Lung cancer lifetime prediction system based on prognosis factor multi-data fusion

Abstract

The invention discloses a lung cancer lifetime prediction system based on prognosis factor multi-data fusion, and belongs to the technical field of lung cancer prognosis prediction, and the system comprises a multi-source data collection module which is used for collecting prognosis multi-source data of a patient; the multi-source data processing module is used for carrying out preprocessing and feature extraction on the prognosis multi-source data of the patient; the multi-source data fusion module is used for carrying out cross-modal alignment and fine-grained fusion on the extracted multi-modal feature vectors, and capturing a dependency relationship between modals based on a hierarchical attention mechanism to form patient prognosis fusion data; and the survival analysis and prediction module is used for analyzing the prognosis fusion data of the patient according to the lung cancer lifetime prediction model, automatically predicting the lifetime of the patient and displaying the lifetime in a visual form. The problems that existing lung cancer lifetime prediction is low in accuracy and cannot provide support for personalized treatment are solved. The lung cancer lifetime prediction accuracy can be improved, and support can be provided for personalized treatment.

Description

Technical Field

[0001] This invention relates to the field of lung cancer prognosis prediction technology, specifically to a lung cancer survival prediction system based on the fusion of multiple prognostic factors. Background Technology

[0002] Lung cancer is one of the malignant tumors with the highest incidence and mortality rates worldwide. Its prognosis is affected by a variety of factors, including pathological type, stage, patient age, gender, genetic characteristics, treatment methods, and psychological and economic status. Traditional prognostic prediction methods mainly rely on single-modality data, such as clinicopathological features or gene expression data. However, the predictive power of a single modality is limited and it is difficult to fully reflect the complex biological characteristics of lung cancer, resulting in low accuracy in predicting lung cancer survival and failing to support personalized treatment.

[0003] In recent years, with the development of multi-omics technology and artificial intelligence algorithms, prognostic prediction methods based on multi-data fusion have gradually become a research hotspot. Multi-omics data can reveal the heterogeneity of cancer at the molecular level, while multimodal data can supplement the correlation between biological characteristics and clinical information. Therefore, lung cancer survival prediction systems based on multi-data fusion of prognostic factors are particularly important. Summary of the Invention

[0004] The purpose of this invention is to provide a lung cancer survival prediction system based on the fusion of multiple prognostic factors, which can improve the accuracy of lung cancer survival prediction, provide support for personalized treatment, and solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A lung cancer survival prediction system based on multi-data fusion of prognostic factors includes: The multi-source data acquisition module is used to collect imaging data, clinical data, pathological data and genomic data of patients' prognosis to form multi-source data of patients' prognosis; The multi-source data processing module is used to preprocess and extract features from multi-source patient prognosis data based on natural language processing, and to determine 3D depth features, clinical embedding features, pathological visual features and pathway perception features. The multi-source data fusion module is used to perform cross-modal alignment and fine-grained fusion of the extracted multimodal feature vectors, and capture intermodal dependencies based on a hierarchical attention mechanism to form patient prognosis fusion data; The survival analysis and prediction module is used to build a lung cancer survival prediction model based on artificial intelligence. It analyzes the patient prognosis fusion data based on the lung cancer survival prediction model and automatically predicts the patient's survival time, and displays it in a visual form.

[0006] Preferably, the multi-source data acquisition module includes: The image acquisition unit is used to acquire imaging data for patient prognosis, including CT and PET-CT images; The clinical data collection unit is used to collect clinical data on patient prognosis, including age, gender, TNM stage, ECOG score, smoking history, treatment history, and laboratory test results. The pathology acquisition unit is used to collect pathological data for patient prognosis, including tissue type, degree of differentiation, PD-L1 expression, and Ki-67 index. The genome acquisition unit is used to collect genomic data for patient prognosis, including driver gene mutations and tumor mutational burden.

[0007] Preferably, the multi-source data processing module includes: The data cleaning unit is used to clean the multi-source patient prognosis data, remove noise from the multi-source patient prognosis data, identify outliers in the multi-source patient prognosis data based on bias test, query missing values ​​in the multi-source patient prognosis data based on Python, and evaluate the outliers and missing values ​​in the multi-source patient prognosis data. If the outliers and missing values ​​are useful for predicting lung cancer survival, the outliers are corrected and the missing values ​​are filled; otherwise, the outliers and missing values ​​are deleted. The data transformation unit is used to transform multi-source patient prognostic data based on Z-score standardization, converting multi-source patient prognostic data with different dimensions and distribution ranges into a unified scale, eliminating the dimensional differences between multi-source patient prognostic data, and forming standardized multi-source patient prognostic data.

[0008] Preferably, the multi-source data processing module further includes: The feature extraction unit is used to extract features from multi-source patient prognostic data; Feature extraction from multi-source patient prognostic data includes: Feature extraction is performed on the image data, the tumor area is automatically segmented and extended to the peritumoral area, and depth features are extracted based on the 3D-SEResNet network to comprehensively capture the three-dimensional spatial heterogeneity of the tumor and extract radiomics features to form 3D depth features. Feature extraction is performed on clinical data, clinical text is processed based on a natural language processing model, and structured data is converted into high-dimensional feature vectors to form clinical embedding features; Feature extraction is performed on pathological data. High-resolution whole-slice images are cut into multiple small image patches. Each image patch is encoded using a pre-trained convolutional neural network. Through an attention-based multi-instance learning mechanism, the most important image regions for prognosis are automatically identified and weighted, and aggregated into pathological visual features. Feature extraction is performed on genomic data, and a pathway-aware multimodal framework is used for processing. Based on known biological knowledge, gene expression data is mapped to 186 specific biological pathways, narrowing the semantic gap between genomic data and imaging or pathological data, enabling information from different modalities to interact and forming pathway-aware features.

[0009] Preferably, the multi-source data fusion module includes: The feature fusion unit is used to deeply fuse the extracted multimodal feature vectors; Information interaction is carried out within the modality based on the self-attention mechanism. The unpaired label contrastive learning method is used to align the visual semantics of image patches with the functional semantics of biological pathways in the feature space. Following the biological prior that genotype determines phenotype, the model is guided by biological pathway features to find key pathological image regions related to them, thereby achieving fine-grained fusion of gene and pathological data. Based on the hierarchical attention mechanism, CT image features and clinical embedded features are used as queries, keys, and values. Through attention weight calculation, the dependency relationship between image manifestations and clinical symptoms and test indicators is deeply explored. The features after cross-attention fusion are spliced ​​with the processed pathological features. Through the multi-head attention mechanism, global and cross-modal related features are extracted to generate patient prognostic fusion data.

