Clinical performance verification system based on infectious lung disease diagnosis model

Abstract

The invention discloses a clinical performance verification system based on an infectious lung disease diagnosis model, and relates to the technical field of medical field, and the clinical performance verification system comprises a data collection module which is used for collecting and processing data by adopting a standardization process according to data types of different dimensions, and carrying out cleaning and standardization processing on all multi-modal data, the system has the advantages that in the aspect of data collection, multi-dimensional data such as patient baseline information, laboratory detection and iconography are covered, standardization processing is carried out on all kinds of data, and the data collection efficiency is improved. For example, laboratory detection data are uniformly converted into international standard units, iconography data are stored in a DICOM format, metadata are added, data quality is ensured, the model evaluation module adopts a multi-modal fusion technology, image, text, structured data and other data are fused, and model problems can be found and optimized in time.

Description

Technical Field

[0001] The present invention relates to the technical field of medical care, and more specifically, to a clinical performance verification system based on an infectious pulmonary disease diagnosis model. Background Art

[0002] Currently, the clinical diagnosis and treatment of infectious pulmonary diseases face three core challenges, which are intertwined and together constitute a technical bottleneck that urgently needs to be broken through. First is the dilemma of "racing against time". Clinical practice of novel coronavirus infection has made us deeply realize that the time window from the onset of the first symptom in a patient to the need for life support is extremely limited, with an average of only 10 days of the golden treatment period. During this short period, clinicians need to quickly complete a series of key decisions such as disease condition assessment, prediction of disease progression, and formulation of treatment plans. Such urgent time pressure not only tests the clinicians' clinical judgment ability but also highlights the obvious deficiency of existing prediction tools in terms of response speed. Second is the complexity brought about by the "domino effect". Infectious pulmonary diseases often rapidly induce systemic inflammatory responses through a series of cascading reactions such as cytokine storms, and then lead to multiple organ dysfunction. Clinical data show that up to 30% of critically ill patients will develop refractory shock, and the fatality rate reaches 45%-60%. Facing such a complex pathophysiological process, it is difficult for a single-dimensional monitoring and prediction mode to accurately grasp the progression law of the disease. Finally is the contradiction between "individualization - standardization". Existing scoring systems such as APACHE II and SOFA adopt static comprehensive criteria and cannot fully consider the individual differences of patients. This "one-size-fits-all" assessment method not only has a time lag but also is difficult to capture the unique disease conditions of each patient, which greatly affects the accuracy and timeliness of clinical decisions. Therefore, we propose a clinical performance verification system based on an infectious pulmonary disease diagnosis model. Summary of the Invention

[0003] The purpose of the present invention is to provide a clinical performance verification system based on an infectious pulmonary disease diagnosis model.

[0004] To solve the problems raised in the above background art, the present invention provides the following technical solutions: A clinical performance verification system based on an infectious pulmonary disease diagnosis model, including

[0005] A data acquisition module, which adopts a standardized process to collect and process data for different types of data in different dimensions, and cleans and standardizes all multimodal data to ensure the consistency, integrity, and accuracy of the data;

[0006] A data management module, which constructs a standard database for infectious pulmonary diseases, sets different access levels according to different users, anonymizes the personal privacy information of patients, and at the same time uses encryption technology to ensure data security;

[0007] The model evaluation module constructs a model in layers around microbial detection and host information, uses multi-modal fusion technology to extract and fuse features from data such as images, texts, and structured data, and simultaneously uses self-attention mechanism to fuse multi-modal data, outputs the diagnosis of the cause and the category of pathogens through a multi-task learning framework, introduces a feature-based gradient boosting machine to enhance the model performance, and evaluates the model using indicators such as accuracy;

[0008] The result display and feedback module inputs the collected clinical data into the infectious pulmonary disease diagnosis model, compares the model output results with the actual situation, and calculates accuracy, precision, and recall to comprehensively understand the effectiveness and reliability of the model in the diagnosis of infectious pulmonary diseases. At the same time, the results of various indicators of model evaluation are visually displayed in the form of intuitive charts and reports, facilitating researchers and clinicians to quickly understand the model performance.

[0009] As a further solution of the present invention: the scope of data collected by the data collection module covers the patient's baseline information, laboratory test data, imaging data, intervention measures, and clinical outcomes. Among them, the patient's laboratory test data (such as C-reactive protein, procalcitonin, inflammatory factors, etc.) will be uniformly converted into international standard units, and the imaging data (such as CT and X-ray films) will be stored in DICOM format to ensure cross-hospital system compatibility. Metadata such as data source, collection time, and disease type will be attached during all data collection processes to facilitate subsequent data retrieval and correlation analysis. The specific steps of image data feature extraction are as follows:

[0010] Adjust the pixel values to the standard range through normalization processing to reduce the differences between images:

[0011]

[0012] Among them, I is the original image data, I min and I max are the minimum and maximum values of the image data respectively. The purpose of this operation is to provide stable output data for subsequent analysis. In addition, data augmentation technology is used to diversify the images.

