Gynecological and obstetric information processing method and system based on Internet

Abstract

The invention relates to the technical field of information processing, in particular to a gynaecology and obstetrics information processing method and system based on the Internet, and the method comprises the following steps: collecting user browsing records and interaction records based on user behavior data, carrying out the statistical analysis of user behavior characteristics, and generating a user preference analysis result; and based on the user preference analysis result, identifying the health management theme of the obstetrics and gynecology department through deep learning. By analyzing the user behavior data, the service individuation degree and the content correlation are improved. And the gynaecology and obstetrics health management theme identified through deep learning directly reflects user demands. Group analysis is carried out by using similar user data, and content recommendation is carried out by using a collaborative filtering algorithm, so that the recommendation accuracy and individuation level are improved. In combination with personalized suggestions generated by natural language processing, the user participation degree is further improved, and the optimized content is fed back and circulated through the user, so that timely updating and deep personalization of the education content are guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of information processing, and particularly to an Internet-based obstetrics and gynecology information processing method and system. Background Art

[0002] The technical field of information processing mainly involves data collection, storage, processing, analysis, transmission, and display. This field includes multiple steps of data processing, aiming to improve the availability and comprehensibility of information by using various algorithms and computational methods. The obstetrics and gynecology information processing method can provide personalized health management suggestions and educational content for patients.

[0003] However, the existing technologies generally adopt static data sets, which limit the speed of service updates and the depth of personalization, and fail to fully reflect the immediate changes and emerging needs of users. In addition, the existing technologies are also insufficient in the in-depth analysis of user behavior, affecting the accuracy of content recommendation and personalized services, which may lead to a decline in user experience and a lag in information services. This is particularly obvious in highly specialized medical and health fields such as obstetrics and gynecology, affecting the acceptance of health management content by patients. Summary of the Invention

[0004] The purpose of the present invention is to solve the disadvantages existing in the prior art, and to propose an Internet-based obstetrics and gynecology information processing method and system.

[0005] To achieve the above purpose, the present invention adopts the following technical solutions. The Internet-based obstetrics and gynecology information processing method includes the following steps:

[0006] Based on user behavior data, collect user browsing records and interaction records, statistically analyze user behavior characteristics, and generate user preference analysis results; based on the user preference analysis results, identify obstetrics and gynecology health management topics through deep learning to obtain obstetrics and gynecology health topic preference results;

[0007] Based on the obstetrics and gynecology health topic preference results, collect data of similar users, conduct user group analysis, identify similar user groups, and generate similar user group results; use the similar user group results to perform content recommendation through collaborative filtering algorithm to obtain a personalized content recommendation list;

[0008] Based on the personalized content recommendation list, collect associated data of educational content and management, combine natural language processing, generate personalized suggestions, and obtain preliminary personalized educational content; based on the preliminary personalized educational content, conduct user feedback analysis and adjustment to obtain customized educational content;

[0009] Based on the customized educational content, regularly update educational content and management suggestions to obtain user education results.

[0010] Preferably, the steps for obtaining the user preference analysis result are as follows:

[0011] Extract the page click sequence, access depth, stay time, and return visit frequency according to the user's browsing record and interaction record, and generate a behavior feature matrix;

[0012] Based on the behavior feature matrix, calculate the user's behavior feature score, and the calculation formula is:

[0013]

[0014] where F(u) is the behavior feature score of user u, x i is the click-through rate of the user on behavior i, y i is the access depth of behavior i, z is the total number of user behavior categories, w is the user's behavior volatility index, and n represents the total number of behavior feature parameters;

[0015] Based on the behavior feature score, perform classification and aggregation analysis on the behavior features, and combine the access frequency and the overall distribution of the behavior feature scores to generate the user preference analysis result.

[0016] Preferably, the steps for obtaining the preference result of the obstetrics and gynecology health theme are as follows:

[0017] Based on the user preference analysis result, deploy a deep learning model, perform model training, and obtain the model training result;

[0018] Based on the model training result, apply the deep learning model to identify the obstetrics and gynecology health management theme, and obtain the preliminary obstetrics and gynecology health theme preference data;

[0019] Based on the preliminary obstetrics and gynecology health theme preference data, perform data refinement and analysis, and perform preference operations to obtain the obstetrics and gynecology health theme preference result.

[0020] Preferably, the steps for obtaining the similar user group result are as follows:

[0021] According to the preference result of the obstetrics and gynecology health theme, calculate the feature mapping similarity between users, and the expression is:

[0022]

[0023] where R ij is the feature mapping similarity between user i and user j, f ik is the interaction frequency of user i on feature k, g jk is the theme distribution density of user j on feature k, h ik is the normalized value of the browsing depth of user i on feature k, t jkThe standardized value of the residence time of user j on feature k, where m represents the total number of feature dimensions;

[0024] Based on the feature mapping similarity, use the hierarchical clustering method to group user features, extract the aggregation results of similar user features, and generate the results of similar user groups.

[0025] Preferably, the steps for obtaining the personalized content recommendation list are as follows:

[0026] Based on the results of the similar user groups, calculate the content recommendation interest score, and the expression is:

[0027]

[0028] Among them, P ij is the interest score of user i in content j, B ik is the content matching degree of user i on feature k, C jk is the performance value of content j on feature k, D ik is the time series value of the click behavior feature of user i, E jk is the interaction mode feature value of content j, F ij is the basic interaction frequency between user i and content j, and q is the total dimension of preference features;

[0029] Based on the content recommendation interest score, sort the interest priorities of the recommended content, screen and combine the content to obtain the personalized content recommendation list.

[0030] Preferably, the steps for obtaining the preliminary personalized education content are as follows:

[0031] According to the personalized content recommendation list, extract the education content and management data associated with the user's needs, screen the features of the education content and management data, and establish a content association table to generate the preliminary matching results of the education content and management data;

[0032] Based on the preliminary matching results of the education content and management data, combine natural language processing to analyze the semantic structure of the education content and the logical features of the management data, extract the association relationship between the semantics and the logic, and generate the preliminary personalized recommendation data;

[0033] Based on the preliminary personalized recommendation data, combine the user's interest points to improve the personalized content and obtain the preliminary personalized education content.