[0010] Preferably, the survival analysis and prediction module includes: Model building unit, used to build lung cancer survival prediction models; Based on the needs of lung cancer survival prediction, historical data on lung cancer survival prediction were collected and divided into training set and test set. The deep learning model is trained using a training set, enabling it to learn lung cancer survival prediction behavior autonomously from the training set and automatically predict the patient's survival probability in the next 1, 3, and 5 years, thus determining the lung cancer survival prediction model. The lung cancer survival prediction model was tested using a test set. The performance of the lung cancer survival prediction model was evaluated based on the consistency index, calibration curve and ROC curve. The model was used to determine whether it could automatically predict the survival probability of patients in the next 1 year, 3 years and 5 years and to determine the model test evaluation results. When the lung cancer survival prediction model cannot automatically predict the patient's survival probability in the next 1 year, 3 years and 5 years, the parameters of the lung cancer survival prediction model are adjusted and iteratively optimized until the lung cancer survival prediction model can automatically predict the patient's survival probability in the next 1 year, 3 years and 5 years, and the optimal lung cancer survival prediction model is determined.

[0011] Preferably, the survival analysis and prediction module includes: The analysis and prediction unit is used to analyze patient prognostic fusion data and automatically predict patient survival. Deploy the lung cancer survival prediction model and place it in the actual patient lung cancer survival prediction environment; Patient prognostic fusion data is input into the lung cancer survival prediction model. The model analyzes the patient prognostic fusion data and automatically predicts the patient's survival probability in the next 1 year, 3 years, and 5 years, thus determining the patient's lung cancer survival prediction result.

[0012] Preferably, it also includes a visualization module for displaying the patient's lung cancer survival prediction report in a visual form, including the patient's lung cancer survival prediction results and follow-up recommendations.

[0013] Preferably, a training set is used to train the deep learning model, enabling the deep learning model to autonomously learn lung cancer survival prediction behavior from the training set, automatically predicting the patient's survival probability in the next 1 year, 3 years, and 5 years, and determining the lung cancer survival prediction model, including: Obtain a training dataset, which includes core features and follow-up time series data; A dynamic coefficient Cox regression module is constructed based on a time-dependent regression coefficient function. The regression coefficients are fitted using a piecewise cubic spline function, and the parameters are solved using maximum likelihood estimation. The core features are input, and the survival risk function is output. A temporal Bayesian network module is constructed, dividing time into multiple discrete intervals and building sub-networks. The EM algorithm is used in conjunction with follow-up data to learn time-dependent parameters. A discretized state transition probability matrix is ​​introduced to characterize the probability of variable changes over time, and the survival risk function is output. An adaptive weighted fusion strategy is adopted, dynamically adjusting the weight coefficients based on the prediction errors of the two modules on the same time grid in the training data. The fusion yields the basic survival risk function and the basic survival probabilities for 1 year, 3 years, and 5 years, completing the construction of the basic survival prediction unit. Multiple nodes are defined; initial edge weights between nodes are calculated based on Pearson correlation coefficient, clinical guideline logical correlation strength, or mutual information value; a nonlinear time decay function is introduced to construct a mechanism interaction graph with dynamic edge weights; an improved graph attention network is adopted, which aggregates neighborhood node information through a multi-head attention mechanism, introduces a mechanism contribution attention layer to learn time-dependent mechanism contribution weights, and outputs mechanism interaction embedding features; the mechanism interaction embedding features are used as additional input to supplement the basic survival prediction unit, and the basic survival risk function is corrected to obtain the updated survival risk function, thus completing the construction of the multi-mechanism collaborative inference layer; Using the spline coefficients of the basic survival prediction unit, the transition probability of the temporal Bayesian network, and the attention weights of the graph attention network as optimization objects, we define the patient individual feature vector, follow-up time, C-index prediction performance index, and feature distribution offset as the state space, construct a multi-objective reward function, and complete the construction of the dynamic adaptive optimization layer. Using patient follow-up time-series data as input, a time-series feature sequence is constructed, including baseline survival risk values ​​at each time point, mechanism interaction embedding features, and clinical follow-up indicators. An LSTM and attention pooling architecture is employed, with the LSTM layer capturing long-term dependencies of time-series features and the attention pooling layer enhancing the weights of key time-point features. This outputs continuous survival predictions, 1-year / 3-year / 5-year survival probabilities and confidence intervals, and risk inflection points. Based on the slope features of the risk evolution curve and the baseline risk score, a dynamic risk stratification standard is constructed, completing the construction of the risk evolution prediction layer. The training dataset was grouped by patient, and an end-to-end joint training strategy was adopted. The basic survival prediction unit, multi-mechanism collaborative inference layer, dynamic adaptive optimization layer and risk evolution prediction layer were synchronously optimized by sharing a loss function. When the C-index of the validation set exceeded 0.75 and the p-value of the risk stratification Log-rank test was <0.01, the lung cancer survival prediction model was obtained.

[0014] Preferably, based on a hierarchical attention mechanism, CT image features and clinical embedded features are used as queries, keys, and values. Through attention weight calculation, the dependencies between imaging manifestations and clinical symptoms and laboratory indicators are deeply mined. The features fused through cross-attention are then combined with processed pathological features. A multi-head attention mechanism is used to extract global, cross-modal related features to generate patient prognostic fusion data, including: Using tumor region features from enhanced CT images as query vectors and clinical embedding adaptation features as key vectors, cross-attention weights are calculated, and CT-guided clinical enhancement features are obtained through weighted summation. Using CT-guided clinical enhancement features as query vectors and CT image adaptation features as key vectors, the reverse cross-attention weights are calculated and combined with the lesion region segmentation mask of CT images to obtain clinically guided CT enhancement features. CT-guided clinical enhancement features are element-wise added to clinical-guided CT enhancement features, and then dimensionally fused using convolutional kernels to obtain CT-clinical cross-fusion features. Key region features of pathological images, weighted biological pathway features, and CT-clinical cross-fusion features are concatenated along the channel dimension to obtain initial multimodal features. An adaptive gating unit adjusts the channel proportions of each modality feature. The gated multimodal features are input into a multi-head attention module. This module contains k parallel self-attention heads, each learning cross-modal associations in different dimensions. The outputs of the k attention heads are concatenated and linearly transformed to obtain global association features. An autoencoder is used to denoise the global association features to generate patient prognostic fusion data.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention collects imaging data, clinical data, pathological data, and genomic data related to patient prognosis to form multi-source prognostic data. Based on natural language processing, it preprocesses and extracts features from the multi-source prognostic data, automatically segments the tumor region and extends it to the peritumoral region, captures the three-dimensional spatial heterogeneity of the tumor, and forms 3D depth features. Based on a natural language processing model, it processes clinical text and converts structured data into high-dimensional feature vectors to form clinical embedding features. A pre-trained convolutional neural network is used to encode each image patch, automatically identifies and weights the most important image regions for prognosis, and aggregates them into pathological visual features. A pathway-aware multimodal framework is used to narrow the semantic gap between genomic data and imaging or pathological data to form pathway-aware features that can comprehensively reflect the complex biological characteristics of lung cancer.

