A method, device, and readable medium for auxiliary diagnosis of respiratory tract infection pathogens

Abstract

This invention discloses an auxiliary diagnostic method, device, and readable medium for respiratory tract infection pathogens, relating to the field of medical data processing. The method includes: acquiring key features related to respiratory tract infection in a person to be diagnosed and inputting them into a respiratory tract infection pathogen auxiliary diagnostic model; the key features are first processed by a trained respiratory tract infection site prediction model to obtain infection site prediction results; the infection site prediction results and key features are input into the trained respiratory tract infection pathogen prediction model to obtain pathogen prediction results, which are then input into a trained mixed infection prediction model to obtain a mixed infection prediction probability; the infection site prediction results and pathogen prediction results are input into a trained severe illness risk prediction model to obtain a severe illness risk prediction probability; and the presence of mixed infection and severe illness risk in the person to be diagnosed is determined based on the mixed infection prediction probability and the severe illness risk prediction probability, respectively. This invention addresses the problem of low diagnostic accuracy in existing diagnostic methods.

Description

Technical Field

[0001] This invention relates to the field of medical data processing, and more specifically to a method, device, and readable medium for auxiliary diagnosis of respiratory tract infection pathogens. Background Technology

[0002] The common cold (upper respiratory tract infection), bronchitis, and pneumonia are the three most common types of respiratory infections in childhood. These diseases are mostly caused by viruses, bacteria, or atypical pathogens, and their clinical manifestations often overlap, primarily presenting with nonspecific symptoms such as cough and fever. Therefore, differential diagnosis and pathogen identification based solely on clinical presentation are often insufficient, requiring imaging examinations to pinpoint the site of infection. Since the treatment regimens for different infection sites and their causative pathogens vary significantly, early and accurate diagnosis is crucial for improving child prognosis and avoiding the misuse of antibiotics. Children face greater challenges in diagnosing respiratory infections due to their incompletely developed immune systems, atypical clinical presentations, and limited ability to describe their symptoms, making them more prone to developing severe illness.

[0003] Currently, commonly used clinical methods for diagnosing respiratory pathogens each have significant limitations: microbial culture methods are time-consuming and have limited sensitivity; while PCR technology improves detection speed, its coverage of pathogens is limited, and its ability to identify mixed infections is insufficient; imaging examinations (such as X-rays) carry radiation exposure risks and are poorly accessible in primary healthcare institutions; routine laboratory tests (such as complete blood counts) can indicate the presence of infection, but lack pathogen specificity, requiring reliance on clinicians' experience for judgment. These bottlenecks hinder the rapid and accurate classification and treatment of respiratory infections, especially in pediatric patients.

[0004] In recent years, artificial intelligence (AI) technologies, represented by machine learning (ML) and deep learning (DL), have developed rapidly and are increasingly being applied in the healthcare field. Their powerful high-dimensional data processing and pattern recognition capabilities are particularly well-suited for mining the complex clinical information contained in electronic health records (EHRs), and have gradually become an important pillar in biomedical research and the development of auxiliary diagnostic tools. By integrating multi-source, multi-modal clinical data (such as demographic characteristics, laboratory results, and imaging reports), AI models can capture multi-level feature associations that are difficult for the human brain to recognize, thereby establishing disease prediction and classification models and making it possible to achieve precision medicine—the entire process of "prevention-early diagnosis-subtyping-targeted therapy."

[0005] Several studies have attempted to apply AI technology to the diagnosis and prediction of respiratory infections, with some models showing good performance in both adult and pediatric populations. However, existing methods still have significant limitations: most models are designed for single pathogens and fail to adequately consider the common mixed infections in clinical practice; there is also a lack of diagnostic tools specifically designed for children that integrate multiple indicators for joint assessment.

[0006] It is particularly important to note that respiratory infections in their early stages are often difficult to distinguish based on symptoms alone. Although most upper respiratory tract infections are self-limiting and have a short course, while lower respiratory tract infections (such as bronchitis and pneumonia) may progress continuously and have prolonged symptoms, their early clinical manifestations are highly similar. "Cough progressing to pneumonia" is actually the progression of the infection itself, not the cough causing pneumonia. Therefore, timely and accurate identification of the site of infection and potential pathogens in the early stages of infection is crucial for implementing targeted interventions, halting disease progression, and improving prognosis.

[0007] Against this backdrop, developing a diagnostic method that can comprehensively utilize routine clinical and laboratory data, especially applicable to pediatric populations, and simultaneously identify single and mixed respiratory pathogen infections has significant clinical value and application prospects. Current respiratory pathogen diagnostic methods suffer from the following limitations:

[0008] 1. Limited diagnostic accuracy: Traditional clinical laboratory test results cannot accurately reflect an individual's infection status, resulting in low accuracy in predicting infection risk. Existing respiratory pathogen diagnostic models mostly rely on single types of clinical data, making it difficult to comprehensively capture the complex patterns of pathogen infection. Especially for mixed infections and atypical pathogens (such as mycoplasma, adenovirus, etc.), the sensitivity and specificity of the models decrease significantly, easily leading to misdiagnosis or missed diagnosis.

[0009] 2. Poor model generalization and interpretability: Although machine learning models (such as deep learning) perform well on the training set, they still have significant shortcomings in issues such as black-box decision-making, feature collinearity, and overfitting. The generalization ability of the model in children of different populations, regions, and age groups has not been fully validated, and the diagnostic results lack clinically understandable interpretations, reducing doctors' trust in the system.

[0010] 3. High barriers to clinical applicability and engineering deployment: Existing AI diagnostic systems often rely on high-quality, large-scale labeled data, but the data quality in primary healthcare institutions varies greatly, and labeling resources are scarce, leading to a decline in model performance in practical applications. At the same time, the systems have high hardware computing power requirements, and the inference speed cannot meet the needs of real-time outpatient diagnosis, making it difficult to promote them in resource-constrained scenarios.

[0011] 4. Overly Simplified Predictive Models: Many current respiratory infection diagnostic models are based on simple linear models. Whether using statistical or machine learning methods, these models often exhibit significant errors when handling high-dimensional clinical data. Simple linear models perform poorly when the relationship between features and diseases is complex. Electronic health records typically contain a large number of features, and linear models may fail to capture the nonlinear relationships and interactions between these features. Simple linear models are sensitive to collinearity of features, which often leads to model instability and increased prediction errors. Summary of the Invention

[0012] The purpose of this application is to provide an auxiliary diagnostic method, device, and readable medium for respiratory tract infection pathogens in response to the aforementioned technical problems.

[0013] In a first aspect, the present invention provides an auxiliary diagnostic method for respiratory tract infection pathogens, comprising the following steps:

[0014] A feature representation architecture based on ensemble learning is constructed. The feature representation architecture includes several voting ensemble models. The prediction results of all voting ensemble models are fused using a soft voting method to obtain the output result of the feature representation architecture.

[0015] A respiratory infection pathogen auxiliary diagnostic model is constructed, which includes a trained respiratory infection site prediction model, a trained respiratory infection pathogen prediction model, a trained mixed infection prediction model, and a trained severe disease risk prediction model. The respiratory infection site prediction model adopts a feature representation architecture, and the respiratory infection pathogen prediction model is a loop structure containing the feature representation architecture.

[0016] Key features related to respiratory infection in individuals seeking diagnosis are acquired and input into a respiratory infection pathogen-assisted diagnostic model. These key features are first processed by a trained respiratory infection site prediction model to obtain infection site prediction results. The infection site prediction results and key features are then input into a trained respiratory infection pathogen prediction model to obtain pathogen prediction results. These pathogen prediction results are then input into a trained mixed infection prediction model to obtain a mixed infection prediction probability. Finally, the infection site prediction results and pathogen prediction results are input into a trained severe illness risk prediction model to obtain a severe illness risk prediction probability. Based on the mixed infection prediction probability and the severe illness risk prediction probability, the presence of mixed infection and severe illness risk in the individuals seeking diagnosis is determined, respectively. Finally, the pathogen and infection site of the individuals seeking diagnosis are confirmed by combining the infection site prediction results and the pathogen prediction results.

