A data processing model training method for lung cancer risk prediction

Abstract

This invention relates to the field of data processing technology and discloses a data processing model training method for lung cancer risk prediction. The method involves collecting historical data containing multiple lung cancer clustered familial units as training and validation sets, constructing a familial relationship graph based on the familial relationship data, and using a graph attention network to calculate the association strength to obtain a familial clustering index vector. Genetic marker data and environmental exposure data are processed separately to obtain genetic feature vectors and environmental exposure feature vectors. Multimodal feature fusion is performed on the familial clustering index vector, genetic feature vector, and environmental exposure feature vector to obtain a fused feature vector. The fused feature vector is used as input, and the disease label is used as a supervision signal to construct a TCN-SENet-Transformer model for training. This invention improves the model's ability to learn and fit lung cancer risk patterns.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically to a data processing model training method for lung cancer risk prediction. Background Technology

[0002] Lung cancer is one of the malignant tumors with the highest incidence and mortality rates worldwide. Early screening and diagnosis are key to improving patient survival rates. Currently, lung cancer risk prediction models are mostly based on single-dimensional data and do not fully consider family clustering characteristics. Lung cancer has a clear familial genetic tendency, and the disease status of relatives within the family has important reference value for individual risk assessment. However, existing technologies mostly use simple statistical methods to calculate family incidence rates, which cannot accurately capture the association and interaction effects between relatives.

[0003] Meanwhile, existing models suffer from insufficient multimodal data fusion: they fail to effectively integrate lung anatomical features and temporal dynamic features of environmental exposure, resulting in incomplete characterization of risk factors; during model training, hyperparameter optimization often employs methods such as grid search and simple Bayesian optimization, which suffer from low search efficiency and a tendency to get trapped in local optima; and the prediction models are mostly single-algorithm models, making it difficult to simultaneously capture the sequential dependencies and nonlinear interactions between features, thus requiring improvement in prediction accuracy. Summary of the Invention

[0004] The purpose of this invention is to solve the above-mentioned problems by designing a data processing model training method for lung cancer risk prediction.

[0005] This invention provides a data processing model training method for lung cancer risk prediction, the method comprising the following steps:

[0006] Historical data containing multiple lung cancer clusters in family units were collected as training and validation sets. The training set included genetic marker data of family members, quantified environmental exposure data, family relationship data, and disease status labels.

[0007] Based on the aforementioned family relationship data, a family relationship graph is constructed, and a graph attention network is used to calculate the weighted clustering coefficient and the strength of the association of kinship diseases for each node, thereby obtaining a family clustering index vector.

[0008] The genetic marker data and the environmental exposure data are processed separately to obtain genetic feature vectors and environmental exposure feature vectors;

[0009] Multimodal feature fusion is performed on the family clustering index vector, genetic feature vector, and environmental exposure feature vector to obtain a fused feature vector;

[0010] The TCN-SENet-Transformer model is constructed and trained by using the fused feature vector as input and the disease label as the supervision signal. During the training process, Bayesian optimization combined with the Memetic algorithm is used to adaptively adjust the model hyperparameters, with the goal of maximizing the validation set AUC value.

[0011] Optionally, in the first implementation of the present invention, the step of constructing a pedigree graph based on the pedigree data and using a graph attention network to calculate the weighted clustering coefficient and the strength of kinship disease association for each node to obtain a pedigree clustering index vector includes:

[0012] Each family member is treated as a node in the graph. An undirected pedigree graph is constructed with families as the unit. Initial weights are assigned to the edges between nodes based on the degree of kinship. Disease status information corresponding to each node is collected to obtain basic features.

[0013] Initialize the parameters of the graph attention network, construct the loss function, and update the parameters of the graph attention network using the backpropagation algorithm until the graph attention network converges. The graph attention network includes a feature mapping layer, an attention calculation module, and a dual-task output layer. The dual-task output layer is used to simultaneously output the individual disease probability of each node and the family clustering index vector of each node.

[0014] The feature mapping layer of a graph attention network maps the basic features of each node into a low-dimensional feature vector.

[0015] Obtain the neighboring nodes of each target node in the undirected kinship graph. Input the low-dimensional feature vector of the target node and the low-dimensional feature vector of each neighboring node into the attention calculation module. First, obtain the concatenated feature vector through linear transformation. Then, calculate the original attention coefficient between the target node and each neighboring node based on the concatenated feature vector.

[0016] The original attention coefficients are normalized using the Softmax function in the attention calculation module to obtain normalized attention coefficients. The normalized attention coefficients of each neighbor node are multiplied by the probability of illness of the neighbor node, and the sum is divided by the total number of neighbor nodes of the target node to obtain the weighted clustering coefficient of the target node. The normalized attention coefficients of each neighbor node are added to the probability of illness of the neighbor node to obtain the association strength value of the target node affected by the illness status of relatives.

[0017] By using the dual-task output layer of the graph attention network, the weighted clustering coefficient and association strength of each target node are combined in sequence to form the family clustering index vector corresponding to the target node. At the same time, the individual disease probability of each node is output for training supervision and prediction.

[0018] Optionally, in a second implementation of the present invention, the step of processing the genetic marker data and the environmental exposure data respectively to obtain the genetic feature vector and the environmental exposure feature vector includes:

[0019] For genetic marker data, the PRSice-2 algorithm was used to calculate polygenic risk scores, and the polygenic risk scores were used as genetic feature vectors.

[0020] Obtain environmental exposure sample data containing time series attributes, and obtain the training dataset after preprocessing. The preprocessing process includes at least outlier cleaning and timestamp alignment and normalization.

[0021] Construct a BiLSTM network and configure its hyperparameters, which include at least the number of hidden layer units, time step, batch size, and initial learning rate.

[0022] The training dataset is input into the BiLSTM network step by step. The error between the actual observations and the predicted values ​​output by the BiLSTM network is used as the objective. The Adam optimizer is used for backpropagation training. The loss change is monitored after each training round. The BiLSTM network is considered to have converged when the loss change no longer decreases for 10 consecutive rounds. The actual observations are the actual measurements of environmental exposure variables in the training dataset at each time step. The environmental exposure variables include at least the cumulative smoking index, alcohol intake, and residential industrial pollution index.

[0023] After normalizing the environmental exposure data, it is input into the trained BiLSTM network to obtain the environmental exposure feature vector.

[0024] Optionally, in a third implementation of the present invention, the step of calculating a polygenic risk score using the PRSice-2 algorithm on the genetic marker data and using the polygenic risk score as a genetic feature vector includes:

[0025] Obtain summary statistics on lung cancer-related SNPs, including at least the SNP name, chromosomal location, allele, effect size, and p-value.

[0026] Set a threshold range for the p-value of SNP screening, divide it into multiple consecutive p-value intervals, and for each p-value threshold interval, screen out SNPs with p-values ​​less than the current threshold from the statistical data and match them with SNPs in the genetic marker data to obtain the intersection SNPs.

[0027] Extract the effect values ​​of the intersection SNPs in the statistical data, and use the corresponding effect values ​​as the weights of the SNPs;

[0028] For each individual in the genetic marker data, calculate the polygenic risk score at the current p-value threshold based on the genotype and weight of their intersection SNPs;

[0029] The polygenic risk scores calculated for each individual under all p-value thresholds were standardized, and cross-validation was used to evaluate the association strength between the polygenic risk scores and lung cancer phenotypes under different p-value thresholds. The polygenic risk score corresponding to the p-value threshold with the strongest association was selected as the individual's final polygenic risk score.

