A severe case risk prediction method and device based on LLM and machine learning fusion

Abstract

The application discloses a severe risk prediction method and device based on LLM and machine learning fusion, and the method comprises the following steps: firstly, acquiring physiological parameter data information set of a user; then, pre-processing and task routing processing the physiological parameter data information set to obtain pre-processed physiological parameter data information set and target prediction model type information; then, performing risk prediction and feature attribution analysis processing on the pre-processed physiological parameter data information set and the target prediction model type information to obtain risk prediction result information and feature attribution result information set; then, performing medical literature consistency verification processing on the feature attribution result information set and the target prediction model type information to obtain verified feature attribution result information set; finally, processing the verified feature attribution result information set and the risk prediction result information to obtain a target risk assessment report information.

Description

Technical Field

[0001] This invention relates to the field of medical data processing technology, and in particular to a method and device for predicting the risk of severe illness based on the fusion of LLM and machine learning. Background Technology

[0002] With the rapid development of artificial intelligence technology in the field of medical data processing, critical illness risk prediction technology based on machine learning models has been widely used. Existing critical illness risk prediction systems are mostly built upon supervised learning models such as XGBoost, LightGBM, or LSTM, which are trained on historical medical data to predict the risk of severe illness. However, some technical problems still need to be solved in practical applications.

[0003] Existing critical illness risk prediction technologies suffer from the following technical shortcomings: First, prediction methods based on a single machine learning model essentially rely on correlation fitting based on historical data distribution. When data bias exists in the training data, the model is prone to capturing statistical spurious correlation features and misclassifying them as valid features, leading to a decrease in the reliability of the prediction results. Second, while general-purpose large language models perform well in text processing, their parameter weights are fixed at the end of training, failing to internalize the latest medical knowledge updates and exhibiting insufficient numerical calculation accuracy when processing continuous physiological parameters. Third, existing interpretability technologies such as SHAP and LIME can only provide explanations at the feature attribution level, unable to introduce external authoritative knowledge to verify the consistency of model conclusions. When the model gives incorrect attributions based on data bias, existing technologies cannot automatically detect and correct them. Furthermore, existing technologies struggle to effectively integrate the high-precision numerical prediction capabilities of machine learning with the semantic understanding capabilities of large language models, resulting in limited overall performance of the prediction system. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a critical illness risk prediction method and device based on the fusion of LLM and machine learning. By integrating the high-precision numerical prediction capability of machine learning models with the semantic understanding capability of large language models, and introducing a medical literature knowledge base for consistency verification, the invention solves the technical problems of insufficient reliability of prediction results and inability to automatically verify feature attribution in the prior art.

[0005] To address the aforementioned technical problems, a first aspect of this invention discloses a method for predicting the risk of severe illness based on the fusion of LLM and machine learning, the method comprising: S1, Obtain the user's physiological parameter data set; S2, preprocess and task routing process the physiological parameter data information set to obtain the preprocessed physiological parameter data information set and target prediction model type information; S3, perform risk prediction and feature attribution analysis on the preprocessed physiological parameter data information set and the target prediction model type information to obtain risk prediction result information and feature attribution result information set respectively; S4, Perform medical literature consistency verification on the feature attribution result information set and the target prediction model type information to obtain the verified feature attribution result information set respectively; S5, process the verified feature attribution result information set and the risk prediction result information to obtain the target risk assessment report information.

[0006] As an optional implementation, in the first aspect of the present invention, the preprocessing and task routing processing of the physiological parameter data information set to obtain the preprocessed physiological parameter data information set and target prediction model type information includes: S21, perform data cleaning processing on the physiological parameter data information set to obtain the first physiological parameter data information set; S22, perform missing value filling processing on the first physiological parameter data information set to obtain the second physiological parameter data information set; S23, normalize the second physiological parameter data information set to obtain a preprocessed physiological parameter data information set; S24, Perform task routing processing on the preprocessed physiological parameter data information set to obtain target prediction model type information.

[0007] As an optional implementation, in the first aspect of the present invention, the normalization processing of the second physiological parameter data information set to obtain a preprocessed physiological parameter data information set includes: Using the normalized first calculation model, each physiological parameter data in the second physiological parameter data set is normalized to obtain the preprocessed physiological parameter data set. The first normalized calculation model is as follows: ; ; ; In the formula, The normalized value of the physiological parameter of the j-th physiological parameter data in the preprocessed physiological parameter data set. The physiological parameter value of the j-th physiological parameter data in the second physiological parameter data set. Let be the mean of the j-th physiological parameter data in a preset reference dataset. Let be the standard deviation of the j-th physiological parameter data in the preset reference dataset. and These are the first and third quartiles of the j-th physiological parameter data in a preset reference dataset, respectively. Let be the kurtosis value of the j-th physiological parameter data in a preset reference dataset. For adaptive fusion weights, The sigmoid scaling factor. Kurtosis sensitivity coefficient Kurtosis threshold It is a smoothing constant. M represents the total amount of physiological parameter data.

[0008] As an optional implementation, in the first aspect of the present invention, the step of performing task routing processing on the preprocessed physiological parameter data information set to obtain target prediction model type information includes: S241, The preprocessed physiological parameter data information set is extracted and processed to obtain the key indicator data information set; S242, perform matching processing on the key indicator data information set and the preset indicator threshold condition set to obtain the matching result information set; S243, perform model type determination processing on the matching result information set to obtain target prediction model type information.

[0009] As an optional implementation, in the first aspect of the present invention, the step of performing risk prediction and feature attribution analysis on the preprocessed physiological parameter data information set and the target prediction model type information to obtain risk prediction result information and feature attribution result information set, respectively, includes: S31, Based on the preset machine learning model pool, perform model loading processing on the target prediction model type information to obtain a machine learning prediction model; S32, using the machine learning prediction model, perform model inference processing on the preprocessed physiological parameter data information set to obtain the risk probability value; S33, Perform risk level classification on the risk probability value to obtain risk level information; S34, integrate the risk probability value and the risk level information to obtain risk prediction result information; S35, using the machine learning prediction model, perform feature attribution analysis on the preprocessed physiological parameter data information set to obtain a feature attribution result information set.

[0010] As an optional implementation, in the first aspect of the present invention, the step of using the machine learning prediction model to perform feature attribution analysis on the preprocessed physiological parameter data information set to obtain a feature attribution result information set includes: S351, using a feature attribution calculation model, feature attribution analysis is performed on the machine learning prediction model and the preprocessed physiological parameter data information set to obtain attribution value information; The feature attribution calculation model is as follows: ; In the formula, The attribution value information is the first The attribution values ​​of feature i are given by F, where F is the set of all features and S is a subset of features excluding feature i. The output value of the machine learning prediction model when feature i is included. The output value of the machine learning prediction model when feature i is not included. Let S be the number of features in subset S. The total number of features; S352, Perform interactive enhancement processing on the attribution value information to obtain enhanced attribution value information; the enhanced attribution value information includes several attribution contribution values; S353, sort the enhanced attribution value information to obtain sorted attribution value information; S354, Select the top N features and their corresponding attribution contribution values ​​from the sorted attribution value information, and integrate the top N features and their corresponding attribution contribution values ​​with the preprocessed physiological parameter data information set to obtain the feature attribution result information set.

