A personalized evidence-based knowledge pushing method and system based on physician behavior portrait

Abstract

The present application relates to the technical field of computer, especially to a personalized evidence-based knowledge pushing method and system based on doctor behavior portrait, the method comprises the following steps: constructing a multi-dimensional doctor behavior portrait by continuously collecting the doctor's department attribute, historical query record, attention disease spectrum and browsing behavior data; when the doctor initiates a query, performing feature matching and correlation score calculation on the initial evidence candidate set based on the portrait; reordering the search results according to the score to generate a personalized evidence recommendation list that meets the doctor's specialty field characteristics and historical attention theme; at the same time, using time decay and interest weight dynamic updating strategy to make the portrait reflect the doctor's latest interest changes in real time. The present application solves the technical problem that the existing evidence-based knowledge retrieval ignores the deep shaping effect of user characteristics on retrieval intention, significantly improves the matching accuracy between evidence-based evidence and the doctor's actual professional needs.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a method and system for personalized evidence-based knowledge delivery based on physician behavioral profiles. Background Technology

[0002] In clinical practice and medical research, evidence-based medicine knowledge retrieval systems serve as crucial tools for physicians to obtain clinical evidence and are widely used to support treatment decisions. However, due to differences in department affiliation, professional field, treatment habits, and knowledge structure, different physicians exhibit significantly individualized needs for evidence-based evidence.

[0003] Common evidence-based knowledge retrieval methods primarily rely on keyword matching or semantic vector similarity calculation. They generate search results rankings by constructing inverted indexes or pre-trained language models to score the relevance of queries to candidate documents. However, these methods generally treat the retrieval process as an isolated match between queries and documents, neglecting the profound shaping effect of user characteristics on search intent.

[0004] Due to the lack of dynamic modeling and integration of physician behavioral profiles, departmental characteristics, and historical query preferences, existing methods struggle to reorder and adaptively match search results based on the individual characteristics of different physicians. This results in a mismatch between the returned evidence list and the physician's actual professional needs, failing to provide accurate and personalized evidence-based knowledge services. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for personalized evidence-based knowledge delivery based on physician behavioral profiles, in order to solve the technical problems mentioned above.

[0006] One aspect of the present invention provides a method for personalized evidence-based knowledge delivery based on physician behavioral profiles, comprising: Continuously collect doctors' department attribute information, historical query records, disease spectrum data, and browsing behavior data, and construct multi-dimensional doctor behavior profiles based on the data; When a doctor initiates a query request through the evidence-based knowledge retrieval system, the system obtains the search keywords of the query request and retrieves an initial set of candidate evidence that matches the search keywords from a preset evidence-based database. The candidate evidence in the initial evidence candidate set is matched with the multidimensional doctor behavior profile to obtain the correlation score between each candidate evidence and the doctor profile. Based on the relevance score, the initial evidence candidate set is reordered to generate a personalized evidence recommendation list that conforms to the characteristics of the doctor's specialty field and historical topics of interest, and the personalized evidence recommendation list is pushed to the doctor's terminal for display.

[0007] In some embodiments, the continuous collection of doctors' departmental attribute information, historical query records, disease spectrum data, and browsing behavior data, and the construction of a multi-dimensional doctor behavior profile based on the data, includes: Collect doctors' department identifiers, professional titles, and professional field tags as static profile features; Collect doctors’ search keywords, search timestamps, and search result click records submitted within a preset time period as dynamic behavioral characteristics; The disease types, research directions of interest, and frequently accessed evidence categories that doctors actively collect or mark in the evidence-based knowledge retrieval system are used as interest and preference features. The duration of doctors' browsing of evidence summaries, scrolling speed, and number of repeated visits were collected as behavioral pattern features. The static profile features, dynamic behavior features, interest preference features, and behavior pattern features are fused to obtain a fused multidimensional doctor behavior profile vector.

[0008] In some embodiments, the feature fusion processing of the static profile features, the dynamic behavior features, the interest preference features, and the behavior pattern features includes: Categorical features are quantified using one-hot encoding or label encoding. Numerical features are quantified using standardization or normalization methods; Textual features are semantically encoded using a pre-trained language model to obtain fixed-dimensional semantic vectors; The encoded features are weighted and summed or concatenated according to a preset fusion strategy to generate a fused multidimensional doctor behavior profile vector.