[0030] The final polygenic risk scores of all individuals are arranged in order to form a genetic feature vector.

[0031] Optionally, in the fourth implementation of the present invention, the step of inputting the normalized environmental exposure data into the trained BiLSTM network to obtain the environmental exposure feature vector includes:

[0032] Z-score normalization is performed on each feature of the environmental exposure data. A one-dimensional convolutional layer is used to extract local features from the normalized data. Feature maps are generated by sliding the convolutional kernel along the time dimension.

[0033] The convolutional output is input into a bidirectional LSTM layer to process the forward and reverse directions of the sequence, and the hidden states of the two directions at each time step are connected.

[0034] After normalizing the output of each time step of the bidirectional LSTM layer, the Dropout layer is used to regularize the normalized output.

[0035] The attention mechanism is applied to the output of the bidirectional LSTM layer at all time steps, weighted and averaged to calculate the attention weights, and then weighted to obtain the attention feature vector:

[0036]

[0037]

[0038] in, Indicates attention weights, Indicates attention score, Represents an exponential function. Indicates the total number of time steps. Represents the normalization factor. Represents the attention feature vector. Indicates the first Output at each time step;

[0039] The attention feature vector is input into the fully connected layer for nonlinear transformation to generate the final environmental exposure feature vector.

[0040] Optionally, in a fifth implementation of the present invention, the step of performing multimodal feature fusion on the pedigree clustering index vector, genetic feature vector, and environmental exposure feature vector to obtain a fused feature vector includes:

[0041] By mapping the pedigree clustering index vector, genetic trait vector, and environmental exposure trait vector to the same dimension, the transformed pedigree clustering index vector is obtained. Transformed genetic feature vector Transformed environmental exposure feature vector ;

[0042] Transformed mode vectors , , The concatenation is performed to obtain the concatenated feature vector. and splicing feature vectors After performing a linear transformation, the intermediate vector is obtained. The attention weights for each modality are calculated by applying the Softmax function to the intermediate vector. , , intermediate vector The formula is:

[0043]

[0044] in, Represents the weight matrix. Indicates bias. Attention weights representing familial clustering characteristics Attention weights representing genetic traits Attention weights representing environmental exposure characteristics;

[0045] The transformed modal vectors are weighted and summed according to the attention weights to obtain the fused feature vector.

[0046] Optionally, in the sixth implementation of the present invention, the step of using the fused feature vector as input and the disease label as a supervision signal to construct the TCN-SENet-Transformer model for training includes:

[0047] A TCN module is constructed, which uses dilated causal convolutional layers to extract temporal features from the fused feature vector. After each convolution, the features are normalized and activated to output TCN features.

[0048] The SENet module is constructed, and the channel descriptors are obtained by global average pooling of the TCN features output by the TCN module. The channel attention weights are calculated through a fully connected layer containing compression ratio parameters. The channel attention weights are multiplied by the TCN features by channel to output the enhanced features.

[0049] A Transformer module is constructed, which takes the enhanced features output by the SENet module as the sequence input, calculates the feature association within the sequence through a multi-head self-attention mechanism, and outputs Transformer features after processing by a feedforward neural network. The feedforward neural network is a two-layer fully connected layer structure, with the first layer using the ReLU activation function and the second layer not using an activation function.

[0050] The Transformer features are input into a fully connected layer, and the predicted incidence rate of lung cancer is output through a Sigmoid activation function to construct the TCN-SENet-Transformer model.

[0051] Optionally, in the seventh implementation of the present invention, the step of adaptively adjusting the model hyperparameters using Bayesian optimization combined with the Memetic algorithm during training to optimize the model with the goal of maximizing the validation set AUC value includes:

[0052] Define the hyperparameter search space of the model to be optimized, which includes the number of convolutional kernels of the TCN module, the compression ratio of the SENet module, and the number of multi-head attention heads and the learning rate of the Transformer module;

[0053] The mapping relationship between the model hyperparameter search space and the validation set AUC is modeled based on the Gaussian process model. The hyperparameter combination with the largest expected improvement is selected as the initial solution to obtain the initial hyperparameter combination.

[0054] The Memetic algorithm is applied to the initial hyperparameter combination for local optimization. Neighborhood solutions are generated through mutation operators, and the AUC value of each neighborhood solution on the validation set is calculated to obtain the optimal neighborhood solution.

[0055] The optimal neighborhood solution is substituted into the TCN-SENet-Transformer model, with the training set data as input and the disease label as the supervision signal. The loss is calculated using the cross-entropy loss function, and the model parameters are updated through backpropagation.

[0056] After each training round, the AUC value is calculated using the validation set. If the current AUC value does not exceed the historical maximum value, the hyperparameters are repeatedly adjusted. If the AUC value does not improve after a preset number of rounds, training is stopped.

[0057] Optionally, in the eighth implementation of the present invention, the step of applying the Memetic algorithm to the initial hyperparameter combination for local optimization, generating neighborhood solutions through mutation operators, and calculating the AUC value of each neighborhood solution on the validation set to obtain the optimal neighborhood solution includes:

[0058] Determine the current combination of initialization hyperparameters to be optimized. , This indicates the number of hyperparameters and sets local optimization parameters, including the number of neighborhood solution generation. Variable Asynchronous Long Range Number of local search iterations , This represents the minimum length of the variable asynchronous operation. Indicates the maximum value of variable asynchronous length;

[0059] For initialization hyperparameter combination Perform the mutation operation to generate an initial neighborhood solution, at the... In the next iteration, the initial neighborhood solution Each hyperparameter is subjected to a preset mutation amount within a variable asynchronous length range to generate sub-neighborhood solutions. ,in ;

[0060] Calculate the sub-neighborhood solution If the validation set AUC value is better than the initial neighborhood solution... Then update for After reaching the required number of local search iterations, a neighborhood solution optimized by the local search is obtained.

[0061] Substitute the initial hyperparameter combination and the neighborhood solution optimized by local search into the TCN-SENet-Transformer model respectively, train the model using the same training set, and then calculate the corresponding AUC value using the validation set.

[0062] Compare the AUC values ​​of all solutions and select the solution with the largest AUC value as the optimal neighborhood solution for local optimization by the Memetic algorithm.

[0063] Optionally, in the ninth implementation of the present invention, the initialization hyperparameter combination... Performing mutation operations to generate initial neighborhood solutions includes:

[0064] For continuous hyperparameters in the initialization hyperparameter combination Calculate the mutated value using the following formula:

[0065]

[0066] in, Indicates the mutated th New values ​​for each hyperparameter Indicating the first hyperparameter in the initialization hyperparameter combination Initial values ​​of each hyperparameter Indicates variable asynchronous length, Indicates the direction factor.

[0067] Compared with the prior art, the present invention has the following advantages and technical effects:

[0068] 1. By constructing a pedigree graph and combining it with a graph attention network, the weighted clustering coefficient of nodes and the strength of the association between relatives' illnesses can be calculated. This allows for a deeper exploration of risk association patterns among family members, making the characterization of pedigree clustering characteristics more targeted and accurate, and providing more reliable pedigree dimension support for risk prediction.

[0069] 2. By extracting features from genetic marker data and time-series environmental exposure data respectively, and combining them with pedigree clustering index vectors to carry out multimodal feature fusion, we can comprehensively cover the three core risk dimensions of genetics, environment and pedigree, and achieve a three-dimensional and all-round characterization of lung cancer risk factors, thereby improving the completeness of risk factor representation.