[0017] As a preferred option, the construction process of the voting integration model is as follows:

[0018] Gradient boosting decision tree model, light gradient boosting machine model, random forest model and extreme gradient boosting model are selected as base learners. Each base learner is trained by a hyperparameter search mechanism based on multi-objective co-evolution strategy to obtain trained base learners.

[0019] A soft voting ensemble mechanism is used to weight and fuse the predicted probabilities output by all trained base learners to form an initial voting ensemble model.

[0020] Each trained base learner in the initial voting ensemble model is retrained using the bagging method to obtain several voting ensemble models.

[0021] As a preferred approach, the hyperparameter search mechanism based on a multi-objective co-evolution strategy proceeds as follows:

[0022] By embedding the shared and unique hyperparameters of all base learners into a high-dimensional composite search space, a collaborative optimization objective function is constructed, as shown in the following equation:

[0023] ;

[0024] in, Represents the set of hyperparameters for all base learners. Let represent the hyperparameters of the gradient boosting decision tree model, the lightweight gradient boosting machine model, the random forest model, and the extreme gradient boosting model, respectively. This represents the set of hyperparameters used when the maximum value is reached. This represents the dynamic weights of the m-th base learner. Represents the validation set. Let be the area under the curve performance metric of the m-th base learner on the validation set. The regularization coefficient is . The regularization term, which characterizes the hyperparameter complexity, is expressed as follows:

[0025] ;

[0026] in, These represent the first, second, and third levels of punishment, respectively. The m-th base learner represents the set of hyperparameters of the m-th base learner. The values ​​of the hyperparameters, Represents the first base learner of all base learners. The baseline values ​​for each hyperparameter. Denotes the set of hyperparameters of the m-th base learner. The number of all hyperparameters, Represents the m-th base learner A function with a time complexity of . This represents the measurement of the m-th base learner. and the nth base learner The function for predicting variance is expressed as follows:

[0027] ;

[0028] in, Let represent the predicted probability vector of the m-th base learner on the validation set. This represents the Pearson correlation coefficient;

[0029] In the t-th training epoch, its corresponding dynamic weights in the t-th training epoch. Calculated using the following formula:

[0030] ;

[0031] in, This represents the area under the curve (AUC) of the m-th base learner on the validation set during the t-th training epoch. This represents the area under the curve (AUC) of the nth base learner on the validation set during the t-th training epoch. Scaling factor For temperature parameters;

[0032] An improved Bayesian optimization algorithm is used to perform a global search of the hyperparameter set in a high-dimensional composite search space. Multi-objective optimization is then performed based on a collaborative optimization objective function to determine the optimal hyperparameter set.

[0033] Preferably, the respiratory infection pathogen prediction model adopts a K-layer recurrent pathogen prediction architecture. This architecture uses a feature representation framework. In the first round, the infection site prediction result, key features, and zero vector are concatenated and input into the first-layer pathogen prediction architecture to obtain the pathogen prediction probability for the first round. In the current round, the infection site prediction result, key features, and the pathogen prediction probability from the previous round are concatenated and input into the current round's pathogen prediction architecture to obtain the pathogen prediction probability for the current round, as shown in the following formula:

[0034] ;

[0035] ;

[0036] in, This represents the pathogen prediction architecture in the k-th round. This represents the pathogen prediction probability output by the pathogen prediction architecture in the k-th round. It is the pathogen prediction probability output by the pathogen prediction architecture in the (k-1)th round. Indicates key features, Indicates the predicted site of infection; This represents the feature obtained by concatenating the infection site prediction result, key features, and pathogen prediction probability of round k-1 in the k-th round; the pathogen prediction probability of round k is taken as the pathogen prediction result.

[0037] The weighted binary cross-entropy loss function is used in the training process of the respiratory infection pathogen prediction model, as shown in the following formula:

[0038] ;

[0039] Where N represents the number of samples in the training data. This represents the true label of the j-th pathogen in the i-th sample of the training data. This represents the probability that the respiratory infection pathogen prediction model predicts the j-th pathogen for the i-th sample. is the weighting coefficient for the j-th type of pathogen.

[0040] Preferably, the pathogen prediction results output by the trained respiratory infection pathogen prediction model are input into the mixed infection prediction model, and the mixed infection prediction probability is calculated using the following formula:

[0041] ;

[0042] in, It is the Sigmoid activation function. This represents the first learnable parameter. Key features The corresponding predicted probability of the existence of the j-th pathogen. Key features The corresponding predicted probability of mixed infection;

[0043] The prediction results of the infection site and pathogen are input into the severe illness risk prediction model, and the predicted probability of severe illness risk is calculated using the following formula:

[0044] ;

[0045] in, and These represent the predicted infection site and the predicted pathogen, respectively. and This represents the second learnable parameter. For random noise, Indicates the probability of predicting the risk of severe illness;

[0046] During the training process of the mixed infection prediction model and the severe illness risk prediction model, the parameters of the trained respiratory infection site prediction model and the trained respiratory infection pathogen prediction model are fixed, and the first learning parameter and the second learnable parameter in the mixed infection prediction model and the severe illness risk prediction model are trained respectively.

[0047] As a preferred option, the key features are multidimensional features composed of key indicators, including age, sex, alanine aminotransferase, aspartate aminotransferase, the ratio of aspartate aminotransferase to alanine aminotransferase, creatine kinase, direct bilirubin, total bilirubin, lactate dehydrogenase, total protein, albumin, globulin, the ratio of albumin to globulin, alkaline phosphatase, gamma-glutamyl transferase, potassium, sodium, chloride, phosphorus, blood urea nitrogen, creatinine, uric acid, carbon dioxide binding capacity, triglycerides, neutrophils, monocytes, monocyte ratio, eosinophils, eosinophil ratio, basophils, basophil ratio, neutrophil ratio, lymphocyte ratio, red blood cell count, hemoglobin, mean corpuscular volume, white blood cells, platelets, mean corpuscular hemoglobin concentration, red blood cell distribution width, mean platelet volume, and platelet distribution width.

[0048] Secondly, the present invention provides an auxiliary diagnostic device for respiratory tract infection pathogens, comprising:

[0049] The architecture building module is configured to build a feature representation architecture based on ensemble learning. The feature representation architecture includes several voting ensemble models. The prediction results of all voting ensemble models are fused using a soft voting method to obtain the output result of the feature representation architecture.

[0050] The model building module is configured to build a respiratory infection pathogen auxiliary diagnostic model. The respiratory infection pathogen auxiliary diagnostic model includes a trained respiratory infection site prediction model, a trained respiratory infection pathogen prediction model, a trained mixed infection prediction model, and a trained severe disease risk prediction model. The respiratory infection site prediction model adopts a feature representation architecture, and the respiratory infection pathogen prediction model is a loop structure containing the feature representation architecture.

[0051] The prediction module is configured to acquire key features related to respiratory infection in individuals seeking diagnosis and input them into a respiratory infection pathogen-assisted diagnostic model. These key features are first processed by a trained respiratory infection site prediction model to obtain infection site prediction results. The infection site prediction results and key features are then input into a trained respiratory infection pathogen prediction model to obtain pathogen prediction results. These pathogen prediction results are then input into a trained mixed infection prediction model to obtain a mixed infection prediction probability. Finally, the infection site prediction results and pathogen prediction results are input into a trained severe illness risk prediction model to obtain a severe illness risk prediction probability. Based on the mixed infection prediction probability and the severe illness risk prediction probability, the module determines whether the individual seeking diagnosis has a mixed infection and a severe illness risk, and combines the infection site prediction results and pathogen prediction results to confirm the pathogen and infection site of the individual seeking diagnosis.

[0052] Thirdly, the present invention provides an electronic device including one or more processors; and a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any implementation of the first aspect.

[0053] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method as described in any of the implementations of the first aspect.

[0054] Fifthly, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the method as described in any of the implementations in the first aspect.

