Lead poisoning post blood lead re-elevation risk prediction system and method based on deep learning

Abstract

This invention relates to a deep learning-based system and method for predicting the risk of re-emergence of blood lead levels after lead poisoning, belonging to the field of intelligent medical technology. The method includes: S1: acquiring raw clinical data of the patient to be evaluated; S2: preprocessing the raw data, including iterative imputation of missing values, nonlinear transformation of skewed continuous variables, and mode imputation and encoding of categorical variables; S3: feature engineering construction, based on the preprocessed data, constructing a low-dimensional engineered feature space with physiological mechanism orientation through cross-dimensional interaction, including the "exposure source-exposure intensity-follow-up trend-organ involvement" chain features and a comprehensive medical index; S4: inputting the constructed features into a pre-trained prediction model, outputting the probability value of re-emergence of blood lead, and performing risk stratification based on the probability value. This invention achieves accurate prediction of re-emergence of blood lead levels in lead poisoning patients after lead chelation therapy.

Description

Technical Field

[0001] This invention belongs to the field of intelligent medical technology and relates to a system and method for predicting the risk of re-emergence of blood lead levels after lead poisoning based on deep learning. Background Technology

[0002] Lead is a silvery-white, soft heavy metal. Due to its high density, low melting point, ease of processing, and good corrosion resistance, it is widely used in fields such as battery manufacturing, gasoline paint, pigments, glazes, plastics, ceramics, explosion-proof materials, and water pipes. However, lead pollution is widespread in human production and daily life, leading to recurrent lead poisoning in children and adults, which has become a critical scientific and technological problem that urgently needs to be solved.

[0003] From a toxicokinetic perspective, lead has an extremely long half-life after entering the human body and can remain in tissues such as bones for decades. Clinically, although chelation therapy can lower blood lead levels in the short term, due to the continuous release of lead from bones (endogenous exposure), patients are highly susceptible to a secondary increase in blood lead levels after treatment. Currently, the clinical management of lead poisoning patients mainly relies on traditional regular check-ups and follow-up recommendations based on physician experience.

[0004] Current techniques typically assess patient recovery based on a single blood lead concentration indicator or simple statistical models (such as linear regression). However, these existing techniques have the following drawbacks and limitations in practical applications: 1) Limited prediction accuracy and difficulty in identifying hidden risks: Traditional assessment methods often only focus on the immediate blood lead value at the end of treatment, ignoring individual differences among patients and the nonlinear correlation between physiological indicators, resulting in a low prediction accuracy for "re-elevation of blood lead" (referring to the recovery of blood lead to 0.96 μmol / L within 6 months after reaching normal levels).

[0005] 2) Insufficient guidance on clinical compliance: Due to the lack of quantitative risk prediction tools, doctors find it difficult to intuitively demonstrate the risk of recurrence to patients (especially families of pediatric patients), resulting in low patient compliance during follow-up after discharge, making it easy to miss the best intervention opportunity, and thus causing irreversible damage to multiple systems and organs throughout the body.

[0006] 3) Low feature utilization: Existing technologies have failed to fully exploit the multi-dimensional clinical data generated during hospitalization (such as medical comprehensive index, interactions between biochemical indicators, etc.), resulting in poor model generalization ability.

[0007] The aforementioned shortcomings and deficiencies stem from the fact that the prognosis of lead poisoning is a complex process influenced by multiple coupled factors, and simple linear analysis cannot handle the complex interactions between clinical variables. Due to the lack of efficient feature engineering techniques and powerful nonlinear modeling tools (such as ensemble learning algorithms), existing prognostic assessment methods struggle to extract highly predictive feature combinations from massive amounts of clinical data, thus limiting their application value in personalized clinical decision-making. Summary of the Invention

[0008] In view of this, the purpose of this invention is to provide a deep learning-based method for predicting the risk of re-emergence of blood lead levels after lead poisoning, solving the problem in the prior art that the risk of re-emergence of blood lead levels after lead poisoning patients cannot be accurately predicted, and achieving accurate prediction of re-emergence of blood lead levels after lead poisoning patients undergoing lead chelation therapy.

