A data processing method for clinical risk prediction

Abstract

The application provides a kind of data processing method for clinical risk prediction, comprising: obtaining the initial physiological state data of target object;Age stratification baseline correction is carried out on vital sign index based on age attribute to obtain age correction derived feature;Composite inflammation ratio feature is calculated based on biochemical examination index, and ordered classification derived feature is generated by discretization conversion according to historical distribution quantile cut point;Structural input feature vector is constructed and input into pre-trained risk assessment model to output prediction probability;Risk stratification result is determined based on target threshold strategy;Individualized attribution information is generated based on feature contribution degree analysis model.The application eliminates physiological baseline difference of specific population and index noise, realizes adaptive stratification and interpretable evaluation of risk state.

Description

Technical Field

[0001] This invention relates to the field of medical information processing technology, and in particular to a data processing method for clinical risk prediction. Background Technology

[0002] Currently, in the field of medical risk assessment, severe pneumonia in children is a significant disease type in pediatric hospitalization and pediatric intensive care. Some children with severe pneumonia may experience clinical deterioration events such as respiratory failure, circulatory instability, and the need for mechanical ventilation or escalation to intensive care shortly after admission. Early identification of patients at risk of short-term deterioration in the early stages of admission (especially within 24 hours) is crucial for enhanced monitoring, initiating critical care consultations, optimizing decisions regarding escalation of intensive care, and rationally allocating medical resources.

[0003] Currently, early risk assessment for patients with severe pneumonia upon admission mainly relies on clinicians' experience, changes in vital signs, laboratory test results, and some manual scoring tools. However, these methods still have certain shortcomings in practical application: First, clinical judgment is influenced by the experience of pediatricians and the high workload in acute and critical care settings, resulting in a certain degree of subjectivity and lag in risk identification; second, children, being in a period of growth and development, are easily affected by physiological baselines and random fluctuations in their single vital signs or laboratory indicators, making it difficult to comprehensively and stably reflect the complex clinical state of children; third, some manual scoring tools require additional data entry or manual calculation, making it difficult to fully integrate into routine clinical workflows; fourth, the output of traditional risk assessment models is mostly black-box results, making it difficult to simultaneously provide individualized risk probabilities, adaptive risk levels, and attribution explanations for key influencing factors, thus limiting their application in clinical communication and escalation of monitoring and treatment. Therefore, there is an urgent need for a medical risk status data processing method that is suitable for acute and critical care scenarios such as severe pneumonia in children, and can automate, interpret, and visualize adaptive stratified assessment of early clinical deterioration risk within 24 hours of admission based on routinely collected clinical data without increasing the burden of additional clinical data entry. Summary of the Invention

[0004] Therefore, it is necessary to provide a data processing method for clinical risk prediction to address the aforementioned technical problems.

[0005] A data processing method for clinical risk prediction includes the following steps:

[0006] Acquire initial physiological state data of the target object, including demographic attributes, vital signs indicators, and biochemical test indicators;

[0007] Based on the age attribute in the demographic attributes, the vital signs indicators are subjected to age-stratified baseline correction to obtain age-corrected derived features.

[0008] The composite inflammation ratio feature is calculated based on the biochemical test indicators, and the composite inflammation ratio feature is discretized according to the historical distribution quantile cut-off point to generate ordered classification derivative features;

[0009] Based on the age-corrected derived features, the ordered classification derived features, and other structured indicators in the initial physiological state data, a structured input feature vector is constructed.

[0010] The structured input feature vector is input into a pre-trained risk assessment model, which outputs the predicted probability of the target object experiencing a target risk event within a preset time window.

[0011] Based on the currently selected target threshold strategy and the predicted probability, the risk stratification result of the target object is determined.

[0012] In one embodiment, after determining the risk stratification result of the target object based on the currently selected target threshold strategy and the predicted probability, the method further includes:

[0013] Based on the feature contribution analysis model, individualized attribution information that influences the predicted probability is generated.

[0014] In one embodiment, the vital signs indicators are subjected to age-stratified baseline correction based on the age attribute among the demographic attributes to obtain age-corrected derived features, including:

[0015] Determine the corresponding target age stratification based on the age attribute of the target object;

[0016] Obtain the historical statistical mean and historical statistical standard deviation corresponding to the target age stratification;

[0017] The difference between the vital signs index and the historical statistical mean is divided by the historical statistical standard deviation to obtain the age-corrected value, which is used as the age-corrected derived feature.

[0018] In one embodiment, the composite inflammation ratio feature is discretized according to the historical distribution quantile cut-off point to generate ordered classification-derived features, including:

[0019] Obtain the first percentile cutoff point, the second percentile cutoff point, and the third percentile cutoff point based on the distribution of the historical training set;

[0020] Based on the relationship between the composite inflammation ratio feature and the first percentile cutoff point, the second percentile cutoff point, and the third percentile cutoff point, the composite inflammation ratio feature is divided into four ordered levels to obtain ordered classification derived features.

[0021] In one embodiment, prior to constructing the structured input feature vector, the method further includes:

[0022] Outlier identification is performed on continuous physiological indicators in the initial physiological state data;

[0023] Based on the percentile thresholds determined by the distribution of the historical training set, bilateral truncation or unilateral upper tail truncation is performed on the identified abnormal continuous physiological indicators.

[0024] For physiological indicators with missing data, a random forest imputation model based on historical training sets is used to impute missing values.

[0025] Numerical coding and conversion of categorical physiological indicators.

[0026] In one embodiment, the risk assessment model is the LightGBM model based on gradient boosting decision trees.

[0027] In one embodiment, determining the risk stratification result of the target object based on the currently selected target threshold strategy and the predicted probability includes:

[0028] In response to receiving a strategy selection instruction, one of the balanced threshold strategy, the high-sensitivity threshold strategy, and the high-specificity threshold strategy is determined as the target threshold strategy;

[0029] Obtain the target risk threshold corresponding to the target threshold strategy;

[0030] When the predicted probability is greater than or equal to the target risk threshold, the risk stratification result is determined to be a high-risk state; when the predicted probability is less than the target risk threshold, the risk stratification result is determined to be a low-risk state.

[0031] In one embodiment, based on a feature contribution analysis model, individualized attribution information affecting the predicted probability is generated, including:

[0032] The contribution value of each input feature in the structured input feature vector is calculated using the feature contribution analysis model.

[0033] The input features with a contribution value greater than zero are identified as the first target feature identifiers that boost the prediction probability, and the corresponding contribution direction is identified as a positive contribution.

[0034] The input features with a contribution value less than zero are identified as the second target feature identifiers that reduce the prediction probability, and the corresponding contribution direction is identified as a negative contribution.

