Deep learning-based neonatal rare disease clinical feature auxiliary diagnosis system

Abstract

The application discloses a kind of based on deep learning's neonatal rare disease clinical feature auxiliary diagnosis system, it is related to medical diagnosis technical field, to solve multidimensional high-dimensional data feature fusion problem, break through the generalization ability limitation of traditional diagnostic model, accurately capture latent feature combination, reduce the technical problem of neonatal rare disease misdiagnosis risk, including: medical database module, for storing medical training sample and the medical data of patient to be diagnosed;Deep learning system is used to extract features, encode, match and diagnose to the medical data;The diagnostic result visualization tool includes visualization tool and text tool.The application breaks through the generalization limitation of traditional diagnostic model by feature embedding module fusion multidimensional high-dimensional data, accurately capture latent feature combination;Training feature coding module reduces the risk of overfitting, for the complex case of methylmalonic acidemia, the diagnostic accuracy is greatly improved compared with traditional method.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical diagnosis, more particularly to a new-born rare disease clinical feature auxiliary diagnosis system based on deep learning. BACKGROUND

[0002] The new-born rare disease has various clinical features, high data dimension and scarce samples. For example, the new-born methylmalonic acidemia has various clinical manifestations, including non-specific symptoms such as vomiting, lethargy and dyspnea, and involves multi-dimensional high-dimensional data such as genes, metabolite concentrations and imaging, and the clinical samples are extremely scarce.

[0003] The traditional diagnosis method relies on manual screening of features and doctor's experience, and it is difficult to effectively fuse multi-dimensional data such as gene sequence variation, blood biochemical index abnormality and brain MRI image features, and there are technical bottlenecks of insufficient data feature fusion and insufficient diagnosis model generalization ability. This leads to the difficulty of capturing the hidden feature combination (such as the association between specific gene mutation and metabolite concentration abnormality), which easily causes missed diagnosis or misdiagnosis. In view of this, a new-born rare disease clinical feature auxiliary diagnosis system based on deep learning is proposed. SUMMARY

[0004] The present application aims to provide a new-born rare disease clinical feature auxiliary diagnosis system based on deep learning to solve the problem of multi-dimensional high-dimensional data feature fusion, break through the limitation of traditional diagnosis model generalization ability, accurately capture the hidden feature combination and reduce the risk of missed diagnosis and misdiagnosis of new-born rare diseases.

[0005] To solve the above technical problems, the present application provides the following technical scheme: a new-born rare disease clinical feature auxiliary diagnosis system based on deep learning, comprising:

[0006] A medical database module for storing medical training samples and medical data of patients to be diagnosed;

[0007] A deep learning system for feature extraction, encoding, matching and diagnosis of the medical data;

[0008] The diagnosis result visualization tool includes a visualization tool and a text tool for displaying the diagnosis result in a visual form and disease risk judgment result;

[0009] The deep learning system comprises:

[0010] A feature embedding module for processing medical training samples and patients to be diagnosed, outputting training embedding vectors and a to-be-diagnosed embedding vector;

[0011] a training feature encoding module configured to obtain a training feature vector corresponding to the medical training sample according to the training embedding vector

[0012] a classification label encoding module configured to obtain a classification label embedding vector corresponding to the medical training sample according to the classification label

[0013] a feature matching module configured to match the training feature vector with the classification label embedding vector , the to-be-diagnosed embedding vector and the training feature description result to obtain a first diagnosis result and a second diagnosis result ; a classification effect evaluation unit configured to evaluate the judgment result of the deep learning system and the disease risk judgment result.

[0014] The feature embedding module of the deep learning system is used to perform dimension reduction, coding and dimension increase processing on the multi-modal data of gene sequencing data, plasma metabolite spectrum and brain MRI image of neonatal methylmalonic acidemia, to generate a training embedding vector and a to-be-diagnosed embedding vector , which breaks through the limitation of shallow fusion of data features in traditional methods, and the feature matching module further realizes the deep mapping of disease features through the precise matching of the training feature vector with the classification label embedding vector . For the complex cases of methylmalonic acidemia, the diagnosis accuracy is greatly improved compared with traditional methods.