[0011] As an optional implementation, in the first aspect of the present invention, processing the verified feature attribution result information set and the risk prediction result information to obtain target risk assessment report information includes: S51, Perform a summary generation process on the risk prediction result information to obtain risk prediction summary information; S52, perform detail generation processing on the verified feature attribution result information set to obtain feature analysis detail information; S53, perform inconsistent feature filtering processing on the verified feature attribution result information set to obtain correction warning information; S54, perform intervention scheme retrieval processing on the target prediction model type information to obtain recommended scheme information; S55, the risk prediction summary information, the feature analysis details information, the correction warning information and the recommended scheme information are combined and processed to generate the target risk assessment report information.

[0012] A second aspect of this invention discloses a critical illness risk prediction device based on the fusion of LLM and machine learning, the device comprising: The acquisition module is used to acquire the user's physiological parameter data set; The first processing module is used to preprocess and task routing the physiological parameter data information set to obtain the preprocessed physiological parameter data information set and target prediction model type information. The second processing module is used to perform risk prediction and feature attribution analysis on the preprocessed physiological parameter data information set and the target prediction model type information to obtain risk prediction result information and feature attribution result information set respectively. The consistency verification module is used to perform medical literature consistency verification on the feature attribution result information set and the target prediction model type information to obtain the verified feature attribution result information set respectively. The report generation module is used to process the verified feature attribution result information set and the risk prediction result information to obtain the target risk assessment report information.

[0013] A third aspect of this invention discloses another critical illness risk prediction device based on the fusion of LLM and machine learning, the device comprising: Processor 301; A memory 302 containing executable program code is coupled to the processor 301; The processor 301 calls the executable program code stored in the memory 302 to execute some or all of the steps of the critical illness risk prediction method based on the fusion of LLM and machine learning disclosed in the first aspect of the present invention.

[0014] The fourth aspect of the present invention discloses a computer-readable storage medium storing computer instructions, which, when invoked, are used to execute some or all of the steps of the critical illness risk prediction method based on the fusion of LLM and machine learning disclosed in the first aspect of the present invention.

[0015] Compared with the prior art, the embodiments of the present invention have the following beneficial effects: In this embodiment of the invention, the user's physiological parameter data set is first acquired; then, the physiological parameter data set is preprocessed and task routing is performed to obtain a preprocessed physiological parameter data set and target prediction model type information; next, risk prediction and feature attribution analysis are performed on the preprocessed physiological parameter data set and target prediction model type information to obtain risk prediction result information and feature attribution result information set; then, medical literature consistency verification is performed on the feature attribution result information set and target prediction model type information to obtain a verified feature attribution result information set; finally, the verified feature attribution result information set and risk prediction result information are processed to obtain target risk assessment report information. It can be seen that this application achieves high-precision numerical prediction and semantic understanding through the integration of machine learning models and large language models; by introducing a medical literature knowledge base for consistency verification, it can automatically detect and correct erroneous attributions caused by data deviations in the model; and through the task routing mechanism, it realizes automated selection and on-demand invocation of prediction models, improving the system's processing efficiency and prediction accuracy. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating a critical illness risk prediction method based on the fusion of LLM and machine learning, as disclosed in an embodiment of the present invention. Figure 2 This is a schematic diagram of a critical illness risk prediction device based on the fusion of LLM and machine learning, as disclosed in an embodiment of the present invention. Figure 3 This is a schematic diagram of another critical illness risk prediction device based on the fusion of LLM and machine learning disclosed in an embodiment of the present invention. Detailed Implementation

[0018] 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 of the present invention. 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 are within the scope of protection of the present invention.

[0019] The terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion.

[0020] Specific features, structures, or characteristics described in connection with the embodiments may be included in at least one embodiment of the invention. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment that is mutually exclusive with other embodiments.

[0021] Example 1 Please see Figure 1 , Figure 1 This is a flowchart illustrating a critical illness risk prediction method based on the fusion of LLM and machine learning, as disclosed in an embodiment of the present invention. Figure 1 The described critical illness risk prediction method based on the fusion of LLM and machine learning is applied to medical data processing systems, such as local servers or cloud servers in medical institution data centers, etc., and the embodiments of this invention are not limited thereto. Figure 1 As shown, this critical illness risk prediction method based on the fusion of LLM and machine learning includes: S1, obtain the user's physiological parameter data set.

[0022] It should be noted that the physiological parameter data set includes several physiological parameter data information, each of which includes a parameter name, parameter value, and collection timestamp. The physiological parameter data information may include, but is not limited to, blood oxygen saturation, heart rate, blood pressure, body temperature, respiratory rate, lactate level, creatinine level, bilirubin level, and blood gas analysis data. For example, the physiological parameter data set includes 34 physiological indicator data.

[0023] S2, preprocess and task routing are performed on the physiological parameter data information set to obtain the preprocessed physiological parameter data information set and the target prediction model type information.

[0024] It should be noted that the preprocessing includes data cleaning, missing value imputation, and data normalization.

[0025] It should be noted that the task routing process is a classification process based on key indicators. It determines the user's risk type by using preset indicator threshold conditions, thereby identifying the target prediction model to be invoked. For example, when a key indicator meets a preset threshold condition, the task is routed to the corresponding prediction model. It should also be noted that the target prediction model type information refers to the type identifier of the critical illness risk prediction model that should be used for this task, as determined by the task routing process. This identifier is used to load the corresponding machine learning prediction model from a preset machine learning model pool. The target prediction model type information may include MODS prediction models, ARDS prediction models, hemorrhagic shock prediction models, etc. The above task routing process can be implemented by calling the Large Language Model API interface.

[0026] S3, perform risk prediction and feature attribution analysis on the preprocessed physiological parameter data information set and the target prediction model type information to obtain risk prediction result information and feature attribution result information set respectively.

[0027] It should be noted that the risk prediction processing refers to using a machine learning prediction model corresponding to the target prediction model type information to perform model inference processing on the preprocessed physiological parameter data set to obtain risk probability values ​​and risk level information. The machine learning prediction model is a pre-trained supervised learning model, which can employ XGBoost, LightGBM, Random Forest, LSTM, or other machine learning models; this embodiment of the invention is not limited to any particular model. The risk prediction result information includes risk probability values ​​and risk level information. For example, the risk prediction result information can be expressed as "the probability of MODS occurring within the next 24 hours is 85%, and the risk level is high risk."