[0009] To achieve the above objectives, the present invention provides the following technical solution: Option 1: A deep learning-based method for predicting the risk of re-emergence of blood lead levels after lead poisoning, comprising the following steps: S1: Obtain raw clinical data of the patient to be evaluated, including exposure history, symptoms and signs, laboratory test results, intensity of chelation therapy, and follow-up trends in blood lead levels; S2: Preprocess the raw data, including iterative imputation of missing values, nonlinear transformation of skewed continuous variables, and mode imputation and encoding of categorical variables; S3: Feature engineering construction, specifically: based on preprocessed data, construct a low-dimensional engineered feature space with physiological mechanism orientation through cross-dimensional interaction, including the chain features of "exposure source-exposure intensity-follow-up trend-organ involvement" and medical comprehensive index; S4: Input the constructed features into a pre-trained random forest, support vector machine, logistic regression or XGBoost binary classifier prediction model, output the probability value of blood lead recurrence, and perform risk stratification based on the probability value.

[0010] Furthermore, in step S1, the raw data includes exposure history, symptoms and signs, laboratory test indicators, intensity of lead chelation therapy, and follow-up blood lead change trends.

[0011] Furthermore, step S2 specifically includes: First, during the training phase, identifying fields such as name and ID number are removed to prevent the model from learning spurious associations related to non-medical mechanisms. At the same time, threshold filtering is performed on variables with excessively high missing rates and unclear clinical significance to reduce noise dimensionality. For numerical variables, during the training phase, iterative imputation is used to fill in missing values ​​using multivariate relationships to alleviate distribution shifts caused by univariate imputation. For categorical variables, the mode is used for imputation and encoding during the training phase to ensure model usability. For mixed-type fields, the type is unified before entering the encoding process to reduce parsing errors caused by inconsistent input. For significantly skewed continuous variables and key indicators with threshold effects, nonlinear transformations and interaction term construction are combined during the training phase to enable the model to express risk patterns that better conform to physiological mechanisms with fewer splits in the tree structure. During the deployment phase, a more robust engineering strategy is adopted. When individual inputs are missing, reasonable default values ​​and safe division rules are used to complete feature construction, thereby ensuring system availability and avoiding interruptions during inference.

[0012] Furthermore, in step S3, the low-dimensional engineering feature space includes: Lead exposure intensity and accumulation signal characteristics, including lead exposure intensity Blood lead level changes and trends in blood lead levels ;

[0013]

[0014]

[0015] in, For hospital admission, blood lead levels Blood lead levels at the most recent follow-up visit; For safety constants, The interval from admission to follow-up; Signs of hematopoietic system involvement, including hemoglobin With calcium ratio and red blood cell distribution width With mean erythrocyte volume ratio or product ,in; Liver function and bilirubin metabolism-related signaling features, including aspartate aminotransferases With alanine aminotransferase ratio Total bilirubin With alanine aminotransferase ratio and direct bilirubin With total bilirubin ratio ; Symptom-environment interaction features: Symptom counts coupled with environmental exposure are constructed by binarizing or hierarchically encoding information such as abdominal pain, abdominal tenderness, and lead-producing factories near residences.

[0016] Furthermore, in step S3, among the lead exposure intensity and accumulation signal characteristics, if there are multiple follow-up blood lead levels... Then, further trend fitting is performed.

[0017] Pick As a trend characteristic, interactive elements are added to express "exposure persistence × management behavior":

[0018] in, As the starting point for fitting, To observe fluctuations or residuals, Number of follow-up visits.

[0019] Option 2: A deep learning-based system for predicting the risk of relapse into elevated blood lead levels after lead poisoning, comprising: The data extraction module is used to extract predictive feature variables from electronic medical records or medical datasets; The data preprocessing module is used to perform missing value imputation, noise dimension filtering, nonlinear transformation, and feature encoding. The feature engineering module is used to construct interactive features and comprehensive medical indices with clinical logic. The risk assessment module has a built-in random forest, support vector machine, logistic regression, or XGBoost binary classifier prediction model to output risk probabilities.

[0020] Solution 3: A computer-readable storage medium having a computer program stored thereon, characterized in that the program, when executed by a processor, implements the method described in Solution 1.

[0021] The beneficial effects of this invention are as follows: 1) Data preprocessing: This invention employs various techniques to process different types of data, such as iterative imputation of missing values, nonlinear transformation of skewed continuous variables, and mode imputation and encoding of categorical variables, to ensure data integrity and system availability.

[0022] 2) Deep Feature Engineering: This invention employs different feature construction methods to construct more interactive features for lead exposure intensity and accumulation signals, hematopoietic system involvement signals, liver function and bilirubin metabolism-related signals, and symptoms and environmental cues. This breaks through the limitations of traditional single indicators. The constructed interactive features explicitly introduce the chain information of "exposure source - exposure intensity - follow-up trend - organ involvement", ensuring consistent and reproducible calculation rules for model training and deployment. This enables the model to obtain stable and interpretable prior guidance in small sample scenarios of lead poisoning.

