Respiratory infectious disease risk early warning method and system

Abstract

The invention relates to the technical field of medical care informatics, in particular to a respiratory infectious disease risk early warning method and system. The method comprises the following steps: acquiring a respiratory tract sample of a mutant virus infected person; the method comprises the following steps: extracting a mutant virus from a respiratory tract sample, carrying out genome sequencing on the mutant virus, positioning starting and ending positions of a spike protein coding gene in a genome by virtue of a biological information comparison method, extracting a spike protein coding sequence, and translating the spike protein coding sequence into an amino acid sequence; inputting the amino acid sequence into an immune escape prediction model to predict the immune escape score of the mutant virus; and performing risk assessment based on the immune escape score in combination with clinical monitoring information, and outputting a respiratory infectious disease risk early warning value. According to the invention, the risk of respiratory infectious diseases can be early warned better.

Description

Technical Field

[0001] The present application relates to the field of medical care informatics technology, and in particular to a respiratory infectious disease risk warning method and system. Background Art

[0002] Outbreaks of respiratory infectious diseases have become a major threat to human health and social stability. The current respiratory infectious disease surveillance system primarily relies on post-event responses, identifying and assessing the threat level of viral variants through epidemiological statistics, clinical observations, and laboratory testing. This delayed monitoring approach often results in public health authorities delaying the implementation of prevention and control measures until respiratory infectious viruses have begun to spread widely, often missing the optimal window for early intervention.

[0003] At the same time, existing threat assessment and risk warning methods for respiratory viruses rely on large-scale population infection data and long-term clinical observations. It usually takes several weeks to accurately judge the immune escape ability and pathogenicity of a new variant. This time lag may lead to irreversible consequences when facing highly transmissible and highly pathogenic viruses.

[0004] Therefore, there are still some problems with the existing respiratory infectious disease risk warning methods. Summary of the Invention

[0005] In order to solve the above technical problems or at least partially solve the above technical problems, the present application provides a respiratory infectious disease risk warning method and system, which can better warn of the risks of respiratory infectious diseases.

[0006] In a first aspect, the present application proposes a respiratory infectious disease risk warning method, which comprises the following steps: Obtain respiratory samples from people infected with the mutant virus; Extracting mutant viruses from respiratory samples, performing genome sequencing on the mutant viruses, locating the start and end positions of the spike protein-encoding gene in the genome using a bioinformatics comparison method, extracting the spike protein-encoding sequence, and translating the spike protein-encoding sequence into an amino acid sequence; Inputting the amino acid sequence into an immune escape prediction model to predict the immune escape score of the mutant virus; Based on the immune escape score and clinical monitoring information, a risk assessment is performed to output a respiratory infectious disease risk warning value.

[0007] Optionally, the immune escape prediction model includes: Multi-channel feature extraction module, including: The ESM-2 protein language model is used to extract the evolutionary semantic features of the amino acid sequence; PSIPRED structure prediction tool, used to extract the one-dimensional structural features of the amino acid sequence; a fusion feature module, connected to the multi-channel feature extraction module, for fusing the evolutionary semantic features and the one-dimensional structural features to obtain a fusion feature; An attention mechanism module, including a Transformer encoder, connected to the fusion feature module, for encoding the fusion feature and learning the interaction relationship between amino acid sites through a multi-head self-attention mechanism to obtain sequence information processed according to attention weighting; The prediction output module includes a fully connected layer connected to the attention mechanism module, and is used to output the immune escape score according to the sequence information processed by attention weighting.

[0008] Optionally, performing risk assessment based on the immune escape score in combination with clinical monitoring information to output a respiratory infectious disease risk warning value includes the following steps: Clinical monitoring information includes chest CT images and clinical symptom information; Obtaining chest CT images and clinical symptom information of patients infected with the mutant virus; Inputting the immune escape score, chest CT image, and clinical symptom information into a multimodal risk assessment model, wherein the multimodal risk assessment model includes an evaluation head and a fusion head; Generate a medical assessment text through the evaluation head to distinguish the relative contributions of viral threat factors and individual susceptibility factors; Outputting, by the fusion head, a standardized score and a dynamic importance weight of each modality based on the medical assessment text, wherein the standardized score includes a predicted immune escape score, a CT severity score, and a clinical severity score; The dynamic importance weight includes the positive importance weight and negative importance weight of each mode; The respiratory infectious disease risk warning value is calculated and output based on the standardized score and the dynamic importance weight.

