An adverse drug reaction prediction system using an electronic medical record of skin

Abstract

The present application relates to the field of medical information technology, in particular to a kind of adverse drug reaction prediction system using skin electronic medical record, the system obtains the skin electronic medical record of a number of historical patients, by analyzing skin electronic medical record, determine the basic correlation degree of each feature and its effective coefficient, further determine the correlation degree of each feature, based on the correlation degree, determine the reference weight of a number of decision trees trained using the skin electronic medical record of historical patients;The skin electronic medical record of the patient to be predicted is input into a number of decision trees, and the output result of the decision tree is weighted using the reference weight, so as to obtain the prediction result of the adverse drug reaction of the patient to be predicted.The present application determines the reference weight of the decision tree by self-adaptation, improves the prediction accuracy of the adverse drug reaction of the patient to be predicted.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical information technology, and in particular to an adverse drug reaction prediction system using skin electronic medical records. BACKGROUND

[0002] Adverse drug reaction, also known as adverse drug reaction (ADR), is an adverse reaction that occurs during treatment with a drug in the prescribed dose, which is unrelated to the treatment purpose and has a causal relationship with the use of the drug. Common adverse drug reactions in skin diseases include drug eruption, photosensitivity, pigmentation, and hair loss. During drug treatment for skin disease patients, due to individual factors (such as allergy, genetics, environment), drug interactions (such as antibiotics combined with antifungal drugs, which may enhance skin allergic reactions or photosensitivity), and the stimulating nature of the drug itself, patients may experience adverse drug reactions. When a doctor develops a drug treatment plan for a skin disease patient, predicting adverse drug reactions in skin disease patients can help remind the doctor to pay attention to the possibility of adverse reactions caused by the drug, so as to assist the doctor in adjusting the drug treatment plan or informing the patient of the adverse drug reactions.

[0003] In the prior art, when predicting adverse drug reactions in skin disease patients, a classification model can be constructed based on machine learning methods such as random forests and support vector machines, and the contents of the patient's skin electronic medical record can be used to train the classification model, and the trained classification model can be used to predict the probability of adverse reactions in patients using a certain drug. When using random forests to construct classification models for adverse drug reaction prediction, the output results of all decision trees in the random forest are averaged to obtain the final prediction result. However, since each decision tree in the random forest uses a random subset of all patient features, and each feature has a different degree of association with adverse reactions, the output results of different decision trees have different degrees of influence on the prediction result, which results in a large prediction error when the output results of all decision trees in the random forest are directly averaged to obtain the final prediction result, ultimately affecting the accuracy of adverse drug reaction prediction for skin disease patients. SUMMARY

[0004] The purpose of the present application is to provide an adverse drug reaction prediction system using skin electronic medical records, which solves the problem of poor accuracy of adverse drug reaction prediction for skin disease patients.

[0005] To solve the above technical problems, in a first aspect, the present application provides an adverse drug reaction prediction system using skin electronic medical records, which comprises:

[0006] a data acquisition module configured to acquire electronic skin records of a plurality of historical patients, wherein the electronic skin records include diagnosis information, and some of the electronic skin records further include re-visit information, and wherein the diagnosis information and the re-visit information include occurrence of each feature and occurrence of adverse drug reactions of the patients, and wherein the features include patient symptoms and drug components used in treatment plans;

[0007] a basic correlation degree acquisition module configured to count occurrence of each feature and occurrence of adverse drug reactions of the patients in the diagnosis information of all the historical patients, and determine a basic correlation degree of each feature by combining differences in occurrence of each feature and differences in occurrence of adverse drug reactions of the patients in the diagnosis information and the re-visit information of the same historical patient;

[0008] an effective coefficient acquisition module configured to train a plurality of decision trees by using the electronic skin records of the historical patients, determine an effective coefficient of the basic correlation degree of each feature according to distribution of historical patients who have adverse drug reactions before and after a target feature divides a parent node in the decision trees, and a number of layers of the parent node divided by each feature in the decision trees;

[0009] a correlation degree acquisition module configured to determine a correlation degree of each feature according to the basic correlation degree and the effective coefficient of each feature;

[0010] a weight acquisition module configured to determine a reference weight of the decision trees according to the correlation degrees of all the features included in the decision trees;

[0011] a prediction module configured to input an electronic skin record of a patient to be predicted into the plurality of decision trees, and weight output results of the decision trees by using the reference weight, so as to obtain a prediction result of adverse drug reactions of the patient to be predicted.

[0012] In some possible implementation manners, in combination with the first aspect, the basic correlation degree acquisition module includes:

[0013] a frequency difference acquisition unit configured to count occurrence of a target feature and occurrence of adverse drug reactions of the patients in the diagnosis information of all the historical patients, and determine a frequency difference of the target feature, wherein the frequency difference reflects a frequency difference of occurrence of adverse drug reactions of the patients when the target feature occurs or does not occur;

[0014] a change coefficient acquisition unit configured to determine a change coefficient of a target feature according to a total number of historical patients whose diagnosis information and re-visit information in the electronic skin records have differences in occurrence of the target feature and differences in occurrence of adverse drug reactions of the patients;

[0015] The basic correlation degree acquisition unit is configured to determine the basic correlation degree of the target item feature according to a difference size between a frequency difference of the target item feature and frequency differences of other item features and a variation coefficient of the target item feature.

[0016] In some possible implementation manners, the frequency difference acquisition unit comprises:

[0017] The first frequency determination unit is configured to determine a ratio of a number of all historical patients in the diagnostic information who have the target item feature and who have an adverse drug reaction to a number of all historical patients in the diagnostic information who have the target item feature, to obtain a first frequency.

[0018] The second frequency determination unit is configured to determine a ratio of a number of all historical patients in the diagnostic information who do not have the target item feature and who have an adverse drug reaction to a number of all historical patients in the diagnostic information who do not have the target item feature, to obtain a second frequency.

[0019] The frequency difference determination unit is configured to determine an absolute value of a difference between the first frequency and the second frequency, to obtain the frequency difference of the target item feature.

[0020] In some possible implementation manners, the basic correlation degree acquisition unit comprises:

[0021] The average frequency difference determination unit is configured to determine an average frequency difference according to an overall distribution level of the frequency differences of all item features.

[0022] The frequency difference relative value determination unit is configured to determine a ratio of the frequency difference of the target item feature to the average frequency difference, to obtain a frequency difference relative value.

[0023] The frequency difference relative value increment determination unit is configured to determine a product of the frequency difference relative value and the variation coefficient of the target item feature, to obtain a frequency difference relative value increment.

[0024] The basic correlation degree determination unit is configured to normalize an addition value of the frequency difference relative value and the frequency difference relative value increment, to obtain the basic correlation degree of the target item feature.

[0025] In some possible implementation manners, the effective coefficient acquisition module comprises:

[0026] The partition evaluation obtaining unit is configured to determine a partition evaluation of the target item feature on the decision tree according to a difference between a proportion of historical patients with adverse drug reactions in each pair of sibling nodes obtained after the target item feature partitions a parent node in the decision tree and a proportion of historical patients with adverse drug reactions in the parent node, and a difference between proportions of historical patients with adverse drug reactions in each sibling node in each pair of sibling nodes.

