Animal vaccination risk prediction method and system based on cross-modal large model

Abstract

The invention provides an animal vaccination risk prediction method and system based on a cross-modal large model, and relates to the technical field of vaccination risk prediction. Comprising a multi-modal data acquisition and integration module, a cross-modal feature extraction and semantic alignment module, a pre-training model construction and transfer learning module, a risk prediction and decision support module and a model verification and iterative optimization module. Misjudgment caused by same surface symptoms can be effectively avoided, in animal epidemic disease diagnosis, similar symptoms may be caused by different viruses, a traditional method is easily interfered by surface phenomena, and the optimized method can be used for diagnosis of animal epidemic diseases by constructing a causal diagram, analyzing real causes of symptoms and stripping confounding factors such as environment and physiological status. Through the means, the model can more accurately distinguish virus types, the misjudgment risk is remarkably reduced, and a reliable basis is provided for a subsequent vaccination strategy.

Description

Technical Field

[0001] The present invention belongs to the technical field of vaccination risk prediction, and more specifically, relates to an animal vaccination risk prediction method and system based on a cross-modal large model. Background Art

[0002] The animal vaccination risk prediction method based on a cross-modal large model is an innovative solution that integrates multi-source data, such as text, images, sensor signals, etc., and uses artificial intelligence technology to accurately predict the risks of animals after vaccination. This method integrates multi-modal data such as animal physiological characteristics, vaccine characteristics, and environmental factors to construct a unified semantic representation space, thereby realizing dynamic evaluation of risk indicators such as adverse reactions and immune effects after vaccination. The animal vaccination risk prediction method based on a cross-modal large model provides an intelligent solution for animal disease prevention and control by integrating multi-source data, constructing a unified semantic space, and realizing dynamic risk assessment. Its core advantage lies in breaking through the limitations of single modal analysis and realizing a paradigm shift from passive disinfection to active prediction. However, the implementation of the technology still needs to solve challenges such as data quality, model interpretability, and ethical norms. In the future, interdisciplinary cooperation and standardization construction are needed to promote the widespread application of this method on a global scale.

[0003] However, during the implementation of the above technical solution, at least the following technical problems were found:

[0004] In the risk prediction of animal vaccination, if two viruses show the same symptoms, the model may misjudge due to the similar surface features in the multimodal data. For example, both porcine reproductive and respiratory syndrome virus and swine influenza virus may cause respiratory symptoms such as fever, cough, and difficulty breathing, and abnormal white blood cell counts may appear in blood tests. Lung CT images in visual data may also show similar ground-glass lesions. If the model only relies on shallow multimodal feature matching such as symptoms and physiological indicators, and fails to deeply explore the differences in viral genomes, specific antibodies of host immune responses, or epidemiological backgrounds, it may classify individuals carrying different viruses into the same risk category. If a vaccine for virus A is used for animals infected with virus B, it may not stimulate effective immunity due to antigen mismatch, resulting in the disease still occurring after vaccination, or even accelerated virus mutation due to immune pressure.

[0005] In view of this, the existing structure and deficiencies are studied and improved, and an animal vaccination risk prediction method and system based on a cross-modal large model is provided. Summary of the Invention

[0006] In order to solve the above technical problems, the present invention provides an animal vaccination risk prediction method and system based on a cross-modal large model to solve the above problems.

[0007] The animal vaccination risk prediction system based on a cross-modal large model includes a multimodal data acquisition and integration module, a cross-modal feature extraction and semantic alignment module, a pre-training model construction and transfer learning module, a risk prediction and decision support module, and a model verification and iterative optimization module. The details are as follows:

[0008] Multimodal data collection and integration module: comprehensively collects multi-type animal data, including body temperature, body surface characteristics, vaccine instructions, and viral gene information, and processes the raw data;

[0009] Cross-modal feature extraction and semantic alignment module: This module extracts semantic, visual, and dynamic change features from text, images, and time series data. Through contrastive learning and adversarial training, it maps the different modal features into a unified semantic space and uses attention mechanisms or graph neural networks to explore potential connections between data.

[0010] Pre-trained model construction and transfer learning module: This module understands the inherent patterns of cross-modal data, then transfers the model to animal vaccination scenarios and fine-tunes it based on specific data from regional vaccination history.

