Artificial intelligence-based research and statistical analysis scheme recommendation method

Abstract

This invention relates to the field of machine learning technology and proposes a method for recommending research and statistical analysis schemes based on artificial intelligence. The method includes: collecting disease information data matrices and drug information data matrices; obtaining the correlation between each element in the disease information data matrix and other elements; obtaining the hierarchy of each element; obtaining the hierarchy coefficient of each element in the disease information data matrix and drug information data matrix; obtaining the hierarchy change curve of each row of each matrix; obtaining the corrected correlation and correlation feature vector of each row between the two matrices; constructing a neural network based on the disease information data matrix and drug information data matrix; obtaining the pruning degree of each neuron and pruning it based on the data matrix and correlation feature vector in the neurons; training the neural network and performing scheme recommendations. This invention aims to solve the problem that repetitive information affects the learning accuracy of the neural network during the intelligent recommendation process of training schemes through a neural network.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and more specifically to a method for recommending research and statistical analysis schemes based on artificial intelligence. Background Technology

[0002] In the current healthcare system, doctors need to formulate the best treatment plan based on disease diagnosis and drug information, while patients also need to choose the most suitable treatment plan based on this information. However, due to the complexity of medical knowledge, doctors and patients often find it difficult to grasp all the relevant information in a short period of time. Therefore, it is often necessary to combine machine learning methods to extract useful features and build recommendation models for intelligent recommendation of treatment plans. In addition, a large amount of drug information is stored in the database of the healthcare system, and each drug has corresponding indications. By analyzing the medication regimens in the database, corresponding indication reports are generated to assist physicians in formulating treatment plans for patients.

[0003] During the training of neural networks, pruning is often necessary due to limited computational resources and to avoid learning repetitive information. This pruning process involves removing some neurons and their corresponding edges. These neurons often contain repetitive or useless information, which can lead to incorrect learning and overfitting during training. Therefore, neurons containing repetitive or useless information must be removed. Summary of the Invention

[0004] This invention provides a method for recommending research and statistical analysis schemes based on artificial intelligence, to solve the problem that repetitive information affects the accuracy of neural network learning in existing intelligent recommendation schemes trained by neural networks. The specific technical solution adopted is as follows:

[0005] One embodiment of the present invention provides a method for recommending research and statistical analysis schemes based on artificial intelligence, the method comprising the following steps:

[0006] Collect disease information data and drug information data from several patients to obtain disease information data matrix and drug information data matrix;

[0007] Based on the word vectors obtained by word segmentation of each element in the disease information data matrix, the relationship between each element and other elements is obtained; the hierarchy of each element is obtained based on the relationship between elements; the elements are clustered based on the relationship between elements and the hierarchy of elements to obtain the hierarchy coefficient of each element in the disease information data matrix; the hierarchy coefficient of each element in the drug information data matrix is ​​obtained.

[0008] Based on the hierarchical coefficients of each row of elements in each matrix, obtain the hierarchical change curve of each row of elements in each matrix; combine the similarity between word vectors of elements at the same position in two matrices to obtain the corrected correlation and correlation feature vector of each row between the two matrices; construct a neural network based on the disease information data matrix and the drug information data matrix; obtain the pruning degree of each neuron based on the data matrix and correlation feature vector in the neuron.

[0009] Pruning is performed based on the degree of pruning of neurons to train the neural network.

[0010] Optionally, the relationship between each element and other elements is obtained using the following method:

[0011] For any element in the disease information data matrix, the text data corresponding to the element is processed by Jieba word segmentation to obtain several words of the element and the part of speech of each word. The word2vec model is used to convert each word into a word vector to obtain the word vector of each word of the element.

[0012] Taking the i-th element as the baseline, for the first word segment in the i-th element, obtain the cosine similarity between the word vector of each word segment in the l-th element that has the same part of speech as the first word segment and the word vector of the first word segment in the i-th element. Take the word segment corresponding to the word vector with the largest cosine similarity and combine it with the first word segment in the i-th element as a combination, which is denoted as the word combination of the first word segment in the i-th element in the l-th element. Obtain the word combination of each word segment in the i-th element in the l-th element. Based on the word combination of each word segment in each element in each other element and the cosine similarity between the words in the word combination, obtain the connection between each element and other elements.