[0055] Compared with the prior art, the present invention has the following beneficial effects:

[0056] (1) The respiratory infection pathogen auxiliary diagnosis method proposed in this invention constructs a respiratory infection pathogen auxiliary diagnosis model and adopts a multi-layer stacked integration framework and hyperparameter optimization strategy. It can effectively integrate the prediction results of multiple high-performance machine learning models and significantly improve the identification ability of mixed infections and rare pathogens. The model is trained and validated on a large-scale dataset and shows excellent generalization performance and stability. It can provide reliable and interpretable pathogen diagnosis output in different medical scenarios. It simultaneously realizes four tasks within the same framework: infection site identification, pathogen classification, mixed infection identification, and severe risk warning, truly achieving a highly efficient diagnosis mode of "one model with multiple outputs". It can be seamlessly embedded into hospital information systems (HIS), laboratory information systems (LIS), or mobile medical platforms to provide a full-process auxiliary diagnosis and treatment solution for children's respiratory infections, from early diagnosis and classification assessment to risk prediction.

[0057] (2) The respiratory infection pathogen auxiliary diagnosis method proposed in this invention comprehensively utilizes multi-source features such as demographic data, vital signs, and laboratory test results (e.g., blood routine and inflammatory markers) to overcome the limitations of single data types and comprehensively capture the physiological state and infection characteristics of children. This multimodal fusion strategy enhances the model's ability to distinguish complex infection phenotypes, and is particularly suitable for respiratory infections with atypical manifestations and diverse pathogens. Built-in interpretable algorithms such as SHAP (SHapley Additive ex Planations) and LIME can clearly demonstrate the contribution and direction of each feature to the individual prediction results, assisting doctors in understanding the model's inference logic. The output results include pathogen probability distribution, mixed infection risk labeling, and confidence assessment, effectively enhancing clinical trust and decision-making transparency.

[0058] (3) The respiratory infection pathogen auxiliary diagnosis method proposed in this invention is based on routine electronic health records (EHR) and biochemical blood routine data, without the need for additional complex tests or special sampling, which greatly improves clinical usability and ease of operation. Through efficient feature engineering and ensemble learning architecture, it can quickly extract key discriminative information from multi-dimensional clinical indicators, realize early identification of pathogens and determination of infection sites, significantly shorten the diagnosis cycle, and is suitable for rapid decision support in outpatient and emergency scenarios. Attached Figure Description

[0059] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0060] Figure 1 This is a schematic flowchart illustrating an auxiliary diagnostic method for respiratory infection pathogens according to an embodiment of this application.

[0061] Figure 2 This is a schematic diagram of the model architecture of the respiratory tract infection pathogen auxiliary diagnosis method according to an embodiment of this application;

[0062] Figure 3 This is a schematic diagram of an auxiliary diagnostic device for respiratory infection pathogens according to an embodiment of this application;

[0063] Figure 4 A schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0065] Figure 1 An embodiment of this application illustrates an auxiliary diagnostic method for respiratory infection pathogens, comprising the following steps:

[0066] S1. Construct a feature representation architecture based on ensemble learning. The feature representation architecture includes several voting ensemble models. Use soft voting to fuse the prediction results of all voting ensemble models to obtain the output result of the feature representation architecture.

[0067] In a specific embodiment, the construction process of the voting integration model is as follows:

[0068] Gradient boosting decision tree model, light gradient boosting machine model, random forest model and extreme gradient boosting model are selected as base learners. Each base learner is trained by a hyperparameter search mechanism based on multi-objective co-evolution strategy to obtain trained base learners.

[0069] A soft voting ensemble mechanism is used to weight and fuse the predicted probabilities output by all trained base learners to form an initial voting ensemble model.

[0070] Each trained base learner in the initial voting ensemble model is retrained using the bagging method to obtain several voting ensemble models.

[0071] In a specific embodiment, the hyperparameter search mechanism based on the multi-objective co-evolution strategy proceeds as follows:

[0072] By embedding the shared and unique hyperparameters of all base learners into a high-dimensional composite search space, a collaborative optimization objective function is constructed, as shown in the following equation:

[0073] ;

[0074] in, Represents the set of hyperparameters for all base learners. Let represent the hyperparameters of the gradient boosting decision tree model, the lightweight gradient boosting machine model, the random forest model, and the extreme gradient boosting model, respectively. This represents the set of hyperparameters used when the maximum value is reached. This represents the dynamic weights of the m-th base learner. Represents the validation set. Let be the area under the curve performance metric of the m-th base learner on the validation set. The regularization coefficient is . The regularization term, used to characterize the hyperparameter complexity, is expressed as follows:

[0075] ;

[0076] in, These represent the first, second, and third levels of punishment, respectively. The m-th base learner represents the set of hyperparameters of the m-th base learner. The values ​​of the hyperparameters, Represents the first base learner of all base learners. The baseline values ​​for each hyperparameter. Denotes the set of hyperparameters of the m-th base learner. The number of all hyperparameters, Represents the m-th base learner A function with a time complexity of . This represents the measurement of the m-th base learner. and the nth base learner The function for predicting variance is expressed as follows:

[0077] ;

[0078] in, Let represent the predicted probability vector of the m-th base learner on the validation set. This represents the Pearson correlation coefficient;

[0079] In the t-th training epoch, its corresponding dynamic weights in the t-th training epoch. Calculated using the following formula:

[0080] ;

[0081] in, This represents the area under the curve (AUC) of the m-th base learner on the validation set during the t-th training epoch. This represents the area under the curve (AUC) of the nth base learner on the validation set during the t-th training epoch. Scaling factor For temperature parameters;

[0082] An improved Bayesian optimization algorithm is used to perform a global search of the hyperparameter set in a high-dimensional composite search space. Multi-objective optimization is then performed based on a collaborative optimization objective function to determine the optimal hyperparameter set.

[0083] For details, please refer to Figure 2The feature representation architecture proposed in the embodiments of this application is a multi-layer stacked architecture, which integrates four cutting-edge machine learning algorithm models, including Gradient Boosting Decision Tree (GBDT) model, Light Gradient Boosting Machine (LightGBM) model, Random Forest (RF) model, and Extreme Gradient Boosting (XGBoost) model. To comprehensively improve the overall diagnostic performance of the model, the embodiments of this application adopt a unified hyperparameter optimization strategy for multi-model collaboration. This method breaks through the limitations of traditional independent parameter tuning of a single model. By constructing a joint optimization framework, it simultaneously performs efficient collaborative tuning of the four tree models: Gradient GBDT, RF, LightGBM, and XGBoost. The embodiments of this application design a hyperparameter search mechanism based on a multi-objective co-evolution strategy. It embeds the shared and unique hyperparameters (such as learning rate, tree depth, number of leaf nodes, feature sampling ratio, etc.) of the four base learners into a unified high-dimensional composite search space. An improved Bayesian optimization algorithm is used to globally explore the high-dimensional composite search space and construct a collaborative optimization objective function. The regularization term in the collaborative optimization objective function is used to control the overall model complexity and alleviate overfitting.

[0084] The first of all base learners in this regularization term The baseline value of the hyperparameter can be the th hyperparameter of all base learners. The average of the hyperparameters, The term will penalize hyperparameters that deviate from the baseline value to prevent the model from becoming overly complex due to a certain parameter becoming extreme.

[0085] It is a measure of the first The complexity is a function of the complexity of each base learner. For tree models (such as GBDT, RF, LightGBM, XGBoost), this can be the total number of leaf nodes in all trees (n). leaves ), the total number of all trees (n) estimators ), average tree depth (max) depth Linear or logarithmic combinations of indicators such as ).

[0086] It is a measure of two base learners Sum base learners This is a function predicting the diversity. Subtracting this term means that the greater the diversity between models, the smaller the penalty (equivalent to a reward), which helps the ensemble model achieve better performance. A common metric is to calculate the correlation coefficient or cosine similarity of the models' predictions on the validation set. The diversity reward can be defined as... . These are three hyperparameters used for trade-offs, controlling the strength of parameter value penalties, structural complexity penalties, and model diversity rewards in the total penalty term, respectively. They can also be set through meta-optimization or based on prior knowledge.

