Aluminum exposure neurotoxicity risk prediction method and device based on lncrna expression profile

Abstract

The application discloses an aluminum exposure neurotoxicity risk prediction method based on an LncRNA expression profile, and comprises the following steps: collecting peripheral blood samples of aluminum exposure workers and control groups; extracting total RNA of the peripheral blood samples, constructing an LncRNA sequencing library, and performing high-throughput sequencing to obtain LncRNA expression profile data; screening LncRNA markers with significant differences in the LncRNA expression profile; constructing an aluminum exposure neurotoxicity risk prediction model by using a machine learning algorithm; and performing risk assessment and hierarchical intervention on new aluminum exposure individuals by using the prediction model, wherein the peripheral blood LncRNA expression profile is used as an objective detection index, and compared with a subjective neuropsychological scale, the method is more accurate and sensitive, and is simple to operate and suitable for large-scale population screening. The machine learning method is used for modeling, and the best feature combination can be found from massive high-dimensional omics data, so that the limitations of traditional statistical methods are overcome.

Description

Technical Field

[0001] This invention relates to the technical field of methods for predicting the neurotoxicity risk of aluminum exposure, and more specifically, to a method and apparatus for predicting the neurotoxicity risk of aluminum exposure based on LncRNA expression profiles. Background Technology

[0002] Aluminum (Al) is one of the most abundant metallic elements in the Earth's crust and is widely used in many industries such as aerospace, construction, electronics, and packaging. However, long-term excessive exposure to aluminum and its compounds may cause a series of occupational diseases, such as aluminum pneumoconiosis and aluminum osteopathy.

[0003] The most prominent harm of occupational aluminum poisoning is damage to the nervous system. Numerous studies have shown that inhaled aluminum exposure can lead to central nervous system disorders such as cognitive impairment and decreased motor coordination.

[0004] Neurotoxicity is the main manifestation of occupational aluminum poisoning, but its pathogenesis is not fully understood. Previous studies have suggested that aluminum exposure can induce increased oxidative stress levels in brain tissue, leading to mitochondrial dysfunction and neuronal apoptosis. Some researchers have also found that aluminum ions can competitively bind to calcium channels on the presynaptic membrane, inhibiting neurotransmitter release and resulting in reduced synaptic plasticity. Our previous work revealed that epigenetic modifications are involved in regulating aluminum neurotoxicity; aluminum exposure can cause changes in the methylation level of neuronal DNA, thereby affecting the expression of some important genes.

[0005] Although neurotoxicity is a prominent feature of occupational aluminum poisoning, the screening and intervention for aluminum-exposed workers still faces many challenges. First, existing screening methods mainly rely on neuropsychological tests, such as the Wechsler Adult Intelligence Scale and the Auditory-Verbal Learning Test. These methods are complex and time-consuming, unsuitable for large-scale population screening, and their results are easily influenced by individual subjective factors, resulting in low sensitivity and specificity. Second, due to the lack of convenient and sensitive early warning indicators, aluminum poisoning patients often present with obvious clinical symptoms by the time of diagnosis, missing the optimal treatment window. Furthermore, the complexity of the neural apparatus and the high heterogeneity of cognitive impairment make the development of targeted interventions and the evaluation of their effectiveness extremely difficult. Therefore, there is an urgent need to develop rapid, accurate, and objective early prediction methods for occupational aluminum poisoning to achieve precise screening and timely intervention for key populations. Summary of the Invention

[0006] This invention addresses the aforementioned technical problems by proposing a method and device for predicting the neurotoxicity risk of aluminum exposure based on LncRNA expression profiles. The method first uses high-throughput sequencing to analyze the LncRNA expression profiles of occupational groups, then constructs a neurotoxicity risk prediction model using machine learning algorithms, predicts the probability of cognitive impairment based on an individual's LncRNA expression pattern, and finally implements targeted prevention and control measures for high-risk groups.

[0007] This invention provides a method for predicting the neurotoxicity risk of aluminum exposure based on lncRNA expression profiling, comprising the following steps:

[0008] Peripheral blood samples were collected from aluminum-exposed workers and control groups; total RNA was extracted from the peripheral blood samples, a lncRNA sequencing library was constructed and high-throughput sequencing was performed to obtain lncRNA expression profile data; lncRNA biomarkers with significant differential expression in the lncRNA expression profile were screened; a risk prediction model for aluminum exposure neurotoxicity was constructed using machine learning algorithms; and the prediction model was used to conduct risk assessment and stratified intervention for new aluminum-exposed individuals.

[0009] Specifically, the step of screening LncRNA biomarkers includes: preprocessing the LncRNA expression profile data to obtain an expression matrix E = {E...} ij}, where E ij Let represent the expression level of LncRNAj in sample i. The Boruta algorithm is used for initial screening of candidate features, with a threshold of Thres, and N iterations. The importance of the original features is evaluated by constructing random shadow features, resulting in a candidate LncRNA feature set F = {j|Hit}. j >Thres}, where Hit j The calculation formula is:

[0010]

[0011] In the formula, r j s represents the original feature, and s represents the shading feature. The indicator function is used; further, a swarm intelligence optimization method is employed to screen the optimal LncRNA combination G. * Candidate features are encoded as binary vectors, the population is randomly initialized, and selection, crossover, and mutation are iteratively performed, using the Matthews correlation coefficient (MCC) as the fitness function, until the stopping condition is met, and the optimal combination is output. The objective function is:

[0012]

[0013] Where ||·||0 represents the L0 norm, k is the number of target features, and L is the sample label vector.

[0014] Specifically, the steps for constructing the risk prediction model include: performing adaptive semi-supervised oversampling balancing on the selected LncRNA expression data, and for each minority class sample x i From its k nearest neighbors N(x) i Select appropriate samples from ) Generate a new sample x using interpolation ij :

[0015]

[0016] By estimating sample x i Local minority class probability density at [location] Adaptive adjustment of the number of synthesized samples r i :

[0017]

[0018] In the formula, For indicator function l x Let |P| be the sample label, |P| be the number of minority class samples, and γ be the balance factor. A classifier based on an attention mechanism and a multi-scale convolutional neural network (AMCNN) is constructed, using the LncRNA expression matrix as input to extract multi-scale local features.