[0013] As a further solution of the present invention: the model evaluation module includes a feature engineering module, a multi-modal fusion model module, and a prediction model module. The feature engineering module adopts data processing methods of multi-modal feature extraction, adaptive Fourier transform, and feature space fusion, and the method of multi-modal feature extraction is as follows;

[0014] When processing structured data and feature extraction, in order to ensure that data with different dimensions can be compared on the same scale, we perform standardization processing on these data;

[0015] X std = (X - μ X ) / σ X

[0016] where X represents the original data, μ X represents the mean, and σ X represents the standard deviation. In this standardization process, the dimensional differences between different features are effectively eliminated, enabling all data to be compared and modeled on a consistent basis;

[0017] For the standardized data, we adopted a variety of advanced machine learning methods for feature extraction. Among them, the multiple learning methods include algorithms such as the random forest model and the support vector machine, thereby identifying the most important features;

[0018] F ml = MLFeatureExtraction(X std )

[0019] where F ml is the extracted feature, MLFeatureExtraction is the machine learning feature extraction method, and X std is the structured data after standardization. These features are further input into the random forest model to extract the important features in the structured data, and then integrated and output:

[0020]

[0021] Integrate the multi-time point data from outpatient clinics, emergency departments, and ICUs. Subsequently, perform standardization or normalization processing on the integrated data. Then, extract the change trends and time series features, including heart rate changes and white blood cell counts. At the same time, GBM (Gradient Boosting Machine), an ensemble learning method, gradually reduces the prediction error. And in each round of training, GBM will fit the residuals generated by the previous round of the model to optimize the prediction ability of the model. The GBM model is trained with the goal of minimizing the loss function. At this time, assume the training set is X, the target variable is F, and the predicted value of the model is obtained according to the corresponding mechanism;

[0022]

[0023] where h m (X) is the m-th decision tree, and α m is the weight.

[0024] As a further solution of the present invention: the data management module adopts a distributed storage structure to ensure the security and efficient management of data, and supports multi-user and multi-hospital access simultaneously through network interconnection. In terms of data permission management, different access levels are set according to user roles, and patient personal privacy information is anonymized and encrypted for storage and transmission.

[0025] As a further solution of the present invention: the specific steps of the feature engineering module for processing and extracting time series data using adaptive Fourier decomposition are as follows;

[0026] For the structured data (time series data) statistically calculated daily, we apply adaptive Fourier decomposition (AFD) for feature extraction, and the daily statistical results are used as the input of the time series, specifically as follows;

[0027] X time-series =DailyStatistics(X)

[0028] Where X time-series represents the transformed time series data, and this formula expresses the process of converting the original structured data into time series data through daily statistical operations. AFDAFD decomposes the signal into multiple orthogonal basis functions recursively, and these basis functions are adaptively selected according to the spectral characteristics of the signal to ensure maximizing the extraction of key features in the signal. Among them, each basis function is adaptively selected according to the specific spectral characteristics of the signal. Denote the standardized time series data as the signal f(t). First, follow the maximum selection principle of energy tracking matching to select the initial basis function g0(t) that best matches the signal characteristics. Subsequently, calculate the projection of the signal on the initial basis function 〈f,g0〉, where the inner product here is the inner product in the Euclidean space, and thus calculate the residual signal:

[0029] r1(t)=f(t)-〈f,g0〉g0(t)

[0030] Repeat the above process to obtain a series of basis functions, calculate the projection and residual respectively until the stop condition is met (where the conditions met include but are not limited to the energy of the residual signal being lower than the threshold):

[0031] rk +1 (t)=r k (t)-〈r k ,g k 〉g k (t)

[0032] Finally, sum up the projection results of all basis functions to obtain the AFD reconstruction result of the original signal:

[0033]

[0034] In this way, the AFD extracts the main spectral features of the signal, reflecting the hidden periodic and aperiodic components in the signal. We further input the features extracted by the AFD into a Long Short-Term Memory (LSTM) model for learning and predicting time series features. The LSTM model can effectively handle long-term dependencies and capture the dynamic change features in the time series, thus improving the prediction ability for complex time series data, especially having significant advantages in biomedical signal analysis. The specific formula is as follows;

[0035] F afd = AFD(X time-series ), F afd-lstm = LSTM(F afd )

[0036] Among them, (X time-series ) represents the input time series data. AFD is a function name. F afd is the result obtained after processing the time series data by the AFD function. It may be the extracted feature vector or the data form after a certain transformation, and will be used as the input for the next calculation. LSTM represents the Long Short-Term Memory network, and F afd-lstm is the output result obtained by inputting F afd into the LSTM network.

[0037] As a further solution of the present invention: The multi-modal fusion model module is responsible for integrating data from different modalities, such as images, texts, time series, and structured data. At the same time, to achieve effective fusion of multi-modal data, a self-attention mechanism is adopted to integrate image features, text features, structured data features, and the associated features extracted by the prediction model and the large language model into a feature matrix;

[0038] F concat = [F img-reduced , F txt-fixed , F str-selected , F afd-lstm , F LLM

[0039] The feature fusion depends on the self-attention mechanism (Self-Attention). This mechanism determines the importance of features in different modalities for the current task by calculating the correlation weights between each feature and other features. The specific operations are as follows:

[0040] Q = W Q F concat , K = W K F concat , V = W V ​F concat

[0041] The query vector is used to characterize the association mode between the features of the current modality and other modalities. The key vector is used to capture the feature representations of other modalities, and the value vector represents the weights and information of the features themselves. At this time, the dot product operation between the query vector and the key vector reflects the correlation between the features of different modalities. After the dot product result is normalized by the Softmax function, the weights are obtained;

[0042]

[0043] Among them, Q represents the query vector, which is used to characterize the association mode between the features of the current modality and other modalities. K represents the key vector, which is used to capture the feature representations of other modalities. V represents the value vector, which represents the weights and information of the features themselves. QK T is expressed as the dot product operation of the transpose of Q and K, where d k is the dimension of K.