[0028] It should be noted that the feature attribution analysis process refers to using interpretability algorithms to analyze the feature importance of the prediction results of the machine learning model, and to determine the contribution of each input feature to the prediction results. The feature attribution analysis process can employ the SHAP (SHapley Additive exPlanations) algorithm, the LIME (Local Interpretable Model-agnostic Explanations) algorithm, or other interpretable algorithms, and this embodiment of the invention is not limited thereto.

[0029] It should be noted that the feature attribution result information set includes several feature attribution result information sets, each of which includes a feature name, feature value, attribution contribution value, and contribution direction information. The attribution contribution value indicates the degree of influence of the feature on the prediction result; the contribution direction information indicates whether the feature contributes positively or negatively. In this embodiment of the invention, the feature attribution result information set includes the attribution results of Top-N key features, where N is a positive integer, preferably N equal to 10.

[0030] S4, Perform medical literature consistency verification on the feature attribution result information set and the target prediction model type information to obtain the verified feature attribution result information set respectively.

[0031] It should be noted that the aforementioned medical literature consistency verification process involves comparing the feature attribution results of the machine learning model with the medical knowledge in the medical literature knowledge base to detect any inconsistent attribution results and correct any inconsistencies. This medical literature consistency verification process is implemented by calling the Large Language Model API interface.

[0032] It should be noted that the verified feature attribution result information set includes several verified feature attribution result information sets. Each verified feature attribution result information set includes a feature name, feature value, attribution contribution value, contribution direction information, verification result information, and correction description information. The verification result information indicates whether the feature attribution result is consistent with medical literature knowledge, and the value is consistent (1) or inconsistent (0). The correction description information is used to record the correction content and basis when the verification result is inconsistent.

[0033] S5, process the verified feature attribution result information set and the risk prediction result information to obtain the target risk assessment report information.

[0034] It should be noted that the above processing refers to integrating the risk prediction results and the verified feature attribution results into a structured risk assessment report. The target risk assessment report includes risk prediction results, key feature analysis information, a summary of verification results, and recommended intervention plans. The recommended intervention plans are obtained from a medical literature knowledge base.

[0035] As can be seen, the critical illness risk prediction method based on the fusion of LLM and machine learning described in this embodiment of the invention first obtains the user's physiological parameter data set; then, the physiological parameter data set is preprocessed and task routing is performed to obtain the preprocessed physiological parameter data set and the target prediction model type information; next, risk prediction and feature attribution analysis are performed to obtain risk prediction results and feature attribution results; then, medical literature consistency verification is performed to obtain the verified feature attribution results; finally, report generation is performed to obtain the target risk assessment report information. This method achieves accurate prediction of critical illness risk by integrating the high-precision numerical prediction capability of machine learning models with the semantic understanding capability of large language models; by introducing a medical literature knowledge base for consistency verification, it can automatically detect and correct erroneous attributions caused by data bias in the model, improving the reliability and interpretability of the prediction results.

[0036] In an optional embodiment, the preprocessing and task routing of the physiological parameter data set to obtain the preprocessed physiological parameter data set and target prediction model type information includes: S21, perform data cleaning processing on the physiological parameter data information set to obtain the first physiological parameter data information set; S22, perform missing value filling processing on the first physiological parameter data information set to obtain the second physiological parameter data information set; S23, normalize the second physiological parameter data information set to obtain a preprocessed physiological parameter data information set; S24, Perform task routing processing on the preprocessed physiological parameter data information set to obtain target prediction model type information.

[0037] It should be noted that the data cleaning process includes deleting abnormal data that exceeds a preset value range and deleting duplicate data records. For example, for heart rate values, if a value less than 20 or greater than 300 is detected, it is determined to be abnormal data and deleted.

[0038] It should be noted that the missing value filling process can be carried out in one of the following ways: mean filling, median filling, forward filling, backward filling, or multiple interpolation filling.

[0039] It should be noted that the normalization process can employ Min-Max normalization, Z-Score normalization, or other normalization methods.

[0040] In an optional embodiment, the normalization process of the second physiological parameter data set to obtain a preprocessed physiological parameter data set includes: Using the normalized first calculation model, each physiological parameter data in the second physiological parameter data set is normalized to obtain the preprocessed physiological parameter data set. The first normalized calculation model is as follows: ; ; ; In the formula, The normalized value of the physiological parameter of the j-th physiological parameter data in the preprocessed physiological parameter data set. The physiological parameter value of the j-th physiological parameter data in the second physiological parameter data set. Let be the mean of the j-th physiological parameter data in a preset reference dataset. Let be the standard deviation of the j-th physiological parameter data in the preset reference dataset. and These are the first and third quartiles of the j-th physiological parameter data in a preset reference dataset, respectively. Let be the kurtosis value of the j-th physiological parameter data in a preset reference dataset. For adaptive fusion weights, The sigmoid scaling factor. Kurtosis sensitivity coefficient Kurtosis threshold It is a smoothing constant. M represents the total amount of physiological parameter data.

[0041] It should be noted that the preset reference dataset is a large-scale sample dataset of physiological parameters collected in advance, used to provide the statistical distribution characteristics of each physiological parameter. The mean, standard deviation, first quartile, third quartile, and kurtosis are all statistical quantities calculated from the preset reference dataset. Specifically, the first quartile is the value located at the 25th percentile after all values ​​of the j-th physiological parameter in the reference dataset are arranged in ascending order, and the third quartile is the value located at the 75th percentile.

[0042] It should be noted that the first normalization calculation model achieves adaptive normalization processing of physiological parameter data with different distribution characteristics by integrating Sigmoid transform and quartile normalization. When the data kurtosis... When it is high (outliers exist), When the data approaches 0, robust normalization based on quartiles is used; when the data distribution is close to a normal distribution, Approaching 1, a smooth normalization based on Sigmoid is employed. The parameters are adaptively calculated based on the actual distribution range of the data. Their function is to control the scaling ratio of the Sigmoid function, ensuring that the normalized data falls within the interval [0.05, 0.95]. According to the formula, The typical value range is from 0.1 to 10.0. In this embodiment of the invention, The value ranges from 0.5 to 2.0, with a preferred value of 1.0; The value ranges from 2.0 to 4.0, with a preferred value of 3.0; The range of values ​​is to The preferred value is .

[0043] As can be seen, the critical illness risk prediction method based on the fusion of LLM and machine learning described in the embodiments of the present invention improves the quality and consistency of input data by performing data cleaning, missing value imputation and normalization on physiological parameter data, thus providing a high-quality data foundation for subsequent risk prediction processing.