[0019] C i =ReLU(Conv1d(E,h) i C = [C1, C2, ..., C k ]

[0020] The sample representation is adaptively adjusted through an attention layer:

[0021] A = Softmax(tanh(CW1+b1)W2)

[0022]

[0023] In the formula, h i Let b1, b1, and w2 be the size of the i-th convolutional kernel, and W1, b1, and W2 be the parameters of the attention layer; Perform global average pooling and generate sample prediction probabilities through a fully connected layer.

[0024]

[0025] Minimize the cross-entropy loss function to train the network parameters θ:

[0026]

[0027] Among them, y i The true labels for the samples.

[0028] Specifically, the steps of training the prediction model further include: optimizing the hyperparameters using an evolutionary-gradient hybrid optimization algorithm, and dividing the hyperparameters into a discrete space Θ. d and continuous space Θ c An initial population II = {θ1, θ2, ..., θ3} is formed by randomly sampling each sample. N Iteratively select, recombine, mutate, and perform gradient descent to update the population and output the optimal hyperparameter combination θ. * Discrete variables employ two-point crossover and random mutation, while continuous variables employ arithmetic crossover and Gaussian mutation, and are updated using gradient descent.

[0029]

[0030] In the formula, η is the learning rate. Let θ be the gradient of the loss function with respect to θ′.

[0031] Specifically, the step of training the prediction model further includes:

[0032] The generalization performance of the prediction model is evaluated using the K-fold cross-validation method. The dataset is randomly divided into K mutually exclusive subsets. Each subset is used as the test set in turn, and the remaining K-1 subsets are used as the training set. The model performance is evaluated on the test set, and the average evaluation index is calculated.

[0033] The stability of the prediction model is evaluated using the Y-Randomization method. R random datasets are generated, the features are kept unchanged, the sample labels are randomly shuffled, the model is trained and its performance is evaluated, the performance distribution of the random model is obtained, and the performance of the real model is compared with it to evaluate the reliability of the model.

[0034] Specifically, the application steps of the prediction model include: obtaining newly collected blood samples from aluminum-exposed workers, extracting total RNA and detecting lncRNA expression profiles; inputting the new sample expression profile data into the trained prediction model to obtain the risk prediction probability of the sample.

[0035]

[0036] Based on predicted probability The samples are divided into high-risk groups by comparing them with a preset threshold τ. or low-risk group And take corresponding health management measures.

[0037] Specifically, the method further includes a model interpretation step: interpreting the decision-making mechanism of the prediction model using the SHAP framework, and measuring the contribution of each LncRNA feature to the model prediction based on the Shapley value; Shapley value φj Defined as the weighted average of the marginal contribution of feature j across all feature permutations:

[0038]

[0039] Where S is a subset of features, F is the complete set of features, and f S (x) represents the predicted value of model f for sample x when only feature S is considered; SHAP approximates the Shapley value by introducing a linear function q and solving the additive feature attribution equation:

[0040]

[0041] Finally, the importance indicators of each lncRNA were obtained.

[0042] Specifically, before the step of constructing the risk prediction model, the method further includes: performing batch normalization on the LncRNA expression matrix to reduce the impact of batch effects.

[0043] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for predicting the neurotoxicity risk of aluminum exposure.

[0044] A device for predicting the neurotoxicity risk of aluminum exposure based on lncRNA expression profiling includes:

[0045] The data acquisition module is used to collect peripheral blood samples from aluminum-exposed workers and control groups.

[0046] The bioinformatics analysis module is used to extract total RNA from the blood sample, construct a LncRNA sequencing library, perform high-throughput sequencing, and analyze the LncRNA expression profile data.

[0047] The LncRNA screening module is used to screen for significantly differentially expressed LncRNA biomarkers from the LncRNA expression profile, including preliminary screening using the Boruta algorithm and further selection of the best LncRNA combination using a swarm intelligence optimization algorithm.

[0048] The data balancing module is used to balance the number of samples of different classes using the adaptive semi-supervised oversampling method ASEMI.

[0049] The model training module is used to build a risk prediction model based on attention mechanism and multi-scale convolutional neural network AMCNN, and to optimize the model hyperparameters using the evolutionary gradient hybrid optimization algorithm EGHO.

[0050] The model evaluation module is used to evaluate the generalization performance and stability of the prediction model using K-fold cross-validation and Y-Randomization methods.

[0051] The model application module is used to input newly collected LncRNA expression profile data of aluminum-exposed individuals into the prediction model to predict their neurotoxicity risk probability, and to perform stratified management and intervention for individuals based on the prediction results.

[0052] The model interpretation module is used to interpret the decision-making mechanism of the prediction model using the SHAP framework and to calculate the Shapley importance values ​​of key LncRNA biomarkers.

[0053] The present invention has the following beneficial effects:

[0054] (1) Using peripheral blood LncRNA expression profile as an objective detection indicator, it is more accurate and sensitive than subjective neuropsychological scales, and is easy to operate, making it suitable for large-scale population screening.

[0055] (2) By using machine learning methods to model, the best feature combination can be discovered from massive high-dimensional omics data, overcoming the limitations of traditional statistical methods.

[0056] (3) Various algorithm innovations were introduced in the feature selection, sample balancing and model building stages, such as the combination of Boruta-swarm intelligence optimization, adaptive semi-supervised oversampling, and evolutionary-gradient hybrid optimization, which significantly improved the classification performance and robustness of the model.

[0057] (4) The prediction model based on attention mechanism and multi-scale convolutional neural network can extract deep features of LncRNA expression profile from both sequence and function levels, which is superior to conventional machine learning classifiers.

