A method and system for evaluating the prognosis risk stratification of tumor patients based on deep learning and multi-omics data

Abstract

The application relates to the technical field of artificial intelligence, and discloses a tumor patient prognosis risk stratification evaluation method and system based on deep learning and multi-omics data. The method comprises the following steps: collecting and pre-processing multi-omics data; extracting each omics specificity representation vector through a multi-path parallel network; generating a global prognosis atlas by fusing features by using a cross-modal attention mechanism; inputting a survival risk prediction model to output a risk score and a survival curve; and realizing risk stratification according to a threshold. The system comprises a data acquisition and preprocessing module, a feature mapping subsystem, a cross-modal fusion engine and a risk stratification calculation terminal. The application significantly improves the accuracy, robustness and interpretability of prognosis evaluation by deeply fusing multi-omics information and a deep learning architecture, and provides a scientific basis for clinical precise treatment.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a method and system for stratifying the prognostic risk of cancer patients based on deep learning and multi-omics data. Background Technology

[0002] With the rapid development of biomedical technology, precision medicine has become a core concept in cancer diagnosis and treatment, aiming to achieve personalized treatment plans through in-depth analysis of patients' biomarkers. Prognostic risk assessment for cancer patients, as a crucial link in precision medicine, directly affects the selection of treatment plans, adjustment of drug dosages, and formulation of postoperative follow-up plans, playing a vital role in improving patient survival rates and optimizing the allocation of medical resources.

[0003] The integrated analysis of multi-omics data provides a comprehensive perspective for understanding the heterogeneity and complex evolution of tumors, encompassing biological information at multiple levels, including the genome, transcriptome, proteome, and metabolome. Utilizing the powerful nonlinear feature extraction capabilities of deep learning to fuse massive, high-dimensional multi-omics data has become an important research direction in the field of tumor prognostic assessment, aiming to accurately predict patient clinical outcomes by constructing high-precision predictive models.

[0004] Existing prognostic risk stratification assessment methods for cancer patients suffer from several drawbacks: First, insufficient data integration depth, often relying solely on single-omics information, makes it difficult to capture synergistic regulatory relationships and complex interaction networks across different biological levels. Second, given the inherent high dimensionality, sparsity, and heterogeneity of multi-omics data, traditional statistical models or shallow machine learning methods are prone to the curse of dimensionality, limiting model generalization ability and predictive accuracy. Third, the feature selection process lacks sufficient consideration of biological logic, making it difficult to effectively remove noisy data and retain key features strongly correlated with prognosis. Finally, the model's interpretability is poor, failing to reveal differences in molecular characteristics among different risk groups. This leads to significant uncertainty and limitations in guiding clinical decision-making with existing prognostic assessment results. To address the shortcomings of traditional assessment methods, such as insufficient data integration depth, limited model generalization ability, lack of biological logic in feature selection, and poor model interpretability, the prognostic risk stratification assessment method for cancer patients based on deep learning and multi-omics data proposed in this invention is particularly important. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for assessing the prognostic risk stratification of cancer patients based on deep learning and multi-omics data. This method and system can effectively solve the problems mentioned in the background. This invention is committed to overcoming the technical defects of existing technologies, such as insufficient data integration depth, the curse of dimensionality caused by high-dimensional data, lack of biological logic in feature selection, and poor interpretability of model evaluation results, through multi-dimensional biological data integration and deep nonlinear feature mining. This will enable accurate identification and scientific stratification of the survival risk of cancer patients.

[0006] To achieve the above objectives, the technical solution adopted by this invention is as follows: a method for stratifying the prognostic risk of tumor patients based on deep learning and multi-omics data, comprising the following steps: S1. Collect multi-omics raw data from tumor patients, and perform denoising and standardization preprocessing on the multi-omics raw data to obtain a multi-omics standard feature set; S2. Construct a multi-path parallel feature extraction network to perform deep feature mapping on each dimension of the multi-omics standard feature set and extract the specific representation vector of each omics. S3. Based on the cross-modal attention mechanism, dynamic weight allocation and nonlinear fusion are performed on the specific representation vectors of each omics to construct a global prognostic feature map; S4. Input the global prognostic feature map into the preset survival risk prediction model for iterative training and evaluation, and output the patient's survival probability distribution curve and risk score. S5. The risk score is compared with the preset risk stratification threshold to classify the prognostic risk level of cancer patients.

[0007] Preferably, step S1 specifically includes the following steps: S11. Obtain gene mutation data, transcriptome expression data, deoxyribonucleic acid methylation data, and clinical phenotype data of cancer patients through biological sample detection equipment; S12. Perform sparsity processing on the gene mutation data, retain the mutation sites with a frequency higher than a preset threshold, and convert them into a binary encoding matrix; S13. Perform logarithmic transformation and mean normalization on the transcriptome expression data to eliminate sequencing depth bias between different samples; S14. Perform quality control on the deoxyribonucleic acid methylation data, remove probe sites with a missing value ratio higher than a certain ratio, and use the nearest neighbor algorithm to complete the remaining missing values.

[0008] Preferably, step S2 specifically includes the following steps: S21. Input the processed omics data into the corresponding residual connection encoders respectively; S22. Using the multilayer linear mapping unit and nonlinear activation function in the residual connection encoder, low-dimensional biological features are mapped to a high-dimensional feature embedding space. S23. In the feature mapping process, a dropout layer is introduced to randomly deactivate some neuron nodes according to a predetermined probability, so as to enhance the robustness of the model to high-dimensional sparse data. S24. Output the hidden layer state of each omics data in a unified dimension, as the specific representation vector of each omics.

[0009] Preferably, step S3 specifically includes the following steps: S31. Calculate the correlation matrix between the specific characterization vectors of each omics to measure the strength of synergistic effects between different biological levels. S32. Using the transformation logic of query vector, key vector and value vector, calculate the attention distribution weight of each set of learning features for the final prognosis contribution. S33. Weighted summation of each omics feature according to the attention distribution weights to capture the nonlinear interaction relationships between different omics; S34. The weighted features are compressed and refined through a fully connected layer to generate a global prognostic feature map that represents the overall biological state of the patient.