[0016] 2. This invention performs cross-modal alignment and fine-grained fusion on the extracted multimodal feature vectors, captures intermodal dependencies based on a hierarchical attention mechanism, forms patient prognostic fusion data, constructs a lung cancer survival prediction model based on artificial intelligence, analyzes the patient prognostic fusion data according to the lung cancer survival prediction model, and automatically predicts the patient's survival probability in the next 1 year, 3 years, and 5 years, determines the patient's lung cancer survival prediction results, and displays them in a visual form, which can improve the accuracy of lung cancer survival prediction and provide support for personalized treatment. Attached Figure Description

[0017] Figure 1 This is a block diagram of the lung cancer survival prediction system based on multi-data fusion of prognostic factors according to the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] To address the limitations of existing single-modal predictions, which fail to fully reflect the complex biological characteristics of lung cancer and thus result in low accuracy in predicting lung cancer survival, thus hindering support for personalized treatment, please refer to [link to relevant documentation]. Figure 1 This embodiment provides the following technical solution: The lung cancer survival prediction system based on multi-source data fusion of prognostic factors includes: a multi-source data acquisition module, a multi-source data processing module, a multi-source data fusion module, and a survival analysis and prediction module.

[0020] The multi-source data acquisition module is used to collect imaging data, clinical data, pathological data and genomic data of patients' prognosis to form multi-source data of patient prognosis.

[0021] In this embodiment, the multi-source data acquisition module includes: The image acquisition unit is used to acquire imaging data for patient prognosis, including CT and PET-CT images; The clinical data collection unit is used to collect clinical data on patient prognosis, including age, gender, TNM stage, ECOG score, smoking history, treatment history, and laboratory test results. The pathology acquisition unit is used to collect pathological data for patient prognosis, including tissue type, degree of differentiation, PD-L1 expression, and Ki-67 index. The genome acquisition unit is used to collect genomic data for patient prognosis, including driver gene mutations and tumor mutational burden.

[0022] Among them, the multi-source data processing module is used to preprocess and extract features from multi-source patient prognosis data based on natural language processing, and to determine 3D depth features, clinical embedding features, pathological visual features and pathway perception features.

[0023] In this embodiment, the multi-source data processing module includes: The data cleaning unit is used to clean the multi-source patient prognosis data, remove noise from the multi-source patient prognosis data, identify outliers in the multi-source patient prognosis data based on bias test, query missing values ​​in the multi-source patient prognosis data based on Python, and evaluate the outliers and missing values ​​in the multi-source patient prognosis data. If the outliers and missing values ​​are useful for predicting lung cancer survival, the outliers are corrected and the missing values ​​are filled; otherwise, the outliers and missing values ​​are deleted. The data transformation unit is used to transform multi-source patient prognostic data based on Z-score standardization, converting multi-source patient prognostic data with different dimensions and distribution ranges into a unified scale, eliminating the dimensional differences between multi-source patient prognostic data, and forming standardized multi-source patient prognostic data.

[0024] In this embodiment, the multi-source data processing module further includes: The feature extraction unit is used to extract features from multi-source patient prognosis data and transform multi-source patient prognosis data in different formats into high-quality deep feature vectors. Feature extraction from multi-source patient prognostic data includes: Feature extraction is performed on image data to automatically segment the tumor region and extend it to the peritumoral region, such as 3mm, 6mm or 9mm. Studies have shown that the heterogeneity of peritumoral tissue has an important impact on prognosis. Based on the 3D-SEResNet network, depth features are extracted. This network embeds the Squeeze-and-Excitation (SE) module into the classic 3D ResNet, which can adaptively calibrate the importance of feature channels, thereby more comprehensively capturing the three-dimensional spatial heterogeneity of the tumor and extracting radiomics features, such as texture and wavelet features, and thus forming 3D depth features. Feature extraction is performed on clinical data, clinical text is processed based on a natural language processing model, and structured data is converted into high-dimensional feature vectors to form clinical embedding features; Feature extraction is performed on pathological data. High-resolution whole-slice images are cut into multiple small image patches. Each image patch is encoded using a pre-trained convolutional neural network. Through an attention-based multi-instance learning mechanism, the most important image regions for prognosis are automatically identified and weighted, and aggregated into pathological visual features. Feature extraction is performed on genomic data, and a pathway-aware multimodal framework is used for processing. Based on known biological knowledge, gene expression data is mapped to 186 specific biological pathways, narrowing the semantic gap between genomic data and imaging or pathological data, enabling information from different modalities to interact and forming pathway-aware features.

[0025] The multi-source data fusion module is used to perform cross-modal alignment and fine-grained fusion of the extracted multimodal feature vectors, and capture intermodal dependencies based on a hierarchical attention mechanism to form patient prognosis fusion data.

[0026] In this embodiment, the multi-source data fusion module includes: The feature fusion unit is used to deeply fuse the extracted multimodal feature vectors; Based on the self-attention mechanism, information interaction is carried out within the modality (such as between different pathological image patches or between different biological pathways). For pathological and genomic data, a label-free contrastive learning method is used to align the visual semantics of image patches with the functional semantics of biological pathways in the feature space. Following the biological prior that genotype determines phenotype, the model is guided by biological pathway features to find key pathological image regions related to them, thereby achieving fine-grained fusion of gene and pathological data. Based on the hierarchical attention mechanism, CT image features and clinical embedded features are used as queries, keys, and values. Through attention weight calculation, the dependency relationship between image manifestations and clinical symptoms and test indicators is deeply explored. The features after cross-attention fusion are spliced ​​with the processed pathological features. Through the multi-head attention mechanism, global and cross-modal related features are extracted to generate patient prognostic fusion data.

[0027] The survival analysis and prediction module is used to build a lung cancer survival prediction model based on artificial intelligence. It analyzes the patient prognosis fusion data based on the lung cancer survival prediction model and automatically predicts the patient's survival time, which is then displayed in a visual form.