[0035] Based on the first target feature identifier, the second target feature identifier, and the absolute value of their corresponding contribution values, the individualized attribution information, which includes feature ranking and contribution direction, is generated.

[0036] In one embodiment, the feature contribution analysis model is the SHAP model, and the contribution value is the SHAP value.

[0037] A data processing system for clinical risk prediction, used to implement the data processing method for clinical risk prediction as described above, includes:

[0038] The data acquisition module is used to acquire the initial physiological state data of the target object, which includes demographic attributes, vital signs indicators and biochemical test indicators.

[0039] The feature correction module is used to perform age-stratified baseline correction on the vital signs indicators based on the age attribute in the demographic attributes, and obtain age-corrected derived features.

[0040] The feature derivation module is used to calculate the composite inflammation ratio feature based on the biochemical test indicators, and to discretize the composite inflammation ratio feature according to the historical distribution quantile cut-off point to generate ordered classification derivation features;

[0041] The feature construction module is used to construct a structured input feature vector based on the age-corrected derived features, the ordered classification derived features, and other structured indicators in the initial physiological state data.

[0042] The risk assessment module is used to input the structured input feature vector into a pre-trained risk assessment model and output the predicted probability of the target object experiencing a target risk event within a preset time window.

[0043] The risk stratification module is used to determine the risk stratification result of the target object based on the currently selected target threshold strategy and the predicted probability.

[0044] Compared to existing technologies, the advantages and beneficial effects of this invention are as follows: This invention provides a data processing method for medical risk status. By performing stratified baseline correction of vital sign indicators based on age attributes and discretizing composite inflammation ratio features using historical distribution quantile cutoff points, it eliminates physiological baseline differences caused by the growth and development of specific populations and random noise from single indicators at the data source. This constructs a more robust structured input feature vector, thereby improving the accuracy and stability of the risk assessment model prediction. By supporting the dynamic selection of multiple target threshold strategies such as balanced, high-sensitivity, and high-specificity, the risk stratification results can adaptively match the differentiated decision-making needs under different medical resource conditions and application scenarios. By introducing a feature contribution analysis model to generate individualized attribution information including feature ranking and contribution direction, it breaks the black-box limitation of the model and enhances the interpretability and clinical understandability of the assessment results. Furthermore, by solidifying preprocessing parameters, thresholds, and model parameters into fixed parameters predetermined based on historical training sets, it effectively prevents information leakage in the actual application stage and ensures the generalization ability and reliability of the model in cross-time and cross-hospital validation. Attached Figure Description

[0045] Figure 1 This is a flowchart illustrating a data processing method for clinical risk prediction in one embodiment.

[0046] Figure 2 This is a schematic diagram of the structure of a data processing system for clinical risk prediction in one embodiment. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0048] It should be noted that, unless otherwise defined, the technical or scientific terms used in one or more embodiments of this specification should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in one or more embodiments of this specification do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word covers the element or object listed following the word and its equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0049] In one embodiment, such as Figure 1 As shown, a data processing method for clinical risk prediction is provided, including the following steps:

[0050] Step S101: Obtain the initial physiological state data of the target object, which includes demographic attributes, vital signs indicators and biochemical test indicators.

[0051] Specifically, the target group refers to individuals requiring risk status assessment, and the initial physiological status data refers to the raw clinical data collected within a preset assessment time window. In this embodiment, the target group is hospitalized children with severe pneumonia. The raw clinical data in this embodiment can be automatically captured through hospital information systems and electronic medical record systems, or manually entered through terminal devices. Demographic attributes include at least basic information such as age and gender; vital signs indicators include, but are not limited to, parameters reflecting basic physiological functions such as body temperature, heart rate, respiratory rate, blood pressure, and blood oxygen saturation; biochemical test indicators cover laboratory test results such as complete blood count, liver and kidney function, electrolytes, coagulation function, and inflammatory markers. Standardized acquisition of multi-source heterogeneous data provides a complete and timely data foundation for subsequent feature engineering, ensuring the comprehensiveness and objectivity of the information used for risk assessment.

[0052] Step S102: Based on the age attribute in the demographic attributes, perform age-stratified baseline correction on the vital signs indicators to obtain age-corrected derived features.

[0053] Specifically, based on the age attribute in the demographic attributes, the vital signs indicators are subjected to age-stratified baseline correction to eliminate the interference of differences in children's age-related physiological baselines, thereby obtaining standardized age-corrected derived features.

[0054] Based on this, age-stratified baseline correction is performed on the vital signs indicators according to the age attribute in the demographic attributes to obtain age-corrected derived features, including:

[0055] Determine the corresponding target age stratification based on the age attribute of the target object;

[0056] Obtain the historical statistical mean and historical statistical standard deviation corresponding to the target age stratification;

[0057] The difference between the vital signs index and the historical statistical mean is divided by the historical statistical standard deviation to obtain the age-corrected value, which is used as the age-corrected derived feature.

[0058] Specifically, the age stratification is matched according to the age attributes of the target children; among them, the heart rate index adopts the Pediatric Advanced Life Support (PALS) age stratification, and the respiratory rate index adopts the World Health Organization (WHO) pneumonia-specific age stratification.

[0059] The statistical mean μ and standard deviation σ of the historical training set corresponding to the target age stratification are retrieved. All statistical parameters are determined and fixed only based on the training set and do not participate in the iteration of the validation set and actual application data throughout the process to avoid information leakage.

[0060] By correction formula Age-corrected values ​​were calculated to generate age-corrected derived features, including age-corrected heart rate (HR_AgeZ) and age-corrected respiratory rate (Breathe_AgeZ); where X represents the child's original vital sign measurements. This age-corrected value characterizes the standard deviation of the target subject's current physiological indicators from the average level of the same age group, thus mathematically eliminating the physiological baseline drift caused by aging. This makes target subjects of different ages comparable within the same feature space, providing a pure pathological signal free from age-related confounding factors for subsequent risk assessment models.

[0061] Step S103: Calculate the composite inflammation ratio feature based on the biochemical test indicators, and discretize the composite inflammation ratio feature according to the historical distribution quantile cut-off point to generate ordered classification derivative features.

[0062] Specifically, single biochemical test indicators are often significantly affected by individual differences, testing errors, and transient physiological states, resulting in high levels of random noise. Based on biochemical test indicators, composite inflammation ratios are calculated, and these features are discretized according to fixed distribution quantile cutoff points in the historical training set to generate ordered classification-derived features. Composite inflammation ratio features include the C-reactive protein / albumin ratio (CAR), the derived neutrophil-to-lymphocyte ratio (dNLR), and the C-reactive protein / lymphocyte ratio (CLR). A composite inflammation ratio can only be calculated if all components of the ratio have been collected and their results are available during the initial assessment upon admission. If any component is missing or not updated, the composite inflammation ratio is directly marked as missing.

