CN111061700A

CN111061700A - Hospitalizing migration scheme recommendation method and system based on similarity learning

Info

Publication number: CN111061700A
Application number: CN201911101258.XA
Authority: CN
Inventors: 史玉良; 程林; 管永明; 张晖; 吕梁; 姜诚
Original assignee: Dareway Software Co ltd
Current assignee: Dareway Software Co ltd
Priority date: 2019-11-12
Filing date: 2019-11-12
Publication date: 2020-04-24

Abstract

The disclosure discloses a hospitalizing migration scheme recommendation method and system based on similarity learning, wherein the method comprises the following steps: a training phase and an application phase; the training phase comprises: acquiring historical hospitalizing data of all medical insurance personnel, and training a hospitalizing migration scheme recommendation model based on the acquired data; the application phase comprises the following steps: recommending the recommendation proportions of different hospitalizing migration schemes for the patient to be recommended based on the hospitalizing migration scheme recommendation model, and giving the patient a hospitalizing migration suggestion.

Description

Hospitalizing migration scheme recommendation method and system based on similarity learning

Technical Field

The disclosure relates to the technical field of medical insurance information processing, in particular to a method and a system for recommending a hospitalizing transfer scheme based on similarity learning.

Background

The statements in this section merely provide background information related to the present disclosure and may not constitute prior art.

With the differentiation development of urban medical levels, the popularization of medical knowledge networks in various regions and the rise of the living standard of residents, conditions are provided for hospitalizing individuals to move to hospitals with higher treatment levels and more favorable reimbursement policies, so that the problem of hospitalizing movement is increasingly prominent, and the balance between the utilization rate of medical resources of hospitals is large, and further the waste of medical resources is caused; meanwhile, the development of the internet leads to continuous improvement of information interaction level and cognition of medical level, and the information explosion is caused by the double-edged sword, so that the hospitalizing individual is difficult to master and transversely compare comprehensive real data of various hospitals, and a comprehensive evaluation mechanism for information value and effectiveness is lacked. Therefore, how to analyze the massive medical insurance data, dig out the hospitalizing rules of the insured persons hidden in the medical insurance massive data, construct a hospitalizing behavior recommendation model, provide reliable reference for hospitalizing selection, avoid unnecessary hospitalizing migration, further provide auxiliary decision support for medical insurance fund allocation and medical resource optimization scheduling, and become an important research problem in the medical insurance field.

In the course of implementing the present disclosure, the inventors found that the following technical problems exist in the prior art:

at present, the traditional recommendation method in the field of medical insurance mainly comprises two steps: (1) calculating similarities between patients; (2) and generating a recommendation list for the patient according to the similarity and the historical hospitalizing behaviors of the patient. Based on medical insurance data, many similarity learning methods are proposed. However, these methods do not take into account the effect of temporal information between hospitalization sequences. In recent years, with the advance in the field of medical insurance information technology, deep learning methods have been widely adopted and rapidly developed in patient-like learning, such as an automatic encoder, a Recurrent Neural Network (RNN), a Long Short-Term Memory Neural Network (LSTM), and a Convolutional Neural Network (CNN). In past studies, the inventors found that CNN is superior to RNN and LSTM in data reduction and continuity representation of feature vectors, and that CNN has also demonstrated its superior ability to measure patient similarity. However, the CNN framework ignores the impact of high-dimensional medical insurance data on similarity learning between patients. Therefore, how to utilize medical insurance data with high-dimensional time characteristics for similarity learning has a great challenge.

In the real medical insurance field, the distribution of most data sets is unbalanced, that is, some classes have more examples than other classes, the degree of unbalance of the data sets may reach 10:1, 100:1 or even 1000:1, and in this case, the data distribution rule information of a few classes is often difficult to be fully expressed. Most data imbalance processing algorithms achieve sample classification by learning the distribution trend of boundary samples, because samples on a boundary are more easily misclassified. Such as the Borderline-SMOTE algorithm, randomly adds instances to a few class samples of the boundary using the SMOTE algorithm. However, this algorithm does not assign weights to the boundary instances.