[0010] Preferably, step S4 specifically includes the following steps: S41. Construct a negative log-likelihood loss function based on the proportional risk assumption and use it as the objective function for model optimization. S42. Input the global prognostic feature map into the survival prediction layer and calculate the cumulative risk function of the patient at different time points; S43. The gradient descent algorithm is used to update the model parameters so that the predicted risk ranking output by the model is consistent with the actual observed survival time ranking. S44. Introduce a regularization penalty term to constrain the complexity of model parameters and prevent overfitting for small sample multi-omics data.

[0011] Preferably, the multi-omics raw data covers the complete information chain from the microscopic molecular level to macroscopic clinical manifestations.

[0012] Preferably, the gene mutation data reflects the stability of the tumor genome and the characteristics of driver mutations.

[0013] Preferably, the transcriptome expression data reflects changes in intracellular functional execution and metabolic activity.

[0014] Preferably, the deoxyribonucleic acid methylation data characterizes the regulatory logic at the epigenetic level, and has a significant biological correlation with differential gene expression.

[0015] Preferably, the clinical phenotypic data includes the patient's age, sex, pathological stage, treatment plan, and follow-up survival status, which are used to provide basic background constraints for the deep learning model.

[0016] Preferably, in the preprocessing process, for continuous numerical variables, a preset quantile standardization method is used to transform them to a specific numerical range to ensure the comparability of omics data at different scales in the feature space. Preferably, for categorical variables, one-hot encoding is used to convert them into numerical vectors. Preferably, when dealing with the heterogeneity of multi-omics data, a unified feature indexing system is constructed to ensure that the features of the same patient at different omics levels can be accurately aligned, thereby establishing cross-omics associations.

[0017] Preferably, the residual connection encoder, when extracting features, directly transmits the raw information from the shallow layers to the deep layers through skip connections, effectively mitigating the gradient vanishing problem that may occur in deep neural networks when processing large amounts of biological data. Preferably, the nonlinear activation function is configured as a function model with one-sided inhibition characteristics, which enhances the model's ability to nonlinearly fit complex biological signals. Preferably, the dimension of the high-dimensional space mapping is dynamically adjusted according to the complexity of the input omics data, ensuring that sufficient feature richness is retained while avoiding redundant computation.

[0018] Preferably, the cross-modal attention mechanism automatically learns the importance of different omics, enabling it to identify key omics factors that have the greatest impact on prognosis. Preferably, the attention weights are continuously iteratively optimized during training, reflecting the logic of multi-omics synergistic regulation in biology. Preferably, the nonlinear fusion process is not a simple feature stacking, but rather captures the deep coupling relationships between gene expression and methylation, and mutation and phenotype through interactive mapping.

[0019] Preferably, the survival risk prediction model, when calculating the risk score, not only considers the overall distribution of the entire dataset but also fine-tunes the parameters for the characteristics of different tumor subtypes. Preferably, the risk stratification threshold is set based on the statistical distribution of a large number of clinical samples, and patients are divided into high-risk, medium-risk, and low-risk groups by finding the optimal survival difference cutoff point. Preferably, the survival probability distribution curve uses a parametric model to probabilistically model future survival status, providing clinicians with an intuitive reference for risk trends.

[0020] A prognostic risk stratification assessment system for cancer patients based on deep learning and multi-omics data, used to implement the above method, includes: The data acquisition and preprocessing module is used to acquire and clean the genomic, transcriptomic, and methylation data of cancer patients and output a standardized dataset. The feature mapping subsystem, whose input is connected to the data acquisition and preprocessing module, is used to extract omics-specific features through a multi-path network. A cross-modal fusion engine is used to receive the omics-specific features and generate a global feature vector based on an attention mechanism; The risk stratification calculation terminal is configured to receive the global feature vector and output the risk stratification results and survival prediction curves through the prognostic model.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention constructs a prognostic assessment system with high predictive accuracy and biological interpretability through the deep integration of deep learning frameworks and multi-omics data. First, it overcomes the limitations of single-omics data by integrating multi-dimensional information such as gene mutations, transcriptional expression, and methylation, enabling a more comprehensive characterization of tumor heterogeneity and complex regulatory networks, thus significantly improving the biological completeness of prognostic assessment. Second, the introduced parallel feature extraction network and cross-modal attention mechanism effectively address the high-dimensional sparsity problem inherent in multi-omics data. Dynamic weight allocation achieves automatic focusing on key biomarkers, avoiding the tedious and error-prone manual feature selection process in traditional methods. In summary, this invention demonstrates strong robustness and generalization ability in addressing the complex nonlinear problem of tumor prognostic assessment, providing more accurate risk scores and personalized prognostic stratification for cancer patients, and offering scientific data support for the development of precision treatment plans in clinical practice.

[0022] Furthermore, the application of cross-modal attention mechanisms endows the model with a degree of interpretability. By analyzing the distribution of attention weights, researchers can trace back and identify the omics dimensions and key molecular features that play a decisive role in the prognosis of specific tumors. This has significant auxiliary value for discovering new biomarkers and understanding the molecular mechanisms of tumor malignant progression. This closed loop from data-driven to biological logistic regression makes the evaluation results more clinically convincing.

[0023] Furthermore, the residual connection encoder and regularization constraint strategy employed in this invention solve the overfitting problem commonly encountered in the biomedical field due to relatively small sample sizes and extremely high feature dimensionality. By performing efficient feature dimensionality reduction and abstraction in the latent space, the model can learn highly generalizable prognostic correlation patterns, ensuring consistent evaluation across different centers and batches of data, and improving the reliability of clinical applications.

[0024] Furthermore, by constructing a negative log-likelihood loss function based on survival time, this invention directly models survival risk, rather than through simple classification or regression. This enables the model to handle right-censored data unique to survival analysis, fully utilizes the crucial information of follow-up time, significantly improves the accuracy of predicting long-term survival outcomes for patients, and achieves dynamic risk monitoring over time.