[0063] When all components of the composite inflammation ratio feature are available and their results are visible, the corresponding composite inflammation ratio feature is calculated. If any component is missing or not visible, the corresponding composite inflammation ratio feature is treated as missing. The composite inflammation ratio feature is a higher-order feature derived from multiple original biochemical indicators through mathematical operations. In actual clinical data collection, some indicators are often missing due to differences in the combination of test items, sample hemolysis, or instrument malfunction.

[0064] Based on this, the composite inflammation ratio feature is discretized according to the historical distribution quantile cut-off points to generate ordered classification-derived features, including:

[0065] Obtain the first percentile cutoff point, the second percentile cutoff point, and the third percentile cutoff point based on the distribution of the historical training set;

[0066] Based on the relationship between the composite inflammation ratio feature and the first percentile cutoff point, the second percentile cutoff point, and the third percentile cutoff point, the composite inflammation ratio feature is divided into four ordered levels to obtain the ordered classification derived feature.

[0067] Specifically, three fixed cut points—the 25th percentile, 50th percentile, and 75th percentile—are retrieved based on a pre-determined historical training set.

[0068] The continuous composite inflammation ratio feature was compared with the three percentile cut-off points one by one, and divided into four ordered levels: Q1, Q2, Q3, and Q4, forming ordered classification derived features. All percentile cut-off points were fixed and reused throughout the process to ensure data processing consistency.

[0069] It is particularly important to emphasize that the historical statistical mean and standard deviation relied upon for the age stratification baseline correction, as well as the first, second, and third percentile cutoff points relied upon for the discretization transformation, are all parameters pre-determined based on the historical training set and remain fixed in the current application stage. This means that when processing new target object data, these statistical parameters will not be recalculated or updated based on the new input data; instead, the offline-fixed parameter values ​​will be directly invoked. This parameter fixing mechanism is a key technical means to ensure data processing consistency and prevent information leakage. It ensures that no matter when the target object enters the evaluation process, its feature transformation rules are strictly aligned with the distribution patterns learned during model training, avoiding feature semantic shifts caused by real-time data distribution drift, thereby guaranteeing the stability and reliability of risk assessment results in long-term applications.

[0070] Step S104: Construct a structured input feature vector based on the age-corrected derived features, the ordered classification derived features, and other structured indicators in the initial physiological state data.

[0071] Specifically, the initial physiological state data is standardized and preprocessed, and then a structured input feature vector with uniform dimensions and standardized format is constructed based on the preprocessed age-corrected derived features, ordered classification derived features, and other structured indicators in the initial physiological state data.

[0072] Based on this, before constructing the structured input feature vector, the following is also included:

[0073] Outlier identification is performed on continuous physiological indicators in the initial physiological state data;

[0074] Based on the percentile thresholds determined by the distribution of the historical training set, bilateral truncation or unilateral upper tail truncation is performed on the identified abnormal continuous physiological indicators.

[0075] For physiological indicators with missing data, a random forest imputation model based on historical training sets is used to impute missing values.

[0076] Numerical coding and conversion of categorical physiological indicators.

[0077] Specifically, based on the fixed percentile thresholds of the training set, the vital signs indicators are subjected to double-tailed truncation (the 1st and 99th percentiles are the thresholds), and the laboratory biochemical indicators are subjected to single-tailed truncation (the 99th percentile is the threshold) to remove interference from extreme outlier data.

[0078] Invalid variables with a missing rate exceeding 30% were removed, and the remaining missing variables were imputed using the missForest random forest imputation model built based on the training set.

[0079] Ordinal categorical variables are coded in an ordered manner according to their clinical risk level, while unordered categorical variables are coded using one-hot encoding to complete the numerical conversion.

[0080] The 10 preprocessed basic structured indicators and 5 derived features are integrated to form a standardized input feature vector containing 15 core features. These 15 core features include D-dimer (DD), CAR, dNLR, HR_AgeZ, CLR, International Normalized Ratio (INR), Procalcitonin (PCT), Uric Acid (UA), Total Carbon Dioxide (TCO2), Fibrinogen (FIB), Lactate (Lac), Activated Partial Thromboplastin Time (APTT), pH, Lactate Dehydrogenase (LDH), and Breathe_AgeZ.

[0081] Step S105: Input the structured input feature vector into the pre-trained risk assessment model and output the predicted probability of the target object experiencing a target risk event within a preset time window.

[0082] Specifically, the structured input feature vector is input into a pre-trained risk assessment model, which outputs an individualized predicted probability of early clinical deterioration in the target child within a preset 24-hour time window; the early clinical deterioration is a composite outcome consisting of unplanned transfer to the ICU within 24 hours of admission and all-cause death.

[0083] Based on this, the risk assessment model is the LightGBM model based on gradient boosting decision trees.

[0084] Specifically, the risk assessment model is a LightGBM lightweight gradient boosting decision tree model optimized through multi-model screening. The model's hyperparameters are determined based on five-fold cross-validation combined with grid search on the training set. The core optimal parameters are: learning rate 0.01, number of leaf nodes 23, maximum tree depth 5, feature sampling ratio 0.6, sample sampling ratio 0.6, L1 regularization parameter 0.5, and L2 regularization parameter 3. The model balances the class imbalance problem through a positive class weight multiplier. This model has undergone internal time-based validation and cross-hospital external validation, demonstrating excellent discrimination, calibration, and generalization capabilities.

[0085] Step S106: Based on the currently selected target threshold strategy and the predicted probability, determine the risk stratification result of the target object.

[0086] Specifically, based on the currently selected target threshold strategy and the predicted probability output by the model, the risk stratification results of the target children are accurately determined; this method is configured with three differentiated preset threshold strategies to adapt to different clinical scenarios.

[0087] Based on this, and using the currently selected target threshold strategy and the predicted probability, the risk stratification result of the target object is determined, including:

[0088] In response to receiving a strategy selection instruction, one of the balanced threshold strategy, the high-sensitivity threshold strategy, and the high-specificity threshold strategy is determined as the target threshold strategy;

[0089] Obtain the target risk threshold corresponding to the target threshold strategy;

[0090] When the predicted probability is greater than or equal to the target risk threshold, the risk stratification result is determined to be a high-risk state; when the predicted probability is less than the target risk threshold, the risk stratification result is determined to be a low-risk state.