[0044] In another optional embodiment, the step of performing task routing processing on the preprocessed physiological parameter data set to obtain target prediction model type information includes: S241, The preprocessed physiological parameter data information set is extracted and processed to obtain the key indicator data information set; It should be noted that the extraction process refers to filtering physiological parameter data related to the type of severe illness risk from the preprocessed physiological parameter data set according to a preset list of key indicator names. The key indicator data set includes oxygenation index data, lactate value data, blood pressure data, urine output data, etc. The above extraction process can be implemented through key-value queries, field filtering, or index retrieval, etc., and this embodiment of the invention does not impose any limitations on these methods.

[0045] S242, perform matching processing on the key indicator data information set and the preset indicator threshold condition set to obtain the matching result information set; It should be noted that the set of indicator threshold conditions includes several indicator threshold conditions, each of which includes an indicator name, a threshold range, and a corresponding model type. For example, the indicator threshold conditions can be expressed as follows: when the oxygenation index PaO2 / FiO2 < 300, it corresponds to the ARDS prediction model; when the lactate value > 2 mmol / L, it corresponds to the Sepsis prediction model; and when the systolic blood pressure < 90 mmHg and urine output < 0.5 ml / kg / h, it corresponds to the hemorrhagic shock prediction model.

[0046] It should be noted that the matching process refers to traversing each key indicator data information in the key indicator data information set, comparing its value with the threshold range of the corresponding indicator in the indicator threshold condition set, determining whether the threshold condition is met, and recording the matching result. The matching result information set includes several matching result information, each of which contains the indicator name, indicator value, whether the threshold condition is met, and the corresponding model type.

[0047] S243, perform model type determination processing on the matching result information set to obtain target prediction model type information.

[0048] It should be noted that the model type determination process refers to determining the target prediction model type information to be invoked based on the matching results in the matching result information set that meet the threshold conditions. When multiple model types meet the conditions, the model type with the highest priority can be selected as the target prediction model type information according to a preset priority rule; when no model type meets the conditions, a preset default model type can be selected as the target prediction model type information. The priority rule can be set according to factors such as the urgency of the disease, the severity of the risk, or clinical guideline recommendations.

[0049] As can be seen, the critical illness risk prediction method based on the fusion of LLM and machine learning described in the embodiments of the present invention can automatically select the most suitable prediction model according to the user's specific physiological state through a task routing mechanism based on key indicators, avoiding information redundancy caused by using a general model and improving the pertinence and accuracy of the prediction.

[0050] In another optional embodiment, the step of performing risk prediction and feature attribution analysis on the preprocessed physiological parameter data information set and the target prediction model type information to obtain risk prediction result information and feature attribution result information set, respectively, includes: S31, Based on the preset machine learning model pool, perform model loading processing on the target prediction model type information to obtain a machine learning prediction model; It should be noted that the model loading process refers to retrieving and loading the corresponding machine learning prediction model from a pre-defined machine learning model pool into memory based on the model identifier in the target prediction model type information. The machine learning model pool is a pre-built model repository containing machine learning prediction models trained for different types of severe illnesses, such as MODS prediction models, ARDS prediction models, and Sepsis prediction models. The machine learning model pool can be stored on a local server or a cloud server and accessed through a model loading interface.

[0051] S32, using the machine learning prediction model, perform model inference processing on the preprocessed physiological parameter data information set to obtain the risk probability value; It should be noted that the model inference processing refers to taking the preprocessed physiological parameter data set as input, performing forward propagation calculations through the computational graph of the machine learning prediction model, and obtaining the output result. Specifically, the preprocessed physiological parameter data set contains several physiological parameter data of a single user. These several physiological parameter data are arranged in a preset order to construct a one-dimensional feature vector, which is used as the input of the machine learning prediction model. The machine learning prediction model performs model parameter matrix operations and output activation processing on the one-dimensional feature vector, outputting a single risk probability value. The risk probability value is a floating-point number between 0 and 1, representing the probability that the user will experience this type of severe illness event within a preset time window. The above model inference processing can be implemented using machine learning frameworks such as TensorFlow, PyTorch, ONNXRuntime, or scikit-learn, and this embodiment of the invention is not limited thereto.

[0052] S33, Perform risk level classification on the risk probability value to obtain risk level information; It should be noted that the risk level classification process refers to converting risk probability values ​​into risk levels based on preset risk level classification thresholds. For example, a risk probability value less than 0.3 can be defined as low-risk information, 0.3 to 0.7 as medium-risk information, and greater than 0.7 as high-risk information. The above risk level classification process can be carried out using a fixed threshold segmentation method, a confidence-aware dynamic threshold method, or a condition-based judgment method, etc. Specifically, this embodiment of the invention does not limit the specific methods. The specific threshold settings can be adjusted according to the actual application scenario.

[0053] S34, integrate the risk probability value and the risk level information to obtain risk prediction result information; It should be noted that the integration process refers to combining the risk probability value and the risk level information according to a preset data structure to form structured risk prediction result information.

[0054] S35, using the machine learning prediction model, perform feature attribution analysis on the preprocessed physiological parameter data information set to obtain a feature attribution result information set.

[0055] As can be seen, the critical illness risk prediction method based on the fusion of LLM and machine learning described in the embodiments of the present invention achieves unified management and on-demand invocation of multiple prediction models through the machine learning model pool mechanism, thereby improving the scalability and flexibility of the system.

[0056] In an optional embodiment, the step of using the machine learning prediction model to perform feature attribution analysis on the preprocessed physiological parameter data set to obtain a feature attribution result information set includes: S351, using a feature attribution calculation model, feature attribution analysis is performed on the machine learning prediction model and the preprocessed physiological parameter data information set to obtain attribution value information; The feature attribution calculation model is as follows: ; In the formula, The attribution value information is the first The attribution values ​​of feature i are given by F, where F is the set of all features and S is a subset of features excluding feature i. The output value of the machine learning prediction model when feature i is included. The output value of the machine learning prediction model when feature i is not included. Let S be the number of features in subset S. The total number of features; It should be noted that the features in the formula refer to the various physiological parameter dimensions that are used as inputs to the machine learning prediction model, obtained after constructing feature vectors from the preprocessed physiological parameter data set. The features correspond one-to-one with the physiological parameter data in the preprocessed physiological parameter data set in a preset order: the i-th feature corresponds to the normalized value of the i-th physiological parameter data, where i is the feature index. The set F is a set consisting of the indices of all features, i.e., F = {1, 2, ..., ...} },in This is equal to the number of physiological parameter data points in the preprocessed physiological parameter data set (i.e., the total number of features, consistent with the number of items contained in the physiological parameter data set). The subset S is any subset of F that does not contain index i, i.e. In the formula The summation represents iterating through all subsets S that satisfy the condition. This refers to the output value of the machine learning prediction model when the features corresponding to the indices in S are retained as actual input values, and the remaining features (features corresponding to the indices in F and S) are set to preset baseline values ​​(such as the mean of the corresponding dimension in the reference dataset). This represents the output value of the machine learning prediction model when the input value of feature i is added to S.