[0009] Optionally, the multimodal risk assessment model includes: Shared multimodal encoder, including: A visual encoder of the LLaVA-Med model, configured to extract features from the chest CT image; The linear projection layer is connected to the output of the visual encoder and is used to map the visual encoder output to the language model space; The language model of the LLaVA-Med model is used to directly receive text input and is connected to the output of the linear projection layer, so as to process the projected visual features and the immune escape score and clinical symptom information, and output a fused multimodal feature representation; The evaluation head is connected to the output of the language model and includes: Transformer decoder, used to process multimodal feature representation; A medical attribution analysis unit is connected to the output of the Transformer decoder and outputs a medical assessment text using autoregressive generation. The medical assessment text includes a virus threat text analysis, an imaging text analysis, and an individual susceptibility text analysis. The fusion head is connected to the output of the language model and the output of the evaluation head, including: a text encoder, connected to an output terminal of the evaluation head, for encoding the medical evaluation text into text features; The feature fusion module is connected to the output of the language model and the text encoder to fuse the original multimodal features and the evaluation text features; a score prediction unit, comprising three parallel MLP structures, connected to the output end of the feature fusion module to respectively output the predicted immune escape score, CT severity score, and clinical severity score; The weight calculation unit consists of three parallel MLP structures connected to the output of the feature fusion module to calculate the dynamic importance weight of each modality.

[0010] Optionally, the training step of the multimodal risk assessment model includes: freezing parameters of the shared multimodal encoder; Train the evaluation head: Construct a training dataset, including chest CT images, clinical symptom information, and immune escape scores of patients infected with the mutant virus, as well as a corresponding medical assessment annotation text dataset annotated by medical experts. The medical assessment annotation text includes virus threat annotation text analysis, imaging annotation text analysis, and individual susceptibility annotation text analysis; Freeze all parameters of the fusion head; Inputting the chest CT image, clinical symptom information, and immune escape score into the shared multimodal encoder to obtain a multimodal feature representation; Using the multimodal feature representation to train the Transformer decoder and the medical attribution analysis unit; The cross-entropy loss function is used, and the medical evaluation text annotated by experts is used as the supervision signal to optimize the accuracy and medical professionalism of the generated medical evaluation text; The evaluation head and the fusion head are jointly trained: Construct an expert-annotated standardized score and dynamic importance weight dataset, including the annotated values ​​of the predicted immune escape score, CT severity score, clinical severity score of each modality, and the corresponding dynamic importance weight annotated values; A multi-task loss function is used for optimization, including the cross-entropy loss for medical assessment text generation, the mean squared error loss for standardized score prediction, and the mean squared error loss for dynamic weight calculation. All parameters of the evaluation head and the fusion head are trained simultaneously to obtain a multimodal risk assessment model.

[0011] Optionally, individuals infected with the mutant virus are screened by the following steps: All respiratory samples from pneumonia patients are tested using PCR with at least two targets: Conserved region targets are used to detect regions in the genome of pre-set respiratory viruses that are less susceptible to mutation to confirm viral infection; Characteristic region targets are used to detect gene sequence segments unique to the preset respiratory virus genome; Samples with positive results for the conservative region target detection and negative results for the characteristic region target detection were screened as characteristic target shedding signal samples; The patient corresponding to the characteristic target shedding signal sample is marked as infected with the mutant virus.

[0012] In the second aspect, the present application proposes a respiratory infectious disease risk warning system, which includes a processor and a memory, wherein the memory stores at least one instruction, at least one program, code set or instruction set, and the at least one instruction, at least one program, code set or instruction set is loaded and executed by the processor to implement the respiratory infectious disease risk warning method as described in any one of the first aspects.

[0013] The technical solution provided by this application has the following advantages over the existing technology: One of its beneficial effects and its working principle is: Because traditional respiratory infectious disease monitoring methods rely on epidemiological statistics and clinical observations, it takes weeks to determine the threat level of emerging mutant viruses through the accumulation of a large amount of clinical data. As a result, public health departments are in a passive response state when facing mutant viruses, easily missing the best early intervention window and unable to take effective prevention and control measures in time. There is a technical problem of the rapid spread of sudden infectious diseases causing social harm.

[0014] The present application provides a method for early warning of respiratory infectious disease risks, which obtains respiratory samples from people infected with mutant viruses and performs genome sequencing, uses bioinformatics comparison methods to locate the spike protein encoding gene and translates it into an amino acid sequence, inputs the amino acid sequence into an immune escape prediction model, and uses the association patterns between a large number of known virus variant sequences and immune escape abilities learned by the model to quantitatively predict the immune escape abilities of newly emerging mutant viruses. This is then combined with clinical monitoring information for a comprehensive risk assessment, thereby achieving a forward-looking risk quantification analysis based on viral genetic characteristics.

[0015] Therefore, the technical effect of shortening the time for virus risk identification from several weeks required by traditional methods to within hours after obtaining the genetic sequence has been achieved, realizing the transition from passive monitoring to active early warning, and reducing the health hazards and economic losses caused to society by sudden infectious disease outbreaks.

[0016] The second beneficial effect and its working principle are: This application combines the evolutionary semantic features extracted by the ESM-2 protein language model with the one-dimensional structural features extracted by the PSIPRED structure prediction tool. This not only fully utilizes the semantic information related to the protein sequence context learned by ESM-2, but also supplements the rapidly acquired structural feature information, thus avoiding the huge computational cost brought about by direct three-dimensional structure prediction. At the same time, the multi-head self-attention mechanism of the Transformer encoder is used to learn the complex interaction relationships between amino acid sites. It can automatically identify and assign higher attention weights to key mutation sites that contribute most to immune escape, thereby improving the accuracy of immune escape prediction. It can also point out the specific mutation sites that cause immune escape through the final attention weight, providing clear guidance for subsequent experimental verification and vaccine development.