[0091] Specifically, the specific layering steps include:

[0092] Receive strategy selection instructions from the clinical operation terminal and select the current target threshold strategy from the balanced threshold strategy, high-sensitivity threshold strategy, and high-specificity threshold strategy;

[0093] The balanced threshold strategy is usually determined based on maximizing the Youden Index on the historical validation set, with an optimal threshold of 0.127; the high-sensitivity threshold is determined based on a preset target sensitivity of 0.95, with an optimal threshold of 0.036; the high-specificity threshold is determined based on a preset target specificity of 0.95, with an optimal threshold of 0.423; all thresholds are fixed during the training phase and are not adjusted in actual applications.

[0094] Risk assessment: When the predicted probability is greater than or equal to the current target risk threshold, the risk level is determined to be high; when the predicted probability is less than the current target risk threshold, the risk level is determined to be low.

[0095] It should be noted that the above three strategies are merely illustrative examples. In other embodiments, more types of threshold strategies can be extended according to actual needs. As long as they can respond to external instructions and switch the corresponding judgment logic, they all fall within the protection scope of this invention.

[0096] Based on this, after determining the risk stratification result of the target object according to the currently selected target threshold strategy and the predicted probability, the following steps are also included:

[0097] Based on the feature contribution analysis model, individualized attribution information that influences the predicted probability is generated.

[0098] Specifically, a feature contribution analysis model is used to provide a post-hoc interpretation of the predicted probabilities of the output. This attribution information can quantify the direction and magnitude of the contribution of each input feature to the current prediction result.

[0099] Based on this, and using a feature contribution analysis model, individualized attribution information affecting the predicted probability is generated, including:

[0100] The contribution value of each input feature in the structured input feature vector is calculated using the feature contribution analysis model.

[0101] The input features with a contribution value greater than zero are identified as the first target feature identifiers that boost the prediction probability, and the corresponding contribution direction is identified as a positive contribution.

[0102] The input features with a contribution value less than zero are identified as the second target feature identifiers that reduce the prediction probability, and the corresponding contribution direction is identified as a negative contribution.

[0103] Based on the first target feature identifier, the second target feature identifier, and the absolute value of their corresponding contribution values, the individualized attribution information, which includes feature ranking and contribution direction, is generated.

[0104] Based on this, the feature contribution analysis model is the SHAP model, and the contribution value is the SHAP value.

[0105] Specifically, the explanation steps include:

[0106] The SHAP summation and interpretation model is used to quantify the SHAP contribution value of each feature in the structured input feature vector;

[0107] Features with a SHAP value > 0 are considered positive contribution features, used to increase the probability of clinical deterioration in children; features with a SHAP value < 0 are considered negative contribution features, used to decrease the probability of clinical deterioration in children.

[0108] Based on the absolute value of each feature's SHAP value, the system generates an individualized attribution report that includes core risk impact features, contribution intensity, and contribution direction. It also supports visualization of results through SHAP summary scatter plots and individualized waterfall plots, intuitively presenting the importance of global features and risk drivers for individual patients.

[0109] This invention's risk assessment model utilizes multi-center, large-sample real clinical data to complete the standardized construction, screening, training, and independent validation process. Strict time-series splitting, data isolation, and parameter solidification strategies are implemented throughout to completely prevent information leakage and ensure the model's discriminative power, calibration accuracy, generalization ability, and clinical applicability. The overall model development process includes eleven standardized technical steps (S101–S111). All processing rules, statistical parameters, model hyperparameters, and risk thresholds are determined and reused only on the training set, and are uniformly applied to internal validation, external validation, and actual clinical inference stages.

[0110] S101. Study Subjects and Acquisition of Original Clinical Data

[0111] This retrospective observational study selected hospitalized children with severe pneumonia admitted to the Children's Hospital of Chongqing Medical University from January 1, 2010 to December 31, 2024 as the research subjects. The study retrospectively collected all structured clinical data routinely gathered during the initial assessment phase upon admission. The data types included demographic attributes, vital signs, clinical symptoms, underlying disease markers, and laboratory test results—all accessible at the initial admission visit. After screening using standardized inclusion and exclusion criteria, 15,211 valid hospitalization records were ultimately included for model construction and validation analysis.

[0112] S102, Independent division of research cohorts

[0113] To fully verify the model's time extrapolation generalization ability and cross-hospital portability, this invention adopts a three-cohort independent partitioning scheme: time-split internal validation and cross-hospital external validation. The data in each cohort is completely isolated, non-overlapping, and non-interfering. Specifically, 8234 inpatient records from the Yuzhong campus (2010-2019) are used as the model training set for determining all data rules, feature engineering, feature selection, model training, and hyperparameter optimization. 3754 inpatient records from the Yuzhong campus (2020-2024) serve as the same-hospital time-sequence retention internal validation set, used for candidate model comparison, feature reduction, model selection, and time extrapolation performance evaluation. All 3223 inpatient records from the Liangjiang campus serve as the cross-hospital external validation set, used only for independent generalization performance testing of the final model and not involved in any training or parameter tuning process.

[0114] S103. Unified Definition of Prediction Start Point and Outcome Indicators

[0115] This invention uniformly sets the prediction starting point as the completion time of the initial assessment upon admission, the prediction time window as 24 hours after admission, and the prediction outcome as the composite endpoint of early clinical deterioration, specifically defined as either unplanned transfer to the ICU or all-cause death occurring within 24 hours of admission. Samples exhibiting either of these events are labeled as positive samples, while samples without either event are labeled as negative samples, thus constructing a binary classification supervised learning modeling labeling system. Specifically, out of 15,211 hospitalization records, 1,047 showed early clinical deterioration within 24 hours of admission, an event rate of 6.9%; among these, 1,042 were unplanned transfers to the ICU, and 5 were all-cause deaths within 24 hours of admission. The event rates for the development cohort and the cross-hospital external validation cohort were 6.6% and 7.9%, respectively; the event rates for the training set and the same-hospital time-sequence internal validation set were 7.0% and 5.7%, respectively.

[0116] S104. Standardized Preprocessing of Comprehensive Clinical Data

[0117] This invention, based on the statistical results of the training set data distribution, unifies and solidifies a complete set of data preprocessing rules, which are then reused without distinction in the validation set and actual inference data, achieving full-process data standardization. The preprocessing process covers standardizing indicator names, normalizing units, converting data types, controlling outliers, handling missing values, and encoding categorical variables. For outliers, a two-sided truncation strategy based on the 1st and 99th percentiles of the training set is used for vital signs such as heart rate and respiratory rate, while a one-sided upper-tail truncation strategy based on the 99th percentile of the training set is used for various biochemical test indicators to eliminate interference from extreme outlier samples. For missing data, invalid variables with a missing proportion greater than 30% are uniformly removed. For the remaining valid variables with missing values, a missForest random forest imputation model pre-fitted in the training set is used for standardization and imputation. For categorical variables, ordered encoding is performed based on clinical risk gradients, and one-hot encoding is used to convert unordered categorical indicators to numerical values, unifying the format of all input data.