In response to the problems and challenges presented above, several key issues are addressed: for the medical insurance data with unbalanced distribution, how to carry out data unbalanced processing is carried out, so that better data characteristic distribution rule extraction is achieved; according to the medical insurance data with the time attribute and the high-dimensional characteristics, how to perform measurement learning of the similarity range of the hospitalizing samples is performed, and comprehensive evaluation of hospitalizing resource matching is realized according to the high-similarity sample set, so that a reference basis is provided for hospitalizing migration.

Disclosure of Invention

In order to solve the defects of the prior art, the medical transfer scheme recommendation method and system based on similarity learning are provided in the disclosure; and (3) performing data unbalance processing on the medical insurance data by adopting an Adaptive Borderline-SMOTE (AB-SMOTE) algorithm to generate a new training data set. In order to verify whether the patient needs to take hospitalizing migration behavior, the patient goes to a hospital outside the comprehensive area for hospitalizing. On the basis of a CNN model, a matching matrix is introduced to construct a similarity learning framework, and samples similar to hospitalizing behavior samples of patients are found in populations with hospitalizing migration and populations without hospitalizing migration. And then CNN processing is respectively carried out on the similar samples according to the historical hospitalizing data to obtain the feature expression vectors of the similar samples. And updating the weight value of each feature in the feature vector on the hospital hospitalizing state data by adopting an attention mechanism based on the feature expression vector of the similar sample, and obtaining a comprehensive feature vector influencing the hospitalizing state. And finally, outputting the average treatment effect of treating the diseases in the region by using the comprehensive characteristic vector of the hospitalizing state and the softmax function, and giving a hospitalizing migration suggestion to the patient through treatment effect comparison.

In a first aspect, the present disclosure provides a method for recommending a hospitalizing migration scheme based on similarity learning;

the hospitalizing migration scheme recommendation method based on similarity learning comprises the following steps: a training phase and an application phase;

the training phase comprises: acquiring historical hospitalizing data of all medical insurance personnel, and training a hospitalizing migration scheme recommendation model based on the acquired data;

the application phase comprises the following steps: recommending the recommendation proportions of different hospitalizing migration schemes for the patient to be recommended based on the hospitalizing migration scheme recommendation model.

In a second aspect, the present disclosure also provides a hospitalizing migration scheme recommendation system based on similarity learning;

a hospitalizing migration scheme recommendation system based on similarity learning comprises: a training module and an application module;

the training module configured to: acquiring historical hospitalizing data of all medical insurance personnel, and training a hospitalizing migration scheme recommendation model based on the acquired data;

the application module configured to: recommending the recommendation proportions of different hospitalizing migration schemes for the patient to be recommended based on the hospitalizing migration scheme recommendation model.

In a third aspect, the present disclosure also provides an electronic device comprising a memory and a processor, and computer instructions stored on the memory and executed on the processor, wherein the computer instructions, when executed by the processor, perform the steps of the method of the first aspect.

In a fourth aspect, the present disclosure also provides a computer-readable storage medium for storing computer instructions which, when executed by a processor, perform the steps of the method of the first aspect.

Compared with the prior art, the beneficial effect of this disclosure is:

(1) according to the method, through analysis of recent hospitalizing behavior data, an AB-SMOTE algorithm is adopted to conduct unbalanced processing on historical hospitalizing data, based on the distribution rule of hospitalizing sample data, sampling expansion is conducted on a few types of samples according to the original data distribution density, the distribution rule of the original sample data is maintained, meanwhile, the parameter adjusting difficulty of a subsequent CNN similarity learning frame is reduced, and the problem that the sample similarity distribution rule is difficult to fit due to the fact that the sample data amount is too small is avoided;

(2) the method is based on historical medical insurance data, because medical records are often large in time span, obvious in information stage and unbalanced in high-dimensional data density distribution, the CNN algorithm can identify the whole medical rule based on partial time sequence information, the identification is not influenced by a time range, and the rule distribution of the medical records is not changed while the high-dimensional characteristic data is scaled by convolution, so that a similarity learning framework is constructed by adopting the CNN algorithm to capture the distribution rule of similar people, and the accuracy of similarity clustering of medical samples is improved;

(3) the comprehensive characteristic weight of the comprehensive characteristic in the hospitalizing sample is updated through an attention mechanism, personalized hospitalizing matching is realized through a softmax function, hospitalizing selection recommendation opinions are provided, the hospitalizing requirements are met, meanwhile, unnecessary hospitalizing migration behaviors are avoided, and balanced utilization of medical resources is facilitated.

