CN113241193A

CN113241193A - Drug recommendation model training method, recommendation method, device, equipment and medium

Info

Publication number: CN113241193A
Application number: CN202110609395.5A
Authority: CN
Inventors: 李佳琳; 王健宗
Original assignee: Ping An Technology Shenzhen Co Ltd
Current assignee: Ping An Technology Shenzhen Co Ltd
Priority date: 2021-06-01
Filing date: 2021-06-01
Publication date: 2021-08-10

Abstract

The invention discloses a medicine recommendation model training method, a recommendation method, a device, equipment and a medium. The method comprises the following steps: obtaining model training data, including training drugs, training disease data and training comment data; performing sentiment analysis on the training comment data to obtain a labeled sentiment label; extracting the characteristics of training disease data and training comment data to obtain training text characteristics; obtaining a model training sample based on each training text feature and the labeled emotion label; inputting the model training sample into a feature classifier corresponding to the training text features for model training to obtain an original feature drug recommendation model corresponding to the training text features; and testing the original characteristic drug recommendation model by adopting the model training sample to obtain a target characteristic drug recommendation model corresponding to the training text characteristic. The method can guarantee the accuracy and effectiveness of the target recommended drugs recommended by the target characteristic drug recommendation model.

Description

Drug recommendation model training method, recommendation method, device, equipment and medium

Technical Field

The invention relates to the technical field of intelligent medical treatment, in particular to a medicine recommendation model training method, a recommendation method, a device, equipment and a medium.

Background

When a doctor prescribes a medicine for a patient, the doctor mainly makes a prescription scheme based on the knowledge and experience of the doctor, namely, the knowledge and experience of the doctor are utilized to analyze the conditions of disease symptoms, past clinical history, related medicine performance and the like to determine the final prescription scheme. In this way, whether the medicine prescribed by the doctor is effective for the disease symptoms of the patient depends mainly on the knowledge and experience of the doctor, and whether the medicine prescribed by the doctor and the disease symptoms of the patient are effective cannot be verified, which is not beneficial to ensuring the life health of the patient. The conventional medicine recommendation system is mainly used for carrying out model training according to historical recommended medicines and corresponding disease condition data thereof, and cannot reflect whether the recommended medicines are effective to diseases or not, so that the recommendation accuracy and effectiveness of the medicine recommendation system are low.

Disclosure of Invention

The embodiment of the invention provides a medicine recommendation model training method and device, computer equipment and a storage medium, and aims to solve the problems of low recommendation accuracy and effectiveness of an existing medicine recommendation system.

A method of drug recommendation model training, comprising:

obtaining model training data, wherein the model training data comprises training medicines, training disease data corresponding to the training medicines and training comment data;

performing emotion analysis on training comment data in the model training data to obtain a labeled emotion label corresponding to the model training data;

performing feature extraction on the training disease data and the training comment data to obtain training text features corresponding to the model training data, wherein the training text features comprise Bow features, TF-IDF features, Word2Vec features and manual features;

obtaining model training samples based on each training text feature and the labeled emotion label, storing the model training samples in a feature sample set corresponding to the training text feature, and dividing the model training samples in the feature sample set into a feature training set and a feature test set;

inputting the model training samples in the feature training set into a feature classifier corresponding to the features of the training text for model training, and acquiring an original feature drug recommendation model corresponding to the features of the training text;

and testing the original characteristic drug recommendation model corresponding to the training text characteristic by adopting the model training sample in the characteristic test set to obtain a target characteristic drug recommendation model corresponding to the training text characteristic.

A medication recommendation model training device, comprising:

the model training data acquisition module is used for acquiring model training data, wherein the model training data comprises training medicines, training disease data corresponding to the training medicines and training comment data;

the marked emotion label acquisition module is used for carrying out emotion analysis on training comment data in the model training data and acquiring a marked emotion label corresponding to the model training data;

the training text feature acquisition module is used for extracting features of the training disease data and the training comment data to acquire training text features corresponding to the model training data, wherein the training text features comprise Bow features, TF-IDF features, Word2Vec features and manual features;

the model training sample acquisition module is used for acquiring a model training sample based on each training text feature and the labeled emotion label, storing the model training sample in a feature sample set corresponding to the training text feature, and dividing the model training sample in the feature sample set into a feature training set and a feature test set;

the original recommendation model acquisition module is used for inputting the model training samples in the feature training set into the feature classifier corresponding to the training text features for model training to acquire an original feature drug recommendation model corresponding to the training text features;

and the target recommendation model acquisition module is used for testing the original characteristic drug recommendation model corresponding to the training text characteristics by adopting the model training samples in the characteristic test set to acquire the target characteristic drug recommendation model corresponding to the training text characteristics.

A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, the processor implementing the above drug recommendation model training method when executing the computer program.

A computer-readable storage medium, in which a computer program is stored which, when being executed by a processor, implements the above-mentioned medication recommendation model training method.

In the drug recommendation model training method, the drug recommendation model training device, the computer equipment and the storage medium, emotion analysis is performed on training comment data in the model training data, and a labeled emotion label corresponding to the model training data is obtained, so that whether the treatment effect of a comment staff training the comment data is approved to a relevant drug is determined. Extracting features of the training disease data and the training comment data to obtain training text features such as Bow features, TF-IDF features, Word2Vec features, manual features and the like, and being beneficial to guaranteeing comprehensiveness of the extracted features; and the four training text characteristics, namely the Bow characteristic, the TF-IDF characteristic, the Word2Vec characteristic and the manual characteristic, are respectively combined with the labeling emotion label to form a model training sample, so that the quantity and the acquisition efficiency of the model training sample are favorably ensured. And model training samples in the same characteristic sample set are adopted, and a characteristic classifier corresponding to the training text characteristics is input for model training and testing, so that the training efficiency of the target characteristic drug recommendation model is favorably ensured, and the accuracy and the effectiveness of the target characteristic drug recommendation model in drug recommendation can be ensured.

The embodiment of the invention provides a medicine recommendation method, a medicine recommendation device, computer equipment and a storage medium, and aims to solve the problems of low recommendation accuracy and effectiveness of an existing medicine recommendation system.

A medication recommendation method comprising:

acquiring disease condition data to be processed, and determining N available medicines matched with the disease condition data to be processed according to the disease condition data to be processed;

acquiring M pieces of comment data to be processed corresponding to each available medicine;

performing feature extraction on the disease data to be processed and the comment data to be processed to obtain target text features corresponding to the comment data to be processed, wherein the target text features comprise Bow features, TF-IDF features, Word2Vec features and manual features;

inputting the target text characteristics corresponding to the comment data to be processed into a target characteristic drug recommendation model corresponding to the target text characteristics for analysis processing, and acquiring a comment recommendation score corresponding to each comment data to be processed;

comprehensively processing the comment recommendation scores corresponding to the M pieces of comment data to be processed corresponding to each available medicine to obtain a medicine recommendation score corresponding to each available medicine;

and determining a target recommended medicament according to the medicament recommendation scores corresponding to all the available medicaments.

A medication recommendation device comprising:

the available medicine determining module is used for acquiring the disease data to be processed and determining N available medicines matched with the disease data to be processed according to the disease data to be processed;

the comment data to be processed acquisition module is used for acquiring M pieces of comment data to be processed corresponding to each available medicine;

the target text feature acquisition module is used for performing feature extraction on the disease data to be processed and the comment data to be processed to acquire target text features corresponding to the comment data to be processed, wherein the target text features comprise a Bow feature, a TF-IDF feature, a Word2Vec feature and a manual feature;

the comment recommendation score acquisition module is used for inputting the target text characteristics corresponding to the to-be-processed comment data into a target characteristic medicine recommendation model corresponding to the target text characteristics for analysis processing, and acquiring a comment recommendation score corresponding to each to-be-processed comment data;

the drug recommendation score acquisition module is used for comprehensively processing the comment recommendation scores corresponding to the M pieces of comment data to be processed corresponding to each available drug to acquire the drug recommendation score corresponding to each available drug;

and the target recommended medicine determining module is used for determining the target recommended medicines according to the medicine recommendation scores corresponding to all the available medicines.

A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, the processor implementing the above-mentioned medication recommendation method when executing the computer program.