[0028] In this embodiment, the survival analysis and prediction module includes: Model building unit, used to build lung cancer survival prediction models; Based on the needs of lung cancer survival prediction, historical data on lung cancer survival prediction were collected and divided into training set and test set. The deep learning model is trained using a training set, enabling it to learn lung cancer survival prediction behavior autonomously from the training set and automatically predict the patient's survival probability in the next 1, 3, and 5 years, thus determining the lung cancer survival prediction model. The lung cancer survival prediction model was tested using a test set. The performance of the lung cancer survival prediction model was evaluated based on the consistency index, calibration curve and ROC curve. The model was used to determine whether it could automatically predict the survival probability of patients in the next 1 year, 3 years and 5 years and to determine the model test evaluation results. When the lung cancer survival prediction model cannot automatically predict the patient's survival probability in the next 1 year, 3 years and 5 years, the parameters of the lung cancer survival prediction model are adjusted and iteratively optimized until the lung cancer survival prediction model can automatically predict the patient's survival probability in the next 1 year, 3 years and 5 years, and the optimal lung cancer survival prediction model is determined.

[0029] In this embodiment, the survival analysis and prediction module includes: The analysis and prediction unit is used to analyze patient prognostic fusion data and automatically predict patient survival. Deploy the lung cancer survival prediction model and place it in the actual patient lung cancer survival prediction environment; Patient prognostic fusion data is input into the lung cancer survival prediction model. The model analyzes the patient prognostic fusion data and automatically predicts the patient's survival probability in the next 1 year, 3 years, and 5 years, thus determining the patient's lung cancer survival prediction result.

[0030] Specifically, by using deep learning methods, we integrate and analyze various data types that are routinely available in clinical practice, and explore the deep and fine-grained relationships between them, thereby constructing a more comprehensive, accurate, and clinically interpretable individualized survival prediction model. It has high-precision prediction and its predictive performance is significantly better than traditional models based on a single data modality or simple splicing.

[0031] In this embodiment, a visualization module is also included to display the patient's lung cancer survival prediction report in a visual form, including the patient's lung cancer survival prediction results and follow-up recommendations.

[0032] In this embodiment, a training set is used to train the deep learning model, enabling the model to autonomously learn lung cancer survival prediction behavior from the training set and automatically predict the patient's survival probability in the next 1 year, 3 years, and 5 years, thus determining the lung cancer survival prediction model, including: Obtain a training dataset, which includes core features and follow-up time series data; A dynamic coefficient Cox regression module is constructed based on a time-dependent regression coefficient function. The regression coefficients are fitted using a piecewise cubic spline function, and the parameters are solved using maximum likelihood estimation. The core features are input, and the survival risk function is output. A temporal Bayesian network module is constructed, dividing time into multiple discrete intervals and building sub-networks. The EM algorithm is used in conjunction with follow-up data to learn time-dependent parameters. A discretized state transition probability matrix is ​​introduced to characterize the probability of variable changes over time, and the survival risk function is output. An adaptive weighted fusion strategy is adopted, dynamically adjusting the weight coefficients based on the prediction errors of the two modules on the same time grid in the training data. The fusion yields the basic survival risk function and the basic survival probabilities for 1 year, 3 years, and 5 years, completing the construction of the basic survival prediction unit. Multiple nodes are defined; initial edge weights between nodes are calculated based on Pearson correlation coefficient, clinical guideline logical correlation strength, or mutual information value; a nonlinear time decay function is introduced to construct a mechanism interaction graph with dynamic edge weights; an improved graph attention network is adopted, which aggregates neighborhood node information through a multi-head attention mechanism, introduces a mechanism contribution attention layer to learn time-dependent mechanism contribution weights, and outputs mechanism interaction embedding features; the mechanism interaction embedding features are used as additional input to supplement the basic survival prediction unit, and the basic survival risk function is corrected to obtain the updated survival risk function, thus completing the construction of the multi-mechanism collaborative inference layer; Using the spline coefficients of the basic survival prediction unit, the transition probability of the temporal Bayesian network, and the attention weights of the graph attention network as optimization objects, we define the patient individual feature vector, follow-up time, C-index prediction performance index, and feature distribution offset as the state space, construct a multi-objective reward function, and complete the construction of the dynamic adaptive optimization layer. Using patient follow-up time-series data as input, a time-series feature sequence is constructed, including baseline survival risk values ​​at each time point, mechanism interaction embedding features, and clinical follow-up indicators. An LSTM and attention pooling architecture is employed, with the LSTM layer capturing long-term dependencies of time-series features and the attention pooling layer enhancing the weights of key time-point features. This outputs continuous survival predictions, 1-year / 3-year / 5-year survival probabilities and confidence intervals, and risk inflection points. Based on the slope features of the risk evolution curve and the baseline risk score, a dynamic risk stratification standard is constructed, completing the construction of the risk evolution prediction layer. The training dataset was grouped by patient, and an end-to-end joint training strategy was adopted. The basic survival prediction unit, multi-mechanism collaborative inference layer, dynamic adaptive optimization layer and risk evolution prediction layer were synchronously optimized by sharing a loss function. When the C-index of the validation set exceeded 0.75 and the p-value of the risk stratification Log-rank test was <0.01, the lung cancer survival prediction model was obtained.

[0033] In this embodiment, the core features include: (1) the relative ratio of core immune cells, specifically the ratio of CD8+ T cells to regulatory T cells (Treg) and the macrophage M1 / M2 polarization ratio, which are calculated quantitatively by flow cytometry or immunohistochemistry; (2) the feature values ​​of prognosis-related gene modules, which are obtained by clustering RNA-seq data using the WGCNA algorithm to obtain 10 functional modules, and three key modules with a Cox regression p value <0.01 with overall survival are selected, and the first principal component of the feature gene of the module is taken as the feature value; (3) clinical indicators of tumor stage / age / gender, wherein the tumor stage adopts the 8th edition of the TNM staging standard.

[0034] In this embodiment, the nodes include immune mechanism nodes (CD8+ T cells, Treg cells, macrophage M1 / M2), gene mechanism nodes (PI3K-AKT pathway activity, TP53 mutation status, immune checkpoint expression module), and clinical mechanism nodes (TNM stage, age, serum LDH level, treatment response).