[0017] The third beneficial effect and its working principle are: Risk warning values ​​derived solely from immune escape prediction scores cannot directly guide public health decision-making. This is because clinical practice requires comprehensive consideration of multidimensional information such as the predicted immune escape score, medical imaging findings, and clinical symptom descriptions. However, the importance of different information sources in virus threat assessment varies dynamically with the specific individual case.

[0018] If we rely solely on the subjective judgment of clinicians, cognitive bias and misjudgment are likely to occur due to the limited number of cases they come into contact with and the limitations of their experience.

[0019] If a fixed weight fusion method is used, it will be impossible to accurately distinguish the relative contributions of the virus threat itself and individual susceptibility factors. It is easy to be misled by the severe symptoms caused by individual vulnerability and ignore the real virus threat signals.

[0020] Therefore, the existing technology has the technical problem of being unable to intelligently and dynamically adjust the weights of each information source to output an accurate comprehensive risk assessment.

[0021] This application constructs a multimodal risk assessment model with an evaluation-fusion dual-head architecture based on the LLaVA-Med multimodal large model. First, the evaluation head is used to generate professional medical evaluation text based on multimodal medical information. By generating structured medical evaluation text, the rich details of the multimodal input information are fully retained and utilized, avoiding the information loss problem caused by the traditional method of directly mapping multimodal features into numerical scores. The model enables the model to perform medical reasoning analysis like a clinician, and distinguish the relative contribution of viral threat factors and individual susceptibility factors in the current clinical manifestations. The fusion head then outputs the standardized score and dynamic importance weight of each modality based on this professional medical evaluation text that includes a complete reasoning process. A positive dynamic importance weight indicates that the modality supports the current threat assessment, while a negative dynamic importance weight indicates the opposite. Finally, an accurate comprehensive risk warning value is calculated based on the standardized score and dynamic importance weight of each modality.

[0022] Therefore, the respiratory infectious disease risk warning method provided in this application can better warn of the risk of respiratory infectious diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 A flow chart of the respiratory infectious disease risk warning method provided in an embodiment of the present application. DETAILED DESCRIPTION

[0024] The technical solution in this application will be described below with reference to the accompanying drawings.

[0025] The following description sets forth many specific details to facilitate a thorough understanding of the present application. However, the present application may also be implemented in other ways than those described herein. It is apparent that the embodiments described in the specification are only some of the embodiments of the present application, not all of them. It should be noted that the embodiments of the present application and the features therein may be combined with each other unless there is a conflict.

[0026] First, this application proposes a respiratory infectious disease risk warning method, referring to Figure 1 As shown, the respiratory infectious disease risk warning method includes the following steps: S1: Obtain respiratory samples from patients infected with the mutant virus; Specifically, people infected with the mutant virus are screened through the following steps: All respiratory samples from pneumonia patients are tested using PCR with at least two targets: Conserved region targets are used to detect regions in the genome of pre-set respiratory viruses that are less susceptible to mutation to confirm viral infection; Characteristic region targets are used to detect gene sequence segments unique to the preset respiratory virus genome; Samples with positive results for the conservative region target detection and negative results for the characteristic region target detection were screened as characteristic target shedding signal samples; The patient corresponding to the characteristic target shedding signal sample is marked as infected with the mutant virus.

[0027] S2: Extract mutant viruses from respiratory samples, perform genome sequencing on the mutant viruses, locate the start and end positions of the spike protein encoding gene in the genome through bioinformatics comparison methods, extract the spike protein encoding sequence, and translate the spike protein encoding sequence into an amino acid sequence.

[0028] Specifically, viral nucleic acid is extracted from respiratory samples of patients infected with the mutant virus screened out in step S1 using a viral RNA extraction kit, the RNA is converted into cDNA through a reverse transcription reaction, and then the entire viral genome is sequenced using the Illumina high-throughput sequencing platform.

[0029] Specifically, the spike protein encoding gene was precisely located using bioinformatics alignment methods. The following steps were used: The viral genome sequence obtained by sequencing was compared with the reference sequence in the NCBI viral genome database by BLAST to determine the specific type and lineage of the virus; Based on the alignment results, the start and end positions of the spike protein-encoding gene in the newly sequenced genome were located using the corresponding reference genome annotation information using the MUSCLE sequence alignment algorithm; For example, for the new coronavirus, the spike protein encoding gene is usually located in the 21563-25384 position range of the genome, while for other respiratory viruses, it is located according to their specific genomic structure.

[0030] Based on the positioning results, the complete coding sequence of the spike protein is accurately extracted from the complete genome sequence.

[0031] The spike protein coding sequence is translated according to the standard genetic code table to obtain the corresponding amino acid sequence.

[0032] The translated amino acid sequence was saved in the standard FASTA format.