A computer-readable storage medium, in which a computer program is stored which, when being executed by a processor, carries out the above-mentioned medication recommendation method.

According to the medicine recommendation method, the device, the computer equipment and the storage medium, feature extraction is carried out on each to-be-processed comment data and to-be-processed disease data, four target text features, namely, a Bow feature, a TF-IDF feature, a Word2Vec feature and a manual feature, are determined, then the four target text features are input into a target feature medicine recommendation model for analysis and processing, and a comment recommendation score corresponding to each to-be-processed comment data is determined, so that the comprehensiveness of the comment recommendation score is favorably guaranteed, and the accuracy and the effectiveness for evaluating whether an available medicine can treat the current patient suffering from a disease are improved. And then, comprehensively processing the comment recommendation scores corresponding to the M pieces of to-be-processed comment data corresponding to each available medicine, and acquiring the medicine recommendation score corresponding to each available medicine, so that the medicine recommendation scores comprehensively consider the comment recommendation scores corresponding to all the to-be-processed comment data, and the accuracy and the effectiveness of evaluating whether the available medicine can treat the current patient suffering from the disease are guaranteed. And finally, determining a target recommended medicament based on the medicament recommendation scores corresponding to all the available medicaments so as to ensure the accuracy and effectiveness of the target recommended medicament for treating diseases corresponding to the to-be-treated disease data and assist doctors in completing a prescription scheme.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings needed to be used in the description of the embodiments of the present invention will be briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art that other drawings can be obtained according to these drawings without inventive labor.

FIG. 1 is a diagram of an application environment of a medication recommendation system in accordance with an embodiment of the present invention;

FIG. 2 is a flowchart of a method for training a medication recommendation model according to an embodiment of the present invention;

FIG. 3 is another flow chart of a method for training a medication recommendation model in accordance with an embodiment of the present invention;

FIG. 4 is another flow chart of a method for training a medication recommendation model in accordance with an embodiment of the present invention;

FIG. 5 is a flowchart of a method for recommending medications according to an embodiment of the present invention;

FIG. 6 is another flow chart of a method for medication recommendation in an embodiment of the present invention;

FIG. 7 is another flow chart of a method for medication recommendation in an embodiment of the present invention;

FIG. 8 is a schematic diagram of a drug recommendation model training device in accordance with an embodiment of the present invention;

FIG. 9 is a schematic view of a medication recommendation device in accordance with an embodiment of the present invention;

FIG. 10 is a schematic diagram of a computer device according to an embodiment of the invention.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are some, not all, embodiments of the present invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

According to the medicine recommendation model training method provided by the embodiment of the invention, the medicine recommendation model training method can be applied to an application environment shown in fig. 1. Specifically, the drug recommendation model training method is applied to a drug recommendation system, the drug recommendation system comprises a client and a server shown in fig. 1, and the client and the server are in communication through a network and are used for achieving drug recommendation model training and drug recommendation by using a drug recommendation model obtained through training so as to guarantee the symptomatic treatment effect of a recommended target recommended drug on a specific disease. The client is also called a user side, and refers to a program corresponding to the server and providing local services for the client. The client may be installed on, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server may be implemented as a stand-alone server or as a server cluster consisting of a plurality of servers.

In an embodiment, as shown in fig. 2, a method for training a medication recommendation model is provided, which is described by taking the server in fig. 1 as an example, and includes the following steps:

s201: model training data is obtained, wherein the model training data comprises training medicines, training disease data corresponding to the training medicines and training comment data.

S202: and carrying out emotion analysis on training comment data in the model training data to obtain a labeled emotion label corresponding to the model training data.

S203: and extracting the characteristics of the training disease data and the training comment data to obtain training text characteristics corresponding to the model training data, wherein the training text characteristics comprise a Bow characteristic, a TF-IDF characteristic, a Word2Vec characteristic and a manual characteristic.

S204: and obtaining model training samples based on each training text characteristic and the labeled emotion label, storing the model training samples in a characteristic sample set corresponding to the training text characteristic, and dividing the model training samples in the characteristic sample set into a characteristic training set and a characteristic testing set.

S205: and inputting the model training samples in the feature training set into a feature classifier corresponding to the features of the training text for model training to obtain an original feature drug recommendation model corresponding to the features of the training text.

S206: and inputting the model training samples in the feature training set into a feature classifier corresponding to the features of the training text for model training to obtain an original feature drug recommendation model corresponding to the features of the training text.

The model training data is data for performing model training. Training drugs refer to drug names recorded in model training data. The training disease data refers to disease data used for recording diseases of patients in model training data, and can be, but is not limited to, medical history texts of patients. The training comment data is used in the model training data to record patient comments on the training medication.

As an example, in step S201, the server may obtain, through a web crawler or other online and offline technology, training comment data for a certain patient to score whether a disease corresponding to training medicine treatment training condition data is effective, so as to obtain a set of model training data, where each set of model training data includes training medicine, training condition data corresponding to the training medicine, and training comment data, and the training comment data may include or may not include a satisfaction score. The satisfaction score is the score that a particular platform scores with the patient for whether the training medication is effective for the disease.

The marked emotion labels are used for reflecting the emotional tendency of the patient to the training medicine, and comprise positive labels and negative labels. The positive label is a label used for reflecting the positive identity degree of a patient on whether the patient can treat the disease suffered by the patient with the training drug. The negative label is a label used for reflecting the negative repudiation attitude of the patient on whether the training drug can treat the suffered disease.

As an example, in step S202, after obtaining the model training data, the server may obtain the labeled emotion label corresponding to the model training sample by performing emotion analysis on the training comment data in the model training data. For example, the training comment data can be analyzed using a neural network model for analyzing emotional tendencies to determine labeled emotional tags corresponding to the model training data. For another example, emotion analysis can be performed on the satisfaction degree score to obtain a labeled emotion label corresponding to the model training sample.

In this example, step S202, that is, performing emotion analysis on training comment data in the model training data to obtain a labeled emotion label corresponding to the model training data, includes: if the training comment data in the model training data contain the satisfaction score, performing emotion analysis based on the satisfaction score to obtain a labeled emotion label corresponding to the model training data; and if the training comment data in the model training data do not contain the satisfaction degree score, performing emotion analysis on the training comment data in the model training data by adopting an emotion analysis model to obtain a labeled emotion label corresponding to the model training data.

As an example, the server may compare the satisfaction score to an emotion score threshold; if the satisfaction degree score is larger than the emotion score threshold value, marking the emotion label as a forward label; and if the satisfaction degree score is not larger than the situation score threshold value, marking the emotion label as a negative label. The emotion score threshold is a preset threshold for evaluating the scores of the positive and negative tags. In the example, the satisfaction score and the emotion score threshold are compared to determine the mode of marking the emotion label for emotion analysis, the processing process is simple and convenient, and the emotion analysis processing efficiency is improved.

The emotion analysis model is a model which is trained on the basis of a neural network model in advance and is used for realizing emotion analysis, and emotion analysis (sentiment analysis) herein superficially refers to emotion mining and analysis on text, images, audio, video and even cross-modal data by using a computer technology.

As an example, the server may use a pre-trained emotion analysis model to perform emotion analysis on training comment data in the model training data, and obtain a labeled emotion tag corresponding to the model training data, so as to determine whether a comment person who trains the comment data agrees with a treatment effect of a related medicine.

The training text features refer to features obtained by feature extraction of text data (i.e., training condition data and training comment data) in model training data. In this example, the training text features include four feature types, Bow feature, TF-IDF feature, Word2Vec feature, and Manual feature.

The Bow feature refers to a feature determined by processing text data in model training data based on the Bow (Bag of words, abbreviated as "Bag of words") technology, and the Bow technology is 1) an algorithm used in natural language processing and is responsible for counting the times of all tokens in comments or documents. A term or token may be referred to as a word (unigram) or any subjective number of words. In this scenario, a (1,2) n-gram range is selected.

The TF-IDF characteristic refers to a characteristic determined by processing text data in the model training data by adopting a TF-IDF algorithm. TF-IDF is a commonly used weighting technique for information retrieval and data mining, where words are provided with negligible weight. The importance of a word is proportional to the number of times it appears in the article and inversely proportional to the number of times it appears in the corpus. The calculation mode can effectively avoid the influence of the common words on the keywords, and improves the correlation between the keywords and the articles. The Frequency (TF) may be referred to as the likelihood of locating a word in the document, i.e., TF (t, d) log (1+ freq (t, d)). Inverse Document Frequency (IDF) is the number of times a particular term appears in the entire corpus as opposed to it, which captures the particular term in a document, i.e.