[0035] In this embodiment, a dynamic coefficient Cox regression module is constructed based on a time-dependent regression coefficient function. The regression coefficients are fitted using a cubic spline function, and the parameters are solved using maximum likelihood estimation. The core features are input, and the survival risk function is output. A temporal Bayesian network module is constructed, dividing time into multiple discrete intervals and constructing sub-networks. The EM algorithm is used in conjunction with follow-up data to learn time-dependent parameters. A state transition probability matrix is ​​introduced to characterize the probability of variable changes over time, and the survival risk function is output. An adaptive weighted fusion strategy is adopted, dynamically adjusting the weight coefficients based on the time-dependent prediction errors of the two modules in the training data. The fusion yields the basic survival risk function and the basic survival probabilities for 1 year, 3 years, and 5 years, completing the construction of the basic survival prediction unit. This includes: adopting a dual-model collaborative architecture of "dynamic coefficient Cox regression + temporal Bayesian network" as the core of the model's basic survival prediction; the dynamic coefficient Cox regression module design: introducing a time-dependent regression coefficient function... The coefficient changes dynamically with the follow-up time t, which is adapted to the dynamic characteristics of lung cancer prognosis: ;in, Piecewise cubic spline function fitting is used. , The basis functions are B-spline functions, with M=3 internal nodes (the node positions are optimized through cross-validation to the 25%, 50%, and 75% quantiles of the total follow-up time). Solve using maximum likelihood estimation; These are the coefficient parameters of the B-spline basis functions; Let be the conditional risk function, representing the instantaneous mortality risk rate for a given feature x at time t; The baseline risk function represents the basic risk when all covariates are zero; Let be the time-dependent regression coefficient of the i-th covariate; Let i be the value of the i-th covariate (feature); For B-spline basis functions; This represents the number of nodes inside the B-spline. The total number of covariates (features) is used; input variables are selected from core prognostic factors screened by LASSO-Cox, including the relative ratio of core immune cells, prognostic-related gene module feature values, and clinical indicators of tumor stage / age / gender; the temporal Bayesian network module design: the follow-up time is discretized into N equally probable intervals, and a sub-network is constructed for each interval; network nodes contain core prognostic factors and survival status (survival / death), and the edge weights are the conditional probabilities of different time intervals. The temporal dependency parameters are learned by combining the EM algorithm with follow-up data; a discrete state transition probability matrix is ​​introduced to characterize the probability of discrete state changes of variables in different time intervals; to align with the COX module, the risk of discrete intervals is extended to continuous time through linear interpolation. ; The risk function output by the temporal Bayesian network; The floor function of time t; The time interval is rounded up; Dual-model collaborative fusion: An adaptive weighted fusion strategy is adopted, with the weight coefficient α(t) dynamically adjusted with time t, and optimization is performed based on the negative log-likelihood error of the two modules in the same time grid in the training data: ; This is the basic survival risk function after fusion; The risk function output by the Cox regression model; Here, is the risk function output by the temporal Bayesian network; where, The weighting coefficients are adaptive; the basic survival risk function is output after fusion. and the basic survival probability in 1 year / 3 years / 5 years ; This is the survival probability function.

[0036] In this embodiment, multiple nodes are defined; initial edge weights between nodes are calculated based on Pearson correlation coefficient, clinical logical association strength, or mutual information value; a time decay factor is introduced to construct a mechanism interaction graph with dynamic edge weights; an improved graph attention network is used to aggregate neighborhood node information through a multi-head attention mechanism; a mechanism contribution attention layer is introduced to learn time-dependent mechanism contribution weights, and mechanism interaction embedding features are output; the mechanism interaction embedding features are used as additional input to supplement the basic survival prediction unit, and the basic survival risk function is corrected to obtain the updated survival risk function, thus completing the construction of the multi-mechanism collaborative inference layer, including: mechanism interaction graph construction; graph node definition: including three types of nodes—immune mechanism nodes (CD8+ T cells, Treg cells, macrophages). M1 / M2), gene mechanism nodes (PI3K-AKT pathway activity, TP53 mutation status, immune checkpoint expression module), and clinical mechanism nodes (TNM stage, age, serum LDH level, treatment response); edge weight definition: edge weights between nodes of the same class: between immune cells based on Spearman rank correlation coefficient (more suitable for nonlinear relationships), between gene modules based on pathway co-occurrence frequency, and between clinical indicators based on NCCN guideline logical association strength (0-1 normalization); edge weights between nodes of different classes: calculated based on conditional mutual information (MI) values, such as the MI value between immune cells and gene modules, and the MI value between gene modules and clinical indicators, estimated using the k-nearest neighbor method; dynamic edge update: introducing a nonlinear time decay function, the edge weights are dynamically adjusted with the follow-up time t. ;in, The initial correlation strength, =0.05 represents the global decay rate (optimized through grid search), and g(t) = log(1+t) is the time nonlinear transformation function; GNN inference module design: An improved graph attention network (GAT) is adopted, where each node aggregates information from neighboring nodes through a 4-head attention mechanism, and learns node embedding features under the interaction of learning mechanisms: ;in, For node i in the th... Embedded representation of +1 layer GNN Let the weight of the k-th attention head be... Let be the set of neighboring nodes of node i. Let be the attention weight of node j to node i in the k-th attention head. Let be the weight matrix of the k-th head in the l-th layer. The LeakyReLU activation function is used; a mechanism contribution attention layer is introduced to automatically learn the dynamic contribution weights of different mechanisms (immune, genetic, clinical) to survival prediction. ;in, Let m be the contribution weight of mechanism m at time t. The node embedding mean of mechanism m. The transformation matrix is... For attention vector parameters, The current time; For transpose; ; Embed the feature matrix for the nodes of mechanism m; Mechanism interaction embedding features; Collaboration with the base prediction unit: The mechanism interaction embedding features output by the GNN are used as additional input to supplement the base survival prediction unit, modifying the base survival risk function: ;in, A two-layer MLP (128 elements, ReLU activation) is used to ensure that the correction term satisfies the non-negativity constraint of the survival analysis; This is the modified survival risk function.

[0037] In this embodiment, the spline coefficients of the basic survival prediction unit, the transition probabilities of the temporal Bayesian network, and the attention weights of the graph attention network are used as optimization objects. The patient's individual feature vector, follow-up time, prediction error, and data distribution offset are defined as the state space. A multi-objective reward function is constructed, including: introducing a reinforcement learning (RL) framework to construct a dynamically optimized agent for model parameters, adapting in real-time to data distribution offsets and individual patient differences; the agent is defined as follows: based on the model's core parameters (the spline coefficients of the dynamic Cox algorithm)... The optimization objects are the temporal Bayesian transition probability matrix and the GNN attention weights W, and the agent's action space is the parameter fine-tuning vector. The amplitude is limited to ±5% of the parameter value); State space definition: including patient individual feature vector (after standardization), current follow-up time t (normalized to [0,1]), current model prediction performance (measured by C-index), data distribution offset (calculated by Wasserstein distance between feature mean and variance to avoid the infeasibility of KL divergence in high dimensions); Reward function design: adopting a multi-objective reward function: ;in, To reinforce the learning reward function value, it is dimensionless. C-index(t) is the time-series C-index at time t, and ΔC-index(t) is the difference in C-index before and after parameter adjustment. For L1 norm regularization terms, =0.6, =0.3, =0.1 is the weight coefficient (determined through Bayesian optimization); Fine-tuning vector for parameters.

