Elderly sepsis index model, senile sepsis prediction model construction method and senile sepsis prediction system

Abstract

The invention relates to the technical field of medical data mining, and particularly discloses an elderly sepsis index model, a prediction model construction method and a prediction system.The method comprises the steps that a development queue data set and an independent verification queue data set of elderly sepsis patients at the age larger than or equal to 60 years old are determined according to an intensive care database; performing sepsis death and prognosis importance sorting on candidate inflammation-immune composite indexes of the development queue data set according to a total correlation feature selection algorithm; superposing and combining preset basic indexes, each index in the importance inflammation marker indexes and a preset importance scoring standard; comparing the scoring results of the multiple groups of joint optimization index models, and evaluating the scoring results; and determining the joint optimization index model with the optimal evaluation result as the senile sepsis index model. According to the construction method of the senile sepsis index model provided by the invention, the index model suitable for senile sepsis can be obtained based on the existing index database.

Description

Technical Field

[0001] This invention relates to the field of medical data mining technology, and in particular to a method for constructing an indicator model for elderly sepsis, a method for constructing a 28-day mortality risk prediction model for elderly sepsis, and a 28-day mortality risk prediction system for elderly sepsis. Background Technology

[0002] Sepsis, a common and highly fatal disease in critical care medicine, has a significantly higher 28-day mortality rate in the elderly (≥60 years old) due to factors such as decreased physiological reserves, weakened immune function, and the coexistence of multiple underlying diseases. In clinical practice, severity scoring tools such as SOFA (Sequential Organ Failure Assessment), APACHE II, and SAPS II are commonly used for prognostic assessment. Existing studies suggest that inflammatory / immune markers (such as the systemic immune inflammatory index SII) constructed based on routine blood cell counts and biochemical indicators are associated with the risk of death from sepsis and can be used for prognostic assessment.

[0003] However, existing technologies have the following main problems when predicting the 28-day mortality risk of elderly patients with sepsis:

[0004] (1) There is a lack of a foundation for building models based on large-sample elderly sepsis cohorts. Most previous studies have focused on "adult sepsis as a whole" or "general ICU population", mixing patients of different age groups in the analysis. There are few large-sample cohorts specifically built for elderly sepsis patients, and even fewer systematic designs such as "development cohort combined with independent validation cohort". Therefore, existing models are unable to fully reflect the characteristics of elderly patients such as immune aging and multiple co-occurrence of diseases, and their predictive performance and external generalization in the elderly population are still unclear.

[0005] (2) There is a lack of systematic screening and optimization process for inflammatory composite indicators and traditional scores. Although some literature has suggested that a single inflammatory indicator is associated with sepsis mortality, most of them are isolated indicator analyses with limited sample sizes. There is a lack of a systematic process for comparing and combining multiple inflammatory composite indicators such as PNI, PWR, NPR, LMR, PAR, dNLR, and SII with traditional scores such as SOFA under a unified framework, which ultimately leads to insufficient model specificity and intelligence.

[0006] Therefore, how to obtain an indicator model applicable to elderly sepsis based on existing indicator databases has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] This invention provides a method for constructing an indicator model for elderly sepsis, a method for constructing a 28-day mortality risk prediction model for elderly sepsis, and a 28-day mortality risk prediction system for elderly sepsis, solving the problem in related technologies that it is impossible to obtain an indicator model applicable to elderly sepsis based on existing indicator databases.

[0008] As a first aspect of the present invention, a method for constructing an indicator model of sepsis in the elderly is provided, comprising:

[0009] A development cohort dataset and an independent validation cohort dataset for elderly sepsis patients aged ≥60 years were determined based on an intensive care unit (ICU) database. The data in the development cohort dataset and the independent validation cohort dataset were obtained from two completely different ICU databases. The development cohort dataset included at least candidate inflammatory and immune composite indicators based on nutritional status, coagulation / platelet ratio, immune cell balance, and comprehensive inflammatory burden dimensions.

[0010] The candidate inflammatory and immune composite indicators in the development cohort dataset are ranked by importance for sepsis mortality and prognosis using a fully relevant feature selection algorithm to obtain important inflammatory marker indicators.

[0011] The preset basic indicators are superimposed and combined with each of the important inflammatory marker indicators and the preset importance scoring criteria to obtain multiple sets of joint optimization indicator models, wherein the preset basic indicators include at least age and lactate.

[0012] The scoring results of multiple joint optimization index models are compared, and the scoring results are evaluated based on the independent validation cohort dataset.

[0013] The optimal joint optimization index model based on the evaluation results is determined as the elderly sepsis index model. The indexes in the elderly sepsis index model include at least age, lactate, and SII.

[0014] Furthermore, the method for constructing the sepsis index model for the elderly also includes the following steps prior to determining the joint optimized index model with the best evaluation results as the sepsis index model for the elderly:

[0015] Single-indicator prediction models were constructed for each candidate inflammatory and immune composite index in the development cohort dataset.

[0016] The single-indicator prediction model is selected based on the AUC and the best Yoden index of each single-indicator prediction model for comprehensive performance evaluation.

[0017] Based on the comprehensive performance evaluation results of the single-indicator prediction model, candidate inflammatory and immune composite indicators whose comprehensive performance of a single indicator meets the preset performance threshold are determined as core inflammatory marker indicators. The core inflammatory marker indicators include at least one candidate inflammatory and immune composite indicator.

[0018] The consistency of the indicators in the optimal joint optimization indicator model based on the core inflammatory marker indicators is verified.

[0019] Furthermore, consistency verification is performed on the indicators in the optimal joint optimization indicator model based on the core inflammatory marker indicators, including:

[0020] The preset basic indicators, the core inflammatory marker indicators, and the preset importance scoring criteria are all superimposed and combined to obtain multiple sets of core inflammatory joint indicator models.

[0021] A comprehensive evaluation was conducted on each group of core inflammation joint indicator models to obtain the optimal core inflammation indicator model;

[0022] The indicators in the core inflammation optimal indicator model are compared with the indicators in the joint optimization indicator model that yields the best current evaluation results.

[0023] If the indicators in the core inflammation optimal indicator model are consistent with the indicators in the joint optimization indicator model with the current evaluation result, then the joint optimization indicator model with the current evaluation result can be determined as the elderly sepsis indicator model.

[0024] Otherwise, a reassessment mechanism is triggered and the above process is repeated until the indicators in the core inflammation optimal indicator model are consistent with the indicators in the joint optimization indicator model that is currently the best in the assessment results.

[0025] Furthermore, the candidate inflammatory and immune composite indicators in the development cohort dataset are ranked by importance for sepsis mortality and prognosis based on a fully relevant feature selection algorithm, including:

[0026] The candidate inflammatory and immune composite indicators in the development queue dataset are used to construct a candidate inflammatory feature set.

[0027] Each column in the candidate inflammation feature set is randomly rearranged to generate shadow candidate features, and the shadow candidate features are used together with the original features to train a random forest.

[0028] The Z-value of each feature in the candidate inflammation feature set is calculated based on the feature importance score output by the random forest.

[0029] Compare the Z-value of the original feature with the maximum Z-value of all shadow features;

[0030] If the absolute value of the difference between the Z value of the original feature and the maximum Z value of all shadow features is greater than a preset threshold, the importance of the original feature is determined to be either important or irrelevant. Specifically, if the original feature is greater than the maximum value of all shadow features, the importance of the original feature is determined to be important; if the original feature is less than the maximum value of all shadow features, the importance of the original feature is determined to be irrelevant.

[0031] If the absolute value of the difference between the Z value of the original feature and the maximum Z value of all shadow features is not greater than a preset threshold, then the importance of the original feature is marked as pending.

[0032] Repeat the above process iteratively to identify candidate inflammatory and immune composite indicators from all development cohort datasets that are marked as important as important inflammatory marker indicators.

[0033] Furthermore, the scoring results of multiple joint optimization index models are compared, and the scoring results are evaluated based on an independent validation cohort dataset, including:

[0034] Calculate the AUC score and Brier score of multiple joint optimization index models respectively;

[0035] Based on the AUC score and Brier score of each set of joint optimization index models, construct the ROC curve, calibration curve and DCA of the set of joint optimization index models;

[0036] The ROC curve, calibration curve, and DCA of each joint optimization index model were evaluated based on the independent validation cohort dataset to obtain the discrimination, calibration, and net DCA benefit evaluation results of each joint optimization index model.

[0037] The discrimination, calibration, and DCA net benefit evaluation results of all joint optimization index models are compared, and the joint optimization index model with the best discrimination, calibration, and DCA net benefit evaluation results is determined as the joint optimization index model with the best evaluation results.

[0038] Furthermore, a development cohort dataset of elderly sepsis patients aged ≥60 years was identified based on the intensive care database, including:

[0039] Obtain routine blood cell counts and biochemical indicators for each elderly sepsis patient aged ≥60 years recorded in the intensive care monitoring database within a preset time period after admission;

[0040] Based on the aforementioned routine blood cell count and biochemical indicators, the following indices were determined: systemic immune inflammation index, platelet-to-albumin ratio, monocyte-to-lymphocyte ratio, neutrophil-to-lymphocyte ratio, systemic inflammatory response index, systemic inflammation index, derived neutrophil-to-lymphocyte ratio, lymphocyte-to-monocyte ratio, neutrophil-to-albumin ratio, neutrophil-to-platelet ratio, platelet-to-leukocyte ratio, and prognostic nutritional index.

[0041] A development cohort dataset was constructed based on the systemic immune inflammatory index, platelet-to-albumin ratio, monocyte-to-lymphocyte ratio, neutrophil-to-lymphocyte ratio, systemic inflammatory response index, systemic inflammatory index, derived neutrophil-to-lymphocyte ratio, lymphocyte-to-monocyte ratio, neutrophil-to-albumin ratio, neutrophil-to-platelet ratio, platelet-to-leukocyte ratio, and prognostic nutritional index.