[0058] (5) The SHAP framework was used to explain the intrinsic mechanism of the model and identify the LncRNA biomarkers that contribute the most to the risk of neurotoxicity. This not only helps to elucidate the molecular mechanism of aluminum poisoning, but also provides a direction for subsequent targeted drug development.

[0059] (6) This method enables hierarchical management and precise prevention and control of aluminum-exposed populations. High-risk populations can be promptly transferred from hazardous positions and receive personalized health education, while low-risk populations can avoid over-medicalization and save medical resources.

[0060] In summary, the novel aluminum neurotoxicity risk prediction method developed in this invention not only overcomes the shortcomings of existing technologies, enabling early warning and individualized prevention and control of aluminum poisoning, but also has significant application value in improving the health of working populations and reducing the burden of occupational diseases. Furthermore, it provides new insights for machine learning-assisted diagnosis of complex diseases and has broad prospects for widespread application. The successful implementation of this invention will promote the deep integration of artificial intelligence technology with occupational disease prevention and control practices, exploring new pathways for the application of precision medicine in the field of occupational health. Attached Figure Description

[0061] Figure 1 This is a flowchart of the aluminum exposure neurotoxicity risk prediction method based on LncRNA expression profiles according to the present invention.

[0062] Figure 2 This is a schematic diagram of the framework of the aluminum exposure neurotoxicity risk prediction device based on LncRNA expression profile of the present invention. Detailed Implementation

[0063] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of the aluminum exposure neurotoxicity risk prediction method based on LncRNA expression profiles proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0064] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0065] Please see Figure 1 This invention proposes a method for predicting the neurotoxicity risk of aluminum exposure based on lncRNA expression profiling, aiming to address the problems of outdated occupational aluminum poisoning screening methods, lack of early warning indicators, and difficulty in individualized intervention in existing technologies. This invention can be applied to health monitoring and occupational disease prevention for high-risk occupational groups in aluminum smelting, aluminum welding, and aluminum processing. It obtains the lncRNA expression profile of peripheral blood from employees through high-throughput sequencing, and uses machine learning algorithms to construct an aluminum poisoning risk prediction model to accurately predict the probability of neurotoxicity in employees. Based on this prediction, targeted interventions can be implemented for high-risk groups, thereby significantly improving the occupational health of workers exposed to aluminum dust.

[0066] Specifically, the technical solution of the present invention is as follows:

[0067] Step 1. Collect peripheral blood samples from aluminum-exposed workers and control groups. In practical applications, workers in high-risk positions such as aluminum electrolysis and aluminum welding can be identified through employee health records and job risk assessments. A portion of these workers are randomly selected for sampling, while a healthy control group matched for age, gender, and other factors is recruited from within the company or surrounding communities. Each subject has 5ml of venous blood collected, anticoagulated with EDTA, and stored at 4℃. Generally, the aluminum exposure group and the control group each have at least 100 participants to ensure the reliability of subsequent statistical analysis.

[0068] Step 2. Extract total RNA from peripheral blood samples, construct a lncRNA sequencing library, and perform high-throughput sequencing. Total RNA was extracted from the obtained blood samples using TRIzol reagent. 2 μg of RNA was used for purity and integrity testing, ensuring the RNA integrity count (RIN) > 7.0. Ultra TM LncRNA sequencing libraries were constructed using an RNA library construction kit, strictly following the standard workflow of the Illumina HiSeq sequencing platform. Paired-end 150bp sequencing was performed, with each sample containing at least 20 million reads to ensure the quality of the LncRNA expression profile data. FastQC was used to assess sequencing data quality, and Cutadapt was used to remove adapters and low-quality reads. Clear reads were aligned to the human reference genome (GRCh38) using HISAT2, transcripts were assembled using StringTie, and new incRNA transcripts were predicted using the PipeLineR incRNA identification workflow. IncRNA expression quantification and differential expression analysis were performed using DESeq2, screening incRNAs with Illumina HiSeq2 FCI > 1 and FDR < 0.05 as differentially expressed genes, resulting in the LncRNA expression matrix E = {E...} ij}, where matrix element E i j represents the expression abundance of LncRNA j in sample i, typically measured using FPKM (fragments per dry base of a gene per million mapped reads) or TPM (transcripts from a gene per million transcripts). This step provides a high-quality feature matrix for subsequent modeling and also allows for the screening of aluminum-induced LncRNA expression abnormalities in the enterprise population, providing clues for elucidating the mechanism of aluminum poisoning.

[0069] Step 3. Screening for significantly differentially expressed lncRNA biomarkers in the lncRNA expression profile. Using a feature selection strategy combining the Boruta algorithm and swarm intelligence optimization, a set of key lncRNA biomarkers that simultaneously satisfy statistical significance and biological relevance are identified from the high-dimensional lncRNA expression profile. This reduces the difficulty of subsequent modeling and also helps to reveal the molecular mechanisms of aluminum toxicity.

[0070] First, the Boruta algorithm is used to initially screen the LncRNA expression matrix E. Boruta evaluates the importance of the original features by constructing random shading features. After N iterations, the candidate feature set F is selected based on the importance score Hit. The formula for calculating Hit is as follows:

[0071]

[0072] r j Indicates the original lncRNA characteristics, s represents random shading characteristics. This is an indicator function. When Hit j If the value exceeds a preset threshold Thres, LncRNA j is included in the candidate set F. The threshold Thres can be flexibly set according to actual needs, and is usually between 0.7 and 0.9.

[0073] Then, a swarm intelligence optimization method is used to further search for the optimal LncRNA combination G from the candidate set F. * Candidate LncRNAs are encoded into binary vectors of length |F|, and N initial solutions are randomly generated as a population. Then, the population is iterated continuously, with evolutionary operations such as selection, crossover, and mutation performed. The Matthews correlation coefficient (MCC) is used as the fitness function to evaluate the performance of each solution until the maximum number of iterations or fitness convergence is met, thus terminating the process.