[0087] During training, an efficient optimization process combining mini-batch gradient descent and early stopping is employed. Each iteration randomly samples 32 samples from the training set as a batch, updating the parameters of the four base learners in parallel. After each batch training, the AUC loss of each base learner on the current validation set is calculated, and the hyperparameter configuration is updated through backpropagation and a multi-objective optimization algorithm. During training, the dynamic weighting strategy employed in this application's embodiments is quantified and adaptively adjusted based on the real-time performance of each base learner on the validation set. The core of this strategy is to dynamically allocate limited training resources and attention to base learners with better performance or greater potential, while maintaining appropriate differences among base learners to ensure ensemble effectiveness. The scaling factor η in the dynamic weights is used to control the sensitivity of performance differences to weight allocation. The larger the η value, the more obvious the weight advantage obtained by the high-performance model; the temperature parameter is used to smooth the weight distribution.

[0088] Furthermore, embodiments of this application design a two-layer stacked ensemble architecture to improve prediction robustness and accuracy. The first layer consists of four heterogeneous ensemble algorithms, including the Gradient GBDT model, the RF model, the LightGBM model, and the XGBoost model. The second layer introduces a soft voting ensemble mechanism, which weights and fuses the predicted probabilities output by the first layer to construct a voting ensemble model.

[0089] ;

[0090] in, The voting ensemble model indicates that key features Predicted as category The probability, Indicates the first Each base learner will use key features Predicted as category The probability, The weight coefficients optimized by grid search satisfy... To enhance model stability and generalization ability, embodiments of this application employ a Bootstrap sampling strategy to construct multiple sub-training sets, train multiple sub-models in parallel, and reduce variance through ensemble analysis, as shown in the following equation:

[0091] ;

[0092] in, The number of sub-models, For the first Sub-model, The output of the feature representation architecture.

[0093] The respiratory infection site prediction model and respiratory infection pathogen prediction model in the embodiments of this application both use a feature representation architecture, only the input and output dimensions are different.

[0094] S2. Construct an auxiliary diagnostic model for respiratory tract infection pathogens. The auxiliary diagnostic model for respiratory tract infection pathogens includes a trained respiratory infection site prediction model, a trained respiratory infection pathogen prediction model, a trained mixed infection prediction model, and a trained severe disease risk prediction model. The respiratory infection site prediction model adopts a feature representation architecture, and the respiratory infection pathogen prediction model is a loop structure containing the feature representation architecture.

[0095] In a specific embodiment, the respiratory infection pathogen prediction model adopts a K-layer recurrent pathogen prediction architecture. This architecture uses a feature representation framework. In the first round, the infection site prediction result, key features, and zero vector are concatenated and input into the first-layer pathogen prediction architecture to obtain the pathogen prediction probability for the first round. In the current round, the infection site prediction result, key features, and the pathogen prediction probability from the previous round are concatenated and input into the current round's pathogen prediction architecture to obtain the pathogen prediction probability for the current round, as shown in the following formula:

[0096] ;

[0097] ;

[0098] in, This represents the pathogen prediction architecture in the k-th round. This represents the pathogen prediction probability output by the pathogen prediction architecture in the k-th round. It is the pathogen prediction probability output by the pathogen prediction architecture in the (k-1)th round. Indicates key features, Indicates the predicted site of infection; This represents the feature obtained by concatenating the infection site prediction result, key features, and pathogen prediction probability of round k-1 in the k-th round; the pathogen prediction probability of round k is taken as the pathogen prediction result.

[0099] The weighted binary cross-entropy loss function is used in the training process of the respiratory infection pathogen prediction model, as shown in the following formula:

[0100] ;

[0101] Where N represents the number of samples in the training data. This represents the true label of the j-th pathogen in the i-th sample of the training data. This represents the probability that the respiratory infection pathogen prediction model predicts the j-th pathogen for the i-th sample. is the weighting coefficient for the j-th type of pathogen.

[0102] In a specific embodiment, the pathogen prediction results output by the trained respiratory infection pathogen prediction model are input into the mixed infection prediction model, and the mixed infection prediction probability is calculated using the following formula:

[0103] ;

[0104] in, It is the Sigmoid activation function. This represents the first learnable parameter. Key features The corresponding predicted probability of the existence of the j-th pathogen. Key features The corresponding predicted probability of mixed infection;

[0105] The prediction results of the infection site and pathogen are input into the severe illness risk prediction model, and the predicted probability of severe illness risk is calculated using the following formula:

[0106] ;

[0107] in, and These represent the predicted infection site and the predicted pathogen, respectively. and This represents the second learnable parameter. For random noise, Indicates the probability of predicting the risk of severe illness;

[0108] During the training process of the mixed infection prediction model and the severe illness risk prediction model, the parameters of the trained respiratory infection site prediction model and the trained respiratory infection pathogen prediction model are fixed, and the first learning parameter and the second learnable parameter in the mixed infection prediction model and the severe illness risk prediction model are trained respectively.

[0109] Specifically, embodiments of this application transform the multi-label classification task of respiratory infections into multiple binary correlation sub-tasks. A respiratory infection pathogen-assisted diagnostic model is constructed based on a total of 42 key indicators. This model is used to complete four tasks: respiratory infection site prediction, respiratory infection pathogen prediction, mixed infection prediction, and severe illness risk prediction, corresponding to the respiratory infection site prediction model, respiratory infection pathogen prediction model, mixed infection prediction model, and severe illness risk prediction model, respectively. For the respiratory infection site prediction model, a feature representation architecture is adopted. The key features, represented as a 42-dimensional vector, composed of the 42 key indicators, are input into the model to predict the infection site, resulting in a 3-dimensional vector. This infection site prediction result corresponds to the classification result of whether the lungs, upper respiratory tract, and bronchi are infected.

[0110] The infection site prediction results output by the respiratory infection site prediction model are used as important features and input into the respiratory infection pathogen prediction model to achieve multi-level feature fusion. The respiratory infection pathogen prediction model is implemented through a K-layer recurrent pathogen prediction architecture. The 3D infection site prediction results and 42-dimensional key features are input into the respiratory infection pathogen prediction model. In the first round, the 3D infection site prediction results and 42-dimensional key features are concatenated with the zero vector and passed through the first round pathogen prediction architecture to obtain the pathogen prediction probability represented as a 22-dimensional vector in the first round. In subsequent iterations, the 22-dimensional pathogen prediction probability obtained in the previous round is concatenated with the 3D infection site prediction results and 42-dimensional key features and then input into the pathogen prediction architecture of the current round to obtain the 22-dimensional pathogen prediction probability of the current round. The above process is repeated, and the 22-dimensional pathogen prediction probability of the last round is used as the pathogen prediction result. The pathogen prediction results include a classification structure for whether 22 pathogens have caused infection. The parameters of the pathogen prediction architecture used in each round are determined separately during training, so the parameters are not the same.

[0111] Traditional models predict each pathogen independently, neglecting ecological relationships such as co-occurrence, competition, or exclusion among microorganisms. In the embodiments of this application, the pathogen prediction probability from the previous round... This is input as a "memory vector." This vector encodes the model's current overall assessment of the probability of the presence of all 22 pathogens. In the next round, the respiratory infection pathogen prediction model can learn the inter-pathogen relationship patterns implicit in this "memory vector," thereby making more synergistic predictions that align with clinical realities and microbiological principles.

[0112] The standardized dataset was divided into a training set and a validation set in a 7:3 ratio. The training set was used for model training, and the validation set was used to evaluate model performance and prevent overfitting. This respiratory pathogen prediction model uses weighted binary cross-entropy as the loss function for this multi-label classification task. The weight coefficients of the j-th pathogen in this loss function are... To mitigate class imbalance, weighting is typically based on the proportion of pathogens in the training set (e.g., inverse frequency weighting). The AdamW optimizer is employed, which introduces weight decay on top of the Adam algorithm to improve generalization and alleviate overfitting. Dropout and L2 regularization are used in the network; during training, the validation set loss is monitored, and training is stopped early if there is no decrease for five consecutive epochs to prevent overfitting.