[0023] 3) Multi-model integration: The prediction model of this invention selects multiple models such as random forest, support vector machine, logistic regression or XGBoost binary classifier, and the prediction performance is comparable.

[0024] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0025] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is a flowchart of the deep learning-based method for predicting the risk of re-emergence of blood lead levels after lead poisoning, as described in this invention. Figure 2 This is a block diagram of the deep learning-based system for predicting the risk of re-emergence of blood lead levels after lead poisoning, as described in this invention. Figure 3 This is a comparison chart of the performance of the prediction models; Figure 4 A nomogram for predicting the risk of relapse in blood lead levels; Figure 5 For model calibration and clinical utility evaluation results; Figure 6 This is a clinical subgroup analysis based on predictive models and re-emergence status. Detailed Implementation

[0026] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0027] Example 1: Please see Figure 1 This embodiment provides a deep learning-based method for predicting the risk of re-emergence of blood lead levels after lead poisoning, specifically including the following steps: Step 1: Data extraction module: Extract several predictive feature variables from previous electronic medical records or medical datasets.

[0028] Step 2: Data Preprocessing: First, during the training phase, identifying fields such as name and ID number are removed to prevent the model from learning spurious associations related to non-medical mechanisms. Threshold filtering is applied to variables with excessively high missing values ​​and unclear clinical significance to reduce noise. For numerical variables, iterative imputation is used during training to fill in missing values ​​using multivariate relationships, mitigating distribution shifts caused by univariate imputation. For categorical variables, the mode is used for imputation and encoding during training to ensure model usability. For mixed-type fields, the type is unified before encoding to reduce parsing errors caused by inconsistent input. For significantly skewed continuous variables and key indicators with threshold effects, nonlinear transformations and interaction term construction are combined during training to enable the model to express risk patterns more consistent with physiological mechanisms with fewer splits in the tree structure. During deployment, a more robust engineering strategy is adopted. When individual inputs are missing, reasonable default values ​​and safe division rules are used to complete feature construction, ensuring system availability and avoiding inference interruptions.

[0029] Step 3: Feature Engineering Construction: The goal of feature engineering is to compress information scattered across exposure history, symptoms and signs, laboratory indicators, treatment intensity, and follow-up behavior into a low-dimensional engineered feature space with a mechanism orientation. The final deployed model uses 39 core engineered features. To differentiate it from the common practice in "general disease classification" of using only static test values / symptoms, this invention explicitly introduces the chain information of "exposure source—exposure intensity—follow-up trend—organ involvement" during front-end feature construction, and ensures reproducible computational rules consistent with training and deployment.

[0030] Blood lead level recorded upon admission: The most recent follow-up blood lead level was The number of follow-up visits was The interval from admission to follow-up is and order This is a safety constant to avoid division by zero and extreme amplification.

[0031] (1) Regarding lead exposure intensity and accumulation signals, the core calculation can be expressed as follows: The engineering characteristics are constructed around the framework of “level-change-trend-management behavior”:

[0032]

[0033]

[0034] in, For lead exposure intensity, This represents the change in blood lead levels. This shows the trend of blood lead levels.

[0035] If multiple follow-up blood lead levels are present Then, further trend fitting is performed.

[0036] Pick As a trend characteristic, interactive elements were added to express "exposure persistence × management behavior".

[0037] (2) Regarding the signals of hematopoietic system involvement, hemoglobin ( Erythroid parameters such as hematocrit (Hct), mean corpuscular volume (MCV), and red blood cell distribution width coefficient of variation (RDW-CV) are used, and their ratios and interactions enhance the expression of "lead-related anemia / erythroid alterations," for example...

[0038] Nonlinear transformations are applied to skewed continuous variables, and second-order interactions are added to characterize threshold effects and synergistic anomalies, thereby improving the separability of complex physiological relationships in small samples.

[0039] (3) Regarding liver function and bilirubin metabolism-related signals, the ratio and proportion characteristics of transaminases (ALT / AST) and bilirubin (distinguishing between total and direct bilirubin if necessary) are constructed as the core, and interact with blood lead levels or trend characteristics.