[0074] The Boruta algorithm evaluates each feature individually, without considering the synergistic effect between LncRNAs. To further optimize feature combinations, a swarm intelligence optimization strategy is designed: candidate LncRNAs are encoded as binary vectors, where 1 indicates selection and 0 indicates non-selection. The vector length is equal to the number of candidate LncRNAs |F|. N LncRNA combinations are randomly generated as the initial population. A Support Vector Machine (SVM) is used as the evaluator to calculate the predictive power of each LncRNA combination, and the Matthews correlation coefficient (MCC) is used as the fitness value. Selection, crossover, and mutation operations are iteratively performed to update the population.

[0075] Selection: Using roulette wheel selection, individuals with high fitness have a higher probability of being selected.

[0076] Crossover: Randomly select two individuals and swap the encoding segments at the corresponding positions according to the crossover probability Pc.

[0077] Mutation: For each individual, randomly flip 0 and 1 according to the mutation probability Pm.

[0078] When the number of iterations reaches T or the optimal fitness does not improve after t consecutive iterations, the iteration stops and the LncRNA combination with the highest fitness is output.

[0079] This process can be represented by the following matrix: Let the i-th individual be G. i =(g i1 ,g i2 ,...,g i|F| The population matrix is ​​G, and the fitness vector is f = (f1, f2, ..., f3). N ),

[0080] but

[0081] The solution with the highest fitness is ultimately output as the optimal LncRNA combination G. * The objective function for optimization is as follows:

[0082]

[0083] Where ||·||0 represents the L0 norm, i.e., the number of non-zero elements in the vector, and k is the number of lncRNAs to be selected. L is the sample category label vector, with aluminum-exposed samples labeled as 1 and control samples labeled as 0. This method considers the synergistic effect between lncRNAs while selecting features, thus possessing stronger biological interpretability. This step can be applied to enterprises with a high incidence of occupational diseases caused by aluminum, rapidly identifying key biomarkers from a massive number of lncRNAs and guiding the optimal allocation of limited medical resources.

[0084] Step 4. Construct a neurotoxicity risk prediction model for aluminum exposure based on lncRNA expression features. Addressing the characteristics of lncRNA data—high dimensionality, limited sample size, and difficult classification—this invention innovatively designs the AMCNN prediction model and proposes ASEMI oversampling and EGHO optimization strategies, significantly improving the accuracy and robustness of aluminum poisoning risk assessment. AMCNN takes the expression matrix E as input and first automatically extracts multi-level features of the lncRNA sequence through multi-scale one-dimensional convolution; then, it dynamically adjusts the feature representation of the samples through an attention mechanism to highlight salient segments, and finally maps the samples to class probability outputs through pooling and cascaded fully connected layers. Considering that aluminum poisoning patients are far fewer than healthy individuals, the ASEMI algorithm synthesizes new minority class samples through adaptive interpolation to alleviate class imbalance.

[0085]

[0086] Next, estimate the synthetic sample x. ij Local minority class probability density in the region

[0087]

[0088] Among them, l x Let x be the label of sample x, and σ be the bandwidth of the Gaussian kernel. Finally, each original minority class sample x is adaptively adjusted based on the estimated probability density. i oversampling factor r i :

[0089]

[0090] In the formula, |P| represents the minority class sample set, and γ is a hyperparameter controlling the degree of balance. Compared with the traditional SMOTE method, ASEMI can better approximate the true data distribution. For x i The nearest neighbor, r represents the minority class density of the region where the interpolated sample is located. i This is the oversampling factor.

[0091] The balanced LncRNA expression matrix is ​​input into the AMCNN model for training. The AMCNN network structure includes:

[0092] First, the input representation matrix E is convolved using multiple one-dimensional convolution kernels of different scales to extract local features. The window size of the i-th convolution kernel is h. i The convolution result is denoted as C. i After activation, you will get:

[0093] C i =ReLU(Conv1d(E,h) i C = [C1, C2, ..., C k ]

[0094] Next, the feature representation of the samples is adaptively adjusted through an attention mechanism. Let C have an n×d shape, and introduce an attention weight matrix A:

[0095] A = Softmax(tanh(CW1+b1)W2)

[0096] in, Multiplying A element-wise by C yields the weighted feature matrix.

[0097]

[0098] For the weighted characteristic matrix Average pooling is performed along the sample dimension, and the predicted probability of each sample is output through a fully connected layer.

[0099]

[0100] Minimize the cross-entropy loss function to optimize model parameters θ:

[0101]

[0102] In the formula, y i Let be the true label of the i-th sample.

[0103] Furthermore, the EGHO algorithm combines evolutionary algorithms and gradient descent to optimize AMCNN hyperparameters, such as kernel size, number of neurons in fully connected layers, learning rate, and regularization coefficients, in a discrete-continuous hybrid search space. EGHO divides the search space into a discrete space Θ. d and continuous space Θ c For the discrete part, classical evolutionary algorithms are used for optimization, including selection, crossover, and mutation operations; for the continuous part, gradient descent is used for local search.

[0104]

[0105] In the formula The objective function (i.e., cross-entropy loss) is represented by this function. The gradient of the continuous parameter θ′ is defined by η, where η is the learning rate. This step can be applied to high-risk enterprises for occupational aluminum poisoning, using past employee health monitoring data to build a predictive model, thus shifting from "post-incident treatment" to "pre-incident prevention." Enterprises can regularly collect blood samples from employees, detect key lncRNA expression profiles, and input them into the trained model to obtain the probability of disease for each employee, thereby developing personalized prevention and control plans. High-risk workers should receive enhanced occupational health training, shorter medical examination periods, and, if necessary, adjustments to their work positions; while low-risk workers can maintain routine management, avoiding excessive medical intervention. This model can be embedded in the enterprise's occupational health information management system, forming a humanized and intelligent occupational disease prevention and control system of "sequencing-prediction-intervention."