[0011] Risk prediction and decision support module: This module inputs the integrated and processed multimodal data into the model, calculates and outputs risk assessment results, and modifies the predictions based on real-time environmental and population epidemiological data. The decision-making basis is displayed through a heat map visualization tool.

[0012] Model validation and iterative optimization module: Use historical vaccination data to verify the model's prediction accuracy, compare the actual adverse reaction rate with the predicted value, analyze the causes of data missing and modal correlation errors, continuously access new data on the effects of new vaccines and emerging epidemics, combine expert feedback, and update model parameters and feature weights through online learning or regular retraining to continuously improve model performance.

[0013] Preferably, the risk prediction and decision support module introduces causal diagrams and intervention algorithms to clarify the causal relationship between virus type (V), symptoms (S), and environmental factors (E), avoiding misjudging false associations caused by the environment as virus characteristics;

[0014] V→S←E, the arrows represent causal relationships.

[0015] Preferably, Do-Calculus is used to calculate the probability of symptoms after intervening with the virus type, such as forcibly setting it to virus A, to eliminate the confounding effect of environmental factors:

[0016] P(S|do(V=A))=∑ E P(S|V=A,E)*P(E);

[0017] P(S|do(V=A)) is the probability of symptom S when the virus type is forced to be A;

[0018] ∑ E P(S|V=A,E) is the conditional probability of symptom S when virus A is known and environmental factor E is present;

[0019] P(E) is the prior probability of the environmental factor E. By calculating the difference between P(S|do(V=A)) and P(S|do(V=B)), the actual infected virus type can be more accurately determined.

[0020] Preferably, the risk prediction and decision support module uses cosine similarity and feature importance weights to quantify the feature differences of different viruses in multimodal data and enhance the ability to distinguish.

[0021] Preferably, the multimodal data of each virus (such as molecular characteristics, symptom characteristics, and environmental characteristics) are encoded into a feature vector:

[0022] Characteristic vector of virus A: x A =[x A1 ,x A2 ,…,x An ];

[0023] Virus B’s characteristic vector: x B =[x B1 ,x B2 ,…,x Bn ];

[0024] x Ai and x Bi They represent the characteristic values ​​of virus A and B in the i-th mode, such as the genome sequence code of virus A and the mean body temperature of virus B.

[0025] Preferably, a feature importance weight vector w = [w1, w2, ..., w3] is introduced to calculate the weighted cosine similarity:

[0026]

[0027] Weight w i The larger the value is, the more important the i-th mode is for distinguishing viruses. The closer the similarity value is to 0, the greater the difference between the two viruses in multimodal features, which reduces the probability of misjudgment.

[0028] Preferably, the risk prediction and decision support module introduces Bayesian reasoning and confidence intervals to quantify the uncertainty of model predictions and avoid hasty decisions when there is insufficient evidence.

[0029] Preferably, assuming the existence of viruses A and B, the posterior probability is calculated based on the multimodal observation data D:

[0030]

[0031] P(V=A|D): The posterior probability that the virus type is A under the observation data D;

[0032] P(D|V=A): The likelihood probability of observing data D when the virus type is A;

[0033] P(V=A): Prior probability that the virus type is A.

[0034] Preferably, confidence intervals of the posterior probability are calculated, and if the confidence intervals of the two viruses overlap, manual review or additional testing is triggered.

[0035] A method for implementing an animal vaccination risk prediction system based on a cross-modal macro model comprises the following steps:

[0036] Step 1: The farm collects physiological data such as body temperature and heart rate through smart collars. If the body temperature of some individuals is found to be persistently above the normal range, high-definition cameras will capture visual data of the animals experiencing rapid breathing and abnormal redness of the skin. Simultaneously, text data from the livestock log will indicate recent problems with insufficient ventilation in the pens and high stocking density. At this time, the veterinarian will collect blood samples from the sick animals and obtain molecular data for viral nucleic acid testing. The system will then denoise and normalize this data and align it in time and space based on individual identifiers, for example, associating physiological data and image data from the same cow.

[0037] Step 2: Use NLP technology to extract semantic features from livestock logs and test report text data, such as "poor ventilation" and "suspected respiratory infection." Use CNN to extract visual features such as reddened skin areas and abnormal respiratory rates from visual images of livestock. Use LSTM to analyze temporal patterns of body temperature and heart rate, such as periodic increases in body temperature fluctuations. Use contrastive learning to map different modal features, such as "high density" in text and "clustering" in images, into a unified semantic space. Use graph neural networks to explore potential associations between environmental factors and physiological symptoms.