The product of The Frequency (TF) and the Inverse Document Frequency (IDF) (i.e., TF x IDF) is determined as the importance and relevance of a vocabulary in the document.

The Word2Vec feature is a feature determined by processing text data in the model training data by adopting the Word2Vec technology. The Word2Vec technique refers to Word embedding from a large corpus using various deep learning models to convert "non-calculable" and "unstructured" words into "calculable" and "structured" vectors. The main idea of Word2Vec technology is to arrange Word vectors in a vector space according to the meanings of words, and finally, the purpose is to find out words with similar meanings in a data set, wherein the words are close to each other in the vector space.

The manual features refer to features acquired by means other than feature engineering technology, specifically features determined manually by a user, and include but are not limited to features formed by the user screening out keywords related to drug recommendation from text data of model training data. As an example, the server may select a specific keyword from text data of the model training data by using a keyword matching algorithm in a keyword information table of keywords related to drug recommendation configured in advance by the user, and determine the specific keyword as a manual feature.

As an example, in step S203, the server may use a feature engineering technology such as Bow, TF-IDF, and Word2Vec to perform feature extraction on text data in the model training data, that is, perform feature extraction on training disorder data and training comment data, and obtain training text features such as Bow feature, TF-IDF feature, and Word2Vec feature, so as to perform model training by using Bow feature, TF-IDF feature, and Word2Vec feature in the following. In this example, in order to ensure the comprehensiveness of feature extraction, the server may further extract manual features from the training condition data and the training comment data in a manner closely related to the manual operation of the user besides the feature engineering technologies such as Bow, TF-IDF, and Word2Vec, specifically including but not limited to the following manners: one is that the server can acquire manual characteristics which are input by the user and used for manually marking and determining training disease data and training comment data. Secondly, the server can adopt a keyword matching algorithm to extract the characteristics of the specific keywords of the training disease data and the training comment data so as to obtain the manual characteristics.

In this example, the server performs feature extraction on training condition data and training comment data to obtain training text features corresponding to model training data, and the method specifically includes: (1) and performing text preprocessing on the training disease data and the training comment data, wherein the text preprocessing includes but is not limited to text preprocessing operations such as deleting specific symbols, segmenting words, removing stop words and the like, and obtaining original text segmentation words. (2) And (3) performing feature extraction on all original text participles by adopting feature engineering technologies such as Bow, TF-IDF, Word2Vec and the like to obtain training text features such as Bow feature, TF-IDF feature, Word2Vec feature and the like. (3) And manually labeling the original text participles by adopting a manual labeling or keyword matching algorithm to obtain the training text characteristic of the manual characteristic.

In the example, training text features such as the Bow feature, the TF-IDF feature, the Word2Vec feature and the manual feature can be obtained by extracting features of training condition data and training comment data, so that the training text features corresponding to model training data are more comprehensive, the matching degree of finally obtained medicines recommended by the medicine recommendation model and diseases suffered by patients is higher, and the accuracy and the effectiveness of the model recommended medicines are improved.

The model training sample is a sample which needs to be input into the feature classifier for model training. The feature classifier is a classifier for constructing a target feature drug recommendation model. The feature sample set is a set used for storing all model training samples corresponding to a certain training text feature, and in this example, the feature sample set includes a Bow feature sample set, a TF-IDF feature sample set, a Word2Vec feature sample set, and a manual feature sample set. The feature training set is a set of model training samples stored for model training. The feature test set is a collection of model training samples stored for model testing.

As an example, in step S204, the server may first obtain a model training sample based on each training text feature and the labeled emotion label. For example, a model training sample can be formed based on the "Bow feature and the labeled emotion label", a model training sample can be formed based on the "TF-IDF feature and the labeled emotion label", a model training sample can be formed based on the "Word 2Vec feature and the labeled emotion label", and a model training sample can be formed based on the "manual feature and the labeled emotion label". Then, the server may store the model training samples in a feature sample set corresponding to the training text features, for example, store the model training samples formed by "Bow feature and label emotion label" in the Bow feature sample set, store the model training samples formed by "TF-IDF feature and label emotion label" in the TF-IDF feature sample set, store the model training samples formed by "Word 2Vec feature and label emotion label" in the Word2Vec feature sample set, and store the model training samples formed by "manual feature and label emotion label" in the manual feature sample set. Finally, the server may randomly divide all feature sample sets into a feature training set and a feature testing set based on the same division ratio (e.g., 8: 2), so as to ensure that the number of samples in all feature training sets is the same and the number of samples in all feature testing sets is the same.

The feature classifier corresponding to the training text features refers to a classifier matched based on the feature attributes of the training text features.

As an example, in step S205, the server may input the Bow feature and the TF-IDF feature into any one of a logistic regression, a polynomial naive bayes algorithm, a stochastic gradient descent, a support vector classifier, a perceptron classification and a ridge regression classification, respectively, to perform model training, so as to obtain an original feature drug recommendation model corresponding to the Bow feature and an original feature drug recommendation model corresponding to the TF-IDF feature. The server can input the Word2Vec features and the manual features into any one of the decision tree, the random forest, the LGBM and the Catboost classifier respectively for model training, and an original feature drug recommendation model corresponding to the Word2Vec features and an original feature drug recommendation model corresponding to the manual features are obtained. In this example, the Bow feature and the TF-IDF feature are generally very sparse matrices, and when the matrices are applied to feature classifiers (such as a decision tree) corresponding to two training text features, namely, the Word2Vec feature and the manual feature, the time consumption of the model training process is long. On the contrary, when two training text features, namely the Word2Vec feature and the manual feature, are applied to the feature classifier corresponding to the Bow feature and the TF-IDF feature, the model training effect is poor. In this example, each training text feature is subjected to model training by using a feature classifier corresponding to the feature attribute of the feature classifier, which is helpful for ensuring the training effect and the processing timeliness of the original feature drug recommendation model corresponding to each training text feature.

As an example, in step S206, the server needs to test the original feature drug recommendation model corresponding to the feature of the training text by using the model training sample in the feature test set, and obtains a model test result, which specifically includes: inputting model training samples in the feature test set into an original feature drug recommendation model corresponding to the feature of the training text for emotion analysis, and acquiring a predicted emotion label corresponding to each model training sample output by the original feature drug recommendation model; and carrying out comprehensive analysis according to whether the labeled emotion label corresponding to each model training sample is consistent with the predicted emotion label or not so as to obtain a model test result, wherein the model test result comprises a test passing mode and a test failing mode. Then, when the model test result is that the test is passed, the server can determine the original characteristic drug recommendation model corresponding to the training text characteristic as the target characteristic drug recommendation model corresponding to the training text characteristic, which is helpful for ensuring the accuracy and effectiveness of subsequent drug recommendation based on the target characteristic drug recommendation model.

In this example, since the training text features include the Bow feature, the TF-IDF feature, the Word2Vec feature, and the manual feature, the target feature drug recommendation model corresponding to the training text features specifically includes a Bow feature drug recommendation model, a TF-IDF feature drug recommendation model, a Word2Vec feature drug recommendation model, and a manual feature drug recommendation model.

In the drug recommendation model training method provided by this embodiment, emotion analysis is performed on training comment data in model training data, and a labeled emotion label corresponding to the model training data is obtained, so as to determine whether a comment person who trains the comment data agrees with a treatment effect of a relevant drug. The training disease data and the training comment data are subjected to feature extraction, and training text features such as Bow features, TF-IDF features, Word2Vec features, manual features and the like are obtained, so that the comprehensiveness of the extracted features is guaranteed; and the four training text characteristics, namely the Bow characteristic, the TF-IDF characteristic, the Word2Vec characteristic and the manual characteristic, are respectively combined with the labeled emotion label to form a model training sample, so that the quantity and the acquisition efficiency of the model training sample are favorably ensured. Model training samples in the same characteristic sample set are adopted, and a characteristic classifier corresponding to training text characteristics is input for model training and testing, so that the training efficiency of the target characteristic drug recommendation model is guaranteed, and the accuracy and the effectiveness of the target characteristic drug recommendation model in drug recommendation can be guaranteed.