[0025] Furthermore, the systematic architecture design of this invention enables the method to have good scalability, and it can seamlessly integrate multimodal data such as proteomics, metabolomics or radiomics according to the development of subsequent biological detection technologies. By adding corresponding parallel feature extraction paths, the evaluation efficiency of the model can be further enhanced.

[0026] Furthermore, the automated preprocessing pipeline and stratification threshold setting mechanism reduce the reliance on operators' professional statistical knowledge, achieving standardization and intelligence in the assessment process. Clinicians only need to input the raw test data to obtain a high-confidence risk stratification report, greatly optimizing the process and efficiency of clinical decision-making.

[0027] Furthermore, by conducting differential analysis of the molecular profiles of patients in different risk groups, this invention not only provides risk scores but also helps to reveal the activation status of signaling pathways unique to high-risk groups, providing in-depth pharmacological clues for subsequent targeted drug screening and treatment regimen adjustments, truly achieving a leap from risk assessment to treatment guidance.

[0028] Furthermore, through continuous iterative training on massive clinical samples, this invention enables the parameter distribution within the model to gradually approximate the real biological laws. It not only performs excellently in known tumor types, but also demonstrates superior predictive sensitivity compared to traditional clinical statistical models for rare tumors or complex cases with high heterogeneity.

[0029] Furthermore, the system, through a modular hardware interface, can efficiently exchange data with existing laboratory information management systems and biological sequencing platforms, ensuring the real-time nature and security of the data flow, and laying an engineering foundation for conducting prognostic risk assessments in large-scale clinical cohorts.

[0030] Furthermore, by introducing a dropout layer and an adaptive learning rate adjustment strategy, the model can automatically balance the convergence speed and the search for the global optimum during training, effectively avoiding the local optimum trap, ensuring the mathematical rigor of the generated risk assessment model and its effectiveness in biological practice, and ultimately achieving a scientific, accurate, and efficient assessment of the survival risk of cancer patients. Attached Figure Description

[0031] Figure 1This is a schematic diagram of the overall technical architecture of the deep learning-based prognostic risk stratification assessment method for tumor patients proposed in this invention. Figure 2 This is a schematic diagram of the core principle framework of nonlinear fusion of multi-omics features based on cross-modal attention mechanism in this invention; Figure 3 This is a flowchart illustrating the logical process of multi-omics raw data preprocessing and multi-path parallel feature mapping extraction in this invention. Figure 4 This is a flowchart illustrating the logical process of iterative training and risk score calculation for the survival risk prediction model in this invention. Detailed Implementation

[0032] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0033] Example 1 This embodiment provides a method for stratifying the prognostic risk of cancer patients based on deep learning and multi-omics data. This method aims to provide accurate survival risk prediction and scientific intervention basis for clinical medicine by deeply integrating the cross-dimensional biological information of cancer patients.

[0034] In step S1, the system's primary task is to collect multi-omics raw data from cancer patients. This process involves comprehensive data capture from the microscopic molecular level to the macroscopic clinical phenotype. Specifically, step S11 uses high-throughput sequencing equipment, mass spectrometry, and a clinical information system to acquire the patient's raw sequence information and clinical records. Gene mutation data mainly records genetic alterations such as single nucleotide variants, insertions, or deletions in the patient's genome; these alterations are often driving factors in tumor development and progression. Transcriptome expression data reflects the transcriptional activity of genes at specific times and spaces, demonstrating the real-time metabolic state of cells. Deoxyribonucleic acid methylation data, as an important component of epigenetics, reveals the deep logic of gene expression regulation by describing the methylation modification status of cytosine bases. Clinical phenotype data includes the patient's age, sex, pathological stage, time of diagnosis, follow-up status, and specific treatment plan; these data provide necessary background anchors for the model.

[0035] In step S12, the system performs sparsity processing to address the high-dimensional sparsity of the gene mutation data. Specifically, the frequency of each mutation site in all samples is counted, and low-frequency mutations occurring below a preset proportion (e.g., 1%) are considered background noise or non-universal features and removed. The remaining high-frequency mutation sites are converted into a binary encoding matrix; if a patient has a mutation at a specific site, that position is marked as 1 in the matrix, otherwise it is marked as 0. This processing method effectively transforms discrete gene mutation information into a structured logical expression that can be directly processed by a computer, reducing the dimensionality pressure on the feature space.

[0036] Step S13 focuses on the standardization of transcriptome expression data. Due to systematic errors between samples caused by library construction and varying sequencing depths during sequencing, the system first performs a logarithmic transformation on the raw expression counts to compress the dynamic range of the data and mitigate the interference of extreme values ​​on model training. Subsequently, using a mean normalization algorithm, the mean and standard deviation of expression for each gene in all samples are calculated. By subtracting the mean and dividing by the standard deviation, the expression levels are mapped to a distribution space centered at 0 with a standard deviation of 1, thereby eliminating scale differences between different batches of data.

[0037] Step S14 performs quality control on the deoxyribonucleic acid methylation data. The system automatically scans tens of thousands of probe sites, removing entire rows of probes with missing values ​​exceeding 20%. For the remaining few missing values, the system does not simply fill in zeros, but instead uses a nearest neighbor algorithm. This algorithm infers and completes the missing site by calculating the average of a preset number of neighboring samples closest to the target sample in the feature space. This approach preserves the continuity and integrity of the biological signal to the greatest extent possible. For continuous numerical variables, such as certain biochemical indicators in clinical data, this embodiment uses quantile normalization. This method arranges and maps the data to preset normal or uniform distribution quantiles, ensuring that omics features at different scales have equivalent numerical magnitudes before entering the deep learning network. For categorical variables such as gender and pathological stage, one-hot encoding is used to transform them into multidimensional orthogonal vectors composed of 0s and 1s, avoiding erroneous numerical relationships between categories.

[0038] In step S2, the system constructs a multi-path parallel feature extraction network.