[0038] In this embodiment, patient follow-up time-series data is used as input to construct a time-series feature sequence containing baseline survival risk values, mechanism interaction embedding features, and clinical follow-up indicators at each time point. An LSTM and attention pooling architecture is employed; the LSTM layer captures long-term dependencies of time-series features, and the attention pooling layer strengthens the feature weights at key time points, outputting continuous survival prediction values, 1-year / 3-year / 5-year survival probabilities and confidence intervals, and risk inflection points. Based on the slope features of the risk evolution curve and the baseline risk score, a dynamic risk stratification standard is constructed. The effectiveness of the stratification is verified using Kaplan-Meier survival curves, completing the construction of the risk evolution prediction layer, including: LSTM time-series prediction module design: Input sequence construction: The patient's follow-up time-series data is used as input, including baseline survival risk values ​​at each time point, mechanism interaction embedding features (dimensionality reduced to 64), and clinical follow-up indicators (such as RECIST standard tumor size changes and treatment response grades). Missing values ​​are handled using multiple imputation to construct a time-series feature sequence {X_1, X_2, ..., X_T} (T is the number of follow-up time points); Network structure design: An "LSTM + attention pooling" architecture is adopted. Two layers of LSTM (128 units, Dropout rate 0.3) capture the long-term dependencies of time-series features, and the attention pooling layer strengthens the feature weights of key time points (such as 3 months and 12 months after treatment). ;in, These are the attention vector parameters; This is the attention transformation matrix; The LSTM hidden state at time t; The mean of all hidden states; Attention weights at time point t; Output layer design: Outputs the patient's future survival risk evolution curve, including: continuous survival prediction (mean ± standard deviation calculated by Monte Carlo Dropout sampling 100 times); 1-year / 3-year / 5-year survival probabilities and 95% confidence intervals; risk inflection point prediction (defined as the slope of the risk curve). (Time exceeding 0.05 and lasting for two time points); Dynamic optimization of risk stratification: based on the slope characteristic (k) of the risk evolution curve and the basic risk score at t=0. , construct a dynamic risk stratification standard: low-risk group: basic risk score < Q1(70%) and risk slope k < 0.01; medium-risk group: Q1(70%) ≤ basic risk score ≤ Q2(70%) or 0.01 ≤ k ≤ 0.03; high-risk group: basic risk score > Q2(70%) and k > 0.03; where Q1(70%) = 111.07 and Q2(70%) = 135.40 are the 70% quantiles of the basic risk scores in the training set respectively; the stratification results are verified by the Kaplan-Meier survival curve and Log-rank test (p < 0.05), and if the verification fails, the quantile threshold is automatically adjusted.

[0039] The working principle and beneficial effects of the above technical solution are as follows: multi-module fusion and adaptive weighting strategy, combining the advantages of different methods, accurately capturing various influencing factors, and improving the prediction accuracy of 1-year, 3-year, and 5-year survival probabilities; the multi-mechanism collaborative inference layer mines feature interactions and time-dependent relationships, dynamically adjusts multi-objective parameters, and ensures the stable and effective operation of the model in various scenarios; outputs continuous survival period, survival probability, confidence interval, risk turning point, and stratification standard, assisting clinical precise decision-making and reducing the psychological burden of patients and their families.

[0040] In this embodiment, based on the hierarchical attention mechanism, the CT image features and clinical embedding features are used as query, key, and value respectively. Through the calculation of attention weights, the dependence relationship between the imaging manifestations, clinical symptoms, and test indicators is deeply mined, and the features after cross-attention fusion are spliced with the processed pathological features. The global and cross-modal association features are extracted through the multi-head attention mechanism to generate patient prognosis fusion data, including: Taking the tumor region features of the enhanced CT image as the query vector and the clinically embedded adaptation features as the key vector, calculate the cross-attention weights, and obtain the CT-guided clinical enhancement features through weight weighting. Taking the CT-guided clinical enhancement features as the query vector and the CT image adaptation features as the key vector, calculate the reverse cross-attention weights, and combine the lesion region segmentation mask of the CT image to obtain the clinically guided CT enhancement features. CT-guided clinical enhancement features are element-wise added to clinical-guided CT enhancement features, and then dimensionally fused using convolutional kernels to obtain CT-clinical cross-fusion features. Key region features of pathological images, weighted biological pathway features, and CT-clinical cross-fusion features are concatenated along the channel dimension to obtain initial multimodal features. An adaptive gating unit adjusts the channel proportions of each modality feature. The gated multimodal features are input into a multi-head attention module. This module contains k parallel self-attention heads, each learning cross-modal associations in different dimensions. The outputs of the k attention heads are concatenated and linearly transformed to obtain global association features. An autoencoder is used to denoise the global association features to generate patient prognostic fusion data.

[0041] In this embodiment, CT image adaptation features are used as query vectors, and clinical embedding adaptation features are used as key vectors. Cross-attention weights are calculated, and CT-guided clinical enhancement features are obtained through weighted summation, including: CT image adaptation features Obtained through global average pooling as query vector Clinical embedding adaptation features as a key vector and value vector Calculate the cross-attention weights CT-guided clinical enhancement features were obtained through weighted averaging. Screening key clinical indicators related to CT imaging findings; This is the CT→clinical cross-attention weight vector.

[0042] In this embodiment, CT-guided clinical enhancement features are used as the query vector, CT image adaptation features are used as the key vector, reverse cross-attention weights are calculated, and combined with the spatial attention mask of CT images, clinically guided CT enhancement features are obtained, including: CT-guided clinical enhancement features. as query vector CT image adaptation features After being projected onto d dimensions by a 1×1 convolution, it becomes the key vector. and value vector Calculate the reverse cross-attention weights , The clinical-to-CT cross-attention weight map is generated; and combined with the lesion region segmentation mask of the CT image (automatically generated by the nnU-Net model, Dice coefficient > 0.85), the clinical-guided CT enhancement features are obtained. , where ⊙ represents element-wise multiplication, and the reshape operation flattens the weight map.

[0043] In this embodiment, CT-guided clinical enhancement features are added element-wise to clinical enhancement features, and then dimensionality is fused using a convolution kernel to obtain CT-clinical cross-fusion features, including: CT-guided clinical enhancement features... Compared with clinically guided CT enhancement features Element-level addition is performed, followed by nonlinear fusion using two layers of MLP (128→64 units, ReLU activation) to obtain CT-clinical cross-fusion features. .