[0070] 3. Adaptive hyperparameter adjustment is achieved by combining Bayesian optimization with the Memetic algorithm. This approach leverages the global search advantage of Bayesian optimization to improve optimization efficiency while utilizing the local fine-grained search capability of the Memetic algorithm to discover the optimal hyperparameter combination. The goal is to maximize the AUC value on the validation set, thereby significantly improving the scientific rigor and effectiveness of hyperparameter optimization.

[0071] 4. Construct an integrated TCN-SENet-Transformer model, utilizing TCN to capture temporal dependencies of features, SENet to enhance the channel response of key features, and Transformer to mine nonlinear correlations between features. This solves the problem that a single algorithm model cannot capture multiple types of features simultaneously, significantly improving the model's ability to learn and fit lung cancer risk patterns, making the prediction results more accurate and reliable.

[0072] 5. Improve the efficiency and accuracy of identifying high-risk groups for lung cancer, provide data support for developing individualized lung cancer screening and follow-up programs, help improve the efficiency of early screening and diagnosis of lung cancer, and have important practical significance for reducing lung cancer mortality and improving patient survival rates. Attached Figure Description

[0073] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0074] Figure 1A schematic diagram of the first embodiment of the data processing model training method for lung cancer risk prediction provided by the present invention;

[0075] Figure 2 A schematic diagram of a second embodiment of the data processing model training method for lung cancer risk prediction provided by the present invention;

[0076] Figure 3 This is a schematic diagram of a third embodiment of the data processing model training method for lung cancer risk prediction provided by the present invention. Detailed Implementation

[0077] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” or “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, apparatus, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0078] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 A schematic diagram of the first embodiment of the data processing model training method for lung cancer risk prediction provided by this invention. The method specifically includes the following steps:

[0079] Step 101: Collect historical data containing multiple lung cancer clustered family units as training and validation sets;

[0080] In this embodiment, the training set and validation set include genetic marker data of family members, quantified environmental exposure data, family relationship data, and labels indicating whether the disease has occurred.

[0081] In this embodiment, the quantified environmental exposure data refers to the data obtained after assessing and quantifying various environmental risk factors that an individual is exposed to during their life cycle. These environmental exposure factors mainly include, but are not limited to, long-term smoking history, air pollution levels, drinking habits, intake of carcinogens in the diet, and industrial pollution in the place of residence. Information such as the time, frequency, and intensity of exposure behaviors are obtained, and combined with regional environmental monitoring data, the qualitative descriptions are transformed into calculable numerical indicators, such as "cumulative smoking index" and "annual average PM2.5 exposure dose". The quantified data can effectively reflect the degree of impact of environmental factors on individual health.

[0082] In this embodiment, historical data from multiple lung cancer clustered family units were selected and confirmed as the model training and validation sets. The selected family units must meet the following core conditions: at least one lung cancer patient diagnosed by pathology, information on relatives of three generations or more, completion of chest CT screening data collection, and complete environmental exposure factor investigation records and peripheral blood biomarker detection data. This ensures that the data can comprehensively support the analysis of the interaction patterns between genetic and environmental factors in the family population.

[0083] Step 102: Construct a pedigree graph based on pedigree data, and use a graph attention network to calculate the weighted clustering coefficient and the strength of the association of kinship diseases for each node, so as to obtain the pedigree clustering index vector;

[0084] Step 103: Process the genetic marker data and environmental exposure data respectively to obtain the genetic feature vector and the environmental exposure feature vector;

[0085] In this embodiment, the PRSice-2 algorithm is used to calculate the polygenic risk score for the genetic marker data, and the polygenic risk score is used as the genetic feature vector. The environmental exposure data is normalized and then input into the trained BiLSTM network to obtain the environmental exposure feature vector. The PRSice-2 algorithm is a statistical tool used to calculate the polygenic risk score (PRS) to assess the risk of an individual developing a complex disease due to the accumulation of multiple common genetic variations.

[0086] In this embodiment, environmental exposure sample data containing time-series attributes is acquired, and preprocessed operations such as outlier cleaning, timestamp alignment, and numerical normalization are performed to construct a training dataset. A BiLSTM network is constructed, the structure of which includes: a bidirectional LSTM layer to capture the forward and backward temporal dependencies of the environmental exposure data; the weight parameters of the BiLSTM network are initialized, and hyperparameters are configured: the number of hidden layer units, time step size, batch size, and initial learning rate, and the initial state of the bidirectional LSTM is set to a zero vector; the preprocessed time-series data is input into the BiLSTM network step by step, and the forward and backward hidden states of the bidirectional LSTM are concatenated; the error between the actual observations and the predicted values ​​output by the BiLSTM network is used as the target, and the training dataset is trained... During training, the BiLSTM network outputs a predicted value at each time step. The goal of training is to make these predicted output values ​​close to the actual observed values ​​at the corresponding time steps in the training dataset. The actual observed values ​​are the actual measurements of environmental exposure variables in the training dataset at each time step. Environmental exposure variables include at least the cumulative smoking index, alcohol intake, and residential industrial pollution index. This error is used as the loss function, and backpropagation is driven by the Adam optimizer to continuously adjust the network parameters of the BiLSTM. The change in loss is monitored after each training round, and the BiLSTM network is considered to have converged when the change in loss no longer decreases for 10 consecutive rounds. After training, the final hidden state of the BiLSTM network is output as the feature vector of environmental exposure.

[0087] In this embodiment, the final hidden state refers to the concatenation result of the hidden state vectors output by the bidirectional LSTM layer at the last time step. For a time series of length T, the forward LSTM transmits information sequentially from time step 1 to T, and its hidden state at time step T represents the accumulation of information from the past to the present. The backward LSTM transmits information in reverse from time step T to 1, and its hidden state at time step T contains the influence of future information on the current moment. By concatenating the hidden states of these two directions at the last time step, a vector that integrates complete temporal context information is obtained, which is the final hidden state of the BiLSTM network.

[0088] In this embodiment, a summary statistical data set of lung cancer-related SNPs is obtained, wherein the statistical data includes at least the SNP name, chromosomal location, allele, effect value, and p-value; a p-value threshold range for SNP screening is set, and multiple consecutive p-value intervals are divided. For each p-value threshold interval, SNPs with p-values ​​less than the current threshold are screened from the statistical data and matched with SNPs in the genetic marker data to obtain intersection SNPs; the effect values ​​of the intersection SNPs in the statistical data are extracted, and the corresponding effect values ​​are used as the weights of the SNPs; for each individual in the genetic marker data, a polygenic risk score under the current p-value threshold is calculated based on the genotype and weight of its intersection SNPs; the polygenic risk scores calculated for each individual under all p-value thresholds are standardized, and cross-validation is used to evaluate the association strength between the polygenic risk scores and the lung cancer phenotype under different p-value thresholds, and the polygenic risk score corresponding to the p-value threshold with the strongest association is selected as the individual's final polygenic risk score; the final polygenic risk scores of all individuals are arranged sequentially to form a genetic feature vector.

[0089] In this embodiment, summary statistics of lung cancer-related single nucleotide polymorphisms (SNPs) validated by large-scale lung cancer association studies are obtained. The SNP name is used to uniquely identify each SNP site, the chromosomal location specifies the exact coordinates of the SNP in the genome, the alleles distinguish between effect alleles and non-effect alleles, and the effect value, such as the β coefficient or odds ratio (OR), quantifies the degree of influence of SNPs on the risk of lung cancer incidence. The p-value reflects the statistical significance of the association between SNPs and lung cancer phenotypes. This ensures that the data has undergone strict quality control and is free from problems such as significant batch effects and genotyping errors.