[0039] In step S21, the standardized genome, transcriptome, methylome, and clinical features are fed into four independent residual connection encoders. Each encoder is a specific feature mining path designed to capture deep nonlinear correlations within that dimension.

[0040] In step S22, each encoder contains multiple layers of linear mapping units. Linear information transfer is achieved by element-wise multiplication and summation of the input feature vector and the weight matrix. The result is then fed into a non-linear activation function configured as a modified linear unit with unilateral inhibition characteristics. This function simulates the excitation threshold mechanism of biological neurons, allowing only signals exceeding a certain intensity to pass through, thereby enhancing the model's ability to fit complex patterns. During this process, low-dimensional biological features are progressively mapped to a higher-dimensional feature embedding space, making the originally intertwined biological signals separable in the latent space.

[0041] Step S23 introduces a dropout layer during feature mapping. In each iteration of model training, the dropout layer randomly disconnects some neurons according to a preset probability (e.g., 30%), forcing the model to no longer rely on certain strong features, but instead learn more robust distributed representations. This mechanism is crucial for handling the common problems of small sample sizes and high dimensionality in multi-omics data, effectively preventing overfitting. Finally, in step S24, the hidden layer states of each path are output in a unified dimension, serving as omics-specific representation vectors, laying the foundation for subsequent cross-modal fusion.

[0042] Step S3 involves the implementation of the cross-modal attention mechanism.

[0043] Step S31 involves calculating the association matrix between different omics-specific characterization vectors. This association is not a simple numerical summation, but rather a measure of the strength of synergistic effects between different biological levels by calculating the dot product similarity between two feature vectors. For example, when the methylation level of a gene is highly correlated with its transcriptional expression level in prognostic assessment, the association matrix will show a high value at that position.

[0044] Step S32 utilizes the transformation logic of query vectors, key vectors, and value vectors. The system first transforms the representation vector of each omics through three different linear transformations to generate corresponding query, key, and value components. By calculating the inner product of the query component of the current omics with the key components of all other omics, and processing it using an exponential normalization function, a set of attention distribution weights reflecting the contribution of different omics to the final prognosis is obtained. This set of weights dynamically reflects whether gene mutations are dominant or transcriptional regulation plays a core role in different patient individuals.

[0045] Step S33 performs a weighted summation of the value vectors of each omics based on the obtained attention distribution weights. This process achieves nonlinear fusion; it is not a simple feature stacking but rather captures the deep interactions between different omics. Step S34 further compresses and refines the fused high-dimensional vector through a fully connected layer. The fully connected layer integrates the scattered collaborative signals into a global prognostic feature map characterizing the patient's overall biological state through many-to-many weighted connections. This map contains comprehensive prognostic information from molecular to clinical levels.

[0046] In step S4, the system inputs the global prognostic feature map into the survival risk prediction model.

[0047] Step S41 constructs a negative log-likelihood loss function based on the proportional hazards assumption. The design logic of this function is to focus not only on whether a patient experiences death or relapse, but also on the order in which these events occur. The loss function calculates the consistency deviation between the risk score output by the model and the actual observed survival time ranking.

[0048] Step S42 sends the feature map into the survival prediction layer, which calculates the cumulative risk function of the patient at different time points through exponential mapping.

[0049] Step S43 employs the gradient descent algorithm. In each training round, the system calculates the gradient direction of the loss function with respect to tens of thousands of parameters within the model, and updates the parameters in reverse according to a preset learning rate step size. The goal of the update is to ensure that the predicted scores of high-risk patients are consistently higher than those of low-risk patients, thereby achieving more accurate risk ranking.

[0050] Step S44 introduces a regularization penalty term, which punishes overly complex connections in the model by constraining the L2 norm of the parameters, thereby further improving the model's generalization performance on unknown datasets.

[0051] Finally, step S5 categorizes the prognostic risk levels of cancer patients. The system compares the continuous risk score output by the model with a preset risk stratification threshold. This threshold is determined by analyzing survival differences in a large-scale historical retrospective cohort to find the tangent point at which the difference in survival curves between the two groups reaches statistical significance. Thus, patients are classified into high-risk, intermediate-risk, and low-risk groups. Simultaneously, the system generates survival probability distribution curves, which depict the decay trend of the patient's expected survival probability over the next 3, 5, or even 10 years.

[0052] This invention also provides a prognostic risk stratification assessment system for cancer patients based on deep learning and multi-omics data. This system includes a data acquisition and preprocessing module, a feature mapping subsystem, a cross-modal fusion engine, and a risk stratification computing terminal. The data acquisition and preprocessing module connects to the sequencer and hospital database through a standardized data interface, responsible for cleaning, completing, and normalizing the raw data. The feature mapping subsystem consists of multiple high-performance computing units that run the residual connection encoders of each omics in parallel. The cross-modal fusion engine is responsible for performing complex attention mechanism calculations. The risk stratification computing terminal presents the final analysis results to clinicians in the form of a visual report.

[0053] Example 2 Building upon Example 1, this example further explores how the present invention can achieve more accurate prognostic assessment when dealing with highly heterogeneous non-small cell lung cancer patients by optimizing data preprocessing logic and feature fusion strategies.

[0054] The prognosis of non-small cell lung cancer is significantly affected by mutations in multiple driver genes, while the transcriptional state of the tumor microenvironment also plays a crucial role. In step S11, this embodiment specifically enhances the sampling depth of key driver sites such as the epidermal growth factor receptor gene and the anaplastic lymphoma kinase gene. In the gene mutation processing of step S12, to highlight the importance of driver mutations, the system introduces weighted binary encoding. For known oncogenic driver sites, their value in the matrix is ​​no longer a simple 1, but rather a coefficient value greater than 1 is assigned based on the pathogenicity weight of the mutation in clinical literature, thereby guiding the model to focus on these key signals during the feature extraction stage.