[0013] Optionally, the specific methods for obtaining the relationship between each element and other elements include:

[0014]

[0015] Where, ω i,l J represents the correlation factor between the i-th element and the l-th element. i f represents the number of parts of speech corresponding to all words in the i-th element. i (j) represents the weight value of the j-th part of speech in the i-th element, U i (j) represents the number of sub-segments of the j-th part of speech in the i-th element, c i,l,u (j) represents the cosine similarity between the u-th segment of the j-th part of speech in the i-th element and the segmentation combination of the l-th element;

[0016] Obtain the relationship factor between the i-th element and every other element except the i-th element, perform linear normalization on all relationship factors, and denote the result as the relationship between the i-th element and every other element except the i-th element.

[0017] Optionally, the method for obtaining the hierarchy of each element based on the relationships between elements includes:

[0018] For the i-th element, if the correlation between the i-th element and any other element is greater than the correlation threshold, the element is taken as the associated element of the i-th element. Several associated elements of the i-th element are obtained. The ratio of the number of associated elements of the i-th element to the total number of elements in the disease information data matrix is ​​denoted as the hierarchy of the i-th element.

[0019] Optionally, the specific method for obtaining the hierarchical coefficient of each element in the disease information data matrix includes:

[0020] For any element, the mean of the correlation between the element and every other element is denoted as the correlation mean of the element; the absolute difference between the hierarchical relationship of any two elements and the absolute difference between the correlation mean are obtained, and the Euclidean norm is calculated for the two absolute difference values. The Euclidean norm is used as the distance measure between the two elements; all elements in the disease information data matrix are clustered, and the distance measure is the distance measure between any two elements that have been obtained, resulting in several clusters;

[0021] For any cluster, the hierarchical mean of all elements in the cluster is used as the hierarchical factor of the cluster. The hierarchical factor of each cluster is obtained, and all clusters are arranged in ascending order according to the hierarchical factor. The resulting sequence is called the cluster hierarchical sequence. For any element, the order value of the cluster to which the element belongs in the cluster hierarchical sequence is used as the hierarchical coefficient of the element.

[0022] Optionally, the specific method for obtaining the hierarchical variation curve of each row element of each matrix is ​​as follows:

[0023] A two-dimensional coordinate system is constructed with the horizontal axis representing the column number of the elements and the vertical axis representing the hierarchical coefficient of the elements. For any row element in the disease information data matrix, each element in the row is converted into a coordinate point in the two-dimensional coordinate system according to the hierarchical coefficient of the element and the column it belongs to. Connecting the adjacent coordinate points on the horizontal axis yields a curve, which is denoted as the hierarchical change curve of the element in that row of the disease information data matrix. The hierarchical change curves of each row element in the disease information data matrix and the drug information data matrix are obtained.

[0024] Optionally, the specific method for obtaining the corrected correlation and correlation feature vector of each row between the two matrices includes:

[0025] Based on the elements in the same position in the disease information data matrix and drug information data matrix of each row, and the word vectors of several word segments of the elements, obtain several element combinations in each row and the similarity of each element combination.

[0026]

[0027] Among them, R v s represents the corrected correlation of row v between the two matrices. v H represents the Pearson correlation coefficient between the hierarchical variation curve of the element in the v-th row of the disease information data matrix and the hierarchical variation curve of the element in the v-th row of the drug information data matrix. v z represents the number of combinations of elements in the v-th row. v,$ This indicates the similarity of the h-th element combination in the v-th row;

[0028] Obtain the corrected correlation of each row between the two matrices, arrange the corrected correlations of all rows in the order of the rows, and form a vector, which is denoted as the correlation feature vector.

[0029] Optionally, the specific method for obtaining the combinations of elements in each row and the similarity of each combination of elements is as follows:

[0030] For row v, the two elements in the same position in the disease information data matrix and the drug information data matrix of row v are combined as one element to obtain several element combinations of row v; for any element combination, the word vector of each word segment of the element belonging to the disease information data matrix in the element combination is obtained, and the cosine similarity is obtained with the word vector of each word segment of the element belonging to the drug information data matrix. The maximum cosine similarity is taken as the similarity of the element combination.

[0031] Optionally, the specific method for constructing the neural network based on the disease information data matrix and the drug information data matrix includes:

[0032] The disease information data matrix is ​​used as the training dataset, with each row of the disease information data matrix as the input and the corresponding row of the drug information data matrix as the output. A neural network is constructed, and the cross-entropy function is used as the loss function.

[0033] Optionally, the degree of pruning for each neuron can be obtained using the following method:

[0034] For the data matrix in the p-th neuron of the neural network, perform SVD decomposition on the data matrix to obtain several eigenvectors and their corresponding eigenvalues; the pruning degree τ of the p-th neuron. + The calculation method is as follows:

[0035]

[0036] Among them, M + E represents the number of eigenvectors obtained after SVD decomposition of the data matrix in the p-th neuron. +,0 E represents the eigenvalue corresponding to the m-th eigenvector of the data matrix in the p-th neuron. +,012 This represents the maximum eigenvalue among all eigenvectors corresponding to the data matrix in the p-th neuron. Let represent the m-th eigenvector of the data matrix in the p-th neuron, γ represent the correlation eigenvector, and cos(,) represent the cosine similarity between the two vectors.