[0044] As a further solution of the present invention: the prediction model module adopts a multi-task learning framework, and in order to output two parts of the etiological diagnosis and the pathogen category, the total loss function of the multi-task learning includes the cross-entropy loss and the regression loss:

[0045] L = αL1 + βL2

[0046] Among them, L1 and L2 represent the cross-entropy loss, which is used for the classification task:

[0047]

[0048] Among them, y i is the true label of the sample, indicating the true situation of the sample belonging to a certain category, taking values of (0\1), and p i is the probability that the model predicts that the sample belongs to the jth category.

[0049] As a further solution of the present invention: when the result display and feedback module evaluates the diagnostic model of infectious pulmonary diseases based on multi-modal data, the key verification indicators include accuracy (Accuracy), precision (Precision), recall (Recall), and F1-score, as follows;

[0050] Accuracy (Accuracy), which represents the proportion of samples correctly classified by the model in the total samples, and the formula is:

[0051]

[0052] Among them, TP is the true positive, TN is the true negative, FP is the false positive, and FN is the false negative. The accuracy rate can quickly reflect the overall performance of the model;

[0053] Precision measures the proportion of samples predicted as positive by the model that are actually positive. The formula is:

[0054]

[0055] Recall represents the proportion of positive cases identified by the model among the actual positive cases. The formula is:

[0056]

[0057] F1-score is the harmonic mean that comprehensively considers precision and recall. The formula is:

[0058]

[0059] Through the comprehensive evaluation of the above indicators, the effectiveness and reliability of the model in the diagnosis of infectious pulmonary diseases can be comprehensively understood.

[0060] Adopting the above technical solution, compared with the prior art, the beneficial effects of the present invention are as follows:

[0061] 1. In data collection, the present invention covers multi-dimensional data such as patient baseline information, laboratory tests, and imaging. And various data are standardized. For example, laboratory test data are uniformly converted into international standard units, and imaging data are stored in DICOM format. At the same time, metadata are added to ensure data quality. The model evaluation module uses multi-modal fusion technology to fuse image, text, structured and other data, captures the correlation of different modal features by using the self-attention mechanism, and also introduces a feature-based gradient boosting machine to enhance the model performance. The result display and feedback module comprehensively evaluates the model through indicators such as accuracy rate, precision, recall, and F1-score, and can timely discover model problems and optimize them;

[0062] 2. The present invention focuses on key clinical decision-making nodes through the model evaluation module, provides accurate decision-making support for doctors, and integrates a dynamic adaptive prediction model established with multi-dimensional data, which can early warn high-risk patients, help doctors reasonably arrange medical resources, avoid resource waste and over-medical treatment, and at the same time provide strong decision-making support for hierarchical diagnosis and treatment, promote the rational allocation of medical resources, and promote the efficient operation of the medical system, generating important social and economic values. Description of the Drawings

[0063] Figure 1 It is a schematic diagram of the system process in the embodiment of the present invention. Detailed Embodiment

[0064] The following further describes the specific embodiments of the present invention in conjunction with the accompanying drawings. It should be noted here that the description of these embodiments is used to help understand the present invention, but does not constitute a limitation on the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0065] Embodiment 1. Please refer to the attached Figure 1 , the present invention provides a technical solution: a clinical performance verification system based on an infectious pulmonary disease diagnosis model, including

[0066] A data acquisition module, which adopts a standardized process to collect and process data for different dimensions of data types, and cleans and standardizes all multimodal data to ensure the consistency, integrity, and accuracy of the data;

[0067] A data management module, which constructs a standard database for infectious pulmonary diseases, sets different access levels according to different users, anonymizes the personal privacy information of patients, and uses encryption technology to ensure data security;

[0068] A model evaluation module, which hierarchically constructs a model around microbial detection and host information, uses multimodal fusion technology to extract and fuse features from data such as images, texts, and structures, and simultaneously uses a self-attention mechanism to fuse multimodal data, outputs the cause diagnosis and pathogen category through a multi-task learning framework, introduces a feature-based gradient boosting machine to enhance the model performance, and evaluates the model using indicators such as accuracy;

[0069] A result display and feedback module, which inputs the collected clinical data into the infectious pulmonary disease diagnosis model, compares the model output results with the actual situation, and calculates the accuracy, precision, and recall rate to comprehensively understand the effectiveness and reliability of the model in the diagnosis of infectious pulmonary diseases. At the same time, the results of various indicators of the model evaluation are visually displayed in the form of intuitive charts and reports, facilitating researchers and clinicians to quickly understand the model performance.

[0070] Please refer to the attached Figure 1 , the scope collected by the data acquisition module covers the patient's baseline information, laboratory test data, imaging data, intervention measures, and clinical outcomes. Among them, the patient's laboratory test data (such as C-reactive protein, procalcitonin, inflammatory factors, etc.) will be uniformly converted to international standard units, and the imaging data (such as CT and X-ray films) will be stored in DICOM format to ensure compatibility across hospital systems. Metadata such as data source, collection time, and disease type will be attached during all data collection processes to facilitate subsequent data retrieval and correlation analysis. The specific steps for feature extraction of imaging data are as follows:

[0071] The pixel values are adjusted to the standard range through normalization to reduce the differences between images:

[0072]

[0073] Among them, I is the original image data, I min and I max are the minimum and maximum values of the image data respectively. The purpose of this operation is to provide stable output data for subsequent analysis. In addition, data augmentation techniques are used to diversify the images. The model evaluation module includes a feature engineering module, a multi-modal fusion model module, and a prediction model module. The feature engineering module uses data processing methods such as multi-modal feature extraction, adaptive Fourier transform, and feature space fusion. The method of multi-modal feature extraction is as follows;

[0074] When processing structured data and feature extraction, in order to ensure that data with different dimensions can be compared on the same scale, we standardized these data;

