Deep learning-based children viral encephalitis classification system
The deep learning-based classification system for viral encephalitis in children, utilizing multimodal data processing and counterfactual baseline vector reconstruction, solves the problem of misdiagnosis in the context of epidemics using traditional diagnostic methods, achieving high accuracy in encephalitis identification and reducing the risk of misdiagnosis.
Patent Information
- Application Number
- CN202511742856.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-25
- Publication Date
- 2026-02-24
AI Technical Summary
During outbreaks of respiratory infectious diseases such as influenza or mycoplasma, traditional diagnostic methods are unable to effectively distinguish between simple systemic infections and those accompanied by encephalitis, leading to misdiagnosis or insufficient diagnostic confidence. In particular, it is difficult to capture subtle encephalitis-specific features in complex systemic symptom environments.
A deep learning-based classification system for viral encephalitis in children is adopted. Serological indicators, cerebrospinal fluid biochemical indicators, and epidemiological background intensity factors are obtained through the data acquisition module. The feature decoupling module generates a normalized multimodal joint feature vector and reconstructs the counterfactual baseline vector. The symptom masking coefficient and encephalitis-specific residual feature vector are calculated. The comprehensive decision trust entropy is calculated in combination with the risk assessment module, and the diagnosis is output through the closed-loop control module.
It effectively eliminates environmental noise interference, improves the accuracy of encephalitis identification, and reduces the risk of misdiagnosis. It adaptively responds to background noise of epidemics through symptom masking coefficient and dynamic feature amplification mechanism, and optimizes the efficiency of diagnosis and treatment by integrating the comprehensive decision trust entropy index of model uncertainty and environmental deviation risk.
Smart Images
Figure CN121565461A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent medical auxiliary diagnosis and data processing technology, specifically a classification system for viral encephalitis in children based on deep learning. Background Technology
[0002] With the increasing sophistication of the diagnosis and treatment of viral encephalitis in children, the complexity of clinical diagnosis has increased significantly. This complexity mainly stems from the diversity and nonspecificity of the clinical manifestations of the children, especially in the differential diagnosis in the context of a high incidence of epidemics. Currently, diagnosis generally relies on doctors to make a comprehensive judgment based on serological indicators, cerebrospinal fluid biochemical indicators, and clinical manifestations. However, during outbreaks of respiratory infectious diseases such as influenza or mycoplasma, children often have severe systemic inflammatory responses. This strong environmental background noise often masks the weak neurological symptoms in the early stages of viral encephalitis, resulting in a high degree of overlap in the clinical characteristics of the two. Traditional diagnostic methods usually lack the ability to dynamically calibrate by combining real-time epidemiological data, making it difficult to effectively distinguish between simple systemic infection and the complex situation of encephalitis. This leads to problems such as misjudgment or insufficient diagnostic confidence in a highly interfering environment. Therefore, how to effectively eliminate the interference of environmental background noise and accurately capture the subtle encephalitis-specific characteristics from complex systemic symptoms has become an urgent problem to be solved in this field. Summary of the Invention
[0003] To address the aforementioned technical problems, this invention discloses a deep learning-based classification system for viral encephalitis in children. Specifically, the technical solution of this invention includes: The data acquisition module is used to obtain serological indicators, cerebrospinal fluid biochemical indicators, and concurrent epidemiological background intensity factors of the target children. The feature decoupling module is used to call preset historical statistical data to normalize serological indicators and cerebrospinal fluid biochemical indicators, and generate a normalized multimodal joint feature vector by combining the epidemic background intensity factor; calculate the symptom masking coefficient based on the epidemic background intensity factor; reconstruct the counterfactual baseline vector based on the normalized multimodal joint feature vector using a preset variational autoencoder; calculate the difference between the normalized multimodal joint feature vector and the counterfactual baseline vector, and use the symptom masking coefficient to weight the difference to generate an encephalitis-specific residual feature vector. The risk assessment module is used to map the encephalitis-specific residual feature vector to the probability of encephalitis using a preset classifier; and to calculate the comprehensive decision trust entropy based on the probability of encephalitis, the symptom masking coefficient, and the norm of the encephalitis-specific residual feature vector. The closed-loop control module is used to compare the comprehensive decision trust entropy with a preset stability threshold. If the comprehensive decision trust entropy is less than or equal to the stability threshold, the probability of encephalitis and the binary classification result are output. If the overall decision trust entropy is greater than the stability threshold, output a clinical feature significance map.