[0090] Based on the distribution characteristics of the association strength of lung cancer-related SNPs, a p-value threshold range for SNP screening is set, typically from 5e-8 (strong association) to 1 (weak association). This range is divided into multiple consecutive p-value intervals using equidistant or logarithmic scale division methods, such as 5e-8, 1e-6, 1e-4, 1e-2, 0.1, and 1 intervals. For each p-value threshold interval, using the current interval's p-value threshold as the screening condition, SNP sites with p-values ​​less than the threshold are selected from the acquired SNP summary statistics. Simultaneously, all SNP site information contained in the genetic marker data is extracted. The selected SNPs are then compared with the SNPs in the genetic marker data using allele consistency verification and name matching. SNPs with mismatched alleles or names are removed, ultimately yielding the SNPs shared by both.

[0091] For each intersection SNP corresponding to a p-value threshold interval, the effect value corresponding to each intersection SNP is accurately extracted from the aggregated statistical data of lung cancer-related SNPs. If the aggregated data contains multiple effect value types, the β coefficient is uniformly selected as the standard effect value. The effect value of each extracted intersection SNP is directly used as the weight of that SNP in the current p-value threshold interval. The positive or negative sign of the weight reflects the promoting or inhibiting effect of the SNP on lung cancer risk, and the absolute value of the weight reflects the strength of the SNP's influence on lung cancer risk.

[0092] For each family member in the genetic marker data, the genotypes of their intersecting SNPs are first standardized and encoded, with non-effect allele homozygotes encoded as 0, heterozygotes as 1, and effect allele homozygotes as 2. Then, for the current p-value threshold range, the genotype codes and weights of all intersecting SNPs of an individual are multiplied one by one according to the calculation logic of polygenic risk score (PRS) = Σ (genotype code of an individual's SNP × weight of the corresponding SNP). All product results are summed to obtain the polygenic risk score of the individual under the current p-value threshold range.

[0093] Z-score standardization was performed on the polygenic risk scores calculated for each individual across all p-value threshold intervals to eliminate dimensional differences in PRS across different threshold intervals. A 5-fold stratified cross-validation method was used, with the lung cancer incidence label as the dependent variable and the standardized PRS as the independent variable, to evaluate the association strength between PRS and lung cancer phenotype at each p-value threshold. The association strength indices corresponding to all p-value thresholds were compared, and the p-value threshold with the highest association strength was selected. The standardized PRS of individuals at this threshold was then used as their final polygenic risk score.

[0094] Following the original data order of family members in the training or validation set, the final polygenic risk scores of all individuals are sequentially arranged to form a one-dimensional genetic feature vector. The dimension of this vector is consistent with the number of individuals in the training or validation set, and each element in the vector corresponds to the quantitative result of the genetic risk of lung cancer for an individual.

[0095] In this embodiment, each feature of the environmental exposure data is Z-score normalized, and a one-dimensional convolutional layer is used to extract local features from the normalized data. The feature map is generated by sliding the convolutional kernel in the time dimension.

[0096] The convolutional output is input into a bidirectional LSTM layer to process the forward and reverse directions of the sequence, and the hidden states of the two directions at each time step are connected.

[0097] After normalizing the output of each time step of the bidirectional LSTM layer, the Dropout layer is used to regularize the normalized output.

[0098] The attention mechanism is applied to the output of the bidirectional LSTM layer at all time steps, weighted and averaged to calculate the attention weights, and then weighted to obtain the attention feature vector:

[0099]

[0100]

[0101] in, Indicates attention weights, Indicates attention score, Represents an exponential function. Indicates the total number of time steps. Represents the normalization factor. Represents the attention feature vector. Indicates the first Output at each time step;

[0102] The attention feature vector is input into the fully connected layer for nonlinear transformation to generate the final environmental exposure feature vector.

[0103] Step 104: Perform multimodal feature fusion on the family clustering index vector, genetic feature vector, and environmental exposure feature vector to obtain a fused feature vector;

[0104] In this embodiment, the pedigree clustering index vector, genetic feature vector, and environmental exposure feature vector are mapped to the same dimension to obtain the transformed pedigree clustering index vector. Transformed genetic feature vector Transformed environmental exposure feature vector ;

[0105] Transformed mode vectors , , The concatenation is performed to obtain the concatenated feature vector. and splicing feature vectors After performing a linear transformation, the intermediate vector is obtained. The attention weights for each modality are calculated by applying the Softmax function to the intermediate vector. , , intermediate vector The formula is:

[0106]

[0107] in, Represents the weight matrix. Indicates bias. Attention weights representing familial clustering characteristics Attention weights representing genetic traits Attention weights representing environmental exposure characteristics;

[0108] The transformed modal vectors are weighted and summed according to the attention weights to obtain the fused feature vector.

[0109] In this embodiment, the calculation process of the Softmax function is to first perform an exponential operation on each element in an input vector containing multiple real numbers, and then divide each exponential result by the sum of all exponential results, thereby converting the original value into a probability distribution. The resulting output vector has each element between 0 and 1, and the sum of all elements is 1.

[0110] In this embodiment, a linear transformation is used to map the pedigree clustering index vector, genetic feature vector, and environmental exposure feature vector to the same dimension. A weight matrix is ​​set for each input vector of different dimensions. The two-dimensional pedigree clustering index vector is mapped to the d dimension through a 2×d matrix, and the one-dimensional genetic feature vector and environmental exposure feature vector are mapped to the same d dimension through a 1×d matrix, respectively, thus realizing the alignment and unified representation of features of different original dimensions.

[0111] Step 105: Using the fused feature vector as input and the disease label as the supervision signal, construct the TCN-SENet-Transformer model for training. During the training process, Bayesian optimization combined with the Memetic algorithm is used to adaptively adjust the model hyperparameters, with the goal of maximizing the validation set AUC value to optimize the model.

[0112] In this embodiment, a hyperparameter search space for the model to be optimized is defined, which includes the number of convolutional kernels in the TCN (Temporal Convolutional Network) module, the compression ratio of the SENet (Squeeze-and-Excitation Network) module, and the number of multi-head attention heads and learning rate of the Transformer module. A mapping relationship between the model hyperparameter search space and the validation set AUC is modeled based on a Gaussian process model. The hyperparameter combination with the largest expected improvement is selected as the initial solution, resulting in the initial hyperparameter combination. The Memetic Algorithm is applied to the initial hyperparameter combination for local optimization. Neighborhood solutions are generated through mutation operators, and the AUC (Area Under the Validation Set) of each neighborhood solution is calculated. The optimal neighborhood solution is obtained by taking the area under the curve (AUC). The optimal neighborhood solution is then substituted into the TCN-SENet-Transformer model, using the training set data as input and the disease label as the supervision signal. The loss is calculated using the cross-entropy loss function, and the model parameters are updated through backpropagation. After each training round, the AUC value is calculated using the validation set. If the current AUC value does not exceed the historical maximum value, the hyperparameters are repeatedly adjusted. If the AUC value does not improve after a preset number of rounds, training is stopped.

[0113] In this embodiment, the number of convolutional kernels in the TCN module is set to a search range of 32~256 based on the requirements of temporal feature extraction, with candidate values ​​divided by a stride of 32; the compression ratio of the SENet module is set to a search range of 8~32 based on the requirements of channel feature enhancement, with candidate values ​​divided by a stride of 4; the number of multi-head attention heads in the Transformer module is set to a search range of 4~16 based on the feature association capture capability, with candidate values ​​divided by a stride of 2; and the learning rate is set to a search range of 1e-5~1e-3 considering model convergence efficiency and stability, with candidate values ​​divided by a logarithmic scale.