Drawings

The accompanying drawings, which are incorporated in and constitute a part of this application, illustrate embodiments of the application and, together with the description, serve to explain the application and are not intended to limit the application.

Fig. 1 is an overall flowchart of a medical consultation migration recommendation method according to an embodiment of the disclosure;

FIG. 2 is a flow chart of data imbalance processing based on the AB-SMOTE algorithm according to an embodiment of the present disclosure;

fig. 3 is a data flow processing diagram of patient similarity learning with CNN as a core according to an embodiment of the present disclosure;

fig. 4 is a flowchart of a process of recommending hospitalization migration behavior based on CNN and attention mechanism according to an embodiment of the present disclosure;

fig. 5 is a flowchart of an embodiment of recommending a migration action according to an embodiment of the present disclosure.

Detailed Description

It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the disclosure. Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

It is noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of example embodiments according to the present application. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, and it should be understood that when the terms "comprises" and/or "comprising" are used in this specification, they specify the presence of stated features, steps, operations, devices, components, and/or combinations thereof, unless the context clearly indicates otherwise.

The first embodiment provides a hospitalizing migration scheme recommendation method based on similarity learning;

As shown in fig. 1, 4, and 5, as one or more embodiments, the training phase includes: S1-S7;

s1: acquiring historical hospitalizing data of all medical insurance personnel from medical insurance data;

s2: carrying out unbalance processing on the acquired historical hospitalizing data so as to obtain a training set;

s3: inputting the training set into a first Convolutional Neural Network (CNN), and outputting a plurality of similar patient groups, wherein each similar patient group comprises a plurality of similar patients;

s4: performing feature extraction on the historical hospitalizing data of each similar patient in each similar patient group by using a second Convolutional Neural Network (CNN) to obtain a feature expression vector of each similar patient in each similar patient group;

s5: updating the weight value of each characteristic value in the characteristic expression vector for the characteristic expression vector of each similar patient in each similar patient group by adopting an attention mechanism, and carrying out weighted summation on each characteristic value to obtain a comprehensive characteristic vector of each similar patient in each similar patient group;

s6: inputting the comprehensive characteristic vector of each similar patient in each similar patient group into a softmax function, and outputting a recommended proportion value of each similar patient for transferring to a different hospital or not recommending the transfer;

s7: calculating a loss function of the output value of the softmax function, and training the learning parameters of the second convolutional neural network CNN by adopting a back propagation algorithm to complete the training of the second convolutional neural network CNN;

as one or more embodiments, the applying phase includes: S8-S11;

s8: acquiring hospitalizing data of a patient to be recommended, inputting the hospitalizing data and a training set of the patient to be recommended into a first Convolutional Neural Network (CNN), and outputting a similar patient group of the patient to be recommended;

s9: performing feature extraction on the historical hospitalizing data of each similar patient in the similar patient group of the patient to be recommended by using the trained second convolutional neural network CNN to obtain a feature representation vector of each similar patient in the similar patient group of the patient to be recommended;

s10: updating the weight value of each characteristic value in the characteristic expression vector by adopting an attention mechanism, and carrying out weighted summation on each characteristic value to obtain a comprehensive characteristic vector of each similar patient in the similar patient group of the patient to be recommended;

s11: and inputting the comprehensive characteristic vector of each similar patient in the similar patient group of the patient to be recommended into the softmax function, and outputting recommended proportion values of the patient to be recommended, which are transferred to different hospitals or recommended proportion values of the patient not to be recommended, to be transferred.

As one or more embodiments, in S1, historical hospitalizing data of all medical insurance personnel is obtained from medical insurance data; wherein, historical hospitalization data comprises: basic data, hospital data and hospitalization behavior data; the basic data, including: age, gender, income, type of insurance, location of insurance, disease category, distance to hospitalization, and industry category; the hospital data comprises: hospital name, hospital class, location of hospital, days of hospitalization, amount of hospitalization expenditure, and hospitalization diagnostic items; the hospitalizing behavior data comprises: length of stay, time of discharge, medical program, total cost and follow-up records.