[0004] Furthermore, the feature decoupling module generates a normalized multimodal joint feature vector, including: Use serological and cerebrospinal fluid biochemical indicators; Access historical statistical data, which includes the mean vector of serum indicators, the standard deviation vector of serum indicators, the mean vector of cerebrospinal fluid indicators, and the standard deviation vector of cerebrospinal fluid indicators. The serological indicators were standardized using the mean vector and standard deviation vector of the serological indicators. The biochemical indicators of cerebrospinal fluid (CSF) are standardized using the mean vector and standard deviation vector of CSF indicators. The standardized serological indicators, standardized cerebrospinal fluid biochemical indicators, and epidemic background intensity factors are cascaded to generate a normalized multimodal joint feature vector.
[0005] Furthermore, the feature decoupling module calculates the symptom masking coefficient, including: Call the epidemic background intensity factor, the preset Sigmoid steepness parameter, and the preset baseline prevalence threshold; Calculate the difference between the epidemic background intensity factor and the baseline prevalence threshold; Calculate the product of the Sigmoid steepness parameter and the difference; The product is used to calculate the numerical value using the Sigmoid function, and the numerical value is defined as the symptom masking coefficient. The symptom masking coefficient is used to quantify the degree of interference of environmental noise on specific diagnosis.
[0006] Furthermore, the feature decoupling module reconstructs the counterfactual baseline vector, including: Invoke the normalized multimodal joint feature vector; The normalized multimodal joint feature vector is input into the probabilistic encoder network of the variational autoencoder to generate the latent variable distribution; The latent variable distribution is input into the probability decoder network of the variational autoencoder to generate reconstructed data; Define the reconstructed data as a counterfactual baseline vector; The variational autoencoder is trained using a non-encephalitis dataset, and the counterfactual baseline vector represents the theoretical index value when the child only has a systemic infection.
[0007] Furthermore, the feature decoupling module generates encephalitis-specific residual feature vectors, including: Call the preset feature attention vector and preset environment gain adjustment coefficient; Calculate the difference vector between the normalized multimodal joint eigenvector and the counterfactual baseline vector; The difference vector is weighted element-wise using the feature attention vector to generate a weighted difference vector. The dynamic amplification factor is calculated based on the symptom masking coefficient and the environmental gain adjustment coefficient. The weighted difference vector is amplified by a dynamic amplification factor to generate an encephalitis-specific residual feature vector.
[0008] Furthermore, the risk assessment module maps the encephalitis-specific residual feature vector to the probability of encephalitis incidence, including: Call the encephalitis-specific residual feature vector, the preset classifier weight matrix, and the preset classifier bias term; Linear transformation is performed on the encephalitis-specific residual feature vector; The Softmax function is used to normalize the result after linear transformation to generate the probability of encephalitis.
[0009] Furthermore, the risk assessment module calculates the comprehensive decision trust entropy, including: Call the preset unit alignment coefficient; Shannon entropy is calculated based on the probability of encephalitis incidence, and is defined as the uncertainty component of model classification. Calculate the L2 norm of the encephalitis-specific residual eigenvectors; The product of the dimensional alignment coefficient, the symptom masking coefficient, and the L2 norm is calculated, and the product is defined as the environmental feature deviation risk component. The sum of the uncertainty component of the calculation model and the deviation risk component of environmental characteristics is defined as the comprehensive decision trust entropy.
[0010] Furthermore, the closed-loop control module outputs results when the comprehensive decision trust entropy is less than or equal to the stability threshold, including: The current diagnostic conclusion is considered to be within the confidence range; Output the probability of encephalitis generated by the risk assessment module; The child's category is determined based on the probability of encephalitis, and a binary classification result is generated.