[0042] As another aspect of the present invention, a method for constructing a 28-day mortality risk prediction model for elderly patients with sepsis is provided, comprising:

[0043] Obtain the indicator data from the elderly sepsis indicator model, wherein the elderly sepsis indicator model is constructed according to the construction method of the elderly sepsis indicator model described above, and the indicator data includes at least age, lactate, and SII;

[0044] The index data in the elderly sepsis index model are modeled using various machine learning algorithms to obtain various elderly sepsis index machine learning models, which include at least Logistic regression model, random forest model, support vector machine model and XGBoost model;

[0045] Cross-validation was performed on various machine learning models for sepsis in the elderly, and the AUC, calibration curves and DCA of various machine learning models for sepsis in the elderly were compared in an independent development cohort dataset to determine the random forest model as the model to be trained and optimized.

[0046] The indicator data in the elderly sepsis index model is used as input data to perform model distillation on the model to be trained and optimized, thereby obtaining a 28-day mortality risk prediction model for elderly sepsis. The 28-day mortality risk prediction model for elderly sepsis can be called to realize the prediction of mortality risk in elderly sepsis over 28 days.

[0047] Furthermore, the indicator data from the elderly sepsis index model are used as input data to perform model distillation on the model to be trained and optimized, including:

[0048] The index data of the elderly sepsis index are input into the random forest model for fitting, and the predicted probability corresponding to the index data of each elderly sepsis index is obtained, and the predicted probability is used as a soft label.

[0049] Using the index data of the elderly sepsis index as independent variables and the soft label as the output target, a Logistic regression model is trained to minimize the difference between the predicted probability of the Logistic regression model and the predicted probability output by the random forest model.

[0050] The trained Logistic regression model was selected as the predictive model for the 28-day mortality risk of elderly patients with sepsis.

[0051] As another aspect of the present invention, a 28-day mortality risk prediction system for elderly patients with sepsis is provided, comprising: a user terminal and a server terminal, wherein the user terminal is communicatively connected to the server terminal, the user terminal includes at least a user interaction module, and the server terminal includes at least an algorithm engine module.

[0052] The user interaction module is used to acquire the predicted sepsis index data of elderly patients input by the user, and to display the 28-day mortality risk prediction results of elderly sepsis patients to the user.

[0053] The algorithm engine module is used to input the data of the sepsis index to be predicted into the 28-day mortality risk prediction model for sepsis in the elderly to predict the mortality risk and obtain the 28-day mortality risk prediction result for sepsis in the elderly; wherein the 28-day mortality risk prediction model for sepsis in the elderly is constructed according to the construction method of the 28-day mortality risk prediction model for sepsis in the elderly described above.

[0054] Furthermore, the client-side also includes a request encapsulation and validation module, and the server-side also includes a model management module and a result encapsulation and return module.

[0055] The model management module is used to store, optimize, and update the 28-day mortality risk prediction model for elderly patients with sepsis.

[0056] The request encapsulation and verification module is used to encapsulate the user-inputted data on the sepsis index of elderly patients to be predicted into a standardized JSON file according to predefined field specifications.

[0057] The result encapsulation and return module is used to encapsulate the 28-day mortality risk prediction results of elderly sepsis patients into a JSON file and return it to the user interaction module.

[0058] The method for constructing an indicator model for elderly sepsis provided in this invention involves identifying a development cohort dataset and an independent validation cohort dataset of elderly sepsis patients aged ≥60 years using an intensive care unit database. Then, based on candidate inflammatory and immune composite indicators from the development cohort dataset, a fully relevant feature algorithm is used to select and rank the importance of these indicators to obtain important inflammatory marker indicators. Multiple sets of jointly optimized indicator models are obtained by combining these important inflammatory marker indicators with preset basic indicators. These jointly optimized indicator models are scored and evaluated to obtain the optimal jointly optimized indicator model. Finally, this optimal jointly optimized indicator model is determined as the indicator model for elderly sepsis, thus completing the construction of the elderly sepsis indicator model. This method for constructing an elderly sepsis indicator model uses a large-sample critical care dataset as the basis for model construction and systematically screens and optimizes candidate inflammatory and immune composite indicators, thereby achieving the goal of obtaining an indicator model applicable to elderly sepsis based on an existing indicator database. Attached Figure Description

[0059] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the following detailed description to explain the invention, but do not constitute a limitation thereof.

[0060] Figure 1 A flowchart illustrating the method for constructing the sepsis index model for the elderly provided by this invention.

[0061] Figure 2 This is a flowchart illustrating the importance ranking of candidate inflammatory and immune composite indicators based on a fully relevant feature selection algorithm provided by the present invention.

[0062] Figure 3 The flowchart provided by this invention describes the process of evaluating multiple sets of joint optimization index models to obtain the optimal joint optimization index model.

[0063] Figure 4 A flowchart illustrating the method for constructing a 28-day mortality risk prediction model for elderly patients with sepsis provided by this invention.

[0064] Figure 5 The structural block diagram of the 28-day mortality risk prediction system for elderly sepsis patients provided by this invention.

[0065] Figure 6 The flowchart of the 28-day mortality risk prediction system for elderly sepsis patients provided by this invention. Detailed Implementation

[0066] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0067] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0068] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of the invention described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0069] This embodiment provides a method for constructing a sepsis index model in the elderly. Figure 1 This is a flowchart of a method for constructing an elderly sepsis index model according to an embodiment of the present invention, such as... Figure 1 As shown, it includes:

[0070] S100. Determine a development cohort dataset and an independent validation cohort dataset for elderly sepsis patients aged ≥60 years based on the intensive care database. The data in the development cohort dataset and the independent validation cohort dataset are obtained from two completely different intensive care databases. The development cohort dataset includes at least candidate inflammatory and immune composite indicators based on nutritional status, coagulation / platelet ratio, immune cell balance, and comprehensive inflammatory burden dimensions.

[0071] In this embodiment of the invention, a development cohort and an independent validation cohort of elderly patients with sepsis aged ≥60 years can be constructed based on a intensive care unit database through rigorous screening. Candidate inflammatory and immune composite indicators are constructed from the development cohort dataset. These candidate inflammatory and immune composite indicators can characterize the inflammatory and immune status of elderly patients from multiple dimensions, including nutritional status (PNI), coagulation / platelet ratio (PWR, PAR, AISI), immune cell balance (NLR, MLR, LMR, dNLR, SIRI, NPAR, NPR), and total inflammatory burden (SII).

[0072] It should be noted that the embodiments of this invention adopt a two-stage construction scheme of development and external validation, which can simultaneously ensure the quality of model construction and cross-population generalization ability. Specifically, the data source for the development cohort dataset can be the MIMIC-IV v2.0 database (after screening, a total of 8,935 cases were included in the study), which is mainly used to complete model development and indicator screening: that is, in elderly sepsis patients aged ≥60 years, the predictive value of various inflammatory markers is systematically calculated and compared, the importance of features is ranked, and on this basis, different algorithm models are trained and compared to determine the indicator model most suitable for elderly sepsis patients.

[0073] To avoid overfitting caused by repeatedly screening variables and tuning parameters within the same database, which can lead to model performance being limited to the training samples, this invention introduces an independent external validation cohort dataset. For example, a cohort of elderly sepsis patients from the Fourth People's Hospital of Zigong City, China (including 345 cases) can be used. The discriminative power of the determined indicator combination and the model is directly tested in independent samples, evaluating its stability and generalizability across different regions, healthcare systems, and population characteristics. This construction scheme, combining the development cohort dataset and the independent validation cohort dataset, not only reduces bias caused by "chance correlations" and improves the reproducibility of conclusions, but also provides crucial methodological basis and safety assurance for the subsequent engineering of the model into an online tool for clinical use.

[0074] S200. Based on the full correlation feature selection algorithm, the candidate inflammatory and immune composite indicators of the development cohort dataset are ranked according to their importance in terms of sepsis mortality and prognosis to obtain important inflammatory marker indicators.

[0075] In this embodiment of the invention, candidate inflammatory and immune composite indicators in the aforementioned development cohort dataset are ranked by importance using a fully correlated feature selection algorithm to obtain important inflammatory marker indicators. Specifically, candidate inflammatory and immune composite indicators can be screened using a correlation method.

[0076] S300. The preset basic indicators are superimposed and combined with each of the important inflammatory marker indicators and the preset importance scoring criteria to obtain multiple sets of joint optimization indicator models, wherein the preset basic indicators include at least age and lactate.

[0077] In this embodiment of the invention, preset basic indicators age and lactate are superimposed and combined with each of the aforementioned important inflammatory marker indicators, and also superimposed and combined with a preset importance scoring standard, thereby obtaining multiple sets of joint optimized indicator models. For example, if the aforementioned important inflammatory marker indicators include 7 indicators, then age and lactate are superimposed and combined with each of these 7 indicators to obtain 7 sets of joint optimized indicator models. In addition, age and lactate are also superimposed and combined with the preset importance scoring standard to obtain one set of joint optimized indicator models. Therefore, after this superposition and combination, a total of 8 sets of joint optimized indicator models are obtained.

[0078] S400. Compare the scoring results of multiple joint optimization index models and evaluate the scoring results based on the independent validation cohort dataset;

[0079] In this embodiment of the invention, the scoring results obtained after scoring the above-mentioned multiple sets of joint optimization index models are compared, and then the scoring results are evaluated based on the independent validation queue dataset.

[0080] It should be understood that the embodiments of the present invention test the model’s generalization ability on new samples that were not used in training by repeatedly evaluating the ROC in completely independent validation queues, thereby reducing the risk of overfitting and improving the reproducibility and clinical usability of the conclusions.