[0106] Step 5. Evaluate the generalization performance and stability of the prediction model. K-fold cross-validation and Y-Randomization are used to comprehensively evaluate the model's predictive ability for unknown samples and its ability to identify spurious associations, providing objective evidence for practical applications. In K-fold cross-validation, the data is randomly divided into K parts. One part is used as the test set, and the remaining K-1 parts are used as the training set in turn. The experiment is repeated K times, and the average performance is evaluated, which reduces model variance and improves reliability. In Y-Randomization, multiple random datasets are generated by artificially shuffling sample labels. The performance difference between the original model and the randomized model is compared, which eliminates the risk of overfitting and ensures the reliability of the prediction results. If the performance of the original model is significantly better than that of the randomized distribution, its prediction results can be considered reliable, rather than spurious associations caused by overfitting data. Enterprises can internally validate the model based on local employee data, while also conducting independent testing with external data, and continuously optimize and iterate in practical applications to continuously improve the prediction level.

[0107] Step 6. Apply the trained model to predict the risk of aluminum poisoning in new samples. For newly hired employees or those who have not been previously screened, collect venous blood, extract RNA, and detect the expression levels of key lncRNAs. Substituting this data into the model allows for rapid prediction of their risk probability.

[0108]

[0109] In the formula, x is the LncRNA expression vector of the new sample, and θ * These are the optimal model parameters obtained in step 4. Then, based on the predicted probabilities... The sample is divided into a high-risk group based on its relationship with a preset threshold τ. or low-risk group The threshold τ can be flexibly set according to actual needs, balancing sensitivity and specificity. For high-risk groups, interventions such as regular physical examinations, environmental monitoring, and improvements to protective facilities should be implemented, while routine health management should be maintained for low-risk groups. When an employee... When the number of high-risk individuals is significantly higher than the average for employees in the same position, the risk of illness should be taken seriously, and clinical consultation and occupational disease diagnosis should be conducted promptly. When the number or proportion of high-risk individuals within the company exceeds the warning threshold, an investigation of occupational hazard factors should be carried out in a timely manner, and protective facilities should be improved and the working environment optimized. This step enables companies to accurately identify key targets for prevention and control, carry out targeted occupational disease prevention and control work, and maximize the protection of workers' health.

[0110] Step 7. Explain the decision-making mechanism of the prediction model. To understand how the model works and explore the importance of each LncRNA biomarker, this invention uses the SHAP (SHapley Additive exPlanations) framework to interpret the prediction results of AMCNN. SHAP is based on the concept of Shapley value in game theory, measuring the contribution of each feature to the model output. Formally, for a sample x... i Shapley value φ j Defined as the weighted average of the marginal contribution of feature j across all feature subsets S:

[0111]

[0112] F represents the entire feature set, f S (x) represents the prediction of model f for x when only the feature subset S is considered. Directly calculating the Shapley value is very complex. To improve efficiency, SHAP introduces a linear function g as an approximation of f:

[0113]

[0114] z′∈{0,1} M This is an indicator vector for whether a feature appears. SHAP solves for the optimal feature attribution weight φ by minimizing the mean squared errors of g and f and adding a sparse regularization term. j :

[0115]

[0116] Using SHAP values This step visually presents the magnitude and direction of each lncRNA's contribution to predicting high / low risk. It reveals key lncRNA biomarkers that determine the risk of aluminum poisoning, providing insights into the molecular mechanisms of aluminum poisoning and potential therapeutic targets for subsequent drug development. Companies can regularly provide feedback on SHAP analysis results to occupational disease prevention and control departments, offering empirical evidence for optimizing occupational disease regulatory policies.

[0117] Step 8. Perform batch normalization preprocessing on the LncRNA expression matrix. This operation aims to standardize the data of each batch to a distribution with a mean of 0 and a variance of 1, thereby reducing the difference in data distribution between different batches, accelerating model convergence, and improving training stability. Let the input of the i-th batch be B = {x1, x2, ..., x...} m The formula for batch normalization is:

[0118]

[0119] Where, μ B and Let be the mean and variance of batch B, respectively, and let ∈ be a small constant to prevent the denominator from being zero. γ and β are learnable scaling and translation parameters, ensuring that the normalized data retains some expressive power. In practice, methods such as Min-Max scaling should be used to pre-normalize LncRNA expression levels to eliminate batch-to-batch systematic bias and improve model applicability.

[0120] The aluminum poisoning prediction method involved in this invention is packaged into a computer program product. This program can be stored on various computer-readable media (such as CDs, USB flash drives, and external hard drives), and can be read and executed by a computer to automatically predict the risk of aluminum poisoning. The program should preferably be written in a cross-platform language (such as Java or Python) and implemented with a software architecture of "interactive process design + modular algorithm library + visual operation interface," ensuring that users can flexibly call it without programming knowledge and that it can be customized according to actual needs. Enterprises can purchase this software product and deploy it on their own servers, with professional personnel performing daily maintenance, thereby achieving continuous monitoring and intelligent early warning of aluminum poisoning risks.

[0121] Please refer to Figure 2 The present invention further provides a device for predicting the neurotoxicity risk of aluminum exposure based on lncRNA expression profiles. This device includes the following modules:

[0122] Data acquisition module 1: used to collect peripheral blood samples from aluminum-exposed workers and control groups, and to sample and preserve them.

[0123] Bioinformatics analysis module 2: used to extract total RNA from the blood sample, construct a LncRNA sequencing library, generate LncRNA expression profile data using a high-throughput sequencing platform, and perform data analysis such as quality control, alignment, and quantification, and finally generate an LncRNA expression matrix for subsequent modeling.

[0124] LncRNA screening module 3: Initially screens candidate LncRNA features using the Boruta algorithm, and then uses a swarm intelligence optimization algorithm to search for the best combination of LncRNAs to discover key LncRNA biomarkers.

[0125] Data balancing module 4: Used to balance the number of samples of different categories using the ASEMI adaptive oversampling method, thereby mitigating the adverse effects of sample imbalance on modeling.

[0126] Model Training Module 5: Construct and train an aluminum poisoning risk prediction model based on AMCNN, and use the EGHO algorithm to optimize the model hyperparameters to finally obtain a high-performance risk predictor.