[0044] In this embodiment, pathological key region features, weighted biological pathway features, and CT-clinical cross-fusion features are concatenated along the channel dimension to obtain initial multimodal features. An adaptive gating unit adjusts the channel proportions of each modality feature. The gated multimodal features are then input into a multi-head attention module. This multi-head attention module contains K parallel self-attention heads, each learning cross-modal associations in different dimensions. The outputs of the K attention heads are concatenated and linearly transformed to obtain global association features, including: pathological key region features... (Tumor invasion margin features were extracted using ResNet-50 and reduced to 32 dimensions via PCA), weighted biological pathway features (Based on the KEGG pathway database, the activity of 30 lung cancer-related pathways was weighted according to the fold change in gene expression), CT-clinical cross-fusion features. Concatenate the features along the channel dimension to obtain the initial multimodal features. The channel ratio of each modality feature is adjusted by an adaptive gating unit (simplifying the gating mechanism). ,in For the Sigmoid function, For the gated weight matrix, The gated multimodal features are shown; ⊙ represents element-wise multiplication; The following characters represent the gated multimodal features. The input is a multi-head attention module, which contains eight parallel self-attention heads. Each attention head learns different dimensions of cross-modal associations (heads 1-2: intra-modal local associations, heads 3-5: inter-modal global associations, heads 6-8: pathway-pathology-imaging causal associations). The output of each attention head is: ; , , The query, key, and value transformation matrix for the i-th attention head is used. The global association features are obtained by concatenating and linearly transforming the outputs of the eight attention heads. Denoising of F_global was performed using an autoencoder (encoder 128→64 units, decoder 64→128 units, reconstruction loss weight 0.3) to generate patient prognostic fusion data. And it is concatenated with the core features described in claim 9 as the final input.

[0045] The working principle and beneficial effects of the above technical solution are as follows: Based on the hierarchical attention mechanism, cross-attention weights are calculated to deeply explore the dependency relationship between CT image features and clinical embedded features, and fully integrate image manifestations and clinical information; CT-clinical cross-fusion features and pathological features are stitched together, and the channel ratio is adjusted by an adaptive gating unit. Combined with a multi-head attention module to learn cross-modal associations, the complementarity of multimodal features is enhanced; after obtaining the global association features, an autoencoder is used for noise reduction to generate high-quality patient prognostic fusion data, providing strong support for accurate prognostic assessment.

[0046] In summary, multi-data fusion methods based on prognostic factors provide a more comprehensive and accurate solution for predicting lung cancer survival. By integrating multi-omics and multi-modal data and combining deep learning with survival analysis methods, predictive performance can be significantly improved.

[0047] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0048] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A lung cancer survival prediction system based on multi-data fusion of prognostic factors, characterized in that, include: The multi-source data acquisition module is used to collect imaging data, clinical data, pathological data and genomic data of patients' prognosis to form multi-source data of patients' prognosis; The multi-source data processing module is used to preprocess and extract features from multi-source patient prognosis data based on natural language processing, and to determine 3D depth features, clinical embedding features, pathological visual features and pathway perception features. The multi-source data fusion module is used to perform cross-modal alignment and fine-grained fusion of the extracted multimodal feature vectors, and capture intermodal dependencies based on a hierarchical attention mechanism to form patient prognosis fusion data; The survival analysis and prediction module is used to build a lung cancer survival prediction model based on artificial intelligence. It analyzes the patient prognosis fusion data based on the lung cancer survival prediction model and automatically predicts the patient's survival time, and displays it in a visual form.

2. The lung cancer survival prediction system based on multi-data fusion of prognostic factors according to claim 1, characterized in that, The multi-source data acquisition module includes: The image acquisition unit is used to acquire imaging data for patient prognosis, including CT and PET-CT images; The clinical data collection unit is used to collect clinical data on patient prognosis, including age, gender, TNM stage, ECOG score, smoking history, treatment history, and laboratory test results. The pathology acquisition unit is used to collect pathological data for patient prognosis, including tissue type, degree of differentiation, PD-L1 expression, and Ki-67 index. The genome acquisition unit is used to collect genomic data for patient prognosis, including driver gene mutations and tumor mutational burden.

3. The lung cancer survival prediction system based on multi-data fusion of prognostic factors according to claim 2, characterized in that, The multi-source data processing module includes: The data cleaning unit is used to clean the multi-source patient prognosis data, remove noise from the multi-source patient prognosis data, identify outliers in the multi-source patient prognosis data based on bias test, query missing values ​​in the multi-source patient prognosis data based on Python, and evaluate the outliers and missing values ​​in the multi-source patient prognosis data. If the outliers and missing values ​​are useful for predicting lung cancer survival, the outliers are corrected and the missing values ​​are filled; otherwise, the outliers and missing values ​​are deleted. The data transformation unit is used to transform multi-source patient prognostic data based on Z-score standardization, converting multi-source patient prognostic data with different dimensions and distribution ranges into a unified scale, eliminating the dimensional differences between multi-source patient prognostic data, and forming standardized multi-source patient prognostic data.

4. The lung cancer survival prediction system based on multi-data fusion of prognostic factors according to claim 3, characterized in that, The multi-source data processing module further includes: The feature extraction unit is used to extract features from multi-source patient prognostic data; Feature extraction from multi-source patient prognostic data includes: Feature extraction is performed on the image data, the tumor area is automatically segmented and extended to the peritumoral area, and depth features are extracted based on the 3D-SEResNet network to comprehensively capture the three-dimensional spatial heterogeneity of the tumor and extract radiomics features to form 3D depth features. Feature extraction is performed on clinical data, clinical text is processed based on a natural language processing model, and structured data is converted into high-dimensional feature vectors to form clinical embedding features; Feature extraction is performed on pathological data. High-resolution whole-slice images are cut into multiple small image patches. Each image patch is encoded using a pre-trained convolutional neural network. Through an attention-based multi-instance learning mechanism, the most important image regions for prognosis are automatically identified and weighted, and aggregated into pathological visual features. Feature extraction is performed on genomic data, and a pathway-aware multimodal framework is used for processing. Based on known biological knowledge, gene expression data is mapped to 186 specific biological pathways, narrowing the semantic gap between genomic data and imaging or pathological data, enabling information from different modalities to interact and forming pathway-aware features.

5. The lung cancer survival prediction system based on multi-data fusion of prognostic factors according to claim 4, characterized in that, The multi-source data fusion module includes: The feature fusion unit is used to deeply fuse the extracted multimodal feature vectors; Information interaction is carried out within the modality based on the self-attention mechanism. The unpaired label contrastive learning method is used to align the visual semantics of image patches with the functional semantics of biological pathways in the feature space. Following the biological prior that genotype determines phenotype, the model is guided by biological pathway features to find key pathological image regions related to them, thereby achieving fine-grained fusion of gene and pathological data. Based on the hierarchical attention mechanism, CT image features and clinical embedded features are used as queries, keys, and values. Through attention weight calculation, the dependency relationship between image manifestations and clinical symptoms and test indicators is deeply explored. The features after cross-attention fusion are spliced ​​with the processed pathological features. Through the multi-head attention mechanism, global and cross-modal related features are extracted to generate patient prognostic fusion data.