[0055] In step S13, when processing transcriptome data, to address the RNA degradation problem caused by necrotic regions commonly found in lung cancer tissues, the system adds a sample quality assessment step before standardization. By calculating the distribution entropy of all gene expression levels, low-quality samples with overly flat expression profiles and lacking biological information are identified, and their weights in the final prediction are weakened. In the feature mapping step S22, to accommodate the complex cellular composition within lung cancer tissues, the number of layers in the residual connection encoder is increased to 8, and batch normalization layers are added between each layer to accelerate the convergence speed of the deep network and ensure that gradient explosion does not occur during signal propagation across multiple layers.

[0056] In this embodiment, the cross-modal attention mechanism in step S32 is optimized for the biological logic specific to lung cancer. Considering that gene methylation often affects prognosis by inhibiting gene expression, the system specifically enhances the contrast calculation between methylation feature vectors and corresponding transcript feature vectors when calculating the association matrix. When attention weights are assigned to methylation pathways, the system automatically retrieves transcriptional pathways with physical proximity and strengthens the negative feedback association logic between them during feature fusion. This feature fusion with prior biological knowledge enables the generated global prognostic feature map to more accurately reflect the molecular pathogenesis of lung cancer patients.

[0057] In the model iteration process of step S43, this embodiment introduces an adaptive learning rate strategy. In the early stages of training, a larger learning rate is used to quickly traverse the fluctuating regions of the non-convex loss function; as the number of training epochs increases, the learning rate decays exponentially, allowing the model to be fine-tuned near local optima. This dynamic adjustment mechanism ensures that the model can capture subtle prognostic differences between lung cancer subtypes.

[0058] The risk stratification results generated in step S5 are further correlated with specific adjuvant chemotherapy recommendations in this embodiment. High-risk patients not only receive lower predicted survival probabilities, but the system also identifies potential molecular pathways contributing to their high risk based on key contributing factors in the feature map. If a patient's risk score is primarily driven by a specific gene pathway, the risk stratification terminal will prompt clinicians to consider the possibility of corresponding targeted therapies, thus achieving a closed-loop process from prognostic assessment to treatment plan development.

[0059] Example 3 This embodiment focuses on describing the hardware implementation details of the present invention in a distributed computing environment and the stability guarantee mechanism in large-scale clinical cohort processing. When processing trillions of multi-omics data from thousands of patients, the computing power of a single machine is often insufficient to meet real-time requirements.

[0060] The evaluation system in this embodiment adopts a distributed deployment approach based on a cloud computing architecture. The data acquisition and preprocessing module is deployed on edge computing nodes, responsible for performing preliminary format verification and noise reduction before data upload. This design reduces the pressure on network transmission and ensures the localized anonymization of patient privacy data. In the preprocessing pipeline of step S1, multi-threaded concurrent processing technology is used to distribute the standardization tasks of different omics data to different processor cores. For example, the binaryization of gene mutations and the normalization of transcriptomes are physically parallel, which greatly shortens the time spent in the data preparation stage.

[0061] The feature mapping subsystem consists of multiple GPU accelerators. In step S2, the forward propagation operation of the residual connected encoder is mapped to the streaming multiprocessor of the GPU, leveraging the advantages of massively parallel computing to process matrix multiplication. To ensure hardware stability, in addition to the dropout layer in step S23, a hardware monitoring loop is added. When the temperature or memory usage of a computing unit exceeds a safe threshold, the system automatically migrates the computing task to a backup node, ensuring uninterrupted training.

[0062] In the fusion engine of step S3, to address the issue of the massive computational cost of the cross-modal attention mechanism, this embodiment introduces memory optimization techniques. By using half-precision floating-point numbers to store intermediate feature vectors, memory usage is reduced by half with almost no loss of model accuracy, enabling the system to process more omics dimensions within limited GPU memory. The training process in step S4 is performed using a distributed parameter server model. Multiple computing nodes jointly maintain a set of model weight parameters. The gradients calculated by each node are aggregated and averaged on the parameter server before being distributed back to the nodes. This architecture ensures global consistency in model convergence under large-scale data training.

[0063] The risk stratification calculation terminal is designed as a highly available web server, providing services to terminals in different hospitals through a standardized application programming interface. In the threshold comparison step S5, the system not only relies on preset static thresholds but also uses dynamic calibration technology to fine-tune the stratification cut-off points based on the distribution of the center's historical data. This means the assessment system has self-evolution capabilities, continuously optimizing the accuracy of its risk definitions as more data is accessed.

[0064] Furthermore, this embodiment also designs a unified feature indexing system to address data heterogeneity. At the physical storage level, by establishing a distributed key-value database with the patient's unique identification code as the key and the standard feature sets of each omics as the values, it ensures that gene, transcription, and clinical information can achieve millisecond-level precise alignment when performing cross-modal fusion, eliminating prognostic assessment bias caused by data alignment errors.

[0065] Example 4 This embodiment details the robust compensation mechanism of the present invention for low-quality samples with missing omics data or high levels of sequencing noise, ensuring the universality of the evaluation method in complex clinical settings.

[0066] In real-world clinical scenarios, not all patients can undergo comprehensive genomic, transcriptomic, and methylation testing simultaneously. When the data collected in step S1 has gaps at the omics level (e.g., a patient only has transcriptomic and clinical data), the system automatically activates a virtual feature generation mechanism. This mechanism, based on existing massive multi-omics association databases, utilizes generative adversarial networks to infer the probability distribution of missing methylation based on known clinical phenotypes and transcriptional features, generating a virtual feature vector with a confidence mask to participate in subsequent calculations.

[0067] In the feature extraction stage of step S2, a feature confidence evaluation unit is added inside the residual connection encoder. This unit dynamically adjusts the initial weight of the path in subsequent fusion based on the quality score of the input raw data (such as sequencing quality score, RNA integrity value). For low-quality paths, the encoder outputs a high-entropy hidden layer state, transmitting a signal to the subsequent attention mechanism, indicating that the information uncertainty in this dimension is high.