[0118] S105. Multi-dimensional candidate feature extraction and derived feature construction

[0119] Structured clinical variables, first available during the initial admission assessment phase, were extracted from preprocessed clinical data as candidate predictive variables. These candidate predictive variables were derived from routine clinical data collected during the initial admission assessment phase, including demographic information, vital signs, underlying disease information, and biochemical indicators. Demographic information included age and sex; vital signs included body temperature, heart rate, respiratory rate, blood pressure, and blood oxygen saturation; underlying disease information was identified based on ICD-10 codes related to previous illnesses and was categorized as 0, 1, or greater than 1 based on the number of underlying diseases; and biochemical indicators included complete blood count, biochemical markers, inflammatory markers, liver function, kidney function, electrolytes, coagulation function, and blood gas analysis indicators.

[0120] A composite inflammatory index is calculated based on laboratory test results that are available and visible during the initial assessment phase upon admission, and age-corrected vital signs variables are constructed as candidate derived features. The composite inflammatory index includes, but is not limited to, the C-reactive protein / albumin ratio (CAR), the derived neutrophil-to-lymphocyte ratio (dNLR), and the C-reactive protein / lymphocyte ratio (CLR).

[0121] Using the 25th, 50th, and 75th percentiles in the training set as cutoff points, the continuous composite inflammation index was divided into four levels: Q1, Q2, Q3, and Q4, and these levels were incorporated as ordered categorical variables into subsequent feature selection and modeling. These quartile cutoff points were determined only in the training set and remained fixed across the internal validation set, external validation set, and practical application phase.

[0122] HR_AgeZ is calculated according to the Pediatric Advanced Life Support (PALS) age stratification, and Breathe_AgeZ is calculated according to the World Health Organization (WHO) pneumonia-related age stratification. For the [missing information]... For example, for a child with a specific vital sign within their age stratum, the value is... The mean of the training set for the corresponding age group is The standard deviation is Calculate the age-corrected Z-scores of corresponding vital signs according to the preset age strata:

[0123]

[0124] To ensure consistency in the predicted time series, only variables that are available and visible at the initial admission assessment stage are included. Variables not available at this stage, or those collected but not visible at the time of the initial assessment, are treated as missing. It should be noted that the number of initially entered variables may exceed the number of features input to the final model. Some of the original variables are used to calculate the composite inflammation index or age-corrected vital signs variables. After preprocessing, derived feature calculation, quartile transformation, and feature selection, the final input feature vector required by the model is formed.

[0125] S106: Two-stage feature screening combined with clinical prior constraints

[0126] Predefined candidate variables are pre-screened using the Boruta algorithm. The Boruta algorithm identifies candidate variables associated with early clinical deterioration outcomes by constructing randomly shuffled shadow variables corresponding to the original variables and comparing the importance of the original and shadow variables in the random forest model. For each original candidate variable, a randomly permuted shadow variable is generated. Both the original and shadow variables are then input into the random forest model, and an importance score is calculated for each variable. If the importance score of an original variable is significantly higher than that of its shadow variable over multiple iterations, the variable is considered important and retained. If the importance score of an original variable is lower than that of its shadow variable, the variable is removed. Variables with unstable outcomes are considered provisional variables, and their retention is determined based on subsequent screening results and clinical significance.

[0127] Based on the Boruta algorithm's pre-screening, LASSO regression is used to further compress features. LASSO regression uses early clinical deterioration within 24 hours of admission as a binary outcome variable. By introducing an L1 regularization penalty term, the coefficients of some variables are compressed to 0, thus achieving variable screening. The optimal penalty parameter is determined through five-fold cross-validation. In a preferred embodiment, the solution corresponding to lambda.1se is selected as the data-driven feature selection result to improve model simplicity and stability.

[0128] Based on the data-driven screening results, and further incorporating prior clinical knowledge and its potential importance in assessing acute deterioration in pediatric patients, lactate, international normalized ratio, and blood urea nitrogen were included as a definitive feature set in subsequent model development. It should be noted that this clinically defined feature set is used for candidate model development and comparison; the final inclusion of these variables in the model is determined by a comprehensive evaluation of subsequent model performance, feature reduction results, and clinical deployment feasibility.

[0129] After feature selection, multiple candidate machine learning models are constructed based on the training set. These candidate models include, but are not limited to, CatBoost, LightGBM, XGBoost, GBM, Random Forest, AdaBoost, Logistic Regression, Support Vector Machine, KNN, and Neural Networks. Each candidate model is trained using clinical features available during the initial assessment of the child upon admission as input, and early clinical deterioration within 24 hours of admission as the predictive label. Early clinical deterioration is defined as a composite outcome, including unplanned transfer to the ICU or all-cause death within 24 hours of admission. Hospitalization records showing either of these outcomes are defined as events, while hospitalization records not showing any of these outcomes are defined as non-events.

[0130] The hyperparameters of the candidate models were optimized using a combination of five-fold cross-validation and grid search. Model training and hyperparameter optimization were performed solely on the training set. The candidate algorithms were compared primarily based on AUC, with AUPRC used as a supplementary metric for low event rate scenarios. A comprehensive evaluation was also conducted, incorporating calibration performance, decision curve analysis, model complexity, number of variables, and feasibility for clinical deployment.

[0131] S107: Model Comparison and Final Model Selection

[0132] Performance comparisons of candidate models were conducted within a time-series validation set within the same hospital area. A comprehensive evaluation was performed considering model discrimination, calibration accuracy, net clinical benefit, alarm efficiency, model complexity, and clinical deployability. The model with superior overall performance and suitability for clinical deployment was selected as the foundation algorithm for subsequent streamlined modeling. Model performance was evaluated across three dimensions: discrimination, calibration accuracy, and clinical utility. Discrimination metrics included ROC curve, AUC, PRC curve, and AUPRC; AUC was the primary discrimination metric, while AUPRC served as a supplementary metric for low event rate scenarios. Threshold-dependent metrics included accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and F1 score. Calibration accuracy was evaluated using Brier scores and calibration curves. Clinical utility was evaluated using decision curve analysis.