[0081] S500. The joint optimization index model with the best evaluation results is determined as the elderly sepsis index model. The indexes in the elderly sepsis index model include at least age, lactate and SII.

[0082] Based on the above evaluation results, the optimal joint optimization index model was determined to be the elderly sepsis index model. In this embodiment of the invention, the indexes in the elderly sepsis index model include at least age, lactate, and SII.

[0083] Therefore, the method for constructing an indicator model for elderly sepsis provided in this embodiment of the invention determines a development cohort dataset and an independent validation cohort dataset of elderly sepsis patients aged ≥60 years using an intensive care unit database. Then, based on the candidate inflammatory and immune composite indicators in the development cohort dataset, a fully relevant feature algorithm is used to select and rank the importance of these indicators to obtain important inflammatory marker indicators. Multiple sets of joint optimized indicator models are obtained by combining these important inflammatory marker indicators with preset basic indicators. These joint optimized indicator models are scored and evaluated to obtain the optimal joint optimized indicator model. Finally, this optimal joint optimized indicator model is determined as the indicator model for elderly sepsis, thus completing the construction of the elderly sepsis indicator model. This method for constructing an elderly sepsis indicator model uses a large-sample critical care dataset as the basis for model construction and systematically screens and optimizes candidate inflammatory and immune composite indicators, thereby achieving the goal of obtaining an indicator model applicable to elderly sepsis based on an existing indicator database.

[0084] In this embodiment of the invention, a development cohort dataset of elderly sepsis patients aged ≥60 years is determined based on an intensive care database, including:

[0085] 1) Obtain routine blood cell counts and biochemical indicators for each elderly sepsis patient aged ≥60 years recorded in the intensive care monitoring database within a preset time period after admission;

[0086] 2) Based on the routine blood cell count and biochemical indicators, determine the systemic immune inflammation index, platelet-to-albumin ratio, monocyte-to-lymphocyte ratio, neutrophil-to-lymphocyte ratio, systemic inflammatory response index, systemic inflammation index, derived neutrophil-to-lymphocyte ratio, lymphocyte-to-monocyte ratio, neutrophil-to-albumin ratio, neutrophil-to-platelet ratio, platelet-to-leukocyte ratio, and prognostic nutritional index;

[0087] 3) A development cohort dataset was constructed based on the systemic immune inflammatory index, platelet-to-albumin ratio, monocyte-to-lymphocyte ratio, neutrophil-to-lymphocyte ratio, systemic inflammatory response index, systemic inflammatory index, derived neutrophil-to-lymphocyte ratio, lymphocyte-to-monocyte ratio, neutrophil-to-albumin ratio, neutrophil-to-platelet ratio, platelet-to-leukocyte ratio, and prognostic nutritional index.

[0088] Specifically, in this embodiment of the invention, the process of constructing candidate inflammatory and immune composite indicators in the development cohort dataset may include: based on routine blood cell counts and biochemical indicators available in the first 24 hours after admission, constructing the following candidate inflammatory and immune composite indicators in the development cohort (all blood cell counts are 10). 9 / L, albumin is g / L): Systemic Immune-Inflammation Index (SII) = Platelet count × Neutrophil count / Lymphocyte count; Platelet-Albumin Ratio (PAR) = Platelet count / Albumin; Monocyte-Lymphocyte Ratio (MLR) = Monocyte count / Lymphocyte count; Neutrophil-Lymphocyte Ratio (NLR) = Neutrophil count / Lymphocyte count; Systemic Inflammation Response Index (SIRI) = Neutrophil count × Monocyte count / Lymphocyte count; Aggregate Index of Systemic Inflammation (ARI) Inflammation (AISI) = Neutrophil count × Platelet count × Monocyte count / Lymphocyte count; Derived Neutrophil-Lymphocyte Ratio (dNLR) = Neutrophil count / (White blood cell count - Neutrophil count); Lymphocyte-Monocyte Ratio (LMR) = Lymphocyte count / Monocyte count; Neutrophil-to-Albumin Ratio (NPAR) = Neutrophil count / Albumin; Neutrophil-Platelet Ratio (NPR) = Neutrophil count / Platelet count; Platelet-White Blood Cell Ratio (PWR) = Platelet count / White blood cell count; Prognostic Nutritional Index (PNI) = Albumin + 5× lymphocyte count. These indicators characterize the inflammatory-immune status of elderly patients from multiple dimensions, including nutritional status (PNI), coagulation / platelet ratio (PWR, PAR, AISI), immune cell balance (NLR, MLR, LMR, dNLR, SIRI, NPAR, NPR), and total inflammatory burden (SII).

[0089] In this embodiment of the invention, the candidate inflammatory and immune composite indicators of the development cohort dataset are ranked by importance for sepsis mortality and prognosis according to a fully relevant feature selection algorithm, specifically as follows: Figure 2 As shown, it includes:

[0090] S210. Construct a candidate inflammatory feature set from the candidate inflammatory immune composite indicators in the development queue dataset;

[0091] Specifically, the importance of all candidate inflammatory and immune composite indicators and clinical variables was ranked. Using 28-day mortality as the outcome, a feature set was constructed in the development cohort: F = {PNI, PWR, NPR, LMR, PAR, dNLR, SII, NLR, SIRI, AISI, NPAR, SOFA score, age, lactate, ...}, and the Boruta algorithm (with random forest as the base model) was used to rank the feature importance.

[0092] S220. Randomly rearrange each column in the candidate inflammation feature set to generate shadow candidate features, and train a random forest together with the shadow candidate features and their original features.

[0093] Specifically, each column in F is randomly rearranged to generate "shadow features," which are then used to train a random forest together with the original features.

[0094] S230. Calculate the Z value of each feature in the candidate inflammation feature set based on the feature importance score output by the random forest;

[0095] Specifically, the Z-value of each feature is calculated based on the feature importance score output by the random forest. In this embodiment of the invention, the Z-value = (mean of feature importance score - mean of all shadow feature importance scores) / (standard deviation of feature importance score). Simply put, the Z-value is the standardized score of the feature's importance score relative to its own volatility, centered on the mean of the shadow features.

[0096] S240. Compare the Z-value of the original feature with the maximum Z-value of all shadow features;

[0097] Specifically, the Z-value of the true feature (i.e., the original feature) is compared with the maximum Z-value of all shadow features.

[0098] S250. If the absolute value of the difference between the Z value of the original feature and the maximum Z value of all shadow features is greater than a preset threshold, then the importance of the original feature is determined to be marked as important or irrelevant. If the original feature is greater than the maximum value of all shadow features, then the importance of the original feature is determined to be marked as important. If the original feature is less than the maximum value of all shadow features, then the importance of the original feature is determined to be marked as irrelevant.

[0099] S260. If the absolute value of the difference between the Z value of the original feature and the maximum Z value of all shadow features is not greater than a preset threshold, then the importance of the original feature is marked as pending.

[0100] Specifically, when the Z-value of the true feature is significantly higher than that of the shadow feature, it is judged as "important"; when it is significantly lower, it is judged as "irrelevant"; and when it is in between, it is judged as "pending".

[0101] S270. Repeat the above process iteratively to identify the candidate inflammatory immune composite indicators of all development cohort datasets that are marked as important as important inflammatory marker indicators.

[0102] Specifically, through multiple rounds of iteration, stable conclusions regarding the importance of features were obtained. The final results of this invention's embodiments show that in the ≥60-year-old sepsis cohort, the importance of seven composite indicators—PNI, PWR, NPR, LMR, PAR, dNLR, and SII—was significantly higher than the SOFA score (the SOFA score is clinically used to assess sepsis mortality and prognosis), while the importance of AISI, SIRI, NPAR, and NLR was lower than the SOFA score. Furthermore, age and lactate were consistently marked as "important" features in Boruta, also demonstrating higher importance than the SOFA score.

[0103] In this embodiment of the invention, in order to further verify the accuracy of the labeling of the above-mentioned important features, the method for constructing the sepsis index model for the elderly further includes the following step before determining the joint optimized index model with the best evaluation results as the sepsis index model for the elderly:

[0104] 1) Construct single-indicator prediction models for each candidate inflammatory and immune composite index in the development cohort dataset;

[0105] Specifically, in this embodiment of the invention, a separate prediction model is constructed for each candidate inflammatory and immune composite indicator in the aforementioned development cohort dataset. A univariate logistic regression model is constructed using a certain indicator X (such as SII) as the independent variable: logit(p) = α0 + α1 × X, where p represents the 28-day mortality probability.

[0106] 2) Select a single-indicator prediction model based on the AUC and the best Youden index of each single-indicator prediction model to conduct a comprehensive performance evaluation of the single-indicator prediction model;

[0107] Specifically, for each single-indicator prediction model, the AUC (Area Under the ROC Curve), the optimal Youden index, sensitivity, and specificity are calculated, and the ROC (Receiver Operating Characteristic Curve) curve is plotted to facilitate a comprehensive performance evaluation of each candidate inflammatory and immune composite indicator.

[0108] 3) Based on the comprehensive performance evaluation results of the single-index prediction model, candidate inflammatory and immune composite indices whose comprehensive performance of a single index meets the preset performance threshold are determined as core inflammatory marker indices. The core inflammatory marker indices include at least one candidate inflammatory and immune composite indices.

[0109] Specifically, the comprehensive performance evaluation results of all the above single-indicator prediction models are compared with the preset performance threshold to determine the core inflammatory marker that can meet the preset performance threshold. It should be understood that the core inflammatory marker includes at least one candidate inflammatory immune complex indicator.

[0110] 4) Verify the consistency of the indicators in the optimal joint optimization indicator model based on the core inflammatory marker indicators.