[0009] In some embodiments, the step of performing feature matching calculations between each candidate piece of evidence in the initial candidate evidence set and the multidimensional doctor behavior profile to obtain a correlation score between each candidate piece of evidence and the doctor profile includes: Extract the text content features, evidence source features, evidence quality rating features, and evidence publication time features of each candidate piece of evidence; The feature vectors of candidate evidence are compared with the multidimensional doctor behavior profile vectors to calculate the basic relevance score. Based on the professional field label of the doctor's current department, calculate the field matching degree between each candidate piece of evidence and the doctor's professional field; Based on the distribution of physicians' historical topics of interest, the thematic similarity between each candidate piece of evidence and the physicians' historical topics of interest is calculated. Based on the behavioral patterns of doctors, calculate the behavioral fit between each candidate piece of evidence and the doctors' browsing preferences. The basic relevance score, the domain matching degree, the topic similarity degree, and the behavior fit degree are weighted and fused to obtain the final relevance score.

[0010] In some embodiments, the vector similarity calculation employs a cosine similarity algorithm.

[0011] In some embodiments, the step of reordering the initial evidence candidate set based on the relevance score to generate a personalized evidence recommendation list that conforms to the characteristics of the physician's specialty field and historical topics of interest includes: The candidate evidence is sorted in descending order of relevance score to generate an initial ranking result. A diversity test is performed on the first preset number of candidate evidences in the initial sorting result. When the topic similarity of adjacent candidate evidences exceeds a preset threshold, the sorting position of subsequent candidate evidences is reduced. The quality of evidence is verified on the adjusted ranking results. When the quality rating of candidate evidence is lower than the minimum quality threshold, its ranking position is adjusted to below the preset lower limit. Output the final sorted list of personalized evidence recommendations.

[0012] In some embodiments, performing a diversity test on the first preset number of candidate pieces of evidence in the initial ranking result further includes: Calculate the topic similarity between adjacent candidate evidence; Consider the diversity of evidence sources and avoid making the evidence in the recommended list too concentrated in the same journal or database.

[0013] In some embodiments, the method further includes: dynamically updating the multidimensional doctor behavior profile; when a doctor generates new query behavior or browsing behavior, adjusting the weight values ​​of corresponding features in the doctor profile vector according to a preset profile update strategy.

[0014] In some embodiments, the feature vectors used in the feature matching calculation are obtained by semantically encoding the doctor's behavior data and evidence content through a pre-trained language model, which is fine-tuned based on a large-scale corpus in the medical field.

[0015] Another aspect of the present invention provides a personalized evidence-based knowledge delivery system based on physician behavioral profiles, comprising at least a profile construction module, an evidence retrieval module, a feature matching module, and a re-ranking module, wherein: The profile building module is used to continuously collect doctors' department attribute information, historical query records, disease spectrum data, and browsing behavior data, and build a multi-dimensional doctor behavior profile based on the above data. The evidence retrieval module is used to obtain the search keywords of the query request when a doctor initiates a query request through the evidence-based knowledge retrieval system, and retrieve an initial set of candidate evidence that matches the search keywords from a preset evidence-based evidence database. The feature matching module is used to perform feature matching calculations between each candidate piece of evidence in the initial evidence candidate set and the multidimensional doctor behavior profile to obtain a correlation score between each candidate piece of evidence and the doctor profile. The re-ranking module is used to re-rank the initial evidence candidate set based on the relevance score, generate a personalized evidence recommendation list that conforms to the characteristics of the doctor's specialty field and historical focus topics, and push the personalized evidence recommendation list to the doctor's terminal for display.