[0033] S3: inputting the amino acid sequence into an immune escape prediction model to predict the immune escape score of the mutant virus; Specifically, the immune escape prediction model includes: Multi-channel feature extraction module, including: The spike protein amino acid sequence obtained in step S2 is simultaneously input into the following two parallel feature extraction channels: The ESM-2 protein language model is used to extract the evolutionary semantic features of the amino acid sequence.

[0034] Specifically, the pre-trained ESM-2-1.3B model is used. This model has been unsupervised pre-trained on a large number of protein sequences and can extract deep evolutionary semantic features of amino acid sequences.

[0035] The input spike protein amino acid sequence is first tokenized and then input into the ESM-2 model, which consists of a 33-layer Transformer encoder and can output a 1280-dimensional deep evolutionary semantic feature vector for each amino acid site.

[0036] PSIPRED structure prediction tool is used to extract the one-dimensional structural features of the amino acid sequence.

[0037] Specifically, PSIPRED is a protein secondary structure prediction algorithm based on statistics and machine learning. It directly takes the amino acid sequence as input and predicts the secondary structure of the entire protein sequence by analyzing the local and global characteristic patterns of the amino acid sequence.

[0038] The spike protein amino acid sequence obtained in step S2 is input into the PSIPRED tool to predict the secondary structure distribution of the entire sequence, including the distribution patterns of structural types such as α-helices, β-sheets, and random coils within the sequence. The prediction results are represented using a 3D one-hot encoding, where each amino acid sequence corresponds to a 3D vector ([1, 0, 0] for α-helices, [0, 1, 0] for β-sheets, and [0, 0, 1] for random coils). This provides the model with quick one-dimensional structural features without the computationally expensive 3D structure prediction.

[0039] The fusion feature module is connected to the multi-channel feature extraction module and is used to fuse the evolutionary semantic features and the one-dimensional structural features to obtain fusion features.

[0040] Specifically, the 1280-dimensional semantic features extracted by ESM-2 and the 3D and 1D structural features predicted by PSIPRED are simply concatenated into a vector matrix to form a 1283-dimensional fused feature vector.

[0041] The attention mechanism module includes a Transformer encoder connected to the fusion feature module, which is used to encode the fusion features and learn the interaction relationship between amino acid sites through a multi-head self-attention mechanism to obtain sequence information processed according to attention weighting.

[0042] Specifically, the Transformer encoder adopts a Transformer encoder architecture with 6 encoder layers.

[0043] The 1283-dimensional fused feature vector sequence received from the fusion feature module is first position-encoded using a sinusoidal position encoding method to obtain a position-encoded feature sequence.

[0044] The position-encoded feature sequence is input into the Transformer encoder, and each encoder layer contains a multi-head self-attention sublayer and a feedforward neural network sublayer.

[0045] In the multi-head self-attention sub-layer, 8 attention heads are used to process the feature sequence in parallel.

[0046] After being processed by the 6-layer encoder, the encoded feature sequence and attention weight matrix are output.

[0047] The encoded feature sequence is weighted averaged according to the attention weight matrix to obtain sequence information processed according to the attention weight.

[0048] The prediction output module includes a fully connected layer connected to the attention mechanism module, and is used to output the immune escape score according to the sequence information processed by attention weighting.

[0049] Specifically, the prediction output module: The fully connected layer inputs the sequence information processed according to the attention weighting into a fully connected neural network consisting of two layers. The first layer is a 512-dimensional hidden layer (using ReLU activation function and Dropout regularization for feature transformation), and the second layer is the output layer (using Sigmoid activation function).

[0050] The final output is an immunity escape score in the range of [0, 1], where 0 indicates no escape capability and 1 indicates complete escape.

[0051] The training process of the immune escape prediction model includes the following steps: Training data for immune escape prediction models was collected from the publicly available Deep Mutation Scan (DMS) dataset, which contains the results of large-scale mutation experiments on the coronavirus spike protein. Specifically, these datasets experimentally measure the effects of thousands of single-point mutations or combinations of mutations on the virus's ability to evade neutralizing antibodies. Each experimental entry contains the mutated spike protein amino acid sequence and the corresponding experimentally measured neutralization escape score.

[0052] It should be noted that the target object in the embodiment of the present application is a coronavirus respiratory virus, so only the DMS data set of the coronavirus is collected in the embodiment of the present application. In other embodiments, those skilled in the art can expand the data set to deal with different viruses.

[0053] The amino acid sequence of the spike protein of the mutant virus is used as the input of the model, and the experimentally measured neutralizing antibody escape score (normalized to the range of [0, 1], where 0 indicates no escape ability and 1 indicates complete escape of neutralizing antibodies) is used as the label for supervised learning training to obtain the immune escape prediction model.

[0054] S4: Perform risk assessment based on the immune escape score combined with clinical monitoring information, and output a respiratory infectious disease risk warning value.