It should be understood that in S1, medical history data is obtained based on the social security system in a certain area of a certain city as an application example, and comparative analysis is performed on the medical history data from the internal medical areas and the external medical areas, and medical records of three related hospitals are comprehensively evaluated by taking tuberculosis, pneumonia and pulmonary lymphatic metastasis (rare diseases) as specific cases, wherein the internal hospitals in the certain area are comprehensive second-grade-A hospitals and the external hospitals in the certain area are comprehensive third-grade-A hospitals and third-grade-A hospitals in the special oncology department. The historical hospitalization data of 2015-plus-2017 years are used as training samples, the data of 2018 years are used as experimental samples, and as shown in table 1, statistics examples of the historical healing sample data sets of hospitals are shown.

TABLE 1 basic statistics of hospitalization category data sets

Based on the medical samples, medical history data of the social security system is extracted, wherein the medical history data comprises 72 types of characteristic items, and the medical records are 21179026 times.

As one or more embodiments, in S2, performing imbalance processing on the acquired historical medical data, so as to obtain a training set; the unbalanced processing is realized by using an Adaptive boundary sampling algorithm (AB-SMOTE) algorithm.

It should be understood that in S2, the mass medical insurance data is processed by data imbalance, the AB-SMOTE algorithm is used, the number of instances that need to be added to each boundary sample is calculated by an adaptive method based on the minority medical samples of the boundary, the adaptive sampling assigns weights according to the learning difficulty level of each minority boundary sample, and τ is distributed by density_iThe number of the generated examples of each few types of boundary samples is automatically determined as a standard, the algorithm obviously improves the unbalanced distribution of the data, and meanwhile, the accuracy of constructing the classifier based on the sample data distribution rule in the later period is improved.

Specifically, as shown in fig. 2, in S2, the generation process of the data imbalance processing based on the AB-SMOTE algorithm is as follows:

s21: defining all medical codes in the medical insurance data as c₁,c₂,…,c_|C|E.g., C, where | C | is the number of hospitalization codes; suppose there are N patients, the Nth patient has T⁽ⁿ⁾And (6) recording medical treatment. A patient can be represented by a hospitalization sequence as

Each medical record contains a set of feature vectors x ∈ R^|C|(ii) a Thus, each patient record can be mapped to a matrixWherein the horizontal dimension represents the corresponding hospitalization event and the vertical dimension represents the visit; an (i, j) entity of the matrix equal to 1 indicates the hospitalization code C of the patient_jAt time V_iIs observed. For CNN operation, each patient has a fixed access length

S22: standardizing the medical behavior sample data, and carrying out normalization processing on the original medical insurance data X by adopting a min-max standardization method, wherein the value of characteristic data is [0,1 ];

s23: classifying a minority of classes of the raw data, each

Determining a series of nearest sample sets called S from the raw data_min-NNAnd is and

where S represents a sample set, S_minA few sample sets are represented:

s24: for each X_iDetermining samples S in the nearest neighbor that belong to the majority class_majI.e. satisfy | S_min-NN∩S_majNumber of instances of |;

if X is_iSatisfies the formula (1), then X_iBelong to boundary sample set Danger;

wherein m is the number of the nearest sample sets;

if x_iAll nearest neighbors are of majority class, i.e. | S_min-NN∩S_maj|＝m，

Then X_iConsidered as a Noise set Noise;

in other cases X_iBelongs to a safety sample set Safe;

s25: first assume that a few classes in the boundary sample set have d_numOne sample, then, it is necessary to generate s × d_minSamples (s is an integer from 1 to k); then for each sample X in the boundary sample set_iFinding out k nearest neighbors, and calculating tau according to Euclidean distance_iAs shown in the formula (2),

wherein, Delta_iIs the number of samples in the k nearest neighbors that belong to the majority class, Z is a normalization constant, and ∑ τ_i＝1；

Calculating each sample X in the boundary sample set according to formula (3)_iNumber of instances g to be generated_i，

g_i＝τ_i×E_x(3)

Wherein E is_x＝s×d_numIndicating the number of samples generated.

Finally, applying SOMTE algorithm to each sample X in the boundary sample set_iA corresponding number of instances is generated to obtain a new training data set.

As one or more embodiments, after S1 and before S2, the method further includes: preprocessing data; the data preprocessing comprises the following steps: data cleaning, missing data completion, data definition and data storage.