[0037] The beneficial effects of this invention are as follows: This invention trains and prunes a neural network by semantic analysis and recognition of disease information data and drug information data of several patients, thereby improving the accuracy of the neural network's learning effect, avoiding repeated learning that leads to reduced training efficiency, and thus improving the accuracy of treatment plan recommendations. Specifically, the invention segments the text data corresponding to the elements in the matrix into words, and quantifies the connections between elements based on the similarity between the word vectors of the segments. Based on these connections, the hierarchy of the elements is obtained to reflect the generality of the information contained in the elements. The elements are then clustered based on these connections, and the hierarchical coefficients of the elements are obtained based on the clustering results, representing the degree of generality of the information contained in the elements. Then, by analyzing the temporal changes of the hierarchical coefficients of the elements in the same row in the two matrices, and combining the similarity of the combinations of elements in the same position, the correlation of each row between the two matrices is corrected, and a correlation feature vector is obtained. A neural network is constructed based on the two matrices. By performing feature analysis on the data matrix in the neurons, and combining the correlation feature vector, the degree of pruning is quantified by the distribution of correlation information, thereby reflecting the importance of neurons in the neural network learning process and providing a basis for subsequent pruning judgments. Finally, the neurons are pruned and the neural network is trained to achieve treatment plan recommendations and assist physicians in formulating treatment plans. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a schematic diagram of a method for recommending research and statistical analysis schemes based on artificial intelligence, provided in one embodiment of the present invention. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] Please see Figure 1 The diagram illustrates a flowchart of a method for recommending research and statistical analysis schemes based on artificial intelligence, provided by an embodiment of the present invention. The method includes the following steps:

[0042] Step S001: Collect disease information data and drug information data from several patients to obtain a disease information data matrix and a drug information data matrix.

[0043] The purpose of this embodiment is to train a neural network by collecting disease information data and drug information data from several patients, and then to recommend a treatment plan. Therefore, it is necessary to collect disease information data and drug information data from patients first.

[0044] Preferably, this embodiment obtains the patient's disease information and medication information data through the patient's medical records at different times in the hospital database, collecting data from the most recent 1000 patients. For any patient in the data collection, all of the patient's medical records at the hospital are collected. Each medical record corresponds to a treatment cycle, i.e., the time of the visit and the duration of medication required after that visit, which are combined into a single treatment cycle. The disease information data for each date within the treatment cycle is the disease information determined for that visit, and the medication information data for each date is determined according to the duration of medication required for the corresponding drug. For example, if a patient needs to take one drug until January 31 and another drug until February 15, then the patient's medication information data for January 31 includes both drugs, while the patient's medication information data for February 1 includes only one drug. Thus, for each date in the treatment cycle corresponding to each patient's medical record, disease information data and medication information data are obtained. It should be noted that both the disease information data and the medication information data are text data.

[0045] Furthermore, the earliest date of the first visit in the medical records of all patients collected is obtained, and this date is used as the date corresponding to the first column of the disease information data matrix; the latest date of the last day of the treatment cycle in the most recent medical records of all patients collected is obtained, and this date is used as the date corresponding to the last column of the disease information data matrix; each column in the disease information data matrix corresponds to a date from the first column to the last column, and adjacent columns correspond to adjacent dates, that is, starting from the date corresponding to the first column, each day corresponds to a column until the date corresponding to the last column; each row of the matrix contains several disease information data for one patient, where the patients are arranged in the order of collection; therefore, for any patient in the data collection, The disease information data for each date in the treatment cycle corresponding to several medical records of the patient is filled into the elements of the corresponding date column of the patient's row. For example, if there is a medical record on January 1st and medication is taken until February 15th, then the disease information data for February 1st is the same as the medical record for January 1st. If some dates are not within any of the patient's treatment cycles, then the elements of the corresponding date column of the patient's row are replaced with 0. For example, if the patient's next medical record is on March 1st, then the disease information data for February 20th is replaced with 0. By filling the elements of each patient's row into the disease information data matrix in the above manner, and adding 0 to dates that do not belong to the patient's treatment cycle, the disease information data matrix is ​​obtained.