[0038] Step 3: Pre-train a large cross-modal model based on a vast amount of publicly available data from veterinary medical literature and vaccine trial reports to learn the underlying patterns between different data modalities. The model is then fine-tuned using the farm's historical disease data and vaccination records to adapt it to the farm's breeding environment and livestock breed characteristics, optimizing its ability to predict vaccination risk.

[0039] Step 4: Model construction. Causal diagram analysis found that although livestock showed symptoms of fever and cough, insufficient ventilation in the pens was a potential confounding factor leading to respiratory symptoms. Through Do-Calculus calculation, after excluding the interference of ventilation factors, it was found that the actual probability of viral infection was lower than the initial judgment, avoiding the misjudgment of symptoms caused by environmental stress as specific viral infection. The two suspected viruses were analyzed. The clinical symptoms of the two were similar, but molecular detection data showed that bovine influenza virus had specific hemagglutinin gene fragments in blood samples, while bovine infectious rhinotracheitis virus showed elevated interleukin-6 levels. The model assigned these key molecular characteristics. High weight, calculating weighted cosine similarity, found that the multimodal characteristics of the current sick animal were less different from those of IBRV, thus preliminarily judging that the latter may have been infected. Based on Bayesian reasoning, the posterior probability of infection with IBRV was calculated to be 70%, with a confidence interval of ±15%. Because this probability was close to the threshold and had high uncertainty, the system triggered a manual review mechanism, prompting the veterinarian to isolate and culture the sick animal for virus testing. Ultimately, the test confirmed that the animal was infected with IBRV. The model recommended the appropriate vaccine and vaccination dose based on the virus type and the health status of the individual livestock, and generated an early warning prompt, suggesting that the livestock in the same group be isolated and observed.

[0040] Step 5: Subsequently, the prediction accuracy is verified by comparing the actual epidemic development with the model prediction results. If it is found that the model's risk prediction for pregnant livestock is significantly biased, analysis may be that the physiological data characteristics of pregnancy are not fully incorporated into the model. Therefore, specific data of pregnant livestock are supplemented, and the feature weights are adjusted in combination with veterinary clinical experience. The model is updated through online learning to improve the prediction accuracy for special groups.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] In the present invention, by introducing causal reasoning and multimodal feature weighted comparison mechanism, it can effectively avoid misjudgment caused by the same surface symptoms. In the diagnosis of animal diseases, different viruses may cause similar symptoms. Traditional methods are easily interfered by surface phenomena. The optimized method constructs a causal graph to analyze the real cause of the symptoms, removes confounding factors such as environment and physiological status, and locates the true causal relationship between the virus and symptoms. At the same time, the key differences in multimodal data, such as molecular detection and imaging features, are weighted to highlight the differences in deep characteristics of viruses such as genes and antibodies. For example, it can distinguish the differences in genome sequences of different viruses behind similar respiratory symptoms. Through these means, the model can more accurately distinguish virus types, significantly reduce the risk of misjudgment, and provide a reliable basis for subsequent vaccination strategies.

[0043] In the present invention, the reliability of its own predictions is quantified through Bayesian reasoning or confidence interval calculation. When the uncertainty is high, such as when the probabilities of predicting two viruses are close and the confidence intervals overlap, or when the prediction result is lower than the set threshold, the system automatically triggers manual review or supplementary testing procedures to avoid blind decisions caused by insufficient data or feature confusion. For example, when the model has difficulty determining the specific type of virus that infects an animal, it will prompt the veterinarian to perform more accurate tests such as genetic sequencing. This combination of model automation and human intervention can not only give full play to the efficiency of AI, but also ensure the safety and scientific nature of decision-making, reducing the waste of epidemic prevention resources and the risk of epidemic spread due to misjudgment. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a flow chart of the animal vaccination risk prediction method based on the cross-modal large model of the present invention. DETAILED DESCRIPTION

[0045] The following embodiments of the present invention are described in further detail with reference to the accompanying drawings and examples. The following examples are used to illustrate the present invention but are not intended to limit the scope of the present invention.