[0113] To determine mixed infections, embodiments of this application obtain a predicted probability of mixed infection by jointly inferring the probabilities of infection with different pathogens. This predicted probability is then compared with a mixed infection threshold. If the predicted probability is greater than the threshold, a mixed infection is confirmed, and the top M pathogens from the pathogen prediction results output by the respiratory infection pathogen prediction model are output as the mixed infection result. If the predicted probability is less than or equal to the threshold, a single infection is determined, not a mixed infection. Therefore, the pathogen with the highest probability value from the pathogen prediction results output by the respiratory infection pathogen prediction model is selected. The system outputs the infection site prediction result and combines it with the site with the highest probability from the infection site prediction result as the infection site output. Finally, by embedding the infection site prediction result output by the respiratory infection site prediction model and the pathogen prediction result output by the respiratory infection pathogen prediction model, the embodiments of this application construct a severe illness risk prediction model. This model adopts a multimodal logistic regression framework to predict the probability of severe illness risk. The predicted probability of severe illness risk is compared with a severe illness risk threshold. If the predicted probability of severe illness risk is greater than the severe illness risk threshold, it is determined that there is a risk of severe illness; if the predicted probability of severe illness risk is less than or equal to the severe illness risk threshold, it is determined that there is no risk of severe illness. This comprehensive architecture significantly improves the model's discrimination ability and system interpretability in complex multi-label scenarios, and is particularly suitable for the comprehensive diagnosis and treatment of pediatric respiratory infections.

[0114] S3. Obtain key features related to respiratory infection in the person to be diagnosed and input them into the respiratory infection pathogen-assisted diagnostic model. The key features are first processed by a trained respiratory infection site prediction model to obtain infection site prediction results. The infection site prediction results and key features are then input into a trained respiratory infection pathogen prediction model to obtain pathogen prediction results. The pathogen prediction results are then input into a trained mixed infection prediction model to obtain mixed infection prediction probability. The infection site prediction results and pathogen prediction results are then input into a trained severe illness risk prediction model to obtain severe illness risk prediction probability. Based on the mixed infection prediction probability and severe illness risk prediction probability, it is determined whether the person to be diagnosed has mixed infection and severe illness risk, respectively. The pathogen and infection site of the person to be diagnosed are confirmed by combining the infection site prediction results and pathogen prediction results.

[0115] In specific embodiments, the key features are multidimensional features composed of key indicators, including age, sex, alanine aminotransferase, aspartate aminotransferase, the ratio of aspartate aminotransferase to alanine aminotransferase, creatine kinase, direct bilirubin, total bilirubin, lactate dehydrogenase, total protein, albumin, globulin, the ratio of albumin to globulin, alkaline phosphatase, gamma-glutamyl transferase, potassium, sodium, chloride, phosphorus, blood urea nitrogen, creatinine, uric acid, carbon dioxide binding capacity, triglycerides, neutrophils, monocytes, monocyte ratio, eosinophils, eosinophil ratio, basophils, basophil ratio, neutrophil ratio, lymphocyte ratio, red blood cell count, hemoglobin, mean corpuscular volume, white blood cells, platelets, mean corpuscular hemoglobin concentration, red blood cell distribution width, mean platelet volume, and platelet distribution width.

[0116] Specifically, the respiratory infection pathogen-assisted diagnostic model is deployed and inferred. In the embodiments of this application, the collected indicator parameters are first acquired and the data is preprocessed to obtain the preprocessed indicator parameters, which include physiological indicators, biochemical indicators, and blood routine indicators; the physiological indicators include age and gender, etc.

[0117] Biochemical tests: Venous blood samples are collected from subjects in anticoagulant tubes containing EDTA. The samples are gently inverted 8-10 times to avoid hemolysis and clotting. Within 2 hours of collection, the samples are centrifuged at 3000 rpm for 10 minutes at 4°C to obtain plasma. The plasma is immediately aliquoted and stored at -80°C until testing. Before testing, frozen plasma samples are slowly thawed at 4°C, thoroughly mixed, and placed in the sample tray of the fully automated biochemical analyzer. The system automatically aspirates samples according to a preset program, mixes them with the corresponding reagents, measures absorbance changes at a specific wavelength, and calculates the concentrations of various biochemical indicators. Key indicators include, but are not limited to, liver function (e.g., ALT, AST), kidney function (e.g., creatinine, blood urea nitrogen), blood lipids (total cholesterol, triglycerides), blood glucose, and C-reactive protein (CRP). Each batch of samples is accompanied by high and low quality control samples to ensure instrument accuracy and result reliability. Test data are automatically collected and transmitted to the data analysis terminal via a Laboratory Information Management System (LIMS). Professional statistical software or laboratory-developed systems are used for data cleaning, standardization, and outlier identification, ultimately generating a structured report that includes sample number, test items, measured values, reference intervals, and outlier indicators.

[0118] Complete blood count (CBC): Venous blood samples were collected from the subjects using vacuum blood collection tubes (containing EDTA anticoagulant). After collection, the tubes were gently inverted 8-10 times to ensure thorough mixing of the anticoagulant and blood, while strictly avoiding vigorous shaking that could cause hemolysis. The samples were tested within 2 hours of collection. The mixed anticoagulant whole blood sample was placed on the sample rack of a fully automated five-part differential hematology analyzer. The instrument automatically aspirated the sample and used various detection principles, including sheath flow impedance, laser scattering, and fluorescence staining, to count and classify blood cells. Key parameters included white blood cell (WBC) count and five-part differential percentage, red blood cell (RBC) count, hemoglobin (HGB) concentration, hematocrit (HCT), platelet (PLT) count, and calculation of relevant red blood cell indices (such as MCV, MCH, and MCHC). The analytical process strictly adhered to laboratory quality control standards. Two levels of quality control materials were used daily for instrument calibration and verification to ensure the accuracy and stability of the test results.

[0119] Data cleaning and normalization: Based on the above data, data cleaning and normalization are crucial steps in data analysis, aiming to ensure data quality and consistency for subsequent analysis and interpretation. Missing values ​​are handled using the KNN imputation interpolation method. If there are too many missing values, the feature or sample is removed to avoid biasing the analysis results. The data is logarithmically transformed, and each value is summed to allow for zero values. Then, it is scaled to unit variance centered on the mean to reduce skewness. This ensures that each feature has a mean of 0 and a standard deviation of 1.

[0120] Data Balancing and Oversampling: When data is imbalanced (i.e., the ratio of positive to negative results differs significantly), model performance may be biased and problematic. Therefore, to optimize the machine learning algorithm of the embodiments of this application, an oversampling technique is applied using Synthetic Minority Oversampling (SMOTE) to balance the dataset. This technique has been shown to help improve model accuracy without biasing research results. SMOTE is a statistical technique that increases the number of cases in a dataset to balance the dataset. It achieves this by adding new cases to the minority of cases without affecting the number of the majority of cases. This algorithm samples the feature space of each target class and its 5 nearest neighbors, and then generates new cases that combine the features of the target case with the features of its nearest neighbors.

[0121] Specifically, it generates new data from sample data by oversampling a small sample set. The synthesis strategy finds sample X from the interface residuals. i k nearest neighbors, using X i(near) Let near ∈ {1,…,k'}. Then, a sample X is generated by KNN. i(nn) Combine a random number ζ1 between 0 and 1 to create a new sample X. i1 .

[0122] ;

[0123] Repeat the above process N' times until N' new samples are obtained: X i(new) , new∈1,...,N'. This method increases the percentage of the minority class in the dataset. SMOTE is applied only to the training set and does not oversample the test set, thus preserving the natural frequency of results.

[0124] To find a model that provides optimal predictive performance for respiratory infection pathogens, embodiments of this application combine two public intensive care databases and three real-world datasets, including populations without significant serious illnesses and patients diagnosed with respiratory infections (pneumonia, upper respiratory tract infections, etc.). Features, after data cleaning and normalization, are used for feature selection in model construction. Feature selection is a crucial step in building an efficient predictive model, significantly improving its performance and interpretability, and feature importance helps in understanding the clinical significance of features. The feature selection steps are as follows:

[0125] 1) Multicollinearity screening:

[0126] Since multicollinearity between features can affect prediction accuracy, Spearman correlation analysis removes the feature with lower correlation to the result from the dataset when two features are highly correlated (correlation coefficient > 0.9).

[0127] 2) Machine learning feature selection:

[0128] Data from healthy individuals and patients with respiratory infections were grouped, and a random forest model was used for feature selection. To select features using the random forest method, the contribution of each feature to reducing the model's impurity must first be calculated. Specifically, the feature importance in a random forest is calculated by examining the contribution of a feature's split nodes in the tree to reducing model impurity. This process involves multiple steps and ultimately helps to select the most representative features.