[0011] Furthermore, the closed-loop control module outputs a result when the comprehensive decision trust entropy is greater than the stability threshold, including: The current diagnostic conclusion is determined to be in a high-risk misclassification range, and the output of binary classification results is suppressed. Calculate the gradient of the probability of encephalitis with respect to the normalized multimodal joint feature vector; Calculate the Hadamard product of the gradient and the normalized multimodal joint eigenvector; The Hadamard product is defined as a clinical feature saliency map, which is used to indicate clinical indicators that lead to high trust entropy.
[0012] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention effectively solves the problem of misdiagnosis caused by the overlap of encephalitis symptoms and systemic infection symptoms by introducing a counterfactual baseline vector reconstruction mechanism based on variational autoencoders. The system uses a VAE trained only on non-encephalitis datasets to generate theoretical indicators when the child only has a systemic infection, and performs difference calculations with the actually collected normalized multimodal joint feature vectors. This subtraction operation successfully eliminates the interference of non-specific common features such as fever and conventional inflammation, generating a pure encephalitis-specific residual feature vector, forcing the classifier to focus on the real neuropathological changes in cerebrospinal fluid and serum, thereby improving the accuracy of identifying viral encephalitis under complex clinical manifestations.
[0013] 2. This invention innovatively constructs a symptom masking coefficient and a dynamic feature amplification mechanism, enabling it to adaptively cope with background noise from concurrent epidemics such as influenza. By collecting epidemic background intensity factors in real time, the system uses the Sigmoid function to calculate the symptom masking coefficient, quantifying the degree of interference of environmental noise on diagnosis. Based on this coefficient, a reverse compensation mechanism is triggered, and combined with the environmental gain adjustment coefficient, the system can dynamically weight and amplify the feature difference. This means that during the peak of an epidemic, the model will automatically suppress the weights of general indicators that are greatly affected by the environment and amplify the weak specific encephalitis features, effectively avoiding the decline in the model's generalization ability or the generation of a large number of false positives due to fluctuations in external epidemiological data.
[0014] 3. This invention breaks through the limitations of traditional models that rely solely on predicted probabilities for classification, and proposes a comprehensive decision-making trust entropy index that integrates model uncertainty and environmental deviation risk. This index not only calculates Shannon entropy based on the probability of encephalitis, but also introduces an environmental feature deviation risk component that combines the L2 norm and the symptom masking coefficient. This dual verification mechanism can sensitively identify high-risk samples with abnormal data distribution or model indecisiveness. By quantifying the credibility of the diagnosis, the system can issue warnings when features are ambiguous, preventing hasty binary diagnostic conclusions when confidence is insufficient, thereby significantly reducing the risk of medical accidents.
[0015] 4. This system incorporates a closed-loop control module based on a trust entropy threshold, optimizing the diagnostic efficiency of human-machine collaboration. When the comprehensive decision trust entropy is less than the threshold, the system outputs a diagnostic result; however, when the trust entropy is too high, indicating a potential risk of misjudgment, the system automatically suppresses the output of binary classification results and instead calculates and generates a clinical feature saliency map. This map uses the Hadamard product of gradient and feature vector to visually highlight key clinical indicators that increase model uncertainty. This uncertainty-as-an-explanation mechanism not only avoids misleading doctors but also accurately guides them to focus on abnormal data, assisting them in making targeted secondary assessments. Attached Figure Description
[0016] The present invention will be further explained below with reference to the accompanying drawings and embodiments: Figure 1 This is a system structure diagram of the present invention. Detailed Implementation
[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.