In an embodiment, as shown in fig. 3, before step S204, that is, before inputting the model training samples in the feature training set into the feature classifier corresponding to the training text features for model training and obtaining the original feature drug recommendation model corresponding to the training text features, the drug recommendation model training method further includes:

s301: and counting the positive samples, the negative samples and the total samples in the feature training set.

S302: and acquiring a positive and negative sample difference value according to the positive sample number and the negative sample number.

S303: and acquiring a sample deviation proportion according to the positive and negative sample difference value and the total sample number.

S304: and if the sample deviation proportion is larger than the deviation proportion threshold value, performing oversampling processing on the model training samples in the characteristic training set by adopting an oversampling technology to obtain updated model training samples.

The deviation proportion threshold is a preset proportion threshold used for evaluating a standard for determining the quantity balance or unbalance of the positive and negative samples, and can be represented by Pm, namely, the proportion threshold is used for evaluating whether the quantity of the model training samples carrying the positive labels and the quantity of the model training samples carrying the negative labels are balanced.

As an example, in step S301, the server counts the number of all model training samples carrying forward labels in the feature training set, and determines the number as the number of positive samples, which can be represented by Sp. The server also counts the number of all model training samples carrying negative labels in the feature training set, determines the number as the negative sample number, and can be represented by Sn. Next, the server may further obtain a total sample number of the feature training set according to the positive sample number Sp and the negative sample number Sn, which may be represented by Ssum, where Ssum is Sp + Sn.

As an example, in step S302, the server may obtain a positive and negative sample difference of the feature training set according to the positive sample number Sp and the negative sample number Sn, where the positive and negative sample difference may be an absolute value of a difference between the positive sample number Sp and the negative sample number Sn, and may be represented by Sd, and then Sd is abs (Sp-Sn).

As an example, in step S303, the server may determine the sample deviation ratio Pd ═ Sd/Ssum ═ abs (Sp-Sn)/(Sp + Sn) according to the positive and negative sample difference Sd and the total number of samples Ssum. In this example, the sample deviation ratio may reflect the quantity deviation degree of model training samples carrying positive labels and negative labels in the feature training set, and may be used to reflect the balance degree of model training samples carrying two kinds of label emotion labels.

As an example, in step S304, the server may perform the following steps: (1) the sample deviation ratio Pd is compared to a deviation ratio threshold Pm. (2) If the sample deviation ratio Pd is greater than the deviation ratio threshold Pm, the sample deviation ratio Pd is determined to be large, the quantity deviation degree of the model training samples of the positive labels and the negative labels is large, at this time, the server needs to adopt an oversampling technology to perform oversampling processing on the model training samples in the feature training set, and obtain updated model training samples. For example, the model training samples in the feature training set are oversampled by using an oversampling technique (including but not limited to the SMOTE technique), and updated model training samples are obtained to prevent the problem of imbalance between positive and negative samples. The SMOTE technology is to synthesize new minority class samples, and the synthesis strategy is to randomly select a sample b from the nearest neighbor of each minority class sample a and then randomly select a point on a connecting line between a and b as the newly synthesized minority class sample. (3) If the sample deviation ratio Pd is larger than the deviation ratio threshold Pm, the sample deviation ratio Pd is determined to be smaller, the quantity deviation degree of model training samples of the positive labels and the negative labels is smaller, and the quantity of the positive samples and the negative samples is more balanced, at this moment, the model training samples in the feature training set can be executed, the model training samples are input into the feature classifier corresponding to the features of the training text for model training, and the processing process of obtaining the original feature medicine recommendation model corresponding to the features of the training text is obtained.

In the example, a sample deviation proportion is obtained through statistics by utilizing the number of positive samples, the number of negative samples and the total number of samples in the feature training set, when the sample deviation proportion is larger than a deviation proportion threshold value, the sample imbalance standard is determined to be met, at the moment, an oversampling technology is adopted to perform oversampling processing on the model training samples in the feature training set, updated model training samples are obtained, the balance of the number of the positive samples and the negative samples in the feature training set is favorably ensured, and the accuracy and the effectiveness of medicine recommendation of a target feature medicine recommendation model obtained through training are further ensured.

In an embodiment, as shown in fig. 4, in step S205, that is, testing the original feature drug recommendation model corresponding to the training text feature by using the model training samples in the feature test set, and acquiring the target feature drug recommendation model corresponding to the training text feature includes:

s401: and inputting the model training samples in the characteristic test set into the original characteristic drug recommendation model corresponding to the training text characteristics for identification, and acquiring the predicted emotion labels corresponding to the model training samples.

S402: and counting the positive correct times, the negative correct times, the positive error times and the negative error times according to the labeled emotion label and the predicted emotion label corresponding to the model training sample.

S403: and acquiring the current precision, the current withdrawal rate, the current F1score, the current accuracy and the current AUC of the original characteristic drug recommendation model corresponding to the training text characteristics according to the positive correct times, the negative correct times, the positive error times and the negative error times.

S404: and obtaining a model test result of the original characteristic drug recommendation model corresponding to the training text characteristic according to the current precision, the current withdrawal rate, the current F1score, the current accuracy and the current AUC.

S405: and if the model test result is that the test is passed, determining the original characteristic drug recommendation model corresponding to the training text characteristic as the target characteristic drug recommendation model corresponding to the training text characteristic.

The emotion label prediction method comprises the steps of predicting an emotion label by using an original characteristic drug recommendation model corresponding to a training text characteristic, and processing a model training sample in a characteristic test set to output the emotion label.

As an example, in step S401, the server inputs the model training samples in the feature test set one by one into the original feature drug recommendation models corresponding to the features of the training text for identification, and obtains the predicted emotion labels corresponding to the model training samples, so as to perform comparative analysis by using the predicted emotion labels and the labeled emotion labels corresponding to the model training samples, and further evaluate whether the original feature drug recommendation models corresponding to the features of the training text meet the standard.

As an example, in step S402, the server performs comparison analysis on the labeled emotion tag and the predicted emotion tag corresponding to the model training sample, and obtains an analysis result corresponding to each model training sample. For example, if the labeled emotion label and the predicted emotion label corresponding to the model training sample are both forward labels, the analysis result corresponding to the model training sample is forward correct; if the marked emotion label and the predicted emotion label corresponding to the model training sample are negative labels, the analysis result corresponding to the model training sample is negative correct; if the negative label of the labeled emotion label corresponding to the model training sample and the predicted emotion labels are positive labels, the analysis result corresponding to the model training sample is a positive error; if the emotion label positive label is labeled corresponding to the model training sample, and the predicted emotion labels are negative labels, the analysis result corresponding to the model training sample is a negative error.

In this example, after obtaining the analysis result corresponding to each model training sample, the server performs statistical analysis on the analysis results corresponding to all model training samples in the feature test set, determines the number of forward correct analysis results corresponding to the model training samples as the number of forward correct times, and uses Tp to represent that the forward label occurrence times are correctly predicted by the original feature drug recommendation model. And determining the number of negative-direction correct analysis results corresponding to the model training samples as negative-direction correct times, and expressing by adopting Tn to reflect that the original characteristic drug recommendation model correctly predicts the occurrence times of the negative-direction labels. And determining the number of forward errors as the number of forward error times according to the analysis result corresponding to the model training sample, and expressing by adopting Fp to reflect the occurrence times of the forward label predicted by the original characteristic drug recommendation model. And determining the number of negative errors as the number of negative errors in the analysis result corresponding to the model training sample as the number of negative errors, and expressing the number of times of occurrence of the negative label predicted by the original characteristic drug recommendation model by adopting Fn.

The current precision refers to the precision determined according to the analysis result of the original characteristic drug recommendation model, and is expressed by Pre below. The current withdrawal rate refers to a withdrawal rate determined according to the analysis result of the original characteristic drug recommendation model, and is represented by Rec below. The current accuracy refers to the accuracy determined according to the analysis result of the original characteristic drug recommendation model, and is expressed by Acc. The current F1score refers to the F1score determined from the analysis of the original characteristic drug recommendation model, denoted as F1score below. The current AUC refers to the AUC determined according to the analysis result of the original characteristic drug recommendation model, and the AUC (Area Under cut) is defined as the Area enclosed by the ROC Curve and the coordinate axis and can be used for reflecting the authenticity of the model test.