[0075] X std =(X - μ X ) / σ X

[0076] Among them, X represents the original data, μ X represents the mean value, and σ X represents the standard deviation. In this standardization process, the dimensional differences between different features are effectively eliminated, enabling all data to be compared and modeled on a consistent basis;

[0077] For the standardized data, we used a variety of advanced machine learning methods for feature extraction. Among them, a variety of learning methods include algorithms such as random forest models and support vector machines to identify the most important features;

[0078] F ml = MLFeatureExtraction(X std )

[0079] Among them, F ml is the extracted feature, MLFeatureExtraction is the machine learning feature extraction method, and X std is the standardized structured data. These features are further input into the random forest model to extract the important features in the structured data, and then integrated and output:

[0080]

[0081] Integrate multi-timepoint data from outpatient clinics, emergency departments, and ICUs, and then perform standardization or normalization on the integrated data. Next, extract trend and time series features from it, including heart rate changes and white blood cell counts. At the same time, use GBM (Gradient Boosting Machine), an ensemble learning method, to gradually reduce the prediction error. And in each round of training, GBM will fit the residuals generated by the previous round of the model to optimize the model's prediction ability. The GBM model is trained with the goal of minimizing the loss function. At this time, assume the training set is X, the target variable is F, and the predicted value of the model is obtained according to the corresponding mechanism;

[0082]

[0083] where h m (X) is the m-th decision tree, and α m is the weight. The data management module adopts a distributed storage structure to ensure data security and efficient management, and supports multi-user and multi-hospital simultaneous access through network interconnection. And in terms of data permission management, different access levels are set according to user roles, and patient personal privacy information is anonymized and encrypted for storage and transmission. The feature engineering module uses adaptive Fourier decomposition for time series data processing and feature extraction. The specific steps are as follows;

[0084] For the structured data (time series data) statistically calculated daily, we apply adaptive Fourier decomposition (AFD) for feature extraction, and the daily statistical results are used as the input of the time series. Specifically as follows;

[0085] X time-series =DailyStatistics(X)

[0086] where X time-series represents the transformed time series data, and this formula expresses the process of converting the original structured data into time series data through daily statistical operations. And AFDAFD decomposes the signal into multiple orthogonal basis functions recursively. These basis functions are adaptively selected according to the spectral characteristics of the signal to ensure maximizing the extraction of key features in the signal. Among them, each basis function is adaptively selected according to the specific spectral characteristics of the signal. Denote the standardized time series data as the signal f(t). First, select the initial basis function g0(t) that best matches the signal characteristics according to the maximum selection principle of energy tracking matching. Subsequently, calculate the projection of the signal on the initial basis function 〈f, g0〉. Here, the inner product is the inner product in the Euclidean space. Thus, calculate the residual signal:

[0087] r1(t) = f(t) - 〈f, g0〉g0(t)

[0088] Repeat the above process to obtain a series of basis functions, and calculate the projection and residual respectively until the stop condition is met (where the conditions that can be met include but are not limited to the energy of the residual signal being lower than the threshold):

[0089] r k+1 (t) = r k (t) - 〈r k , g k 〉g k (t)

[0090] Finally, sum up the projection results of all basis functions to obtain the AFD reconstruction result of the original signal:

[0091]

[0092] In this way, AFD extracts the main spectral features of the signal, reflecting the hidden periodic and aperiodic components in the signal. We further input the features extracted by AFD into a long short-term memory network (LSTM) model for learning and predicting time series features. The LSTM model can effectively handle long-term dependencies and capture the dynamic change features in the time series, thereby improving the prediction ability for complex time series data, especially having significant advantages in biomedical signal analysis. The specific formula is as follows;

[0093] F afd = AFD(X time-series )、F afd-lstm = LSTM(F afd )

[0094] Among them, (X time-series ) represents the input time series data. AFD is a function name. F afd is the result obtained after processing the time series data by the AFD function. It may be the extracted feature vector or the data form after a certain transformation, which will be used as the input for the next calculation. lstm represents the long short-term memory network (Long-ShortTermMemorynetwork). F afd-lstm is the output result obtained by inputting F afd into the LSTM network.

[0095] In this embodiment, in order to deeply mine the key information in medical data, this study adopts large language model technology and focuses on the feature extraction of two types of data: one is the medical record text information, and the other is the laboratory test results. The large language model has powerful semantic understanding and reasoning capabilities, can automatically identify and extract the key features in different data sources, and associate this information to provide a high-dimensional and semantically rich feature representation for model training, as follows:

[0096] For medical record texts (such as symptoms, medical history, treatment plans, etc.), large language models can deeply understand the medical entities and their relationships in the text. For example, the model can identify the patient's symptoms, such as "fever, cough", and infer its possible causes and development trends. By combining external medical knowledge bases, the model can automatically supplement the implicit association information in the text and construct a feature representation with a deeper semantic level. The information extracted by the model from the medical record includes:

[0097] · Symptoms: such as "fever, shortness of breath"

[0098] · Medical history: such as "chronic obstructive pulmonary disease (COPD)"

[0099] · Medications: such as "azithromycin"

[0100] Combined with external retrieval, the model can associate this information. For example, infer the potential impact of medical history on medications, thereby generating comprehensive features. The output feature representation is:

[0101] F LLM-text = LLM(T medical , K external )

[0102] where T medical represents the medical record text data, and k external is the relevant information retrieved from external medical knowledge bases and the consultation experiences of multiple famous doctors. The feature vector generated in this process reflects both the explicit information in the medical record and the implicit associations inferred by the model;