[0018] Example 1 Please see Figure 1 A deep learning-based classification system for viral encephalitis in children, comprising: The data acquisition module is used to obtain serological indicators, cerebrospinal fluid biochemical indicators, and concurrent epidemiological background intensity factors of the target children. The feature decoupling module is used to call preset historical statistical data to normalize serological indicators and cerebrospinal fluid biochemical indicators, and generate a normalized multimodal joint feature vector by combining the epidemic background intensity factor; calculate the symptom masking coefficient based on the epidemic background intensity factor; reconstruct the counterfactual baseline vector based on the normalized multimodal joint feature vector using a preset variational autoencoder; calculate the difference between the normalized multimodal joint feature vector and the counterfactual baseline vector, and use the symptom masking coefficient to weight the difference to generate an encephalitis-specific residual feature vector. The risk assessment module is used to map the encephalitis-specific residual feature vector to the probability of encephalitis using a preset classifier; and to calculate the comprehensive decision trust entropy based on the probability of encephalitis, the symptom masking coefficient, and the norm of the encephalitis-specific residual feature vector. The closed-loop control module is used to compare the comprehensive decision trust entropy with a preset stability threshold. If the comprehensive decision trust entropy is less than or equal to the stability threshold, output the probability of encephalitis and the binary classification result. If the comprehensive decision trust entropy is greater than the stability threshold, output a clinical feature significance map; This embodiment illustrates a deep learning-based classification system for viral encephalitis in children. This system solves the problem of capturing weak neurological symptoms under strong background noise by integrating environmental epidemiological data and counterfactual reasoning mechanisms. The data acquisition module, serving as the system's standardized input interface, is responsible for aggregating multi-source heterogeneous clinical data. This module acquires in real-time serological indicators, cerebrospinal fluid biochemical indicators, and concurrent epidemiological background intensity factors for the target children. Defined as the weighted moving average of the positive rates of influenza and mycoplasma in fever clinics in the region within 72 hours before the time of the target child's visit, it is used to characterize the level of external environmental noise at the current time point; The feature decoupling module performs environmental noise elimination and specific pathological feature extraction; this module calls preset historical statistical data to normalize serological indicators and cerebrospinal fluid biochemical indicators, and generates a normalized multimodal joint feature vector by combining the epidemic background intensity factor; the module calculates the symptom masking coefficient to quantify the degree of environmental interference based on the epidemic background intensity factor; the module uses a variational autoencoder trained on a non-encephalitis dataset to reconstruct the counterfactual baseline vector based on the normalized multimodal joint feature vector; The counterfactual baseline vector represents the theoretical index state of the child when he / she has only a systemic infection and no encephalitis complications; the module calculates the difference between the normalized multimodal joint feature vector and the counterfactual baseline vector, and uses the symptom masking coefficient to weight the difference to generate an encephalitis-specific residual feature vector; The risk assessment module, as a decision inference unit, uses a pre-defined classifier to map the encephalitis-specific residual feature vector to the probability of encephalitis. To prevent misjudgment caused by extreme data, the module calculates the comprehensive decision trust entropy based on the probability of encephalitis, the symptom masking coefficient, and the norm of the encephalitis-specific residual feature vector. The closed-loop control module will combine the decision trust entropy with the preset stability threshold. Comparison; stability threshold The value of is determined by calculating the 95th quantile of the trust entropy distribution of correctly classified samples in the validation set to ensure statistical significance in the high confidence interval. When the comprehensive decision trust entropy is less than or equal to the stability threshold, the system determines that the diagnosis is credible and outputs the probability of encephalitis and the binary classification result. When the comprehensive decision trust entropy is greater than the stability threshold, the system determines that the diagnosis is in the high-risk misjudgment interval, suppresses the output of classification results, and outputs a clinical feature significance map instead.
[0019] Example 2 The feature decoupling module generates a normalized multimodal joint feature vector, including: Use serological and cerebrospinal fluid biochemical indicators; Access historical statistical data, which includes the mean vector of serum indicators, the standard deviation vector of serum indicators, the mean vector of cerebrospinal fluid indicators, and the standard deviation vector of cerebrospinal fluid indicators. The serological indicators were standardized using the mean vector and standard deviation vector of the serological indicators. The biochemical indicators of cerebrospinal fluid (CSF) are standardized using the mean vector and standard deviation vector of CSF indicators. The standardized serological indicators, standardized cerebrospinal fluid biochemical indicators, and epidemic background intensity factors are concatenated to generate a normalized multimodal joint feature vector. The feature decoupling module generates normalized multimodal joint feature vectors. The process follows the logic of multi-source heterogeneous data fusion; Module calls serological indicators and cerebrospinal fluid biochemical indicators ; Retrieve historical statistical data, which includes a vector of mean serum markers derived from a historical database of non-encephalitis cases. Standard deviation vector of serum indicators cerebrospinal fluid index mean vector Vector of standard deviation of cerebrospinal fluid indicators ; The module constructs a normalized multimodal joint feature vector based on the Z-score normalization method. The calculation formula is as follows: ; in, This represents a vector concatenation operation, which concatenates the components into a high-dimensional column vector. The background intensity factor of the concurrent epidemic is directly mapped to a dimension of the feature space, enabling subsequent models to have environmental perception capabilities.