[0015] Preferably, the medical database module includes a medical training sample medical database and a to-be-diagnosed patient medical database.

[0016] The medical training sample medical database includes medical training samples, and each medical training sample includes a training feature vector, a classification label and a rare disease category.

[0017] The to-be-diagnosed patient medical database includes to-be-diagnosed patients, each corresponding to a to-be-diagnosed feature vector.

[0018] ​​​​​​The classification labels include training positive labels and training negative labels, the training positive labels are rare disease samples, and the training negative labels are non-rare disease samples.

[0019] Preferably, the feature embedding module includes a first full connection layer, a first pooling layer, an encoder, a second full connection layer, a second pooling layer, and a feature embedding layer.

[0020] The first full connection layer has an input dimension of and an output dimension of , and is configured to reduce the dimension of a first feature vector matrix to obtain a first full connection sample matrix with a dimension of ;

[0021] The first pooling layer is configured to pool the first full connection sample matrix to obtain a first pooling feature vector of medical training samples ;

[0022] The encoder is configured to calculate an encoding feature vector corresponding to the medical training sample according to the first pooling feature vector ;

[0023] The second full connection layer is configured to increase the dimension of each encoding feature vector to obtain an increased dimension feature vector , and has an input dimension of and an output dimension of ;

[0024] The second pooling layer is configured to calculate a second pooling feature vector according to the increased dimension feature vector

[0025] The feature embedding layer is configured to calculate a training embedding vector representing the clinical features of the medical training sample according to the second pooling feature vector .

[0026] Preferably, the medical training sample feature vectors of the medical training sample form a first feature vector matrix , the to-be-diagnosed sample feature vectors of the to-be-diagnosed patient form a second feature vector matrix , and the training embedding vector and the to-be-diagnosed embedding vector together form a first embedding vector matrix .

[0027] Preferably, the training feature encoding module comprises a first fully connected network and a second fully connected network, the first fully connected network and the second fully connected network are of the same structure and different network parameters;

[0028] The input dimension of the first fully connected network is , and the output dimension is , the input is the first embedded feature matrix , and the output is the third pooled feature matrix ;

[0029] The input dimension of the second fully connected network is , and the output dimension is , the input is the third pooled feature matrix , and the output is the fourth pooled feature matrix .

[0030] Preferably, the training feature encoding module fuses the fourth pooled feature matrix into a training feature description result , and the fusion method is:

[0031] ;

[0032] In the formula, is the fourth pooled feature matrix output by the second fully connected network.

[0033] Preferably, the first diagnosis result includes the rare disease probability of the medical training sample , the non-rare disease probability of the medical training sample , and the disease risk judgment result.

[0034] Wherein, the method for calculating the rare disease probability of the medical training sample is:

[0035] ;

[0036] In the formula, is a sigmoid function, is the training feature description result corresponding to the medical training sample, ​corresponding to the first medical training sample. corresponding to the first medical training sample.

[0037] Preferably, the disease risk judgment result is obtained by a threshold calculation part, and the threshold calculation part obtains a threshold value according to the diagnosis result and the diagnosis result , wherein the threshold value is a threshold value of the first medical training sample.

[0038] If the diagnosis result is low, the disease risk judgment result is low risk; if the diagnosis result is high, the disease risk judgment result is high risk.

[0039] Preferably, the second diagnosis result includes a rare disease probability of the patient to be diagnosed and a non-rare disease probability of the patient to be diagnosed , and specifically:

[0040] ;

[0041] In the formula, is a normalized exponential function.

[0042] Preferably, the diagnosis result is composed of the second diagnosis result and the disease risk judgment result, that is:

[0043] ;

[0044] In the formula, is a preset probability fluctuation threshold value, is a vector product of a fourth pooling feature matrix corresponding to the training embedding vector and a diagnosis embedding vector to be diagnosed , and is a reference value related to the second diagnosis result.