[0027] The partition efficiency obtaining unit is configured to determine a partition efficiency of the target item feature according to the partition evaluation of the target item feature on the decision tree and a number of layers of the parent node partitioned by the target item feature in the decision tree.

[0028] The effective coefficient obtaining unit is configured to determine an effective coefficient of the basic correlation degree of the target item feature according to a difference between the partition efficiency of the target item feature and partition efficiencies of other item features.

[0029] In some possible implementation manners, in combination with the first aspect, the partition evaluation obtaining unit includes:

[0030] The first proportion difference determining unit is configured to determine a difference between a maximum value of the proportion of historical patients with adverse drug reactions in each pair of sibling nodes obtained after the target item feature partitions a parent node in the decision tree and the proportion of historical patients with adverse drug reactions in the parent node, to obtain a first proportion difference.

[0031] The second proportion difference determining unit is configured to determine a difference between the proportions of historical patients with adverse drug reactions in each pair of sibling nodes obtained after the target item feature partitions a parent node in the decision tree, to obtain a second proportion difference.

[0032] The partition evaluation determining unit is configured to determine an overall distribution level of product values of the first proportion difference and the second proportion difference corresponding to each pair of sibling nodes obtained after the target item feature partitions the parent node in the decision tree, to determine the partition evaluation of the target item feature on the decision tree.

[0033] In some possible implementation manners, in combination with the first aspect, the partition efficiency obtaining unit includes:

[0034] The first partition efficiency determining unit is configured to determine a product of the partition evaluation of the target item feature on the decision tree and a number of layers of the parent node partitioned by the target item feature in the decision tree, to obtain the partition efficiency of the target item feature on a single decision tree.

[0035] The second partition efficiency determining unit is configured to determine an overall distribution level of the partition efficiency of the target item feature on all decision trees, to determine the partition efficiency of the target item feature.

[0036] With the first aspect above, in some possible implementation manners, the effective coefficient obtaining unit comprises:

[0037] an average partition efficiency determining unit configured to determine an average partition efficiency according to an overall distribution level of the partition efficiencies of all item features;

[0038] a partition efficiency difference determining unit configured to determine a multiplication value of the average partition efficiency and a set effective threshold, and then determine a difference between the partition efficiency of the target item feature and the multiplication value, to obtain a partition efficiency difference;

[0039] an effective coefficient determining unit configured to input the partition efficiency difference into a ReLU function, and output an effective coefficient of the basic correlation degree of the target item feature from the ReLU function.

[0040] With the first aspect above, in some possible implementation manners, the correlation degree obtaining module comprises:

[0041] a correlation degree growth value obtaining unit configured to determine a product of the basic correlation degree of each item feature and the effective coefficient thereof, to obtain a correlation degree growth value;

[0042] a correlation degree obtaining unit configured to normalize an added value of the basic correlation degree of each item feature and the correlation degree growth value, to obtain the correlation degree of each item feature.

[0043] With the first aspect above, in some possible implementation manners, the weight obtaining module comprises:

[0044] an average correlation degree obtaining unit configured to determine an average correlation degree corresponding to the decision tree according to an overall distribution level of the correlation degrees of all item features contained in the decision tree;

[0045] a reference weight obtaining unit configured to normalize the average correlation degree corresponding to the decision tree, to obtain a reference weight of the decision tree, and an added value of reference weights of all the decision trees is equal to a numerical value 1.

[0046] To solve the above technical problem, the second aspect of the present application further provides an adverse drug reaction prediction method using skin electronic medical records, which comprises:

[0047] obtaining skin electronic medical records of a plurality of historical patients, wherein the skin electronic medical records comprise diagnosis information, and part of the skin electronic medical records further comprise re-visit information, the diagnosis information and the re-visit information comprise occurrence of each item feature and occurrence of adverse drug reactions of patients, and the each item feature comprises patient symptoms and drug components used in a treatment scheme;

[0048] count the occurrence of each feature in the diagnosis information of all historical patients and the occurrence of adverse drug reactions of patients, and determine the basic correlation degree of each feature according to the difference between the occurrence of each feature in the diagnosis information and the reexamination information of the same historical patient and the difference between the occurrence of adverse drug reactions of patients;

[0049] train a plurality of decision trees by using the electronic medical records of the historical patients, determine the effective coefficient of the basic correlation degree of each feature according to the distribution of historical patients who have adverse drug reactions before and after the target feature divides the parent node in the decision tree and the number of layers of the parent node divided by each feature in the decision tree;

[0050] determine the correlation degree of each feature according to the basic correlation degree and the effective coefficient of each feature;

[0051] determine the reference weight of the decision tree according to the correlation degree of all features contained in the decision tree;

[0052] input the electronic medical record of the patient to be predicted into the plurality of decision trees, and weight the output results of the decision trees by using the reference weight, so as to obtain the prediction result of the adverse drug reactions of the patient to be predicted.

[0053] To solve the above technical problems, in a third aspect, the present application also provides an adverse drug reaction prediction device using electronic medical records of skin, comprising a memory and a processor. The memory is used to store executable program code, and the processor is used to call and run the executable program code from the memory, so that the device executes the module implementation steps in the first aspect or any one of the possible implementation manners of the first aspect.

[0054] To solve the above technical problems, in a fourth aspect, the present application also provides a computer program product, which comprises computer program code. When the computer program code runs on a computer, it makes the computer execute the module implementation steps in the first aspect or any one of the possible implementation manners of the first aspect.

[0055] To solve the above technical problems, in a fifth aspect, the present application also provides a computer readable storage medium, which stores computer program code. When the computer program code runs on a computer, it makes the computer execute the module implementation steps in the first aspect or any one of the possible implementation manners of the first aspect.

[0056] The present application has the following beneficial effects: by obtaining the electronic medical records of the skin of a plurality of historical patients, the electronic medical records of the skin including diagnosis information, part of the electronic medical records of the skin also including re-consultation information, the diagnosis information and the re-consultation information including the occurrence of each feature and the occurrence of adverse drug reactions of patients, each feature including patient symptoms and drug components used in the treatment scheme; by counting the occurrence of each feature and the occurrence of adverse drug reactions of patients in the diagnosis information of all historical patients, and combining the differences in the occurrence of each feature and the differences in the occurrence of adverse drug reactions of patients in the diagnosis information and the re-consultation information of the same historical patient, the basic correlation degree of each feature is determined; at the same time, the electronic medical records of the skin of historical patients are used to train a plurality of decision trees, and based on the distribution of historical patients who have adverse drug reactions before and after the target feature divides the parent node in the decision tree, and the number of layers of the parent node divided by each feature in the decision tree, the effective coefficient of the basic correlation degree of each feature is determined; based on the basic correlation degree and the effective coefficient of each feature, the correlation degree of each feature is determined, and based on the correlation degree, the reference weight of the plurality of decision trees trained by the electronic medical records of the skin of historical patients is determined; the electronic medical record of the skin of the patient to be predicted is input into the plurality of decision trees, and the output results of the decision trees are weighted by using the reference weight, so as to obtain the prediction result of the adverse drug reactions of the patient to be predicted. The present application adaptively determines the reference weight of the decision tree, and effectively improves the prediction accuracy of the adverse drug reactions of the patient to be predicted. BRIEF DESCRIPTION OF DRAWINGS