[0040] (4) Regarding symptom and environmental cues, information such as abdominal pain, abdominal tenderness, thinning hair, and the presence of lead-producing factories near the residence are uniformly binarized or hierarchically coded, and symptom counts and symptom-environment interactions are constructed to highlight the coupling relationship between exposure sources and clinical manifestations, for example...

[0041] Based on this, a comprehensive risk score with clinical rules as its framework (and subsequent...) is introduced. Using consistent sub-items as input features, the model obtains stable and interpretable prior guidance in small-sample lead poisoning scenarios.

[0042] Step 4: During the training phase, implement and compare several commonly used classifiers and ensemble models, including linear models, bagged tree models, support vector machines, and boosting tree models, with ROC-AUC and other probabilistic discriminant metrics as the primary selection criterion. The final deployment model uses the XGBoost binary classifier, which is then fixed in the model file as a parameterized object that can be loaded offline. The main reason for choosing XGBoost is its strong ability to express nonlinear relationships and high-order interactions, its ability to gradually focus on high-information-gain split structures through gradient boosting in small sample sizes, and its good engineering controllability and inference efficiency, making it suitable for the response constraints of desktop offline applications. The key hyperparameters of the deployment model are set to 150 trees, a maximum depth of 4, and a learning rate of 0.05 to achieve a balance between expressive power and the risk of overfitting. These parameters are then used to train the prediction model, and continuous tuning and optimization are performed to find the optimal parameters.

[0043] Step 5: The model output is in probabilistic form. Comprehensive risk score.

[0044] in, Score the severity of symptoms. For environmental exposure risk scoring, Scoring for abnormal laboratory indicators Rate the treatment response or management behavior.

[0045] Suppose the model is based on the input feature vector x The generated internal score The risk probability is then obtained by the Sigmoid transformation.

[0046] in, The value, ranging from 0 to 1, represents the probability of a readmission. To adapt to different management scenarios, this invention decouples the probability output from the decision threshold. During the training phase, the threshold can be selected on the validation set based on the Youden index or a cost-sensitivity criterion, thereby achieving an interpretable trade-off between missed diagnoses and misdiagnoses.

[0047] The system also provides a multi-level risk stratification strategy, mapping probabilities to four levels: low risk, medium risk, high risk, and very high risk, to generate recommendations for follow-up cycles and intervention priorities. This design ensures the model remains stable as a continuous risk estimator, while the thresholds and stratification, as strategy layers, can be flexibly adjusted under different departments and follow-up resource conditions.

[0048] The software embeds model inference and feature construction as stable interfaces within the desktop application. Model files are stored in Flask format and support offline loading, with feature names and order remaining consistent. The system shares the same set of feature construction semantics and prediction interfaces in both single-input and batch import paths, avoiding distribution shifts caused by differences in data channels. For missing or anomalous inputs, the system ensures the continuous operation of the inference process through safe numerical conversion, default value rollback, and denominator lower bound control, and writes prediction records to a local database to support subsequent trend analysis and reassessment. Simultaneously, the patient has a time-series model, and the time-series indicator trend fit is as follows:

[0049] It provides trend fitting and risk extrapolation capabilities for multiple historical prediction records of the same patient, enabling the model to expand from a one-time assessment to a dynamic management tool that supports continuous follow-up.

[0050] Example 2: Please see Figure 2 This embodiment provides a deep learning-based system for predicting the risk of re-emergence of blood lead levels after lead poisoning, including: The data extraction module is used to extract predictive feature variables from electronic medical records or medical datasets; The data preprocessing module is used to perform missing value imputation, noise dimension filtering, nonlinear transformation, and feature encoding. The feature engineering module is used to construct interactive features and comprehensive medical indices with clinical logic. The risk assessment module has a built-in random forest, support vector machine, logistic regression, or XGBoost binary classifier prediction model to output risk probabilities.

[0051] Example 3: 1. Comparative Experiment: Performance Comparison of Predictive Models Re-emergence of blood lead was defined as a blood lead level returning to ≥0.96 μmol / L within 6 months after chelation therapy had reached a normal level (<0.48 μmol / L). This experiment used three machine learning algorithms (logistic regression, random forest, and support vector machine) to develop a predictive model and integrated feature engineering to create 105 interaction features and a comprehensive medical index. The dataset was randomly divided into a training group (75%, n=122) and a test group (25%, n=41). Performance curves (AUC), sensitivity, specificity, and clinical utility metrics were evaluated using the area under receiver operating characteristics (AUC), as shown in Table 1. Figure 3 .