[0034] Preferably, the steps for obtaining the customized education content are as follows:

[0035] Based on the preliminary personalized education content, collect user feedback data, including the user's acceptance of the content, the reading completion rate, and the modification suggestions put forward by the user, and generate the user feedback analysis results;

[0036] Based on the analysis results of the user feedback, adjust the structure and expression of the educational content to generate customized educational content.

[0037] Preferably, the steps for obtaining the user education result are as follows:

[0038] Based on the customized educational content, monitor the interaction of users in the educational content, collect user satisfaction data, and generate an evaluation result of the educational content;

[0039] According to the evaluation result of the educational content, identify the educational content that needs to be updated or improved, and combine with the current educational theory and practical achievements to update the educational content and generate a draft of the updated educational content;

[0040] Based on the draft of the updated educational content, conduct user testing and feedback loops.

[0041] The present invention provides an information processing system, including:

[0042] A user data collection module, which obtains the browsing records and interaction data of users from websites and applications, conducts item statistics on the data, and generates an analysis of user behavior characteristics;

[0043] A theme recognition and analysis module, which uses the data in the analysis of user behavior characteristics, processes the features through deep learning, identifies keywords and classifies them into health management themes related to obstetrics and gynecology, and outputs the theme preference results of obstetrics and gynecology;

[0044] A group user analysis module, based on the theme preference results of obstetrics and gynecology, collects user behavior data, conducts pattern matching and clustering analysis, and obtains the characteristics of similar user groups;

[0045] A content recommendation module, which applies the characteristics of similar user groups, selects and ranks relevant educational content and management suggestions through collaborative filtering of user behavior and preferences, and generates a personalized content recommendation list;

[0046] A customized educational content generation module, according to the personalized content recommendation list, combines user feedback, and uses natural language processing to adjust the educational content to generate and update customized educational content.

[0047] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0048] Through the analysis of user behavior data, the present invention improves the degree of service personalization and content relevance. The theme of obstetrics and gynecology health management identified through deep learning directly reflects user needs. By using the data of similar users for group analysis and the collaborative filtering algorithm for content recommendation, the accuracy and personalization level of the recommendation are improved. Combining the personalized suggestions generated by natural language processing further improves user engagement. By optimizing the content through the user feedback loop, the timely update and deep personalization of educational content are ensured, not only improving the processing efficiency, but also achieving the precise management and dynamic update of obstetrics and gynecology information, enhancing the user experience and satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a schematic diagram of the steps of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0050] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0051] Please refer to Figure 1 , the present invention provides a technical solution, an Internet-based method for processing obstetrics and gynecology information, including the following steps:

[0052] Based on user behavior data, collect user browsing records and interaction records, statistically analyze user behavior characteristics, and generate user preference analysis results; based on the user preference analysis results, identify the theme of obstetrics and gynecology health management through deep learning to obtain the obstetrics and gynecology health theme preference results;

[0053] Based on the obstetrics and gynecology health theme preference results, collect the data of similar users, conduct user group analysis, identify similar user groups, and generate similar user group results; use the similar user group results to perform content recommendation through the collaborative filtering algorithm to obtain a personalized content recommendation list;

[0054] Based on the personalized content recommendation list, collect the associated data of educational content and management, combine natural language processing, generate personalized suggestions, and obtain preliminary personalized educational content; based on the preliminary personalized educational content, conduct user feedback analysis and adjustment to obtain customized educational content;

[0055] Based on the customized educational content, regularly update the educational content and management suggestions to obtain user education results.

[0056] The steps for obtaining the user preference analysis results are as follows:

[0057] Extract the page click sequence, access depth, stay time, and return visit frequency based on the user's browsing history and interaction records, and generate a behavior feature matrix;

[0058] Based on the behavior feature matrix, calculate the behavior feature score of the user. The calculation formula is:

[0059]

[0060] Among them, F(u) is the behavior feature score of user u, x i is the click-through rate of the user on behavior i, y i is the access depth of behavior i, z is the total number of user behavior categories, w is the behavior volatility index of the user, and n represents the total number of behavior feature parameters;

[0061] Based on the behavior feature score, conduct a classification and aggregation analysis of the behavior features, and combine the overall distribution of the access frequency and the behavior feature score to generate the user preference analysis result.

[0062] Specifically, according to the page click sequence, access depth, stay time, and return visit frequency information included in the user's browsing history and interaction records, number these records in chronological order, read the page access identifier, access duration, and access level and other elements in each record, and refer to the average stay duration interval obtained from the actual website traffic monitoring in advance, such as 30 seconds to 600 seconds, and mark the abnormal stay records below 30 seconds or exceeding 600 seconds as outliers for subsequent differentiation in statistics. Subsequently, map the page access level value to the level mapping list of the page in the website structure. The mapping list is sorted out from the actual level situation of the website structure, so as to obtain the specific depth value of each access. At the same time, record the time interval of each return visit, and register the return visit times outside less than 1 hour and more than 24 hours in the additional table of the return visit index. The return visit index is quantified according to the user's daily repeated visit situation to ensure that the user behavior can be accurately characterized according to each access time interval and access path. Merge all the processed page click sequences, access depths, stay durations, and return visit frequency data into a unified matrix, and fill them into the corresponding rows and columns in the order of access time. Finally, obtain the behavior feature matrix.

[0063] The advantage of the formula is that by simultaneously considering two types of indicators, namely the click-through rate and the access depth, and combining the denominator balance of the number of behavior categories and the volatility index, it can quantify the user behavior characteristics in a more intuitive way, so as to more accurately distinguish the interest preferences of different users in subsequent analysis; Next, assign definite values to all the parameters in the formula and conduct actual example calculations. First, count three groups of behavior feature parameters in the user's browsing history, let n = 3, let x 1 = 0.32, y 1 = 2, x2 = 0.47, y 2 = 3, x 3 = 0.58, y 3 = 1. These six values are all obtained after normalizing a large amount of log data. Then, by observing the distribution of users on different types of pages on-site, the total number of behavior categories is set to z = 5, and the volatility index w = -2 is obtained through the change range of the user's access behavior at different times over 30 consecutive days. When substituting into the formula for calculation, first calculate the numerator:

[0064] (0.32) 2 + 2 2 = 0.1024 + 4 = 4.1024

[0065] (0.47) 2 + 3 2 = 0.2209 + 9 = 9.2209

[0066] (0.58) 2 + 1 2 = 0.3364 + 1 = 1.3364

[0067]

[0068] Then calculate the denominator:

[0069] z + |w| = 5 + |-2| = 5 + 2 = 7

[0070] Finally, we get:

[0071]

[0072] This result shows that the user behavior characteristic score is approximately 1.447. A relatively high value often means that the user is relatively active in terms of click-through rate and access depth and has a relatively large number of behavior categories. The absolute value of the volatility index regulates it in the denominator. Therefore, the numerical result obtained in this step can provide a basis for subsequent user type classification.