As an example, in step S403, the server calculates the statistically analyzed positive correct times Tp, negative correct times Tn, positive error times Fp, and negative error times Fn by using index calculation formulas corresponding to different evaluation indexes, and obtains the current precision Pre, the current withdrawal rate Rec, the current F1score F1score, the current accuracy Acc, and the current AUC of the original feature drug recommendation model corresponding to the training text features.

In this example, the server employs a precision calculation formula

And calculating the forward correct times Tp and the forward error times Fp to obtain the current precision Pre.

In this example, the server employs a revocation rate calculation formula

And calculating the positive correct times Tp and the negative error times Fn to obtain the current withdrawal rate Rec.

In this example, the server employs the F1score calculation formula

And calculating the current precision Pre and the current withdrawal rate Rec to obtain a current F1score F1 score.

In this example, the server employs an accuracy calculation formula

And calculating the positive correct times Tp, the negative correct times Tn, the positive error times Fp and the negative error times Fn to obtain the current accuracy Acc.

In this example, the server is based on

And (3) establishing an X-axis coordinate under the ROC curve, establishing a Y-axis coordinate under the ROC curve based on the current withdrawal rate Rec, calculating the area enclosed by the coordinate axes under the ROC curve, and acquiring the current AUC.

As an example, in step S404, after the server evaluates the index actual measurement values of the indexes, such as the current precision Pre, the current withdrawal rate Rec, the current F1score F1score, the current accuracy Acc, and the current AUC, of the original feature drug recommendation model corresponding to the training text feature, the server compares the index actual measurement value of each evaluation index with the corresponding index evaluation threshold to determine whether the corresponding evaluation index meets the standard, and obtains an index evaluation result corresponding to each evaluation index, where the index evaluation result includes two types, that is, an index meeting the standard and an index not meeting the standard. If the index evaluation results corresponding to all the evaluation indexes reach the index standard, obtaining a model test result passing the test; and if the index evaluation results corresponding to at least one evaluation index do not reach the standard, obtaining the model test result which does not pass the test.

In this example, the server compares the current precision Pre with a preset precision evaluation threshold; if the current precision Pre is greater than the precision evaluation threshold, obtaining an index evaluation result of which the precision reaches the standard; and if the current precision Pre is not greater than the precision evaluation threshold, obtaining an index evaluation result of which the precision does not reach the standard. The precision evaluation threshold is a preset threshold used for judging whether the current precision reaches the standard.

In this example, the server compares the current revocation rate Rec with a preset revocation rate evaluation threshold; if the current withdrawal rate Rec is larger than the withdrawal rate evaluation threshold, acquiring an index evaluation result of which the withdrawal rate reaches the standard; and if the current withdrawal rate Rec is not greater than the withdrawal rate evaluation threshold, acquiring an index evaluation result of which the withdrawal rate does not reach the standard. Wherein the withdrawal rate evaluation threshold is a preset threshold for evaluating whether the current withdrawal rate meets the standard.

In this example, the server compares the current F1score, F1score, to a pre-set F1score evaluation threshold; if the current F1score F1score is larger than the F1score evaluation threshold, acquiring an index evaluation result of the F1score reaching the standard; and if the current F1score F1score is not more than the F1score evaluation threshold, acquiring an index evaluation result that the F1score does not reach the standard. Wherein the F1score evaluation threshold is a preset threshold for evaluating whether the current F1score meets the standard.

In this example, the server compares the current accuracy Acc with a preset accuracy evaluation threshold; if the current accuracy Acc is greater than the accuracy evaluation threshold, acquiring an index evaluation threshold of which the accuracy reaches the standard; and if the current accuracy Acc is not greater than the accuracy evaluation threshold, acquiring an index evaluation threshold with the accuracy not reaching the standard. The accuracy evaluation threshold is a preset threshold for evaluating whether the current accuracy reaches the standard.

In this example, the server compares the current AUC with a preset AUC evaluation threshold; if the current AUC is larger than the AUC evaluation threshold, acquiring an index evaluation result of the AUC reaching the standard; and if the current AUC is not greater than the AUC evaluation threshold, acquiring an index evaluation result that the AUC does not reach the standard. The AUC evaluation threshold is a preset threshold for evaluating whether the current AUC reaches the standard.

As an example, in step S205, the model test result of the original characteristic drug recommendation model corresponding to the training text feature obtained by the server may be a test-passing model or a test-failing model, and the server determines the model test result as the original characteristic drug recommendation model that passes the test and as the target characteristic drug recommendation model corresponding to the training text feature, so as to ensure that the five evaluation indexes, that is, the current precision, the current withdrawal rate, the current F1score, the current accuracy and the current AUC, corresponding to all the target characteristic drug recommendation models all reach the standard, thereby ensuring the accuracy and the effectiveness of drug recommendation using the target characteristic drug recommendation model.

In an embodiment, as shown in fig. 5, a method for recommending a medication is provided, which is described by taking the server shown in fig. 1 as an example, and includes the following steps:

s501: acquiring the data of the disease to be processed, and determining N available medicines matched with the data of the disease to be processed according to the data of the disease to be processed.

S502: and acquiring M pieces of to-be-processed comment data corresponding to each available medicine.

S503: and performing feature extraction on the disease data to be processed and the comment data to be processed to obtain target text features corresponding to the comment data to be processed, wherein the target text features comprise Bow features, TF-IDF features, Word2Vec features and manual features.

S504: and inputting the target text characteristics corresponding to the comment data to be processed into a target characteristic medicine recommendation model corresponding to the target text characteristics for analysis processing, and acquiring a comment recommendation score corresponding to each comment data to be processed.

S505: and comprehensively processing the comment recommendation scores corresponding to the M pieces of to-be-processed comment data corresponding to each available medicine to obtain the medicine recommendation score corresponding to each available medicine.

S506: and determining the target recommended medicament according to the medicament recommendation scores corresponding to all the available medicaments.

Wherein the condition data to be processed is condition data reflecting a disease currently suffered by the patient. Available medications refer to medications that are systematically recorded and can be used to treat the current patient's disease.

As an example, in step S501, the server may receive the data of the to-be-processed medical condition input by the user, query the system database according to the data of the to-be-processed medical condition, and determine N available drugs matching the data of the to-be-processed medical condition, where the number of the available drugs is N, and N is greater than or equal to 1. For example, for the to-be-processed condition data corresponding to a heart disease, the medicines which can be used for treating the heart disease and are recorded on the system are medicine a, medicine B, medicine C, medicine D and medicine E, and then medicine a, medicine B, medicine C, medicine D and medicine E are available medicines.

The comment data to be processed is comment data which needs to be processed, and the comment data is particularly a comment on whether a certain available medicine can be used for symptomatic treatment of a disease condition corresponding to the comment data to be processed.

As an example, in step S502, the server may query the system database according to the available drugs to obtain all the to-be-processed comment data corresponding to each available drug, where each to-be-processed comment data may reflect feedback of an existing patient on whether the available drugs can symptomatically treat the disease symptoms corresponding to the to-be-processed disease symptom data, and therefore, whether the available drugs can symptomatically treat the disease corresponding to the to-be-processed disease symptom data may be comprehensively determined by analyzing all the to-be-processed comment data. In this example, the number of all the to-be-processed comment data corresponding to each available medication is M, where M ≧ 1.

The target text features are text features determined by feature extraction of the disease data to be processed and the comment data to be processed. In this example, the target text feature includes several features of the Bow feature, the TF-IDF feature, the Word2Vec feature, and the manual feature.

As an example, in step S503, after acquiring the disease state data to be processed and the comment data to be processed, the server may adopt a feature engineering technology such as Bow, TF-IDF, Word2Vec, and the like to perform feature extraction on the disease state data to be processed and the comment data to be processed, and acquire target text features such as Bow feature, TF-IDF feature, Word2Vec feature, and the like; and extracting the characteristics of the disease data to be processed and the comment data to be processed by adopting a manual label mode or a keyword matching algorithm and the like to obtain the target text characteristic of manual characteristics.

The target emotion label is a target characteristic drug recommendation model corresponding to a target text characteristic, and is an emotion label output by processing the target text characteristic corresponding to the comment data to be processed. The emotion label can be a positive label or a negative label.