[0057] In order to more clearly illustrate the technical solutions and advantages of the embodiments of the present application or the prior art, the drawings needed in the following embodiment or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0058] Figure 1 A structure schematic diagram of an adverse drug reaction prediction system using electronic medical records of the skin according to an embodiment of the present application;

[0059] Figure 2 A step flowchart of an adverse drug reaction prediction method using electronic medical records of the skin according to an embodiment of the present application;

[0060] Figure 3 A structure schematic diagram of a decision tree according to an embodiment of the present application;

[0061] Figure 4 A structure schematic diagram of an adverse drug reaction prediction device using electronic medical records of the skin according to an embodiment of the present application. DETAILED DESCRIPTION

[0062] To clearly illustrate the technical features of the present application, the following will describe the present application in detail with specific embodiments and in conjunction with the drawings.

[0063] Embodiments of the present application will now be described in detail with reference to the accompanying drawings. Although a certain embodiment of the present application is shown in the drawings, it is understood that the present application can be carried out in various forms and should not be interpreted as being limited to the embodiments set forth herein; instead, these embodiments are provided so as to enable a more thorough and complete understanding of the present application. It is understood that the drawings and embodiments of the present application are for exemplary purposes only and should not be construed as limiting the scope of the present application.

[0064] It should be understood that each of the steps in the method embodiments of the present application can be performed in different orders and / or in parallel. In addition, the method embodiments can include additional steps and / or omit the steps shown. The scope of the present application is not limited in this respect.

[0065] The term "comprising" and variations thereof as used herein are used inclusively, i.e., "comprising but not limited to". The term "based on" is "based at least in part on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". Related definitions are given throughout the description.

[0066] It should be noted that the terms "first", "second", and so on used in the present application are only used to distinguish different devices, modules or units, and are not intended to limit the order or interdependence of the functions performed by these devices, modules or units.

[0067] In the embodiments of the present application, although the operations or steps are described in a specific order in the accompanying drawings, it should not be understood as requiring the operations or steps to be performed in the specific order or in a serial order, or requiring all of the shown operations or steps to be performed to obtain a desired result. In the embodiments of the present application, the operations or steps can be performed in series; the operations or steps can be performed in parallel; or a part of the operations or steps can be performed.

[0068] At the same time, it can be understood that the data involved in the technical solutions of the present application (including but not limited to the data itself, the acquisition or use of the data) should comply with the requirements of the corresponding laws, regulations and relevant provisions. Unless otherwise defined, all technical and scientific terms used in the present application have the same meaning as understood by those skilled in the art to which the present application belongs, and all parameters or indicators in the formulas involved in the present application are normalized values that eliminate the dimension influence.

[0069] In order to solve the problem of poor accuracy of adverse drug reaction prediction for existing skin disease patients, the embodiments of the present application provide an adverse drug reaction prediction system using skin electronic medical records. The system is essentially a software system composed of various modules that realize corresponding functions, and the corresponding structural diagram is shown in Figure 1 The core of the system is to realize an adverse drug reaction prediction method using skin electronic medical records. Each module in the system corresponds to each step in the method, and the corresponding flowchart is shown in Figure 2 The modules of the system will be described in detail below in combination with the specific steps in the method.

[0070] The data acquisition module is used to acquire skin electronic medical records of a plurality of historical patients. The skin electronic medical records include diagnosis information, and part of the skin electronic medical records also include re-consultation information. The diagnosis information and the re-consultation information include the occurrence of each feature and the occurrence of adverse drug reactions of patients. The features include patient symptoms and drug components used in treatment plans.

[0071] Specifically, in order to facilitate subsequent construction of an adverse drug reaction prediction model, a plurality of historical skin electronic medical record samples need to be acquired first. In this embodiment, skin electronic medical records of a plurality of historical patients are acquired, and the skin electronic medical records contain diagnosis information of the historical patients. The number of historical patients can be reasonably set according to needs, for example, set to 1000 historical patients. At the same time, if the historical patients have re-consultation records, the skin electronic medical records of these historical patients also contain re-consultation information of the historical patients.

[0072] The diagnosis information and the re-visit information contained in the electronic medical records of the skin of the several historical patients include the occurrence of each feature and the occurrence of adverse drug reactions of the patients. Each feature includes patient symptoms and features in the drug treatment plan. Each feature of the patient symptoms includes whether the patient has a certain skin manifestation or is allergic to a certain drug, and the skin manifestations include itching, rash, desquamation, exudation, etc. Each feature of the drug treatment plan refers to whether a certain drug component is used in the treatment plan, and the drug component includes chloramphenicol, erythromycin, mupirocin, loratadine, etc. By inputting each feature in the diagnosis information and the re-visit information contained in the electronic medical record of the skin as a binary variable, taking whether there is a certain skin manifestation as an example, when there is a certain skin manifestation, the feature is quantified as “1”, and when there is no certain skin manifestation, the feature is quantified as “0”. The same quantification processing is performed on whether the patient is allergic to a certain drug and whether a certain drug component is used. If the patient is allergic to a certain drug or uses a certain drug component, the corresponding feature is quantified as “1”, otherwise the corresponding feature is quantified as “0”. For example, the patient uses loratadine and does not use mupirocin, so the corresponding variable of loratadine is quantified as 1, and the corresponding variable of mupirocin is quantified as 0.

[0073] In this way, by quantifying each feature in the diagnosis information and the re-visit information contained in the electronic medical record of the skin, the quantified value of each feature in the diagnosis information and the re-visit information contained in the electronic medical record of the skin can be obtained. When the quantified value is 1, it means that the corresponding feature appears, and when the quantified value is 0, it means that the corresponding feature does not appear. It should be understood that the electronic medical records of the skin of the several historical patients are all medical record data obtained with the permission of the patients.

[0074] The basic correlation degree acquisition module is configured to count the occurrence of each feature in the diagnosis information of all historical patients and the occurrence of adverse drug reactions of the patients, and determine the basic correlation degree of each feature by combining the differences in the occurrence of each feature in the diagnosis information and the re-visit information of the same historical patient and the differences in the occurrence of adverse drug reactions of the patients.

[0075] Specifically, the electronic medical record data of the skin contains a large amount of patient feature information, and the information is usually high-dimensional, diversified, and has a complex nonlinear relationship. Random forest can automatically process various features, adapt to the diversity of data, extract effective feature information for prediction by integrating multiple decision trees. Therefore, the embodiment establishes an adverse drug reaction prediction model based on random forest. In the process of establishing the adverse drug reaction prediction model based on random forest, a plurality of sub-training sets are extracted from the training set, and each sub-training set trains a decision tree. After training a plurality of decision trees, the current to-be-predicted quantity is input into each decision tree to obtain a plurality of output results, and a combiner is constructed to combine the output results of each decision tree as the output of the final random forest.