[0045] Compared with the prior art, the present application has the following beneficial effects:

[0046] 1. The present application breaks through the limitation of shallow fusion of data features in traditional methods by using a feature embedding module of a deep learning system to perform dimension reduction, coding and dimension elevation on multi-modal data such as gene sequencing data, plasma metabolite spectrum and brain MRI image of neonatal methylmalonic acidemia, to generate training embedding vectors and diagnosis embedding vectors that can represent cross-modal feature association. ​precise match with the classification label embedding vector Deep mapping of disease characteristics is achieved. For complex cases of methylmalonic acidemia, the diagnostic accuracy is significantly improved compared to traditional methods.

[0047] 2、The application also trains the feature encoding module to adopt The fully connected network with the same group structure but independent parameters is used to extract features in parallel, and the comprehensive feature description result is generated by mean fusion, which significantly reduces the overfitting risk of a single network to a specific data distribution. This design effectively addresses the challenge of rare disease sample scarcity. Even in the case of diseases such as methylmalonic acidemia, where the sample size is small, the differentiated learning of multiple network parameters can still extract common patterns from the data, significantly improving the adaptability of the model in small sample scenarios.

[0048] 3、The application also uses a diagnosis result visualization tool, a threshold calculation and a probability fluctuation threshold to convert the diagnosis result into quantitative information containing risk levels and present it in a visual chart. This mechanism not only ensures the accuracy of the diagnosis, but also solves the problem of ambiguous risk judgment in traditional diagnosis, assisting doctors in completing case evaluation in a short time, and greatly improving the efficiency of the traditional process. In addition, the classification effect evaluation unit uses tools such as confusion matrix and ROC curve to quantify the performance of the model, forming a "diagnosis-evaluation-optimization" closed loop to continuously enhance the response ability of the system to new rare diseases. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 is the system architecture diagram of the application; DETAILED DESCRIPTION

[0050] Example 1: As shown in the figure, the application relates to a new neonatal rare disease clinical feature auxiliary diagnosis system based on deep learning, which comprises a medical database module, a deep learning system and a diagnosis result visualization tool. Figure 1

[0051] In the embodiment of the application, the medical database module comprises a medical training sample medical database and a patient to be diagnosed medical database, which are used to store medical training samples and medical data of patients to be diagnosed;

[0052] The medical training sample medical database comprises one medical training sample, which comprises a training feature vector, a classification label and a rare disease category;

[0053] The classification label comprises a training positive label and a training negative label, the training positive label is a rare disease sample, and the training negative label is a non-rare disease sample;

[0054] ​The medical database of the to-be-diagnosed patient comprises a to-be-diagnosed patient, and each to-be-diagnosed patient corresponds to a to-be-diagnosed feature vector;

[0055] In an embodiment of the present application, the deep learning system comprises a feature embedding module, a training feature encoding module, a classification label encoding module, a feature matching module, and a classification effect evaluation unit, which are used for feature extraction, encoding, matching, and diagnosis of the medical data;

[0056] The feature embedding module is used for processing medical training samples and to-be-diagnosed patients, and outputs training embedding vectors and to-be-diagnosed embedding vectors;

[0057] The medical training samples correspond to medical training sample feature vectors, which form a first feature vector matrix , and the to-be-diagnosed patients correspond to to-be-diagnosed sample feature vectors, which form a second feature vector matrix ;

[0058] In an embodiment of the present application, the feature embedding module comprises a first full connection layer, a first pooling layer, an encoder, a second full connection layer, a second pooling layer, and a feature embedding layer;

[0059] The first full connection layer has an input dimension of , , where d represents the original dimension of the input feature vector, and an output dimension of , is the dimension after dimension reduction by the first full connection layer, which is used for dimension reduction of the first feature vector matrix to obtain a first full connection sample matrix with a dimension of , i.e. , where is a mapping function of the first full connection layer;

[0060] The first pooling layer is used for pooling the first full connection sample matrix to obtain first pooling feature vectors of the medical training samples , and the pooling operation aims to reduce the dimension of the medical data and extract key features;

[0061] The encoder is used for calculating encoding feature vectors corresponding to the medical training samples according to the first pooling feature vectors , i.e. ;