As an example, in step S504, the server may input four target text features, namely, a Bow feature, a TF-IDF feature, a Word2Vec feature, and a manual feature, corresponding to the comment data to be processed into target feature drug recommendation models matched with the target text features respectively for recognition, and obtain a target emotion tag output by each target feature drug recommendation model. For example, feature extraction is carried out on a piece of comment data to be processed of an available drug A and disease data to be processed, and four target text features, namely, a Bow feature, a TF-IDF feature, a Word2Vec feature and a manual feature, are obtained; and respectively inputting the Bow feature, the TF-IDF feature, the Word2Vec feature and the manual feature into a Bow feature drug recommendation model, a TF-IDF feature drug recommendation model, a Word2Vec feature drug recommendation model and a manual feature drug recommendation model, and acquiring target emotion labels output by four target feature recommendation models corresponding to each piece of comment data to be processed, wherein the target emotion labels can be represented by 1 or 0, wherein 1 is a positive label, and 0 is a negative label. Then, the server can comprehensively process the four target emotion tags of the same to-be-processed comment data by adopting a pre-configured comment scoring rule to obtain a comment recommendation score corresponding to each to-be-processed comment data, so that the comprehensiveness of the comment recommendation scores is favorably ensured, and the accuracy and the effectiveness for evaluating whether the available medicines can treat the current patient suffered from the disease are improved.

The drug recommendation score is the data of the disease to be processed by the pointer, and finally the score for recommending a certain available drug is determined.

As an example, in step S505, the server may perform comprehensive processing on the review recommendation scores of the M to-be-processed review data of the same available medicine, and obtain a medicine recommendation score corresponding to each available medicine. For example, the server may perform a mean processing on the M review recommendation scores to determine an average of the M review recommendation scores as a drug recommendation score corresponding to each available drug, which helps to ensure the accuracy and effectiveness of the available drugs in evaluating whether the available drugs can treat the current patient's disease.

As an example, in step S506, the server may sort the drug recommendation scores corresponding to all the available drugs according to the obtained drug recommendation scores corresponding to all the available drugs, select the first few available drugs with higher drug recommendation scores, and determine the first few available drugs as the target recommended drugs, so that the user selects corresponding target recommended drugs according to actual conditions.

In the medicine recommendation method provided by the embodiment, each piece of comment data to be processed and the piece of disease data to be processed are subjected to feature extraction, four target text features, namely, a Bow feature, a TF-IDF feature, a Word2Vec feature and a manual feature, are determined, then the four target text features are input into a target feature medicine recommendation model for analysis and processing, and a comment recommendation score corresponding to each piece of comment data to be processed is determined, so that the comprehensiveness of the comment recommendation score is favorably ensured, and the accuracy and the effectiveness for evaluating whether an available medicine can treat a disease of a current patient are improved. And then, the comment recommendation scores corresponding to the M pieces of to-be-processed comment data corresponding to each available medicine are comprehensively processed, and the medicine recommendation score corresponding to each available medicine is obtained, so that the comment recommendation scores corresponding to all the to-be-processed comment data are comprehensively considered by the medicine recommendation score, and the accuracy and the effectiveness of the medicine recommendation score for evaluating whether the available medicine can treat the current patient suffering from the disease are guaranteed. And finally, determining the target recommended medicament based on the medicament recommendation scores corresponding to all available medicaments so as to ensure the accuracy and effectiveness of the target recommended medicament for treating diseases corresponding to the to-be-treated disease data and assist doctors in completing a prescription scheme.

In an embodiment, as shown in fig. 6, in step S504, that is, inputting a target text feature corresponding to comment data to be processed into a target feature drug recommendation model corresponding to the target text feature for analysis processing, and acquiring a comment recommendation score corresponding to each comment data to be processed includes:

s601: and inputting the target text characteristics corresponding to the comment data to be processed into the target characteristic medicine recommendation models corresponding to the target text characteristics for analysis processing, and acquiring the target emotion labels output by each target characteristic medicine recommendation model.

S602: and obtaining a comment recommendation score corresponding to each comment data to be processed according to the target emotion labels output by all the target characteristic medicine recommendation models.

As an example, in step S601, the server may input four target text features, namely, a Bow feature, a TF-IDF feature, a Word2Vec feature, and a manual feature, corresponding to the comment data to be processed into target feature drug recommendation models matched with the target text features respectively for recognition, and obtain a target emotion tag output by each target feature drug recommendation model. For example, feature extraction is carried out on a piece of comment data to be processed of an available drug A and disease data to be processed, and four target text features, namely, a Bow feature, a TF-IDF feature, a Word2Vec feature and a manual feature, are obtained; and respectively inputting the Bow feature, the TF-IDF feature, the Word2Vec feature and the manual feature into a Bow feature drug recommendation model, a TF-IDF feature drug recommendation model, a Word2Vec feature drug recommendation model and a manual feature drug recommendation model, and acquiring target emotion labels output by four target feature recommendation models corresponding to each piece of comment data to be processed, wherein the target emotion labels can be represented by 1 or 0, wherein 1 is a positive label, and 0 is a negative label.

As an example, in step S602, four target text features, namely, a Bow feature, a TF-IDF feature, a Word2Vec feature and a manual feature, extracted from each piece of comment data to be processed are respectively input to the target text features for analysis processing, and target emotion tags output by four target feature drug recommendation models are obtained. Then, the server can perform comprehensive processing on the target emotion tags output by the target characteristic drug recommendation model to determine the comment recommendation score corresponding to each piece of comment data to be processed. For example, the server may perform numerical conversion on the target emotion labels output by the four target characteristic drug recommendation models to 1/0 emotion label values in this form, or other emotion label values; and then, carrying out weighting processing on the emotion label numerical values corresponding to the four target emotion labels to obtain comment recommendation scores corresponding to the comment data to be processed.

In the example, four target text features, namely, the Bow feature, the TF-IDF feature, the Word2Vec feature and the manual feature, extracted from each piece of comment data to be processed are respectively input into a corresponding target feature drug recommendation model for analysis processing, four target emotion tags are obtained, then the four target emotion tags are comprehensively processed, so that a comment recommendation score is obtained, the comment recommendation score is associated with the four Bow features, the TF-IDF feature, the Word2Vec feature and the manual feature, the comprehensiveness of the comment recommendation score is favorably guaranteed, and the accuracy and the effectiveness for evaluating whether available drugs can treat the current patient suffering from diseases are improved.

In an embodiment, as shown in fig. 7, step S505 is to perform comprehensive processing on the review recommendation scores corresponding to the M to-be-processed review data corresponding to each available medication to obtain a medication recommendation score corresponding to each available medication, where the method includes:

s701: and counting the number of the to-be-processed comment data corresponding to each available medicine, and determining the number of the single-medicine comments corresponding to each available medicine.

S702: and acquiring a comment quantity normalization value corresponding to each available medicine according to the comment quantity of the single medicine corresponding to the N available medicines.

S703: and processing the comment recommendation scores corresponding to the M pieces of comment data to be processed corresponding to each available medicine to obtain a single medicine score corresponding to each available medicine.

S704: and taking the drug recommendation score corresponding to each available drug according to the comment quantity normalized value and the single drug score corresponding to each available drug.

As an example, in step S701, the server needs to count the number M of the to-be-processed review data corresponding to each available medication, and determine the number of the single medication reviews corresponding to each available medication. In this example, the number of single drug reviews corresponding to the ith available drug is defined as Mi.

As an example, in step S702, the server needs to use a normalization processing formula to normalize the number Mi of the single drug reviews corresponding to the N available drugs, and obtain a normalized value of the number of reviews corresponding to each available drug. The normalization processing formula is

Ms is a normalized value of the number of reviews and K is a normalized range, which may be 1, 10, 100, or other threshold. Understandably, the normalization processing is carried out on the single-medicine comment quantity Mi corresponding to the N available medicines, so that the finally obtained comment quantity normalization value is favorably ensured, the normalization processing is related to the single-medicine comment quantity of all the available medicines, the comprehensive processing is favorably carried out on the medicine recommendation score from the comment quantity dimension of the comment data to be processed, the comprehensiveness of the finally obtained medicine recommendation score is further ensured, and the accuracy and the effectiveness for evaluating whether the available medicines can treat the current patient suffered from the disease are improved.