As one or more embodiments, in S3, the first convolutional neural network CNN is a modified convolutional neural network; the first convolutional neural network CNN, including: two parallel branch lines;

a first branch comprising: the system comprises a first input layer, a first convolution layer, a first pooling layer, a first weighted summary layer and a first vector representation layer;

a second leg comprising: a second input layer, a second convolutional layer, a second pooling layer, a second weighted summary layer, and a second vector representation layer;

the first output value of the first vector representation layer and the second output value of the second vector representation layer are input into a matching matrix, and the matching matrix outputs similarity representation of the first output value and the second output value;

the similarity of the first output value and the second output value is expressed and input into a softmax function, and the softmax function outputs the similarity; and classifying the patients with the similarity larger than a set threshold value into one category.

Wherein the first input layer is used for inputting patient data of a patient A; the first convolution layer is used for carrying out feature extraction on patient data of a patient A; the first pooling layer is used for pooling the result of the feature extraction; the first weighting and summarizing layer is used for carrying out weighting summation on the result output by the pooling layer; the first vector representation layer is used for representing the result after weighted summation;

wherein the second input layer is used for inputting the patient data of the patient B; the second convolutional layer is used for carrying out feature extraction on the patient data of the patient B; the second pooling layer is used for pooling the result of the feature extraction; the second weighting and summarizing layer is used for carrying out weighting summation on the result output by the pooling layer; the second vector representation layer is used for representing the result after weighted summation;

wherein, the matching matrix is a symmetric matrix with m rows and n columns, which is used for converting the similarity vectors of the patient A and the patient B.

It should be understood that, based on the step S2, performing data imbalance processing on the historical training samples, on the basis of the first convolutional neural network CNN framework, a matching matrix is introduced to construct a similarity learning framework to implement similar learning of the patient. As shown in FIG. 3, the framework firstly maps the high-dimensional sparse feature matrix of the patient into a low-dimensional dense matrix through an Embedding layer, and then divides the low-dimensional dense matrix into a plurality of sub-matrices. Single-sided convolution and maximum pooling is applied to each sub-matrix. The feature vectors of the sub-matrices are then aggregated to form the final feature representation vector for the patient. The feature representation vector of the patient obtains a similarity feature vector through the matching matrix and the conversion layer. Finally, the similarity feature vector is passed through the softmax layer to obtain the patient's similarity probability.

As one or more embodiments, the S3 includes:

s31: the original high-dimensional sparse feature matrix ignores the relation between the hospitalizing sequences, so that the feature matrix is high-dimensional andand (4) sparse. For less relationship between feature dimension and learning sequence, ReLU function is used to map feature matrix to a vector space. Each hospitalizing sequence v_iIs mapped to a vector x_i∈R^dAs follows:

x_i＝ReLU(W_vv_i+b_v) (4)

ReLU(x)＝max{x,0} (5)

where d denotes the mapping dimension, W_vAnd b_vThe learned feature matrix and basis vector are represented. After the embedding operation, the embedding matrix X epsilon R of each patient can be obtained^d。

S32: and acquiring a low-dimensional sparse vector based on S31, and performing convolution operation to obtain a feature set with strong low-latitude correlation. Assume a convolution kernel from visit vector x_iTo x_i+h-1Code for seeking medical treatment c_iUsing c_i＝ReLU(W_c·x_i+h-1+b_c). The convolution kernel is applied to each time window { x }_1:h,x_2:h+1,…,x_t-h+1:tGenerating a medical treatment coding vector mapping c ═ c₁,c₂,…,c_t-h+1In which c ∈ R^t-h+1. Since there are a total of m convolution kernels, m vector maps can be obtained. The output of the convolutional layer is applied to the pooling layer. The maximum pooling is applied to c as c '═ max { c }, where c' denotes the maximum value of the corresponding convolution kernel. The output connections of all the convolution kernels after pooling form a vector characterization h epsilon R^mAnd h is the vector characterization h of the original embedding matrix X.

S33: and (4) obtaining a vector characterization h based on S32, and introducing a matching matrix for similarity learning, as follows:

S＝h_AMh_B(6)

wherein h is_AShown is a vector representation of patient a; h is_BRepresented is a vector representation of patient B; the matching matrix M belongs to R^m ^×mIs symmetrical.