[0073] In the formula, x i represents the click-through rate value of the user on behavior i, which is obtained from the ratio of the total number of clicks on a specific page or function button to the total number of views by the user within a certain period of time. By recording user interaction logs item by item and classifying and counting the total number of clicks and the total number of displays for each behavior item, and then taking the ratio of the total number of clicks to the total number of displays, x can be obtained i ;

[0074] y iDenote the access depth on behavior i, which is usually quantified by using the hierarchical values that can be clearly counted in the website structure. For example, consider the homepage as depth 1, the page after an inward jump as depth 2, and so on. To obtain this parameter, it is necessary to segment and track the access path and record the occurrences of pages at different depths during multiple visits. After summarization, organize the distribution data of the user's stay on each hierarchical page;

[0075] z is the total number of behavior categories of the user, which is determined according to the number of different types of pages clicked or accessed by the user within a certain period. For this purpose, it is necessary to organize all page categories within the statistical period, mark the categories with user interactions as used, and then calculate the total number of categories and assign it to z;

[0076] w is the behavior volatility index of this user, which depends on the amplitude of the change in the access frequency or interaction intensity of the user over consecutive days. It can be obtained by calculating the variance or standard deviation of the daily access times. Compare the obtained standard deviation with a certain normalization benchmark to represent the degree of volatility in a numerical way, and finally use it as the value of w;

[0077] Based on the behavior feature scores, conduct a classification and aggregation analysis of the behavior features. Combine the overall distribution of the access frequency and behavior feature scores. First, select the behavior feature scores obtained previously as the input parameters for the aggregation analysis, and determine the differences among users in terms of browsing depth, click activity, return visit frequency, etc. by comparing the distributions of the scores in different threshold intervals. Divide users with scores less than 1.0 into the low-activity group, whose corresponding access frequency is usually less than 1 time per day and the page stay time is concentrated in the interval of 30 seconds to 60 seconds. Classify users with scores between 1.0 and 2.0 into the medium-activity group. The access frequency of this group is mostly between 1 and 3 times per day and the page stay time can reach 60 seconds to more than 180 seconds. Define users with scores greater than 2.0 as the high-activity group. The access frequency of this group can often reach more than 3 times per day and the page stay time is concentrated in the range of 180 seconds to 600 seconds. Then, according to the performance of each group in multi-dimensional data such as stay duration, browsing depth, and return visit interval, conduct a secondary combined statistics according to the high and low access frequencies and the differences in feature scores. During the statistical process, according to the pre-set hierarchical thresholds, for example, define the interval from layer 1 to layer 5 of the access depth as the low-depth range, from layer 6 to layer 10 as the medium-depth range, and above layer 10 as the high-depth range, and combine the situations such as the stay duration being less than 30 seconds and more than 600 seconds for division. Through the above refined analysis process, finally obtain the user preference analysis result.

[0078] The steps to obtain the preference result of the obstetrics and gynecology health theme are as follows:

[0079] Based on the user preference analysis result, deploy a deep learning model, perform model training, and obtain the model training result;

[0080] Based on the model training results, apply the deep learning model to identify the themes of obstetrics and gynecology health management, and obtain the preliminary data of obstetrics and gynecology health theme preferences;

[0081] Based on the preliminary data of obstetrics and gynecology health theme preferences, conduct data refinement and analysis, perform preference operations, and obtain the results of obstetrics and gynecology health theme preferences.

[0082] Specifically, based on the user preference analysis results, by extracting core parameters such as the user's click behavior, page access level, and interaction frequency, and organizing these parameters into input feature vectors suitable for deep learning, after setting the mapping relationship between the feature vector dimension and the sample label, first configure the initial structure of the model, including defining the number of network layers and the number of neurons in each layer, etc. Then, identify the records above a certain set frequency and the records below this frequency from the previously obtained user click behavior distribution, distinguish their numerical ranges at different access depths, and divide all samples into a training set and a validation set. The proportion of the training set is usually between 60% and 80%, and the proportion of the validation set is the remaining part. Input the samples in the training set into the deep learning model in sequence and perform forward propagation and backward propagation. In forward propagation, calculate the weighted sum of each neuron layer by layer and generate the output result. In backward propagation, update the neuron weights according to the difference between the output result and the actual label. Converge the parameters of each layer through continuous multiple rounds of iteration. If the error of the validation set is outside a certain preset range, the learning rate needs to be numerically adjusted in the next round of iteration, either increased to improve the convergence speed or decreased to maintain stable changes. These learning rate values are selected from a more appropriate range based on the records of multiple historical trainings. Then, gradually compare the accuracy results presented on the validation set in each round of iteration. When the accuracy reaches a certain set threshold for several consecutive rounds, the model can be regarded as stable. This threshold is determined by recording the convergence performance of multiple iterations in the previously collected training logs and is set to an accuracy range higher than 80% or 90%. Finally, output the converged network weights and fix them in the parameter data required for subsequent inference processes to obtain the model training results.

[0083] Based on the model training results, read the fixed network weights and network layer configurations after training is completed. Use the previously generated feature vector layout to integrate new test data. Match the click time interval data, page level distribution data, interaction duration distribution data, etc. from the user preference analysis results with the model training results, and then input them into the deep learning model item by item to perform forward calculations. Record the output values of each neuron and the final output of the network. In the output stage, map the key topic tags, diagnosis and treatment process tags, auxiliary content tags, etc. related to obstetrics and gynecology health management to the output results. Perform a threshold comparison on the output values that meet the predefined tags. Consider the output values greater than or equal to the threshold as important topic information recognized by the network. The threshold is set by examining the balance relationship between the accuracy rate and false alarm rate of topic recognition on the validation set during the previous training. For example, if statistical analysis shows that the optimal performance is in the range of 0.5 to 0.7, then 0.6 can be selected as the threshold. Filter out the topic information below this threshold, and then summarize the topic tags that pass the threshold in combination with the associated sub-tags to obtain preliminary obstetrics and gynecology health topic preference data.