[0076] Since the features used by each decision tree in the random forest are a part of all the features of the skin disease patients randomly extracted, the random extraction manner of the random forest is affected by the corresponding association relationship between different features in the electronic medical record of the skin disease patients and the adverse drug reaction, resulting in weakening of the feature association relationship in the random division process. By analyzing the correlation degree of each feature and the occurrence of adverse drug reactions according to the electronic medical records of all historical patients, the reference weight of each decision tree is determined according to the features contained in each decision tree, and the output results of each decision tree are weighted using the reference weight, which can effectively improve the prediction accuracy of the random forest.

[0077] Considering that the electronic medical record of the skin disease patient contains a plurality of features, the manifestations of a plurality of features can cause the patient to have an adverse drug reaction, and the manifestations of the adverse drug reaction are related to the skin surface thickness, ultraviolet sensitivity and other features of the patient. Different features such as different skin properties have different correlation degrees with the manifestations of the adverse drug reaction. When the patient has a skin desquamation or peeling symptom, the skin surface of the patient becomes thin and the skin becomes more sensitive. When the skin is stimulated by the outside world, the patient is prone to have an adverse drug reaction such as skin rash. The correlation degree between different features and adverse drug reactions can be reflected by the statistical result difference in the electronic medical record of all historical patients.

[0078] Further, the above basic correlation degree acquisition module comprises:

[0079] A frequency difference acquisition unit is configured to count the occurrence of the target feature and the occurrence of the adverse drug reaction of the patient in the diagnosis information of all historical patients, and determine the frequency difference of the target feature. The frequency difference reflects the frequency difference of the occurrence of the adverse drug reaction of the patient when the target feature appears or does not appear.

[0080] a change coefficient acquisition unit configured to determine a change coefficient of the target item feature according to whether there is a difference between whether the target item feature appears in the diagnosis information and the re-visit information of the same historical patient and whether there is a difference between whether the adverse drug reaction of the patient appears in the historical patient;

[0081] a basic correlation degree acquisition unit configured to determine a basic correlation degree of the target item feature according to a difference between the frequency difference of the target item feature and the frequency difference of the other item feature and the change coefficient of the target item feature.

[0082] Specifically, since different features in the electronic medical record of the skin disease patient have different degrees of correlation with whether the adverse drug reaction appears, some features can only cause adverse drug reactions in certain specific environments. For example, some drugs that cause photosensitivity usually cause adverse reactions only under high-intensity ultraviolet radiation, while tetracycline and other drugs can activate the body's immune system, causing the human immune system to produce an allergic reaction to the drug, which can easily cause adverse reactions. By statistically analyzing the probability of whether the adverse drug reaction appears in each feature of all historical patients, the basic correlation degree of each feature can be obtained to reflect the frequency of the adverse drug reaction under different manifestations of each feature.

[0083] When the frequency difference of the adverse drug reaction when the manifestation of a certain feature appears or does not appear is large, the basic correlation degree of the feature is larger, which reflects that the feature is more likely to be associated with the manifestation of the adverse drug reaction. If the presence or absence of a certain feature has little effect on the frequency of the adverse drug reaction, the basic correlation degree of the feature is smaller.

[0084] Further, the frequency difference acquisition unit includes: a first frequency determination unit configured to determine a ratio of a number of all historical patients in which the target item feature appears in the diagnosis information and the adverse drug reaction of the patient appears to a number of all historical patients in which the target item feature appears in the diagnosis information, to obtain a first frequency; a second frequency determination unit configured to determine a ratio of a number of all historical patients in which the target item feature does not appear in the diagnosis information and the adverse drug reaction of the patient appears to a number of all historical patients in which the target item feature does not appear in the diagnosis information, to obtain a second frequency; and a frequency difference determination unit configured to determine an absolute value of a difference between the first frequency and the second frequency, to obtain the frequency difference of the target item feature.

[0085] In this embodiment, the first frequency is determined as the number of historical patients in which the target item feature appears in the diagnosis information and the adverse drug reaction of the patient appears, and the second frequency is determined as the number of historical patients in which the target item feature does not appear in the diagnosis information and the adverse drug reaction of the patient appears. The target item feature is taken as an example, the first frequency of the target item feature is determined as the number of historical patients in which the target item feature appears in the diagnosis information and the adverse drug reaction of the patient appears, and the second frequency of the target item feature is determined as the number of historical patients in which the target item feature does not appear in the diagnosis information and the adverse drug reaction of the patient appears. The frequency difference of the target item feature is calculated as the absolute value of the difference between the first frequency and the second frequency. The calculation formula of the frequency difference of the target item feature is:

[0086] ​

[0087] wherein: represents the number of all historical patients who have the th feature in the diagnosis information and have adverse drug reactions, i.e., the number of all historical patients who have the th feature with a quantified value of 1 and have adverse drug reactions; represents the number of all historical patients who have the th feature in the diagnosis information, i.e., the number of all historical patients who have the th feature with a quantified value of 1; represents the number of all historical patients who do not have the th feature in the diagnosis information and have adverse drug reactions, i.e., the number of all historical patients who do not have the th feature with a quantified value of 0 and have adverse drug reactions; represents the number of all historical patients who do not have the th feature in the diagnosis information, i.e., the number of all historical patients who do not have the th feature with a quantified value of 0.

[0088] In the above formula, when the first frequency and the second frequency are significantly different, it indicates that the frequency of the occurrence of adverse drug reactions in patients with or without the th feature is significantly different, and the th feature is more likely to be associated with the manifestation of adverse drug reactions.

[0089] Since the electronic medical record of the skin of the historical patient will present a phased increase in content as the reexamination proceeds, including diagnosis information, medication information, etc., its general condition is that according to the manifestation of skin drug reactions, the category of drug types is added or subtracted or the dosage is changed, and then the diagnosis features of the patient change. At this time, if the same patient has several features change, the situation of adverse drug reactions also changes, such as from none to some or from some to none, which indicates that these features are more related to adverse drug reactions.

[0090] In this embodiment, the th feature is taken as the target feature, and the calculation formula of the change coefficient of the th feature is as follows:

[0091] ;

[0092] wherein: represents the total number of historical patients including reexamination information in the electronic medical record of the skin; represents the number of historical patients including reexamination information in the electronic medical record of the skin who have the a difference marker value of whether the patient adverse drug reaction appears in the diagnosis information and the reexamination information of the skin electronic medical record of the historical patient, when the patient adverse drug reaction appears in both the diagnosis information and the reexamination information, the difference marker value is 1, otherwise the difference marker value is 0; , otherwise the difference marker value is 0; ; represents a difference marker value of whether the i-th feature appears in the diagnosis information and the reexamination information of the skin electronic medical record of the historical patient, when the i-th feature appears in both the diagnosis information and the reexamination information, the difference marker value is 1, otherwise the difference marker value is 0; , otherwise the difference marker value is 0; ; represents a linear normalization function.