[0046] See also Figure 1 The present invention provides an animal vaccination risk prediction system based on a cross-modal large model. The system includes a multimodal data acquisition and integration module, a cross-modal feature extraction and semantic alignment module, a pre-training model construction and transfer learning module, a risk prediction and decision support module, and a model verification and iterative optimization module. The specific details are as follows:

[0047] Multimodal data collection and integration module: comprehensively collects multi-type animal data, including body temperature, body surface characteristics, vaccine instructions, and viral gene information, and processes the raw data;

[0048] Cross-modal feature extraction and semantic alignment module: This module extracts semantic, visual, and dynamic change features from text, images, and time series data. Through contrastive learning and adversarial training, it maps the different modal features into a unified semantic space and uses attention mechanisms or graph neural networks to explore potential connections between data.

[0049] Pre-trained model construction and transfer learning module: This module understands the inherent patterns of cross-modal data, then transfers the model to animal vaccination scenarios and fine-tunes it based on specific data from regional vaccination history.

[0050] Risk prediction and decision support module: This module inputs the integrated and processed multimodal data into the model, calculates and outputs risk assessment results, and modifies the predictions based on real-time environmental and population epidemiological data. The decision-making basis is displayed through a heat map visualization tool.

[0051] Model validation and iterative optimization module: Use historical vaccination data to verify the model's prediction accuracy, compare the actual adverse reaction rate with the predicted value, analyze the causes of data missing and modal correlation errors, continuously access new data on the effects of new vaccines and emerging epidemics, combine expert feedback, and update model parameters and feature weights through online learning or regular retraining to continuously improve model performance.

[0052] The risk prediction and decision support module introduces causal diagrams and intervention algorithms to clarify the causal relationship between virus type (V), symptoms (S), and environmental factors (E), avoiding misjudging false associations caused by the environment as virus characteristics;

[0053] V→S←E, the arrows represent causal relationships.

[0054] Do-Calculus is used to calculate the probability of symptoms after intervening with the virus type (such as forcibly setting it to virus A) to eliminate the confounding effect of environmental factors:

[0055] P(S|do(V=A))=∑ E P(S|V=A,E)*P(E);

[0056] P(S|do(V=A)) is the probability of symptom S when the virus type is forced to be A;

[0057] ∑ E P(S|V=A,E) is the conditional probability of symptom S when virus A is known and environmental factor E is present;

[0058] P(E) is the prior probability of the environmental factor E. By calculating the difference between P(S|do(V=A)) and P(S|do(V=B)), the actual infected virus type can be more accurately determined.

[0059] The risk prediction and decision support module uses cosine similarity and feature importance weights to quantify the feature differences of different viruses in multimodal data and enhance the ability to distinguish.

[0060] Encode the multimodal data of each virus, such as molecular characteristics, symptom characteristics, and environmental characteristics, into feature vectors:

[0061] Characteristic vector of virus A: x A =[x A1 ,x A2 ,...,x An ];

[0062] Virus B’s characteristic vector: x B =[x B1 ,x B2 ,...,x Bn ];

[0063] x Ai and x Bi They represent the characteristic values ​​of virus A and B in the i-th mode, such as the genome sequence code of virus A and the mean body temperature of virus B.

[0064] Introduce the feature importance weight vector w = [w1, w2, ..., w3] and calculate the weighted cosine similarity:

[0065]

[0066] Weight w i The larger the value is, the more important the i-th mode is for distinguishing viruses. The closer the similarity value is to 0, the greater the difference between the two viruses in multimodal features, which reduces the probability of misjudgment.

[0067] The risk prediction and decision support module introduces Bayesian reasoning and confidence intervals to quantify the uncertainty of model predictions and avoid hasty decisions when there is insufficient evidence.

[0068] Assuming that viruses A and B exist, calculate the posterior probability based on multimodal observation data D:

[0069]

[0070] P(V=A|D): The posterior probability that the virus type is A under the observation data D;

[0071] P(D|V=A): The likelihood probability of observing data D when the virus type is A;

[0072] P(V=A): Prior probability that the virus type is A.

[0073] Calculate the confidence interval of the posterior probability. If the confidence intervals of the two viruses overlap, manual review or additional testing will be triggered.