[0103] Laboratory test results (such as C-reactive protein (CRP), procalcitonin (PCT), etc.) are important bases for diagnosing and evaluating the condition. The large language model standardizes these data and associates them with the information in the medical record text to generate a composite feature representation. For example, an increase in CRP and PCT may indicate inflammation or infection, and these data can be combined with relevant symptoms in the medical record (such as fever) to form a more diagnostically valuable feature vector. By combining laboratory data and medical record text, the large language model can capture the trends and potential associations of pathological changes, providing support for subsequent condition prediction. The output feature representation form is:

[0104] F LLM-lab = LLM(F CRP , F PCT K external )

[0105] F LLM = F LLM-text + F LLM-lab

[0106] where F CRP and FPCT Data features representing CRP and PCT, K external is the relevant information retrieved from external medical knowledge bases and the medical experience of multiple renowned doctors. The feature vectors generated in this process can integrate multiple data sources, providing a richer description of clinical features. Finally, the data features extracted by the two large models will be fused. To achieve an accurate diagnosis of infectious pulmonary diseases, a multi-task learning framework is adopted. To output two parts: the confirmed cause and the pathogen category, the total loss function of multi-task learning includes cross-entropy loss and regression loss:

[0107] L = αL1 + βL2

[0108] where L1 and L2 represent cross-entropy loss, used for classification tasks:

[0109]

[0110] We use the Adam optimizer for training. The Adam optimizer adjusts parameter updates through an adaptive learning rate, improving the convergence speed and stability of the model. The parameter update formula is as follows:

[0111]

[0112] where m t and v t are the first and second moment estimates of the gradient respectively, and the update rules are as follows:

[0113] m t = β1m t-1 +(1 - β1)g t

[0114]

[0115] where g t is the gradient, β1 and β1 are momentum terms, and η is the learning rate;

[0116] In addition, in addition to the core model, a Feature-based Gradient Boosting Machine (FGBM) is introduced to further enhance prediction accuracy and robustness, and it combines traditional gradient boosting and feature engineering, which is particularly suitable for dealing with complex multi-modal data;

[0117] By combining a decision tree (as a weak learner) with the features extracted by the self-attention layer, the non-linear relationships and interactions between features can be captured, which may not be fully expressed by linear models. FGBM makes the final prediction by combining the output results of multiple decision trees, reducing the bias and variance of the model, thus improving the accuracy and stability of the prediction. During the training process, the FGBM model updates the parameters by minimizing the loss function, usually in combination with gradient descent and regularization techniques to prevent overfitting. The decision trees in FGBM can capture the complex interactions between different modalities, improving the prediction ability of the model in practical applications.

[0118] When in use, in terms of data collection, it covers multi-dimensional data such as patient baseline information, laboratory tests, and imaging. And standardize various types of data. For example, laboratory test data is uniformly converted to international standard units, and imaging data is stored in DICOM format. At the same time, additional metadata is attached to ensure data quality. The model evaluation module uses multi-modal fusion technology to fuse data such as images, texts, and structured data, uses the self-attention mechanism to capture the correlation of features in different modalities, and also introduces a feature-based gradient boosting machine to enhance the model performance. The result display and feedback module comprehensively evaluates the model through indicators such as accuracy, precision, recall, and F1-score, and can timely detect model problems and optimize them.

[0119] Embodiment 2. Please refer to the appendix Figure 1 , the present invention provides a technical solution: a clinical performance verification system for an infectious pulmonary disease diagnosis model, including

[0120] A data collection module, for different types of data in different dimensions, uses a standardized process to collect and process the data, and cleans and standardizes all multi-modal data to ensure the consistency, integrity, and accuracy of the data;

[0121] A data management module, constructs a standard database for infectious pulmonary diseases, sets different access levels according to different users, anonymizes the personal privacy information of patients, and at the same time uses encryption technology to ensure data security;

[0122] A model evaluation module, constructs a model in layers around microbial detection and host information, uses multi-modal fusion technology to extract and fuse features from data such as images, texts, and structured data, at the same time uses the self-attention mechanism to fuse multi-modal data, outputs the etiological diagnosis and pathogen category through a multi-task learning framework, introduces a feature-based gradient boosting machine to enhance the model performance, and evaluates the model using indicators such as accuracy;

[0123] The result display and feedback module inputs the collected clinical data into the infectious pulmonary disease diagnosis model, compares the output result of the model with the actual situation, and calculates the accuracy, precision, and recall rate to comprehensively understand the effectiveness and reliability of the model in the diagnosis of infectious pulmonary diseases. At the same time, the results of various indicators evaluated by the model are visually displayed in the form of intuitive charts and reports, facilitating researchers and clinicians to quickly understand the model performance.

[0124] Please refer to the appendix Figure 1 , the multi-modal fusion model module is responsible for integrating data from different modalities, such as images, texts, time series, and structured data. At the same time, to achieve the effective fusion of multi-modal data, a self-attention mechanism is adopted to integrate the image features, text features, structured data features, and associated features extracted by the prediction model and the large language model into a feature matrix;

[0125] F concat =[F img-reduced ,F txt-fixed ,F str-selected ,F afd-lstm ,F LLM

[0126] The feature fusion depends on the self-attention mechanism (Self-Attention). This mechanism determines the importance of features from different modalities for the current task by calculating the correlation weights between each feature and other features. The specific operations are as follows:

[0127] Q=W Q F concat 、K=W K F concat 、V=W V F concat

[0128] The query vector is used to characterize the association mode of the features of the current modality with other modalities. The key vector is used to capture the feature representations of other modalities. The value vector represents the weights and information of the features themselves. At this time, through the dot product operation between the query vector and the key vector, the correlation between features from different modalities is reflected. After the dot product result is normalized by the Softmax function, the weights are obtained:

[0129]