[0062] a second full connection layer, configured to perform dimension lifting on each encoded feature vector to obtain a dimension-lifted feature vector, and set an input dimension of the dimension-lifted feature vector to , , is a feature dimension of an encoder output, and an output dimension is , that is, the dimension-lifted feature vector , wherein is a mapping function of the second full connection layer;

[0063] a second pooling layer, configured to calculate, according to the dimension-lifted feature vector, a plurality of second pooling feature vectors , that is, ;

[0064] a feature embedding layer, configured to calculate, according to the second pooling feature vectors, a training embedding vector representing a clinical feature corresponding to a medical training sample , that is, , wherein is a mapping function of the feature embedding layer; In an embodiment of the present application, the training embedding vector

[0065] and the to-be-diagnosed embedding vector together constitute a first embedding vector matrix , that is, , wherein is a mapping function for processing the to-be-diagnosed feature vector, and has the same structure as the medical training sample processing , wherein represents a matrix is a row, column real matrix; In the above process, ,

[0066] , , , are all full connection network structures, and use ReLu as an activation function, and a dropout parameter is 0.5. Dropout is a technique to prevent overfitting, and 0.5 means that 50% of neuron connections are randomly discarded during training. In the above process, ReLu is used as an activation function, and a dropout parameter is 1, that is, all neuron connections are discarded during training (only during training, and not discarded during testing).

[0067] ​Training feature encoding module, used to embed vectors according to training , calculated to obtain The training feature vector corresponding to the medical training samples ;

[0068] In an embodiment of the present invention, the training feature encoding module includes The first fully connected network and A second fully connected network; The structures of the first fully connected networks are the same, and the network parameters of the first fully connected networks are different; The structures of the second fully connected networks are the same, and the network parameters of the second fully connected networks are different;

[0069] The input dimension of the first fully connected network is , the output dimension is , whose input is the first embedding feature matrix (By training embedding vector ), the output is the third pooling feature matrix ,Right now , where is the mapping function of the first fully connected network, the activation function is Relu, and the dropout parameter is 0.5;

[0070] The input dimension of the second fully connected network is , the output dimension is , whose input is the third pooling feature matrix , the output is the fourth pooling feature matrix ,and ,Right now , where is the mapping function of the second fully connected network, the activation function is Relu, and the dropout parameter is 0.5;

[0071] Among them, the training feature encoding module will The fourth pooling feature matrix Fusion is the result of training feature description , the fusion method can be summation, averaging and other common feature fusion methods, such as , where For the The fourth pooled feature matrix output by the second fully connected network;

[0072] Classification label encoding module, used to calculate and obtain The classification label embedding vector corresponding to the medical training samples , the classification labels include training positive labels and training negative labels, where the classification label encoding module maps the classification label to an embedding vector through a fully connected network or other methods;

[0073] Feature matching module, used to train feature vectors and the classification label embedding vector Match and get the first diagnosis result , and embed the diagnosis into the vector And the training feature description results Match and calculate to obtain the second diagnosis result ;

[0074] In an embodiment of the present invention, the first diagnosis result Including The probability of rare diseases in medical training samples Hedi The probability of non-rare diseases in medical training samples and disease risk assessment results;

[0075] Among them, calculate the The probability of rare diseases in medical training samples The method is: , where is the sigmoid function, For the The training feature description results corresponding to the medical training samples are For the The classification label embedding vector corresponding to the medical training samples;

[0076] Among them, the disease risk judgment result is obtained by the threshold calculation part, which is calculated based on the diagnosis result. Calculate the threshold , threshold For the The threshold of medical training samples, if , then the disease risk judgment result is low risk; if , then the disease risk judgment result is high risk;

[0077] Among them, the second diagnosis result Probability of including rare disease patients awaiting diagnosis and the probability of non-rare diseases in patients to be diagnosed , specifically , where is the normalized exponential function;

[0078] Classification effect evaluation unit, used to evaluate the judgment results of the deep learning system The confusion matrix method is used to evaluate the judgment results of the deep learning system. The confusion matrix is ​​used to intuitively display the matching between the predicted results of the classification model and the actual results. The combination of ROC curve and AUC, or the combination of receiver operating characteristic curve and AUC, is used to evaluate the disease risk judgment results of the deep learning system. The ROC curve is used to show the performance of the classifier at different thresholds, and the AUC (area under the curve) is used to quantify the overall performance of the classifier. The larger the AUC value, the better the performance.