[0111] Specifically, consistency verification is performed on the indicators in the optimal joint optimization indicator model based on the core inflammatory marker indicators, including:

[0112] 41) The preset basic indicators, the core inflammatory marker indicators, and the preset importance scoring criteria are all superimposed and combined to obtain multiple sets of core inflammatory joint indicator models;

[0113] Specifically, the core inflammatory markers are combined with clinical variables such as age, lactate, and SOFA to construct a multi-group core inflammatory marker model.

[0114] 42) A comprehensive evaluation was conducted on each group of core inflammation joint indicator models to obtain the optimal core inflammation indicator model;

[0115] Specifically, the AUC, calibration curve, and decision curve analysis (DCA) results of each core inflammation joint indicator model are comprehensively evaluated to obtain the optimal core inflammation indicator model.

[0116] 43) Perform consistency verification between the indicators in the core inflammation optimal indicator model and the indicators in the joint optimization indicator model with the current evaluation result as the best;

[0117] 44) If the indicators in the core inflammation optimal indicator model are consistent with the indicators in the joint optimization indicator model with the best current evaluation result, then the joint optimization indicator model with the best current evaluation result can be determined as the elderly sepsis indicator model.

[0118] 45) Otherwise, trigger the re-evaluation mechanism and repeat the above process until the indicators in the core inflammation optimal indicator model are consistent with the indicators in the joint optimization indicator model that is currently the best in the evaluation results.

[0119] Specifically, the re-evaluation mechanism may include: expanding the candidate set of composite inflammatory indicators or adjusting the candidate combinations according to preset rules, re-entering the composite inflammatory indicators with the top N single-indicator AUCs as candidate core indicators for joint combination evaluation, and re-comparing multiple sets of optimal core inflammatory indicator models without changing the evaluation indicator system (AUC) until the indicators in the optimal core inflammatory indicator model are consistent with the indicators in the current optimal joint optimization indicator model.

[0120] It should be understood that, in this embodiment of the invention, when the indicators in the core inflammation optimal indicator model are consistent with the indicators in the currently optimal joint optimization indicator model, it indicates that the inflammation information obtained by single indicator screening is retained in the joint optimization stage, thereby verifying the consistency and rationality of the currently optimal joint optimization indicator model between the "single indicator level" and the "joint optimization level", making it possible to determine it as the elderly sepsis indicator model; otherwise, if the two are inconsistent, the candidate inflammation indicator set and the core inflammation joint indicator model are re-evaluated by triggering a re-evaluation mechanism to ensure that the finally determined elderly sepsis indicator model has clear decision-making basis, reproducible termination conditions, and stable predictive value and clinical interpretability.

[0121] In this embodiment of the invention, the scoring results of multiple sets of joint optimization index models are compared, and the scoring results are evaluated based on an independent validation queue dataset, such as... Figure 3 As shown, it includes:

[0122] S410. Calculate the AUC score and Brier score of multiple joint optimization index models respectively.

[0123] S420. Construct the ROC curve, calibration curve, and DCA of each set of joint optimization index models based on the AUC score and Brier score.

[0124] S430. Evaluate the ROC curve, calibration curve and DCA of each joint optimization index model based on the independent validation queue dataset, and obtain the discrimination, calibration and net DCA benefit evaluation results of each joint optimization index model.

[0125] S440. Compare the discrimination, calibration, and DCA net benefit evaluation results of all joint optimization index models, and determine the joint optimization index model with the best discrimination, calibration, and DCA net benefit evaluation results as the joint optimization index model with the best evaluation results.

[0126] In this embodiment of the invention, each indicator in the important inflammatory markers and the preset importance scoring criteria are superimposed with age and lactate, respectively, to screen and obtain the optimal combined optimization indicators. Specifically, based on the Boruta analysis mentioned above, "age" and "lactate" also have high importance. In this embodiment of the invention, age and lactate are further combined with the above seven inflammatory markers and SOFA scores to construct a series of combined models, such as: age + lactate + SII: logit(p) = β0 + β1 × age + β2 × lactate + β3 × SII or age + lactate + PNI, age + lactate + PAR, etc. By comparing the AUC and Brier scores of each model in the development cohort, and evaluating their ROC curves, calibration curves, and DCA (Decision Curve Analysis) in the independent validation cohort, the discriminative ability and clinical net benefit of different combinations are comprehensively judged. The final results showed that the combination of "age + lactate + SII" was superior to combinations such as "age + lactate + other inflammatory markers" and "age + lactate + SOFA" in terms of discrimination, calibration and net DCA benefit. Therefore, it was identified as the core combined indicator that best fits the sepsis risk phenotype of the elderly.

[0127] Therefore, the method for constructing an elderly sepsis index model provided by this invention involves constructing a development cohort dataset from an intensive care unit database, determining candidate inflammatory and immune composite indicators, sorting the candidate inflammatory and immune composite indicators using the Boruta algorithm to obtain important inflammatory marker indicators, and then performing individual modeling and comparison, as well as constructing a joint model by superimposing age and lactate. This process ultimately forms a disease-specific optimized prediction model with "age + lactate + SII" as its core, obtaining an elderly sepsis index model applicable to elderly sepsis, and effectively solving the problem of "lack of a systematic method for the selection and combination of elderly-specific inflammatory indicators" in the prior art.

[0128] As another embodiment of the present invention, a method for constructing a 28-day mortality risk prediction model for elderly patients with sepsis is provided, wherein, as Figure 4 As shown, it includes:

[0129] S10. Obtain the indicator data in the elderly sepsis indicator model, wherein the elderly sepsis indicator model is constructed according to the construction method of the elderly sepsis indicator model described above, and the indicator data includes at least age, lactate and SII.

[0130] S20. The index data in the elderly sepsis index model are modeled using multiple machine learning algorithms to obtain multiple elderly sepsis index machine learning models. The multiple elderly sepsis index machine learning models include at least Logistic regression model, random forest model, support vector machine model and XGBoost model.

[0131] S30. Cross-validate multiple machine learning models for sepsis indicators in the elderly, and compare the AUC, calibration curves and DCA of multiple machine learning models for sepsis indicators in the independent development cohort dataset to determine the random forest model as the model to be trained and optimized.

[0132] S40. Using the index data in the elderly sepsis index model as input data, perform model distillation on the model to be trained and optimized to obtain a 28-day mortality risk prediction model for elderly sepsis. The 28-day mortality risk prediction model for elderly sepsis can be called to realize the prediction of mortality risk for elderly sepsis in 28 days.

[0133] After determining "age, lactate, and SII" as the core input features, this invention employs various machine learning algorithms, including Logistic Regression, Random Forest, Support Vector Machine (SVM), and XGBoost, to train models on a cohort of elderly sepsis patients aged ≥60 years. Performance is then compared in an independent validation cohort to select the optimal model and perform model distillation. The implementation methods of each algorithm are briefly described below:

[0134] (1) Logistic Regression Model: Logistic regression is used to construct a linear relationship between 28-day mortality (binary) and continuous independent variables. Using "age, lactate, and SII" as independent variables, a linear prediction is established:

[0135] logit(p) = β0 + β1 × age + β2 × lactate + β3 × SII,

[0136] p = 1 / (1 + exp(−logit(p))),

[0137] Where p represents the 28-day mortality probability, and the parameters β0 to β3 are obtained through maximum likelihood estimation. In this invention, Logistic regression is first used as the baseline interpretable model, and then a model distillation strategy is employed based on random forest to finally obtain the explicit formula deployed in the online tool:

[0138] logit(p_RF_SII) = −5.894 + 0.056 × age + 0.447 × lactate + 0.0001 × SII,

[0139] The 28-day mortality rate is p = 1 / (1 + exp(−logit(p_RF_SII))).

[0140] (2) Random Forest Model

[0141] Random forests belong to the ensemble learning method, which improves the robustness and generalization ability of a model by integrating multiple decision trees. In this embodiment of the invention, multiple training subsets are constructed using bootstrap sampling with replacement, and a binary decision tree is trained on each subset. At each node split, a subset of candidate features are randomly selected from all features, and the purity improvement of different split points is evaluated using indicators such as the Gini index, and the optimal split is selected. Finally, the average prediction of all trees is used as the death probability of the sample. The comparison results show that random forests perform best overall on indicators such as AUC and decision curve (DCA), and are therefore selected as the "teacher model," with their predicted probabilities used for subsequent model distillation.

[0142] (3) Support Vector Machine (SVM) model

[0143] Support Vector Machines (SVMs) are used to handle binary classification problems that may be nonlinearly separable. This invention employs a C-SVC model with a Radial Basis Function (RBF). By introducing the kernel function, the input features are mapped to a high-dimensional space, where the maximum margin hyperplane is found. A penalty coefficient C controls the balance between margin size and misclassification. After training, the decision value for each sample is obtained and converted into a 28-day mortality probability using methods such as Platt scaling. This probability is then used for unified comparison with other models on ROC, calibration curves, and DCA.

[0144] (4) XGBoost model

[0145] XGBoost is an efficient implementation of Gradient Boosting Decision Tree, which fits the negative gradient of the loss function by iteratively stacking multiple CART trees. This embodiment uses log loss as the objective function, fitting the residual from the previous iteration in each iteration, and introducing regularization terms to reduce tree structure complexity (such as L2 penalties for the number of leaf nodes and leaf node weights) to prevent overfitting. XGBoost calculates the gain of each candidate split using a second-order Taylor expansion of the loss function, utilizing first-order gradient and second-order Hessian information, and selects the feature and split point with the largest gain. The scores from all trees are summed and converted into a 28-day mortality probability using the Sigmoid function.