[0133] Ten candidate models were compared using an internal validation set that retained the time sequence within the same campus. The results showed that gradient boosting models such as CatBoost, LightGBM, GBM, and XGBoost generally performed best. Specifically, CatBoost had an AUC of 0.855, LightGBM 0.853, GBM 0.850, and XGBoost 0.845; LightGBM had an AUPRC of 0.363, demonstrating superior performance in low event rate scenarios.

[0134] At the preset Youden threshold, LightGBM's accuracy, sensitivity, specificity, positive predictive value, negative predictive value, F1 score, and Brier score were 0.737, 0.828, 0.732, 0.158, 0.986, 0.265, and 0.052, respectively.

[0135] Among the candidate models, the primary model was determined not solely based on the highest AUC, but rather by comprehensively considering AUC, AUPRC, calibration performance, decision curve, sensitivity requirements, feature reduction stability, alarm efficiency, number of variables, and ease of clinical deployment. Taking all these factors into account, LightGBM was selected as the foundational algorithm for subsequent simplified modeling.

[0136] S108: Feature Reduction and Final Model Determination

[0137] Among the top-performing candidate models, the consistency of SHAP feature ranking is further compared to evaluate the stability of key signals. Within the optimal algorithm framework, nested feature subset models are constructed according to the mean absolute SHAP value from high to low. The discriminative power, precision-recall performance, sensitivity, alarm efficiency, and model simplicity of models with different feature counts are evaluated in the same hospital area and time-series internal validation set.

[0138] After determining LightGBM as the base algorithm, nested feature subset models of different sizes were constructed based on the SHAP ranking results of LightGBM, starting with DD and CAR. As the number of included features increased, the model's discriminative power gradually improved, and then plateaued after the first 15 features.

[0139] Considering predictive performance, the number of variables, sensitivity requirements in early warning scenarios, and feasibility of clinical deployment, the 15-feature LightGBM model was ultimately selected as the primary analysis model; an extended model incorporating more high-ranking features can be used for sensitivity analysis or supplementary validation.

[0140] The final model is a LightGBM model built based on 15 clinical features. These 15 clinical features are shown in Table 1 below:

[0141] 1 DD D-dimer mg / L FEU Continuous variables 2 CAR C-reactive protein / albumin ratio, after quartile conversion - Derived ordered classification features 3 dNLR Absolute neutrophil count / (white blood cell count - absolute neutrophil count), after quartile transformation - Derived ordered classification features 4 HR_AgeZ Age-corrected heart rate Z value Derived continuous variables 5 CLR C-reactive protein / absolute lymphocyte ratio, after quartile transformation - Derived ordered classification features 6 INR International Normalized Ratio - Continuous variables 7 PCT Procalcitonin ng / mL Continuous variables 8 UA uric acid μmol / L Continuous variables 9 TCO2 Total carbon dioxide mmol / L Continuous variables 10 FIB fibrinogen g / L Continuous variables 11 Lac lactic acid mmol / L Continuous variables 12 APTT Activated partial thromboplastin time Second Continuous variables 13 pH pH - Continuous variables 14 LDH lactate dehydrogenase U / L Continuous variables 15 Breathe_AgeZ Age-corrected respiratory rate Z value Derived continuous variables

[0142] Table 1 Clinical characteristics

[0143] Among them, CAR, dNLR, and CLR are ordered classification features calculated from the original laboratory indicators and then transformed by the quartile cut points of the training set; HR_AgeZ and Breathe_AgeZ are age-corrected derived features calculated from age and corresponding vital signs.

[0144] The final risk assessment model is a 15-feature LightGBM model. The hyperparameters of the LightGBM model were determined through five-fold cross-validation combined with grid search, and its search range and optimal parameters are shown in Table 2 below:

[0145] Learning rate 0.001、0.003、0.005、0.01、0.02、0.05 0.01 Number of leaf nodes 7、15、23、31、63 23 Maximum tree depth 3、4、5、6、8 5 Feature sampling ratio 0.4、0.6、0.8、1.0 0.6 Sample sampling ratio 0.4、0.6、0.8、1.0 0.6 Bagging frequency 1、3、5、7 1 Minimum number of samples for leaf nodes 10、20、30、50、80、100 50 Minimum Splitting Gain 0、0.01、0.05、0.1、0.2 0.05 L1 regularization parameter 0、0.5、1、3、5 0.5 L2 regularization parameter 0、1、3、5、10 3 Positive category weight multiplier 0.1、0.2、0.3、0.5、0.8 0.2, pos_weight=2.262038

[0146] Table 2 Search range and optimal parameters

[0147] The positive category weight is used to address the category imbalance caused by the low incidence of early clinical deterioration events within 24 hours of admission.

[0148] A 20-feature LightGBM extended model can also be constructed for sensitivity analysis or to supplement the evaluation of clinical scenarios with more complete variable entries. In an external validation set across multiple hospitals, the 20-feature LightGBM extended model had an AUC of 0.842 (95% CI 0.817–0.866), an AUPRC of 0.372 (95% CI 0.313–0.435), and a Brier score of 0.064. Compared to the 15-feature model, the 20-feature extended model had similar overall discriminative power, but with an increased number of variables; therefore, the 15-feature model was preferred as the primary analysis model.

[0149] S109: Model Validation and Evaluation Metrics

[0150] The final risk assessment model was applied to the same-hospital time-series internal validation set and the cross-hospital external validation set to evaluate the model's discrimination, calibration, net clinical benefit, alarm efficiency, and risk stratification performance.

[0151] For a given risk threshold, the model prediction results can form the following confusion matrix:

[0152] Early clinical deterioration actually occurred TP FN No early clinical deterioration actually occurred. FP TN

[0153] Among them, TP indicates a high-risk condition that actually resulted in early clinical deterioration, FP indicates a high-risk condition that did not actually result in early clinical deterioration, FN indicates a low-risk condition that actually resulted in early clinical deterioration, and TN indicates a low-risk condition that did not actually result in early clinical deterioration.

[0154] The formula for calculating accuracy is:

[0155] The formula for calculating sensitivity is:

[0156] The formula for calculating specificity is:

[0157] The formula for calculating the Positive Predictive Value (PPV) is:

[0158] The formula for calculating the Negative Predictive Value (NPV) is:

[0159] The formula for calculating the F1 value is:

[0160] To evaluate the alarm efficiency of the model in early clinical warning scenarios, the NNA (Number of Alarms Not Detected) metric was used. NNA represents the average number of alarms triggered to identify one truly positive case, calculated according to the hospital record hierarchy, and its formula is: ,Right now .