[0079] The diagnostic result visualization tool includes a visualization tool and a text tool. The visualization tool uses a medical data visualization method to display the diagnostic results in a visual form. and disease risk assessment results, such as displaying rare disease probability distribution, risk level, and other information through charts;

[0080] In an embodiment of the present invention, the diagnosis result The second diagnosis and the disease risk judgment results, namely:

[0081] ;

[0082] Where, It is a preset probability fluctuation threshold used to define the reasonable range of disease risk judgment. To train the embedding vector The corresponding fourth pooling feature matrix Embedded vector to be diagnosed The vector product of , is a reference value related to the second diagnosis result.

[0083] Example 2: This example builds a test model for neonatal methylmalonic acidemia based on multicenter clinical data. The following presents the key module parameter settings, data processing flow, and diagnostic results:

[0084] 1. Basic data;

[0085]

[0086] 2. Feature embedding module processing flow;

[0087]

[0088] Processing of patients to be diagnosed: Use the same process to process 20 samples to be diagnosed and generate embedding vectors to be diagnosed (Dimension ).

[0089] III. Training the feature encoding module Group Parallel Network

[0090]

[0091] IV. Diagnosis result calculation example

[0092] 1. First diagnosis result

[0093] Rare disease probability:

[0094] : Classification label embedding vector (positive label set as [1, 0], negative label set as [0, 1], mapped to 10-dimensional vector by full connection layer)

[0095] : sigmoid function, output range (0, 1)

[0096] Risk judgment:

[0097] Threshold Statistically derived from training data (set to 0.6)

[0098] If , it is determined as high risk; otherwise, it is low risk.

[0099] 2. Second diagnosis result

[0100] Probability calculation:

[0101] Output as a two-dimensional vector , representing the probability of rare disease and non-rare disease respectively.

[0102] Diagnosis result integration:

[0103]

[0104] Where (vector product), is the fluctuation threshold.

[0105] V. Visualization result

[0106]

[0107] VI. Performance evaluation index

[0108]

[0109] ​​​Summary: Through the constructed diagnostic model of neonatal methylmalonic acidemia, the feasibility of key parameter analysis is preliminarily verified, and a quantifiable technical path is provided for clinical diagnosis. In practical application, it is recommended to dynamically calibrate the parameters combined with real clinical data, and to optimize the model performance through cross-validation. The optimized diagnostic model can significantly improve the detection efficiency and reduce the misdiagnosis rate, providing efficient and reliable technical support for early accurate diagnosis and intervention of neonatal methylmalonic acidemia.

[0110] The embodiments of the present application are disclosed, but not limited to the preferred embodiments, and those skilled in the art can easily understand the spirit of the present application according to the above embodiments, and make different inferences and changes, as long as they do not deviate from the spirit of the present application, which are within the protection scope of the present application.