[0074] A method for implementing an animal vaccination risk prediction system based on a cross-modal macro model comprises the following steps:

[0075] Step 1: The farm collects physiological data such as body temperature and heart rate through smart collars. If the body temperature of some individuals is found to be persistently above the normal range, high-definition cameras will capture visual data of the animals experiencing rapid breathing and abnormal redness of the skin. Simultaneously, text data from the livestock log will indicate recent problems with insufficient ventilation in the pens and high stocking density. At this time, the veterinarian will collect blood samples from the sick animals and obtain molecular data for viral nucleic acid testing. The system will then denoise and normalize this data and align it in time and space based on individual identifiers, for example, associating physiological data and image data from the same cow.

[0076] Step 2: Use NLP technology to extract semantic features from livestock logs and test report text data, such as "poor ventilation" and "suspected respiratory infection." Use CNN to extract visual features such as reddened skin areas and abnormal respiratory rates from visual images of livestock. Use LSTM to analyze temporal patterns of body temperature and heart rate, such as periodic increases in body temperature fluctuations. Use contrastive learning to map different modal features, such as "high density" in text and "clustering" in images, into a unified semantic space. Use graph neural networks to explore potential associations between environmental factors and physiological symptoms.

[0077] Step 3: Pre-train a large cross-modal model based on a vast amount of publicly available data from veterinary medical literature and vaccine trial reports to learn the underlying patterns between different data modalities. The model is then fine-tuned using the farm's historical disease data and vaccination records to adapt it to the farm's breeding environment and livestock breed characteristics, optimizing its ability to predict vaccination risk.

[0078] Step 4: Model construction. Causal diagram analysis found that although livestock showed symptoms of fever and cough, insufficient ventilation in the pens was a potential confounding factor leading to respiratory symptoms. Through Do-Calculus calculation, after excluding the interference of ventilation factors, it was found that the actual probability of viral infection was lower than the initial judgment, avoiding the misjudgment of symptoms caused by environmental stress as specific viral infection. The two suspected viruses were analyzed. The clinical symptoms of the two were similar, but molecular detection data showed that bovine influenza virus had specific hemagglutinin gene fragments in blood samples, while bovine infectious rhinotracheitis virus showed elevated interleukin-6 levels. The model assigned these key molecular characteristics. High weight, calculating weighted cosine similarity, found that the multimodal characteristics of the current sick animal were less different from those of IBRV, thus preliminarily judging that the latter may have been infected. Based on Bayesian reasoning, the posterior probability of infection with IBRV was calculated to be 70%, with a confidence interval of ±15%. Because this probability was close to the threshold and had high uncertainty, the system triggered a manual review mechanism, prompting the veterinarian to isolate and culture the sick animal for virus testing. Ultimately, the test confirmed that the animal was infected with IBRV. The model recommended the appropriate vaccine and vaccination dose based on the virus type and the health status of the individual livestock, and generated an early warning prompt, suggesting that the livestock in the same group be isolated and observed.

[0079] Step 5: Subsequently, the prediction accuracy is verified by comparing the actual epidemic development with the model prediction results. If it is found that the model's risk prediction for pregnant livestock is significantly biased, analysis may be that the physiological data characteristics of pregnancy are not fully incorporated into the model. Therefore, specific data of pregnant livestock are supplemented, and the feature weights are adjusted in combination with veterinary clinical experience. The model is updated through online learning to improve the prediction accuracy for special groups.

[0080] Working principle:

[0081] The first step is to introduce causal reasoning and multimodal feature weighted comparison mechanism, which can effectively avoid misjudgment caused by the same surface symptoms. In the diagnosis of animal diseases, different viruses may cause similar symptoms. Traditional methods are easily interfered by surface phenomena. The optimized method constructs a causal graph to analyze the real cause of the symptoms, removes confounding factors such as the environment and physiological status, and locates the true causal relationship between the virus and the symptoms. At the same time, the key differences in multimodal data, such as molecular detection and imaging features, are weighted to highlight the differences in deep characteristics of the virus such as genes and antibodies. For example, it can distinguish the differences in the genome sequences of different viruses behind similar respiratory symptoms. Through these means, the model can more accurately distinguish virus types, significantly reduce the risk of misjudgment, and provide a reliable basis for subsequent vaccination strategies.