[0057] It should be noted that the feature attribution calculation model obtains a fair allocation of feature importance by calculating the marginal contribution of each feature across all possible feature combinations. The feature attribution calculation model guarantees that the sum of the feature attribution values ​​equals the difference between the model output and the benchmark output, demonstrating completeness and consistency.

[0058] S352, Perform interactive enhancement processing on the attribution value information to obtain enhanced attribution value information; the enhanced attribution value information includes several attribution contribution values; It should be noted that the above-mentioned interaction enhancement processing can use an interaction enhancement feature attribution calculation model, or it can use a feature interaction attribution method based on integral gradient, a second-order interaction attribution method based on Hessian matrix, or a local interaction attribution method based on LIME. Specifically, the embodiments of the present invention are not limited.

[0059] The interaction-enhanced feature attribution calculation model is as follows: ; ; ; ; In the formula, The attribution contribution value of the i-th feature in the enhanced attribution value information. The attribution value is the attribution value of the i-th feature in the attribution value information. The feature interaction value between the i-th feature and the j-th feature. The interaction item fusion coefficient, For symbolic functions, This is the output value of the model when both features i and j are included. This is the final attribution value after stability adjustment. Let be the standard deviation of the attribution value of the i-th feature in the attribution value information obtained through Bootstrap sampling. The mean of the standard deviations of the attribution values ​​of all features in the attribution value information is given, and v is the stability penalty strength parameter. This is the smoothing constant.

[0060] It should be noted that the interactive enhancement feature attribution calculation model introduces feature interaction values. The stability adjustment mechanism addresses the issues of traditional methods neglecting synergistic effects between features and unstable attribution values. When two features have the same attribution direction (both risk factors or both protective factors) and exhibit a positive interaction effect, the interaction term enhances the attribution value. The stability adjustment mechanism evaluates the variance of the attribution value through bootstrap sampling and penalizes attribution values ​​with large variance (instability), improving the reliability of the attribution results. In this embodiment of the invention, The value of is in the range of 0.1 to 0.3, with a preferred value of 0.2; the value of v is in the range of 0.5 to 2.0, with a preferred value of 1.0; The range of values ​​is to The preferred value is .

[0061] S353, sort the enhanced attribution value information to obtain sorted attribution value information; It should be noted that the sorting process refers to arranging each feature and its corresponding attribution contribution value in descending order based on the absolute value of the attribution contribution value in the enhanced attribution value information, so that the feature with the greatest contribution is placed first. The sorted attribution value information maintains a one-to-one correspondence between feature indices and attribution contribution values.

[0062] S354, Select the top N features and their corresponding attribution contribution values ​​from the sorted attribution value information, and integrate the top N features and their corresponding attribution contribution values ​​with the preprocessed physiological parameter data information set to obtain the feature attribution result information set.

[0063] It should be noted that the integration process refers to, for each selected feature and its corresponding attribution contribution value, retrieving the corresponding feature name and feature value from the preprocessed physiological parameter data set according to the feature index, and combining the attribution contribution value and the contribution direction information determined by the positive or negative sign of the attribution contribution value to form feature attribution result information. The above N feature attribution result information sets are then combined. The feature attribution result information includes feature name, feature value, attribution contribution value, and contribution direction information; the contribution direction information is determined according to the positive or negative sign of the attribution contribution value, with a positive attribution contribution value indicating a positive contribution and a negative attribution contribution value indicating a negative contribution. N is a positive integer used to control the number of key features selected. In this embodiment, the value of N ranges from 5 to 20, with a preferred value of 10. Selecting Top-N key features can focus on the most important influencing factors while reducing the computational complexity of subsequent verification processing.

[0064] As can be seen, by implementing the critical illness risk prediction method based on the fusion of LLM and machine learning described in the embodiments of the present invention, the contribution of each physiological parameter to the prediction results can be quantified through SHAP feature attribution analysis, providing a clear analytical object for medical literature consistency verification.

[0065] In another optional embodiment, the step of performing medical literature consistency verification on the feature attribution result information set and the target prediction model type information to obtain the verified feature attribution result information set includes: S41, perform index database matching processing on the target prediction model type information to obtain a medical literature vector index database; It should be noted that the medical literature vector index library is a pre-built vector database containing vectorized representations of authoritative medical literature for different disease types. For example, for Sepsis, the medical literature vector index library contains vectorized indexes of authoritative guidelines such as the *Surviving Sepsis Campaign Guidelines*; for ARDS, the medical literature vector index library contains vectorized indexes of *ARDSNet Protocols* and related Berlin definition literature. The above index library matching process can be performed using key-value mapping table queries, retrieval of the correspondence between model types and index library paths in configuration files, or relational database queries, etc. Specifically, this embodiment of the invention does not limit the specific methods used.

[0066] S42, preset feature index s=1, preset total number of features is the number N of feature attribution result information in the feature attribution result information set; S43, construct and process the s-th feature attribution result information in the feature attribution result information set to obtain the s-th document retrieval query text; S44, Perform vector similarity retrieval processing on the sth document retrieval query text and the medical document vector index library to obtain the sth related document fragment information set; It should be noted that the vector similarity retrieval processing described herein is a document retrieval method based on Retrieval-Augmented Generation (RAG) technology. The implementation process of the vector similarity retrieval processing includes: first, converting the document retrieval query text into a vector representation using a text embedding model; then, calculating the similarity between the query vector and the document vector in a vector index library; and finally, returning the top few document fragments with the highest similarity. The similarity calculation can employ cosine similarity, Euclidean distance, or other vector distance metrics.

[0067] S45, perform consistency comparison processing on the sth relevant document fragment information set and the sth feature attribution result information to obtain the sth verification result information; It should be noted that the consistency comparison process refers to logically comparing the feature attribution results of the machine learning model with the relevant literature fragment information set. For example, if the machine learning model shows that a certain feature is a protective factor (negative attribution contribution value), while the relevant literature fragment information set indicates that the feature should be a risk factor in pathophysiology, then it is determined to be inconsistent. The consistency comparison process is implemented by calling the large language model API interface, whereby the large language model performs logical reasoning and judgment on the feature attribution conclusions based on the relevant literature fragment information set.

[0068] S46, determine whether the s-th verification result information is inconsistent, and obtain the first judgment result; When the first judgment result is yes, the sth relevant document fragment information set is subjected to correction and explanation generation processing to obtain the sth correction and explanation information; When the first judgment result is negative, a consistent description message is generated, and the consistent description message is determined to be the s-th correction description message; It should be noted that the consistency description information is a preset fixed string, indicating that the feature attribution is consistent with medical literature and requires no correction. When the first judgment result is negative, the fixed string is used as the s-th correction description information.