[0046] Furthermore, the dates corresponding to each column in the drug information data matrix are the same as those in the disease information data matrix, and the patients corresponding to each row are the same as those in the disease information data matrix. Following the method for obtaining the disease information data matrix, the drug information data matrix is ​​filled with the drug information data for each date in the treatment cycle corresponding to each patient's medical record, thus obtaining the drug information data matrix.

[0047] Thus, by collecting the medical records of several patients, a disease information data matrix and a drug information data matrix were constructed and obtained.

[0048] It should be noted that because the drug information data matrix and the corresponding disease information data matrix have a dynamic temporal distribution, they are more representative than drug and disease information from a single period, thus the trained neural network performs better. To utilize this dynamic temporal distribution, the neural network is trained by analyzing the dynamic changes between the drug and disease information data. Therefore, in the neural network pruning process, if a neuron contains more dynamic changes, it indicates that the neuron should be retained. At the same time, in the judgment process, since there are certain connections between some disease information, for example, some are for major disease categories and some are for minor categories of diseases within the same major disease category, that is, there are hierarchical characteristics between different data in the disease information data matrix, it is necessary to analyze the dynamic changes under the influence of hierarchy.

[0049] Step S002: Based on the word vectors obtained by word segmentation of each element in the disease information data matrix, obtain the relationship between each element and other elements; obtain the hierarchy of each element based on the relationship between elements; cluster the elements based on the relationship between elements and the hierarchy of elements to obtain the hierarchy coefficient of each element in the disease information data matrix; obtain the hierarchy coefficient of each element in the drug information data matrix.

[0050] Preferably, in one embodiment of the present invention, the relationship between each element and other elements is obtained based on the word vectors obtained by word segmentation of each element in the disease information data matrix. The specific method includes:

[0051] For any element in the disease information data matrix, the text data corresponding to the element is processed by Jieba word segmentation to obtain several words of the element and the part-of-speech tag of each word. The word2vec model is used to convert the word vector of each word. In this embodiment, the dimension of the word vector is set to 10, so the word vector of each word of the element is obtained. The word vector and part-of-speech tag of several words of each element are obtained according to the above method. Jieba word segmentation and word2vec model are existing methods and will not be described in detail in this embodiment.

[0052] Furthermore, taking the i-th element as the baseline, for the first word segment in the i-th element, obtain the cosine similarity between the word vector of each word segment with the same part of speech as the first word in the l-th element (where l≠i) and the word vector of the first word segment in the i-th element. Take the word segment corresponding to the word vector with the largest cosine similarity and combine it with the first word segment in the i-th element, denoted as the word combination of the first word segment in the i-th element in the l-th element; obtain the word combination of each word segment in the i-th element in the l-th element. Then, the method for calculating the connection between the i-th element and the l-th element is as follows:

[0053]

[0054] Where, ω i,l J represents the correlation factor between the i-th element and the l-th element. i f represents the number of parts of speech corresponding to all words in the i-th element. i (j) represents the weight value of the j-th part of speech in the i-th element. In this embodiment, nouns are considered more important in medical scenarios, so the weight value of nouns is set to 0.8, the weight values ​​of adjectives and adverbs are set to 0.6, and the weight values ​​of other parts of speech are set to 0.4. In other embodiments, the implementer can set the weight values ​​of parts of speech according to the actual situation; U i (j) represents the number of sub-segments of the j-th part of speech in the i-th element, c i,l,u (j) represents the cosine similarity between the u-th segment of the i-th element under the j-th part of speech and the segmentation combination of the l-th element; obtain the connection factor between the i-th element and each other element except the i-th element according to the above method, perform linear normalization on all connection factors, and record the result as the connection between the i-th element and each other element except the i-th element; obtain the connection between each element and other elements according to the above method.

[0055] It should be noted that the cosine similarity of word segments under the same part of speech is averaged, and combined with the weight value of each part of speech, the connection between elements is comprehensively quantified. The larger the weight value, the larger the average cosine similarity of word segments under the part of speech, the more similar the distribution of the corresponding elements under the corresponding part of speech, and the greater the connection. By analyzing the cosine similarity between word segments of the same part of speech of different elements, the similarity of the distribution of different elements under each part of speech is characterized, and thus the correlation between different elements is obtained.

[0056] Preferably, in one embodiment of the present invention, the hierarchical nature of each element is obtained based on the relationships between elements, including the following specific method:

[0057] A connection threshold is preset. In this embodiment, the connection threshold is described as 0.65. For the i-th element, if the connection between the i-th element and any other element is greater than the connection threshold, the element is regarded as the associated element of the i-th element. Several associated elements of the i-th element are obtained. The ratio of the number of associated elements of the i-th element to the total number of elements in the disease information data matrix is ​​recorded as the hierarchy of the i-th element. The hierarchy of each element is obtained.