[0146] Through systematic comparison of the above four models in the development cohort and independent validation cohort of elderly patients with sepsis aged ≥60 years, this invention determines that the random forest model has the best prediction performance. Using its prediction probability as a "soft label", a simple and well-defined logistic regression formula is obtained by distillation on the same feature set (age, lactate, SII). This achieves a final prediction model with "performance close to random forest + highly interpretable form", providing a clear prediction foundation for subsequent encapsulation into algorithm microservices and deployment in hospital AI matrix platforms.

[0147] In this embodiment of the invention, the indicator data in the elderly sepsis index model is used as input data to perform model distillation on the model to be trained and optimized, including:

[0148] 1) Input the index data of the elderly sepsis index into the random forest model for fitting, obtain the predicted probability corresponding to the index data of each elderly sepsis index, and use the predicted probability as a soft label;

[0149] Specifically, 1) a random forest model is trained on the development queue and used as a teacher model; the input feature data of the elderly sepsis index model is input into the random forest model to obtain the prediction probability p_RF corresponding to each sample, and the prediction probability p_RF is used as a soft label.

[0150] 2) Train a Logistic regression model using the index data in the elderly sepsis index as independent variables and the soft label as the output target, so as to minimize the difference between the predicted probability of the Logistic regression model and the predicted probability output by the random forest model.

[0151] Specifically, 2) Using the input feature data as independent variables and the soft label p_RF as the target output, train a Logistic Regression Student Model so that its prediction probability p_LR approximates the prediction probability p_RF of the Random Forest Model; wherein, the distilled regression coefficients are obtained by minimizing the cross-entropy loss function between p_RF and p_LR, and the cross-entropy loss function can be expressed as:

[0152] L = − Σ_i [ p_RF,i × log(p_LR,i) + (1 − p_RF,i) × log(1 − p_LR,i) ];

[0153] 3) The trained Logistic regression model was selected as the prediction model for the 28-day mortality risk of elderly patients with sepsis;

[0154] Specifically, after training, a distilled Logistic regression model is obtained and deployed as an interpretable online model. This allows the model to maintain predictive performance close to that of a random forest model while also outputting explicit linear coefficients and calculation formulas, facilitating clinical interpretation and engineering implementation.

[0155] It should be understood that multi-algorithm modeling is performed based on the optimal combination of indicators, and the final interpretable formula is obtained through model distillation. Specifically, using the three variables "age, lactate, and SII" as input features, various machine learning models such as Logistic Regression, Random Forest, Support Vector Machine (SVM), and XGBoost are constructed. Hyperparameters are optimized through 5-fold cross-validation and grid search, and the AUC, calibration curves, and DCA of each model are compared in an independent validation queue. The comparison results show that the Random Forest model performs best overall in terms of discrimination and clinical net benefit, but its internal structure consists of multiple decision trees, resulting in relatively poor interpretability, which is not conducive to intuitively explaining the contribution of each variable to risk to doctors in clinical scenarios. To balance predictive performance and interpretability, this embodiment of the invention adopts a model distillation strategy: using the predicted probability of the Random Forest model in the development queue as a "soft label," and still using the three variables "age, lactate, and SII" as input, a simple Logistic Regression model is fitted from top to bottom, so that the Logistic Regression approximates the output of the Random Forest as closely as possible. By minimizing the cross-entropy loss between the random forest prediction probability and the logistic regression prediction probability, the distilled regression coefficients are obtained, forming the explicit formula for the final deployment in the online tool:

[0156] logit(p_RF_SII) = −5.894 + 0.056 × age + 0.447 × lactate + 0.0001 ×SII,

[0157] Here, p_RF_SII represents the 28-day mortality risk score given by the distillation model, and the corresponding 28-day mortality rate is: 28-day mortality rate p = 1 / (1 + exp(−logit(p_RF_SII))). In the independent validation cohort, the Logistic Regression model obtained by distillation performs very close to the original Random Forest model in terms of AUC, calibration curve, and DCA. It also provides clear linear coefficients, making it easier for doctors to understand the direction and weight of "age, lactate, and SII" on mortality risk, and also facilitating efficient implementation and maintenance in the form of fixed formulas in the algorithm microservice.

[0158] In summary, the method for constructing a 28-day mortality risk prediction model for elderly sepsis patients provided by this invention addresses the problem of the lack of disease-specific models for predicting the 28-day mortality risk of elderly sepsis patients and the difficulty in engineering applications. First, based on a large intensive care database, a development cohort of elderly sepsis patients (≥60 years old) is rigorously screened and constructed, and an independent validation cohort is constructed using a local database. On this basis, Boruta feature selection combined with multiple machine learning algorithms (such as Logistic Regression, Random Forest, SVM, XGBoost) is used to systematically screen and optimize the 28-day mortality risk prediction model. This not only enables the acquisition of indicator models applicable to elderly sepsis based on existing indicator databases, but also enables the construction of a 28-day mortality risk prediction model for elderly sepsis patients based on these models. This effectively fills the technical gap in the existing technology of lacking a disease-specific cohort of elderly sepsis patients based on a large sample and constructing and validating a prediction model using systematic feature selection and multi-model comparison.

[0159] As another embodiment of the present invention, a 28-day mortality risk prediction system 10 for elderly patients with sepsis is provided, wherein, as Figure 5 As shown, it includes: a user terminal 100 and a server terminal 200, wherein the user terminal 100 is communicatively connected to the server terminal 200, the user terminal 100 includes at least a user interaction module 110, and the server terminal 200 includes at least an algorithm engine module 210.

[0160] User interaction module 110 is used to obtain the data of sepsis indicators to be predicted for elderly patients input by the user, and to display the 28-day mortality risk prediction results of sepsis in elderly patients to the user.

[0161] The algorithm engine module 210 is used to input the data of the sepsis index to be predicted into the 28-day mortality risk prediction model for sepsis in the elderly to predict the mortality risk and obtain the 28-day mortality risk prediction result for sepsis in the elderly; wherein the 28-day mortality risk prediction model for sepsis in the elderly is constructed according to the construction method of the 28-day mortality risk prediction model for sepsis in the elderly described above.

[0162] The present invention provides a 28-day mortality risk prediction system for elderly patients with sepsis. In its algorithm engine module, the system uses the previously described method for constructing a 28-day mortality risk prediction model for elderly patients with sepsis to predict mortality risk based on the target elderly sepsis indicator data, thus obtaining the 28-day mortality risk prediction result. This system encapsulates the 28-day mortality risk prediction model for elderly patients, constructed using the aforementioned method, into an algorithm microservice that follows a unified REST interface specification. In a clinical environment, doctors only need to input a few core indicators (such as age, lactate, and inflammation composite indicators) to obtain the patient's 28-day mortality probability, risk stratification results, and corresponding explanatory prompts with a single click on the same interface. This provides integrated clinical decision support for the early identification and individualized intervention of elderly patients with sepsis.

[0163] In this embodiment of the invention, the user terminal 100 further includes a request encapsulation and verification module 120, and the server terminal 200 further includes a model management module 220 and a result encapsulation and return module 230.

[0164] Model management module 220 is used to store, optimize and update the 28-day mortality risk prediction model for elderly sepsis patients;

[0165] The request encapsulation and verification module 120 is used to encapsulate the user-inputted data of sepsis indicators in the elderly into a standardized JSON file according to the predefined field specifications.

[0166] The result encapsulation and return module 230 is used to encapsulate the 28-day mortality risk prediction result of elderly sepsis into a JSON file and return it to the user interaction module 110.

[0167] In this embodiment of the invention, the user interaction module 110 can be specifically deployed on the front end of the "AI Matrix Library" of the hospital's AI matrix platform to render the "28-day mortality risk prediction for elderly sepsis patients" algorithm card and pop-up prediction interface. This user interaction module 110 provides a structured data input form, primarily supporting the input of core variables (such as age, lactate, and inflammatory composite indicators) for elderly sepsis patients aged ≥60 years. It also performs basic validation of required fields, numerical ranges, and data types locally to prevent invalid or abnormal requests from reaching the backend. After obtaining the prediction results, the user interaction module 110 displays the model output to clinicians using various visualization methods, including numbers (probability values, Logit values), color coding (low / medium / high risk), risk labels, and concise explanatory text, achieving a unified interactive experience of "completing input and viewing results on a single page."

[0168] In this embodiment of the invention, the request encapsulation and verification module 120 is located between the front end and the platform gateway, and is responsible for encapsulating the input parameters collected by the user interaction module 110 into a unified and standardized JSON request message. Firstly, on the front end, it organizes data according to predefined field naming and structure (including key patient variables, model identifiers, request serial numbers, etc.). Secondly, on the gateway side, it performs integrity and compliance checks on the message, verifying whether the parameter types, value ranges, and required fields meet the interface specifications, ensuring that every request entering the algorithm engine module originates from the ≥60-year-old sepsis prediction scenario and conforms to the expected format, thus providing a guarantee for the stability and security of subsequent model inference.

[0169] Specifically, the server also includes a platform gateway module 240 and a platform management module 250. The platform gateway module 240 serves as the unified entry point (API Gateway) for this system to connect to the hospital's AI matrix platform, responsible for routing prediction requests initiated from the front end to the corresponding "28-day mortality risk prediction for elderly sepsis" algorithm microservice. The platform gateway module 240 works in conjunction with the platform management module 250 to complete user login authentication, permission verification, and access control; it also implements functions such as request rate limiting, load balancing, exception handling, and unified error code return to prevent a single model service failure from affecting the overall stability of the platform. By parsing the model identifier carried in the request path, the platform gateway module 240 can support the parallel access of multiple prediction models, reserving routing capabilities for future expansion of other sepsis-related models while prioritizing support for the disease-specific model of this invention.