To guarantee the symmetric constraint of the matching matrix M, it is decomposed into M ═ L^TL, wherein L ∈R^g×m,g<m to ensure lower ranking features. And the medical sample is converted to obtain a similarity vector of the vector by considering the symmetric constraint condition, so that the patient is ensured not to influence the similarity score. First of all, convert h_AAnd h_BIn a vector dimension, by the formula:

wherein W_h∈R^m×m，

Is a bitwise addition operation.

Then H and S are mapped to a fully connected softmax layer resulting in an output probability y' representing the similarity between the two patients. Finally, the cross entropy is used to compute the loss function for y and y':

wherein

The total number of correctly classified patients is shown, since there are N patients

This similarity learning process is end-to-end, and all parameters are updated by back-propagation.

The first convolutional neural network CNN described in S3 is first trained to obtain optimized first convolutional neural network CNN parameters and a matching matrix. The similarity of each test case and all training data is then calculated and ranked using the first convolutional neural network CNN. The data set was randomly divided into a training set, a validation set, and a test set (0.75: 0.1:0.15) for similarity training. During the training process, the first convolutional neural network CNN is implemented using TensorFlow. Adam is used to optimize model parameters. Unlike the normal CNN model, which is imported as a small batch of patients, a similarity framework is trained on a batch of patient pairs to ensure that each patient pair can be measured.

In this example, the performance of the first convolutional neural network CNN is measured using the landed index, purity and NMI. Higher values mean more agreement between the grouped population and the true label, i.e., more similar samples are grouped together, indicating better similarity learning performance. Table 2 describes the comparison results of different similarity learning methods.

TABLE 2 different similarity learning method comparisons

Method of producing a composite material	Index of landed	Purity of	NMI
				Euclidean	0.4743	0.4633	0.0593
Cosine	0.4862	0.4654	0.0582
				GMML	0.5024	0.4822	0.0698
LMNN	0.5778	0.5374	0.1148
				K-means	0.6347	0.6659	0.2316
CNN_triple	0.7351	0.7561	0.3599

Based on the results in table 2, the performance of the first convolutional neural network CNN (CNN _ triple) proposed in this embodiment is superior to other methods.

It should be understood that the step of S4 includes:

and calculating the similarity of the patients based on S3, and combining the second convolutional neural network CNN to realize the doctor-seeking migration recommendation of the patients, namely recommending whether the patients need to have doctor-seeking migration. According to the method, firstly, the similarity between patients is calculated according to the first convolutional neural network CNN introduced in S3, and then people similar to the hospitalizing behavior of the patients are found out from the migrated people and the non-migrated people respectively according to the similarity. Substituting the medical data of similar crowd history as input into a second convolutional neural network CNN, and obtaining the feature vector representation W of the similar crowd through embedding, convolution, pooling and other operations_ic。

As one or more embodiments, it should be understood that the step of S5 includes:

representing vector W based on characteristics of similar population_cAcquiring comprehensive characteristic representation H of similar people by combining probability statistics of hospitalizing distribution data of various hospitals transmitted by a social security system, and updating weighted values of various characteristics by adopting an Attention (Attention) mechanism, namely:

u_t＝tanh(wH_t+b) (8)

δ＝∑_ta_tH_t(10)

wherein w ∈ R^L×|C|And b ∈ R^LRespectively representing a weight matrix and a basic matrix, L is the number of hospitals, C is the disease category, u_tAn attention vector representing hospital-features, T represents the number of attention degrees of each hospital-feature generated at time T, a_tWeight representing normalized Hospital-feature, δ being a_tTo H_tThe adjustment result of the weight; h_tRepresenting the comprehensive characteristic representation of the similar population at the time t; r represents a set of real numbers.

Combining the hospital-feature matching weight obtained by the Attention mechanism, the comprehensive feature vector of each similar patient in each similar patient group

Comprises the following steps:

where N represents the length of the patient's hospitalization sequence.

As one or more embodiments, the S6, comprising:

recommending hospitalizing behaviors of a sample to be predicted and outputting an Attention mechanism

Inputting softmax layer results in output values as follows:

Y＝max(y_A,y_B) (13)

and finally, recommending whether the patient is worthy of hospitalizing migration behavior through comparison of output values.