[0058] It should be noted that if an element has a strong connection with other elements, the more related elements there are, and the more other elements the element contains, the more comprehensive and hierarchical the disease information corresponding to that element becomes.

[0059] Preferably, in one embodiment of the present invention, the elements are clustered according to the relationships between elements and the hierarchical nature of the elements to obtain the hierarchical coefficient of each element in the disease information data matrix. The specific method includes:

[0060] For any element, the mean of the association between the element and every other element is denoted as the association mean of the element. The absolute difference between the hierarchical relationships of any two elements and the absolute difference between the association means are obtained. The Euclidean norm is calculated from these two absolute differences and used as the distance metric between the two elements. The K-means clustering algorithm is used to cluster all elements in the disease information data matrix. The distance metric is the distance between any two elements already obtained. In this embodiment, K is set to 5, resulting in 5 clusters. For any cluster, the mean of the hierarchical relationships of all elements in the cluster is used as the hierarchical factor of the cluster. The hierarchical factor of each cluster is obtained, and all clusters are arranged in ascending order according to their hierarchical factors. The resulting sequence is denoted as the cluster hierarchical sequence. For any element, the order value of the cluster to which the element belongs in the cluster hierarchical sequence is used as the hierarchical coefficient of the element. The hierarchical coefficient of each element in the disease information data matrix is ​​obtained.

[0061] It should be noted that by clustering the elements in the disease information data matrix, and considering both the differences in hierarchy and the differences in the mean of the relationship between the distance measure, the hierarchical relationship between the elements in the obtained clusters is more accurate. Elements in the same cluster have the same or similar hierarchical relationship. Then, the clusters are sorted according to the size of the mean of hierarchy in the cluster, and the hierarchy coefficient of each element is obtained. The larger the hierarchy factor of the cluster, the larger the hierarchy coefficient.

[0062] Furthermore, following the above method, the relationships between elements and the hierarchy of elements in the drug information data matrix are obtained, thereby obtaining the hierarchical coefficients of the elements in the drug information data.

[0063] Thus, by segmenting the text data corresponding to the elements in the matrix, and quantifying the connections between elements based on the similarity between the word vectors of the segmented words, the hierarchy of elements is obtained based on the connections, which reflects the generality of the information contained in the elements. The elements are then clustered based on the connections, and the hierarchical coefficient of the elements is obtained based on the clustering results, which characterizes the degree of generality of the information contained in the elements.

[0064] Step S003: Obtain the hierarchical change curve of each row element in each matrix based on the hierarchical coefficient of each row element in each matrix; combine the similarity between word vectors of elements at the same position in the two matrices to obtain the corrected correlation and correlation feature vector of each row between the two matrices; construct a neural network based on the disease information data matrix and the drug information data matrix; obtain the pruning degree of each neuron based on the data matrix and correlation feature vector in the neuron.

[0065] Preferably, in one embodiment of the present invention, the method for obtaining the hierarchical change curve of each row element of each matrix based on the hierarchical coefficient of each row element in each matrix includes:

[0066] A two-dimensional coordinate system is constructed with the horizontal axis representing the column number of the elements and the vertical axis representing the hierarchical coefficient of the elements. For any row element in the disease information data matrix, each element in the row is converted into a coordinate point in the two-dimensional coordinate system according to the hierarchical coefficient of the element and the column it belongs to. Connecting the adjacent coordinate points on the horizontal axis yields a curve, which is denoted as the hierarchical change curve of the element in that row of the disease information data matrix. The hierarchical change curves of each row element in the disease information data matrix and the drug information data matrix are obtained.

[0067] It should be noted that since the elements at the same position in the two matrices correspond in terms of date, the elements at the same position should also be related in terms of hierarchy. Therefore, it is necessary to obtain the hierarchical change curve of the elements in the same row and perform dynamic analysis based on the hierarchical change curve. By analyzing the hierarchical change of the elements in the same row of the two matrices in terms of time sequence, the correlation of each row between the two matrices can be corrected, which provides a basis for the subsequent quantification of the degree of pruning.

