the efficiency and interpretability of clinical decision-making assistance in the treatment of squamous cell lung carcinoma.

[0084] Specifically, refer to Figure 7 The visual clinical decision-making support interface displays the following information: basic clinical variable information of the target patient (age, gender, TNM stage (clinical tumor stage), white blood cell count, and PD-L1 expression level), intervention strategy options, and their arbitration scores (including triggered arbitration logic labels, such as "model interpretation conflict," "variable failure elimination," and "pseudo-high risk correction"). The intervention strategy recommendation card displays the content of the intervention strategy after optimizing the execution process and content configuration.

[0085] It should also be added that during actual use, the visual clinical decision-making support interface can export individualized prediction results and intervention recommendations into standardized structured electronic files (such as HL7 / CDA format), and embed them into the existing hospital HIS system for doctors to review, or import them into the existing hospital's EMR system as patient follow-up data, so that doctors can obtain information more efficiently when formulating treatment plans, while improving the fluency of cross-departmental collaboration.

[0086] As a preferred embodiment, the prediction system of the present invention can be implemented in hardware by a general computing device (such as a server or workstation) and a corresponding storage unit, referring to Figure 8 It can be seen that the SHAP-interpretable survival prediction system for squamous cell lung cancer is divided into several functional modules, which work together as follows:

[0087] The data acquisition module is used to obtain the patient's squamous cell lung cancer clinical data set and test indicators from the hospital's HIS / EMR system and laboratory information system (LIS) data source. The squamous cell lung cancer clinical data set includes age, gender, TNM stage, white blood cell count, and PD-L1 expression level. The module then outputs the squamous cell lung cancer clinical data set, where the test indicators refer to the critical threshold table.

[0088] The data preprocessing and dynamic feature table construction module is used to preprocess the existing data processing of the lung squamous cell carcinoma clinical dataset, and to filter and label the cleaned data in combination with the "variable critical threshold table". Then, the physiological feature information table that is dynamically updated at time nodes is constructed based on the filtered and labeled data.

[0089] The model training and prediction module is used to train and predict multiple survival analysis models online, and output the predicted survival probability of each model for the target patient within a specified time window (such as 30 days, 90 days).

[0090] The SHAP interpretation and weight calculation module is used to call the corresponding SHAP algorithm (TreeSHAP, Linear SHAP, DeepSHAP, Permutation Importance SHAP) to obtain the SHAP value of each input feature, count and normalize the average absolute SHAP value of each feature, generate a global feature importance distribution, perform weight enhancement on features that exceed the critical threshold, and output a list of normalized and threshold-optimized feature weights.

[0091] The multi-model fusion and dynamic weight allocation module is used to use the meta-learner in the Stacking framework to perform nonlinear combination of the prediction results of each model, use the "feature uniqueness score" obtained by SHAP as the decision basis for the integration weight, realize dynamic weighting, and output the final survival probability prediction and corresponding model contribution after weighted fusion.

[0092] The arbitration logic module is used to calculate the comprehensive priority score of each intervention plan according to the established weights (0.5:0.3:0.2) (model explanation conflict measurement index, invalid variable elimination ratio index, and rationality verification index), and trigger high-confidence execution, manual review or re-feature extraction, and output a priority list of intervention strategies with confidence labels.

[0093] The intervention strategy generation and management module is used to generate intervention plans (such as immunotherapy, chemotherapy adjustments, and infection prevention) based on core characteristics (PD-L1, TNM, white blood cell count, and age), optimize the sorting of plans, and design the execution process according to arbitration logic, ultimately recommending intervention strategy cards and their implementation steps.

[0094] The visual clinical decision-making support interface module is used to display basic patient information, model prediction curves, feature importance, intervention strategies and their arbitration labels (such as "model conflict" and "false high-risk correction") in the form of charts and cards, and output interactive web interfaces and structured reports.

[0095] The system management and interface module is used to maintain user permissions, log records, model updates, and retraining scheduling, and provides a RESTful API interface for deep integration with third-party systems (HIS, LIS, EMR).

[0096] Although embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is limited by the accompanying embodiments and their equivalents.