[0114] A Gaussian process model is used to probabilistically model the mapping relationship between the model's hyperparameter search space and the validation set AUC values. Hyperparameter combinations are used as input variables, and validation set AUC values ​​are used as output variables. The kernel function of the Gaussian process is used to characterize the correlation between hyperparameters and the nonlinear relationship between hyperparameters and AUC values. In the modeling process, a small number of initial hyperparameter combinations are randomly selected from the hyperparameter search space, and their corresponding validation set AUC values ​​are calculated as training data to train the Gaussian process model. Then, the expected AUC improvement of each unevaluated hyperparameter combination is calculated using the expected improvement criterion. The hyperparameter combination with the largest expected improvement is selected as the initial solution. This initial solution takes into account the global search scope and is likely close to the optimal hyperparameter region, thus obtaining the final initialized hyperparameter combination.

[0115] The Memetic algorithm is applied to the obtained initial hyperparameter combinations for local optimization. For discrete hyperparameters, a random neighborhood mutation operator is used to randomly select neighboring values ​​from the candidate value set of the corresponding hyperparameter as the mutated values. All generated neighborhood solutions are substituted into the TCN-SENet-Transformer model. After training with the training set, the AUC value of each neighborhood solution on the validation set is calculated. The AUC values ​​of all neighborhood solutions are compared, and the neighborhood solution with the largest AUC value is selected as the optimal neighborhood solution.

[0116] The hyperparameter combinations corresponding to the obtained optimal neighborhood solutions are configured one by one into the corresponding modules of the TCN-SENet-Transformer model to complete the initialization of the model hyperparameters. The fused feature vector of the training set is used as input data, and the lung cancer incidence labels of family members (0 indicates no disease and 1 indicates disease) are used as supervision signals. The binary cross-entropy loss function is used to calculate the loss value between the model prediction value and the true incidence label. The gradient of the loss function with respect to the parameters of each layer of the model is calculated through the backpropagation algorithm. The Adam optimizer is used to update the model parameters along the gradient descent direction to achieve iterative optimization of the model parameters.

[0117] After each round of model parameter updates, the fused feature vector of the validation set is input into the TCN-SENet-Transformer model in the current training state to obtain the predicted lung cancer incidence probability value for each sample in the validation set. Based on the predicted value and the true incidence label of the validation set, the validation set AUC value for the current round is calculated. The current AUC value is compared with the historical maximum AUC value recorded during training. If the current AUC value does not exceed the historical maximum value, the model is re-modeled based on the Gaussian process model, a new initial hyperparameter combination is obtained, and the model hyperparameters and parameters are updated after local optimization using the Memetic algorithm, and training continues iteratively. At the same time, a preset number of rounds is set, such as 10 rounds. If the validation set AUC value does not exceed the historical maximum value after 10 consecutive rounds of training, the model is considered to have converged, the training process is stopped, and the model parameters and hyperparameter combinations in the current training state are saved.

[0118] In this embodiment, the current combination of initialization hyperparameters to be optimized is determined. , This indicates the number of hyperparameters and sets local optimization parameters, including the number of neighborhood solution generation. Variable Asynchronous Long Range Number of local search iterations , This represents the minimum length of the variable asynchronous operation. Indicates the maximum value of variable asynchronous length; for the combination of initialization hyperparameters Perform the mutation operation to generate an initial neighborhood solution, at the... In the next iteration, the initial neighborhood solution Each hyperparameter is subjected to a preset mutation amount within a variable asynchronous length range to generate sub-neighborhood solutions. ,in ; Calculate the sub-neighborhood solution If the validation set AUC value is better than the initial neighborhood solution... Then update for After reaching the required number of local search iterations, a neighborhood solution optimized by local search is obtained. The initial hyperparameter combination and the neighborhood solution optimized by local search are substituted into the TCN-SENet-Transformer model, and the model is trained using the same training set. The corresponding AUC value is then calculated using the validation set. The AUC values ​​of all solutions are compared, and the solution with the largest AUC value is selected as the optimal neighborhood solution for local optimization by the Memetic algorithm.

[0119] In this embodiment, the focus is on continuous hyperparameters in the initialization hyperparameter combination. Calculate the mutated value using the following formula:

[0120]

[0121] in, Indicates the mutated th New values ​​for each hyperparameter Indicating the first hyperparameter in the initialization hyperparameter combination Initial values ​​of each hyperparameter Indicates variable asynchronous length, Indicates the direction factor.

[0122] In this embodiment, a trained and optimized TCN-SENet-Transformer model is obtained. This model has fully learned the association patterns between multi-dimensional risk factors such as familial clustering, genetics, and environment and the incidence of lung cancer. The new individual familial data received by the model includes familial relationship data of the new individual and their relatives, peripheral blood biomarker detection data of the new individual, and time-series environmental exposure records of the new individual over a recent period. After receiving the data, it is first cleaned according to the data preprocessing standards of the training phase, such as missing value imputation and format standardization. Then, the familial clustering index vector is obtained sequentially through familial relationship graph construction and graph attention network calculation, and the PRSice-2 algorithm is used to calculate... Genetic feature vectors are obtained, and environmental exposure feature vectors are processed through a BiLSTM network. Then, multimodal attention fusion is performed on the three types of feature vectors to obtain a standardized fused feature vector. The fused feature vector is input into the TCN-SENet-Transformer model. The TCN module extracts temporal dependent features, the SENet module enhances the response of key feature channels, and the Transformer module mines the nonlinear correlation between features. Finally, the features are mapped to values ​​in the 0-1 range through a fully connected layer and a Sigmoid activation function. These values ​​are the predicted lung cancer incidence probability values ​​for new individuals, intuitively quantifying the lung cancer risk of new individuals and providing data support for identifying high-risk groups for early lung cancer screening and developing individualized screening programs.

[0123] In this embodiment of the invention, historical data containing multiple lung cancer clustered family units are collected as training and validation sets. A family relationship graph is constructed based on the family relationship data, and a graph attention network is used to calculate the weighted clustering coefficient and the strength of kinship disease association for each node, resulting in a family clustering index vector. Genetic marker data and environmental exposure data are processed separately to obtain genetic feature vectors and environmental exposure feature vectors. Multimodal feature fusion is performed on the family clustering index vector, genetic feature vector, and environmental exposure feature vector to obtain a fused feature vector. The fused feature vector is used as input, and the disease label is used as a supervision signal to construct a TCN-SENet-Transformer model for training. During the training process, Bayesian optimization combined with the Memetic algorithm is used to adaptively adjust the model hyperparameters, with the goal of maximizing the validation set AUC value. This invention improves the efficiency and accuracy of identifying high-risk groups for lung cancer and helps to improve the efficiency of early screening and diagnosis of lung cancer.

[0124] Please see Figure 2 A schematic diagram of a second embodiment of the data processing model training method for lung cancer risk prediction provided by the present invention is shown below. The method includes:

[0125] Step 201: Construct an undirected pedigree graph with each family member as a node in the graph. Assign initial weights to the edges between nodes based on the degree of kinship. Collect disease status information corresponding to each node to obtain basic features.

[0126] In this embodiment, each family member is treated as an independent node in the graph. An undirected pedigree graph is constructed using a single lung cancer cluster family as an independent unit to ensure that the graphs of different families are independent and do not overlap or become confused. Initial weights are assigned to the connecting edges between nodes based on the degree of blood kinship. The weights of first-degree kinship edges such as parents-children and spouses are set to 0.8, the weights of second-degree kinship edges such as siblings and grandparents-grandchildren are set to 0.5, and the weights of third-degree kinship edges such as uncles and nephews are set to 0.3, thereby quantifying the closeness of kinship. At the same time, the disease status information corresponding to each node is collected, with a pathological diagnosis of lung cancer recorded as 1 and an undiagnosed diagnosis recorded as 0. The disease status information of the nodes and the kinship information corresponding to the edge weights are used together as the basic features for constructing the pedigree graph.