[0093] In the above formula, when more historical patients appear the i-th feature in the diagnosis information and the reexamination information changes, and the corresponding adverse drug reaction also changes, it indicates that the i-th feature is more likely to be associated with the manifestation of the adverse drug reaction. The frequency difference reflects the association between the presence or absence of each feature in the diagnosis information of all skin electronic medical records and the presence or absence of adverse drug reactions. If the patient's adverse drug reaction changes in diagnosis and reexamination, the features with different manifestations in diagnosis and reexamination are more likely to be associated with adverse drug reactions. Therefore, the basic association degree of each feature and adverse drug reactions can be obtained by combining the frequency difference and the reexamination change.

[0094] Further, the basic association degree obtaining unit comprises: an average frequency difference determining unit, configured to determine an average frequency difference according to the overall distribution level of the frequency differences of all features; a frequency difference relative value determining unit, configured to determine a ratio of the frequency difference of the target feature to the average frequency difference to obtain a frequency difference relative value; a frequency difference relative value increment determining unit, configured to determine a product of the frequency difference relative value and the change coefficient of the target feature to obtain a frequency difference relative value increment; and a basic association degree determining unit, configured to normalize the sum of the frequency difference relative value and the frequency difference relative value increment to obtain the basic association degree of the target feature.

[0095] In this embodiment, the i-th feature is taken as the target feature, and the basic association degree of the i-th feature is calculated as follows:

[0096]

[0097] ​​​​​​​​ ;

[0098] in: Indicates the first The coefficient of variation of the feature; Indicates the first Frequency differences of the features; It represents the average frequency difference, that is, the average of the frequency differences of all features; This represents the linear normalization function.

[0099] In the above formula, by calculating the first... The ratio of the frequency difference of a feature to the average frequency difference yields the relative value of the frequency difference. And determine the relative value of the frequency difference with the first The product of the coefficients of change of the features yields the relative increment of the frequency difference. and then utilize The sum of the relative value of the frequency difference and the increment of the relative value of the frequency difference. Normalization is performed to obtain the basic correlation. When the first The larger the relative value of the frequency difference of the feature, and the more significant the difference in the frequency of the first feature. The larger the coefficient of change of a feature term, the more it indicates that the first feature term... The more likely a feature is to be associated with adverse drug reactions, the higher its basic correlation. The larger the value, the better.

[0100] Using the above method, the basic correlation of all features in the electronic skin medical records of all historical patients can be determined.

[0101] The effective coefficient acquisition module is used to train several decision trees using the skin electronic medical records of the historical patients. Based on the distribution of historical patients who had adverse drug reactions before and after the division of the parent node by the target item feature in the decision tree, and the number of the parent node divided by each feature in the decision tree, the effective coefficient of the basic correlation degree of each feature is determined.

[0102] Specifically, according to the acquisition process of the basic correlation degree of each feature according to all the above features, the basic correlation degree of each feature is obtained according to the statistical results of the frequency of the occurrence of adverse drug reactions of each feature in all historical patients. The basic correlation degree can reflect the correlation between the feature and the adverse drug reaction to some extent, but since each historical patient has multiple features, some features may not affect the occurrence of adverse drug reactions, but if they appear frequently and are accompanied by high-correlation features in most medical records, their basic correlation degree will also be high. Therefore, the basic correlation degree obtained only according to the overall statistical results cannot be directly used as the correlation degree of each feature and adverse drug reactions, and further analysis is required to accurately determine the correlation degree of each feature.

[0103] Considering that the data subsets used in training several decision trees using the electronic medical records of the skin of the historical patients are different, that is, different samples and features are used for each decision tree, which can effectively reduce the dependence of the prediction model on a single data sample, and that the analysis in each decision tree can discuss the division of patients when some features are the same and some features are consistent on a smaller scale.

[0104] Further, the above effective coefficient acquisition module comprises:

[0105] The division evaluation acquisition unit is configured to determine a division evaluation of the target feature on the decision tree according to the difference between the proportion of historical patients with adverse drug reactions in each pair of sibling nodes and the proportion of historical patients with adverse drug reactions in the parent node after the target feature divides the parent node in the decision tree, and the difference between the proportion of historical patients with adverse drug reactions in each sibling node in each pair of sibling nodes.

[0106] The division efficiency acquisition unit is configured to determine the division efficiency of the target feature according to the division evaluation of the target feature on the decision tree and the number of layers of the parent node divided by the target feature in the decision tree.

[0107] The effective coefficient acquisition unit is configured to determine the effective coefficient of the basic correlation degree of the target feature according to the difference between the division efficiency of the target feature and the division efficiency of other features.

[0108] Specifically, a plurality of decision trees are trained using information in the electronic medical records of the skin of the historical patients. 80% of the total number of electronic medical records of the skin of all historical patients is recorded as the training set size N, and the remaining 20% of the electronic medical records of the skin is recorded as the test set. The number of categories of all features is recorded as M.

[0109] A subset of size n is randomly selected with replacement from the training set N. Furthermore, m features are randomly selected without replacement from all features M. A decision tree is trained using the selected subset of size n and the m features. Each decision tree outputs whether an adverse drug reaction occurs; if an adverse drug reaction occurs, the decision tree outputs "1", otherwise it outputs "0". The number of decision trees is preset to R. This training process is repeated R times to obtain a total of R decision trees. Since the process of training decision trees is existing technology, it will not be described in detail here.

[0110] Because each layer of a decision tree represents a partition of a set using a single feature, such as Figure 3 As shown, Y represents a patient with a history of adverse drug reactions, and X represents a patient with a history of no adverse drug reactions. As the decision tree deepens, the features between sibling nodes become more consistent, meaning they exhibit consistency across multiple features. For example, patients within the same node at level 4 represent those whose performance across the three previously defined features is consistent. For multiple consistent patients, if a significant difference in the frequency of adverse drug reactions occurs after partitioning based on a particular feature, it indicates a stronger association between that feature and adverse drug reactions.

[0111] Furthermore, the aforementioned partitioning evaluation acquisition unit includes: a first proportion difference determination unit, used to determine the difference between the maximum value of the proportion of historical patients with adverse drug reactions in each pair of sibling nodes obtained after the target feature partitions the parent node in the decision tree and the proportion of historical patients with adverse drug reactions in the parent node, to obtain a first proportion difference; a second proportion difference determination unit, used to determine the difference between the proportions of historical patients with adverse drug reactions in each pair of sibling nodes obtained after the target feature partitions the parent node in the decision tree, to obtain a second proportion difference; and a partitioning evaluation determination unit, used to determine the overall distribution level of the product values ​​of the first proportion difference and the second proportion difference corresponding to each pair of sibling nodes obtained after the target feature partitions the parent node in the decision tree, to determine the partitioning evaluation of the target feature on the decision tree.