[0069] It should be noted that the correction description generation process refers to processing the s-th relevant document fragment information set and the s-th feature attribution result information to generate text information for recording the correction content and basis. This correction description generation process can be implemented by calling the large language model API interface, whereby the large language model generates the correction description based on the document fragment information set and feature attribution result information; alternatively, it can be obtained by formatting the correction suggestions output by the large language model during the consistency comparison stage. Specifically, this embodiment of the invention does not limit the specific implementation.

[0070] It should be noted that determining whether the s-th verification result information is inconsistent means: reading the value of the verification result field in the s-th verification result information, where the value of the verification result field is 1 or 0; when the value of the verification result field is equal to 1, the first judgment result is determined to be yes; when the value of the verification result field is equal to 0, the first judgment result is determined to be no.

[0071] S47, determine whether s is equal to N, and obtain a second determination result; When the second judgment result is negative, increment s by 1 and execute S43; If the second judgment result is yes, execute S48; S48, integrate the verification result information and the correction description information of all features to obtain the verified feature attribution result information set.

[0072] As can be seen, the critical illness risk prediction method based on the fusion of LLM and machine learning described in the embodiments of the present invention can use authoritative medical knowledge to verify and correct the attribution results of the machine learning model through the medical literature consistency verification mechanism based on RAG technology, effectively solving the problem of erroneous attribution caused by data bias in the model and improving the medical reliability of the prediction results.

[0073] In an optional embodiment, the step of performing vector similarity retrieval processing on the s-th document search query text and the medical document vector index to obtain the s-th related document fragment information set includes: S441, The sth document retrieval query text is vectorized to obtain query vector information; It should be noted that the above vectorization processing can be performed using text embedding models, embedding layers of pre-trained language models, or vector representations based on TF-IDF, etc. The specific implementation of this invention is not limited.

[0074] S442, Perform similarity calculation on the query vector information and the document vectors in the medical literature vector index library to obtain a similarity value information set; It should be noted that the above processing can be performed using a multi-scale weighted vector similarity retrieval model, or by using cosine similarity calculation, inner product similarity calculation, or similarity retrieval based on ANN approximate nearest neighbor index. Specifically, the embodiments of the present invention are not limited.

[0075] The multi-scale weighted vector similarity retrieval model is as follows: ; ; ; ; ; In the formula, sim is the similarity value information set, and sim(q,d) is the comprehensive similarity value between the query vector q and the document vector d. For the cosine similarity component, To normalize the Euclidean similarity components, To query the i-th component of the vector, Let be the i-th component of the document vector, and n be the dimension of the vector. The cosine similarity weight coefficient is... It is the generalized average index. As the time-degradation factor, This is the current timestamp. This is the publication timestamp of document d. For the time-related decay period, For decay rate parameters, This is the initial threshold for attenuation. The authority weight of document d.

[0076] It should be noted that the multi-scale weighted vector similarity retrieval model overcomes the limitations of a single similarity measure by fusing cosine similarity and Euclidean similarity. The generalized average index... Controlling the fusion method of the two similarities: when When is the arithmetic mean, when The time-related decay factor tends to reach its maximum value. Older literature is weighted less to ensure that the most recent medical evidence is returned first. (The authority weighting is mentioned here.) The weighting is determined based on the source level of the literature (e.g., Cochrane systematic reviews, clinical guidelines, randomized controlled trials, etc.), with higher-authority literature receiving greater weight. In this embodiment of the invention, The value ranges from 0.5 to 0.8, with a preferred value of 0.7; The value ranges from 1.0 to 3.0, with a preferred value of 2.0; The value ranges from 0.1 to 0.5, with a preferred value of 0.2; The value ranges from 0.3 to 0.7, with a preferred value of 0.5; The value ranges from 0.5 to 1.5, depending on the document level.

[0077] S443, sort and filter all the comprehensive similarity values ​​in the similarity value information set in descending order to obtain the s-th relevant document fragment information set.

[0078] It should be noted that the similarity value information set consists of the comprehensive similarity value between the query vector and each document vector in the medical literature vector index library, and each similarity value in the similarity value information set corresponds one-to-one with a document fragment in the medical literature vector index library according to the index, that is, the j-th similarity value corresponds to the document fragment of the j-th document vector in the index library. The descending sorting and filtering process refers to: firstly, sorting the similarity value information set from largest to smallest according to the comprehensive similarity value (maintaining the correspondence between the similarity value and the document fragment index during sorting), and taking the top K elements with the largest similarity values; then, according to the document fragment index corresponding to the K elements, retrieving the corresponding document fragment content (such as document title, abstract, or text fragment, etc.) from the medical literature vector index library, and forming the s-th related document fragment information set by arranging the above K document fragments in descending order of similarity. The aforementioned descending order sorting can be processed using algorithms such as quicksort, mergesort, heapsort, or index-based indirect sorting; the aforementioned filtering process can be processed using methods such as selecting the top K document fragments with the highest similarity values, filtering based on a similarity threshold, or filtering based on quantiles, etc. Specifically, this embodiment of the invention does not impose limitations. K is a positive integer used to control the number of relevant document fragments returned. In this embodiment of the invention, the value of K ranges from 3 to 10, with a preferred value of 5.

[0079] As can be seen, the critical illness risk prediction method based on the fusion of LLM and machine learning described in the embodiments of the present invention can efficiently retrieve the most relevant literature fragments from a large-scale medical literature database by using cosine similarity calculation for vector retrieval, thus providing accurate medical knowledge support for consistency verification.

[0080] In another optional embodiment, the consistency comparison processing of the s-th relevant document fragment information set and the s-th feature attribution result information to obtain the s-th verification result information includes: S451, perform prompt text construction processing on the sth feature attribution result information and the sth related document fragment information set to obtain consistency comparison prompt text; It should be noted that the format for constructing the consistency comparison prompt text is as follows: "Based on the following medical literature evidence, please determine whether the feature attribution conclusions of the machine learning model are consistent with medical knowledge."

[0081] Machine learning model conclusion: The value of feature [feature name] is [feature value], the SHAP contribution value is [attribution contribution value], and the contribution direction is [positive contribution / negative contribution].

[0082] Literature evidence: [Relevant literature excerpts] Please determine whether this attribution of the characteristic is consistent with evidence from medical literature. If not, please explain why. S452, using a large language model, perform logical reasoning processing on the consistency comparison prompt text to obtain consistency determination result information; It should be noted that the logical reasoning process is implemented by calling the API interface of a large language model. The large language model mentioned above can be DeepSeek, Kimi K2, or the Qwen series, etc.; specifically, this embodiment of the invention does not limit the specific model. The consistency determination result information is output by the large language model, including the determination conclusion (consistent / inconsistent) and the determination basis. When the determination conclusion is inconsistent, the large language model will also output correction suggestions to generate correction explanation information.

[0083] S453, perform verification result determination processing on the consistency determination result information to obtain the s-th verification result information.