[0052] Table 1. Performance Comparison of Test Set Models

[0053] Figure 3 In the table, A is the ROC curve showing that the random forest has the highest AUC (0.850); B is the confusion matrix showing that the random forest correctly identified 22 true negatives and 11 true positives, with very few false diagnoses; C shows that as the number of features increases, the random forest reaches its optimal and stable performance with 35-40 features.

[0054] 2. Clinical efficacy assessment Figure 4 This is a clinical predictive nomogram used to predict the risk of lead relapse. The nomogram includes six key predictors: age (years), serum lead level at admission (μmol / L), serum calcium (mmol / L), hemoglobin (g / L), total chelation therapy dose (g), and abdominal pain. Each predictor corresponds to a score, and the scores are summed to obtain a total score, which corresponds to the predicted probability of lead relapse. The total score of the nomogram ranges from 0 to 220, corresponding to a predicted risk probability range of 0.05 to 0.9. This visualization tool allows clinicians to quickly calculate an individualized risk score for each patient at the bedside by simply summing the scores of each predictor and mapping them to a probability scale to obtain a risk estimate.

[0055] 3. Model Calibration and Decision Curve Figure 5 The results are for model calibration and clinical utility evaluation.

[0056] Figure 5The calibration analysis (Figure A) shows that the model's predicted probabilities are highly consistent with actual observations. The Hosmer-Lemeshow test p-value is 0.749 (not statistically significant), indicating that the model has good calibration. After 100 repeated validations using the bootstrap method, the mean absolute error is only 0.013, further confirming the model's reliability in probability estimation. The sample size of n=557 supports the stability of the results. The calibration curve closely approximates the ideal diagonal, indicating that the model can accurately predict the actual incidence rate at different risk levels.

[0057] Figure 5 Figure B represents the decision curve analysis, showing that the model delivers significant net benefits across various threshold probabilities. Compared to the two extreme strategies of "full treatment" and "no treatment," this model demonstrates superior clinical applicability within a threshold probability range of 0.0 to 0.5, with the greatest net benefit observed at low thresholds (<0.1). The corresponding cost-benefit ratios range from 1:100 to 1:2.5, indicating its application value under diverse clinical decision-making scenarios and resource constraints.

[0058] 4. Prognostic impact and clinical risk stratification This study used all cases (n=163) from the entire cohort to explore prognostic significance. Multiple probability cutoff values ​​were evaluated to optimize the performance metrics for predicting relapse, severe relapse, and early relapse (within 6 months post-discharge) in children with lead poisoning. The final random forest model demonstrated robust performance in predicting relapse, severe relapse, and early relapse throughout the follow-up period and within the first 6 months post-discharge. 22.3%, maximizing the Youden index (0.61), was selected as the optimal probability cutoff value and used as the threshold for clinical subgroup division, defined as "predictive model (+)" and "predictive model (-)". )".

[0059] Figure 6 This is a clinical subgroup analysis based on a predictive model and relapse status. In this analysis, A shows the distribution of clinical outcomes among the four subgroups, indicating a progressive escalation of risk; B shows the Kaplan-Meyer survival curves, demonstrating the differences in relapse-free survival among subgroups (log-rank test, p < 0.001); and C shows a heatmap indicating differences in biomarker expression patterns among subgroups, with lead metabolism, calcium homeostasis, and seasonal risk factors progressively worsening from subgroup 1 to 4.

[0060] 5. Clinical Application The final predictive model is deployed as a web application for clinical use. When clinicians input the desired clinical characteristics, the application automatically predicts the risk of relapse into lead levels in individual pediatric patients. The application also displays the contribution of each characteristic, indicating which factors increase or decrease the risk of relapse for each patient, thus supporting clinical decision-making and personalized follow-up strategies.

[0061] In summary, the machine learning-based blood lead re-escalation prediction model proposed in this invention is specifically designed for children, a higher-risk group among lead poisoning patients, and outperforms traditional clinical methods (AUC=0.850, accuracy=80.5%). This random forest model utilizes 36 readily available clinical features to complete individualized risk assessment within 24 hours of admission. The study innovatively established a four-subgroup risk stratification system, successfully identifying different clinical phenotypes: low-risk patients with good prognosis (22.4%), seemingly stable but occult high-risk patients requiring intensive monitoring (26.5%), key false-negative cases with a severe re-escalation rate as high as 66.7% (6.1%), and clearly high-risk patients with re-escalation and early complications in all groups (40.9%). This comprehensive risk stratification system addresses the clinical challenge of a significantly higher re-escalation rate in children than in adults (39.0% vs 20.2%) through personalized monitoring strategies, implementing intensive follow-up for high-risk patients while adhering to standard management procedures for low-risk patients. This invention promotes a shift in the management of childhood lead poisoning from a passive response to a proactive prediction approach, fills a key gap in current guidelines, and provides evidence-based support and a practical framework for the early identification of latent lead metabolism disorders and the prevention of irreversible neurodevelopmental damage in vulnerable children.