[0068] Preferably, in one embodiment of the present invention, after obtaining the hierarchical change curve of each row element of each matrix, the similarity between the word vectors of elements at the same position in the two matrices is combined to obtain the corrected correlation and correlation feature vector of each row between the two matrices. The specific method includes:

[0069] For row v, the two elements in the same position in both the disease information data matrix and the drug information data matrix are grouped together to obtain several element combinations for row v. For any element combination, the cosine similarity between the word vector of each word segment belonging to the disease information data matrix and the word vector of each word segment belonging to the drug information data matrix is ​​obtained. The maximum cosine similarity is taken as the similarity of the element combination. Then, the corrected correlation R between row v and the two matrices is calculated. v The calculation method is as follows:

[0070]

[0071] Among them, sv H represents the Pearson correlation coefficient between the hierarchical variation curve of the element in the v-th row of the disease information data matrix and the hierarchical variation curve of the element in the v-th row of the drug information data matrix. v z represents the number of combinations of elements in the v-th row. v,$ Let represent the similarity of the combination of elements in row v and h; obtain the corrected correlation of each row between the two matrices according to the above method, arrange the corrected correlations of all rows in the order of the rows, and form a vector, denoted as the correlation feature vector.

[0072] It should be noted that the average of the similarity of the element combinations is used as the correlation benchmark for that row between the two matrices. At the same time, the correlation benchmark needs to be adjusted in combination with the correlation coefficient of the hierarchical change curve. The larger the correlation coefficient, the more similar the hierarchical changes of the elements in the two rows are, and the greater the correlation adjustment and correction is required.

[0073] Preferably, in one embodiment of the present invention, the specific method for constructing a neural network based on a disease information data matrix and a drug information data matrix includes:

[0074] The disease information data matrix is ​​used as the training dataset, with each row of the disease information data matrix as the input and the corresponding row of the drug information data matrix as the output. A neural network is constructed using a CNN network structure and the cross-entropy function is used as the loss function.

[0075] It should be noted that by using the disease information data matrix as input and the drug information data matrix as the corresponding output to train the neural network, the data matrix corresponding to each neuron in the neural network has the same matrix size as the disease information data matrix and the drug information data matrix. Then, by analyzing the features of the data matrices in the neurons and combining them with the correlation feature vectors, the degree of pruning of the neurons can be quantified. If the information distribution obtained from the feature analysis is closer to the correlation feature vector, the more relevant information is contained, and the lower the degree of pruning should be; conversely, the less relevant information is contained, the higher the degree of pruning should be. This improves the learning efficiency of the neural network and avoids repeated learning of repetitive information.

[0076] Preferably, in one embodiment of the present invention, the pruning degree of each neuron is obtained based on the data matrix and correlation feature vector in the neuron, including the following specific method:

[0077] For the data matrix in the p-th neuron of the neural network, the SVD decomposition is performed on the data matrix to obtain several eigenvectors and their corresponding eigenvalues. SVD decomposition is a well-known technique and will not be described in detail in this embodiment; then the pruning degree τ of the p-th neuron is... + The calculation method is as follows:

[0078]

[0079] Among them, M + E represents the number of eigenvectors obtained after SVD decomposition of the data matrix in the p-th neuron. +,0 E represents the eigenvalue corresponding to the m-th eigenvector of the data matrix in the p-th neuron. +,012 This represents the maximum eigenvalue among all eigenvectors corresponding to the data matrix in the p-th neuron. Let represent the m-th eigenvector of the data matrix in the p-th neuron, γ represent the correlation eigenvector, and cos(,) represent the cosine similarity between the two vectors; obtain the pruning degree of each neuron according to the above method.

[0080] It should be noted that since the data matrix in the neuron is the same size as the disease information data matrix and the drug information data matrix, feature analysis is performed through SVD decomposition to obtain eigenvectors. The greater the similarity between the eigenvector and the correlation eigenvector, the more correlation distribution information the eigenvector represents. At the same time, the weight of each eigenvector is adjusted according to the size of the eigenvalue. The larger the eigenvalue and the more correlation distribution information the eigenvector contains, the more correlation distribution information the data matrix contains, and the smaller the degree of pruning should be.

[0081] Thus, by analyzing the temporal changes of the hierarchical coefficients of elements in the same row of the two matrices and combining the similarity of elements in the same position, the correlation of each row between the two matrices is corrected and a correlation feature vector is obtained. Based on the two matrices, a neural network is constructed. By performing feature analysis on the data matrix in the neurons and combining the correlation feature vector, the degree of pruning is quantified by the distribution of correlation information, thereby reflecting the importance of neurons in the neural network learning process and providing a basis for subsequent pruning judgments.

[0082] Step S004: Prune neurons according to their pruning degree, train the neural network, and recommend a solution.

[0083] Preferably, a preset pruning threshold is used. In this embodiment, the pruning threshold is described as 0.48. If the degree of pruning of a neuron in the neural network is greater than the pruning threshold, the neural network structure containing that neuron is pruned, and the neural network constructed above is trained to obtain a trained neural network for scheme recommendation.