[0084] It should be noted that the verification result determination process refers to reading the judgment conclusion (consistent or inconsistent) from the consistency judgment result information, writing the judgment conclusion into the verification result field of the verification result information, and obtaining the s-th verification result information. The s-th verification result information is a data structure containing a verification result field, and the value of the verification result field is 1 (consistent) or 0 (inconsistent), consistent with the judgment conclusion in the consistency judgment result information. When the consistency judgment result information also contains judgment basis or correction suggestions, the judgment basis or correction suggestions can be written into the s-th verification result information or associated storage for use in subsequent correction description generation processing. The above verification result determination process can be implemented through field parsing and assignment, structured output parsing, or key-value mapping, etc., and the specific implementation is not limited in this embodiment of the invention.

[0085] In an optional embodiment, processing the verified feature attribution result information set and the risk prediction result information to obtain the target risk assessment report information includes: S51, Perform a summary generation process on the risk prediction result information to obtain risk prediction summary information; It should be noted that the risk prediction summary information includes core information such as risk type, risk probability value, risk level, and prediction time window. For example, "The probability of this user experiencing MODS within the next 24 hours is 85%, and the risk level is high risk." The above summary generation process can be performed through methods such as field filling of a preset template or text concatenation based on conditional configuration; specifically, this embodiment of the invention does not limit the specific methods.

[0086] S52, perform detail generation processing on the verified feature attribution result information set to obtain feature analysis detail information; It should be noted that the feature analysis details include the numerical values, contribution values, contribution directions, and medical explanations of each key feature. The medical explanations are generated by a large language model based on feature attribution results and literature evidence, transforming abstract numerical indicators into natural language descriptions that conform to medical logic. The above-mentioned detail generation process can be performed through template-based natural language generation or knowledge graph-based reasoning descriptions, etc. Specifically, this embodiment of the invention does not limit the specific methods used.

[0087] S53, perform inconsistent feature filtering processing on the verified feature attribution result information set to obtain correction warning information; It should be noted that the correction warning information is used to remind operators to pay attention to the corrected feature attribution results. When the original attribution conclusion of the machine learning model conflicts with medical knowledge, the correction warning information will clearly indicate the conflicting content and the basis for correction, facilitating verification and judgment by operators. The above-mentioned inconsistent feature screening processing can be carried out based on filtering based on conditional expressions or extraction based on index sets, etc., and the specific implementation of this invention is not limited thereto.

[0088] S54, perform intervention scheme retrieval processing on the target prediction model type information to obtain recommended scheme information; It should be noted that the recommended treatment information is reference information generated based on intervention protocols retrieved from a medical literature knowledge base. For example, for high-risk patients with Sepsis, it might be recommended to "perform 30 ml / kg crystalloid resuscitation within 3 hours, according to the SSC 2021 guidelines." This recommended treatment information is for supplementary reference only. The above-mentioned intervention protocol retrieval process can be performed through vector similarity retrieval, keyword matching retrieval, or queries based on disease type and protocol mapping tables, etc. Specifically, this embodiment of the invention does not limit the specific methods used.

[0089] S55, the risk prediction summary information, the feature analysis details information, the correction warning information and the recommended scheme information are combined and processed to generate the target risk assessment report information.

[0090] It should be noted that the above-mentioned combination generation process can be processed by calling the large language model API interface to convert structured data into natural language reports, concatenating fields based on preset report templates, or assembling structured data based on formats such as JSON / XML. Specifically, the embodiments of the present invention do not limit the specifics.

[0091] As can be seen, the critical illness risk prediction method based on the fusion of LLM and machine learning described in the embodiments of the present invention provides medical personnel with comprehensive and interpretable risk assessment information by generating a structured risk assessment report and integrating the prediction results of machine learning, feature attribution analysis and medical literature verification results.

[0092] Example 2 Please see Figure 2 , Figure 2 This is a schematic diagram of a critical illness risk prediction device based on the fusion of LLM and machine learning, as disclosed in an embodiment of the present invention. Figure 2 The described critical illness risk prediction device based on the fusion of LLM and machine learning is applied to medical data processing systems, such as local servers or cloud servers in medical institution data centers, etc., and the embodiments of this invention are not limited thereto. Figure 2 As shown, this critical illness risk prediction device based on the fusion of LLM and machine learning includes: The acquisition module 201 is used to acquire the user's physiological parameter data information set; The first processing module 202 is used to preprocess and task routing the physiological parameter data information set to obtain the preprocessed physiological parameter data information set and target prediction model type information. The second processing module 203 is used to perform risk prediction and feature attribution analysis on the preprocessed physiological parameter data information set and the target prediction model type information to obtain risk prediction result information and feature attribution result information set respectively. The consistency verification module 204 is used to perform medical literature consistency verification processing on the feature attribution result information set and the target prediction model type information to obtain the verified feature attribution result information set respectively. The report generation module 205 is used to process the verified feature attribution result information set and the risk prediction result information to obtain the target risk assessment report information.

[0093] As can be seen, the critical illness risk prediction device based on the fusion of LLM and machine learning described in the embodiments of the present invention realizes a complete processing flow of machine learning prediction, feature attribution analysis and medical literature consistency verification through the collaborative work of various functional modules, and can output a reliable and interpretable critical illness risk assessment report.

[0094] Example 3 Please see Figure 3 , Figure 3 This is a schematic diagram of another critical illness risk prediction device based on the fusion of LLM and machine learning disclosed in an embodiment of the present invention. Figure 3The described critical illness risk prediction device based on the fusion of LLM and machine learning is applied to medical data processing systems, such as local servers or cloud servers in medical institution data centers, etc., and the embodiments of this invention are not limited thereto. Figure 3 As shown, this critical illness risk prediction device based on the fusion of LLM and machine learning includes: Processor 301; A memory 302 containing executable program code is coupled to the processor 301; The processor 301 calls the executable program code stored in the memory 302 to execute some or all of the steps of the critical illness risk prediction method based on the fusion of LLM and machine learning in Embodiment 1.

[0095] As can be seen, the critical illness risk prediction device based on the fusion of LLM and machine learning described in the embodiments of the present invention can realize the complete processing flow of the critical illness risk prediction method by executing the program code in the memory through the processor, and has high versatility and scalability.

[0096] Example 4 This invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements some or all of the steps of the critical illness risk prediction method based on the fusion of LLM and machine learning described in Embodiment 1.

[0097] Example 5 This invention also provides a computer program product that, when run on a computer, causes the computer to perform some or all of the steps of the critical illness risk prediction method based on the fusion of LLM and machine learning described in Embodiment 1.

[0098] The system embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0099] Through the detailed description of the above embodiments, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electronically erasable rewritable read-only memory (EEPROM), read-only optical disc (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium that can be used to carry or store data.