[0020] Example 3 The feature decoupling module calculates the symptom masking coefficient, including: Call the epidemic background intensity factor, the preset Sigmoid steepness parameter, and the preset baseline prevalence threshold; Calculate the difference between the epidemic background intensity factor and the baseline prevalence threshold; Calculate the product of the Sigmoid steepness parameter and the difference; The product is used to calculate the numerical value using the Sigmoid function, and the numerical value is defined as the symptom masking coefficient. The symptom masking coefficient is used to quantify the degree of interference of environmental noise on specific diagnosis. The feature decoupling module calculates the symptom masking coefficient. The process aims to quantify the degree of interference of environmental noise on specific diagnostics; Module calls epidemic background intensity factor Sigmoid steepness parameter and benchmark prevalence threshold Symptom masking coefficient The calculation formula is as follows: ; Sigmoid steepness parameter It is a constant obtained through calibration using historical data; the calibration process involves constructing an environmental prevalence database covering different historical periods. Corresponding misdiagnosis rate The calibration dataset, in which the misdiagnosis rate Defined as the false positive rate of non-encephalitis febrile patients being misclassified as having encephalitis under a given environmental prevalence, the data was analyzed using logistic regression. By fitting the data, the prediction error is minimized, thus determining the correct prediction. The numerical value; the baseline prevalence threshold. The value is set as the sum of the historical average positive rate at fever clinics during the same period and twice the standard deviation. When the environmental positive rate exceeds this value, A value close to 1 indicates a high-noise mode.
[0021] Example 4 The feature decoupling module reconstructs the counterfactual baseline vector, including: Invoke the normalized multimodal joint feature vector; The normalized multimodal joint feature vector is input into the probabilistic encoder network of the variational autoencoder to generate the latent variable distribution; The latent variable distribution is input into the probability decoder network of the variational autoencoder to generate reconstructed data; Define the reconstructed data as a counterfactual baseline vector; The variational autoencoder was trained using a non-encephalitis dataset, and the counterfactual baseline vector represented the theoretical index value when the child only had a systemic infection. Feature decoupling module reconstructs counterfactual baseline vector The process is based on the anomaly detection principle of generative models; Module calls normalized multimodal joint feature vector ; Probabilistic encoder networks utilizing variational autoencoders Will Mapped to a latent variable distribution; a probabilistic decoder network based on a variational autoencoder constructed using a multilayer perceptron is employed. Mapping the distribution of latent variables back to the original data space generates reconstructed data, i.e., counterfactual baseline vectors. The calculation formula is as follows: ; Parameters of variational autoencoder and The model was trained using only non-encephalitis datasets; for input from children with encephalitis, the model was forced to reconstruct the normal non-encephalitis pattern, thereby... It serves as a standard reference for measuring pathological deviations.