Recommending hospitalizing migration behaviors of the sample to be tested, pushing a recommendation result, and comparing the recommendation result with an actual hospitalizing migration behavior, wherein the accuracy of the method in the hospitalizing migration recommendations is described in table 3:

TABLE 3 accuracy of referral migration behavior recommendations

Based on the results in table 3, the performance of the proposed hospitalization migration recommendation model of this embodiment is superior to other methods. In addition, the proposed method for dealing with the data imbalance problem is superior to other methods.

It should be understood that the S7 further includes: based on the comparison between the recommendation result and the actual result, the Attention mechanism is adjusted by interpretable feedback of the recommendation result, the hospitalizing rule obtained by the bottom-layer personal hospitalizing data statistics is updated and perfected, the weight of each dimension is sorted according to the reverse order, and then K data with the top weight are selected, as shown in the following:

argsort(a_t[；i])[1:K](14)

in the formula, a_t[；i]The attention score of the ith characteristic dimension is represented, the importance degree of each characteristic vector to the recommendation of the hospitalizing behavior can be fed back and adjusted by analyzing K characteristic vectors with the weights being earlier, so that historical sample data is continuously updated, and the accuracy of the method is improved.

As one or more embodiments, after S7 and before S8, the method further includes:

and after the model training is finished, outputting a recommendation result of the training set, comparing the recommendation result with the known actual hospitalizing migration behavior, and feeding back and updating the CNN data information of the second convolutional neural network, so that the data weight value in the model is continuously optimized, and the hospitalizing migration recommendation is continuously perfected.

As one or more embodiments, in the second convolutional neural network CNN, the training of the second convolutional neural network CNN is implemented by using cross entropy as a loss function of the second convolutional neural network CNN.

The second embodiment also provides a hospitalizing migration scheme recommendation system based on similarity learning;

As one or more embodiments, the training module includes:

a first acquisition module configured to: acquiring historical hospitalizing data of all medical insurance personnel from medical insurance data;

a pre-processing module configured to: carrying out unbalance processing on the acquired historical hospitalizing data so as to obtain a training set;

a similar patient group output module configured to: inputting the training set into a first Convolutional Neural Network (CNN), and outputting a plurality of similar patient groups, wherein each similar patient group comprises a plurality of similar patients;

a first feature extraction module configured to: performing feature extraction on the historical hospitalizing data of each similar patient in each similar patient group by using a second Convolutional Neural Network (CNN) to obtain a feature expression vector of each similar patient in each similar patient group;

a first weight update module configured to: updating the weight value of each characteristic value in the characteristic expression vector for the characteristic expression vector of each similar patient in each similar patient group by adopting an attention mechanism, and carrying out weighted summation on each characteristic value to obtain a comprehensive characteristic vector of each similar patient in each similar patient group;

a first recommendation proportion output module configured to: inputting the comprehensive characteristic vector of each similar patient in each similar patient group into a softmax function, and outputting a recommended proportion value of each similar patient for transferring to a different hospital or not recommending the transfer;

a training module configured to: calculating a loss function of the output value of the softmax function, and training the learning parameters of the second convolutional neural network CNN by adopting a back propagation algorithm to complete the training of the second convolutional neural network CNN;

as one or more embodiments, the application module includes:

a second acquisition module configured to: acquiring hospitalizing data of a patient to be recommended, inputting the hospitalizing data and a training set of the patient to be recommended into a first Convolutional Neural Network (CNN), and outputting a similar patient group of the patient to be recommended;

a second feature extraction module configured to: performing feature extraction on the historical hospitalizing data of each similar patient in the similar patient group of the patient to be recommended by using the trained second convolutional neural network CNN to obtain a feature representation vector of each similar patient in the similar patient group of the patient to be recommended;

a second weight update module configured to: updating the weight value of each characteristic value in the characteristic expression vector by adopting an attention mechanism, and carrying out weighted summation on each characteristic value to obtain a comprehensive characteristic vector of each similar patient in the similar patient group of the patient to be recommended;

a second recommendation proportion output module configured to: and inputting the comprehensive characteristic vector of each similar patient in the similar patient group of the patient to be recommended into the softmax function, and outputting recommended proportion values of the patient to be recommended, which are transferred to different hospitals or recommended proportion values of the patient not to be recommended, to be transferred.

In a third embodiment, the present embodiment further provides an electronic device, which includes a memory, a processor, and computer instructions stored in the memory and executed on the processor, where the computer instructions, when executed by the processor, implement the steps of the method in the first embodiment.