As an example, in step S703, the server may process the review recommendation scores corresponding to the M to-be-processed review data corresponding to each available medication, specifically, may perform mean processing on the review recommendation scores corresponding to the M to-be-processed review data, and obtain a single medication score corresponding to each available medication. The single drug score reflects the overall evaluation of the available drug by all of the review data to be processed.

As an example, in step S704, the server may determine, as the drug recommendation score corresponding to each available drug, a product of the normalized value of the number of reviews corresponding to each available drug and the single drug score, so that the drug recommendation score comprehensively considers the overall evaluation of all the review data to be processed on the available drug, and comprehensively processes the drug recommendation score from the review number dimension, thereby ensuring the comprehensiveness of the finally obtained drug recommendation score, and improving the accuracy and effectiveness for evaluating whether the available drug can treat the disease suffered by the current patient.

It should be understood that, the sequence numbers of the steps in the foregoing embodiments do not imply an execution sequence, and the execution sequence of each process should be determined by its function and inherent logic, and should not constitute any limitation to the implementation process of the embodiments of the present invention.

In an embodiment, a drug recommendation model training device is provided, and the drug recommendation model training device corresponds to the drug recommendation model training method in the embodiment one to one. As shown in fig. 8, the medicine recommendation model training apparatus includes a model training data obtaining module 801, an emotion label labeling obtaining module 802, a training text feature obtaining module 803, a model training sample obtaining module 804, an original recommendation model obtaining module 805, and a target recommendation model obtaining module 806. The functional modules are explained in detail as follows:

a model training data obtaining module 801, configured to obtain model training data, where the model training data includes a training drug, training condition data corresponding to the training drug, and training comment data.

And a labeled emotion label acquisition module 802, configured to perform emotion analysis on the training comment data in the model training data, and acquire a labeled emotion label corresponding to the model training data.

And the training text feature acquisition module 803 is used for performing feature extraction on the training condition data and the training comment data to acquire training text features corresponding to the model training data, wherein the training text features comprise a Bow feature, a TF-IDF feature, a Word2Vec feature and a manual feature.

A model training sample obtaining module 804, configured to obtain a model training sample based on each training text feature and the labeled emotion label, store the model training sample in a feature sample set corresponding to the training text feature, and divide the model training sample in the feature sample set into a feature training set and a feature test set.

And an original recommendation model obtaining module 805, configured to input the model training samples in the feature training set into a feature classifier corresponding to the features of the training text to perform model training, so as to obtain an original feature drug recommendation model corresponding to the features of the training text.

And a target recommendation model obtaining module 806, configured to use the model training samples in the feature test set to test an original feature drug recommendation model corresponding to the training text features, so as to obtain a target feature drug recommendation model corresponding to the training text features.

In one embodiment, the drug recommendation model training device further comprises:

and the sample number counting unit is used for counting the positive sample number, the negative sample number and the total sample number in the feature training set.

And the positive and negative sample difference value acquisition unit is used for acquiring a positive and negative sample difference value according to the positive sample number and the negative sample number.

And the sample deviation ratio acquisition unit is used for acquiring a sample deviation ratio according to the positive and negative sample difference value and the total number of samples.

And the oversampling processing unit is used for performing oversampling processing on the model training samples in the characteristic training set by adopting an oversampling technology if the sample deviation proportion is greater than the deviation proportion threshold value, and acquiring the updated model training samples.

In an embodiment, the target recommendation model obtaining module 806 includes:

and the predicted emotion label acquisition unit is used for inputting the model training samples in the feature test set into the original feature drug recommendation model corresponding to the training text features for identification, and acquiring the predicted emotion labels corresponding to the model training samples.

And the frequency counting unit is used for counting the positive correct frequency, the negative correct frequency, the positive error frequency and the negative error frequency according to the marked emotion label and the predicted emotion label corresponding to the model training sample.

And the evaluation index acquisition unit is used for acquiring the current precision, the current withdrawal rate, the current F1score, the current accuracy and the current AUC of the original characteristic drug recommendation model corresponding to the training text characteristics according to the positive correct times, the negative correct times, the positive error times and the negative error times.

And the model test result acquisition unit is used for acquiring a model test result of the original characteristic drug recommendation model corresponding to the training text characteristic according to the current precision, the current withdrawal rate, the current F1score, the current accuracy and the current AUC.

And the target model determining unit is used for determining the original characteristic drug recommendation model corresponding to the training text characteristics as the target characteristic drug recommendation model corresponding to the training text characteristics if the model test result is that the test is passed.

For specific limitations of the drug recommendation model training device, reference may be made to the above limitations of the drug recommendation model training method, which are not described herein again. The modules in the drug recommendation model training device can be wholly or partially realized by software, hardware and a combination thereof. The modules can be embedded in a hardware form or independent from a processor in the computer device, and can also be stored in a memory in the computer device in a software form, so that the processor can call and execute operations corresponding to the modules.

In an embodiment, a medicine recommendation device is provided, and the medicine recommendation device corresponds to the medicine recommendation methods in the embodiments one to one. As shown in fig. 9, the medication recommendation apparatus includes an available medication determination module 901, a pending review data acquisition module 902, a target text feature acquisition module 903, a review recommendation score acquisition module 904, a medication recommendation score acquisition module 905, and a target recommended medication determination module 906. The functional modules are explained in detail as follows:

an available drug determining module 901, configured to acquire the data of the condition to be treated, and determine, according to the data of the condition to be treated, N available drugs that match the data of the condition to be treated.

A to-be-processed comment data obtaining module 902, configured to obtain M pieces of to-be-processed comment data corresponding to each available medicine.

And the target text feature acquisition module 903 is used for performing feature extraction on the disease data to be processed and the comment data to be processed to acquire target text features corresponding to the comment data to be processed, wherein the target text features comprise a Bow feature, a TF-IDF feature, a Word2Vec feature and a manual feature.

And the comment recommendation score obtaining module 904 is configured to input the target text features corresponding to the to-be-processed comment data into the target feature medicine recommendation model corresponding to the target text features for analysis processing, and obtain a comment recommendation score corresponding to each to-be-processed comment data.

The medicine recommendation score obtaining module 905 is configured to perform comprehensive processing on the comment recommendation scores corresponding to the M to-be-processed comment data corresponding to each available medicine, and obtain a medicine recommendation score corresponding to each available medicine.

The target recommended medication determining module 906 determines the target recommended medication according to the medication recommendation scores corresponding to all available medications.

In an embodiment, the comment recommendation score obtaining module 904 includes:

and the target emotion tag acquisition unit is used for inputting the target text features corresponding to the comment data to be processed into the target feature drug recommendation models corresponding to the target text features for analysis and processing, and acquiring the target emotion tags output by each target feature drug recommendation model.

And the comment recommendation score acquisition unit is used for acquiring a comment recommendation score corresponding to each piece of comment data to be processed according to the target emotion tags output by all the target characteristic medicine recommendation models.

In one embodiment, the drug recommendation score obtaining module 905 includes:

and the single medicine comment quantity determining unit is used for counting the quantity of the to-be-processed comment data corresponding to each available medicine and determining the quantity of the single medicine comments corresponding to each available medicine.

And the comment quantity normalized value acquisition unit is used for acquiring a comment quantity normalized value corresponding to each available medicine according to the single medicine comment quantity corresponding to the N available medicines.

And the single medicine score obtaining unit is used for processing the comment recommendation scores corresponding to the M pieces of comment data to be processed corresponding to each available medicine and obtaining the single medicine score corresponding to each available medicine.

And the medicine recommendation score obtaining unit is used for obtaining the medicine recommendation score corresponding to each available medicine according to the comment quantity normalization value corresponding to each available medicine and the single medicine score.

For the specific definition of the medication recommendation device, reference may be made to the above definition of the medication recommendation method, which is not described herein again. The modules in the medicine recommending device can be wholly or partially realized by software, hardware and a combination thereof. The modules can be embedded in a hardware form or independent from a processor in the computer device, and can also be stored in a memory in the computer device in a software form, so that the processor can call and execute operations corresponding to the modules.