[0022] Example 5 The feature decoupling module generates encephalitis-specific residual feature vectors, including: Call the preset feature attention vector and preset environment gain adjustment coefficient; Calculate the difference vector between the normalized multimodal joint eigenvector and the counterfactual baseline vector; The difference vector is weighted element-wise using the feature attention vector to generate a weighted difference vector. The dynamic amplification factor is calculated based on the symptom masking coefficient and the environmental gain adjustment coefficient. The weighted difference vector is amplified using a dynamic amplification factor to generate an encephalitis-specific residual feature vector. The feature decoupling module generates encephalitis-specific residual feature vectors. The process enables the extraction and enhancement of key pathological features from the original signal; The module calls the preset feature attention vector. and environmental gain adjustment coefficient Feature attention vector It is a weight vector constructed based on the prior knowledge of clinical experts; its construction logic is as follows: calculate the coefficient of variation of each indicator in the non-encephalitis fever group, and define the weight as the normalized value of the inverse of the coefficient of variation; as a simplified preferred implementation of the above construction logic, in this embodiment, the weight of the highly specific cerebrospinal fluid indicator dimension is directly set to 1.0, and the weight of the serum indicator dimension which is easily affected by the systemic inflammatory response is directly set to 0.1, so as to reduce the computing power cost of real-time calculation. Module computes normalized multimodal joint eigenvectors With counterfactual baseline vector The difference vector; element-wise weighting of the difference vector using the feature attention vector; based on the symptom masking coefficient. and environmental gain adjustment coefficient The dynamic magnification factor is calculated; to prevent the introduction of excessive estimation variance in the context of extreme high epidemics, the system applies upper and lower bound truncation to the dynamic magnification factor, limiting its value to within a certain range. Within the interval; the weighted difference vector is amplified using a dynamic amplification factor to generate an encephalitis-specific residual feature vector. The calculation formula is as follows: ; in, Hadamard product; feature attention vector Dimensions and Normalized Multimodal Joint Feature Vector Same; Environmental gain adjustment coefficient The values are determined as fixed hyperparameters using a grid search method, that is, traversing the validation set. The value of is selected from the value space that maximizes the recall rate of the model during the flu outbreak.
[0023] Example 6 The risk assessment module maps encephalitis-specific residual feature vectors to the probability of encephalitis incidence, including: Call the encephalitis-specific residual feature vector, the preset classifier weight matrix, and the preset classifier bias term; Linear transformation is performed on the encephalitis-specific residual feature vector; The Softmax function is used to normalize the result after linear transformation to generate the probability of encephalitis. The risk assessment module will use encephalitis-specific residual feature vectors. Mapped to the probability of encephalitis The process employs linear transformation and normalization. Module calls encephalitis-specific residual feature vector Classifier weight matrix and classifier bias term ;right A linear transformation is performed, and the result is normalized using the Softmax function to generate the probability of encephalitis incidence. The calculation formula is as follows: ; Classifier weight matrix and bias terms The results were obtained through supervised learning and optimized on a mixed dataset containing both confirmed and excluded cases.
[0024] Example 7 The risk assessment module calculates the comprehensive decision-making trust entropy, including: Call the preset unit alignment coefficient; Shannon entropy is calculated based on the probability of encephalitis incidence, and is defined as the uncertainty component of model classification. Calculate the L2 norm of the encephalitis-specific residual eigenvectors; The product of the dimensional alignment coefficient, the symptom masking coefficient, and the L2 norm is calculated, and the product is defined as the environmental feature deviation risk component. The sum of the uncertainty component of the calculation model classification and the deviation risk component of environmental characteristics is defined as the comprehensive decision trust entropy. The risk assessment module calculates the comprehensive decision-making trust entropy. The process aims to integrate model uncertainty with environmental risks; Module call unit alignment coefficient encephalitis incidence rate Encephalitis-specific residual eigenvectors Symptom masking coefficient ; based on Calculate the Shannon entropy, which is defined as the uncertainty component of the model classification. ; Calculate the L2 norm of the encephalitis-specific residual eigenvectors Calculate the product of the dimensional alignment coefficient, the symptom masking coefficient, and the L2 norm, and define it as the environmental feature deviation risk component; calculate the sum of the two to obtain the comprehensive decision trust entropy. The calculation formula is as follows: ; Dimensional alignment factor The numerical scale is used to balance the two terms in the formula and to map the feature norm with physical units to a dimensionless risk scalar to prevent a single component from dominating the decision due to differences in value ranges; its value is set as the L2 norm of the encephalitis-specific residual feature vector of all samples in the validation set. The reciprocal of the arithmetic mean is used to ensure that the environmental feature deviation risk component and the model classification uncertainty component are on the same order of magnitude, thereby achieving numerical additivity after dimensionless transformation; the environmental feature deviation risk component is used to detect outlier samples with extremely strong environmental interference and drastic feature abnormalities.
