Serial number	Department's office	ICD coding	Name of disease
				1	Department of respiration	B37.101+J99.8*	Pulmonary fungal infection-pulmonary candidiasis
2	Department of respiration	B44.101+J99.8*	Allergic bronchopulmonary aspergillosis
				3	Department of respiration	B44.102+J99.8*	Pulmonary fungal infection-pulmonary aspergillosis
4	Department of respiration	B45.001+J99.8*	Pulmonary fungal infection-cryptococcosis of lung
				5	Department of respiration	C85.908	Pulmonary lymphoma
6	Department of respiration	C95.0018	Interstitial pneumonia of lymphocytes
				7	Department of respiration	D71xx03	Chronic granulomatosis
8	Department of respiration	D76.015	Histiocytosis of Langerhans cells
				9	Department of respiration	D80.501	Immunodeficiency with high lgM
10	Department of respiration	D81.901	Combined immunodeficiency disease
				11	Department of respiration	D83.901	Common variant immunodeficiency disease
12	Department of respiration	E84.902	Cystic fibrosis
				13	Department of respiration	J47xx01	Bronchiectasis
14	Department of respiration	J67.902	Allergic alveolitis
				15	Department of respiration	J69.001	Aspiration pneumonia
16	Department of respiration	J84.001	Pulmonary alveolar proteinosis
				17	Department of respiration	J84.003	Diffuse alveolar hemorrhage
18	Department of respiration	J84.902	Interstitial pneumonia
				19	Department of respiration	M30.101	Eosinophilic granulomatous vasculitis
20	Department of respiration	G24.902	Primary ciliary dyskinesia

The child disease knowledge base system in this embodiment may include a training set, a validation set, and a test set.

In some embodiments of the present application, a large number of professional terms may be collected from various websites or web pages, such as Baidu encyclopedia/medical research and inquiry, electronic medical record data, etc., to construct a multi-type term dictionary; and word segmentation processing can be performed by using a CRF (conditional random field), a hidden Markov, a deep learning method and the like, and in the embodiment, word segmentation information can be corrected by a pediatric transfer to obtain a corpus in a term dictionary.

Each word in the corpus is represented by a vector, so that the keywords in the corpus can be characterized in vectorization.

Since the vector has a large dimension and less non-0 elements, the vector is stored in a sparse manner. For example, each word is assigned a numerical ID, e.g., "fever" as 3 and "fever" as 5 (assuming a mark from 0). During programming, the number distribution work of each word can be completed through the Hash table. Word vectors may be trained using the Skip-gram model.

Furthermore, the format required for inputting the model can be processed from the text of the electronic medical record, for example, after the electronic medical record is extracted from a database, the format required for inputting the model can be processed; and acquiring image data of CT, ultrasound, MR, pathology, etc. from a system of a hospital, and performing a preprocessing operation (such as normalization or binarization processing) on the image data.

Since the diagnostic information includes text information and image data, for this purpose, the normalization process of the image data may include:

the image data of the training set, the validation set, and the test set are normalized, for example, in this embodiment, the mean and the standard deviation of the training set are calculated first, and then the data of the validation set and the test set are normalized according to the mean and the standard deviation of the training set, so as to process the data set including the classification information and the diagnosis information using a unified standard. In addition, data enhancement operation can be performed on the data of the training set, so that the data volume in the data set processing process is increased, and the training precision in the data set processing process is improved.

The children's disease knowledge base system in this embodiment is a term dictionary using vectorization, one category in each term dictionary may be a multi-dimensional vector, each dimension may be understood as a disease feature in the category, and the weight of the disease features of multiple categories is also represented by one vector.

In addition, the model used in the present embodiment is a BP neural network model. The process of BP neural network training can be illustrated as follows:

referring to fig. 3 and 4, in the diagnosis model, a BP network can be used to build a prediction model of a disease, and the basic process of modeling is as follows: firstly, collecting main influencing factors causing diseases and results of the diseases, such as a training set, a testing set and a verification set of a children disease knowledge base system; then inputting the influencing factors and the disease results into a designed neural network model for repeated training until the network converges (namely, an expected training error is achieved), properly adopting a certain skill in the training process to ensure that the training speed of the network is fastest, the error is minimum and the model is optimal, and finally performing disease prediction by using the established model.

Design of input layer

A BP neural network is constructed whose number of input layer nodes depends on the dimension of the input vector, i.e., the dimension of the feature vector. Therefore, when designing a respiratory disease diagnosis simulation system in the respiratory department, the respiratory disease diagnosis selects variables related to diseases, namely age z, symptoms (cough. z, chest pain 22., fever z, expectoration zs, hemoptysis e, dyspnea z, cold sweat zs and pestle finger z) and examinations (radiographic examination z. and bronchofiberscope examination z) as input data, and the data is preprocessed to obtain the dimension of a feature vector so as to determine the number of nodes of a network input layer.

Design of hidden layer

Hidden layer (hidden layer) node number design

In the BP network, the selection of the number of hidden nodes is very important, which not only has great influence on the performance of the established neural network model, but also can not train the network at all or has poor network performance if the number of the hidden nodes is too small; if the number of hidden nodes is too large, although the system error of the network can be reduced, on one hand, the training time of the network is prolonged, on the other hand, the training is easy to fall into local minimum points and an optimal point cannot be obtained, and the hidden nodes are also the inherent reason for generating 'overfitting' during the training. Therefore, the reasonable hidden node number is determined by a node deletion method and an expansion method under the condition of comprehensively considering the complexity of the network structure and the error magnitude.