[0055] Specifically, the following steps are included: Clinical monitoring information includes chest CT images and clinical symptom information; Obtaining chest CT images and clinical symptom information of patients infected with the mutant virus; Inputting the immune escape score, chest CT image, and clinical symptom information into a multimodal risk assessment model, wherein the multimodal risk assessment model includes an evaluation head and a fusion head; Generate a medical assessment text through the evaluation head to distinguish the relative contributions of viral threat factors and individual susceptibility factors; Outputting, by the fusion head, a standardized score and a dynamic importance weight of each modality based on the medical assessment text, wherein the standardized score includes a predicted immune escape score, a CT severity score, and a clinical severity score; The dynamic importance weight includes the positive importance weight and negative importance weight of each mode; The respiratory infectious disease risk warning value is calculated and output based on the standardized score and the dynamic importance weight.

[0056] Specifically, the multimodal risk assessment model includes: Shared multimodal encoder, including: A visual encoder of the LLaVA-Med model, configured to extract features from the chest CT image; The linear projection layer uses a fully connected layer structure and is connected to the output of the visual encoder to map the visual encoder output to the language model space. The language model of the LLaVA-Med model is used to directly receive text input and is connected to the output of the linear projection layer, so as to process the projected visual features and the immune escape score and clinical symptom information, and output a fused multimodal feature representation; The evaluation head is connected to the output of the language model and includes: Transformer decoder, which uses a 2-layer Transformer decoder structure to process multimodal feature representation; The medical attribution analysis unit is connected to the output of the Transformer decoder and uses a standard language model to generate a header structure. It uses autoregressive generation to output medical assessment text token by token. The medical assessment text includes virus threat text analysis (assessing the pathogenicity of the virus itself), imaging text analysis (analyzing the severity and typicality of CT manifestations), and individual susceptibility text analysis (assessing the patient's vulnerability factors). The generated length is controlled at 300-500 tokens. The fusion head is connected to the output of the language model and the output of the evaluation head, including: A text encoder, connected to the output of the evaluation head, adopts the same encoding architecture as the language model and is used to re-encode the medical evaluation text generated by the medical attribution analysis unit into a text feature vector with the same dimension as the original multimodal feature; The feature fusion module is connected to the output of the language model and the text encoder at the same time. It adopts a combined structure of LayerNorm normalization layer, learnable position encoding layer and multi-head attention layer to deeply fuse the original multimodal features from the language model and the evaluated text features from the text encoder, and output an enhanced fused feature representation; a score prediction unit comprising three parallel MLP structures, each of which consists of a single-layer linear transformation plus a Sigmoid activation function and is connected to the output of the feature fusion module to output the predicted immune escape score, CT severity score, and clinical severity score, respectively; The weight calculation unit contains three parallel MLP structures. Each MLP structure consists of a single-layer linear transformation plus a Tanh activation function. It is connected to the output of the feature fusion module and is used to calculate the dynamic importance weight of each modality.

[0057] Specifically, the training steps of the multimodal risk assessment model include: freezing parameters of the shared multimodal encoder; Train the evaluation head: Construct an evaluation head training dataset, including complete multimodal data of real mutant virus infection cases: Chest CT images (512×512 pixel DICOM format after standardized preprocessing); structured clinical symptom information (including patient age, underlying diseases, symptom description, and physical signs); immune escape score.

[0058] A medical assessment annotation text dataset jointly annotated by a team of medical experts (including infectious disease physicians, radiologists, and virology experts). The medical assessment annotation text adopts a unified structured template, including: virus threat annotation text analysis (threat assessment based on immune escape score and virological knowledge); imaging annotation text analysis (lesion severity and typicality analysis based on CT manifestations); individual susceptibility annotation text analysis (vulnerability assessment based on the patient's baseline status).

[0059] Among them, the multimodal data of real mutant virus infection cases can be obtained by collecting historical case data of known mutant strains such as Alpha, Beta, Gamma, Delta and Omicron, but the immune escape score corresponding to each known mutant strain needs to be artificially set according to the transmission speed and range of the prevalent strain at that time.

[0060] Freeze all parameters of the fusion head; Inputting the chest CT image, clinical symptom information, and immune escape score into the shared multimodal encoder to obtain a multimodal feature representation; Using the multimodal feature representation to train the Transformer decoder and the medical attribution analysis unit; The cross-entropy loss function is adopted, and the medical evaluation annotation text annotated by experts is used as the supervision signal to optimize the generation accuracy and medical professionalism of the medical evaluation text.

[0061] The evaluation head and the fusion head are jointly trained: Construct a dataset of standardized scores and dynamic importance weights annotated by experts, including the standardized scores of each modality and the corresponding dynamic importance weight annotation values; Unfreeze fusion head parameters; The parameters of the evaluation head and the fusion head are optimized simultaneously using a multi-task loss function, including the cross-entropy loss for medical assessment text generation, the mean squared error loss for standardized score prediction, and the mean squared error loss for dynamic weight calculation. All parameters of the evaluation head and the fusion head are trained simultaneously to obtain a multimodal risk assessment model.