[0025] Example 8 The closed-loop control module outputs results when the comprehensive decision trust entropy is less than or equal to the stability threshold, including: The current diagnostic conclusion is considered to be within the confidence range; Output the probability of encephalitis generated by the risk assessment module; The child's category is determined based on the probability of encephalitis, and a binary classification result is generated. The closed-loop control module integrates decision-making trust entropy. Less than or equal to the stability threshold The output logic at that time is as follows; The system determines that the current diagnostic conclusion is within the confidence range; the module directly outputs the probability of encephalitis generated by the risk assessment module. The system determines the category of the child based on the probability of encephalitis and generates a binary classification result. This processing logic ensures real-time diagnosis in low-risk or typical scenarios.
[0026] Example 9 The closed-loop control module outputs results when the comprehensive decision trust entropy is greater than the stability threshold, including: The current diagnostic conclusion is determined to be in a high-risk misclassification range, and the output of binary classification results is suppressed. Calculate the gradient of the probability of encephalitis with respect to the normalized multimodal joint feature vector; Calculate the Hadamard product of the gradient and the normalized multimodal joint eigenvector; The Hadamard product is defined as a clinical feature saliency map, which is used to indicate clinical indicators that lead to high trust entropy. The closed-loop control module integrates decision-making trust entropy. Greater than the stability threshold The output logic at that time is as follows; The system determines that the current diagnostic conclusion falls within the high-risk misclassification range and actively suppresses the output of binary classification results; the module calculates the probability of encephalitis. Compared to normalized multimodal joint feature vector The gradient; calculate the gradient and... The Hadamard layer is defined as a clinical feature saliency map. The calculation formula is as follows: ; Clinical feature saliency maps are used to visualize specific clinical indicators that lead to high trust entropy, assisting doctors in making manual judgments when the model refuses to recognize them.
[0027] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention; any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A deep learning-based classification system for viral encephalitis in children, characterized in that, include: The data acquisition module is used to obtain serological indicators, cerebrospinal fluid biochemical indicators, and concurrent epidemiological background intensity factors of the target children. The feature decoupling module is used to call preset historical statistical data to normalize serological indicators and cerebrospinal fluid biochemical indicators, and generate a normalized multimodal joint feature vector by combining the epidemiological background intensity factor. Calculate the symptom masking coefficient based on the epidemiological background intensity factor; The counterfactual baseline vector is reconstructed based on the normalized multimodal joint feature vector using a pre-defined variational autoencoder; the difference between the normalized multimodal joint feature vector and the counterfactual baseline vector is calculated, and the difference is weighted using the symptom masking coefficient to generate encephalitis-specific residual feature vectors. The risk assessment module is used to map the encephalitis-specific residual feature vector to the probability of encephalitis using a preset classifier; and to calculate the comprehensive decision trust entropy based on the probability of encephalitis, the symptom masking coefficient, and the norm of the encephalitis-specific residual feature vector. The closed-loop control module is used to compare the comprehensive decision trust entropy with a preset stability threshold. If the comprehensive decision trust entropy is less than or equal to the stability threshold, the probability of encephalitis and the binary classification result are output. If the overall decision trust entropy is greater than the stability threshold, output a clinical feature significance map.
2. The deep learning-based classification system for viral encephalitis in children according to claim 1, characterized in that, The feature decoupling module generates a normalized multimodal joint feature vector, including: Use serological and cerebrospinal fluid biochemical indicators; Access historical statistical data, which includes the mean vector of serum indicators, the standard deviation vector of serum indicators, the mean vector of cerebrospinal fluid indicators, and the standard deviation vector of cerebrospinal fluid indicators. The serological indicators were standardized using the mean vector and standard deviation vector of the serological indicators. The biochemical indicators of cerebrospinal fluid (CSF) are standardized using the mean vector and standard deviation vector of CSF indicators. The standardized serological indicators, standardized cerebrospinal fluid biochemical indicators, and epidemic background intensity factors are concatenated to generate a normalized multimodal joint feature vector.