[0112] In this embodiment, the first Taking the target feature as an example, for the first feature... Decision trees, if the first The feature appears in the first In the decision tree, the first Feature of the first Evaluation of decision tree partitioning The calculation formula is:

[0113] ;

[0114] in: represents the maximum value of the proportion of historical patients with adverse drug reactions in each pair of sibling nodes obtained after the parent node is divided by the i-th feature in the j-th decision tree. represents the maximum value of the proportion of historical patients with adverse drug reactions in each pair of sibling nodes obtained after the parent node is divided by the i-th feature in the j-th decision tree. represents the maximum value of the proportion of historical patients with adverse drug reactions in each pair of sibling nodes obtained after the parent node is divided by the i-th feature in the j-th decision tree. Figure 3 represents the maximum value of the proportion of historical patients with adverse drug reactions in each pair of sibling nodes obtained after the parent node is divided by the i-th feature in the j-th decision tree. represents the maximum value of the proportion of historical patients with adverse drug reactions in each pair of sibling nodes obtained after the parent node is divided by the i-th feature in the j-th decision tree. represents the maximum value of the proportion of historical patients with adverse drug reactions in each pair of sibling nodes obtained after the parent node is divided by the i-th feature in the j-th decision tree. represents the maximum value of the proportion of historical patients with adverse drug reactions in each pair of sibling nodes obtained after the parent node is divided by the i-th feature in the j-th decision tree. represents the maximum value of the proportion of historical patients with adverse drug reactions in each pair of sibling nodes obtained after the parent node is divided by the i-th feature in the j-th decision tree. represents the maximum value of the proportion of historical patients with adverse drug reactions in each pair of sibling nodes obtained after the parent node is divided by the i-th feature in the j-th decision tree. represents the maximum value of the proportion of historical patients with adverse drug reactions in each pair of sibling nodes obtained after the parent node is divided by the i-th feature in the j-th decision tree. represents the maximum value of the proportion of historical patients with adverse drug reactions in each pair of sibling nodes obtained after the parent node is divided by the i-th feature in the j-th decision tree. represents the maximum value of the proportion of historical patients with adverse drug reactions in each pair of sibling nodes obtained after the parent node is divided by the i-th feature in the j-th decision tree. represents the maximum value of the proportion of historical patients with adverse drug reactions in each pair of sibling nodes obtained after the parent node is divided by the i-th feature in the j-th decision tree. represents the maximum value of the proportion of historical patients with adverse drug reactions in each pair of sibling nodes obtained after the parent node is divided by the i-th feature in the j-th decision tree. represents the maximum value of the proportion of historical patients with adverse drug reactions in each pair of sibling nodes obtained after the parent node is divided by the i-th feature in the j-th decision tree. represents the maximum value of the proportion of historical patients with adverse drug reactions in each pair of sibling nodes obtained after the parent node is divided by the i-th feature in the j-th decision tree. represents the maximum value of the proportion of historical patients with adverse drug reactions in each pair of sibling nodes obtained after the parent node is divided by the i-th feature in the j-th decision tree. represents the maximum value of the proportion of historical patients with adverse drug reactions in each pair of sibling nodes obtained after the parent node is divided by the i-th feature in the j-th decision tree.

[0115] In the above formula, for each decision tree with the i-th feature, the greater the difference between the maximum value of the proportion of historical patients with adverse drug reactions in each pair of sibling nodes obtained after the parent node is divided by the i-th feature and the proportion of historical patients with adverse drug reactions in the parent node, and the greater the difference between the proportion of historical patients with adverse drug reactions in each pair of sibling nodes, the greater the degree of association between the i-th feature and adverse drug reactions.

[0116]

[0117] ​​​​Furthermore, the above-mentioned partitioning efficiency acquisition unit includes: a first partitioning efficiency determination unit, used to determine the partitioning evaluation of the target item feature on the decision tree and the product of the number of the parent node of the target item feature in the decision tree, to obtain the partitioning efficiency of the target item feature on a single decision tree; and a second partitioning efficiency determination unit, used to determine the overall distribution level of the partitioning efficiency of the target item feature on all decision trees, and to determine the partitioning efficiency of the target item feature.

[0118] In this embodiment, the first Taking the target feature as an example, the first feature is... Efficiency of feature partitioning The calculation formula is:

[0119] ;

[0120] in: Indicates containing the first The number of decision trees for each feature; Indicates the first The first decision tree The parent node divided by the feature is in the th The number of levels in a decision tree; Indicates the first Feature of the first Evaluation of decision tree partitioning; This represents the linear normalization function.

[0121] When a feature has low segmentation efficiency but high basic correlation, it indicates that the high basic correlation of that feature is due to other concurrent features, rather than the feature itself being directly and significantly associated with adverse drug reactions. For example, pruritus is a common symptom in many skin conditions. Pruritus does not directly cause adverse reactions, but may be a symptom of other conditions or an adverse reaction caused by medications treating pruritus. Analyzing the association between this feature and adverse reactions solely through basic correlation analysis may yield an overestimation. Therefore, by considering the differences in segmentation efficiency among features, determining the effective coefficient of basic correlation for each feature, and then combining the basic correlation and its effective coefficient, the degree of association for each feature can be accurately determined.

[0122] Further, the effective coefficient obtaining unit comprises: an average partition efficiency determining unit, configured to determine an average partition efficiency according to an overall distribution level of the partition efficiencies of all the item features; a partition efficiency difference determining unit, configured to determine a multiplication value of the average partition efficiency and a preset effective threshold, and further determine a difference between the partition efficiency of the target item feature and the multiplication value to obtain a partition efficiency difference; and an effective coefficient determining unit, configured to input the partition efficiency difference into a ReLU function, and output the effective coefficient of the basic correlation degree of the target item feature from the ReLU function.

[0123] In the embodiment, the preset effective threshold is 0.5, which limits the lower limit of the partition efficiency to avoid the case that the value of the basic correlation degree is too high when the partition efficiency of an item feature is too low. For example, the first item feature is taken as the target item feature, and the basic correlation degree of the first item feature is taken as the basic correlation degree of the target item feature. The calculation formula of the effective coefficient of the basic correlation degree of the first item feature is as follows:

[0124]

[0125] In the formula, the first item feature is taken as the target item feature, and the basic correlation degree of the first item feature is taken as the basic correlation degree of the target item feature. represents the partition efficiency of the first item feature; represents the preset effective threshold; represents the average value of the partition efficiencies of all the item features, i.e., the average partition efficiency; represents the ReLU function, and the output of the ReLU function is 0 when the input is less than 0, and the output of the ReLU function is the input when the input is greater than or equal to 0. According to the above manner, the effective coefficients of the basic correlation degrees of all the item features in the electronic medical records of the skin of all the historical patients can be determined.

[0126] The correlation degree obtaining module is configured to determine the correlation degree of each item feature according to the basic correlation degree and the effective coefficient of each item feature.

[0127] Specifically, the correlation degree of each item feature can be determined by comprehensively considering the basic correlation degree and the effective coefficient of each item feature, and the higher the basic correlation degree and the effective coefficient are, the higher the correlation degree between the corresponding feature and the adverse drug reaction is.

[0128] Further, the correlation degree obtaining module comprises: a correlation degree growth value obtaining unit, configured to determine the product of the basic correlation degree and the effective coefficient of each item feature to obtain a correlation degree growth value; and a correlation degree obtaining unit, configured to normalize the sum of the basic correlation degree of each item feature and the correlation degree growth value to obtain the correlation degree of each item feature.

[0129] ​​​​

[0130] In this embodiment, the first Taking the target feature as an example, the first feature is... degree of correlation of features The calculation formula is:

[0131] ;

[0132] in: Indicates the first The effective coefficient of the basic correlation degree of the feature; Indicates the first The basic correlation of the features; This represents the linear normalization function.