[0068] In step S3, the cross-modal attention mechanism automatically identifies paths with high-entropy or virtual features when calculating the weight distribution. When calculating the correlation between the query vector and the key vector, the attention mechanism introduces a penalty term to reduce the impact of low-quality paths on the final global feature map. For example, if a patient's transcriptome data is too noisy, the attention weights will spontaneously shift towards more reliable clinical phenotypes or gene mutation data, ensuring that the generated feature map can still stably represent the patient's prognostic status.

[0069] In the model optimization step S4, for noisy or missing data, the system uses a robust loss function instead of the standard negative log-likelihood loss. The robust loss function introduces a regulating factor to reduce the contribution of extreme outliers caused by data quality issues to parameter updates, preventing the model from being biased by individual noisy samples. Simultaneously, in the regularization constraints of step S44, the system dynamically adjusts the strength of the penalty term based on the sample quality classification. Weaker regularization is applied to high-quality samples to extract fine features, while stronger regularization is applied to low-quality samples to suppress noise interference.

[0070] In step S5, the risk stratification output not only provides the risk level but also a reliability assessment score. This score is derived by comprehensively considering data completeness, consistency of model predictions, and adequacy of follow-up time. For assessment reports with low reliability scores, the system recommends that clinicians conduct secondary validation using other routine diagnostic methods. This end-to-end quality control and compensation mechanism enables the present invention to maintain high accuracy and robustness in prognostic assessments in diverse biological testing environments.

[0071] Example 5 This embodiment further illustrates the specific implementation process of the present invention for transfer learning and model generalization across different tumor subtypes (such as various classifications of breast cancer). The prognostic logic of tumors has both commonalities and unique characteristics across different organs.

[0072] First, this embodiment utilizes a multi-path feature extraction network pre-trained on a large-scale pan-cancer dataset as the base model. In step S2, the low-level parameters of the residual connection encoder are locked, preserving its ability to extract general biological fundamental frequency features (such as basic cell cycle regulation, metabolic basic logic, etc.). When processing specific tumor subtypes, the system only fine-tunes the high-level feature mapping units.

[0073] In step S3, the query matrix and key matrix of the cross-modal attention mechanism are reinitialized to accommodate the differences in omics contributions across different cancer types. For example, in breast cancer assessment, receptor expression status in clinical phenotypes has extremely high weight; the attention mechanism automatically learns this prior logic during training and embeds it into the weight allocation process. The nonlinear fusion process in step S33 opens up specific mapping regions in the feature map targeting the hormone regulatory pathways specific to breast cancer.

[0074] In the iterative training of step S4, this embodiment employs a meta-learning strategy. The model no longer simply learns how to predict the survival risk of a particular patient, but rather learns how to quickly adjust its assessment logic when faced with a new subtype. The system constructs multiple small-scale task sets to simulate different tumor scenarios, enabling the survival risk prediction model to possess extremely strong adaptability. When new clinical data is input, the model requires only a very small number of labeled samples to achieve high-precision convergence on that specific subtype.

[0075] In step S5, the risk stratification threshold adopts a multi-center collaborative calibration mode. Due to the bias in patient composition across different medical institutions, the system uses a federated learning framework to calculate the statistical characteristics of risk distribution locally at each center without exchanging raw data, and then uploads these calculations to the central processing unit (CPU) for aggregation. The CPU generates a globally optimized dynamic threshold surface, which is then fed back to each terminal. This process ensures that the risk stratification of this invention has high credibility and consistency across different regions and populations.

[0076] Example 6 This embodiment details the logical interactions and data flow between the various functional modules in this invention. The entire lifecycle, from data acquisition to final report generation, strictly adheres to the pre-defined technical logic chain.

[0077] The data acquisition and preprocessing module serves as the system's entry point, integrating the textual logic of various bioinformatics analysis tools. After acquisition (S11), the data stream enters the quality filtering engine. Following the logic of steps S12 to S14, the filtering engine performs parallel cleaning on four types of heterogeneous data. Gene mutation data flows through a sparsity filtering layer, transforming it into binary tensors; transcriptome data flows through a log-normalization layer; and methylation data flows through a nearest-neighbor completion layer. These standardized tensors are loaded into a unified patient ID index in memory, forming a multi-omics standard feature set.

[0078] Subsequently, the standard feature set is pushed into the buffer of the feature mapping subsystem. Within the feature mapping subsystem, the four residual paths begin non-linear feature mining according to pre-defined activation function logic. The output of each hidden layer state is fed back to the resource scheduler in real time to ensure balanced computational load. After the four-dimensional feature embedding vectors are generated, they are aggregated in the convergence layer of the cross-modal fusion engine.

[0079] The cross-modal fusion engine is the core brain of the system. It establishes complex relationships between these feature vectors by executing steps S31 to S34. Attention distribution weights are calculated in real-time and applied to the value vectors, a process involving large-scale matrix dot product operations. The fused global prognostic feature map is a compact, high-dimensional semantic representation, temporarily stored in the system's global shared memory.

[0080] The risk stratification calculation terminal reads the feature map and performs risk score calculation by calling the parameters of the survival prediction layer. This process follows the prediction logic in step S4, generating a continuous value reflecting the mortality risk density. Subsequently, this value enters the stratification determination logic unit. The determination unit compares the score with dynamic thresholds stored in the database to determine the patient's risk level. Finally, the visualization engine calls the plotting component to plot the probability density function of survival probability over time as a smooth survival curve, and combines it with the patient's molecular characteristics to generate a comprehensive report that includes prognostic stratification, molecular interpretation, and clinical recommendations.

[0081] The entire system operates under the control of a master program, which synchronizes the execution sequence of each module and handles any abnormal interruptions. If the processing time of any stage exceeds a preset watchdog threshold, the master program triggers self-checking logic to locate the bottleneck and attempt to restart that task branch. This rigorous architectural design and logical closed loop ensure that the prognostic risk assessment for cancer patients based on deep learning and multi-omics data can run stably and efficiently in a real clinical production environment.