[0062] Specifically, the total loss function consists of three components: Cross-entropy loss for medical assessment text generation (used to ensure that the evaluation head continues to generate high-quality medical text); Mean squared error loss for normalized score predictions (used to optimize the accuracy of the score prediction unit); Mean squared error loss for dynamic weight calculation (used to optimize the rationality of weight calculation units); The total loss function is = α· + β· + γ· ; Wherein, α, β, and γ are adjustable weight coefficients. In the embodiment of the present application, α is set to 0.2, β is set to 0.4, and γ is set to 0.4.

[0063] Specifically, based on an example, the detailed process of multimodal risk assessment is as follows: Case background: Patient Li, a 24-year-old male, had no underlying diseases; Symptoms include high fever of 39.2°C, severe dry cough, and difficulty breathing; The immune escape score outputted in step S3 is 0.92.

[0064] The patient's corresponding data: Chest CT images (showing diffuse ground-glass opacities with consolidation in both lungs, with the lesion area >70%); Clinical symptom information (28 years old, male; no underlying disease; high fever of 39.2°C for 3 days; severe dry cough and dyspnea; blood oxygen saturation of 86%) and immune escape score (0.92) were input into the shared multimodal encoder; Shared multimodal encoder processing stage: The LLaVA-Med visual encoder extracts high-dimensional visual features from CT images, including medical information such as extensive bilateral lung lesions, diffuse ground-glass opacities, large-scale consolidation, and severe lesion severity. These visual features are mapped to the language model space via a linear projection layer. These features are then fused with textual input features such as those related to extremely high immune escape, young and healthy individuals, and severe symptoms within the LLaVA-Med language model, resulting in a unified multimodal feature representation.

[0065] Evaluation of medical reasoning stage: The multimodal features are processed by a 2-layer Transformer decoder; The medical attribution analysis unit generates medical evaluation text: Virus threat text analysis: This variant has a very strong immune escape ability (0.92), has the ability to completely break through existing immune protection, and shows significant pathogenic potential in healthy people.

[0066] Imaging text analysis: Chest CT showed typical manifestations of acute severe viral pneumonia, with diffuse lesions and consolidation in both lungs. The lesions were extensive and the imaging severity was severe.

[0067] Individual susceptibility text analysis: The patient is a 28-year-old healthy male without any underlying diseases. His current severe symptoms are significantly beyond expectations among healthy young people.

[0068] Fusion head dynamic weight calculation stage: The medical assessment text is processed by the text encoder and then deeply fused with the original multimodal features in the feature fusion module.

[0069] Score prediction unit output: =0.92 (confirmed extremely high escape); =0.88 (severe radiographic changes); =0.85 (severe clinical symptoms).

[0070] Weight calculation unit output: =+0.95 (strongly supports high escape judgment); =+0.85 (severe radiographic changes strongly support threat assessment); =+0.90 (severe clinical manifestations strongly support the judgment of threat escalation).

[0071] Final risk warning value calculation stage: This outputs a high-risk warning value close to the maximum value of 3, indicating a strong virus risk and the need to initiate high-level prevention and control measures.

[0072] In summary, the technical solution provided by the embodiments of the present application has the following advantages compared with the prior art: One of its beneficial effects and its working principle is: Because traditional respiratory infectious disease monitoring methods rely on epidemiological statistics and clinical observations, it takes weeks to determine the threat level of emerging mutant viruses through the accumulation of a large amount of clinical data. As a result, public health departments are in a passive response state when facing mutant viruses, easily missing the best early intervention window and unable to take effective prevention and control measures in time. There are technical problems such as the rapid spread of sudden infectious diseases and the resulting social harm.

[0073] The present application provides a method for early warning of respiratory infectious disease risks, which obtains respiratory samples from people infected with mutant viruses and performs genome sequencing, uses bioinformatics comparison methods to locate the spike protein encoding gene and translates it into an amino acid sequence, inputs the amino acid sequence into an immune escape prediction model, and uses the association patterns between a large number of known virus variant sequences and immune escape abilities learned by the model to quantitatively predict the immune escape abilities of newly emerging mutant viruses. This is then combined with clinical monitoring information for a comprehensive risk assessment, thereby achieving a forward-looking risk quantification analysis based on viral genetic characteristics.

[0074] Therefore, the technical effect of shortening the time for virus risk identification from several weeks required by traditional methods to within hours after obtaining the genetic sequence has been achieved, realizing the transition from passive monitoring to active early warning, and reducing the health hazards and economic losses caused to society by sudden infectious disease outbreaks.

[0075] The second beneficial effect and its working principle are: This application combines the evolutionary semantic features extracted by the ESM-2 protein language model with the one-dimensional structural features extracted by the PSIPRED structure prediction tool. This not only fully utilizes the semantic information related to the protein sequence context learned by ESM-2, but also supplements the rapidly acquired structural feature information, thus avoiding the huge computational cost brought about by direct three-dimensional structure prediction. At the same time, the multi-head self-attention mechanism of the Transformer encoder is used to learn the complex interaction relationships between amino acid sites. It can automatically identify and assign higher attention weights to key mutation sites that contribute most to immune escape, thereby improving the accuracy of immune escape prediction. It can also point out the specific mutation sites that cause immune escape through attention weights, providing clear guidance for subsequent experimental verification and vaccine development.