[0084] When refining and analyzing the data based on the preliminary obstetrics and gynecology health topic preference data, it is necessary to perform frequency statistics on the identified topic tags and sub-tags. Concentrate the user click times, access durations, number of pages stayed, etc. corresponding to these tags into a topic feature set. Then, compare the user engagement indicators corresponding to each topic tag item by item within this set. For example, if the user engagement indicator of a certain topic tag is greater than a certain set range, it indicates that it appears more frequently during the research period. This range is calculated by referring to the records of previous users' attention to common health topics. Consider the range of 20 to 40 times of topic occurrences as the moderate range, and more than 40 times as the high-intensity range. Then, perform grouped sorting on the topic tags in the moderate range and high-intensity range respectively. Through preference operations on the grouped and sorted tags, including operations such as eliminating topics with low user access depth or further disassembling and subdividing high-engagement topics, finally, count the distribution of each tag and output the corresponding tag set information to obtain the obstetrics and gynecology health topic preference results.

[0085] The steps to obtain the results of similar user groups are as follows:

[0086] According to the obstetrics and gynecology health topic preference results, calculate the feature mapping similarity between users. The expression is:

[0087]

[0088] Among them, R ij is the feature mapping similarity between user i and user j, f ik is the interaction frequency of user i on feature k, g jkis the topic distribution density of user j on feature k, h ik is the standardized value of user i's browsing depth on feature k, t jk is the standardized value of user j's stay time on feature k, where m represents the total number of feature dimensions;

[0089] Based on the feature mapping similarity, use the hierarchical clustering method to group user features, extract the aggregation results of similar user features, and generate the results of similar user groups.

[0090] Specifically, the advantage of the formula is that by squaring the interaction frequency and the topic distribution density respectively, taking the absolute value of the difference, and then combining the standardized values of the browsing depth and the stay time to participate in the denominator operation together, the similarity calculation between users is more comprehensive;

[0091] f ik The steps to obtain the parameter are as follows: Record the interaction behavior frequency of user i under feature k through the interaction data obtained previously, and obtain the specific value after normalization processing. For example, count the cumulative click times of user i under a certain obstetrics and gynecology related topic within 30 days, and calculate the distribution frequency by comparing with the click times under all similar topics;

[0092] g jk The steps to obtain the parameter are as follows: Record the topic distribution of user j on feature k. By normalizing the access quantity distribution generated by user j on the same category of topics, the distribution density of this topic can be obtained;

[0093] h ik The steps to obtain the parameter are as follows: Statistically calculate the average browsing depth of user i on the page corresponding to feature k. Calculate according to the actual structure level of the website and the number of jump paths of the user entering the page. Record the deepest level of each access, take the average for multiple accesses, and then perform interval normalization to obtain the value;

[0094] t jk The steps to obtain the parameter are as follows: Statistically calculate the stay time of user j on feature k. Accumulate the duration of each access and divide the statistical interval. Continuously record the cumulative duration to obtain the average stay seconds of this user under this topic, and perform normalization;

[0095] m represents the total number of feature dimensions, which can be obtained by counting the obstetrics and gynecology health topic feature dimensions screened previously. If there are 5 topic categories, then m = 5; if there are 10, then m = 10;

[0096] Calculation process:

[0097] Let m = 3, and let the parameter values of user i and user j under features 1, 2, and 3 be: f i1 = 3.2, fi2 = 5.1, f i3 = 2.7, g j1 = 4.2, g j2 = 5.3, g j3 = 2.4, h i1 = 0.6, h i2 = 0.4,

[0098] h i3 = 0.7, t j1 = 0.5, t j2 = 0.3, t j3 = 0.8;

[0099] Calculate the numerator:

[0100]

[0101] Calculate the denominator:

[0102]

[0103] Obtained:

[0104]

[0105] This result indicates that the feature mapping similarity value between user i and user j is approximately 2.36. When the value is relatively large, it means that the difference in the interaction frequency and theme distribution between the two in the considered features is relatively significant. If this result is compared with the values between other subsequent users, the user groups that are closest or have the greatest gap can be more intuitively located.

[0106] Based on the feature mapping similarity, user features are integrated into an input set for hierarchical clustering. First, the numerical values of the feature mapping similarity of all users are disassembled and extracted. The interaction frequency and browsing depth generated by each user in each topic dimension are used as basic information, and the standardized results of the residence time corresponding to each user are listed item by item, thus forming a matrix covering the mapping relationship between all users and topics. Then, the user numbers are corresponding to the rows of this matrix, and the topics and related indicators are corresponding to the columns. The feature mapping similarity between users is compared with this matrix. During the comparison, users are classified according to a similarity segmentation rule obtained through previous statistics. For example, the range with a similarity less than 1.0 can be regarded as the high similarity interval because it is found in the statistics that the interaction differences presented by users with these numerical values in the topic dimension are generally low. For the user group with a similarity between 1.0 and 2.0, it is regarded as the medium similarity group, and more indicator parameters are compared to further distinguish the differences. The group with a similarity exceeding 2.0 is classified into the low similarity interval. At this time, it is necessary to further count whether these users frequently have high interaction records concentrated on a certain topic, and check these statistical results with the access time period of the users and mark the possible peak access time periods or access page levels. Then, the user numbers within the same interval are hierarchically merged, and the user sets obtained after each aggregation and the list of numbers they contain are recorded. If the number of users within a certain interval after aggregation is still greater than a certain quantity threshold, the similarity matrix is called again for secondary comparison. Referring to the same segmentation rule, the similarity values are arranged and located in the matrix according to the previous statistical results. Finally, multi-level aggregation is completed and the aggregated results of the same type of user features are output to obtain the results of similar user groups.