[0082] In summary, this invention, through multi-level technological innovation, constructs a rigorous, scientific evaluation system with extremely high predictive accuracy, from the bottom-level standardization of multi-omics data, to the mid-level parallel residual feature mining and cross-modal attention fusion, and finally to the top-level survival risk hierarchical modeling. It not only overcomes the shortcomings of existing technologies, such as insufficient data integration depth, the curse of dimensionality, and missing biological logic, but also significantly improves the model's generalization ability and reliability in practical applications by introducing engineering methods such as hardware acceleration, robustness compensation, and transfer learning. This invention provides not only a predictive score, but also a deep and interpretable biological atlas, offering strong technical support for the implementation of precision oncology and possessing profound clinical value and social benefits for improving patient quality of life and optimizing the allocation of medical resources.

[0083] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for stratifying the prognostic risk of cancer patients based on deep learning and multi-omics data, characterized in that, Includes the following steps: S1. Collect multi-omics raw data from tumor patients using biological sample detection equipment. Perform sparsity processing on gene mutation data in the multi-omics raw data, perform logarithmic transformation and mean normalization on transcriptome expression data, perform quality control and missing value completion on deoxyribonucleic acid methylation data, and combine with clinical phenotype data to accurately align omics information of different dimensions in the sample dimension by constructing a unified feature index system to obtain a multi-omics standard feature set. S2. Construct a multi-path parallel feature extraction network, input the omics data of each dimension in the multi-omics standard feature set into the corresponding residual connection encoder, and use the multi-layer linear mapping unit and nonlinear activation function inside the residual connection encoder to map the low-dimensional biological features to the high-dimensional feature embedding space. During the mapping process, some neuron nodes are randomly deactivated by the dropout layer according to a predetermined probability, and the hidden layer state of each omics data in the same dimension is output as the specific representation vector of each omics. S3. Based on the cross-modal attention mechanism, the specific representation vectors of each omics are dynamically weighted and nonlinearly fused. By calculating the dot product similarity between different omics specific representation vectors, an association matrix is ​​constructed to measure the synergistic effect strength between different biological levels. The attention distribution weight of each omics feature to the final prognosis is calculated using the transformation logic of query vector, key vector and value vector. Based on the attention distribution weight, each omics feature is weighted, summed and compressed to generate a global prognosis feature map. S4. Input the global prognostic feature map into the preset survival risk prediction model, construct a negative log-likelihood loss function based on the proportional risk assumption as the model optimization objective, calculate the cumulative risk function of patients at different time points through the survival prediction layer, introduce a regularization penalty term to constrain the complexity of model parameters, update the model parameters through the gradient descent algorithm, so that the risk ranking predicted by the model tends to be consistent with the actual observed survival time ranking, and output the patient's survival probability distribution curve and risk score. S5. Compare the risk score with the preset risk stratification threshold, and classify the tumor patients into different prognostic risk levels based on the comparison results.

2. The method for prognostic risk stratification assessment of tumor patients based on deep learning and multi-omics data according to claim 1, characterized in that, Step S1 specifically includes: S11. Obtain gene mutation data, transcriptome expression data, deoxyribonucleic acid methylation data, and clinical phenotype data of cancer patients, wherein the clinical phenotype data includes the patient's age, gender, pathological stage, treatment plan, and follow-up survival status. S12. Perform sparsity processing on the gene mutation data, count the frequency of each mutation site in all samples, retain mutation sites with a frequency higher than a preset proportion and remove low-frequency mutation sites, and convert the retained sites into a binary encoding matrix. S13. The transcriptome expression data is logarithmically transformed to compress the dynamic range of the data, and the mean normalization algorithm is used to map the expression level to a distribution space centered on a specific value and with a unit standard deviation by calculating the expression mean and standard deviation of all samples on each gene, thereby eliminating sequencing depth bias and systematic error between different samples. S14. Perform quality control on the deoxyribonucleic acid methylation data, remove probe sites with a missing value ratio higher than a preset upper limit, and use the nearest neighbor algorithm to fill in the remaining missing sites by calculating the average value of a preset number of neighboring samples that are closest to the target sample in the feature space. S15. For continuous numerical variables in clinical phenotype data, the quantile standardization method is used to map them to the quantiles of the preset distribution; for categorical variables, one-hot encoding is used to convert them into orthogonal numerical vectors.

3. The method for prognostic risk stratification assessment of tumor patients based on deep learning and multi-omics data according to claim 1, characterized in that, Step S2 specifically includes: S21. Construct a parallel mining structure consisting of genomic pathway, transcriptomic pathway, methylome pathway and clinical pathway, and send the preprocessed data of each omics into the corresponding residual connection encoder. S22. In each mapping unit of the residual connection encoder, the input feature vector is multiplied and accumulated element by element with the weight matrix to realize the linear transmission of information, and the linear output result is sent to the modified linear unit with unilateral inhibition characteristics to perform nonlinear activation, simulating the excitation threshold mechanism of biological neurons. S23. Introduce a skip connection structure in the feature mapping path to directly transmit the shallow raw biological information to the deep network, thereby alleviating the gradient vanishing phenomenon in the deep neural network when processing biological big data. S24. Based on the dimension and complexity of the input omics data, dynamically adjust the mapping dimension of the high-dimensional feature embedding space to ensure that the hidden layer states output by different paths are comparable in the feature space, and output the specific representation vectors of each omics.

4. The method for prognostic risk stratification assessment of tumor patients based on deep learning and multi-omics data according to claim 1, characterized in that, Step S3 specifically includes: S31. The omics-specific representation vectors are transformed by three independent linear transformations to generate corresponding query components, key components and value components. S32. By calculating the inner product of the query component of the current omics with all other omics key components and processing it with an exponential normalization function, the attention distribution weights that reflect the contribution of different omics to the final prognosis are obtained. S33. The value components of each omics are weighted and summed according to the attention distribution weight to capture the nonlinear interaction between different omics. For methylation feature vectors and transcript feature vectors with biological relevance, the coupling logic between the two is strengthened by contrast calculation. S34. Using a fully connected layer, the weighted features are nonlinearly mapped and dimensionally compressed to integrate the collaborative signals scattered in each path into the global prognostic feature map that characterizes the overall biological state of the patient.