[0133] In the above formula, by calculating the first... The product of the basic correlation degree of a feature and its effective coefficient yields the correlation degree growth value. And thus the first Normalize the sum of the basic correlation degree of the feature and the growth value of that correlation degree to obtain the first feature. degree of correlation of features .

[0134] Following the above method, the correlation degree of all features in the electronic skin medical records of all historical patients can be determined. The basic correlation degree of all features is obtained through statistical results. The effective coefficient of the basic correlation degree is determined by assessing the segmentation efficiency of all features. This effective coefficient is then used to correct the basic correlation degree, ultimately yielding the correlation degree of all features. Compared to obtaining the basic correlation degree of all features solely through statistical results, calculating the segmentation efficiency of all features eliminates the influence of some other features, making the correlation between changes in the performance of each feature and adverse drug reactions more reliable.

[0135] The weight acquisition module is used to determine the reference weights of the decision tree based on the degree of correlation of all the features contained in the decision tree.

[0136] Specifically, the above analysis of the correlation between each feature and adverse drug reaction is based on the changes in the proportion of adverse drug reactions. Since different features have different degrees of correlation with adverse drug reactions, and the features in each decision tree are randomly selected, different decision trees have different reference weights for the output results of the random forest.

[0137] Furthermore, the aforementioned weight acquisition module includes:

[0138] An average correlation degree obtaining unit is configured to determine an average correlation degree corresponding to the decision tree according to an overall distribution level of the correlation degrees of all item features contained in the decision tree;

[0139] A reference weight obtaining unit is configured to normalize the average correlation degree corresponding to the decision tree to obtain a reference weight of the decision tree, and the cumulative value of the reference weights of all the decision trees is equal to 1.

[0140] In this embodiment, the average value of the correlation degrees of all item features contained in each decision tree is determined to obtain the average correlation degree corresponding to each decision tree. The cumulative value of the average correlation degrees corresponding to all the decision trees is calculated, and the ratio of the average correlation degree corresponding to each decision tree to the cumulative value of the average correlation degrees is calculated to normalize the average correlation degree corresponding to each decision tree by sum value. The ratio is taken as the reference weight of each decision tree. At this time, the cumulative value of the reference weights of all the decision trees is equal to 1.

[0141] In the above manner, the reference weights of all the decision trees can be determined.

[0142] A prediction module is configured to input the skin electronic medical record of a patient to be predicted into the decision trees, and weight the output results of the decision trees by using the reference weights to obtain a prediction result of adverse drug reactions of the patient to be predicted.

[0143] Specifically, the skin electronic medical record of the patient to be predicted is input into each decision tree to obtain the output result of each decision tree. The output result takes a value of 0 or 1, and the value of 0 indicates that no adverse drug reactions occur, and the value of 1 indicates that no adverse drug reactions occur. The output results of all the decision trees are multiplied by the corresponding reference weights to obtain the weighted output results of all the decision trees. Based on the weighted output results of all the decision trees, the weighted output results of each decision tree are combined by a combiner, and the prediction result of adverse drug reactions of the patient to be predicted can be determined. In this embodiment, the average value of the weighted output results of all the decision trees is determined, and the average value is taken as the probability of adverse drug reactions of the patient to be predicted. If the probability is greater than or equal to 0.5, the prediction result of adverse drug reactions of the patient to be predicted is that the patient to be predicted will have adverse drug reactions, otherwise the prediction result of adverse drug reactions of the patient to be predicted is that the patient to be predicted will not have adverse drug reactions. At the same time, the features involved in the decision trees with the maximum reference weight of the output result of 1 are marked as key features to show the key features that can cause adverse drug reactions of the current patient to the doctor, so as to facilitate the doctor to evaluate the situation of adverse drug reactions of the patient to be predicted and adjust the treatment plan according to the key features.

[0144] Based on the same inventive concept, the embodiments of the present application also provide a method for predicting adverse drug reactions by using electronic medical records of skin, as shown in Figure 2 The method comprises the following steps.

[0145] Obtaining electronic medical records of skin of a plurality of historical patients, wherein the electronic medical records of skin comprise diagnosis information, and part of the electronic medical records of skin further comprise re-visit information, the diagnosis information and the re-visit information comprise occurrence of each feature and occurrence of adverse drug reactions of patients, and the each feature comprises patient symptoms and drug components used in treatment plans;

[0146] Counting occurrence of each feature and occurrence of adverse drug reactions of patients in the diagnosis information of all historical patients, and determining basic correlation degrees of each feature by combining differences in occurrence of each feature and differences in occurrence of adverse drug reactions of patients in the diagnosis information and the re-visit information of the same historical patient;

[0147] Training a plurality of decision trees by using the electronic medical records of skin of the historical patients, determining effective coefficients of the basic correlation degrees of each feature according to distribution of historical patients with adverse drug reactions before and after target feature divides parent nodes in the decision trees, and number of layers of each feature in the decision trees;

[0148] Determining correlation degrees of each feature according to the basic correlation degrees and the effective coefficients of each feature;

[0149] Determining reference weights of the decision trees according to the correlation degrees of all features contained in the decision trees;

[0150] Inputting electronic medical records of skin of a patient to be predicted into the plurality of decision trees, weighting output results of the decision trees by using the reference weights, and thus obtaining a prediction result of occurrence of adverse drug reactions of the patient to be predicted.

[0151] Based on the same inventive concept, the embodiments of the present application also provide a device for predicting adverse drug reactions by using electronic medical records of skin, as shown in Figure 4 The device comprises a memory 401, a processor 402, and a computer program 403 stored in the memory 401 and running on the processor 402, wherein when the processor 402 executes the computer program 403, the system can execute the module implementation steps of any one of the adverse drug reaction prediction systems by using electronic medical records of skin introduced above.

[0152] The embodiment of the present application can divide the functions of the device into function modules according to the step examples of the modules in the system described above, for example, each function module can be obtained, or two or more functions can be integrated into one processing module, and the integrated module can be realized in the form of hardware. It should be noted that the division of the modules in the embodiment is illustrative, and is only a logical function division, and another division mode can be used in actual implementation.

[0153] Based on the same inventive concept, the embodiment of the present application further provides a computer program product, which comprises computer program code, when the computer program code is run on a computer, the computer executes the module implementation steps of any one of the adverse drug reaction prediction systems using skin electronic medical records introduced in the foregoing.

[0154] Based on the same inventive concept, the embodiment of the present application further provides a computer readable storage medium, which stores computer program code, when the computer program code is run on a computer, the computer executes the module implementation steps of any one of the adverse drug reaction prediction systems using skin electronic medical records introduced in the foregoing.