[0129] First, embodiments of this application train a random forest model, which consists of a large number of decision trees, each built on a different subset of samples to reduce overfitting. During the construction of the random forest model, each time the optimal split node is selected, the reduction in impurity for each candidate feature is evaluated. The primary impurity metric used here is Gini impurity, which reflects the purity of the samples in a node. For each feature, the change in Gini impurity (i.e., the reduction in impurity) before and after splitting using that feature across all nodes in all decision trees is calculated. These reductions in impurity are accumulated across the entire decision tree and then averaged across the entire forest to obtain the Mean Decrease Gini value for that feature. Embodiments of this application select the top 500 features with the highest Mean Decrease Gini values ​​for further analysis.

[0130] 3) SHAP Feature Filtering:

[0131] The SHAP (Shapley Additive Explanations) method is used to rank the importance of input features and interpret the results of predictive models, aiming to alleviate the "black box" problem often associated with machine learning (ML) models. This study employs SHAP to analyze the interpretability of predictions from a respiratory pathogen-assisted diagnostic model, quantifying the contribution of each input feature to individual predictions. Based on Shapley value theory in cooperative game theory, this method systematically calculates the marginal contribution of each feature across different feature subsets and performs a weighted average, thus fairly distributing the difference between the model's predictions and the baseline predictions. SHAP analysis provides both global and local interpretability: at the global level, it assesses the ranking of the importance of clinical and laboratory features in disease diagnosis and the direction of their non-linear effects; at the local level, it generates visual explanatory diagrams for individual samples, clearly demonstrating how each feature synergistically leads to the final diagnosis. This method effectively enhances the transparency and clinical credibility of the model's decision-making process, providing doctors with intuitive decision support and supporting the validation and optimization of the model's inference logic.

[0132] Finally, by integrating all the analysis results and taking the intersection of the features selected by the three screening methods, a set of features related to respiratory tract infection was determined and applied to the construction and optimization of a respiratory tract infection pathogen-assisted diagnostic model. Based on multiple screening methods, a final feature set was selected, resulting in 42 key indicators related to respiratory tract infections. These key features were then constructed, optimized, and applied to the model to provide more accurate early diagnosis. This process not only helps clinicians better understand the indicative role of routine examination results in respiratory tract infections but also provides potential biomarkers for personalized health management and disease intervention.

[0133] The multi-level intelligent diagnostic system constructed in this invention integrates various machine learning algorithms and is mainly used for the following three closely related clinical tasks to achieve comprehensive assessment and precise management of respiratory infections in children:

[0134] 1. Respiratory Infection Site Diagnosis: This respiratory infection site prediction model, based on key input features, uses an ensemble tree model (such as XGBoost and LightGBM) and a soft voting fusion strategy to determine the specific site of infection, such as the upper respiratory tract, bronchi, or lungs. Its output is a probability distribution for different sites. This probability distribution not only supports preliminary diagnosis but also serves as key feature input to subsequent respiratory infection pathogen prediction models and severe illness risk prediction models, achieving multi-level information fusion and context-enhanced inference.

[0135] 2. Diagnosis of Respiratory Infection Pathogens and Mixed Infections: This respiratory infection pathogen prediction model employs a multi-label classification architecture, using multiple binary classifiers trained in parallel to identify bacteria, viruses, and atypical pathogens. A mixed infection prediction model is specifically designed, dynamically integrating upstream infection site prediction results with current pathogen prediction results to comprehensively determine the existence of multiple infections. This method significantly improves the ability to identify complex infection scenarios and provides crucial decision-making support for rational drug use.

[0136] 3. Severe Illness Diagnosis: Combining the prediction results of infection site and pathogen, this severe illness risk prediction model can identify high-risk children requiring ICU intervention at an early stage, outputting their probability of developing severe illness, and assisting doctors in prioritizing treatment and allocating resources. The model provides visual interpretations of the results, enhancing its clinical credibility and practical value.

[0137] In summary, the embodiments of this application, through a three-level progressive inference mechanism, realize full-process assisted diagnosis from infection site localization and pathogen identification to risk stratification, significantly improving the overall efficiency and accuracy of pediatric respiratory tract infection diagnosis and treatment.

[0138] The final results show that the respiratory infection pathogen-assisted diagnostic model proposed in the embodiments of this application can accurately distinguish 22 pathogen subtypes and outperforms traditional models in identifying single and mixed infections. In internal and external validation cohorts, the model's mean AUCs for viral, bacterial, and fungal infections were 0.875, 0.860, and 0.903, respectively. For example, influenza virus exhibited extremely high classification performance (AUC = 0.946; Sn: 0.880; Sp: 0.860), and also showed good performance in the additional severe illness prediction model (AUC = 0.900). The respiratory infection pathogen-assisted diagnostic model proposed in the embodiments of this application demonstrates the potential for accurate pathogen identification, which can help physicians improve clinical efficiency and personalize the management of pediatric respiratory infections.

[0139] Table 1. Model performance in identifying viral, bacterial, and fungal subtypes.

[0140] Pathogens Sn Sp Youden's index ACC F1 score AUC AUCexternal1 AUCexternal2 MP 0.820(0.794-0.832) 0.770(0.742-0.794) 0.590(0.547-0.619) 0.797(0.778-0.813) 0.744(0.724-0.766) 0.888(0.874-0.899) 0.882(0.871-0.895) 0.792(0.779-0.807) Chlmydia 0.71(0.645-0.839) 0.71(0.687-0.726) 0.420(0.346-0.545) 0.709(0.687-0.727) 0.820(0.804-0.832) 0.806(0.761-0.860) 0.767(0.670-0.859) 0.631(0.548-0.716) HI 0.750(0.685-0.795) 0.700(0.670-0.720) 0.450(0.386-0.510) 0.703(0.679-0.726) 0.808(0.790-0.827) 0.798(0.766-0.832) 0.738(0.706-0.746) 0.651(0.614-0.693) SP 0.690(0.605-0.758) 0.680(0.652-0.686) 0.370(0.260-0.427) 0.674(0.650-0.681) 0.789(0.772-0.795) 0.769(0.703-0.800) 0.676(0.633-0.698) 0.661(0.629-0.705) SA 0.811(0.740-0.871) 0.790(0.770-0.806) 0.601(0.532-0.669) 0.792(0.774-0.809) 0.876(0.864-0.888) 0.883(0.856-0.911) 0.798(0.782-0.820) 0.690(0.634-0.771) KLP 0.86(0.787-0.912) 0.800(0.786-0.816) 0.660(0.580-0.712) 0.807(0.791-0.818) 0.887(0.875-0.895) 0.887(0.842-0.922) 0.845(0.822-0.902) 0.846(0.700-0.974) PA 0.900(0.796-1.00) 0.890(0.872-0.903) 0.790(0.683-0.888) 0.888(0.871-0.902) 0.939(0.929-0.947) 0.935(0.897-0.964) 0.826(0.764-0.892) 0.879(0.781-0.941) BP 0.902(0.771-0.972) 0.801(0.779-0.820) 0.703(0.557-0.712) 0.801(0.782-0.821) 0.886(0.875-0.900) 0.924(0.895-0.948) 0.886(0.851-0.925) 0.763(0.640-0.870) Ab 0.860(0.700-0.962) 0.830(0.810-0.853) 0.690(0.515-0.786) 0.830(0.811-0.852) 0.906(0.894-0.919) 0.930(0.896-0.952) 0.834(0.706-0.933) 0.811(0.756-0.868) MC 0.831(0.667-0.960) 0.830(0.806-0.842) 0.661(0.504-0.786) 0.826(0.808-0.843) 0.904(0.891-0.913) 0.904(0.826-0.983) 0.764(0.725-0.831) 0.804(0.694-0.909) E. coli 0.820(0.667-0.950) 0.780(0.762-0.803) 0.600(0.453-0.734) 0.782(0.763-0.803) 0.876(0.863-0.889) 0.903(0.8490.949) 0.851(0.811-0.880) 0.882(0.797-0.940) RSV 0.810(0.774-0.851) 0.810(0.772-0.817) 0.620(0.569-0.656) 0.802(0.779-0.821) 0.862(0.847-0.878) 0.896(0.874-0.915) 0.873(0.839-0.898) 0.775(0.725-0.816) IFV 0.880(0.802-0.923) 0.860(0.849-0.874) 0.740(0.654-0.784) 0.863(0.850-0.872) 0.922(0.913-0.928) 0.946(0.921-0.963) 0.942(0.922-0.963) 0.832(0.815-0.845) HAdV 0.750(0.692-0.802) 0.750(0.729-0.774) 0.500(0.429-0.553) 0.752(0.728-0.774) 0.839(0.822-0.854) 0.83(0.796-0.855) 0.730(0.688-0.762) 0.817(0.779-0.846) HRV 0.710(0.668-0.756) 0.700(0.678-0.723) 0.410(0.350-0.452) 0.700(0.679-0.721) 0.781(0.763-0.800) 0.797(0.759-0.818) 0.724(0.693-0.745) 0.679(0.615-0.728) HCoV 0.810(0.746-0.887) 0.800(0.779-0.821) 0.610(0.549-0.695) 0.798(0.781-0.816) 0.878(0.866-0.891) 0.904(0.877-0.929) 0.824(0.792-0.848) 0.817(0.790-0.848) HPIV 0.730(0.667-0.799) 0.720(0.698-0.745) 0.450(0.384-0.527) 0.719(0.698-0.735) 0.820(0.806-0.832) 0.812(0.769-0.847) 0.749(0.679-0.821) 0.661(0.604-0.729) HMPV 0.760(0.646-0.828) 0.760(0.735-0.774) 0.520(0.404-0.587) 0.757(0.736-0.774) 0.855(0.840-0.867) 0.837(0.779-0.887) 0.820(0.731-0.872) 0.827(0.756-0.896) HBoV 0.770(0.620-0.871) 0.730(0.704-0.746) 0.500(0.349-0.597) 0.728(0.709-0.751) 0.836(0.822-0.850) 0.838(0.790-0.888) 0.808(0.731-0.893) 0.796(0.720-0.868) Aspergillus 0.880(0.750,1.00) 0.810(0.796-0.835) 0.690(0.567-0.821) 0.814(0.567-0.821) 0.896(0.886-0.908) 0.908(0.878-0.935) 0.740(0.723-0.757) 0.790(0.725-0.886) Candida 0.842(0.746-0.919) 0.840(0.820-0.851) 0.682(0.597-0.775) 0.838(0.821-0.852) 0.909(0.898-0.917) 0.894(0.850-0.933) 0.747(0.665-0.826) 0.791(0.647-0.919) Otherfungus 0.770(0.571-1.00) 0.74(0.723-0.763) 0.510(0.300-0.738) 0.743(0.722-0.762) 0.851(0.838-0.864) 0.814(0.732-0.903) 0.800(0.676-0.909) 0.783(0.744-0.842)