[0107] The steps for obtaining the personalized content recommendation list are as follows:

[0108] Based on the results of similar user groups, calculate the content recommendation interest score, and the expression is:

[0109]

[0110] Among them, P ij is the interest score of user i in content j, B ik is the content matching degree of user i on feature k, C jk is the performance value of content j on feature k, D ik is the time series value of the click behavior feature of user i, E jk is the interaction mode feature value of content j, F ij is the basic interaction frequency between user i and content j, and q is the total dimension of preference features;

[0111] Based on the content recommendation interest score, sort the interest priorities of the recommended content, screen and combine the content to obtain the personalized content recommendation list.

[0112] Specifically, the advantage of the formula lies in comprehensively considering the content matching degree of the user on the feature, the performance value of the content itself under the corresponding feature, and the difference between the time series value of the user's click behavior feature and the content interaction mode feature value, and adding the basic interaction frequency between the user and the content to the denominator for adjustment, making the calculated interest score more balanced and detailed;

[0113] B ik The steps to obtain parameter B are as follows: Based on the matching situation of user i with the content corresponding to feature k obtained previously, compare the preference information shown by the user under feature k with the attribute dimension of the content here, then count the number of interactions or access depth of the user with similar content within this dimension, and compare it with the overall average of all relevant content. After normalization calculation, B is obtained. ik ;

[0114] C jk The steps to obtain parameter C are as follows: By monitoring the specific presentation form of content j on feature k, quantify its comprehensive performance in aspects such as visualization, text information, and function modules, record the scores of multiple dimensions and then perform weighted averaging, and finally normalize it within a specified range to obtain C. jk ;

[0115] D ik The steps to obtain parameter D are as follows: Conduct time series analysis on the click behavior of user i under feature k. First, record the time and corresponding object of each click, collect the total number of these clicks at fixed intervals daily or weekly, and organize them into a click quantity sequence for different time periods. Then perform smoothing or differencing operations on this sequence, and finally perform normalization processing within a reasonable time range to obtain D. ik ;

[0116] E jk The steps to obtain parameter E are as follows: For the interaction mode feature of content j, collect the average viewing or operation duration and the number of interaction links of this content under feature k, calculate the interaction complexity in combination with click path and dwell depth information, and then compare and evaluate it with similar content to finally obtain E. jk ;

[0117] F ij The steps to obtain parameter F are as follows: Directly count the basic interaction frequency from the interaction records between user i and content j, including the total number of clicks, the number of comments, the number of favorites, etc. of the user for this content, which can be obtained by summarizing each entry in all logs within a period of time;

[0118] q is the total dimension of the preference feature, which depends on the number of user preference classification categories or the number of content attribute divisions set in advance, and is obtained by counting the total number of theme or attribute categories;

[0119] Calculation process:

[0120] Let q = 3; take the content matching degrees B of user i under features 1, 2, and 3 i1 = 0.65, B i2 =

[0121] 0.48, B i3 = 0.82, take the performance values C of content j under features 1, 2, and 3 j1 = 0.54, C j2 = 0.70, C j3 = 0.60, take the click behavior feature time series values D i1 = 2.3, D i2 = 1.8, D i3 = 3.1, interaction mode feature values E j1 = 2.1, E j2 = 2.6, E j3 = 2.4, the basic interaction frequency F between user i and content j ij = 12;

[0122] Calculate the sum of squares in the numerator:

[0123]

[0124] Calculate the differences in the numerator:

[0125] |D ik -E jk | = |2.3 - 2.1| + |1.8 - 2.6| + |3.1 - 2.4|

[0126] Here, each item needs to be calculated separately and then accumulated. However, according to the above definition, these three groups of differences need to be unified in a certain summation method or averaged first, which can be determined as needed specifically. Here, each item is demonstrated:

[0127] |2.3 - 2.1| = 0.2, |1.8 - 2.6| = 0.8, |3.1 - 2.4| = 0.7

[0128] 0.2 + 0.8 + 0.7 = 1.7

[0129] Combine the numerator:

[0130]

[0131] Denominator part:

[0132] F ij + 1 = 12 + 1 = 13

[0133] Comprehensively obtained:

[0134]

[0135] The result shows that the interest score of user i in content j is approximately 0.4094. When this value is close to or greater than 1.0, it can generally be regarded as a relatively significant interest correlation. When it is less than 1.0, it means that the user's fit with this content is not high under the current comprehensive dimension, and more dynamic behavior data can be combined to further segment or explore user preferences.

[0136] Based on the content recommendation interest scores, first read the above calculation results and summarize the interest scores of the user in all recommended contents. Then, compare the content numbers with the corresponding scores item by item. During the comparison process, list the score values of all contents of this user and compare the values with multiple thresholds defined during the statistics. For example, if it exceeds 0.8, it is put into the priority push group; if it is between 0.5 and 0.8, it is included in the secondary push group; if it is less than 0.5, it is regarded as the reserve push group. These thresholds are determined by sorting through previous recommendation result records and analyzing the click times and browsing stay times in user usage feedback. Among the previous records, it was found that when the score is higher than 0.8, the user click-through rate is in the range of 30% to 50% and the stay time is greater than 180 seconds. When the score is between 0.5 and 0.8, the user's click-through rate is only in the range of 10% to 30% and the stay time is generally between 60 seconds and 180 seconds. When the score is below 0.5, the click-through rate is often lower than 10% and the stay time is usually less than 60 seconds. Match each recorded score according to the above grouping criteria to obtain the specific lists of contents in the priority push group, secondary push group, and reserve push group. Then, re-sort the priority push group and secondary push group in descending order, and check whether there are duplicate attributes or overly similar contents among the first several items in the sorting. Mark these duplicate items and do not list them as the current recommendation for the time being. Output the sorted list to the customized recommendation function for combination, and determine the order of presentation in sequence according to the aforementioned score values and grouping results. Finally, obtain the personalized content recommendation list.

[0137] The steps for obtaining preliminary personalized education content are as follows:

[0138] According to the personalized content recommendation list, extract the education content and management data related to the user's needs, screen the features of the education content and management data, and establish a content association table to generate a preliminary matching result of the education content and management data;

[0139] Based on the preliminary matching result of the education content and management data, combine natural language processing to analyze the semantic structure of the education content and the logical features of the management data, extract the association relationship between semantics and logic, and generate preliminary personalized recommendation data;

[0140] Based on the preliminary personalized recommendation data, the personalized content is improved by combining with the user's interest points to obtain the preliminary personalized education content.