[0082] In the second step, uncertainty quantification and threshold decision-making mechanisms make model decisions more scientific. During the prediction process, the model no longer simply outputs results, but quantifies the reliability of its own predictions through Bayesian reasoning or confidence interval calculations. When the uncertainty is high, such as when the probabilities of predicting two viruses are close and the confidence intervals overlap, or when the prediction result is lower than the set threshold, the system automatically triggers manual review or supplementary testing processes to avoid blind decisions caused by insufficient data or feature confusion. For example, when the model has difficulty determining the specific type of virus that infects an animal, it will prompt the veterinarian to perform more accurate tests such as genetic sequencing. This combination of model automation and human intervention can not only give full play to the efficiency of AI, but also ensure the safety and scientific nature of decision-making, reducing the waste of epidemic prevention resources and the risk of epidemic spread due to misjudgment.

[0083] The embodiments of the present invention are presented for purposes of illustration and description and are not intended to be exhaustive or to limit the invention to the disclosed forms. Many modifications and variations will be apparent to those skilled in the art. The embodiments are chosen and described in order to better illustrate the principles of the invention and its practical application and to enable those skilled in the art to understand the invention and design various embodiments with various modifications as suited for specific applications.

Claims

1. An animal vaccination risk prediction system based on a cross-modal large model, characterized by: The system includes a multimodal data acquisition and integration module, a cross-modal feature extraction and semantic alignment module, a pre-training model construction and transfer learning module, a risk prediction and decision support module, and a model verification and iterative optimization module, as follows: Multimodal data collection and integration module: comprehensively collects multi-type animal data, including body temperature, body surface characteristics, vaccine instructions, and viral gene information, and processes the raw data; Cross-modal feature extraction and semantic alignment module: This module extracts semantic, visual, and dynamic change features from text, images, and time series data. Through contrastive learning and adversarial training, it maps the different modal features into a unified semantic space and uses attention mechanisms or graph neural networks to explore potential connections between data. Pre-trained model construction and transfer learning module: This module understands the inherent patterns of cross-modal data, then transfers the model to animal vaccination scenarios and fine-tunes it based on specific data from regional vaccination history. Risk prediction and decision support module: This module inputs the integrated and processed multimodal data into the model, calculates and outputs risk assessment results, and modifies the predictions based on real-time environmental and population epidemiological data. The decision-making basis is displayed through a heat map visualization tool. Model validation and iterative optimization module: Use historical vaccination data to verify the model's prediction accuracy, compare the actual adverse reaction rate with the predicted value, analyze the causes of data missing and modal correlation errors, continuously access new data on the effects of new vaccines and emerging epidemics, combine expert feedback, and update model parameters and feature weights through online learning or regular retraining to continuously improve model performance.

2. The animal vaccination risk prediction system based on a cross-modal large model according to claim 1, characterized in that: The risk prediction and decision support module introduces causal diagrams and intervention algorithms to clarify the causal relationship between virus type (V), symptoms (S), and environmental factors (E); Among them, V→S←E, and the arrows represent causal relationships.

3. The animal vaccination risk prediction system based on a cross-modal large model according to claim 2, characterized in that: Do-Calculus is used to calculate the probability of symptoms after intervening with the virus type (such as forcibly setting it to virus A) to eliminate the confounding effect of environmental factors: P(S|do(V=A))=∑ E P(S|V=A,E)*P(E); Where P(S|do(V=A)) is the probability of symptom S appearing when the virus type is forced to be A; ∑ E P(S|V=A,E) is the conditional probability of symptom S when virus A is known and environmental factor E is present; P(E) is the prior probability of environmental factor E.

4. The animal vaccination risk prediction system based on a cross-modal large model according to claim 1, characterized in that: The risk prediction and decision support module uses cosine similarity and feature importance weights to quantify the feature differences of different viruses in multimodal data.

5. The animal vaccination risk prediction system based on a cross-modal large model according to claim 4, characterized in that: Encode the multimodal data of each virus (such as molecular characteristics, symptom characteristics, and environmental characteristics) into a feature vector: Characteristic vector of virus A: x A =[x A1 ,x A2 ,…,x An ]; Virus B’s characteristic vector: x B =[x B1 ,x B2 ,…,x Bn ]; Among them, x Ai and x Bi Represent the eigenvalues ​​of virus A and B in the i-th mode respectively.

6. The animal vaccination risk prediction system based on a cross-modal large model according to claim 4, characterized in that: Introduce the feature importance weight vector w = [w1, w2, ..., w3] and calculate the weighted cosine similarity: Weight w i The larger the value is, the more important the i-th mode is for distinguishing viruses. The closer the similarity value is to 0, the greater the difference between the two viruses in multimodal features, which reduces the probability of misjudgment.