5. The method for prognostic risk stratification assessment of tumor patients based on deep learning and multi-omics data according to claim 1, characterized in that, Step S4 specifically includes: S41. Construct a negative log-likelihood loss function that includes survival time information and censored state information, compare the consistency between the risk score output by the model and the ranking of actual survival time, and calculate the consistency bias. S42. During the model iteration process, parameters are fine-tuned according to the characteristics of different tumor subtypes, and an adaptive learning rate strategy is introduced. In the early stage of training, a larger step size is used to cross the fluctuation region of the non-convex loss function, and in the later stage of training, the learning rate is adjusted near the local optimum by exponential decay. S43. Based on the negative log-likelihood loss function, a regularization penalty term based on the second norm is introduced to prevent overfitting of high-dimensional sparse multi-omics data by constraining the size of the model parameters. S44. Using a parametric model, a probability model of future survival is performed, and a survival probability distribution curve describing the patient's expected survival probability over a preset time span is generated based on the calculated cumulative risk function.

6. The method for prognostic risk stratification assessment of tumor patients based on deep learning and multi-omics data according to claim 1, characterized in that, In step S5, the logic for setting the risk stratification threshold includes: S51. Obtain survival data from a large-scale historical retrospective clinical cohort and analyze the distribution of risk scores in the historical retrospective clinical cohort using statistical methods. S52. Find the survival difference cut-off point that makes the difference in survival curves of patients in different groups reach the maximum statistical significance through an iterative search algorithm, and solidify the survival difference cut-off point as the risk stratification threshold; S53. Compare the real-time calculated patient risk score with the risk stratification threshold to classify the patient into a high-risk group, a medium-risk group, or a low-risk group. S54. Identify the molecular pathways that cause a specific patient to be at a high risk level based on the key contributing factors in the global prognostic feature map, and output the corresponding clinical intervention suggestions.

7. The method for prognostic risk stratification assessment of tumor patients based on deep learning and multi-omics data according to claim 1, characterized in that, When there are missing omics dimensions in the collected multi-omics raw data, the method also includes a robust compensation mechanism: Based on a pre-defined multi-omics association database, generative adversarial logic is used to infer the distribution probability of missing dimensions through existing clinical phenotypic features or transcriptional features, and to generate virtual feature vectors with confidence masks. A feature confidence evaluation unit is introduced inside the residual connection encoder to dynamically adjust the initial weight of the path based on the quality score of the input raw data. When executing the cross-modal attention mechanism, the virtual feature vector or low-quality feature path is perceived, and a penalty term is introduced when calculating the weight distribution to reduce the contribution ratio of the virtual feature vector or low-quality feature path to the final global prognosis feature map. During model training, the strength of the regularization penalty term is dynamically adjusted according to the quality classification of the samples. Weak regularization is applied to high-quality samples, while strong regularization is applied to low-quality samples.

8. The method for prognostic risk stratification assessment of tumor patients based on deep learning and multi-omics data according to claim 1, characterized in that, The method achieves generalization across different tumor subtypes through transfer learning, and the specific logic includes: The multi-path parallel feature extraction network was pre-trained on a pan-cancer dataset to lock the underlying parameters for extracting general biological features. For specific tumor subtypes, the high-level feature mapping units of the multi-path parallel feature extraction network are fine-tuned, and the transformation matrix in the cross-modal attention mechanism is reinitialized to adapt to the differences in omics contributions in specific tumor subtypes. Multiple task sets are constructed using a meta-learning strategy to simulate different tumor scenarios, enabling the survival risk prediction model to learn the ability to adjust its assessment logic when facing new subtypes. A federated learning framework is used to aggregate the statistical features of risk distribution from different centers without exchanging the original data, and generate a globally optimized dynamic threshold surface as a calibration benchmark for the risk stratification threshold.

9. A prognostic risk stratification assessment system for cancer patients based on deep learning and multi-omics data, used to implement the method of any one of claims 1 to 8, characterized in that, include: The data acquisition and preprocessing module is used to acquire genomic data, transcriptome data, methylome data and clinical phenotype data of tumor patients through biological sample detection equipment and clinical information system, and to perform noise reduction, sparsification, normalization and missing value completion operations, and output an aligned multi-omics standard feature set. The feature mapping subsystem, whose input is connected to the output of the data acquisition and preprocessing module, is used to construct a multi-path parallel network. It maps the multi-omics standard feature set to a high-dimensional feature embedding space through the residual connection encoder in each path, and extracts the omics-specific representation vectors using nonlinear activation and discard logic. A cross-modal fusion engine is used to receive the omics-specific representation vectors, execute cross-modal attention allocation logic, calculate the correlation between the query vector and the key vector to determine the weight distribution, and perform nonlinear fusion and fully connected compression on the weighted features to generate a global prognostic feature map. The risk stratification calculation terminal is configured to receive the global prognostic feature map, calculate the patient's risk score using a pre-trained survival risk prediction model, generate a survival probability distribution curve through the survival prediction layer, and output a risk stratification result report in combination with a preset risk stratification threshold.

10. A prognostic risk stratification assessment system for tumor patients based on deep learning and multi-omics data according to claim 9, characterized in that, The system is deployed in a distributed architecture: the data acquisition and preprocessing module is deployed on edge computing nodes and configured to complete format verification and localization desensitization before data is uploaded, and to use multi-threaded concurrency technology to process the standardization tasks of heterogeneous omics data. The feature mapping subsystem includes multiple graphics processor accelerators for mapping the forward propagation operation of the residual connected encoder to a streaming multiprocessor to perform parallel matrix multiplication, and is configured with a hardware monitoring loop to perform node migration for computational tasks. The cross-modal fusion engine is equipped with a memory optimization unit, which reduces video memory usage by using half-precision floating-point numbers to store intermediate feature vectors. The risk stratification calculation terminal performs the aggregation and distribution of model parameters through a distributed parameter server mode, and is equipped with a dynamic threshold calibration unit to adjust the stratification logic in real time according to the distribution of the access data.

Images