[0141] Further reference Figure 3 As an implementation of the methods shown in the above figures, this application provides an embodiment of an auxiliary diagnostic device for respiratory tract infection pathogens, which is similar to... Figure 1 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices.

[0142] This application provides an auxiliary diagnostic device for respiratory tract infection pathogens, including:

[0143] Architecture building module 1 is configured to build a feature representation architecture based on ensemble learning. The feature representation architecture includes several voting ensemble models. The prediction results of all voting ensemble models are fused using a soft voting method to obtain the output result of the feature representation architecture.

[0144] Model building module 2 is configured to build a respiratory infection pathogen auxiliary diagnostic model. The respiratory infection pathogen auxiliary diagnostic model includes a trained respiratory infection site prediction model, a trained respiratory infection pathogen prediction model, a trained mixed infection prediction model, and a trained severe disease risk prediction model. The respiratory infection site prediction model adopts a feature representation architecture, and the respiratory infection pathogen prediction model is a loop structure containing the feature representation architecture.

[0145] Prediction module 3 is configured to acquire key features related to respiratory infection in the person to be diagnosed and input them into the respiratory infection pathogen-assisted diagnostic model. The key features are first processed by a trained respiratory infection site prediction model to obtain infection site prediction results. The infection site prediction results and key features are then input into a trained respiratory infection pathogen prediction model to obtain pathogen prediction results. The pathogen prediction results are then input into a trained mixed infection prediction model to obtain mixed infection prediction probability. The infection site prediction results and pathogen prediction results are then input into a trained severe illness risk prediction model to obtain severe illness risk prediction probability. Based on the mixed infection prediction probability and the severe illness risk prediction probability, it is determined whether the person to be diagnosed has mixed infection and severe illness risk, respectively. The pathogen and infection site of the person to be diagnosed are confirmed by combining the infection site prediction results and the pathogen prediction results.

[0146] Figure 4 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present invention. For example... Figure 4 As shown, the electronic device in this embodiment includes a processor 401 and a memory 402; wherein the memory 402 is used to store computer execution instructions; and the processor 401 is used to execute the computer execution instructions stored in the memory to implement the various steps performed by the electronic device in the above embodiment. For details, please refer to the relevant descriptions in the foregoing method embodiments.

[0147] Alternatively, the memory 402 can be either standalone or integrated with the processor 401.

[0148] When the memory 402 is set up independently, the electronic device also includes a bus 403 for connecting the memory 402 and the processor 401.

[0149] This invention also provides a computer storage medium storing computer execution instructions, which, when executed by processor 401, implement the above method.

[0150] This invention also provides a computer program product, including a computer program, which, when executed by a processor 401, implements the above-described method.

[0151] In the embodiments provided by this invention, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or modules, and may be electrical, mechanical, or other forms.

[0152] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to implement the solution of this embodiment according to actual needs.

[0153] Furthermore, the functional modules in the various embodiments of this invention can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The unit formed by the above modules can be implemented in hardware or in the form of hardware plus software functional units.

[0154] The integrated modules implemented as software functional modules described above can be stored in a computer-readable storage medium. These software functional modules, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor 401 to execute some steps of the methods of the various embodiments of this application.

[0155] It should be understood that the processor 401 described above can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor, or the processor 401 can be any conventional processor 401. The steps of the method disclosed in this invention can be directly manifested as the hardware processor 401 executing the steps, or as a combination of hardware and software modules within the processor 401 executing the steps.

[0156] The memory 402 may include high-speed RAM memory, and may also include non-volatile memory NVM, such as at least one disk storage device, and may also be a USB flash drive, portable hard drive, read-only memory, disk or optical disc, etc.

[0157] Bus 403 can be an Industry Standard Architecture (ISA), a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Bus 403 can be divided into address bus, data bus, control bus, etc. For ease of illustration, the bus 403 in the accompanying drawings of this application is not limited to only one bus 403 or one type of bus 403.

[0158] The aforementioned storage medium can be implemented from any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The storage medium can be any available medium accessible to general-purpose or special-purpose computers.

[0159] An exemplary storage medium is coupled to processor 401, enabling processor 401 to read information from and write information to the storage medium. Alternatively, the storage medium can be an integral part of processor 401. Processor 401 and storage medium can reside in application-specific integrated circuits (ASICs). Alternatively, processor 401 and storage medium can exist as discrete components in an electronic device or host device.