In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as shown in fig. 10. The computer device includes a processor, a memory, a network interface, and a database connected by a system bus. Wherein the processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device comprises a nonvolatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of an operating system and computer programs in the non-volatile storage medium. The database of the computer device is used for executing data adopted or generated in a drug recommendation model training method or a drug recommendation method. The network interface of the computer device is used for communicating with an external terminal through a network connection. The computer program is executed by a processor to implement a medication recommendation model training method or a medication recommendation method.

In an embodiment, a computer device is provided, which includes a memory, a processor, and a computer program stored on the memory and executable on the processor, and when the processor executes the computer program, the method for training a drug recommendation model in the foregoing embodiments is implemented, for example, S201-S206 shown in fig. 2, or shown in fig. 3 to 4, which is not described herein again to avoid repetition. Alternatively, when the processor executes the computer program, the functions of each module/unit in the embodiment of the drug recommendation model training apparatus are implemented, for example, the functions of the model training data obtaining module 801, the emotion label labeling obtaining module 802, the training text feature obtaining module 803, the model training sample obtaining module 804, the original recommendation model obtaining module 805, and the target recommendation model obtaining module 806 shown in fig. 8, and are not described herein again to avoid repetition.

In an embodiment, a computer-readable storage medium is provided, where a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, the method for training a drug recommendation model in the foregoing embodiments is implemented, for example, S201-S206 shown in fig. 2, or shown in fig. 3 to 4, which is not described herein again to avoid repetition. Alternatively, when being executed by a processor, the computer program implements functions of each module/unit in the above-mentioned drug recommendation model training apparatus, for example, functions of the model training data obtaining module 801, the emotion label labeling obtaining module 802, the training text feature obtaining module 803, the model training sample obtaining module 804, the original recommendation model obtaining module 805, and the target recommendation model obtaining module 806 shown in fig. 8, and in order to avoid repetition, details are not repeated here.

In one embodiment, a computer device is provided, which includes a memory, a processor, and a computer program stored on the memory and executable on the processor, and when the processor executes the computer program, the method for recommending a drug in the foregoing embodiments is implemented, for example, as shown in S501-S506 in fig. 5, or as shown in fig. 6 to 7, which is not described herein again to avoid repetition. Alternatively, when the processor executes the computer program, the functions of each module/unit in the embodiment of the medication recommendation device are implemented, for example, the functions of the available medication determination module 901, the to-be-processed comment data acquisition module 902, the target text feature acquisition module 903, the comment recommendation score acquisition module 904, the medication recommendation score acquisition module 905, and the target recommended medication determination module 906 shown in fig. 9 are not described herein again to avoid repetition.

In an embodiment, a computer-readable storage medium is provided, and a computer program is stored on the computer-readable storage medium, and when being executed by a processor, the computer program implements the method for recommending a drug in the above embodiments, for example, S501-S506 shown in fig. 5, or S6-7, which is not repeated here to avoid repetition. Alternatively, when being executed by a processor, the computer program implements the functions of the modules/units in the embodiment of the medication recommendation device, for example, the functions of the available medication determination module 901, the to-be-processed comment data acquisition module 902, the target text feature acquisition module 903, the medication recommendation score acquisition module 905, and the target recommended medication determination module 906 shown in fig. 9, and in order to avoid repetition, details are not repeated here.

It will be understood by those skilled in the art that all or part of the processes of the methods of the embodiments described above can be implemented by hardware instructions of a computer program, which can be stored in a non-volatile computer-readable storage medium, and when executed, can include the processes of the embodiments of the methods described above. Any reference to memory, storage, database, or other medium used in the embodiments provided herein may include non-volatile and/or volatile memory, among others. Non-volatile memory can include read-only memory (ROM), Programmable ROM (PROM), Electrically Programmable ROM (EPROM), Electrically Erasable Programmable ROM (EEPROM), or flash memory. Volatile memory can include Random Access Memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms such as Static RAM (SRAM), Dynamic RAM (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDRSDRAM), Enhanced SDRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), Rambus Direct RAM (RDRAM), direct bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

It will be apparent to those skilled in the art that, for convenience and brevity of description, only the above-mentioned division of the functional units and modules is illustrated, and in practical applications, the above-mentioned function distribution may be performed by different functional units and modules according to needs, that is, the internal structure of the apparatus is divided into different functional units or modules to perform all or part of the above-mentioned functions.

The above-mentioned embodiments are only used for illustrating the technical solutions 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 technical features may be equivalently replaced; such modifications and substitutions do not substantially depart from the spirit and scope of the embodiments of the present invention, and are intended to be included within the scope of the present invention.

Claims

1. A drug recommendation model training method is characterized by comprising the following steps:

2. The method for training the drug recommendation model according to claim 1, wherein before the model training samples in the feature training set are input into the feature classifier corresponding to the training text features for model training and the original feature drug recommendation model corresponding to the training text features is obtained, the method for training the drug recommendation model further comprises:

counting the number of positive samples, the number of negative samples and the total number of samples in the feature training set;

acquiring a positive and negative sample difference value according to the positive sample number and the negative sample number;

acquiring a sample deviation proportion according to the positive and negative sample difference value and the total sample number;

and if the sample deviation proportion is larger than the deviation proportion threshold value, adopting an oversampling technology to perform oversampling processing on the model training samples in the characteristic training set to obtain updated model training samples.

3. The method for training the drug recommendation model according to claim 1, wherein the step of testing the original feature drug recommendation model corresponding to the training text features by using the model training samples in the feature test set to obtain the target feature drug recommendation model corresponding to the training text features comprises:

inputting model training samples in the feature test set into an original feature drug recommendation model corresponding to the training text features for recognition, and obtaining predicted emotion labels corresponding to the model training samples;

counting positive correct times, negative correct times, positive error times and negative error times according to the marked emotion label and the predicted emotion label corresponding to the model training sample;

according to the positive correct times, the negative correct times, the positive error times and the negative error times, obtaining the current precision, the current withdrawal rate, the current F1score, the current accuracy and the current AUC of the original characteristic drug recommendation model corresponding to the training text characteristics;

obtaining a model test result of an original characteristic drug recommendation model corresponding to the training text characteristic according to the current precision, the current withdrawal rate, the current F1score, the current accuracy rate and the current AUC;

and if the model test result is that the test is passed, determining the original characteristic drug recommendation model corresponding to the training text characteristic as the target characteristic drug recommendation model corresponding to the training text characteristic.

4. A method for medication recommendation, comprising:

5. The drug recommendation method of claim 4, wherein the step of inputting the target text features corresponding to the to-be-processed comment data into the target feature drug recommendation model corresponding to the target text features for analysis processing to obtain the comment recommendation score corresponding to each of the to-be-processed comment data comprises:

inputting target text characteristics corresponding to the comment data to be processed into target characteristic medicine recommendation models corresponding to the target text characteristics for analysis processing, and acquiring target emotion labels output by each target characteristic medicine recommendation model;

and obtaining a comment recommendation score corresponding to each comment data to be processed according to the target emotion labels output by all the target characteristic medicine recommendation models.

6. The medication recommendation method of claim 4, wherein the step of performing comprehensive processing on the comment recommendation scores corresponding to the M pieces of to-be-processed comment data corresponding to each of the available medications to obtain the medication recommendation score corresponding to each of the available medications comprises:

counting the number of the comment data to be processed corresponding to each available medicine, and determining the number of the comments of the single medicine corresponding to each available medicine;

obtaining a comment quantity normalization value corresponding to each available medicine according to the comment quantity of the single medicine corresponding to the N available medicines;

processing the comment recommendation scores corresponding to the M pieces of comment data to be processed corresponding to each available medicine to obtain a single medicine score corresponding to each available medicine;

and taking the drug recommendation score corresponding to each available drug according to the comment quantity normalized value and the single drug score corresponding to each available drug.

7. A medication recommendation model training device, comprising:

8. A medication recommendation device, comprising:

9. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the medication recommendation model training method according to any one of claims 1 to 3 when executing the computer program or implements the medication recommendation method according to any one of claims 4 to 6 when executing the computer program.

10. A computer-readable storage medium, in which a computer program is stored, which, when being executed by a processor, carries out a medication recommendation model training method according to any one of claims 1 to 3, or which, when being executed by a processor, carries out a medication recommendation method according to any one of claims 4 to 6.