[0141] Specifically, according to the personalized content recommendation list obtained previously, for each piece of recommendation information in the list, the corresponding user demand identification information and the education content theme with a relatively high matching degree are extracted and disassembled, and the records that are positively or frequently associated with these education themes are selected from the management data. The sources and time spans of these records are counted, and the occurrence frequencies of the keywords involved in their content are compared one by one with the concerns corresponding to the user demand identification. The click times and browsing durations of all themes are listed and compared with a pre-set effective range. For example, the average browsing duration is compared with the interval of 20 seconds to 300 seconds. If it is less than 20 seconds or more than 300 seconds, it is marked in the additional table and associated with the corresponding user demand identification for verification. The field names of similar medical records or health data in the management data are matched with the user demand identification, the characteristic fields of both sides are listed and their similarity is numerically processed. The records with a similarity value exceeding a certain preset threshold are marked as highly associated records. This threshold is determined by statistically analyzing large sample data and refining features in the medical knowledge base collected in the early stage. Subsequently, the matching degree values of each field and the label information of the education theme are used to generate a content association table, and the possible duplicate fields or conflicting names in the table are checked item by item. After excluding duplicates or conflicts, the association relationship data is summarized to generate the preliminary matching result of the education content and the management data.

[0142] Based on the preliminary matching result of the education content and the management data, the text information and logical fields of the education content and the management data in the matching result are parsed item by item. First, the text paragraphs of the education content are recorded in a list that can perform semantic structure decomposition, and the fields in the corresponding management data with a relatively high matching degree to its text paragraphs are listed. By performing clause processing on each text paragraph, the concept terms and digital quantification information appearing in each clause are counted. At the same time, referring to the logical fields in the management data, such as security level, access information, or user medical process, etc., the statuses in multiple characteristic dimensions are corresponded one by one with the concept terms in the text paragraph, and the coincidence degree between the two in terms of noun explanation or numerical interval is checked. The records with a coincidence degree exceeding a certain range are marked. This range can be determined according to the comparison table of medical terms and medical processes in the previous analysis. If the value is in the interval of 0 to 0.7, it is regarded as a low coincidence degree, between 0.7 and 0.9 as a medium coincidence degree, and exceeding 0.9 as a high coincidence degree. Subsequently, the clauses with a high coincidence degree are connected with the corresponding management data fields to form a semantic and logical association relationship. During this process, the action sequences or precautions required in the medical process are numbered, and the coupling situation between these numbers and the text clauses is recorded. Finally, all the coupled data entries are output to generate the preliminary personalized recommendation data.

[0143] Based on the preliminary personalized recommendation data, map the points of interest with a relatively high degree of relevance to the user's needs to the corresponding text content and management data fields. Refer to the semantic and logical association relationships obtained in the previous stage, and strengthen the display of key phrases and quantifiable indicators that appear within the clauses. By setting a key marking rule for medical education content, emphasize the frequently occurring or relatively key medical terms, and further associate the inspection processes or management solutions that the user has paid more attention to in the previous feedback with the corresponding semantic paragraphs. Streamline the repeated or approximate descriptions within the associated text, hierarchically combine the topic tags with longer stay times and more click times in the user's behavior characteristics, list the main fields and numerical ranges under this topic, and supplement the data sources or professional explanations for reference when necessary. Subsequently, perform a sentence coherence detection on the processed text information above. If it is found that the word order is chaotic or the field call order is inconsistent, reallocate the sentence order within the text paragraph and ensure that it is consistent with the logical numbers defined previously. Then, summarize all the text paragraphs that echo the user's points of interest and the relevant management data information to obtain the preliminary personalized education content.

[0144] The steps to obtain customized education content are as follows:

[0145] Based on the preliminary personalized education content, collect user feedback data, including the user's acceptance of the content, reading completion rate, and modification suggestions put forward by the user, and generate the user feedback analysis results;

[0146] Combine the user feedback analysis results, adjust the structure and expression of the education content, and generate customized education content.

[0147] Specifically, based on the preliminary personalized education content, first collect the acceptance indicators and reading completion rate values related to this content from the user side, and extract the text modification suggestions put forward for content paragraphs or key points in the user interface. Divide these acceptance indicators into three categories: high, medium, and low, and record the reading time period and click interaction times of each user. Then set a cut-off range for the reading completion rate. For example, mark it when the reading completion rate is in the range of 60% to 80%. When it is lower than 60% or exceeds 80%, write corresponding notes in the log to distinguish different usage scenarios and facilitate subsequent comparison. In addition, compare the text modification suggestions submitted by the user with the previously recorded content paragraphs to check whether there are frequently occurring words or professional terms in the paragraphs, and count whether the occurrence times of these words exceed a certain predetermined standard. This standard is calculated based on the comprehensibility commonly seen by users in multiple investigations within a large range. For example, take the frequency at which a certain type of medical term is understandable to 90% of users as a threshold. If it is statistically found that the occurrence frequency of some words is not within this range or is difficult to understand when compared with the feedback from multiple people, then mark them additionally. Then, centrally summarize information such as the scrolling times, paragraph staying times, and the number of users who put forward modification suggestions shown by all readers during the reading process. Combine the statistical results of the acceptance indicators and the reading completion rate to form a feedback classification table. Conduct additional verification on samples that exceed the specified reading completion rate or are lower than a certain acceptance value, and cross-compare with other data to check whether there are paragraphs with redundant text or lack of logical order. List these paragraphs separately in the classification table. Finally, merge and count paragraphs of the same type or in similar fields in the classification table to obtain the user feedback analysis result.

[0148] Combined with the results of user feedback analysis, for paragraphs in the classification table with low reading completion rates and acceptance scores in a low range, first list the uncommon or professional vocabulary that may exist in the paragraph and compare them with the suggested entries submitted by users one by one. If it is confirmed that there are high-frequency overlapping entries, replace these entries with more easy-to-understand expressions, and rearrange the paragraphs to keep them in a more logical order. Then check whether there are paragraphs with many modification opinions in the acceptance records. If the modification opinions of a paragraph account for more than 30% of the total number of users who read the paragraph, the paragraph is marked as requiring deep optimization. Then, combined with the distribution of user reading completion rates, these paragraphs to be optimized are compared with the high-profile paragraphs in the original educational content. Cross-check the key topics and record whether the paragraph is consistent with the key topic in terms of text description, or whether there are repetitions or contradictions in the language style. If it is found that there is misunderstanding in the paragraph content that may lead to the concentrated opinions of some users, a more detailed explanation will be given here and integrated with the replacement terms selected earlier. Finally, the order of all paragraphs is sorted out in terms of structure, and the main title, subtitle and the sentences and content within the paragraph are connected in sequence to ensure that they match the reading process sequence most frequently clicked by the aforementioned users, and transition words or objective descriptions related to the medical care scenarios are added between sentences. Then the adjusted text content is uniformly numbered in the new version and included in the subsequent maintenance records to generate customized educational content.