[0170] Specifically, the algorithm engine module 210 is the core service of this invention (Prediction Engine Service), responsible for loading and running the optimal machine learning model and its configuration trained on a large sample of elderly sepsis patients aged ≥60 years, using a development cohort and an independent validation cohort. The algorithm engine module 210 performs preprocessing operations such as standardization, feature reconstruction, and missing data handling on the input data verified by the request encapsulation and verification module 120. It then calls the target model from encapsulated models such as Logistic Regression, Random Forest, and XGBoost to complete the inference calculation of the 28-day mortality risk, outputting the corresponding Logit value and mortality probability. Simultaneously, the algorithm engine module 210 stratifies the probability results into risk levels (e.g., low, medium, high) based on the risk thresholds pre-configured in the model management module and generates brief explanatory information to reflect the main contributions of key variables (e.g., lactate, inflammation composite indicators) to the prediction results.

[0171] The results encapsulation and return module 230 is responsible for uniformly structuring the raw calculation results output by the algorithm engine module 210 and constructing a standard JSON response message. This message includes not only the Logit value, 28-day mortality probability, and risk level, but also metadata such as the model version number, prediction timestamp, and model identifier. This facilitates clear display on the front end and allows the platform management module to conduct subsequent auditing and performance review analysis. Through the unified encapsulation design of the results encapsulation and return module 230, the interaction interface between the front end and different model versions remains stable, supporting model upgrades in the background without affecting clinical use.

[0172] The platform management module 250 focuses on the management and maintenance of system operation, and is a crucial supporting component for transforming this invention from a "scientific research model" into a "clinical-grade tool." The platform management module 250 provides unified user authentication, access control, and department / role management functions, ensuring that only qualified medical personnel can access the geriatric sepsis risk prediction tool. The platform management module 250 records user information, call time, model version, response time, and basic result summaries for each call, and collects system operation logs and abnormal events, enabling real-time monitoring and alerting of the algorithm's service status. Through these mechanisms, the platform management module 250 supports a quantitative assessment of the model's long-term operational security, stability, and usage frequency, providing a basis for subsequent model retraining and strategy optimization.

[0173] The model management module 220 is used to centrally manage various prediction models connected to the AI ​​matrix platform, focusing on ensuring the safe iteration and scalability of the 28-day mortality risk prediction model for elderly sepsis patients of this invention throughout its entire lifecycle. The model management module 220 uniformly maintains model files, feature configurations, standardized parameters, and risk threshold configurations, supporting model version control, hot updates, and canary releases, allowing for a smooth switch to a new version model or the introduction of new inflammatory composite indicator combinations without interrupting clinical services. By providing a standardized model loading interface to the algorithm engine module, the model management module 220 ensures the stable operation of the disease-specific model of this invention while reserving expansion space for subsequently incorporating other sepsis or severe illness-related models into the same platform to form a multi-model collaborative decision support system.

[0174] The following is a detailed description of the relationship and workflow of the various modules in the 28-day mortality risk prediction system for elderly sepsis patients according to an embodiment of the present invention.

[0175] (1) User input and request initiation.

[0176] At this stage, clinical medical staff use the user interaction module 110 to select the "28-Day Mortality Risk Prediction for Elderly Patients with Sepsis" card in the front-end interface of the AI ​​Matrix Platform's "AI Matrix Library," and enter core variables such as the patient's age, lactate levels, and inflammatory composite indicators (e.g., SII) in the pop-up form. After completing basic validation of the required local fields and numerical ranges, the user interaction module 110 transmits the qualified data to the request encapsulation and validation module 120. The request encapsulation and validation module 120 is responsible for encapsulating these input parameters into standardized JSON messages according to predefined field specifications and performing integrity and format validation. After successful validation, the request is forwarded to the platform gateway module 240 for subsequent processing.

[0177] (2) Gateway routing and authentication.

[0178] After receiving the prediction request from the request encapsulation and verification module 120, the platform gateway module 240 first collaborates with the platform management module 250 to complete user identity authentication and permission verification based on the doctor's login information and access token, ensuring that the access scenario is a legitimate in-hospital geriatric sepsis risk assessment operation. After successful authentication, the platform gateway module 240 accurately forwards the request to the corresponding "28-day mortality risk prediction for geriatric sepsis" algorithm engine module 210 according to internal routing rules. Simultaneously, the platform gateway module 240 reports key information about this request (such as call time, request identifier, target model ID, response time, etc.) to the platform management module 250, providing a data foundation for system usage statistics, performance monitoring, and audit trails.

[0179] (3) Model reasoning and result generation.

[0180] Upon receiving the authentication request from the platform gateway module 240, the algorithm engine module 210 first queries the model management module 220 and loads the optimal model version and its feature configuration currently effective for the elderly sepsis population (this model originates from a large-sample development queue and an independent validation queue, and is optimized through Boruta feature selection and comparison with multiple models such as Logistic Regression, Random Forest, SVM, and XGBoost). Subsequently, the algorithm engine module 210 performs the same standardization processing and feature construction on the input data as during the training phase, calls the target model for inference calculations, obtains the patient's Logit value and 28-day mortality probability, and automatically classifies the patient into low-risk, intermediate-risk, or high-risk levels based on the risk threshold preset in the model management module 220. Simultaneously, the algorithm engine module 210 generates a brief explanatory text based on the contribution of each input variable to the prediction results. The above calculation results are transmitted to the result encapsulation and return module 230, which encapsulates them into a structured JSON response, attaching metadata such as the model version number, timestamp, and prediction serial number.

[0181] (4) Results return and front-end display.

[0182] After receiving the structured response data from the result encapsulation and return module 230, the platform gateway module 240 processes the response status code and exception information (such as model unavailable, parameter errors, etc.) in a unified manner, forms a standardized HTTP response, and returns it to the front end. The user interaction module 110 parses the response, extracts key fields such as Logit value, 28-day mortality probability, risk level, and explanatory text, and clearly presents the mortality risk level of elderly sepsis patients and a brief explanation in the prediction modality box through various visualization methods such as numerical display, color coding (such as green / yellow / red), icons, and labels. This provides intuitive support for clinicians to make rapid risk identification and intervention decisions at the bedside.

[0183] (5) Platform operation and maintenance and continuous optimization.

[0184] During long-term system operation, the platform management module 250 continuously receives call logs and performance metrics (including call frequency, response time, error rate, user distribution, etc.) reported by the platform gateway module 240 and the algorithm engine module 210. This data supports security auditing, operational statistics, and troubleshooting, and can trigger alarms as needed. The model management and extension module 300, based on the continuously accumulating call data and clinical feedback on elderly sepsis prediction, regularly conducts model evaluation and retraining to create new model versions. These new versions are then smoothly deployed through version control and canary release mechanisms, enabling continuous model optimization and secure iteration. The entire process does not require interruption of the online services of the platform gateway module 240 and the algorithm engine module 210, thus ensuring the high availability and sustainable improvement capabilities of the 28-day mortality risk prediction tool for elderly sepsis in a clinical environment.

[0185] The following is a flowchart ( Figure 6 The method of the present invention can be divided into the following steps in chronological order (taking a single prediction as an example):

[0186] S1: Doctors log in to the AI ​​Matrix Platform. Doctors use their hospital accounts to access the AI ​​Matrix Platform through a unified login portal. The platform completes user authentication and loads their role and permission information to ensure the legitimacy and security of subsequent access to the "28-Day Mortality Risk Prediction for Elderly Patients with Sepsis" tool.

[0187] S2: Select the "28-Day Mortality Risk Prediction for Elderly Patients with Sepsis" algorithm card. On the platform's "AI Matrix Library" page, the doctor browses and identifies multiple algorithm cards, then clicks on the "28-Day Mortality Risk Prediction for Elderly Patients with Sepsis" card corresponding to this invention. This operation triggers the system to pop up a dedicated prediction modal box, preparing for the subsequent input of key information about the elderly sepsis patient.

[0188] S3: Input or Load Sample Data. In the prediction modality box, the doctor inputs the target patient's age, lactate level, SII (Sepsis Intensity II), and other necessary clinical variables. For easy demonstration and training, the system provides a "One-Click Load Sample Data" button, which automatically fills in preset sample values ​​to help users familiarize themselves with the use of the geriatric sepsis-specific model.

[0189] S4: The front-end performs basic validation and encapsulates the request. Before submitting data, the user interaction module performs basic validation on the completeness, numerical range, and data type of each input item, especially verifying whether the age meets the definition for elderly sepsis research. After the validation passes, the user interaction module 110 combines core indicators such as age, lactate, and SII with metadata such as model identifier, user information, and request time into a structured JSON request body, which is then handed over to the request encapsulation and validation module for further processing.

[0190] S5: The platform gateway receives and authenticates the request. The frontend sends a POST request to the platform gateway 600 via HTTPS / HTTP protocol. After receiving the request, the platform gateway module 240 verifies whether the access is qualified to call the geriatric sepsis risk prediction service based on the user token and permission configuration; after successful authentication, the gateway records a summary log of the request and forwards the request to the corresponding algorithm microservice address (i.e., the geriatric sepsis 28-day mortality risk prediction model service of this invention) according to the model identifier and internal routing rules.

[0191] S6: The algorithm microservice parses requests and preprocesses data. The algorithm engine module 210 receives requests from the gateway, parses the input variables in the JSON structure, and performs normalization or standardization preprocessing on variables such as age, lactate, and SII based on statistics obtained in the elderly sepsis development queue during the model training phase (such as mean, standard deviation, and missing value handling strategies), ensuring that the input distribution in the inference phase is consistent with that in the training phase.

[0192] S7: The model calculates the 28-day mortality risk. With the support of the model management module 220, the algorithm engine module 210 loads the optimal predictive model parameter file (e.g., Logistic regression or other optimal-performing machine learning models) optimized for elderly sepsis patients through a large-sample development cohort and independent validation cohort. Based on the preprocessed input data, the patient's Logit value or equivalent risk score is calculated. Subsequently, the Logit value is converted into a 28-day mortality probability through the Sigmoid function or the model's own output layer, and the risk is classified into low, medium, and high levels according to the probability threshold preset in the model management module 220.