[0076] The third beneficial effect and its working principle are: Risk warning values ​​derived solely from immune escape prediction scores cannot directly guide public health decision-making. This is because clinical practice requires comprehensive consideration of multidimensional information, including predicted immune escape scores, medical imaging findings, and clinical symptom descriptions. However, the importance of different information sources in virus threat assessment varies dynamically with specific cases.

[0077] If we rely solely on the subjective judgment of clinicians, cognitive bias and misjudgment are likely to occur due to the limited number of cases they come into contact with and the limitations of their experience.

[0078] If a fixed weight fusion method is used, it will be impossible to accurately distinguish the relative contributions of the virus threat itself and individual susceptibility factors. It is easy to be misled by the severe symptoms caused by individual vulnerability and ignore the real virus threat signals.

[0079] Therefore, the existing technology has the technical problem of being unable to intelligently and dynamically adjust the weights of each information source to output an accurate comprehensive risk assessment.

[0080] This application constructs a multimodal risk assessment model with an evaluation-fusion dual-head architecture based on the LLaVA-Med multimodal large model. First, the evaluation head is used to generate professional medical evaluation text based on multimodal medical information. By generating structured medical evaluation text, the rich details of the multimodal input information are fully retained and utilized, avoiding the information loss problem caused by the traditional method of directly mapping multimodal features into numerical scores. The model enables the model to perform medical reasoning analysis like a clinician, and distinguish the relative contribution of viral threat factors and individual susceptibility factors in the current clinical manifestations. The fusion head then outputs the standardized score and dynamic importance weight of each modality based on this professional medical evaluation text that includes a complete reasoning process. A positive dynamic importance weight indicates that the modality supports the current threat assessment, while a negative dynamic importance weight indicates the opposite. Finally, an accurate comprehensive risk warning value is calculated based on the standardized score and dynamic importance weight.

[0081] Therefore, the respiratory infectious disease risk warning method provided in this application can better warn of the risk of respiratory infectious diseases.

[0082] In the second aspect, the present application proposes a respiratory infectious disease risk warning system, which includes a processor and a memory, wherein the memory stores at least one instruction, at least one program, code set or instruction set, and the at least one instruction, at least one program, code set or instruction set is loaded and executed by the processor to implement the respiratory infectious disease risk warning method as described in any one of the first aspects.

[0083] It should be noted that, in this document, relational terms such as "first" and "second" are used solely to distinguish one entity or operation from another, and do not necessarily require or imply any actual relationship or order between these entities or operations. Furthermore, the terms "include," "comprises," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or device comprising a series of elements includes not only those elements but also other elements not explicitly listed, or includes elements inherent to such process, method, article, or device. Without further limitation, an element defined by the phrase "comprising a..." does not preclude the presence of additional identical elements in the process, method, article, or device comprising the element. Furthermore, in the description of the embodiments of this application, unless otherwise specified, " / " represents or. For example, A / B can represent either A or B. "And / or" herein is merely a description of an associative relationship between associated objects, indicating that three relationships can exist. For example, "A and / or B" can represent: A exists alone, A and B exist simultaneously, or B exists alone. Furthermore, in the description of the embodiments of the present application, “plurality” refers to two or more than two.

[0084] The foregoing description is intended only to provide specific embodiments of the present application, which will enable those skilled in the art to understand and implement the present application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application is not limited to the embodiments described herein, but is intended to be construed in the broadest manner consistent with the principles and novel features disclosed herein.

Claims

1. A respiratory infectious disease risk warning method, characterized in that: The respiratory infectious disease risk early warning method comprises the following steps: Obtain respiratory samples from people infected with the mutant virus; Extracting mutant viruses from respiratory samples, performing genome sequencing on the mutant viruses, locating the start and end positions of the spike protein-encoding gene in the genome using a bioinformatics comparison method, extracting the spike protein-encoding sequence, and translating the spike protein-encoding sequence into an amino acid sequence; Inputting the amino acid sequence into an immune escape prediction model to predict the immune escape score of the mutant virus; Based on the immune escape score and clinical monitoring information, a risk assessment is performed to output a respiratory infectious disease risk warning value.

2. The respiratory infectious disease risk warning method according to claim 1, characterized in that: The immune escape prediction model includes: Multi-channel feature extraction module, including: The ESM-2 protein language model is used to extract the evolutionary semantic features of the amino acid sequence; PSIPRED structure prediction tool, used to extract the one-dimensional structural features of the amino acid sequence; a fusion feature module, connected to the multi-channel feature extraction module, for fusing the evolutionary semantic features and the one-dimensional structural features to obtain a fusion feature; An attention mechanism module, including a Transformer encoder, connected to the fusion feature module, for encoding the fusion feature and learning the interaction relationship between amino acid sites through a multi-head self-attention mechanism to obtain sequence information processed according to attention weighting; The prediction output module includes a fully connected layer connected to the attention mechanism module, and is used to output the immune escape score according to the sequence information processed by attention weighting.