[0149] The steps to obtain user education results are:

[0150] Based on customized educational content, monitor user interaction in educational content, collect user satisfaction data, and generate educational content evaluation results;

[0151] Based on the results of the education content evaluation, identify the education content that needs to be updated or improved, update the education content in combination with current education theory and practice results, and generate a draft of the updated education content;

[0152] Conduct user testing and feedback loops based on the updated draft educational content.

[0153] Specifically, based on the customized educational content, first number the user interaction records related to this educational content item by item, summarize information such as the browsing duration, total number of clicks, and number of comment entries in the records, and compare the core evaluation indicators mentioned therein with a satisfaction evaluation standard extracted from previous data. If the browsing duration and satisfaction score of certain entries exceed the preset thresholds, special marks are made in the checklist. These thresholds are summarized from the data obtained through multiple user tests in the preliminary research. For example, users with a browsing duration concentrated in the interval of 30 seconds to 120 seconds are regarded as having normal attention, those with a browsing duration exceeding 120 seconds and a satisfaction score exceeding 80 points are classified as having high satisfaction. Then, compare the number of comment entries with the preset range. If the number of comment entries is less than 1, it is recorded as lacking interaction data. If it is in the interval of 1 to 5, it is regarded as having a general interaction level. If it exceeds 5, it is classified into the multiple feedback interval. Integrate all these marked contents into the interaction detail list, then calculate the average satisfaction of all users with this educational content within a specific period and list the distribution map, such as whether there are significant differences between weekdays and weekends, and record whether some users have concentratedly mentioned difficulties in paragraph understanding or insufficient number of cases when submitting other subjective opinions or written descriptions. Then, based on these feedback characteristics, associate them with the previously obtained browsing duration, click frequency, and number of comments to form a set of hierarchical statistical summaries. Finally, merge this set of statistical summaries and match them with the reference medical education background materials to observe whether there are relatively concentrated opinions or higher satisfaction biases in certain chapters or specific topic parts. Finally, generate the evaluation results of the educational content based on this information.

[0154] Based on the results of the educational content evaluation, search for chapters or paragraphs that have been marked multiple times or have concentrated opinions in the previous statistical summary. For the issues that users are more concerned about, first list the keywords and sort out the supplementary information that can be found in the relevant literature or medical practice. Then divide the text organization of these chapters and the number of actual cases again. Split the places where the content is not detailed enough or has repeated descriptions. Combine the paragraphs with less readable but necessary information with more illustrative examples for statistics. Evaluate whether the paragraphs with higher difficulty need to be broken down into multiple sub-parts, and use evidence-based medicine cases or concise diagrams to assist. In this process, continue to compare the recommended structures given by current educational theories and practical results. For example, when medical theory When the knowledge exceeds a certain proportion, it is necessary to add explanation items for ordinary readers in the text, or when the information obtained from the field interview records shows that there are multiple cases that highly match the topic, attach relevant real-life situation descriptions after the paragraph, and then when introducing new content, confirm that it is consistent with the previous topic in terms of terminology and expression, and confirm whether the medical terms cited before and after have the same meaning or the same origin in the previously listed general terminology list to avoid multiple names referring to the same concept. After sorting out the text and diagrams of all newly inserted paragraphs, rearrange them in the corresponding chapter positions and attach simple numbers where necessary. After this round of updating, the new teaching structure is refined into a preliminary manuscript to obtain an updated draft of educational content.

[0155] Based on the updated draft of educational content, user tests were conducted on the reorganized content of each chapter, and reading time and feedback were collected during the tests. For each user test record, the specific browsing process and interactive actions were tracked in segments to check whether users focused on specific sections or wandered between multiple adjacent paragraphs multiple times in different time periods. All evaluations on readability and practicality were summarized and matched with the user background information marked when distributing the questionnaire. By comparing the differences in reading focus and feedback types among users with different education levels or different age groups, the parts that most need to be revised again were summarized. Then, representative improvement suggestions were identified in the questionnaire data, and the number of such suggestions was counted to see whether it had reached the set ratio. For example, if more than 3% of users have made similar suggestions in their comments, they will be included in the high-level optimization list separately, and then it will be checked whether there are repeated words or content paragraphs that are not consistent with information coherence. The text will be adjusted or more context will be added to reduce the possibility of reader misunderstanding. Finally, after completing this round of testing, a feedback loop will be formed and further sorted out to obtain the subsequent improvement direction and phased optimization focus, which will be summarized to the editorial team or medical experts for continuous improvement, completing this round of user testing and feedback loop.

[0156] The present invention provides an information processing system, comprising:

[0157] The user data collection module obtains the browsing records and interaction data of users from websites and applications, conducts item statistics on the data, and generates an analysis of user behavior characteristics;

[0158] The theme recognition and analysis module uses the data in the analysis of user behavior characteristics, processes the features through deep learning, identifies keywords and classifies them into health management themes related to obstetrics and gynecology, and outputs the results of obstetrics and gynecology theme preferences;

[0159] The group user analysis module collects user behavior data based on the results of obstetrics and gynecology theme preferences, conducts pattern matching and clustering analysis, and obtains the characteristics of similar user groups;

[0160] The content recommendation module applies the characteristics of similar user groups, selects and ranks relevant educational content and management suggestions through collaborative filtering of user behavior and preferences, and generates a personalized content recommendation list;

[0161] The customized educational content generation module adjusts the educational content using natural language processing according to the personalized content recommendation list and in combination with user feedback, and generates and updates the customized educational content.