[0193] S8: The algorithm engine generates concise risk explanation prompts (such as "significantly elevated lactate" and "high inflammatory burden" as possible major risk factors) based on the contribution of each input variable (e.g., age, lactate, SII, etc.) to the final prediction result, combined with a predefined text template. Subsequently, the system encapsulates the Logit value, 28-day mortality probability, risk level, explanatory text, model version number, prediction timestamp, and other information into a structured JSON response object, which is output by the result encapsulation and return module 230.

[0194] S9: The algorithm microservice returns the encapsulated JSON response to the platform gateway module 240, which then forwards it to the user terminal. The user interaction module 110 clearly displays the results within the prediction modality box using numbers, color coding, and labels, including: Logit value, 28-day mortality rate formatted as a percentage, risk levels differentiated by color, and corresponding text explanations and brief clinical prompts, helping doctors quickly complete risk assessments for elderly sepsis patients.

[0195] S10: After the prediction process is completed, the platform management module 250 records a summary of the prediction call, including the call time, calling user, model version, response time, and risk level results. Throughout this process, strict adherence to hospital and relevant regulations regarding privacy protection is maintained, and no raw data that could identify the patient is recorded. The aforementioned logs and statistical information will be used for subsequent algorithm usage frequency statistics, system performance analysis, and model retraining decisions, providing data support for the continuous optimization of the 28-day mortality risk prediction tool for elderly sepsis patients.

[0196] Based on the structure and prediction process of the 28-day mortality risk prediction system for elderly patients with sepsis described above, this system was applied to predict the 28-day mortality risk of an elderly patient with sepsis. The specific process is as follows:

[0197] First, hospital doctors log in to the AI ​​matrix platform using their unified hospital account. The platform management module 250 completes identity authentication and loads their department and access permission information to ensure legal and secure access. Then, on the "AI Matrix Library" page, the doctor selects the "28-Day Mortality Risk Prediction for Elderly Patients with Sepsis" algorithm card. The system automatically pops up the corresponding prediction modality, leading to a dedicated elderly sepsis risk assessment interface. In the user interface, the doctor enters key indicators collected within the first 24 hours after admission, including age, lactate, and the Systemic Immune-Inflammation Index (SII). SII is calculated from platelet count, neutrophil count, and lymphocyte count using the formula SII = (platelets × neutrophils) / lymphocytes (all units are 10). 9The value was calculated using the formula ( / L). In this embodiment, the patient's age was 67 years old (meeting the criteria for elderly people ≥60 years old), lactate was 0.90 mmol / L, and SII was 2612. After verifying the completeness and numerical range of the input information, the user interaction module 110 on the user side encapsulates the age, lactate, SII, model identifier, user information, and other fields into a standardized JSON request message. After authentication and routing by the platform gateway module 240, the message is forwarded to the "28-day mortality risk prediction for elderly sepsis" algorithm microservice. The algorithm microservice parses the request content and performs preprocessing operations such as standardization on each variable according to the statistics obtained in the large-sample elderly sepsis development queue during the model training phase. In the algorithm engine module 210, the system calls the pre-stored Logistic regression optimal model, which is based on the large-sample elderly sepsis development queue and optimized by random forest distillation, i.e., the 28-day mortality risk prediction model for elderly sepsis, to predict the 28-day mortality risk of the patient. The prediction formula used is: logit(p_RF_SII) = -5.894 + 0.056×age + 0.447×lactic acid + 0.0001×SII, where the 28-day mortality rate p = 1 / (1 + exp(-logit(p_RF_SII))). Substituting the patient's age, lactate, and SII values ​​into the model formula, the 28-day mortality probability of this patient can be calculated to be approximately 0.149. Based on the probability threshold rules preset in the model management module 220, the platform automatically classifies the patient into the corresponding risk level (e.g., intermediate risk) and, combined with the contribution of each variable to the logit, generates a brief risk explanation text and clinical prompts (such as indicating the impact of lactate and inflammation levels on the prediction results). The result encapsulation and return module 230 encapsulates the logit value, 28-day mortality rate, risk level, explanation text, model version number, prediction timestamp, and other information into a unified JSON response message, which is returned to the user terminal via the platform gateway module 240. The user interface displays the results intuitively within the prediction modal frame using numerical values, color coding, and labels. For example, it shows "28-day mortality rate approximately 14.9%" along with the corresponding risk stratification label, providing doctors with a reference for clinical decisions (such as strengthening monitoring or adjusting treatment strategies). Simultaneously, the platform management module 250 records statistical and log information such as the time of the prediction call, the calling user, the model version number, and the response time, but does not store any raw data that could identify the patient, complying with privacy and compliance requirements. Post-hospital follow-up results showed that the patient did not die on day 28 of hospitalization. These prediction results, along with the follow-up outcome, will be incorporated into the ongoing evaluation and optimization of the model's performance, providing real-world data support for the retraining and upgrading of the 28-day mortality risk prediction model for elderly sepsis patients.

[0198] Therefore, this invention provides a 28-day mortality risk prediction system for elderly sepsis patients, addressing the lack of disease-specific models and difficulties in engineering application for 28-day mortality risk prediction in elderly sepsis patients. It offers an online prediction system and method for 28-day mortality risk in elderly sepsis patients based on an AI matrix platform. First, based on a large intensive care database, a development cohort of elderly sepsis patients (≥60 years old) is rigorously selected and constructed. An independent validation cohort is then built using a local database. Based on this, Boruta feature selection combined with multiple machine learning algorithms (such as Logistic Regression, Random Forest, SVM, and XGBoost) is used to systematically screen and optimize the 28-day mortality risk prediction model. Subsequently, the validated optimal model is encapsulated as an algorithm microservice conforming to a unified REST interface specification and connected to the hospital's AI matrix platform through a platform gateway. In a clinical environment, doctors only need to input a few core indicators (such as age, lactate, and inflammation composite indicators) to obtain the patient's 28-day mortality probability, risk stratification results, and corresponding explanatory prompts with a single click on the same interface, thus providing integrated clinical decision support for the early identification and individualized intervention of elderly sepsis patients.

[0199] This invention encapsulates the optimal prediction model, derived from a large-sample cohort of elderly patients with sepsis, into an algorithm microservice that adheres to a unified REST interface and JSON message specification. This microservice is then seamlessly integrated into the hospital's AI matrix platform via a platform gateway module. In the clinical environment, doctors simply log in to the platform, select the "28-day mortality risk prediction for elderly sepsis patients" algorithm card, and input a few core indicators to obtain mortality probability and risk stratification with a single click, without requiring any programming skills or local script execution. This design changes the current situation where machine learning models have long remained confined to R / Python scripts and academic papers, establishing a complete engineering path from "large-sample modeling - model validation - microservice encapsulation - platform integration - clinical use," effectively solving the core pain point of difficulty in deploying models to the front lines of clinical practice.

[0200] This invention employs modern front-end technologies such as Vue3, Vite, and TS on the AI ​​matrix platform front-end, providing a specialized tool for geriatric sepsis in a unified format of "algorithm cards + prediction modal boxes," maintaining a consistent interactive experience with other AI tools on the platform. The system supports "one-click loading of example data," facilitating teaching, training, and rapid demonstrations. In terms of results display, the front-end simultaneously presents Logit values, 28-day mortality rates formatted as percentages, risk levels differentiated by color coding, and explanatory text generated based on the contributions of key variables. This overcomes the shortcomings of traditional tools that only provide single scores or probabilities and lack interpretability, enabling physicians to obtain intuitive and interpretable risk information within seconds, significantly improving the efficiency of clinical decision support and trust in the model's output. By defining a unified REST interface specification and JSON message structure, the 28-day mortality risk prediction model for geriatric sepsis can be smoothly migrated between development, testing, and production deployment environments, and is decoupled from platform gateways, platform management, and other modules. This standardized microservice design not only facilitates the rapid deployment and upgrade of the model in this invention, but also allows the platform to quickly integrate other sepsis-related algorithms or open-source models under the same specifications, avoiding situations where various tools operate independently and have incompatible interfaces. As a result, the platform can gradually build a sepsis AI-assisted decision-making system covering different disease stages (early identification, prognostic assessment, treatment response prediction, etc.), in which this invention provides replicable technical templates and interface specifications.

[0201] This invention fully leverages the unified authentication, permission management, logging, and monitoring statistics functions provided by an AI matrix platform to securely manage and track the entire lifecycle of a tool for predicting the risk of sepsis in the elderly. The platform management module records user information, call time, model version, and response time for each call, meeting auditing and quality management requirements without storing any identifiable patient data. The model management and extension module provides version control, hot updates, and canary release capabilities, enabling models to be upgraded and rolled back without service interruption. Unlike existing decentralized tools that are "maintained independently and difficult to update," this invention, through a platform-level operation and maintenance system, significantly reduces operation and maintenance costs and update risks, ensuring stable operation and continuous optimization of the system in real clinical environments.

[0202] This invention focuses on elderly patients with sepsis, completing a seamless process from large-sample retrospective studies, Boruta feature screening and multi-model comparison, independent validation cohort evaluation, to algorithm microservice encapsulation, integration with an AI matrix platform, and online clinical use. This forms a closed loop of "development cohort - independent validation cohort - algorithm microservice - platform-level deployment - real-world feedback - model re-optimization." Compared to traditional methods relying on manual literature review and scoring, this system significantly improves the efficiency of medical staff, reduces the risk of human error, and enhances the early identification and intervention capabilities for elderly patients with sepsis through more accurate risk prediction. It substantially shortens the translation link from "scientific discovery to clinical application," providing a replicable path for rapidly applying more research findings in the field of geriatric critical care to clinical practice.