[0130] Among them, Q represents the query vector, which is used to characterize the association mode of the features of the current modality with other modalities. K represents the key vector, which is used to capture the feature representations of other modalities. V represents the value vector, which represents the weights and information of the features themselves. QK T represents the dot product operation of the transpose of Q and K, where d k ​is the dimension of K. The prediction model module adopts a multi-task learning framework. In order to output two parts: etiological diagnosis and pathogen category, the total loss function of multi-task learning includes cross-entropy loss and regression loss:

[0131] L = αL1 + βL2

[0132] where L1 and L2 represent cross-entropy loss and are used for classification tasks:

[0133]

[0134] where y i is the true label of the sample, indicating the true situation of the sample belonging to a certain category, taking values (0\1), and p i is the probability that the model predicts the sample belongs to the j-th category. When the result display and feedback module evaluates the diagnostic model of infectious pulmonary diseases based on multi-modal data, the key verification indicators include accuracy, precision, recall, and F1-score, which are specifically as follows;

[0135] Accuracy, which represents the proportion of samples correctly classified by the model in the total samples. The formula is:

[0136]

[0137] where TP is the true positive, TN is the true negative, FP is the false positive, and FN is the false negative. Accuracy can quickly reflect the overall performance of the model;

[0138] Precision, which measures the proportion of samples actually being positive among the samples predicted as positive by the model. The formula is:

[0139]

[0140] Recall, which represents the proportion of positive cases identified by the model in the actual positive cases. The formula is:

[0141]

[0142] F1-score, which comprehensively considers the harmonic mean of precision and recall. The formula is:

[0143]

[0144] Through the comprehensive evaluation of the above indicators, the effectiveness and reliability of the model in the diagnosis of infectious pulmonary diseases can be comprehensively understood;

[0145] In this embodiment, the precision rate can evaluate the reliability of the model in diagnosing infectious pulmonary diseases, which helps reduce the risk of misdiagnosis; the recall rate represents the proportion of positive examples identified by the model among the actual positive examples; the F1-score balances the precision rate and the recall rate and is very suitable for paying attention to the performance of the model under imbalanced categories. By comprehensively evaluating these indicators, the effectiveness and reliability of the model in the diagnosis of infectious pulmonary diseases can be comprehensively understood. In the process of model construction, first, through feature extraction and multi-modal data fusion, core features are extracted from multi-dimensional data such as clinical and imaging data; at the same time, the medical record text information is combined with the extraction of imaging features to enhance the diagnostic and predictive capabilities of the model. Subsequently, the unified feature space integration method of multi-modal data is used to convert data of different modalities into a unified feature space and perform model parameter integration to ensure the overall performance and stability of the model. The model update and optimization module mainly uses the adaptive Fourier decomposition (AFD) method to capture key features in the data and combines multi-modal data fusion technology to achieve the integration of different data sources, thereby providing richer and more accurate inputs for the model. The model will output through multiple tasks. First, it will give the scoring prediction of the risk status of severe pneumonia and respiratory support. On the basis of this prediction output result, combined with the large language model and the expert knowledge graph, it will assist doctors in making precise decision support for ICU needs, help doctors judge whether the patient needs to adjust the medication, conduct dynamic evaluation and provide auxiliary decision-making, and allocate key ICU resources such as ventilators;

[0146] A multi-task learning framework is constructed based on multi-modal data for parallel execution of two core tasks: infectious pulmonary disease risk prediction and ICU requirement prediction. The input data of the model covers imaging information (such as CT or X-ray images), medical record texts (patient symptoms, medical history records), and laboratory test data (such as inflammatory markers like CRP, PCT, etc.). These multi-modal data are first deeply processed by a feature extraction module, and then the information from different data sources is fused through a self-attention mechanism to form a high-dimensional unified feature representation. Based on this unified representation, the model simultaneously learns and predicts for the two tasks. The first task is infectious pulmonary disease risk prediction: its goal is to first determine whether the patient is in a high-risk state of severe pneumonia. The model uses the fused multi-modal features, combines the lesions in the images, clinical symptom descriptions, and laboratory data, and outputs through a classifier whether the patient is in a high-risk state of severe pneumonia and whether ICU intervention is needed; at the same time, a score prediction for respiratory support is made based on multi-modal clinical information to evaluate the necessity of mechanical ventilation for the patient. This task realizes the rapid and accurate identification of diseases by comprehensively analyzing information from multiple data sources, helping doctors grasp the patient's pulmonary health status in a timely manner. The second task is decision support based on Task 1: On the basis of the output result of Task 1, combined with a large language model and an expert knowledge graph, it provides precise decision support for doctors' ICU requirements. The large language model semantically analyzes the prediction results provided by Task 1 and combines them with clinical experience and best practices in the knowledge graph to provide intuitive and interpretable decision-making suggestions. This task can help doctors judge whether the patient needs to adjust the medication, dynamically evaluate and assist in decision-making, allocate key ICU resources such as ventilators, and optimize resource allocation, thereby improving the treatment efficiency and ensuring that patients receive timely and appropriate medical interventions.

[0147] When in use, the model evaluation module focuses on key clinical decision-making nodes, provides precise decision support for doctors, and the dynamic adaptive prediction model established by integrating multi-dimensional data can give early warnings to high-risk patients, helping doctors reasonably arrange medical resources, avoid resource waste and over-medical treatment, and at the same time provide strong decision support for hierarchical medical treatment, promote the rational allocation of medical resources, and drive the efficient operation of the medical system, generating important social and economic values.