In a fourth embodiment, the present embodiment further provides a computer-readable storage medium for storing computer instructions, and the computer instructions, when executed by a processor, perform the steps of the method in the first embodiment.

The above description is only a preferred embodiment of the present application and is not intended to limit the present application, and various modifications and changes may be made by those skilled in the art. Any modification, equivalent replacement, improvement and the like made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. The hospitalizing migration scheme recommendation method based on similarity learning is characterized by comprising the following steps of: a training phase and an application phase;

2. The method of claim 1, wherein the training phase comprises: S1-S7;

s7: and calculating a loss function of the output value of the softmax function, and training the learning parameters of the second convolutional neural network CNN by adopting a back propagation algorithm to finish the training of the second convolutional neural network CNN.

3. The method of claim 1, wherein the application phase comprises: S8-S11;

4. The method as set forth in claim 2, wherein,

in the step S2, performing imbalance processing on the acquired historical hospitalization data to obtain a training set; wherein the unbalanced processing is realized by utilizing an adaptive boundary sampling algorithm AB-SMOTE algorithm;

alternatively, the first and second electrodes may be,

in S3, the first convolutional neural network CNN is an improved convolutional neural network; the first convolutional neural network CNN, including: two parallel branch lines;

the similarity of the first output value and the second output value is expressed and input into a softmax function, and the softmax function outputs the similarity; classifying patients with similarity greater than a set threshold into one category;

5. The method as claimed in claim 2, wherein the step of S5 comprises:

based on the use of similar groups of peopleFeature representation vector W_cAnd acquiring comprehensive characteristic representation H of similar people by combining probability statistics of hospitalizing distribution data of various hospitals transmitted by the social security system, and updating the weighted value of each characteristic by adopting an Attention mechanism, namely:

u_t＝tanh(wH_t+b) (8)

δ＝∑_ta_tH_t(10)

wherein w ∈ R^L×|C|And b ∈ R^LRespectively representing a weight matrix and a basic matrix, L is the number of hospitals, C is the disease category, u_tAn attention vector representing hospital-features, T represents the number of attention degrees of each hospital-feature generated at time T, a_tWeight representing normalized Hospital-feature, δ being a_tTo H_tThe adjustment result of the weight; h_tRepresenting the comprehensive characteristic representation of the similar population at the time t; r represents a set of real numbers;

Comprises the following steps:

wherein N represents the length of the patient hospitalization sequence;

alternatively, the first and second electrodes may be,

the S6, including:

Inputting softmax layer results in output values as follows:

Y＝max(y_A,y_B) (13)

finally, recommending whether the patient is worth hospitalizing migration behavior through comparison of output values;

alternatively, the first and second electrodes may be,

the S7 further includes: based on the comparison between the recommendation result and the actual result, the Attention mechanism is adjusted by interpretable feedback of the recommendation result, the hospitalizing rule obtained by the bottom-layer personal hospitalizing data statistics is updated and perfected, the weight of each dimension is sorted according to the reverse order, and then K data with the top weight are selected, as shown in the following:

argsort(a_t[；i])[1:K](14)

in the formula, a_t[；i]The attention score of the ith characteristic dimension is expressed, and the importance of each characteristic vector to the recommendation of the hospitalizing behavior can be fed back and adjusted by analyzing K characteristic vectors with the weights being earlier, so that historical sample data is continuously updated, and the accuracy of the method is improved;

alternatively, the first and second electrodes may be,

after the S7 and before the S8, the method further comprises:

6. A hospitalizing migration scheme recommendation system based on similarity learning is characterized by comprising: a training module and an application module;

7. The system of claim 6, wherein the training module comprises:

a training module configured to: and calculating a loss function of the output value of the softmax function, and training the learning parameters of the second convolutional neural network CNN by adopting a back propagation algorithm to finish the training of the second convolutional neural network CNN.

8. The system of claim 6, wherein the application module comprises:

9. An electronic device comprising a memory and a processor and computer instructions stored on the memory and executed on the processor, the computer instructions, when executed by the processor, performing the steps of the method of any of claims 1-5.

10. A computer readable storage medium storing computer instructions which, when executed by a processor, perform the steps of the method of any one of claims 1 to 5.