7. The animal vaccination risk prediction system based on a cross-modal large model according to claim 1, characterized in that: The risk prediction and decision support module introduces Bayesian reasoning and confidence intervals to quantify the uncertainty of model predictions.

8. The animal vaccination risk prediction system based on a cross-modal large model according to claim 7, characterized in that: Assuming that viruses A and B exist, calculate the posterior probability based on multimodal observation data D: Where P(V=A|D): the posterior probability that the virus type is A under the observation data D; P(D|V=A): The likelihood probability of observing data D when the virus type is A; P(V=A): Prior probability that the virus type is A.

9. The animal vaccination risk prediction system based on a cross-modal large model according to claim 7, characterized in that: Calculate the confidence interval of the posterior probability. If the confidence intervals of the two viruses overlap, manual review or additional testing will be triggered.

10. A method for implementing the animal vaccination risk prediction system based on a cross-modal large model according to claims 1-9, characterized in that: The method comprises the following steps: Step 1: The farm collects physiological data such as body temperature and heart rate through smart collars. If the body temperature of some individuals is found to be persistently above the normal range, high-definition cameras will capture visual data of the animals experiencing rapid breathing and abnormal redness of the skin. Simultaneously, text data from the livestock log will indicate recent problems with insufficient ventilation in the pens and high stocking density. At this time, the veterinarian will collect blood samples from the sick animals and obtain molecular data for viral nucleic acid testing. The system will then denoise and normalize this data and align it in time and space based on individual identifiers, for example, associating physiological data and image data from the same cow. Step 2: Use NLP technology to extract semantic features from livestock logs and test report text data, such as "poor ventilation" and "suspected respiratory infection." CNN extracts visual features of skin redness and abnormal respiratory rate from visual images of livestock. Long-term time-lapse memory (LSTM) is used to analyze temporal patterns of body temperature and heart rate, such as periodic increases in body temperature fluctuations. Contrastive learning is used to map different modal features, such as "high density" in text and "clustering" in images, into a unified semantic space. Graph neural networks are then used to explore potential associations between environmental factors and physiological symptoms. Step 3: Pre-train a large cross-modal model based on a vast amount of publicly available data from veterinary medical literature and vaccine trial reports to learn the underlying patterns between different data modalities. The model is then fine-tuned using the farm's historical disease data and vaccination records to adapt it to the farm's breeding environment and livestock breed characteristics, optimizing its ability to predict vaccination risk. Step 4: Model construction. Causal diagram analysis found that although livestock showed symptoms of fever and cough, insufficient ventilation in the pens was a potential confounding factor leading to respiratory symptoms. Through Do-Calculus calculation, after excluding the interference of ventilation factors, it was found that the actual probability of viral infection was lower than the initial judgment, avoiding the misjudgment of symptoms caused by environmental stress as specific viral infection. The two suspected viruses were analyzed. The clinical symptoms of the two were similar, but molecular detection data showed that bovine influenza virus had specific hemagglutinin gene fragments in blood samples, while bovine infectious rhinotracheitis virus showed elevated interleukin-6 levels. The model assigned these key molecular characteristics. High weight, calculating weighted cosine similarity, found that the multimodal characteristics of the current sick animal were less different from those of IBRV, thus preliminarily judging that the latter may have been infected. Based on Bayesian reasoning, the posterior probability of infection with IBRV was calculated to be 70%, with a confidence interval of ±15%. Because this probability was close to the threshold and had high uncertainty, the system triggered a manual review mechanism, prompting the veterinarian to isolate and culture the sick animal for virus testing. Ultimately, the test confirmed that the animal was infected with IBRV. The model recommended the appropriate vaccine and vaccination dose based on the virus type and the health status of the individual livestock, and generated an early warning prompt, suggesting that the livestock in the same group be isolated and observed. Step 5: Subsequently, the prediction accuracy is verified by comparing the actual epidemic development with the model prediction results. If it is found that the model's risk prediction for pregnant livestock is significantly biased, analysis may be that the physiological data characteristics of pregnancy are not fully incorporated into the model. Therefore, specific data of pregnant livestock are supplemented, and the feature weights are adjusted in combination with veterinary clinical experience. The model is updated through online learning to improve the prediction accuracy for special groups.

Images