[0161] The lower the NNA, the fewer alarms are needed to identify one true early clinical deterioration patient, and the higher the clinical alarm efficiency.

[0162] The Brier score is used to evaluate the consistency between predicted probabilities and actual outcomes. The formula for its calculation is as follows: .

[0163] in, This represents the model-predicted probability of the i-th patient. This represents the actual outcome, and n represents the sample size.

[0164] Decision curve analysis was used to evaluate the clinical net benefit of the model at different risk thresholds. Let the threshold probability be... The formula for calculating net profit NB is: .

[0165] Within the same hospital's time-series-preserved internal validation set, the 15-feature LightGBM model achieved an AUC of 0.838 (95% CI: 0.810–0.866) and an AUPRC of 0.324 (95% CI: 0.262–0.392). Compared to the 28-feature full model, the 15-feature model showed only a slight performance decrease, but with a significantly reduced number of variables, making it more suitable for clinical deployment.

[0166] In the cross-hospital external validation set, the 15-feature LightGBM model had an AUC of 0.841 (95% CI 0.817–0.866), an AUPRC of 0.377 (95% CI 0.316–0.441), and a Brier score of 0.064. ROC and PR curves showed that the model maintained good discriminative ability in the external cohort. The calibration curve showed that the predicted risk was generally consistent with the observed risk, and the decision curve showed that the model had a clinical net benefit within a threshold probability range of approximately 8%–45%.

[0167] In an external validation set across multiple hospital campuses, the 15-feature LightGBM model correctly identified 203 hospitalization records with early clinical deterioration and 2,244 hospitalization records without early clinical deterioration at a Youden threshold of 0.127, with 52 missed cases and 724 false positives. The corresponding NNA was 4.6, indicating that approximately 5 alerts were triggered on average to identify one child with actual early clinical deterioration.

[0168] S110: Risk Threshold Determination

[0169] Risk stratification thresholds are determined based on the training set and pre-defined clinical application goals. These thresholds include the Youden threshold, high sensitivity threshold, and high specificity threshold. Each threshold is predetermined during the training phase and remains fixed across the internal validation set, external validation set, and practical application phase.

[0170] In the cross-campus external validation set, the 15-feature LightGBM model, with a Youden threshold of 0.127, achieved an accuracy of 0.759, a sensitivity of 0.796, a specificity of 0.756, a positive predictive value of 0.219, a negative predictive value of 0.977, an F1 score of 0.343, and an NNA of 4.6. At a threshold of 0.036 corresponding to a target sensitivity of 0.95, the model achieved a sensitivity of 0.933, a negative predictive value of 0.988, and an NNA of 7.8, making it suitable for early screening scenarios where minimizing false negatives is crucial. At a threshold of 0.423 corresponding to a target specificity of 0.95, the model achieved a specificity of 0.957, a positive predictive value of 0.426, and an NNA of 2.3 in the cross-campus external validation set, making it suitable for scenarios requiring reduced invalid alarms and improved alarm accuracy.

[0171] The LightGBM model ultimately outputs the predicted probability of early clinical deterioration within 24 hours of admission. Based on pre-determined risk thresholds, patients can be classified into low-risk and high-risk groups to assist clinicians in early risk stratification upon admission, adjustment of monitoring intensity, and further treatment decisions.

[0172] S111: Characteristic Contribution Analysis

[0173] Feature contribution analysis (SCA) methods are used to interpret the final model. SCA methods include, but are not limited to, SHAP, feature importance analysis, local interpretation methods, or other model interpretation techniques. SCA can output global feature importance, key feature contributions for individual child predictions, features that increase risk probability, and features that decrease risk probability.

[0174] The SHAP method was used to explain the prediction basis of the 15-feature LightGBM model. In the cross-campus external validation set, ranked by mean absolute SHAP value, D-dimer (DD) contributed the most to the model output, with a mean absolute SHAP value of 0.489; followed by CAR (0.358); dNLR (0.288); HR_AgeZ (0.284); CLR (0.236); INR (0.217); and PCT (0.216). UA, TCO2, FIB, Lac, APTT, pH, LDH, and Breathe_AgeZ also contributed to the model output.

[0175] SHAP meta-analysis shows that higher DD, CAR, dNLR, HR_AgeZ, CLR, INR, PCT, UA, TCO2, APTT, LDH, and Breathe_AgeZ generally correspond to higher predictive risk; lower FIB and lower pH are also associated with higher predictive risk. At the individual level, the main factors driving the increase or decrease in predictive risk can be shown through SHAP waterfall plots or feature contribution lists, thereby improving the interpretability of the model output results.

[0176] This invention provides a data processing method for medical risk status. It performs stratified baseline correction of vital signs based on age attributes and discretizes the composite inflammation ratio feature by combining historical distribution quantile cutoff points. This eliminates physiological baseline differences caused by growth and development in specific populations and random noise from single indicators at the data source, constructing a more robust structured input feature vector, thereby improving the accuracy and stability of the risk assessment model. By supporting the dynamic selection of multiple target threshold strategies such as balanced, high-sensitivity, and high-specificity, the risk stratification results can adaptively match the differentiated decision-making needs under different medical resource conditions and application scenarios. By introducing a feature contribution analysis model to generate individualized attribution information including feature ranking and contribution direction, it breaks the black-box limitation of the model and enhances the interpretability and clinical understandability of the assessment results. Furthermore, by solidifying preprocessing parameters, thresholds, and model parameters into fixed parameters predetermined based on historical training sets, it effectively prevents information leakage in the actual application stage and ensures the generalization ability and reliability of the model in cross-time and cross-hospital validation.

[0177] It should be noted that the method of this embodiment can be executed by a single device, such as a computer or server. The method of this embodiment can also be applied to a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method of this embodiment, and the multiple devices will interact with each other to complete the method described.

[0178] It should be noted that the above description describes some embodiments of the present invention. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims may be performed in a different order than that shown in the above embodiments and still achieve the desired results. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0179] Based on the same inventive concept, corresponding to any of the above embodiments, the present invention also provides a data processing system for clinical risk prediction.

[0180] refer to Figure 2 The data processing system for clinical risk prediction includes:

[0181] Data acquisition module 201 is used to acquire the initial physiological state data of the target object, the initial physiological state data including demographic attributes, vital signs indicators and biochemical test indicators;

[0182] Feature correction module 202 is used to perform age-stratified baseline correction on the vital signs indicators based on the age attribute in the demographic attributes to obtain age-corrected derived features;

[0183] The feature derivation module 203 is used to calculate the composite inflammation ratio feature based on the biochemical test indicators, and to discretize the composite inflammation ratio feature according to the historical distribution quantile cut-off point to generate ordered classification derivation features;

[0184] The feature construction module 204 is used to construct a structured input feature vector based on the age-corrected derived features, the ordered classification derived features, and other structured indicators in the initial physiological state data.