[0127] Model Evaluation Module 6: Utilizes K-fold cross-validation and Y-Randomization to comprehensively evaluate the model's generalization performance and stability.

[0128] Model Application Module 7: It is responsible for receiving new aluminum-exposed worker samples, calling the trained prediction model and outputting the predicted probability of their neurotoxicity risk, and realizing risk stratification and intervention decisions based on preset thresholds.

[0129] Model Interpretation Module 8: Uses the SHAP framework to calculate and present the importance values ​​of key LncRNA biomarkers in the model, revealing the model's internal decision-making mechanism.

[0130] These modules, independently packaged yet collaboratively integrated, constitute an automated and intelligent occupational aluminum poisoning early warning device. Based on lncRNA expression profiling, this device accurately predicts the neurotoxicity risk of aluminum-exposed workers, demonstrating greater specificity and foresight than traditional methods. Furthermore, the proposed machine learning modeling strategy and algorithm improvements overcome the limitations of existing technologies, showcasing originality and superiority in multiple aspects. This device can be embedded into routine occupational disease monitoring and management processes, assisting doctors in screening key populations and providing early intervention. It has significant application value in improving the health of aluminum dust workers and reducing the incidence of occupational diseases.

[0131] The system's modules can operate independently, be flexibly combined, and seamlessly integrated to form a complete prediction process. The system is best designed with a B / S architecture, using a relational database to store data, a web service framework to build the business logic layer, and visual charts to present analysis results, facilitating distributed deployment and remote access. Enterprises can integrate this system into their existing occupational health management information systems, enabling data sharing and coordinated responses with employee health records, occupational disease monitoring, and other related functions.

[0132] In summary, this invention fully utilizes cutting-edge technologies such as high-throughput sequencing and machine learning, taking LncRNA expression profiling as a starting point, to develop a highly practical and targeted aluminum poisoning risk prediction scheme. This scheme can be widely applied in the following scenarios:

[0133] (1) Enterprises in high-risk industries such as aluminum smelting and aluminum welding can use this invention to establish an enterprise-specific occupational disease prediction model, conduct aluminum poisoning risk assessment for all employees, identify key protection targets, and formulate differentiated prevention and control measures to minimize the incidence of occupational diseases.

[0134] (2) Occupational health testing and occupational disease diagnosis institutions can use this invention to conduct early screening for aluminum poisoning among workers in key industries. By tracking and following up on high-risk groups, early cases can be detected and intervened in a timely manner, effectively reducing the disability rate caused by occupational diseases.

[0135] (3) Government functional departments such as disease control centers and safety supervision departments can use this invention to carry out occupational disease supervision and inspection of aluminum industry enterprises in their jurisdiction. By collecting and summarizing the predicted probability distribution of enterprise employees, they can dynamically grasp the overall occupational disease risk level of the industry and provide a scientific basis for formulating and optimizing prevention and control policies.

[0136] (4) Medical institutions and research institutes can use this invention to conduct in-depth research on the pathogenesis and prevention of aluminum poisoning. By conducting functional experiments on key LncRNAs screened in the prediction model, they can elucidate their regulatory role in the occurrence and development of aluminum poisoning, and thereby discover new early diagnostic markers and therapeutic intervention targets.

[0137] Furthermore, the predictive modeling ideas and algorithmic strategies involved in this invention, such as Boruta-swarm intelligence optimization feature selection, ASEMI adaptive oversampling, and EGHO evolutionary-gradient hybrid optimization, can also be extended to the health effect assessment of other occupational hazards, such as coal dust, benzene, and noise. This not only expands the application scope of machine learning in the field of occupational diseases but also opens up new paths for intelligent assisted diagnosis of complex diseases. With the rapid development of molecular biology, artificial intelligence, and other disciplines, especially the continuous emergence of new technologies such as single-cell sequencing, knowledge graphs, and transfer learning, disease risk prediction based on omics big data will become an important development direction for precision medicine. This invention, as an innovative attempt to combine LncRNA expression profiling with machine learning, sets a benchmark for multi-omics data-driven risk assessment of complex diseases and has broad application prospects.

[0138] In summary, this invention provides a novel method for predicting the neurotoxicity risk of aluminum-exposed workers based on LncRNA expression profiles using a machine learning model. A comprehensive intelligent early warning system covering the entire process—from sample collection and data analysis to model building, effect evaluation, and result interpretation—is built around this method. This effectively addresses the technical challenges of outdated aluminum dust occupational disease screening methods and insufficient data for policy formulation, providing new ideas and tools for the precise prevention and control of occupational aluminum poisoning. On the one hand, this invention can be applied to health monitoring of aluminum industry workers, helping companies implement targeted individualized interventions and improve the effectiveness of occupational disease prevention and control. On the other hand, it can be extended to risk assessment of other occupational and environmental exposures, providing scientific references for government departments to formulate exposure limits and optimize management policies. Furthermore, this invention's innovative approach in disease risk prediction opens a new path for intelligent diagnosis of complex diseases driven by multi-omics data. With the rapid advancement of big data and artificial intelligence technologies, this research paradigm is expected to further expand its applications and serve a wider range of healthcare needs.

[0139] To verify the superiority of the method of the present invention, the applicant designed the following embodiments and comparative examples, and conducted detailed performance tests:

[0140] Example 1: The method for predicting the neurotoxicity risk of aluminum exposure based on LncRNA expression profiles proposed in this invention;

[0141] Comparative Example 1: A method for assessing the risk of aluminum poisoning based on traditional biochemical indicators (serum aluminum concentration, urinary aluminum concentration);

[0142] Comparative Example 2: A method for assessing the risk of aluminum poisoning based on neuropsychological tests (such as the Mini-Mental State Examination, MMSE).

[0143] Comparative Example 3: A Method for Predicting Aluminum Poisoning Risk Based on Conventional Machine Learning Algorithms (Random Forest)

[0144] Test dataset: Peripheral blood samples were collected from 500 aluminum-exposed workers and 500 healthy controls from a large aluminum smelting company. 80% of the samples were randomly selected as the training set and 20% as the test set.