[0127] Step 202: Initialize the parameters of the graph attention network, construct the loss function, and update the graph attention network parameters using the backpropagation algorithm until the graph attention network converges;

[0128] In this embodiment, the core parameters of the initialized graph attention network include the node feature embedding matrix, the attention weight matrix, and the bias vector. The dimension of the embedding matrix is ​​preset according to the dimension of the basic features, and the attention weight matrix and the bias vector are generated by random initialization and limited to a reasonable range of values. The basic features of each node are mapped through a preset linear transformation function, which transforms the high-dimensional basic features into a low-dimensional dense feature vector, thereby reducing the complexity of subsequent calculations and retaining the core information in the basic features that is related to lung cancer risk.

[0129] In this embodiment, the graph attention network includes a feature mapping layer, an attention calculation module, and a dual-task output layer. The dual-task output layer is used to simultaneously output the individual disease probability of each node and the familial clustering index vector of each node. During the training of the graph attention network, the parameters in the graph attention network are randomly initialized. The low-dimensional feature vector is extracted through the feature mapping layer in the graph attention network. This layer performs linear transformations or nonlinear mappings on the basic features of each node, compressing them into a low-dimensional representation suitable for graph attention network processing. This mapping process itself is learnable and is jointly optimized with the attention mechanism during training, so that the extracted low-dimensional features can retain individual information. In the attention calculation module, the low-dimensional feature vectors of the target node and its neighboring nodes are concatenated to generate the original attention coefficients. Then, these original attention coefficients are normalized using the Softmax function to obtain the normalized attention coefficients of each neighboring node. Subsequently, the normalized attention coefficients are compared with the disease status of the corresponding neighbors. The weighted clustering coefficient of a node is obtained by multiplying, summing, and dividing by the total number of neighbors. The normalized attention coefficients of each neighboring node are then summed to obtain the association strength value of the target node affected by the disease status of its relatives. The dual-task output layer of the graph attention network receives the result processed by the attention calculation module and maps it to the final familial clustering index vector. It also outputs the individual disease probability of each node for training supervision and prediction. The familial clustering index vector represents the result of the target node in the familial relationship based on the interaction between neighboring nodes and the influence of disease status. To optimize the model, a multi-task loss function is constructed, which includes the cross-entropy loss between the individual disease probability and the true disease label, and the fitting loss between the familial clustering index and the familial clustering degree statistical label. The gradient of the loss with respect to the network parameters is calculated layer by layer through the backpropagation algorithm, and the parameters are updated with the help of the optimizer. This iterative process continues, and the performance of the model is evaluated after each round of training. When the validation performance no longer improves significantly, the graph attention network is considered to have converged, and training is terminated.

[0130] Step 203: Use the feature mapping layer of the graph attention network to map the basic features of each node into a low-dimensional feature vector;

[0131] Step 204: Obtain the neighboring nodes of each target node in the undirected kinship graph. Input the low-dimensional feature vector of the target node and the low-dimensional feature vector of each neighboring node into the attention calculation module. First, obtain the concatenated feature vector through linear transformation, and then calculate the original attention coefficient between the target node and each neighboring node based on the concatenated feature vector.

[0132] In this embodiment, for each node in the undirected kinship graph, it is taken as the target node. The graph structure is traversed to obtain all nodes directly connected to the target node through edges, i.e., the neighbor nodes of the target node. In a fixed order, the low-dimensional feature vector of the target node comes first and the low-dimensional feature vector of the neighbor nodes comes last. The low-dimensional feature vector of the target node is concatenated with the low-dimensional feature vector of each neighbor node in turn to form a high-dimensional concatenated vector with a length equal to the sum of the dimensions of the two. The concatenated vector is input into a preset linear transformation layer, and matrix multiplication is used to transform the high-dimensional concatenated vector into a concatenated feature vector with uniform dimensions.

[0133] In this embodiment, the concatenated feature vector is linearly calculated by the attention calculation module and then substituted into the LeakyReLU activation function for nonlinear transformation. This suppresses the gradient vanishing problem while enhancing the model's ability to capture feature association information. A single numerical value is output through the output layer, which is the original attention coefficient between the target node and its corresponding neighboring nodes in the association of lung cancer risk.

[0134] Step 205: Normalize the original attention coefficients using the Softmax function in the attention calculation module to obtain normalized attention coefficients. Multiply the normalized attention coefficient of each neighbor node by the probability of illness of the neighbor node, sum the results, and divide by the total number of neighbor nodes of the target node to obtain the weighted clustering coefficient of the target node. Add the normalized attention coefficient of each neighbor node to the probability of illness of the neighbor node to obtain the association strength value of the target node affected by the illness status of relatives.

[0135] In this embodiment, the Softmax function is used to normalize all the original attention coefficients corresponding to each target node. The normalized attention coefficient is obtained by calculating the ratio of each original attention coefficient to the sum of all original attention coefficients, ensuring that the sum of the normalized attention coefficients corresponding to all neighboring nodes is 1. The normalized attention coefficient of each neighboring node is multiplied element-wise with the disease status information (1 or 0) corresponding to that neighboring node, and then all the multiplication results are summed to obtain the disease association weighted sum of the target node's neighbors. The weighted sum is divided by the total number of neighboring nodes of the target node to obtain the weighted clustering coefficient of the target node.

[0136] In this embodiment, all neighboring nodes of each target node are filtered out, and neighboring nodes with a disease status information of 1 are selected as the diseased neighboring nodes of the target node. The normalized attention coefficient corresponding to each diseased neighboring node is extracted, and the normalized attention coefficients of all diseased neighboring nodes are summed according to the numerical accumulation rule. The accumulated result is the association strength value of the target node affected by the disease status of relatives. The association strength value quantifies the potential impact of the diseased relatives on the risk of lung cancer in the target node.

[0137] Step 206: Through the dual-task output layer of the graph attention network, the weighted clustering coefficient and association strength of each target node are combined in sequence to form the family clustering index vector corresponding to the target node. At the same time, the individual disease probability of each node is output for training supervision and prediction.

[0138] In this embodiment, the dual-task output layer combines the weighted clustering coefficient and association strength value, which reflect the clustering characteristics of family risk, to generate a family clustering index vector corresponding to each target node, thereby quantifying the lung cancer incidence association characteristics of the node in the entire family network. On the other hand, it simultaneously outputs the individual incidence probability of each node, which is directly aligned with the supervision label in the training phase.

[0139] Please see Figure 3 A schematic diagram of the third embodiment of the data processing model training method for lung cancer risk prediction provided by the present invention, the method comprising:

[0140] Step 301: Construct the TCN module, use dilated causal convolutional layers to extract temporal features from the fused feature vectors, and after each convolutional layer, normalize and activate the features to output the TCN features;

[0141] In this embodiment, the TCN module uses dilated causal convolutional layers to extract temporal features from the input fused feature vector. Based on the distribution characteristics of environmental exposure temporal information and familial risk association temporal features in the fused features, the preset convolutional kernel size is 3, and the dilation rate increases in powers of 2 to ensure that the receptive field is expanded without increasing the computational load. At the same time, the causal convolutional characteristics prevent the leakage of temporal information. After each dilated causal convolutional layer is completed, batch normalization is first performed to eliminate the influence of feature distribution offset on model training. Then, the LeakyReLU activation function is substituted for nonlinear transformation to enhance the model's ability to capture complex association patterns in temporal features. Finally, the TCN features enhanced with temporal features are output.