Claims

1. A deep learning-based neonatal rare disease clinical feature auxiliary diagnosis system, characterized in that, Comprise: a medical database module for storing medical training samples and medical data of patients to be diagnosed; a deep learning system for feature extraction, encoding, matching and diagnosis of the medical data; Diagnostic result visualization tools, including visualization tools and textualization tools, for displaying diagnostic results in a visualized form and disease risk determination results; wherein the deep learning system comprises: Feature embedding module for Medical training samples and Patients to be diagnosed are processed and output training embedding vectors and Embedding vectors to be diagnosed; Training feature encoding module, used to embed vectors according to training , calculated to obtain The training feature vector corresponding to the medical training samples ; The training feature encoding module includes The first fully connected network and A second fully connected network, The first fully connected network with The second fully connected networks have the same structure but different network parameters; The input dimension of the first full connection network is , and the output dimension is , the input of which is the first embedded feature matrix , and the output is the third pooled feature matrix ; The input dimension of the second full connection network is , and the output dimension is , the input of which is the third pooling feature matrix , and the output is the fourth pooling feature matrix ; The training feature coding module fuses a fourth pooling feature matrix into a training feature description result , and a fusion manner is: , wherein, is a fourth pooling feature matrix output by the i th second fully connected network The classification label encoding module is configured to calculate a classification label embedding vector corresponding to each medical training sample according to a classification label ;​ Feature matching module, used to train feature vectors Embedded vector with classification label Matching, embedded vector to be diagnosed And the training feature description results Matching to obtain the first diagnosis result and the second diagnosis ; the first diagnosis result a rare disease probability of the first medical training sample a non-rare disease probability of the first medical training sample a rare disease probability of the first medical training sample a non-rare disease probability of the first medical training sample and a disease risk determination result The disease risk judgment result is calculated by a threshold calculation section, which calculates the threshold value according to the diagnosis result calculates the threshold value , the threshold value is the threshold value of the first medical training sample If , the disease risk judgment result is low risk; if , the disease risk judgment result is high risk; The method for calculating the rare disease probability of the i-th medical training sample is as follows: The method for calculating the rare disease probability of the i-th medical training sample is as follows:​ ； Where, is the sigmoid function, For the The training feature description results corresponding to the medical training samples are For the The classification label embedding vector corresponding to the medical training samples; the second diagnosis result a probability of a rare disease of the patient to be diagnosed and a probability of a non-rare disease of the patient to be diagnosed in particular: ； wherein is a normalized exponential function; a classification effect evaluation unit configured to evaluate a judgment result of the deep learning system and a disease risk judgment result; The diagnostic result Specifically, , wherein, is a preset probability fluctuation threshold, is a training embedding vector corresponding to a fourth pooling feature matrix is a vector product of the to-be-diagnosed embedding vector and the embedding vector, is a reference value related to the second diagnostic result.

2. The deep learning-based neonatal rare disease clinical feature auxiliary diagnosis system according to claim 1, characterized in that, the medical database module comprises a medical training sample medical database and a patient to be diagnosed medical database; The medical training sample medical database includes a medical training sample, the medical training sample including a training feature vector, a classification label, and a rare disease category; The medical database of the patient to be diagnosed comprises a patient to be diagnosed, respectively corresponding to a feature vector to be diagnosed; wherein the classification label comprises a training positive label and a training negative label, the training positive label is a rare disease sample, and the training negative label is a non-rare disease sample.

3. The deep learning-based neonatal rare disease clinical feature auxiliary diagnosis system according to claim 2, characterized in that, The feature embedding module comprises a first full connection layer, a first pooling layer, an encoder, a second full connection layer, a second pooling layer and a feature embedding layer; The first full connection layer is configured to have an input dimension of and an output dimension of , and is used to reduce the dimension of the first feature vector matrix to obtain a first full connection sample matrix with a dimension of ; a first pooling layer configured to pool the first full connection sample matrix to obtain a first pooled feature vector of the medical training sample ;​​ An encoder is configured to obtain a first pooling feature vector from the first feature vector The medical training sample corresponding to the first pooling feature vector is calculated The medical training sample corresponding to the first pooling feature vector is calculated ; a second fully connected layer, configured to obtain an encoded feature vector of each image feature vector dimensionality increasing, to obtain a dimensionality-increased feature vector , and set an input dimension of the second fully connected layer as , and set an output dimension of the second fully connected layer as ; a second pooling layer configured to calculate a second pooling feature vector according to the dimension-increased feature vector ;​ a feature embedding layer configured to obtain a training embedding vector representing a corresponding clinical feature of the medical training sample based on the second pooled feature vector computing a training embedding vector representing a corresponding clinical feature of the medical training sample .

4. The deep learning-based neonatal rare disease clinical feature auxiliary diagnosis system according to claim 3, characterized in that, The medical training sample corresponds to A first feature vector matrix is composed of the medical training sample feature vectors The patient to be diagnosed corresponds to A second feature vector matrix is composed of the to-be-diagnosed sample feature vectors The training embedding vector And the to-be-diagnosed embedding vector Together constitute a first embedding vector matrix .

Images