[0084] It should be noted that the output of the trained neural network is only a suggestion and is intended to assist professional physicians in making judgments.

[0085] Furthermore, neural networks can intelligently generate corresponding treatment plans, i.e., medication plans, based on the doctor's diagnosis of the patient. These medication plans involve multiple drugs, and the hospital's database stores a large amount of data on these drugs, including the indications for each drug. Therefore, in this embodiment, after obtaining the medication plan, an SQL program is used to traverse the database and extract the indications for each drug involved in the medication plan. Based on the obtained indications, a treatment report is generated, containing medication recommendations for the patient and possible symptoms, i.e., several indications obtained from database analysis. The doctor then further diagnoses the patient based on the treatment report, and through the analysis of the medication plan and the indications in the database, assists the doctor in formulating a treatment plan for the patient.

[0086] Thus, through semantic analysis and recognition of disease and drug information data from several patients, the neural network is trained and pruned, thereby improving the accuracy of neural network learning, avoiding repeated learning that leads to reduced training efficiency, and ultimately improving the accuracy of treatment recommendations.

[0087] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for recommending research and statistical analysis schemes based on artificial intelligence, characterized in that, The method includes the following steps: Collect disease information data and drug information data from several patients to obtain disease information data matrix and drug information data matrix; Based on the word vectors obtained by word segmentation of each element in the disease information data matrix, the relationship between each element and other elements is obtained, specifically including: ； in, Indicates the first The element and the first The connection factor of each element, Indicates the first The number of part-of-speech tags corresponding to all word segments in each element Indicates the first The element in the nth element The weight value of each part of speech, Indicates the first The element in the nth element The number of words segmented under each part of speech Indicates the first The element in the nth element Parts of speech, second The word segmentation in the first The cosine similarity between two word segments in a word segmentation combination of elements; Get the The element and the element excluding the first The correlation factor of each element other than the first element is taken, and all correlation factors are linearly normalized. The result is denoted as the first element. The element and the element excluding the first The relationships between all elements except for the first element; The hierarchy of each element is derived from the relationships between elements, the first... The hierarchy of the elements is the first. The ratio of the number of associated elements of an element to the total number of elements in the disease information data matrix; based on the relationships between elements and the hierarchical nature of elements, clustering is performed to obtain the hierarchical coefficient of each element in the disease information data matrix. Specifically, this includes: for any cluster after clustering, the hierarchical mean of all elements in the cluster is used as the hierarchical factor of the cluster; the hierarchical factor of each cluster is obtained; all clusters are arranged in ascending order according to the hierarchical factor, and the resulting sequence is denoted as the cluster hierarchical sequence; for any element, the order value of the cluster to which the element belongs in the cluster hierarchical sequence is used as the hierarchical coefficient of the element; the hierarchical coefficient of each element in the drug information data matrix is ​​obtained. Based on the hierarchical coefficient of each row element in each matrix, obtain the hierarchical change curve of each row element in each matrix; combine the similarity between word vectors of elements at the same position in the two matrices to obtain the corrected correlation and correlation feature vector of each row between the two matrices. Specifically, based on the elements at the same position in the disease information data matrix and drug information data matrix of each row, and the word vectors of several word segments of the elements, obtain several element combinations of each row and the similarity of each element combination. ； in, Indicates the first Corrected correlation of rows between two matrices The disease information data matrix represents the first... The hierarchical change curve of the row element and the first row in the drug information data matrix The Pearson correlation coefficient of the hierarchical variation curve of row elements. Indicates the first The number of combinations of elements in a row. Indicates the first Line number Similarity of combinations of elements; Obtain the corrected correlation of each row between the two matrices, arrange the corrected correlations of all rows in the order of the rows, and form a vector, which is denoted as the correlation feature vector. A neural network is constructed based on disease information data matrices and drug information data matrices; the pruning degree of each neuron is obtained based on the data matrix and correlation feature vectors in the neurons, specifically including: for the first neuron in the neural network... The data matrix of the n neurons is decomposed using SVD to obtain several feature vectors and their corresponding feature values; The degree of pruning of each neuron The calculation method is as follows: ； in, Indicates the first The number of eigenvectors obtained after SVD decomposition of the data matrix in each neuron. Indicates the first The data matrix in the nth neuron The eigenvalues ​​corresponding to each eigenvector Indicates the first The maximum value among all eigenvalues ​​corresponding to the eigenvectors of the data matrix in each neuron. Indicates the first The data matrix in the nth neuron 1 eigenvector Represents the relevance feature vector. express and Cosine similarity; Pruning is performed on neurons to a certain degree to train the neural network; and the trained neural network is used to recommend research and statistical analysis schemes.