In this embodiment, the number of hidden nodes is determined, an "overfitting" phenomenon during training is avoided as much as possible, and a sufficiently high network performance and generalization capability are ensured, and the most basic principle for determining the number of hidden nodes is as follows: on the premise of meeting the precision requirement, a structure which is as compact as possible is obtained, namely the number of hidden nodes is as small as possible. According to empirical formulah is the number of hidden layer nodes, m is the number of input layer nodes, n is the number of output layer nodes, and a is a constant. Alternatively, the empirical formula is h ═ n (number of input layer nodes) + m (number of output layer nodes) + a (constant).

Learning step length

The learning step size affects the stability of the learning process of the BP neural network. The large learning rate may cause the correction amount of the network weight value each time to be too large, and even cause the weight value to exceed a certain error in the correction process, and the minimum value appears irregular jump and does not converge; however, an excessively small learning rate results in an excessively long learning time, but can ensure convergence to a certain minimum value. Generally, the learning rate is selected to be smaller to ensure the convergence (stability) of the learning process, and is usually between 0.01 and 0.8.

Initial connection weight of network

The BP algorithm determines that a plurality of local minimum points generally exist in an error function, and different network initial weights directly determine which local minimum point or global minimum point the BP algorithm converges on.

Determination of a transfer function

The linear transfer functions purelin, tan-sigmoid transfer function tansig and log-sigmoid transfer function logsig which are commonly used in BP neural networks.

1) The linear transfer function is a linear transfer function that calculates the output of the output layer from the input of the network in the format: (x) purelin (x).

2) tangent-Signoid-type transfer functions are used to map the input range of neurons from plus infinity to minus infinity to (-1, 1). The tangent Sigmoid type transfer function tansig () is a differentiable function and is therefore well suited for training a neural network using the BP algorithm, and this embodiment selects this function as the transfer function of the network implicit and output layers. The format is as follows: (x) tan sig (x).

3) A logarithmic Sigmoid type transfer function is used to map the input range of the neuron from positive infinity to negative infinity to (0, 1). The logarithmic Sigmoid type transfer function logsig () is also a differentiable function, suitable for training neural networks using the BP algorithm, in the format: (x) logsig (x).

Design of output layer

The output layer of the BP neural network is only one layer, and the number of nodes is determined by the number of output vectors. If the respiratory research shows that the patients have 10 diseases (pneumonia, asthma, nervous system combination and acute bronchitis … …), the output vector is 10.

Training of neural network systems

The training of the BP network is to continuously adjust the weight of the network by applying the error back-propagation principle to ensure that the sum of the squares of the errors between the output value of the network model and the output value of the known training sample reaches the minimum or is smaller than a certain expected value. A reasonable BP neural network model is established through learning (training) of training concentrated training samples, the trained BP neural network model is utilized, simulation diagnosis is carried out by using test set and verification concentrated data, experimental results are compared with doctor diagnosis with abundant experience, and the diagnosis coincidence rate is more than 90% and is qualified.

Example 2

As shown in fig. 2, the present embodiment provides an intelligent auxiliary diagnosis system for respiratory diseases of children, including:

As shown in fig. 2, the first information collecting module includes:

In practical application, the first information collection module and the display module are simultaneously located on an information display interface of the auxiliary diagnosis system.

Through machine learning, a typical symptom diagnosis model in early occurrence of the respiratory distress/respiratory department is formed. Firstly, in the clinical standard diagnosis and treatment quality control process of the children hospital respiratory department cooperation specialist, a high-sensitivity screen-out reminding for key clinical work elements is formed; secondly, in the key difficult and complicated disease typing work, high-specificity auxiliary accurate diagnosis typing is realized, and auxiliary judgment is made on the state risk and the subsequent outcome of the disease.

The product is simple to operate, has a wide application range, is suitable for various medical health systems, and has disease symptoms including patient age, onset condition, main symptoms, special examinations, family history and auxiliary examinations which are arranged according to respiratory experts of Beijing children hospitals on the left side; the right panel shows the disease diagnosis. The doctor can display the diagnosis result in real time by clicking the symptom expression of the patient and the right column.

The diagnosis result format provided on the right side of the intelligent diagnosis is disease name + diagnosis weight, and the diagnosis results are arranged from large to small according to the weight. Typical symptoms of the disease are listed below for each diagnosis, with blue highlighted symptoms indicating that the current patient is present and grey symptoms indicating that the current patient is not present. If the result of the diagnosis of the disease is difficult miscellaneous disease, the upper right corner gives a red prompt to remind the patient that the patient has the possibility of difficult miscellaneous disease, and the doctor can be helped to identify and diagnose the difficult symptoms.

It should also be noted that the exemplary embodiments mentioned in this patent describe some methods or systems based on a series of steps or devices. However, the present invention is not limited to the order of the above-described steps, that is, the steps may be performed in the order mentioned in the embodiments, may be performed in an order different from the order in the embodiments, or may be performed simultaneously.

The above embodiments may be referred to each other, and the present embodiment does not limit the embodiments.

Finally, it should be noted that: the above-mentioned embodiments are only used for illustrating the technical solution of the present invention, and not for limiting the same; although the present invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some or all of the technical features may be equivalently replaced; and the modifications or the substitutions do not make the essence of the corresponding technical solutions depart from the scope of the technical solutions of the embodiments of the present invention.