[0203] It is understood that the above embodiments are merely exemplary implementations used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A method for constructing an index model of sepsis in the elderly, characterized in that, include: A development cohort dataset and an independent validation cohort dataset were determined based on an intensive care unit (ICU) database for elderly patients with sepsis aged ≥60 years. The data in the development cohort dataset and the independent validation cohort dataset were obtained from two completely different ICU databases. The development cohort dataset included at least candidate inflammatory and immune composite indicators based on nutritional status, coagulation / platelet ratio, immune cell balance, and comprehensive inflammatory burden dimensions. The candidate inflammatory and immune composite indicators in the development cohort dataset are ranked by importance for sepsis mortality and prognosis using a fully relevant feature selection algorithm to obtain important inflammatory marker indicators. The preset basic indicators are superimposed and combined with each of the important inflammatory marker indicators and the preset importance scoring criteria to obtain multiple sets of joint optimization indicator models, wherein the preset basic indicators include at least age and lactate. The scoring results of multiple joint optimization index models are compared, and the scoring results are evaluated based on the independent validation cohort dataset. The optimal joint optimization index model based on the evaluation results is determined as the elderly sepsis index model. The indexes in the elderly sepsis index model include at least age, lactate, and SII.

2. The method for constructing the sepsis index model in the elderly according to claim 1, characterized in that, The method for constructing the sepsis index model for the elderly also includes the following steps prior to determining the joint optimized index model with the best evaluation results as the sepsis index model for the elderly: Single-indicator prediction models were constructed for each candidate inflammatory and immune composite index in the development cohort dataset. The single-indicator prediction model is selected based on the AUC and the best Yoden index of each single-indicator prediction model for comprehensive performance evaluation. Based on the comprehensive performance evaluation results of the single-indicator prediction model, candidate inflammatory and immune composite indicators whose comprehensive performance of a single indicator meets the preset performance threshold are determined as core inflammatory marker indicators. The core inflammatory marker indicators include at least one candidate inflammatory and immune composite indicator. The consistency of the indicators in the optimal joint optimization indicator model based on the core inflammatory marker indicators is verified.

3. The method for constructing the sepsis index model in the elderly according to claim 2, characterized in that, Consistency verification of the indicators in the optimal joint optimization indicator model based on the core inflammatory marker indicators is performed, including: The preset basic indicators, the core inflammatory marker indicators, and the preset importance scoring criteria are all superimposed and combined to obtain multiple sets of core inflammatory joint indicator models. A comprehensive evaluation was conducted on each group of core inflammation joint indicator models to obtain the optimal core inflammation indicator model; The indicators in the core inflammation optimal indicator model are compared with the indicators in the joint optimization indicator model that yields the best current evaluation result. If the indicators in the core inflammation optimal indicator model are consistent with the indicators in the joint optimization indicator model with the best current evaluation result, then the joint optimization indicator model with the best current evaluation result can be determined as the elderly sepsis indicator model. Otherwise, a reassessment mechanism is triggered and the above process is repeated until the indicators in the core inflammation optimal indicator model are consistent with the indicators in the joint optimization indicator model that is currently the best in the assessment results.

4. The method for constructing an indicator model for sepsis in the elderly according to any one of claims 1 to 3, characterized in that, The candidate inflammatory and immune composite indicators in the development cohort dataset were ranked by importance for sepsis mortality and prognosis based on a fully relevant feature selection algorithm, including: The candidate inflammatory and immune composite indicators in the development queue dataset are used to construct a candidate inflammatory feature set. Each column in the candidate inflammation feature set is randomly rearranged to generate shadow candidate features, and the shadow candidate features are used together with the original features to train a random forest. The Z-value of each feature in the candidate inflammation feature set is calculated based on the feature importance score output by the random forest. Compare the Z-value of the original feature with the maximum Z-value of all shadow features; If the absolute value of the difference between the Z value of the original feature and the maximum Z value of all shadow features is greater than a preset threshold, the importance of the original feature is determined to be either important or irrelevant. Specifically, if the original feature is greater than the maximum value of all shadow features, the importance of the original feature is determined to be important; if the original feature is less than the maximum value of all shadow features, the importance of the original feature is determined to be irrelevant. If the absolute value of the difference between the Z value of the original feature and the maximum Z value of all shadow features is not greater than a preset threshold, then the importance of the original feature is marked as pending. Repeat the above process iteratively to identify candidate inflammatory and immune composite indicators from all development cohort datasets that are marked as important as important inflammatory marker indicators.

5. The method for constructing an index model for sepsis in the elderly according to any one of claims 1 to 3, characterized in that, The scoring results of multiple joint optimization index models are compared, and the scoring results are evaluated based on independent validation cohort datasets, including: Calculate the AUC score and Brier score of multiple joint optimization index models respectively; Based on the AUC score and Brier score of each set of joint optimization index models, construct the ROC curve, calibration curve and DCA of the set of joint optimization index models; The ROC curve, calibration curve, and DCA of each joint optimization index model were evaluated based on the independent validation cohort dataset to obtain the discrimination, calibration, and net DCA benefit evaluation results of each joint optimization index model. The discrimination, calibration, and DCA net benefit evaluation results of all joint optimization index models are compared, and the joint optimization index model with the best discrimination, calibration, and DCA net benefit evaluation results is determined as the joint optimization index model with the best evaluation results.

6. The method for constructing an index model for sepsis in the elderly according to any one of claims 1 to 3, characterized in that, A development cohort dataset of elderly sepsis patients aged ≥60 years was identified based on the intensive care unit database, including: Obtain routine blood cell counts and biochemical indicators for each elderly sepsis patient aged ≥60 years recorded in the intensive care monitoring database within a preset time period after admission; Based on the aforementioned routine blood cell count and biochemical indicators, the following indices were determined: systemic immune inflammation index, platelet-to-albumin ratio, monocyte-to-lymphocyte ratio, neutrophil-to-lymphocyte ratio, systemic inflammatory response index, systemic inflammation index, derived neutrophil-to-lymphocyte ratio, lymphocyte-to-monocyte ratio, neutrophil-to-albumin ratio, neutrophil-to-platelet ratio, platelet-to-leukocyte ratio, and prognostic nutritional index. A development cohort dataset was constructed based on the systemic immune inflammatory index, platelet-to-albumin ratio, monocyte-to-lymphocyte ratio, neutrophil-to-lymphocyte ratio, systemic inflammatory response index, systemic inflammatory index, derived neutrophil-to-lymphocyte ratio, lymphocyte-to-monocyte ratio, neutrophil-to-albumin ratio, neutrophil-to-platelet ratio, platelet-to-leukocyte ratio, and prognostic nutritional index.

7. A method for constructing a 28-day mortality risk prediction model for elderly patients with sepsis, characterized in that, include: Obtain indicator data from an elderly sepsis indicator model, wherein the elderly sepsis indicator model is constructed using the method described in any one of claims 1 to 6, and the indicator data includes at least age, lactate, and SII. The index data in the elderly sepsis index model are modeled using various machine learning algorithms to obtain various elderly sepsis index machine learning models, which include at least Logistic regression model, random forest model, support vector machine model and XGBoost model; Cross-validation was performed on various machine learning models for sepsis in the elderly, and the AUC, calibration curves and DCA of various machine learning models for sepsis in the elderly were compared in an independent development cohort dataset to determine the random forest model as the model to be trained and optimized. The indicator data in the elderly sepsis index model is used as input data to perform model distillation on the model to be trained and optimized, thereby obtaining a 28-day mortality risk prediction model for elderly sepsis. The 28-day mortality risk prediction model for elderly sepsis can be called to realize the prediction of mortality risk in elderly sepsis over 28 days.

8. The method for constructing the 28-day mortality risk prediction model for elderly patients with sepsis according to claim 7, characterized in that, Using the index data from the elderly sepsis index model as input data, the model to be trained and optimized is distilled, including: The index data of the elderly sepsis index are input into the random forest model for fitting, and the predicted probability corresponding to the index data of each elderly sepsis index is obtained, and the predicted probability is used as a soft label. Using the index data of the elderly sepsis index as independent variables and the soft label as the output target, a Logistic regression model is trained to minimize the difference between the predicted probability of the Logistic regression model and the predicted probability output by the random forest model. The trained Logistic regression model was selected as the predictive model for the 28-day mortality risk of elderly patients with sepsis.

9. A 28-day mortality risk prediction system for elderly patients with sepsis, characterized in that, include: A client and a server are connected communicatively. The client includes at least a user interaction module, and the server includes at least an algorithm engine module. The user interaction module is used to acquire the predicted sepsis index data of elderly patients input by the user, and to display the 28-day mortality risk prediction results of elderly sepsis patients to the user. The algorithm engine module is used to input the data of the sepsis index to be predicted into the 28-day mortality risk prediction model for sepsis in the elderly to predict the mortality risk and obtain the 28-day mortality risk prediction result for sepsis in the elderly; wherein the 28-day mortality risk prediction model for sepsis in the elderly is constructed by the method for constructing the 28-day mortality risk prediction model for sepsis in the elderly according to claim 8 or 9.

10. The 28-day mortality risk prediction system for elderly patients with sepsis according to claim 9, characterized in that, The client-side application also includes a request encapsulation and validation module, and the server-side application also includes a model management module and a result encapsulation and return module. The model management module is used to store, optimize, and update the 28-day mortality risk prediction model for elderly patients with sepsis. The request encapsulation and verification module is used to encapsulate the user-inputted data of sepsis indicators in the elderly into a standardized JSON file according to predefined field specifications. The result encapsulation and return module is used to encapsulate the 28-day mortality risk prediction results of elderly sepsis patients into a JSON file and return it to the user interaction module.

Images