3. The deep learning-based classification system for viral encephalitis in children according to claim 1, characterized in that, The feature decoupling module calculates the symptom masking coefficient, including: Call the epidemic background intensity factor, the preset Sigmoid steepness parameter, and the preset baseline prevalence threshold; Calculate the difference between the epidemic background intensity factor and the baseline prevalence threshold; Calculate the product of the Sigmoid steepness parameter and the difference; The product is used to calculate the numerical value using the Sigmoid function, and the numerical value is defined as the symptom masking coefficient. The symptom masking coefficient is used to quantify the degree of interference of environmental noise on specific diagnosis.
4. A deep learning-based classification system for viral encephalitis in children according to claim 1, characterized in that, The feature decoupling module reconstructs the counterfactual baseline vector, including: Invoke the normalized multimodal joint feature vector; The normalized multimodal joint feature vector is input into the probabilistic encoder network of the variational autoencoder to generate the latent variable distribution; The latent variable distribution is input into the probability decoder network of the variational autoencoder to generate reconstructed data; Define the reconstructed data as a counterfactual baseline vector; The variational autoencoder is trained using a non-encephalitis dataset, and the counterfactual baseline vector represents the theoretical index value when the child only has a systemic infection.
5. A deep learning-based classification system for viral encephalitis in children according to claim 1, characterized in that, The feature decoupling module generates encephalitis-specific residual feature vectors, including: Call the preset feature attention vector and preset environment gain adjustment coefficient; Calculate the difference vector between the normalized multimodal joint eigenvector and the counterfactual baseline vector; The difference vector is weighted element-wise using the feature attention vector to generate a weighted difference vector. The dynamic amplification factor is calculated based on the symptom masking coefficient and the environmental gain adjustment coefficient. The weighted difference vector is amplified by a dynamic amplification factor to generate an encephalitis-specific residual feature vector.
6. A deep learning-based classification system for viral encephalitis in children according to claim 1, characterized in that, The risk assessment module maps encephalitis-specific residual feature vectors to the probability of encephalitis incidence, including: Call the encephalitis-specific residual feature vector, the preset classifier weight matrix, and the preset classifier bias term; Linear transformation is performed on the encephalitis-specific residual feature vector; The Softmax function is used to normalize the result after linear transformation to generate the probability of encephalitis.
7. A deep learning-based classification system for viral encephalitis in children according to claim 1, characterized in that, The risk assessment module calculates the comprehensive decision trust entropy, including: Call the preset unit alignment coefficient; Shannon entropy is calculated based on the probability of encephalitis incidence, and is defined as the uncertainty component of model classification. Calculate the L2 norm of the encephalitis-specific residual eigenvectors; The product of the dimensional alignment coefficient, the symptom masking coefficient, and the L2 norm is calculated, and the product is defined as the environmental feature deviation risk component. The sum of the uncertainty component of the calculation model and the deviation risk component of environmental characteristics is defined as the comprehensive decision trust entropy.
8. A deep learning-based classification system for viral encephalitis in children according to claim 1, characterized in that, The closed-loop control module outputs results when the comprehensive decision trust entropy is less than or equal to the stability threshold, including: The current diagnostic conclusion is considered to be within the confidence range; Output the probability of encephalitis generated by the risk assessment module; The child's category is determined based on the probability of encephalitis, and a binary classification result is generated.
9. A deep learning-based classification system for viral encephalitis in children according to claim 1, characterized in that, The closed-loop control module outputs results when the comprehensive decision trust entropy is greater than the stability threshold, including: The current diagnostic conclusion is determined to be in a high-risk misclassification range, and the output of binary classification results is suppressed. Calculate the gradient of the probability of encephalitis with respect to the normalized multimodal joint feature vector; Calculate the Hadamard product of the gradient and the normalized multimodal joint eigenvector; The Hadamard product is defined as a clinical feature saliency map, which is used to indicate clinical indicators that lead to high trust entropy.