[0185] The risk assessment module 205 is used to input the structured input feature vector into a pre-trained risk assessment model and output the predicted probability of the target object experiencing a target risk event within a preset time window.

[0186] Risk stratification module 206 is used to determine the risk stratification result of the target object based on the currently selected target threshold strategy and the predicted probability.

[0187] For ease of description, the above system is described by dividing it into various modules based on their functions. Of course, in implementing this invention, the functions of each module can be implemented in one or more software and / or hardware components.

[0188] The system described above is used to implement a data processing method for clinical risk prediction in any of the foregoing embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0189] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention (including the claims) is limited to these examples; within the framework of the invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the embodiments of the invention as described above, which are not provided in the details for the sake of brevity.

[0190] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.

[0191] While specific details have been set forth to describe exemplary embodiments of the invention, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive. Although the invention has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art based on the foregoing description.

[0192] The embodiments of this invention are intended to cover all such substitutions, modifications, and variations falling within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of this invention should be included within the protection scope of this invention.

Claims

1. A data processing method for clinical risk prediction, characterized in that, include: Acquire initial physiological state data of the target object, including demographic attributes, vital signs indicators, and biochemical test indicators; Based on the age attribute in the demographic attributes, the vital signs indicators are subjected to age-stratified baseline correction to obtain age-corrected derived features. The composite inflammation ratio feature is calculated based on the biochemical test indicators, and the composite inflammation ratio feature is discretized according to the historical distribution quantile cut-off point to generate ordered classification derivative features; Based on the age-corrected derived features, the ordered classification derived features, and other structured indicators in the initial physiological state data, a structured input feature vector is constructed. The structured input feature vector is input into a pre-trained risk assessment model, which outputs the predicted probability of the target object experiencing a target risk event within a preset time window. Based on the currently selected target threshold strategy and the predicted probability, the risk stratification result of the target object is determined.

2. The data processing method for clinical risk prediction according to claim 1, characterized in that, After determining the risk stratification result of the target object based on the currently selected target threshold strategy and the predicted probability, the method further includes: Based on the feature contribution analysis model, individualized attribution information that influences the predicted probability is generated.

3. The data processing method for clinical risk prediction according to claim 1, characterized in that, The age-stratified baseline correction of the vital signs indicators based on the age attribute in the demographic attributes yields age-corrected derived features, including: Determine the corresponding target age stratification based on the age attribute of the target object; Obtain the historical statistical mean and historical statistical standard deviation corresponding to the target age stratification; The difference between the vital signs index and the historical statistical mean is divided by the historical statistical standard deviation to obtain the age-corrected value, which is used as the age-corrected derived feature.

4. The data processing method for clinical risk prediction according to claim 1, characterized in that, The discretization transformation of the composite inflammation ratio feature according to the historical distribution quantile cut-off points generates ordered classification derived features, including: Obtain the first percentile cutoff point, the second percentile cutoff point, and the third percentile cutoff point based on the distribution of the historical training set; Based on the relationship between the composite inflammation ratio feature and the first percentile cutoff point, the second percentile cutoff point, and the third percentile cutoff point, the composite inflammation ratio feature is divided into four ordered levels to obtain ordered classification derived features.

5. The data processing method for clinical risk prediction according to claim 1, characterized in that, Before constructing the structured input feature vector, the process also includes: Outlier identification is performed on continuous physiological indicators in the initial physiological state data; Based on the percentile thresholds determined by the distribution of the historical training set, bilateral truncation or unilateral upper tail truncation is performed on the identified abnormal continuous physiological indicators. For physiological indicators with missing data, a random forest imputation model based on historical training sets is used to impute missing values. Numerical coding and conversion of categorical physiological indicators.

6. The data processing method for clinical risk prediction according to claim 1, characterized in that, The risk assessment model is the LightGBM model based on gradient boosting decision trees.

7. The data processing method for clinical risk prediction according to claim 1, characterized in that, The step of determining the risk stratification result of the target object based on the currently selected target threshold strategy and the predicted probability includes: In response to receiving a strategy selection instruction, one of the balanced threshold strategy, the high-sensitivity threshold strategy, and the high-specificity threshold strategy is determined as the target threshold strategy; Obtain the target risk threshold corresponding to the target threshold strategy; When the predicted probability is greater than or equal to the target risk threshold, the risk stratification result is determined to be a high-risk state; when the predicted probability is less than the target risk threshold, the risk stratification result is determined to be a low-risk state.

8. The data processing method for clinical risk prediction according to claim 2, characterized in that, The feature contribution analysis model generates individualized attribution information that influences the predicted probability, including: The contribution value of each input feature in the structured input feature vector is calculated using the feature contribution analysis model. The input features with a contribution value greater than zero are identified as the first target feature identifiers that boost the prediction probability, and the corresponding contribution direction is identified as a positive contribution. The input features with a contribution value less than zero are identified as the second target feature identifiers that reduce the prediction probability, and the corresponding contribution direction is identified as a negative contribution. Based on the first target feature identifier, the second target feature identifier, and the absolute value of their corresponding contribution values, the individualized attribution information, which includes feature ranking and contribution direction, is generated.

9. The data processing method for clinical risk prediction according to claim 8, characterized in that, The feature contribution analysis model is the SHAP model, and the contribution value is the SHAP value.

10. A data processing system for clinical risk prediction, characterized in that, A data processing method for implementing a clinical risk prediction method as described in any one of claims 1-9 includes: The data acquisition module is used to acquire the initial physiological state data of the target object, which includes demographic attributes, vital signs indicators and biochemical test indicators. The feature correction module is used to perform age-stratified baseline correction on the vital signs indicators based on the age attribute in the demographic attributes, and obtain age-corrected derived features. The feature derivation module is used to calculate the composite inflammation ratio feature based on the biochemical test indicators, and to discretize the composite inflammation ratio feature according to the historical distribution quantile cut-off point to generate ordered classification derivation features; The feature construction module is used to construct a structured input feature vector based on the age-corrected derived features, the ordered classification derived features, and other structured indicators in the initial physiological state data. The risk assessment module is used to input the structured input feature vector into a pre-trained risk assessment model and output the predicted probability of the target object experiencing a target risk event within a preset time window. The risk stratification module is used to determine the risk stratification result of the target object based on the currently selected target threshold strategy and the predicted probability.

Images