[0016] Compared with the prior art, the present invention has the following beneficial effects: By continuously analyzing doctors' departmental attributes, historical query records, disease spectrum of interest, and browsing time, a multi-dimensional doctor profile is constructed. This enables the technical effect of re-ranking results based on the profile on the basis of general search results, prioritizing the display of evidence that is strongly related to the doctor's specialty and similar to the topics of interest in the past. This addresses the technical problem of existing evidence-based knowledge retrieval methods neglecting the deep shaping effect of user characteristics on search intent, and overcomes the technical deficiency of traditional methods that cannot adaptively adapt search results according to the individual characteristics of different doctors. It significantly improves the matching accuracy between evidence-based evidence and doctors' actual professional needs, providing doctors with a precise and personalized evidence-based knowledge service experience. Attached Figure Description

[0017] 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.

[0018] Figure 1 This is a flowchart of the personalized evidence-based knowledge delivery method of the present invention; Figure 2 This is the architecture diagram of the personalized evidence-based knowledge delivery system of the present invention. Detailed Implementation

[0019] The following will be based on embodiments of the present invention. Figures 1-2 The technical solutions in the embodiments of the present invention will be clearly and completely described together. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0020] Example This embodiment describes the complete technical implementation process of a personalized evidence-based knowledge delivery method based on physician behavioral profiles in the DeepEvidence intelligent evidence-based decision-making system. It is worth noting that the method is executed based on a data acquisition layer, a profile construction layer, an evidence retrieval layer, a feature matching layer, a re-ranking layer, and a delivery display layer.

[0021] Specifically, the data acquisition layer is responsible for continuously acquiring physician behavior-related data from multiple data sources. This layer establishes data channels with the hospital information system, the DeepEvidence intelligent evidence-based decision-making system's client application, and the data platform to achieve real-time and batch collection of physician behavior data. The data types collected by the data acquisition layer include four main categories: department attribute information, historical query records, disease spectrum data, and browsing behavior data. Department attribute information covers organizational structure information such as the physician's department identification code, department name, and department category, as well as personal attribute information such as the physician's professional title, qualifications, and professional field tags. Historical query records cover all search keywords submitted by the physician through the DeepEvidence intelligent evidence-based decision-making system within a preset time period, timestamps of each search, records of clicks on search results after each search, and records of time spent on the evidence details page after clicking. Disease spectrum data covers explicit interest data such as evidence items actively collected by the physician in the DeepEvidence intelligent evidence-based decision-making system, actively marked disease types, subscribed research directions, and frequently accessed evidence categories. Browsing behavior data includes implicit behavioral signals such as the distribution of time doctors spend browsing the evidence summary list, changes in page scrolling speed, the number of times the same evidence is accessed repeatedly, and the duration of time spent browsing.

[0022] The profile building layer, based on multi-source heterogeneous data acquired by the data acquisition layer, performs the construction and updating of multi-dimensional doctor behavioral profiles. Internally, the profile building layer is divided into a feature extraction module and a feature fusion module. The feature extraction module extracts four types of features from the raw data: static profile features, dynamic behavioral features, interest and preference features, and behavioral pattern features. Static profile features are derived from departmental attribute information, including departmental identifiers, professional title codes, and professional field tags. These features remain stable over a relatively long period and serve as the foundational dimensions of the doctor profile. Dynamic behavioral features are derived from historical query records, including sequences of search keywords submitted by doctors within a preset time period, the timestamp distribution of each search keyword, the click sequence of search results, and the click time distribution. These features reflect recent trends in doctors' information needs. Interest and preference features are derived from disease spectrum data, including evidence identifiers actively collected by doctors and their topic tag distribution, actively marked disease types and their hierarchical structure, subscribed research direction tags, and the distribution of frequently accessed evidence categories. These features reflect doctors' long-term professional interests and explicit preferences. Behavioral pattern features are derived from browsing behavior data, including the distribution of average dwell time when doctors browse evidence summaries, the median and variance distribution of page scrolling speed, the distribution of average number of repeated visits to the same evidence, and the probability distribution of jumping out of different time intervals. These features reflect doctors' information acquisition habits and browsing preference patterns.

[0023] The feature fusion module performs unified vectorization and fusion processing on the four extracted features. The feature fusion module first encodes the features numerically: For categorical features, one-hot encoding or label encoding methods are used to quantify them; For numerical features, standardization or normalization methods are used to quantify them; For textual features, a pre-trained language model is used for semantic encoding to obtain a fixed-dimensional semantic vector.