Claims

1. A multi-model fusion system for predicting survival probability of squamous cell lung cancer, characterized by: include: Data acquisition module, used to obtain clinical data sets and test indicators of patients with lung squamous cell carcinoma; The data preprocessing and dynamic feature table construction module is used to preprocess the existing data of the lung squamous cell carcinoma clinical dataset and dynamically update the physiological feature information table; Model training and prediction module, used for training and online prediction of various survival analysis models; The SHAP interpretation and weight calculation module is used to call the corresponding SHAP algorithm to obtain the average absolute SHAP value of each input feature and perform weight enhancement on features that exceed the critical threshold; The multi-model fusion and dynamic weight assignment module is used to use the meta-learner in the Stacking framework to perform nonlinear combination of the prediction results of each survival analysis model and use the feature uniqueness score obtained by SHAP value as the decision basis for ensemble weights to achieve dynamic weighting; The arbitration logic module is used to calculate the comprehensive priority score of each intervention plan and trigger high-confidence execution, manual review or re-feature extraction, and output a priority list of intervention strategies with confidence labels; The intervention strategy generation and management module is used to generate intervention plans based on core features and optimize the order and execution process of the plans according to arbitration logic; A visual clinical decision-making support interface module, which displays basic patient information, model prediction curves, feature importance, intervention strategies, and arbitration labels in the form of charts and cards; The system management and interface module is used to maintain user permissions, log records, model updates and retraining scheduling, and provides interfaces for integration with third-party systems.

2. The multi-model fusion lung squamous cell carcinoma survival probability prediction system according to claim 1 is characterized by: SHAP algorithms include: TreeSHAP algorithm, linear SHAP algorithm, DeepSHAP algorithm and SHAP approximate calculation method based on permutation importance. The SHAP algorithms are adapted to the corresponding model structure to obtain the average absolute SHAP value.

3. The multi-model fusion lung squamous cell carcinoma survival probability prediction system according to claim 2 is characterized by: The TreeSHAP algorithm is suitable for tree-based SHAP-interpretable gradient boosting models. The linear SHAP interpreter is adapted to the semi-parametric nature of the Cox proportional hazards model based on the survival analysis framework. The DeepSHAP framework is designed for deep temporal attention network models based on deep neural network architectures. The random forest survival regression model uses the SHAP approximation calculation method of the feature permutation method.

4. The multi-model fusion lung squamous cell carcinoma survival probability prediction system according to claim 1 is characterized by: In the arbitration logic module, the model interpretation conflict measurement index determines the conflict situation by comparing the SHAP value direction and absolute value trend of different models. The invalid variable elimination ratio index eliminates variables based on the proportion of variables not included in the physiological characteristic information table in the predicted output. The rationality verification index detects the aggregation trend of the predicted value near the upper limit of the training data through statistical analysis to mark false high risks.

5. The multi-model fusion lung squamous cell carcinoma survival probability prediction system according to claim 4 is characterized by: The model explains the conflict measurement index, invalid variable elimination ratio index and rationality verification index, and obtains the comprehensive priority score of each intervention strategy through weighted calculation. The score range is 0-10 points. When the score is ≥7, it is judged as a high-confidence plan. When the score is in the range of 4-6 points, the manual review mechanism needs to be triggered. When the score is ≤3 points, the feature weights are automatically re-extracted.

6. The multi-model fusion lung squamous cell carcinoma survival probability prediction system according to claim 1, characterized in that: The survival analysis models include the SHAP interpretable gradient boosting model, the Cox proportional hazards model based on the survival analysis framework, the deep temporal attention network model and the random forest survival regression model.

7. The multi-model fusion lung squamous cell carcinoma survival probability prediction system according to claim 1, characterized in that: The core characteristics include: PD-L1, TNM, white blood cell count, age and gender; the intervention options include: surgery + drug therapy, radical chemoradiotherapy ± immunotherapy, and myelosuppressive drug therapy.

8. The multi-model fusion lung squamous cell carcinoma survival probability prediction system according to claim 1 is characterized by: The data acquisition module obtains the target patient's lung squamous cell carcinoma clinical data set and test indicators from the hospital's HIS / EMR system and laboratory information system. The target patient's lung squamous cell carcinoma clinical data set includes age, gender, TNM stage, white blood cell count, and PD-L1 expression level. The test indicators refer to a critical threshold table, which is determined by the treatment guideline.

9. The multi-model fusion lung squamous cell carcinoma survival probability prediction system according to claim 8, characterized in that: The critical threshold is determined based on the critical threshold table. The weight enhancement is performed on the features that exceed the critical threshold by implementing feature weight enhancement for the variables whose values ​​in the feature importance distribution list exceed the feature items corresponding to the critical threshold table. For variables that do not exceed the threshold, the original weight configuration is maintained. The weight enhancement ratio is determined by calculating the percentage of the absolute difference between the feature importance distribution value and the critical threshold.

Images