3. The respiratory infectious disease risk warning method according to claim 1, characterized in that: Performing risk assessment based on the immune escape score combined with clinical monitoring information to output a respiratory infectious disease risk warning value includes the following steps: Clinical monitoring information includes chest CT images and clinical symptom information; Obtaining chest CT images and clinical symptom information of patients infected with the mutant virus; Inputting the immune escape score, chest CT image, and clinical symptom information into a multimodal risk assessment model, wherein the multimodal risk assessment model includes an evaluation head and a fusion head; Generate a medical assessment text through the evaluation head to distinguish the relative contributions of viral threat factors and individual susceptibility factors; Outputting, by the fusion head, a standardized score and a dynamic importance weight of each modality based on the medical assessment text, wherein the standardized score includes a predicted immune escape score, a CT severity score, and a clinical severity score; The dynamic importance weight includes the positive importance weight and negative importance weight of each mode; The respiratory infectious disease risk warning value is calculated and output based on the standardized score and the dynamic importance weight.

4. The respiratory infectious disease risk warning method according to claim 1, characterized in that: The multimodal risk assessment model includes: Shared multimodal encoder, including: A visual encoder of the LLaVA-Med model, configured to extract features from the chest CT image; The linear projection layer is connected to the output of the visual encoder and is used to map the visual encoder output to the language model space; The language model of the LLaVA-Med model is used to directly receive text input and is connected to the output of the linear projection layer, so as to process the projected visual features and the immune escape score and clinical symptom information, and output a fused multimodal feature representation; The evaluation head is connected to the output of the language model and includes: Transformer decoder, used to process multimodal feature representation; A medical attribution analysis unit is connected to the output of the Transformer decoder and outputs a medical assessment text using autoregressive generation. The medical assessment text includes a virus threat text analysis, an imaging text analysis, and an individual susceptibility text analysis. The fusion head is connected to the output of the language model and the output of the evaluation head, including: a text encoder, connected to an output terminal of the evaluation head, for encoding the medical evaluation text into text features; The feature fusion module is connected to the output of the language model and the text encoder to fuse the original multimodal features and the evaluation text features; a score prediction unit, comprising three parallel MLP structures, connected to the output end of the feature fusion module to respectively output the predicted immune escape score, CT severity score, and clinical severity score; The weight calculation unit consists of three parallel MLP structures connected to the output of the feature fusion module to calculate the dynamic importance weight of each modality.

5. The respiratory infectious disease risk warning method according to claim 4, characterized in that: The training steps of the multimodal risk assessment model include: freezing parameters of the shared multimodal encoder; Train the evaluation head: Construct a training dataset, including chest CT images, clinical symptom information, and immune escape scores of patients infected with the mutant virus, as well as a corresponding medical assessment annotation text dataset annotated by medical experts. The medical assessment annotation text includes virus threat annotation text analysis, imaging annotation text analysis, and individual susceptibility annotation text analysis; Freeze all parameters of the fusion head; Inputting the chest CT image, clinical symptom information, and immune escape score into the shared multimodal encoder to obtain a multimodal feature representation; Using the multimodal feature representation to train the Transformer decoder and the medical attribution analysis unit; The cross-entropy loss function is used, and the medical evaluation text annotated by experts is used as the supervision signal to optimize the accuracy and medical professionalism of the generated medical evaluation text; The evaluation head and the fusion head are jointly trained: Construct an expert-annotated standardized score and dynamic importance weight dataset, including the annotated values ​​of the predicted immune escape score, CT severity score, clinical severity score of each modality, and the corresponding dynamic importance weight annotated values; A multi-task loss function is used for optimization, including the cross-entropy loss for medical assessment text generation, the mean squared error loss for standardized score prediction, and the mean squared error loss for dynamic weight calculation. All parameters of the evaluation head and the fusion head are trained simultaneously to obtain a multimodal risk assessment model.

6. The respiratory infectious disease risk warning method according to claim 1, characterized in that: People infected with the mutant virus are screened through the following steps: All respiratory samples from pneumonia patients are tested using PCR with at least two targets: Conserved region targets are used to detect regions in the genome of pre-set respiratory viruses that are less susceptible to mutation to confirm viral infection; Characteristic region targets are used to detect gene sequence segments unique to the preset respiratory virus genome; Samples with positive results for the conservative region target detection and negative results for the characteristic region target detection were screened as characteristic target shedding signal samples; The patient corresponding to the characteristic target shedding signal sample is marked as infected with the mutant virus.

7. Respiratory infectious disease risk warning system, characterized by: The system includes a processor and a memory, wherein the memory stores at least one instruction, at least one program, code set or instruction set, and the at least one instruction, at least one program, code set or instruction set is loaded and executed by the processor to implement the respiratory infectious disease risk warning method as described in any one of claims 1-6.

Images