[0100] Finally, it should be noted that the critical illness risk prediction method and device based on the fusion of LLM and machine learning disclosed in the embodiments of the present invention are merely preferred embodiments of the present invention and are only used to illustrate the technical solutions of the present invention, not to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the risk of severe illness based on the fusion of LLM and machine learning, characterized in that, The method includes: S1, Obtain the user's physiological parameter data set; S2, preprocess and task routing process the physiological parameter data information set to obtain the preprocessed physiological parameter data information set and target prediction model type information; S3, perform risk prediction and feature attribution analysis on the preprocessed physiological parameter data information set and the target prediction model type information to obtain risk prediction result information and feature attribution result information set respectively; S4, Perform medical literature consistency verification on the feature attribution result information set and the target prediction model type information to obtain the verified feature attribution result information set respectively; S5, process the verified feature attribution result information set and the risk prediction result information to obtain the target risk assessment report information.

2. The critical illness risk prediction method based on the fusion of LLM and machine learning according to claim 1, characterized in that, The preprocessing and task routing of the physiological parameter data set yields a preprocessed physiological parameter data set and target prediction model type information, including: S21, perform data cleaning processing on the physiological parameter data information set to obtain the first physiological parameter data information set; S22, perform missing value filling processing on the first physiological parameter data information set to obtain the second physiological parameter data information set; S23, normalize the second physiological parameter data information set to obtain a preprocessed physiological parameter data information set; S24, Perform task routing processing on the preprocessed physiological parameter data information set to obtain target prediction model type information.

3. The critical illness risk prediction method based on the fusion of LLM and machine learning according to claim 2, characterized in that, The normalization process performed on the second physiological parameter data set to obtain a preprocessed physiological parameter data set includes: Using the normalized first calculation model, each physiological parameter data in the second physiological parameter data set is normalized to obtain the preprocessed physiological parameter data set. The first normalized calculation model is as follows: ; ; ; In the formula, The normalized value of the physiological parameter of the j-th physiological parameter data in the preprocessed physiological parameter data set. The physiological parameter value of the j-th physiological parameter data in the second physiological parameter data set. Let be the mean of the j-th physiological parameter data in a preset reference dataset. Let be the standard deviation of the j-th physiological parameter data in the preset reference dataset. and These are the first and third quartiles of the j-th physiological parameter data in a preset reference dataset, respectively. Let be the kurtosis value of the j-th physiological parameter data in a preset reference dataset. For adaptive fusion weights, The sigmoid scaling factor. Kurtosis sensitivity coefficient Kurtosis threshold It is a smoothing constant. M represents the total amount of physiological parameter data.

4. The critical illness risk prediction method based on the fusion of LLM and machine learning according to claim 2, characterized in that, The step of performing task routing processing on the preprocessed physiological parameter data set to obtain target prediction model type information includes: S241, The preprocessed physiological parameter data information set is extracted and processed to obtain the key indicator data information set; S242, perform matching processing on the key indicator data information set and the preset indicator threshold condition set to obtain the matching result information set; S243, perform model type determination processing on the matching result information set to obtain target prediction model type information.

5. The critical illness risk prediction method based on the fusion of LLM and machine learning according to claim 1, characterized in that, The process of performing risk prediction and feature attribution analysis on the preprocessed physiological parameter data set and the target prediction model type information yields risk prediction results and feature attribution results sets, respectively, including: S31, Based on the preset machine learning model pool, perform model loading processing on the target prediction model type information to obtain a machine learning prediction model; S32, using the machine learning prediction model, perform model inference processing on the preprocessed physiological parameter data information set to obtain the risk probability value; S33, Perform risk level classification on the risk probability value to obtain risk level information; S34, integrate the risk probability value and the risk level information to obtain risk prediction result information; S35, using the machine learning prediction model, perform feature attribution analysis on the preprocessed physiological parameter data information set to obtain a feature attribution result information set.

6. The critical illness risk prediction method based on the fusion of LLM and machine learning according to claim 5, characterized in that, The process of using the machine learning prediction model to perform feature attribution analysis on the preprocessed physiological parameter data set yields a feature attribution result set, including: S351, using a feature attribution calculation model, feature attribution analysis is performed on the machine learning prediction model and the preprocessed physiological parameter data information set to obtain attribution value information; The feature attribution calculation model is as follows: ; In the formula, The attribution value information is the first The attribution values ​​of feature i are given by F, where F is the set of all features and S is a subset of features excluding feature i. The output value of the machine learning prediction model when feature i is included. The output value of the machine learning prediction model when feature i is not included. Let S be the number of features in subset S. The total number of features; S352, Perform interactive enhancement processing on the attribution value information to obtain enhanced attribution value information; the enhanced attribution value information includes several attribution contribution values; S353, sort the enhanced attribution value information to obtain sorted attribution value information; S354, Select the top N features and their corresponding attribution contribution values ​​from the sorted attribution value information, and integrate the top N features and their corresponding attribution contribution values ​​with the preprocessed physiological parameter data information set to obtain the feature attribution result information set.

7. The critical illness risk prediction method based on the fusion of LLM and machine learning according to claim 1, characterized in that, The process of processing the verified feature attribution result information set and the risk prediction result information to obtain the target risk assessment report information includes: S51, Perform a summary generation process on the risk prediction result information to obtain risk prediction summary information; S52, perform detail generation processing on the verified feature attribution result information set to obtain feature analysis detail information; S53, perform inconsistent feature filtering processing on the verified feature attribution result information set to obtain correction warning information; S54, perform intervention scheme retrieval processing on the target prediction model type information to obtain recommended scheme information; S55, the risk prediction summary information, the feature analysis details information, the correction warning information and the recommended scheme information are combined and processed to generate the target risk assessment report information.

8. A critical illness risk prediction device based on the fusion of LLM and machine learning, characterized in that, The device includes: The acquisition module is used to acquire the user's physiological parameter data set; The first processing module is used to preprocess and task routing the physiological parameter data information set to obtain the preprocessed physiological parameter data information set and target prediction model type information. The second processing module is used to perform risk prediction and feature attribution analysis on the preprocessed physiological parameter data information set and the target prediction model type information to obtain risk prediction result information and feature attribution result information set respectively. The consistency verification module is used to perform medical literature consistency verification on the feature attribution result information set and the target prediction model type information to obtain the verified feature attribution result information set respectively. The report generation module is used to process the verified feature attribution result information set and the risk prediction result information to obtain the target risk assessment report information.

9. A critical illness risk prediction device based on the fusion of LLM and machine learning, characterized in that, The device includes: Processor 301; A memory 302 containing executable program code is coupled to the processor 301; The processor 301 calls the executable program code stored in the memory 302 to execute the critical illness risk prediction method based on the fusion of LLM and machine learning as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions, which, when invoked, are used for the critical illness risk prediction method based on the fusion of LLM and machine learning as described in any one of claims 1-7.

Images