[0148] Working principle:

[0149] First, using a standardized process, collect multi-dimensional data such as the patient's baseline information, laboratory tests, and imaging. Convert laboratory test data into international standard units, store imaging data in DICOM format, and attach metadata at the same time. Clean and standardize all multi-modal data to ensure data quality, build a standard database for infectious pulmonary diseases, adopt a distributed storage structure, set access levels according to different user roles, anonymize and encrypt patient privacy information to ensure data security and efficient management. Subsequently, around microbial detection and host information, build models in layers, use multi-modal fusion technology, combine self-attention mechanisms, extract and fuse features from multi-source data such as images and texts, use a multi-task learning framework to output the confirmed etiology and pathogen categories, introduce a feature-based gradient boosting machine to enhance performance, evaluate the model using indicators such as accuracy, precision, recall, and F1-score. Input the collected clinical data into the model, and the model outputs through multi-tasks. First, give predictions on the risk of severe pneumonia and respiratory support scores. On this basis, combine large language models and expert knowledge graphs to assist doctors in making decisions on ICU needs, help doctors judge medication adjustments, allocate resources, etc. Visualize the model evaluation index results in the form of charts and reports to facilitate researchers and doctors to understand the model performance, discover problems in a timely manner based on the evaluation results, optimize the model, promote the reasonable allocation of medical resources, and drive the efficient operation of the medical system.

[0150] Those skilled in the art should understand that the above-mentioned modules or steps of the present invention can be implemented by a general-purpose computer device. Optionally, they can be implemented by program codes executable by a computing device. Thus, they can be stored in a storage device and executed by the computing device, or they can be separately made into individual integrated circuit modules, or multiple modules or steps among them can be made into a single integrated circuit module to be implemented. The present invention is not limited to any specific combination of hardware and software.

[0151] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

[0152] Although the specific implementation manners of the present invention are described above in conjunction with the accompanying drawings, it is not a limitation on the protection scope of the present invention. Those skilled in the art should understand that based on the technical solutions of the present invention, various modifications or deformations that do not require creative labor by those skilled in the art are still within the protection scope of the present invention.

Claims

1. A clinical performance verification system based on an infectious lung disease diagnostic model, characterized by: include The data collection module uses standardized processes to collect and process data of different dimensions, and cleans and standardizes all multimodal data to ensure data consistency, completeness, and accuracy; The data management module builds a standard database of infectious lung diseases, sets different access levels for different users, anonymizes the patient's personal privacy information, and uses encryption technology to ensure data security; The model evaluation module builds models in layers around microbial detection and host information, and uses multimodal fusion technology to extract and fuse features from image, text, structured and other data. It also uses a self-attention mechanism to fuse multimodal data, outputs etiology diagnosis and pathogen classification through a multi-task learning framework, introduces a feature-based gradient boosting machine to enhance model performance, and uses accuracy and other indicators to evaluate the model. The result display and feedback module inputs the collected clinical data into the infectious lung disease diagnosis model, compares the model output results with the actual situation, and calculates the accuracy, precision and recall rate to fully understand the effectiveness and reliability of the model in the diagnosis of infectious lung diseases. At the same time, the results of various indicators of the model evaluation are visualized in the form of intuitive charts and reports, which makes it easy for researchers and clinicians to quickly understand the model performance.

2. The clinical performance verification system based on the infectious lung disease diagnostic model according to claim 1, characterized in that: The data acquisition module collects patient baseline information, laboratory test data, imaging data, intervention measures and clinical outcomes. Among them, the patient's laboratory test data (such as C-reactive protein, procalcitonin, inflammatory factors, etc.) will be uniformly converted into international standard units, and imaging data (such as CT and X-rays) will be stored in DICOM format to ensure compatibility across hospital systems. Metadata will be added during all data collection processes, such as data source, collection time, disease type, etc., to facilitate subsequent data retrieval and association analysis. The specific steps for image data feature extraction in this process are as follows: Normalization is used to adjust pixel values ​​to a standard range to reduce differences between images: Among them, I is the original impact data, I min and I max They are the minimum and maximum values ​​of the image data respectively. The purpose of this operation is to provide stable output data for subsequent analysis. In addition, data augmentation technology is used to diversify the image processing.

3. The clinical performance verification system based on the infectious lung disease diagnostic model according to claim 1, characterized in that: The model evaluation module includes a feature engineering module, a multimodal fusion model module and a prediction model module. The feature engineering module adopts a data processing method of multimodal feature extraction, adaptive Fourier and feature space fusion, and the method of multimodal feature extraction is as follows; When processing structured data and feature extraction, we normalized the data to ensure that data of different dimensions were compared on the same scale; X std =(X-μ X ) / s X Among them, X represents the original data, μ X represents the mean, σ X Represents the standard deviation. This standardization process effectively eliminates the dimensional differences between different features, allowing all data to be compared and modeled on a consistent basis; We used a variety of advanced machine learning methods to extract features from the standardized data, including random forest models and support vector machines, to identify the most important features. F ml =MLFeatureExtraction(X std ) Among them, F ml is the extracted feature, MLFeatureExtraction is the machine learning feature extraction method, and X std These features of the standardized structured data are further input into the random forest model to extract the important features of the structured data, integrate and output them: The multi-time point data from outpatient, emergency and ICU are integrated, and then the integrated data are standardized or normalized. Then, the change trend and time series characteristics are extracted from them, including heart rate changes and white blood cell counts. At the same time, GBM (gradient boosting machine) is an integrated learning method to gradually reduce the prediction error. In each round of training, GBM will fit the residuals generated by the previous round of model to optimize the prediction ability of the model. The GBM model is trained with the goal of minimizing the loss function. At this time, assuming that the training set is X and the target variable is F, the prediction value of the model is obtained according to the corresponding mechanism; Among them, h m (X) is the mth decision tree, α m is the weight.