[0145] Detection indicators and methods:

[0146] 1. Accuracy: Number of correctly predicted samples / Total number of samples;

[0147] 2. Sensitivity: Number of correctly predicted positive samples / Total number of actual positive samples;

[0148] 3. Specificity: Number of correctly predicted negative samples / Total number of actual negative samples;

[0149] 4. AUC (Area Under the ROC Curve): Reflects the overall classification performance of the model;

[0150] 5. F1 score: the harmonic mean of precision and recall;

[0151] 6. Running time: The computation time required for model training and prediction;

[0152] The test results are shown in the table below:

[0153] Example 1 0.92 0.89 0.95 0.97 0.91 45 Comparative Example 1 0.75 0.7 0.8 0.82 0.73 10 Comparative Example 2 0.78 0.75 0.81 0.85 0.76 3600 Comparative Example 3 0.85 0.82 0.88 0.91 0.84 30

[0154] Test Result Analysis and Interpretation:

[0155] 1. Accuracy: The accuracy of the method in this invention (0.92) is significantly higher than other methods, exceeding the closest comparative example 3 (random forest) by 7 percentage points. This indicates that the present invention can more accurately distinguish between high-risk and low-risk groups for aluminum poisoning, thus helping to improve the accuracy of screening.

[0156] 2. Sensitivity: The method of this invention has the highest sensitivity (0.89), meaning it can detect more potential cases of aluminum poisoning and reduce the rate of missed diagnoses. This is crucial for the early detection and intervention of occupational diseases.

[0157] 3. Specificity: The method of the present invention has the highest specificity (0.95), indicating that it performs well in identifying healthy people and can effectively reduce misdiagnosis and unnecessary waste of medical resources.

[0158] 4. AUC: The AUC value of the method of this invention (0.97) is close to 1, which is much higher than other methods, indicating that it can maintain excellent classification performance under various threshold settings and has stronger generalization ability.

[0159] 5. F1 score: The F1 score (0.91) of the method of this invention is the highest, indicating that it achieves the best balance between accuracy and recall, and is particularly suitable for dealing with the problem of unbalanced aluminum poisoning risk prediction.

[0160] 6. Running Time: Although the method of this invention (45s) is slightly slower than the traditional biochemical index method (10s) and the random forest method (30s), it is significantly more efficient than the neuropsychological testing method (3600s). Considering its significantly improved predictive performance, this time cost is acceptable.

[0161] Best Practice Implementation: Example 1 (Method of the Invention), these test results fully demonstrate the superiority of the method of the invention:

[0162] 1. Compared with traditional biochemical index methods (Comparative Example 1), the method of this invention shows significant improvements in all indicators, especially an increase of 0.15 in AUC. This indicates that the lncRNA expression profile contains richer information than a single biochemical index and can more comprehensively reflect the impact of aluminum exposure on the human body.

[0163] 2. Compared to neuropsychological testing methods (Comparative Example 2), the method of this invention is not only more accurate, but also significantly reduces the testing time (from 1 hour to 45 seconds). This greatly improves screening efficiency and makes large-scale population screening possible.

[0164] 3. Compared to conventional machine learning methods (Comparative Example 3), the method of this invention shows an improvement of 5-7 percentage points in all metrics. This verifies that the innovations in feature selection, sample balancing, and model construction (such as Boruta-swarm intelligence optimized feature selection, ASEMI oversampling, and AMCNN models) of this invention do indeed bring about performance improvements.

[0165] 4. The method of this invention achieves a high level of sensitivity and specificity (0.89 and 0.95, respectively), a balance that is particularly important for occupational disease screening. It can identify potential patients early while avoiding excessive intervention for healthy workers.

[0166] 5. In terms of runtime, the method of this invention is much faster than manually operated neuropsychological tests, but slightly slower than simple biochemical index measurements. This reflects that the method achieves a good balance between information content and computational complexity, ensuring both the comprehensiveness and accuracy of predictions while meeting the efficiency requirements of practical applications.

[0167] In summary, these test results demonstrate that the method of this invention outperforms existing technologies in terms of accuracy, efficiency, and practicality, providing a more reliable and efficient new approach for predicting the neurotoxicity risk of aluminum-exposed workers. It not only improves the early screening effectiveness of occupational diseases but also provides a scientific basis for the development of personalized protective measures, thus playing a significant role in enhancing occupational health management.

[0168] It should be noted that the technical solution of the present invention is not limited to the specific embodiments described above. Based on the common technical knowledge of those skilled in the art, without departing from the basic principles and spirit of the present invention, specific parameters in the embodiments can be adjusted and optimized, such as sample size, sequencing depth, feature screening threshold, model structure parameters, etc., to achieve the best results in specific application environments. Furthermore, the above-mentioned algorithm modules can also be appropriately tailored and combined according to actual needs, or replaced by other mature algorithms. As long as the overall process can achieve the mapping from expression profile to risk prediction, it should be considered to fall within the protection scope of the present invention.