[0062] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting the risk of re-emergence of blood lead levels after lead poisoning based on deep learning, characterized in that, Includes the following steps: S1: Obtain raw clinical data of the patient to be evaluated; S2: Preprocess the raw data; S3: Feature engineering construction, specifically: based on preprocessed data, constructing a low-dimensional engineered feature space with physiological mechanism orientation through cross-dimensional interaction, including the "exposure source-exposure intensity-follow-up trend-organ involvement" chain features and medical comprehensive index; S4: Input the constructed features into the pre-trained prediction model, output the probability value of blood lead levels rising again, and perform risk stratification based on the probability value.

2. The method for predicting the risk of re-emergence of blood lead levels after lead poisoning based on deep learning according to claim 1, characterized in that, In step S1, the raw data includes exposure history, symptoms and signs, laboratory test indicators, intensity of lead chelation therapy, and follow-up blood lead change trends.

3. The method for predicting the risk of re-emergence of blood lead levels after lead poisoning based on deep learning according to claim 1, characterized in that, Step S2 specifically includes: First, during the training phase, identifying fields are removed, and threshold filtering is performed on variables with excessively high missing rates and unclear clinical significance; for numerical variables, the training phase uses iterative imputation to fill in missing values ​​using multivariate relationships; for categorical variables, the training phase fills in the missing values ​​with the mode and encodes them; for mixed-type fields, the type is unified before entering the encoding process; for significantly skewed continuous variables and key indicators with threshold effects, the training phase combines nonlinear transformations and interaction terms to construct them. During the deployment phase, an engineering strategy is adopted to construct features using reasonable default values ​​and safe division rules when individual inputs are missing.

4. The method for predicting the risk of re-emergence of blood lead levels after lead poisoning based on deep learning according to claim 1, characterized in that, In step S3, the low-dimensional engineering feature space includes: Lead exposure intensity and accumulation signal characteristics, including lead exposure intensity Blood lead level changes and trends in blood lead levels ; in, For hospital admission, blood lead levels Blood lead levels at the most recent follow-up visit; For safety constants, The interval from admission to follow-up; Signs of hematopoietic system involvement, including hemoglobin With calcium ratio and red blood cell distribution width With mean erythrocyte volume ratio or product ,in; Liver function and bilirubin metabolism-related signaling features, including aspartate aminotransferases With alanine aminotransferase ratio Total bilirubin With alanine aminotransferase ratio and direct bilirubin With total bilirubin ratio ; Symptom-environment interaction features: Symptom counts coupled with environmental exposure are constructed by binarizing or hierarchically encoding information on abdominal pain, abdominal tenderness, and lead-producing factories around residential areas.

5. The method for predicting the risk of re-emergence of blood lead levels after lead poisoning based on deep learning according to claim 4, characterized in that, In step S3, among the lead exposure intensity and accumulation signal characteristics, if there are multiple follow-up blood lead levels... Then, further trend fitting is performed. Pick As a trend feature, interactive elements are added to express "exposure persistence × management behavior": in, As the starting point for fitting, To observe fluctuations or residuals, Number of follow-up visits.

6. The method for predicting the risk of re-emergence of blood lead levels after lead poisoning based on deep learning according to claim 1, characterized in that, In step S4, the prediction model includes random forest, support vector machine, logistic regression, or XGBoost binary classifier.

7. A deep learning-based system for predicting the risk of re-emergence of blood lead levels after lead poisoning, characterized in that, include: The data extraction module is used to extract predictive feature variables from electronic medical records or medical datasets; The data preprocessing module is used to perform missing value imputation, noise dimension filtering, nonlinear transformation, and feature encoding. The feature engineering module is used to construct interactive features and comprehensive medical indices with clinical logic. The risk assessment module has a built-in random forest, support vector machine, logistic regression, or XGBoost binary classifier prediction model to output risk probabilities.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the method described in any one of claims 1 to 6.

Images