[0160] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0161] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for auxiliary diagnosis of respiratory tract infection pathogens, characterized in that, Includes the following steps: A feature representation architecture based on ensemble learning is constructed, comprising several voting ensemble models. The construction process of the voting ensemble models is as follows: Gradient boosting decision tree model, lightweight gradient boosting machine model, random forest model, and extreme gradient boosting model are selected as base learners. Each base learner is trained using a hyperparameter search mechanism based on a multi-objective co-evolutionary strategy to obtain trained base learners. The process of the hyperparameter search mechanism based on the multi-objective co-evolutionary strategy is as follows: By embedding the shared and unique hyperparameters of all base learners into a high-dimensional composite search space, a collaborative optimization objective function is constructed, as shown in the following equation: ； in, Represents the set of hyperparameters for all base learners. This represents the set of hyperparameters for the m-th base learner; Let represent the hyperparameters of the gradient boosting decision tree model, the lightweight gradient boosting machine model, the random forest model, and the extreme gradient boosting model, respectively. This represents the set of hyperparameters used when the maximum value is reached. This represents the dynamic weights of the m-th base learner. Represents the validation set. Let be the area under the curve performance metric of the m-th base learner on the validation set. The regularization coefficient is . The regularization term, which characterizes the hyperparameter complexity, is expressed as follows: ； in, These represent the first, second, and third levels of punishment, respectively. The m-th base learner represents the set of hyperparameters of the m-th base learner. The values ​​of the hyperparameters, Represents the first base learner of all base learners. The baseline values ​​for each hyperparameter. Denotes the set of hyperparameters of the m-th base learner. The number of all hyperparameters, Represents the m-th base learner A function with a time complexity of . This represents the measurement of the m-th base learner. and the nth base learner The function for predicting variance is expressed as follows: ； in, Let represent the predicted probability vector of the m-th base learner on the validation set. This represents the Pearson correlation coefficient; In the t-th training epoch, its corresponding dynamic weights in the t-th training epoch. Calculated using the following formula: ； in, This represents the area under the curve (AUC) of the m-th base learner on the validation set during the t-th training epoch. This represents the area under the curve (AUC) of the nth base learner on the validation set during the t-th training epoch. Scaling factor For temperature parameters; An improved Bayesian optimization algorithm is used to perform a global search on the hyperparameter set in the high-dimensional composite search space, and multi-objective optimization is performed based on the collaborative optimization objective function to determine the optimal hyperparameter set. A soft voting ensemble mechanism is used to weight and fuse the predicted probabilities output by all trained base learners to form an initial voting ensemble model. Each trained base learner in the initial voting ensemble model is retrained using the bagging method to obtain several voting ensemble models; the prediction results of all voting ensemble models are fused using the soft voting method to obtain the output result of the feature representation architecture. A respiratory tract infection pathogen auxiliary diagnostic model is constructed, comprising a trained respiratory tract infection site prediction model, a trained respiratory tract infection pathogen prediction model, a trained mixed infection prediction model, and a trained severe illness risk prediction model. The respiratory tract infection site prediction model adopts the aforementioned feature representation architecture, and the respiratory tract infection pathogen prediction model is a recurrent structure containing the feature representation architecture. The respiratory tract infection pathogen prediction model employs a K-layer recurrent pathogen prediction architecture, which also uses the aforementioned feature representation architecture. In the first round, the infection site prediction result, key features, and zero vector are concatenated and input into the first-layer pathogen prediction architecture to obtain the pathogen prediction probability for the first round. In the current round, the infection site prediction result, key features, and the pathogen prediction probability from the previous round are concatenated and input into the current round pathogen prediction architecture to obtain the pathogen prediction probability for the current round, as shown in the following formula: ； ； in, This represents the pathogen prediction architecture in the k-th round. This represents the pathogen prediction probability output by the pathogen prediction architecture in the k-th round. It is the pathogen prediction probability output by the pathogen prediction architecture in the (k-1)th round. Indicates key features, Indicates the predicted site of infection; This represents the feature obtained by concatenating the infection site prediction result, key features, and pathogen prediction probability of round k-1 in the k-th round; the pathogen prediction probability of round k is taken as the pathogen prediction result. Key features related to respiratory infection in the person to be diagnosed are obtained and input into the respiratory infection pathogen-assisted diagnostic model. The key features are first processed by the trained respiratory infection site prediction model to obtain the infection site prediction result. The infection site prediction result and the key features are then input into the trained respiratory infection pathogen prediction model to obtain the pathogen prediction result. The pathogen prediction result is then input into the trained mixed infection prediction model to obtain the mixed infection prediction probability. The infection site prediction result and the pathogen prediction result are then input into the trained severe illness risk prediction model to obtain the severe illness risk prediction probability. Based on the mixed infection prediction probability and the severe illness risk prediction probability, it is determined whether the person to be diagnosed has a mixed infection and a severe illness risk, respectively. Finally, the pathogen and infection site of the person to be diagnosed are confirmed by combining the infection site prediction result and the pathogen prediction result.

2. The method for auxiliary diagnosis of respiratory tract infection pathogens according to claim 1, characterized in that, The weighted binary cross-entropy loss function is used during the training of the respiratory infection pathogen prediction model, as shown in the following equation: ； Where N represents the number of samples in the training data. This represents the true label of the j-th pathogen in the i-th sample of the training data. This represents the probability that the respiratory infection pathogen prediction model predicts the j-th pathogen for the i-th sample. is the weighting coefficient for the j-th type of pathogen.

3. The method for auxiliary diagnosis of respiratory tract infection pathogens according to claim 1, characterized in that, The pathogen prediction results output by the trained respiratory infection pathogen prediction model are input into the mixed infection prediction model, and the mixed infection prediction probability is calculated using the following formula: ； in, It is the Sigmoid activation function. This represents the first learnable parameter. Key features The corresponding predicted probability of the existence of the j-th pathogen. Key features The corresponding predicted probability of mixed infection; The predicted infection site and the predicted pathogen are input into the severe illness risk prediction model, and the predicted probability of severe illness risk is calculated using the following formula: ； in, and The predicted infection site and the predicted pathogen are respectively represented. and This represents the second learnable parameter. For random noise, Indicates the probability of predicting the risk of severe illness; During the training process of the mixed infection prediction model and the severe illness risk prediction model, the parameters of the trained respiratory infection site prediction model and the trained respiratory infection pathogen prediction model are fixed, and the first learning parameter and the second learnable parameter in the mixed infection prediction model and the severe illness risk prediction model are trained respectively.

4. The method for auxiliary diagnosis of respiratory tract infection pathogens according to claim 1, characterized in that, The key features are multidimensional features composed of key indicators, including age, sex, alanine aminotransferase, aspartate aminotransferase, the ratio of aspartate aminotransferase to alanine aminotransferase, creatine kinase, direct bilirubin, total bilirubin, lactate dehydrogenase, total protein, albumin, globulin, the ratio of albumin to globulin, alkaline phosphatase, gamma-glutamyl transferase, potassium, sodium, chloride, phosphorus, blood urea nitrogen, creatinine, uric acid, carbon dioxide binding capacity, triglycerides, neutrophils, monocytes, monocyte ratio, eosinophils, eosinophil ratio, basophils, basophil ratio, neutrophil ratio, lymphocyte ratio, red blood cell count, hemoglobin, mean corpuscular volume, white blood cells, platelets, mean corpuscular hemoglobin concentration, red blood cell distribution width, mean platelet volume, and platelet distribution width.

5. An auxiliary diagnostic device for respiratory tract infection pathogens, characterized in that, The method for auxiliary diagnosis of respiratory tract infection pathogens according to any one of claims 1-4 includes: The architecture building module is configured to build a feature representation architecture based on ensemble learning. The feature representation architecture includes several voting ensemble models. The prediction results of all voting ensemble models are fused using a soft voting method to obtain the output result of the feature representation architecture. The model building module is configured to build an auxiliary diagnostic model for respiratory tract infection pathogens. The auxiliary diagnostic model for respiratory tract infection pathogens includes a trained respiratory infection site prediction model, a trained respiratory infection pathogen prediction model, a trained mixed infection prediction model, and a trained severe illness risk prediction model. The respiratory infection site prediction model adopts the feature representation architecture, and the respiratory infection pathogen prediction model is a loop structure containing the feature representation architecture. The prediction module is configured to acquire key features related to respiratory infection in the person to be diagnosed and input them into the respiratory infection pathogen-assisted diagnostic model. The key features are first processed by the trained respiratory infection site prediction model to obtain an infection site prediction result. The infection site prediction result and the key features are then input into the trained respiratory infection pathogen prediction model to obtain a pathogen prediction result. The pathogen prediction result is then input into the trained mixed infection prediction model to obtain a mixed infection prediction probability. The infection site prediction result and the pathogen prediction result are then input into the trained severe illness risk prediction model to obtain a severe illness risk prediction probability. Based on the mixed infection prediction probability and the severe illness risk prediction probability, it is determined whether the person to be diagnosed has a mixed infection and a severe illness risk, respectively. The pathogen and infection site of the person to be diagnosed are confirmed by combining the infection site prediction result and the pathogen prediction result.

6. An electronic device, comprising: One or more processors; Storage device for storing one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-4.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-4.

Images