[0142] Step 302: Construct the SENet module, perform global average pooling on the TCN features output by the TCN module to obtain channel descriptors, calculate channel attention weights through a fully connected layer containing compression ratio parameters, multiply the channel attention weights by the TCN features by channel, and output the enhanced features.

[0143] In this embodiment, the SENet module first performs global average pooling on the TCN features output by the TCN module. For each channel of the TCN feature, it calculates the average value of all feature values ​​under that channel, compressing the high-dimensional features of each channel into a single value to form a channel descriptor that can represent the global information of the channel. Then, the channel descriptor is input into a fully connected layer network containing a compression ratio parameter. For example, if the preset compression ratio is 16, the channel dimension is first compressed to 1 / 16 of the original dimension through the first fully connected layer, and then the dimension is restored to the original number of channels through the second fully connected layer to obtain the channel attention weights corresponding to each channel. Finally, the attention weights of each channel are multiplied element-wise with the feature map of the corresponding channel in the TCN feature to strengthen important feature channels and suppress irrelevant feature channels, outputting enhanced features with stronger feature expression capabilities.

[0144] Step 303: Construct the Transformer module, take the enhanced features output by the SENet module as the sequence input, calculate the feature association within the sequence through the multi-head self-attention mechanism, and output the Transformer features after processing by the feedforward neural network;

[0145] In this embodiment, the enhanced features output by the SENet module are organized into a sequence according to the time dimension or feature association dimension and used as input to the Transformer module. The Transformer module uses a multi-head self-attention mechanism to calculate the feature association within the sequence. The preset number of attention heads is 8. First, the enhanced features are split into 8 feature subspaces of different dimensions through linear transformation. In each subspace, the query, key, and value matrices are calculated. The feature association weight is obtained by multiplying the transpose of the query and key matrices and normalizing them. Then, it is multiplied with the value matrix to obtain the feature output of a single attention head. Finally, the outputs of the 8 attention heads are concatenated and subjected to linear transformation to obtain the multi-head self-attention features. The features are then input into a feedforward neural network for further processing. The feedforward neural network has a two-layer fully connected structure. The first fully connected layer uses the ReLU activation function for non-linear mapping, and the second fully connected layer does not use an activation function to maintain the linear output of the features. Finally, the Transformer features enhanced by feature association are obtained.

[0146] Step 304: Input the Transformer features into the fully connected layer, and output the predicted value of lung cancer incidence rate through the Sigmoid activation function to construct the TCN-SENet-Transformer model.

[0147] In this embodiment, the Transformer features output by the Transformer module are input into a preset fully connected layer. This fully connected layer maps the high-dimensionality of the Transformer features to a 1-dimensional feature vector through a linear transformation, achieving dimensionality compression from high-dimensional features to predicted values. Subsequently, the 1-dimensional feature vector is substituted into the Sigmoid activation function for calculation, mapping the feature values ​​to the 0-1 interval to obtain a probability prediction value that can directly characterize the likelihood of a new individual developing lung cancer. Through the ordered combination of the TCN module, SENet module, Transformer module, fully connected layer, and activation function, the TCN-SENet-Transformer model is constructed.

[0148] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A data processing model training method for lung cancer risk prediction, characterized in that, The method includes the following steps: Historical data containing multiple lung cancer clusters in family units were collected as training and validation sets. The training set included genetic marker data of family members, quantified environmental exposure data, family relationship data, and disease status labels. Based on the aforementioned family relationship data, a family relationship graph is constructed, and a graph attention network is used to calculate the weighted clustering coefficient and the strength of the association of kinship diseases for each node, thereby obtaining a family clustering index vector. The genetic marker data and the environmental exposure data are processed separately to obtain genetic feature vectors and environmental exposure feature vectors; Multimodal feature fusion is performed on the family clustering index vector, genetic feature vector, and environmental exposure feature vector to obtain a fused feature vector; The TCN-SENet-Transformer model is constructed and trained by using the fused feature vector as input and the disease label as the supervision signal. During the training process, Bayesian optimization combined with the Memetic algorithm is used to adaptively adjust the model hyperparameters, with the goal of maximizing the validation set AUC value.

2. The data processing model training method for lung cancer risk prediction as described in claim 1, characterized in that, The process involves constructing a pedigree graph based on the pedigree data and using a graph attention network to calculate the weighted clustering coefficient and the strength of kinship disease association for each node, resulting in a pedigree clustering index vector, including: Each family member is treated as a node in the graph. An undirected pedigree graph is constructed with families as the unit. Initial weights are assigned to the edges between nodes based on the degree of kinship. Disease status information corresponding to each node is collected to obtain basic features. Initialize the parameters of the graph attention network, construct the loss function, and update the parameters of the graph attention network using the backpropagation algorithm until the graph attention network converges. The graph attention network includes a feature mapping layer, an attention calculation module, and a dual-task output layer. The dual-task output layer is used to simultaneously output the individual disease probability of each node and the family clustering index vector of each node. The feature mapping layer of a graph attention network maps the basic features of each node into a low-dimensional feature vector. Obtain the neighboring nodes of each target node in the undirected kinship graph. Input the low-dimensional feature vector of the target node and the low-dimensional feature vector of each neighboring node into the attention calculation module. First, obtain the concatenated feature vector through linear transformation. Then, calculate the original attention coefficient between the target node and each neighboring node based on the concatenated feature vector. The original attention coefficients are normalized using the Softmax function in the attention calculation module to obtain normalized attention coefficients. The normalized attention coefficients of each neighbor node are multiplied by the probability of illness of the neighbor node, and the sum is divided by the total number of neighbor nodes of the target node to obtain the weighted clustering coefficient of the target node. The normalized attention coefficients of each neighbor node are added to the probability of illness of the neighbor node to obtain the association strength value of the target node affected by the illness status of relatives. By using the dual-task output layer of the graph attention network, the weighted clustering coefficient and association strength of each target node are combined in sequence to form the family clustering index vector corresponding to the target node. At the same time, the individual disease probability of each node is output for training supervision and prediction.

3. The data processing model training method for lung cancer risk prediction as described in claim 1, characterized in that, The process of processing the genetic marker data and the environmental exposure data to obtain genetic feature vectors and environmental exposure feature vectors includes: For genetic marker data, the PRSice-2 algorithm was used to calculate polygenic risk scores, and the polygenic risk scores were used as genetic feature vectors. Obtain environmental exposure sample data containing time series attributes, and obtain the training dataset after preprocessing. The preprocessing process includes at least outlier cleaning and timestamp alignment and normalization. Construct a BiLSTM network and configure its hyperparameters, which include at least the number of hidden layer units, time step, batch size, and initial learning rate. The training dataset is input into the BiLSTM network step by step. The error between the actual observations and the predicted values ​​output by the BiLSTM network is used as the objective. The Adam optimizer is used for backpropagation training. The loss change is monitored after each training round. The BiLSTM network is considered to have converged when the loss change no longer decreases for 10 consecutive rounds. The actual observations are the actual measurements of environmental exposure variables in the training dataset at each time step. The environmental exposure variables include at least the cumulative smoking index, alcohol intake, and residential industrial pollution index. After normalizing the environmental exposure data, it is input into the trained BiLSTM network to obtain the environmental exposure feature vector.