[0162] The above are only the preferred embodiments of the present invention, and do not limit the present invention in other forms. Any person skilled in the relevant art may use the disclosed technical content to make changes or modifications into equivalent embodiments with equivalent changes and apply them to other fields. However, any simple modification, equivalent change, and modification made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention still fall within the protection scope of the technical solution of the present invention.

Claims

1. An Internet-based obstetrics and gynecology information processing method, characterized in that: The following steps are involved: Based on user behavior data, collect user browsing records and interaction records, statistically analyze user behavior characteristics, and generate user preference analysis results; Based on the user preference analysis results, the obstetrics and gynecology health management topic is identified through deep learning to obtain the obstetrics and gynecology health topic preference results; Based on the obstetrics and gynecology health topic preference results, collect data of similar users, conduct user group analysis, identify similar user groups, and generate similar user group results; use the similar user group results to recommend content through collaborative filtering algorithms to obtain a personalized content recommendation list; Based on the personalized content recommendation list, collect the associated data of educational content and management, combine with natural language processing, generate personalized suggestions, and obtain preliminary personalized educational content; Based on the preliminary personalized educational content, performing user feedback analysis and adjustment to obtain customized educational content; Based on the customized educational content, the educational content and management suggestions are updated regularly to obtain user education results.

2. The Internet-based obstetrics and gynecology information processing method according to claim 1, characterized in that: The steps for obtaining the user preference analysis result are: According to the user's browsing history and interaction history, the page click sequence, visit depth, stay time and return visit frequency are extracted to generate a behavior feature matrix; Based on the behavior feature matrix, the user's behavior feature score is calculated using the following formula: Among them, F(u) is the behavioral feature score of user u, x i is the click rate of the user on behavior i, y i is the access depth of behavior i, z is the total number of user behavior categories, w is the user's behavior volatility index, and n represents the total number of behavior feature parameters; Based on the behavior feature scores, the behavior features are classified and aggregated for analysis, and the user preference analysis results are generated by combining the overall distribution of the access frequency and the behavior feature scores.

3. The Internet-based obstetrics and gynecology information processing method according to claim 1, characterized in that: The steps for obtaining the obstetrics and gynecology health topic preference results are as follows: Based on the user preference analysis results, deploy a deep learning model, perform model training, and obtain a model training result; Based on the model training results, a deep learning model is used to identify obstetrics and gynecology health management topics to obtain preliminary obstetrics and gynecology health topic preference data; Based on the preliminary obstetrics and gynecology health topic preference data, data refinement and analysis are performed, and preference operations are performed to obtain obstetrics and gynecology health topic preference results.

4. The Internet-based obstetrics and gynecology information processing method according to claim 1, characterized in that: The steps for obtaining the similar user group results are: According to the obstetrics and gynecology health topic preference results, the feature mapping similarity between users is calculated, and the expression is: Among them, R ij is the feature mapping similarity between user i and user j, f ik is the interaction frequency of user i on feature k, g jk is the topic distribution density of user j on feature k, h ik is the normalized value of user i’s browsing depth on feature k, t jk is the normalized value of user j’s stay time on feature k, and m represents the total number of feature dimensions; Based on the feature mapping similarity, the user features are grouped using a hierarchical clustering method, the aggregation results of similar user features are extracted, and similar user group results are generated.

5. The Internet-based obstetrics and gynecology information processing method according to claim 1, characterized in that: The steps for obtaining the personalized content recommendation list are as follows: Based on the similar user group results, the content recommendation interest score is calculated, and the expression is: Among them, P ij is the interest score of user i for content j, B ik is the content matching degree of user i on feature k, C jk is the performance value of content j on feature k, D ik is the time series value of the click behavior characteristics of user i, E jk is the interaction mode feature value of content j, F ij is the basic interaction frequency between user i and content j, and q is the total dimension of preference features; Based on the content recommendation interest scores, the interest priorities of the recommended contents are sorted, and the contents are filtered and combined to obtain a personalized content recommendation list.

6. The Internet-based obstetrics and gynecology information processing method according to claim 1, characterized in that: The steps for obtaining the preliminary personalized education content are: Extracting educational content and management data associated with user needs based on the personalized content recommendation list, screening the characteristics of educational content and management data and establishing a content association table, and generating preliminary matching results between educational content and management data; Based on the preliminary matching results of the educational content and the management data, the semantic structure of the educational content and the logical characteristics of the management data are analyzed by natural language processing, the association relationship between semantics and logic is extracted, and preliminary personalized recommendation data is generated; Based on the preliminary personalized suggestion data, the personalized content is improved in combination with the user's interests to obtain preliminary personalized educational content.

7. The Internet-based obstetrics and gynecology information processing method according to claim 1, characterized in that: The steps for obtaining the customized educational content are: Based on the preliminary personalized education content, collect user feedback data, including user acceptance of the content, reading completion rate and modification suggestions proposed by users, and generate user feedback analysis results; In combination with the user feedback analysis results, the structure and expression of the educational content are adjusted to generate customized educational content.

8. The Internet-based obstetrics and gynecology information processing method according to claim 1, characterized in that: The steps for obtaining the user education results are: Based on the customized educational content, monitor the user interaction in the educational content, collect user satisfaction data, and generate educational content evaluation results; Based on the educational content evaluation results, identify educational content that needs to be updated or improved, update the educational content in combination with current educational theories and practical achievements, and generate a draft of updated educational content; Based on said updated draft educational content, user testing and feedback loops are conducted.

9. The information processing system of the Internet-based obstetrics and gynecology information processing method according to any one of claims 1 to 8, characterized in that: include: The user data collection module obtains users' browsing history and interaction data from websites and applications, performs item statistics on the data, and generates user behavior feature analysis; The topic identification and analysis module uses the data from the user behavior feature analysis to process features through deep learning, identify keywords and classify them into health management topics related to obstetrics and gynecology, and output obstetrics and gynecology topic preference results; The group user analysis module collects user behavior data based on the obstetrics and gynecology topic preference results, performs pattern matching and cluster analysis, and obtains similar user group characteristics; The content recommendation module applies similar user group characteristics, selects and sorts related educational content and management suggestions through collaborative filtering of user behaviors and preferences, and generates a personalized content recommendation list; The customized educational content generation module adjusts the educational content based on the personalized content recommendation list and user feedback, and uses natural language processing to generate and update customized educational content.

Images