6. The lung cancer survival prediction system based on multi-data fusion of prognostic factors according to claim 5, characterized in that, The survival analysis and prediction module includes: Model building unit, used to build lung cancer survival prediction models; Based on the needs of lung cancer survival prediction, historical data on lung cancer survival prediction were collected and divided into training set and test set. The deep learning model is trained using a training set, enabling it to learn lung cancer survival prediction behavior autonomously from the training set and automatically predict the patient's survival probability in the next 1, 3, and 5 years, thus determining the lung cancer survival prediction model. The lung cancer survival prediction model was tested using a test set. The performance of the lung cancer survival prediction model was evaluated based on the consistency index, calibration curve and ROC curve. The model was used to determine whether it could automatically predict the survival probability of patients in the next 1 year, 3 years and 5 years and to determine the model test evaluation results. When the lung cancer survival prediction model cannot automatically predict the patient's survival probability in the next 1 year, 3 years and 5 years, the parameters of the lung cancer survival prediction model are adjusted and iteratively optimized until the lung cancer survival prediction model can automatically predict the patient's survival probability in the next 1 year, 3 years and 5 years, and the optimal lung cancer survival prediction model is determined.

7. The lung cancer survival prediction system based on multi-data fusion of prognostic factors according to claim 6, characterized in that, The survival analysis and prediction module includes: The analysis and prediction unit is used to analyze patient prognostic fusion data and automatically predict patient survival. Deploy the lung cancer survival prediction model and place it in the actual patient lung cancer survival prediction environment; Patient prognostic fusion data is input into the lung cancer survival prediction model. The model analyzes the patient prognostic fusion data and automatically predicts the patient's survival probability in the next 1 year, 3 years, and 5 years, thus determining the patient's lung cancer survival prediction result.

8. The lung cancer survival prediction system based on multi-data fusion of prognostic factors according to claim 7, characterized in that, It also includes a visualization module, which is used to display the patient's lung cancer survival prediction report in a visual form, including the patient's lung cancer survival prediction results and follow-up recommendations.

9. The lung cancer survival prediction system based on multi-data fusion of prognostic factors according to claim 6, characterized in that, The deep learning model is trained using a training set, enabling it to autonomously learn lung cancer survival prediction behavior from the training set and automatically predict the patient's survival probability in the next 1, 3, and 5 years. This establishes the lung cancer survival prediction model, including: Obtain a training dataset, which includes core features and follow-up time series data; A dynamic coefficient Cox regression module is constructed based on a time-dependent regression coefficient function. The regression coefficients are fitted using a piecewise cubic spline function, and the parameters are solved using maximum likelihood estimation. The core features are input, and the survival risk function is output. A temporal Bayesian network module is constructed, dividing time into multiple discrete intervals and building sub-networks. The EM algorithm is used in conjunction with follow-up data to learn time-dependent parameters. A discretized state transition probability matrix is ​​introduced to characterize the probability of variable changes over time, and the survival risk function is output. An adaptive weighted fusion strategy is adopted, dynamically adjusting the weight coefficients based on the prediction errors of the two modules on the same time grid in the training data. The fusion yields the basic survival risk function and the basic survival probabilities for 1 year, 3 years, and 5 years, completing the construction of the basic survival prediction unit. Multiple nodes are defined; initial edge weights between nodes are calculated based on Pearson correlation coefficient, clinical guideline logical correlation strength, or mutual information value; a nonlinear time decay function is introduced to construct a mechanism interaction graph with dynamic edge weights; an improved graph attention network is adopted, which aggregates neighborhood node information through a multi-head attention mechanism, introduces a mechanism contribution attention layer to learn time-dependent mechanism contribution weights, and outputs mechanism interaction embedding features; the mechanism interaction embedding features are used as additional input to supplement the basic survival prediction unit, and the basic survival risk function is corrected to obtain the updated survival risk function, thus completing the construction of the multi-mechanism collaborative inference layer; Using the spline coefficients of the basic survival prediction unit, the transition probability of the temporal Bayesian network, and the attention weights of the graph attention network as optimization objects, we define the patient individual feature vector, follow-up time, C-index prediction performance index, and feature distribution offset as the state space, construct a multi-objective reward function, and complete the construction of the dynamic adaptive optimization layer. Using patient follow-up time-series data as input, a time-series feature sequence is constructed, including baseline survival risk values ​​at each time point, mechanism interaction embedding features, and clinical follow-up indicators. An LSTM and attention pooling architecture is employed, with the LSTM layer capturing long-term dependencies of time-series features and the attention pooling layer enhancing the weights of key time-point features. This outputs continuous survival predictions, 1-year / 3-year / 5-year survival probabilities and confidence intervals, and risk inflection points. Based on the slope features of the risk evolution curve and the baseline risk score, a dynamic risk stratification standard is constructed, completing the construction of the risk evolution prediction layer. The training dataset was grouped by patient, and an end-to-end joint training strategy was adopted. The basic survival prediction unit, multi-mechanism collaborative inference layer, dynamic adaptive optimization layer and risk evolution prediction layer were synchronously optimized by sharing a loss function. When the C-index of the validation set exceeded 0.75 and the p-value of the risk stratification Log-rank test was <0.01, the lung cancer survival prediction model was obtained.

10. The lung cancer survival prediction system based on multi-data fusion of prognostic factors according to claim 5, characterized in that, Based on a hierarchical attention mechanism, CT image features and clinical embedded features are used as queries, keys, and values. Through attention weight calculation, the dependencies between imaging manifestations and clinical symptoms and laboratory indicators are deeply mined. The features fused through cross-attention are then concatenated with processed pathological features. A multi-head attention mechanism is used to extract global, cross-modal related features, generating patient prognostic fusion data, including: Using tumor region features from enhanced CT images as query vectors and clinical embedding adaptation features as key vectors, cross-attention weights are calculated, and CT-guided clinical enhancement features are obtained through weighted summation. Using CT-guided clinical enhancement features as query vectors and CT image adaptation features as key vectors, the reverse cross-attention weights are calculated and combined with the lesion region segmentation mask of CT images to obtain clinically guided CT enhancement features. CT-guided clinical enhancement features are element-wise added to clinical-guided CT enhancement features, and then dimensionally fused using convolutional kernels to obtain CT-clinical cross-fusion features. Key region features of pathological images, weighted biological pathway features, and CT-clinical cross-fusion features are concatenated along the channel dimension to obtain initial multimodal features. An adaptive gating unit adjusts the channel proportions of each modality feature. The gated multimodal features are input into a multi-head attention module. This module contains k parallel self-attention heads, each learning cross-modal associations in different dimensions. The outputs of the k attention heads are concatenated and linearly transformed to obtain global association features. An autoencoder is used to denoise the global association features to generate patient prognostic fusion data.

Images