[0155] It should be noted that: the above-described embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A system for predicting adverse drug reactions using an electronic medical record of skin, characterized by, The system comprises: a data acquisition module configured to acquire electronic medical records of a plurality of historical patients, wherein the electronic medical records comprise diagnosis information, and some of the electronic medical records further comprise re-visit information, and the diagnosis information and the re-visit information comprise occurrence of each feature and occurrence of adverse drug reactions of the patients, wherein the features comprise patient symptoms and drug components used in treatment plans; a basic correlation degree acquisition module configured to count the occurrence of each feature and the occurrence of adverse drug reactions of the patients in the diagnosis information of all the historical patients, and determine a basic correlation degree of each feature according to differences in the occurrence of each feature and the occurrence of adverse drug reactions of the patients in the diagnosis information and the re-visit information of the same historical patient; an effective coefficient acquisition module configured to train a plurality of decision trees using the electronic medical records of the historical patients, determine an effective coefficient of the basic correlation degree of each feature according to a distribution of historical patients who have adverse drug reactions before and after a target feature divides a parent node in the decision trees, and a number of layers of the parent node divided by each feature in the decision trees; a correlation degree acquisition module configured to determine a correlation degree of each feature according to the basic correlation degree and the effective coefficient of each feature; a weight acquisition module configured to determine a reference weight of the decision trees according to the correlation degrees of all the features included in the decision trees; a prediction module configured to input electronic medical records of a patient to be predicted into the plurality of decision trees, and weight output results of the decision trees using the reference weight, so as to obtain a prediction result of adverse drug reactions of the patient to be predicted; the basic correlation degree acquisition module comprises: a frequency difference acquisition unit configured to count the occurrence of a target feature and the occurrence of adverse drug reactions of the patients in the diagnosis information of all the historical patients, and determine a frequency difference of the target feature, wherein the frequency difference reflects a frequency difference of the occurrence of adverse drug reactions of the patients when the target feature appears or not; a change coefficient acquisition unit configured to determine a change coefficient of a target feature according to a total number of historical patients whose diagnosis information and re-visit information have differences in the occurrence of the target feature and the occurrence of adverse drug reactions of the patients; a basic correlation degree acquisition unit configured to determine a basic correlation degree of a target feature according to a difference between the frequency difference of the target feature and frequency differences of other features, and the change coefficient of the target feature; the effective coefficient acquisition module comprises: a division evaluation acquisition unit configured to determine a division evaluation of a target feature to the decision trees according to a difference between a proportion of historical patients who have adverse drug reactions in each pair of sibling nodes obtained after the target feature divides a parent node in the decision trees and the parent node, and a difference between proportions of historical patients who have adverse drug reactions in each sibling node in the each pair of sibling nodes. The division efficiency obtaining unit is configured to determine the division efficiency of the target item feature according to the division evaluation of the target item feature on the decision tree and the number of layers of the parent node divided by the target item feature in the decision tree. The effective coefficient obtaining unit is configured to determine the effective coefficient of the basic correlation degree of the target item feature according to the difference between the division efficiency of the target item feature and the division efficiency of other item features.

2. The adverse drug reaction prediction system using an electronic skin medical record according to claim 1, characterized by, The frequency difference obtaining unit comprises: The first frequency determining unit is configured to determine a ratio of a number of all historical patients in which the target item feature appears and an adverse drug reaction occurs in the diagnostic information to a number of all historical patients in which the target item feature appears in the diagnostic information, to obtain a first frequency. The second frequency determining unit is configured to determine a ratio of a number of all historical patients in which the target item feature does not appear and an adverse drug reaction occurs in the diagnostic information to a number of all historical patients in which the target item feature does not appear in the diagnostic information, to obtain a second frequency. The frequency difference determining unit is configured to determine an absolute value of a difference between the first frequency and the second frequency, to obtain the frequency difference of the target item feature. 3.The adverse drug reaction prediction system using an electronic skin medical record of claim 1, wherein The basic correlation degree obtaining unit comprises: The average frequency difference determining unit is configured to determine an average frequency difference according to an overall distribution level of the frequency differences of all item features. The frequency difference relative value determining unit is configured to determine a ratio of the frequency difference of the target item feature to the average frequency difference, to obtain a frequency difference relative value. The frequency difference relative value increment determining unit is configured to determine a product of the frequency difference relative value and a change coefficient of the target item feature, to obtain a frequency difference relative value increment. The basic correlation degree determining unit is configured to normalize an addition value of the frequency difference relative value and the frequency difference relative value increment, to obtain the basic correlation degree of the target item feature.

4. The adverse drug reaction prediction system using an electronic skin medical record according to claim 1, characterized by, The division evaluation obtaining unit comprises: The first proportion difference determining unit is configured to determine a difference between a maximum value of a historical patient proportion in which an adverse drug reaction occurs in each pair of sibling nodes obtained after the target item feature divides a parent node in the decision tree and a historical patient proportion in which an adverse drug reaction occurs in the parent node, to obtain a first proportion difference. The second proportion difference determining unit is configured to determine a difference between the historical patient proportions in which an adverse drug reaction occurs in each pair of sibling nodes obtained after the target item feature divides the parent node in the decision tree, to obtain a second proportion difference. The division evaluation determining unit is configured to determine an overall distribution level of a product value of the first proportion difference and the second proportion difference corresponding to each pair of sibling nodes obtained after the target item feature divides the parent node in the decision tree, to determine the division evaluation of the target item feature on the decision tree. 5.The adverse drug reaction prediction system using a skin electronic medical record of claim 1, wherein The division efficiency obtaining unit comprises: The first division efficiency determining unit is configured to determine a product of the division evaluation of the target item feature on the decision tree and the number of layers of the parent node divided by the target item feature in the decision tree, to obtain the division efficiency of the target item feature on a single decision tree. The second split efficiency determination unit is configured to determine an overall distribution level of split efficiency of the target item feature on all decision trees, and determine the split efficiency of the target item feature.

6. The adverse drug reaction prediction system using an electronic skin medical record according to claim 1, wherein The effective coefficient acquisition unit comprises: The average split efficiency determination unit is configured to determine an average split efficiency according to an overall distribution level of split efficiency of all item features; The split efficiency difference determination unit is configured to determine a multiplication value of the average split efficiency and a set effective threshold, and further determine a difference value between the split efficiency of the target item feature and the multiplication value, to obtain a split efficiency difference; The effective coefficient determination unit is configured to input the split efficiency difference into a ReLU function, and output an effective coefficient of the basic correlation degree of the target item feature from the ReLU function.

7. The adverse drug reaction prediction system using an electronic skin medical record according to claim 1, characterized by, The correlation degree acquisition module comprises: The correlation degree growth value acquisition unit is configured to determine a product of the basic correlation degree of each item feature and the effective coefficient thereof, to obtain a correlation degree growth value; The correlation degree acquisition unit is configured to perform normalization processing on an added value of the basic correlation degree of each item feature and the correlation degree growth value, to obtain the correlation degree of each item feature. 8.The adverse drug reaction prediction system using a skin electronic medical record of claim 1, wherein, The weight acquisition module comprises: The average correlation degree acquisition unit is configured to determine an average correlation degree corresponding to the decision tree according to an overall distribution level of the correlation degrees of all item features contained in the decision tree; The reference weight acquisition unit is configured to perform normalization processing on the average correlation degree corresponding to the decision tree, to obtain a reference weight of the decision tree, and an added value of reference weights of all decision trees is equal to a numerical value 1.

Images