Claims

1. A method for predicting the risk of neurotoxicity of aluminum exposure based on LncRNA expression profile, characterized in that The procedure includes the following steps: collecting peripheral blood samples from aluminum-exposed workers and control groups; extracting total RNA from the peripheral blood samples, constructing a lncRNA sequencing library and performing high-throughput sequencing to obtain lncRNA expression profile data; screening for differentially expressed lncRNA biomarkers in the lncRNA expression profile; constructing an aluminum exposure neurotoxicity risk prediction model using machine learning algorithms; and using the prediction model to conduct risk assessment and stratified intervention for new aluminum-exposed individuals. The step of screening LncRNA biomarkers includes: preprocessing the LncRNA expression profile data to obtain an expression matrix. ,in Indicates sample LncRNA The expression level; the Boruta algorithm was used to initially screen candidate features, with a threshold set at . Iteration Next, the importance of the original features was evaluated by constructing random shadow features, resulting in a set of candidate LncRNA features. ,in The calculation formula is: ， In the formula, Original features As a shadow feature, The indicator function is used; further, a swarm intelligence optimization method is employed to screen the optimal LncRNA combination. Candidate features are encoded as binary vectors, the population is randomly initialized, and selection, crossover, and mutation are iteratively performed, using the Matthews correlation coefficient (MCC) as the fitness function, until the stopping condition is met, and the optimal combination is output. The objective function is: ， in, Describing the L0 norm, For target feature number For sample label vectors, For the first in the group Each individual, i.e., a binary coding vector of a candidate LncRNA combination; The steps for constructing the risk prediction model include: performing adaptive semi-supervised oversampling balancing on the selected LncRNA expression data, and for each minority class sample... , from its Neighbor Select appropriate samples New samples are generated using interpolation. ， By estimating the sample Local minority class probability density at Adaptive adjustment of the number of synthesized samples ， ， In the formula, For indicator functions For sample labels, For the minority class sample size, As a balance factor, For the bandwidth of the Gaussian kernel, To control the random coefficients of the interpolation positions and ensure they follow a uniform distribution in [0,1], The baseline oversampling factor is used; a classifier based on attention mechanism and multi-scale convolutional neural network AMCNN is constructed, using the LncRNA expression matrix as input to extract multi-scale local features: ， The sample representation is adaptively adjusted through an attention layer: ， ， In the formula, For the first Each convolutional kernel size, For attention layer parameters, , , , The weight matrix and bias vector of the fully connected layer at the output end; Perform global average pooling and generate sample prediction probabilities through a fully connected layer. ， Minimize the cross-entropy loss function to train network parameters : ， in, The true labels for the samples; The steps for training the prediction model further include: optimizing the hyperparameters using an evolutionary-gradient hybrid optimization algorithm, and dividing the hyperparameters into a discrete space. and continuous space Initial population II was formed by random sampling. Iteratively select, recombine, mutate, and perform gradient descent to update the population and output the optimal combination of hyperparameters. Discrete variables are processed using two-point crossover and random mutation, while continuous variables are processed using arithmetic crossover and Gaussian mutation, and are updated using gradient descent. In the formula, For learning rate, For loss function pairs The gradient; The method further includes a model interpretation step: explaining the decision-making mechanism of the prediction model using the SHAP framework, and measuring the contribution of each LncRNA feature to the model prediction based on the Shapley value; Shapley value Defined as features The weighted average of the marginal contribution under all characteristic arrangements: ， in, For feature subset, For the complete feature set, To consider only features Time model For the sample The predicted value; SHAP introduces a linear function Approximate the Shapley value and solve the additive characteristic attribution equation: ， ， In the formula, To simplify the number of input features, i.e., indicator vectors The dimension; As an indicator vector representing whether each feature appears, its first... The component corresponds to the first One feature; This represents the baseline value for the additive attribution model; (·) represents a sparse regularization term; The feature attribution weights to be solved are used to ultimately obtain the importance index of each LncRNA. .

2. The method as described in claim 1, characterized in that... The step of training the prediction model is followed by: The generalization performance of the prediction model is evaluated using K-fold cross-validation. The dataset is randomly divided into K mutually exclusive subsets. Each subset is used as the test set in turn, and the remaining K-1 subsets are used as the training set. The model performance is evaluated on the test set, and the average evaluation metric is calculated. The stability of the prediction model is evaluated using the Y-Randomization method. R random datasets are generated, the features are kept unchanged, the sample labels are randomly shuffled, the model is trained and its performance is evaluated, the performance distribution of the random model is obtained, and the performance of the real model is compared with it to evaluate the reliability of the model.

3. The method as described in claim 1, characterized in that, The application steps of the prediction model include: obtaining newly collected blood samples from aluminum-exposed workers, extracting total RNA and detecting lncRNA expression profiles; inputting the new sample expression profile data into the trained prediction model to obtain the risk prediction probability of the sample. : ， In the formula, For the lncRNA expression vector of the new sample, To optimize the obtained optimal model parameters, based on the predicted probability... With preset threshold The comparison divided the samples into high-risk groups. or low-risk group And take corresponding health management measures.

4. The method as described in claim 1, characterized in that... Prior to the step of constructing the risk prediction model, the method further includes performing batch normalization on the LncRNA expression matrix to reduce the impact of batch effects.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that... When the program is executed by the processor, it implements the aluminum exposure neurotoxicity risk prediction method according to any one of claims 1-4.

6. A device for predicting the neurotoxicity risk of aluminum exposure based on lncRNA expression profiling, used to implement the method according to any one of claims 1-4, characterized in that... ,include: The data acquisition module is used to collect peripheral blood samples from aluminum-exposed workers and control groups. The bioinformatics analysis module is used to extract total RNA from the blood sample, construct a lncRNA sequencing library, perform high-throughput sequencing, and analyze the lncRNA expression profile data. The LncRNA screening module is used to screen for significantly differentially expressed LncRNA biomarkers from the LncRNA expression profile, including preliminary screening using the Boruta algorithm and further selection of the optimal LncRNA combination using a swarm intelligence optimization algorithm. The data balancing module is used to balance the number of samples from different classes using the adaptive semi-supervised oversampling method ASEMI. The model training module is used to build a risk prediction model based on attention mechanism and multi-scale convolutional neural network AMCNN, and to optimize the model hyperparameters using the evolutionary gradient hybrid optimization algorithm EGHO; The model evaluation module is used to evaluate the generalization performance and stability of the prediction model using K-fold cross-validation and Y-Randomization methods. The model application module is used to input newly collected LncRNA expression profile data of aluminum-exposed individuals into the prediction model to predict their neurotoxicity risk probability, and to perform stratified management and intervention for individuals based on the prediction results. The model interpretation module is used to interpret the decision-making mechanism of the prediction model using the SHAP framework and to calculate the Shapley importance values ​​of key LncRNA biomarkers.

Images