2. The method for recommending research and statistical analysis schemes based on artificial intelligence according to claim 1, characterized in that, The specific method for obtaining the relationship between each element and other elements is as follows: For any element in the disease information data matrix, the text data corresponding to the element is processed by Jieba word segmentation to obtain several words of the element and the part of speech of each word. The word2vec model is used to convert each word into a word vector to obtain the word vector of each word of the element. With the first Using the nth element as a reference, for the nth element... The first word segment in the nth element, get the nth... The word vector of each word in the elements that has the same part of speech as the first word is compared with the word vector of the second word. The cosine similarity of the word vector of the first word segment in the nth element is used to compare the word segment corresponding to the word vector with the largest cosine similarity with the nth element. The first word in each element is treated as a group, denoted as the first word. The first word in the element is in the... Word segmentation and combination of the nth element; obtain the nth element. Each word in the nth element is in the nth... The word segmentation combination of each element; based on the word segmentation combination of each element in each other element, and the cosine similarity between word segments in the word segmentation combination, the relationship between each element and other elements is obtained.

3. The method for recommending research and statistical analysis schemes based on artificial intelligence according to claim 1, characterized in that, The method for obtaining the hierarchy of each element based on the relationships between elements includes: For the The element, if the element If the relationship between an element and any other element is greater than a relationship threshold, then that element is considered the 1st element. Get the associated elements of the i-th element. The element has several associated elements, and the first element will be... The ratio of the number of associated elements of the i-th element to the total number of elements in the disease information data matrix is ​​denoted as the i-th element. The hierarchy of elements.

4. The method for recommending research and statistical analysis schemes based on artificial intelligence according to claim 1, characterized in that, The specific method for obtaining the hierarchical coefficient of each element in the disease information data matrix is ​​as follows: For any element, the mean of the correlation between that element and every other element is denoted as the correlation mean of that element. Obtain the absolute value of the difference between the hierarchical values ​​of any two elements and the absolute value of the difference between their mean values. Calculate the Euclidean norm for the two absolute values ​​of the difference and use the Euclidean norm as the distance measure between the two elements. Cluster all elements in the disease information data matrix, using the distance metric between any two already acquired elements, to obtain several clusters; For any cluster, the hierarchical mean of all elements in the cluster is used as the hierarchical factor of the cluster. The hierarchical factor of each cluster is obtained, and all clusters are arranged in ascending order according to the hierarchical factor. The resulting sequence is called the cluster hierarchical sequence. For any element, the order value of the cluster to which the element belongs in the cluster hierarchical sequence is used as the hierarchical coefficient of the element.

5. The method for recommending research and statistical analysis schemes based on artificial intelligence according to claim 1, characterized in that, The specific method for obtaining the hierarchical change curve of each row element in each matrix is ​​as follows: A two-dimensional coordinate system is constructed with the horizontal axis representing the column number of the elements and the vertical axis representing the hierarchical coefficient of the elements. For any row element in the disease information data matrix, each element in the row is converted into a coordinate point in the two-dimensional coordinate system according to the hierarchical coefficient of the element and the column it belongs to. Connecting the adjacent coordinate points on the horizontal axis yields a curve, which is denoted as the hierarchical change curve of the element in that row of the disease information data matrix. The hierarchical change curves of each row element in the disease information data matrix and the drug information data matrix are obtained.

6. The method for recommending research and statistical analysis schemes based on artificial intelligence according to claim 1, characterized in that, The specific method for obtaining the combinations of elements in each row and the similarity of each element combination is as follows: For the Okay, the first Two elements that occupy the same position in the disease information data matrix and the drug information data matrix are combined as one element to obtain the first element. The combination of several elements in the row; for any combination of elements, obtain the word vector of each word segment of the element belonging to the disease information data matrix in the combination of elements, and the cosine similarity between the word vector of each word segment of the element belonging to the drug information data matrix, and take the maximum cosine similarity as the similarity of the combination of elements.

7. The method for recommending research and statistical analysis schemes based on artificial intelligence according to claim 1, characterized in that, The specific method for constructing a neural network based on the disease information data matrix and the drug information data matrix is ​​as follows: The disease information data matrix is ​​used as the training dataset, with each row of the disease information data matrix as the input and the corresponding row of the drug information data matrix as the output. A neural network is constructed, and the cross-entropy function is used as the loss function.

Images