[0024] The feature fusion module performs weighted summation or concatenation operations on the encoded features according to a preset fusion strategy to generate a fused multidimensional doctor behavior profile vector. The dimensionality of the multidimensional doctor behavior profile vector is set according to the application scenario and computing resource limitations of the doctor profile, typically using a dense vector representation of 128 to 512 dimensions. The weight parameters in the fusion process are initialized according to the importance of different types of features and dynamically adjusted through subsequent profile update processes.

[0025] The profile construction layer also includes a profile quality assessment unit, which verifies the quality of the generated multidimensional doctor behavior profile vectors. This unit calculates a stability index for the profile vectors, assessing whether the changes in the profile vectors are within a preset range when the doctor has not generated new behavioral data within a short period. The profile quality assessment unit also calculates a completeness index for the profile vectors, evaluating whether there are feature dimensions with excessively high missing proportions. When the profile quality assessment unit detects that the quality index of the profile vectors is below a preset threshold, it triggers the profile reconstruction process, re-executing the data acquisition and feature fusion processes to generate updated profile vectors.

[0026] The evidence retrieval layer is activated when a doctor initiates a query request through the DeepEvidence intelligent evidence-based decision-making system. The layer first obtains the doctor's submitted query request through the query interface, then parses out the search keywords. Search keywords may be single words, phrases composed of multiple words, or complete query statements. The evidence retrieval layer internally comprises two functional components: a keyword processing unit and a candidate set generation unit.

[0027] The keyword processing unit performs standardization and semantic expansion on the original search keywords. Standardization includes word segmentation, stop word filtering, case unification, and word form restoration, converting the search keywords into standardized word sequences. Semantic expansion employs three strategies: synonym expansion, hyponym / hypernym expansion, and related word expansion. Synonym expansion, based on medical thesaurus and domain dictionaries, expands the search keywords into their synonymous expressions. Hyponym / hypernym expansion, based on the medical concept hierarchy, expands the search keywords into their hypernyms and hyponyms. Related word expansion, based on co-occurrence analysis and semantic similarity calculation, expands the search keywords into related terms that frequently co-occur in medical corpora. The keyword processing unit generates an expanded keyword set using these three strategies, encompassing different expressions and semantic levels of the doctor's search intent.

[0028] The candidate set generation unit performs a search operation in a pre-defined evidence-based database based on the expanded keyword set, generating an initial candidate set of evidence that matches the search keywords. The evidence-based database pre-stores a massive amount of evidence-based medicine evidence, including structured records of different evidence types such as clinical practice guidelines, systematic reviews, randomized controlled trials, cohort studies, case-control studies, and case reports. Each evidence record includes structured fields such as evidence identifier, evidence title, evidence abstract, evidence text, source journal, publication date, evidence quality rating, and subject classification. The candidate set generation unit uses a combination of inverted index technology and full-text search technology to efficiently search the evidence-based database. The inverted index is used to quickly locate evidence records containing specific keywords, while the full-text search is used to score the relevance of the evidence content. The candidate set generation unit includes evidence records with a search score higher than a preset threshold as members of the initial candidate set.

[0029] The feature matching layer performs feature matching calculations between each candidate piece of evidence in the initial evidence candidate set and the multi-dimensional doctor behavior profile generated by the profile construction layer, obtaining a correlation score between each candidate piece of evidence and the doctor profile. The feature matching layer is internally divided into an evidence feature extraction unit, a matching degree calculation unit, and a score fusion unit.

[0030] The evidence feature extraction unit extracts features from each candidate piece of evidence in the initial evidence candidate set, obtaining a feature vector for the candidate evidence. The features extracted by the evidence feature extraction unit include text content features, evidence source features, evidence quality rating features, and evidence publication time features. Text content features use a pre-trained language model to semantically encode the title, abstract, and body of the candidate evidence, obtaining a fixed-dimensional text semantic vector. The pre-trained language model is fine-tuned based on a large-scale corpus in the medical field, enabling it to accurately understand the semantic meaning and professional context of medical terminology. Evidence source features convert categorical information such as the source journal, source database, and evidence type into a numerical feature vector through encoding. Evidence quality rating features directly use the quality rating values ​​of the candidate evidence and perform normalization processing. Evidence publication time features use the number of days between the evidence publication time and the current time as the feature value and perform a logarithmic transformation to compress the numerical range.