4. The data processing model training method for lung cancer risk prediction as described in claim 3, characterized in that, The genetic marker data were used to calculate a polygenic risk score using the PRSice-2 algorithm, and the polygenic risk score was used as a genetic feature vector, including: Obtain summary statistics on lung cancer-related SNPs, including at least the SNP name, chromosomal location, allele, effect size, and p-value. Set a threshold range for the p-value of SNP screening, divide it into multiple consecutive p-value intervals, and for each p-value threshold interval, screen out SNPs with p-values ​​less than the current threshold from the statistical data and match them with SNPs in the genetic marker data to obtain the intersection SNPs. Extract the effect values ​​of the intersection SNPs in the statistical data, and use the corresponding effect values ​​as the weights of the SNPs; For each individual in the genetic marker data, calculate the polygenic risk score at the current p-value threshold based on the genotype and weight of their intersection SNPs; The polygenic risk scores calculated for each individual under all p-value thresholds were standardized, and cross-validation was used to evaluate the association strength between the polygenic risk scores and lung cancer phenotypes under different p-value thresholds. The polygenic risk score corresponding to the p-value threshold with the strongest association was selected as the individual's final polygenic risk score. The final polygenic risk scores of all individuals are arranged in order to form a genetic feature vector.

5. The data processing model training method for lung cancer risk prediction as described in claim 3, characterized in that, The environmental exposure data, after being normalized, is input into the trained BiLSTM network to obtain an environmental exposure feature vector, including: Z-score normalization is performed on each feature of the environmental exposure data. A one-dimensional convolutional layer is used to extract local features from the normalized data. Feature maps are generated by sliding the convolutional kernel along the time dimension. The convolutional output is input into a bidirectional LSTM layer to process the forward and reverse directions of the sequence, and the hidden states of the two directions at each time step are connected. After normalizing the output of each time step of the bidirectional LSTM layer, the Dropout layer is used to regularize the normalized output. The attention mechanism is applied to the output of the bidirectional LSTM layer at all time steps, weighted and averaged to calculate the attention weights, and then weighted to obtain the attention feature vector: in, Indicates attention weights, Indicates attention score, Represents an exponential function. Indicates the total number of time steps. Represents the normalization factor. Represents the attention feature vector. Indicates the first Output at each time step; The attention feature vector is input into the fully connected layer for nonlinear transformation to generate the final environmental exposure feature vector.

6. The data processing model training method for lung cancer risk prediction as described in claim 1, characterized in that, The multimodal feature fusion of the family clustering index vector, genetic feature vector, and environmental exposure feature vector yields a fused feature vector, including: By mapping the pedigree clustering index vector, genetic trait vector, and environmental exposure trait vector to the same dimension, the transformed pedigree clustering index vector is obtained. Transformed genetic feature vector Transformed environmental exposure feature vector ; Transformed mode vectors , , The concatenation is performed to obtain the concatenated feature vector. and splicing feature vectors After performing a linear transformation, the intermediate vector is obtained. The attention weights for each modality are calculated by applying the Softmax function to the intermediate vector. , , intermediate vector The formula is: in, Represents the weight matrix. Indicates bias. Attention weights representing familial clustering characteristics Attention weights representing genetic traits Attention weights representing environmental exposure characteristics; The transformed modal vectors are weighted and summed according to the attention weights to obtain the fused feature vector.

7. The data processing model training method for lung cancer risk prediction as described in claim 1, characterized in that, The step of constructing and training a TCN-SENet-Transformer model by using the fused feature vector as input and the disease label as a supervision signal includes: A TCN module is constructed, which uses dilated causal convolutional layers to extract temporal features from the fused feature vector. After each convolution, the features are normalized and activated to output TCN features. The SENet module is constructed, and the channel descriptors are obtained by global average pooling of the TCN features output by the TCN module. The channel attention weights are calculated through a fully connected layer containing compression ratio parameters. The channel attention weights are multiplied by the TCN features by channel to output the enhanced features. A Transformer module is constructed, which takes the enhanced features output by the SENet module as the sequence input, calculates the feature association within the sequence through a multi-head self-attention mechanism, and outputs Transformer features after processing by a feedforward neural network. The feedforward neural network is a two-layer fully connected layer structure, with the first layer using the ReLU activation function and the second layer not using an activation function. The Transformer features are input into a fully connected layer, and the predicted incidence rate of lung cancer is output through a Sigmoid activation function to construct the TCN-SENet-Transformer model.

8. The data processing model training method for lung cancer risk prediction as described in claim 7, characterized in that, The process of adaptively adjusting the model's hyperparameters during training using Bayesian optimization combined with the Memetic algorithm, with the goal of maximizing the validation set AUC value, includes: Define the hyperparameter search space of the model to be optimized, which includes the number of convolutional kernels of the TCN module, the compression ratio of the SENet module, and the number of multi-head attention heads and the learning rate of the Transformer module; The mapping relationship between the model hyperparameter search space and the validation set AUC is modeled based on the Gaussian process model. The hyperparameter combination with the largest expected improvement is selected as the initial solution to obtain the initial hyperparameter combination. The Memetic algorithm is applied to the initial hyperparameter combination for local optimization. Neighborhood solutions are generated through mutation operators, and the AUC value of each neighborhood solution on the validation set is calculated to obtain the optimal neighborhood solution. The optimal neighborhood solution is substituted into the TCN-SENet-Transformer model, with the training set data as input and the disease label as the supervision signal. The loss is calculated using the cross-entropy loss function, and the model parameters are updated through backpropagation. After each training round, the AUC value is calculated using the validation set. If the current AUC value does not exceed the historical maximum value, the hyperparameters are repeatedly adjusted. If the AUC value does not improve after a preset number of rounds, training is stopped.

9. The data processing model training method for lung cancer risk prediction as described in claim 8, characterized in that, The process of applying the Memetic algorithm to locally optimize the initial hyperparameter combination, generating neighborhood solutions through mutation operators, and calculating the AUC value of each neighborhood solution on the validation set to obtain the optimal neighborhood solution includes: Determine the current combination of initialization hyperparameters to be optimized. , This indicates the number of hyperparameters and sets local optimization parameters, including the number of neighborhood solution generation. Variable Asynchronous Long Range Number of local search iterations , This represents the minimum length of the variable asynchronous operation. Indicates the maximum value of variable asynchronous length; For initialization hyperparameter combination Perform the mutation operation to generate an initial neighborhood solution, at the... In the next iteration, the initial neighborhood solution Each hyperparameter is subjected to a preset mutation amount within a variable asynchronous length range to generate sub-neighborhood solutions. ,in ; Calculate the sub-neighborhood solution If the validation set AUC value is better than the initial neighborhood solution... Then update for After reaching the required number of local search iterations, a neighborhood solution optimized by the local search is obtained. Substitute the initial hyperparameter combination and the neighborhood solution optimized by local search into the TCN-SENet-Transformer model respectively, train the model using the same training set, and then calculate the corresponding AUC value using the validation set. Compare the AUC values ​​of all solutions and select the solution with the largest AUC value as the optimal neighborhood solution for local optimization by the Memetic algorithm.

10. The data processing model training method for lung cancer risk prediction as described in claim 9, characterized in that, The initialization hyperparameter combination Performing mutation operations to generate initial neighborhood solutions includes: For continuous hyperparameters in the initialization hyperparameter combination Calculate the mutated value using the following formula: in, Indicates the mutated th New values ​​for each hyperparameter Indicating the first hyperparameter in the initialization hyperparameter combination Initial values ​​for each hyperparameter. Indicates variable asynchronous length, Indicates the direction factor.

Images