4. The clinical performance verification system based on the infectious lung disease diagnostic model according to claim 1, characterized in that: The data management module adopts a distributed storage structure to ensure the security and efficient management of data, and supports simultaneous access by multiple users and multiple hospitals through network interconnection. In terms of data authority management, different access levels are set according to user roles, and patients' personal privacy information is anonymized and encrypted for storage and transmission.

5. The clinical performance verification system based on the infectious lung disease diagnostic model according to claim 3, characterized in that: The feature engineering module uses adaptive Fourier decomposition time series data processing and feature extraction. The specific steps are as follows: For structured data (time series data) counted daily, we apply adaptive Fourier decomposition (AFD) for feature extraction, and the daily statistical results are used as the input of the time series, as follows; X time-series =DailyStatistics(X) Among them, X time-series Represents the converted time series data, and this formula expresses the process of converting the original structured data into time series data through daily statistical operations. AFDAFD recursively decomposes the signal into multiple orthogonal basis functions. These basis functions are adaptively selected according to the spectral characteristics of the signal to ensure the maximum extraction of key features in the signal. Each basis function is adaptively selected according to the specific spectral characteristics of the signal. The standardized time series data is denoted as signal f(t). First, the initial basis function g0(t) that best matches the signal characteristics is selected in accordance with the maximum selection principle of energy tracking matching. Then, the projection of the signal on the initial basis function 〈f,g0〉 is calculated. The inner product here is the inner product of the Euclidean space, and the residual signal is calculated: r1(t)=f(t)-〈f,g0〉g0(t) Repeat the above process to obtain a series of basis functions, and calculate the projection and residual respectively until the stopping condition is met (the conditions to be met include but are not limited to the residual signal energy being lower than the threshold): r k+1 (t)=r k (t)-〈r k ,g k 〉g k (t) Finally, the projection results of all basis functions are summed up to obtain the AFD reconstruction result of the original signal: In this way, AFD extracts the main spectral features of the signal, reflecting the periodic and non-periodic components hidden in the signal. We further input the features extracted by AFD into the long short-term memory network (LSTM) model to learn and predict time series features. The LSTM model can effectively handle long-term dependencies and capture dynamic change characteristics in time series, thereby improving the prediction ability of complex time series data, especially in biomedical signal analysis. The specific formula is as follows; F afd =AFD(X time-series )、F afd-lstm =LSTM(F afd ) Among them, (X time-series ) represents the input time series data, AFD is a function name, F afd It is the result obtained after the AFD function processes the time series data. It may be the extracted feature vector or the data form after some transformation, which will be used as the input for the next calculation. LSTM stands for Long Short-Term Memory Network (Long-ShortTermMemorynetwork),F afd-lstm Is F afd The output result after input into the LSTM network.

6. The clinical performance verification system based on the infectious lung disease diagnostic model according to claim 3, characterized in that: The multimodal fusion model module is responsible for integrating data from different modalities, such as images, texts, time series, and structured data. To achieve effective fusion of multimodal data, a self-attention mechanism is used to integrate image features, text features, structured data features, and associated features extracted by the prediction model and the large language model into a feature matrix. F concat =[F img-reduced ,F txt-fixed ,F str-selected ,F afd-lstm ,F LLM ] Feature fusion relies on the self-attention mechanism, which determines the importance of features of different modalities to the current task by calculating the correlation weight between each feature and other features. The specific operation is as follows: Q=W Q F concat 、K=W K F concat 、V=W V F concat The query vector is used to characterize the relationship between the features of the current modality and other modalities. The key vector is used to capture the feature representation of other modalities. The value vector represents the weight and information of the feature itself. At this time, the dot product operation between the query vector and the key vector is used to reflect the correlation between the features of different modalities. The dot product result is normalized by the Softmax function to obtain the weight: Among them, Q represents the query vector, which is used to characterize the relationship between the features of the current modality and other modalities, K represents the key vector, which is used to capture the feature representation of other modalities, and V represents the value vector, which represents the weight and information of the feature itself. QK T It is represented as the dot product operation of the transpose of Q and K. Here is the k is the dimension of K.

7. The clinical performance verification system based on the infectious lung disease diagnostic model according to claim 3, characterized in that: The prediction model module adopts a multi-task learning framework, and in order to output the diagnosis of the cause and the category of the pathogen, the total loss function of the multi-task learning includes cross entropy loss and regression loss: L=αL1+βL2 Among them, L1 and L2 represent cross entropy losses and are used for classification tasks: Among them, y i is the true label of the sample, indicating the true situation that the sample belongs to a certain category, and its value is (0\1), p i is the probability that the model predicts that the sample belongs to the jth class.

8. The clinical performance verification system based on the infectious lung disease diagnostic model according to claim 1, characterized in that: When evaluating the infectious lung disease diagnosis model based on multimodal data, the key verification indicators of the result display and feedback module include accuracy, precision, recall and F1-score, as follows: Accuracy refers to the proportion of samples correctly classified by the model to the total samples. The formula is: Among them, TP is a true positive example, TN is a true negative example, FP is a false positive example, and FN is a false negative example. The accuracy rate can quickly reflect the overall performance of the model; Precision measures the proportion of samples that are actually positive examples among those predicted by the model. The formula is: Recall refers to the ratio of positive examples identified by the model to actual positive examples. The formula is: F1-score, the harmonic mean of precision and recall, is calculated as follows: Through the comprehensive evaluation of the above indicators, we can fully understand the effectiveness and reliability of the model in the diagnosis of infectious lung diseases.

Images