[0031] The matching degree calculation unit calculates the multi-dimensional matching degree between candidate evidence and the doctor's profile. First, the matching degree calculation unit calculates the vector similarity between the feature vector of the candidate evidence and the vector of the multi-dimensional doctor's behavioral profile, using a cosine similarity algorithm to obtain a basic relevance score. In essence, vector similarity calculation captures the degree of semantic association between the textual content of the candidate evidence and the long-term interests and short-term needs expressed by the doctor's profile.

[0032] The matching degree calculation unit also calculates the domain matching degree between candidate evidence and doctor's specialty field based on the doctor's professional field label and the subject classification label of candidate evidence. The domain matching degree calculation takes into account the label hierarchy structure. The highest matching degree is given when the subject classification of candidate evidence and doctor's professional field label belong to the same first-level discipline, the second highest matching degree is given when they belong to the same second-level discipline, and so on.

[0033] The matching degree calculation unit also calculates the topic similarity between candidate evidence and the topics that doctors have historically focused on, based on the distribution of topics that doctors have historically focused on and the topic tag distribution of candidate evidence. The topic similarity is calculated based on the cosine similarity between topic vectors. The topic vectors that doctors have historically focused on are obtained by weighted aggregation of the topic vectors of all evidence that doctors have accessed in the past preset time period. The weights take into account time decay factors and access depth factors.

[0034] The matching degree calculation unit also calculates the behavioral fit between candidate evidence and doctors' browsing preferences based on the characteristics of doctors' behavioral patterns and the presentation characteristics of candidate evidence. The calculation of behavioral fit takes into account factors such as doctors' preference for different evidence presentation formats, doctors' preference for evidence of different lengths, and doctors' preference for evidence from different sources.

[0035] The scoring fusion unit weights and fuses the basic relevance score, domain matching degree, topic similarity, and behavioral fit to obtain the final relevance score. The scoring fusion unit employs a linear weighted fusion strategy, and the final relevance score is expressed as the weighted sum of the scores for each component. The weight parameters in the scoring fusion unit are optimized through supervised learning based on click feedback and satisfaction feedback from historical data to maximize prediction accuracy. The scoring fusion unit also includes a scoring calibration unit, which maps each component score to a uniform numerical range through a scoring transformation function, ensuring the comparability of the fused score across different candidate evidence.

[0036] The re-ranking layer re-ranks the initial evidence candidate set based on the relevance score output by the feature matching layer, generating a personalized evidence recommendation list that conforms to the characteristics of the doctor's specialty field and historical topics of interest. The re-ranking layer is internally divided into an initial ranking unit, a diversity test unit, a quality check unit, and an output unit.

[0037] The initial ranking unit sorts all candidate evidence in the initial evidence candidate set in descending order of relevance score from high to low, generating the initial ranking result. The candidate evidence ranked higher in the initial ranking result has a high degree of matching with the doctor's profile, and evidence that is highly relevant to the doctor's professional needs is given priority.

[0038] The diversity verification unit performs a topic diversity test on the first predetermined number of candidate pieces of evidence in the initial ranking results. The unit calculates the topic similarity between adjacent candidate pieces of evidence; if the topic similarity exceeds a predetermined threshold, the two candidate pieces of evidence are deemed to have excessive topic overlap. The diversity verification unit then lowers the ranking of subsequent candidate pieces of evidence, reducing their position in the recommendation list and thus freeing up display space for candidate evidence on other topics. The diversity verification unit also considers the diversity of evidence sources to avoid the evidence in the recommendation list being overly concentrated in the same journal or database.

[0039] The quality verification unit performs evidence quality verification on the ranking results after diversity adjustment. The unit acquires the quality rating information for each candidate piece of evidence. When the quality rating of a candidate piece of evidence falls below a minimum quality threshold, its ranking is adjusted to below a preset lower limit. While candidate evidence below the preset lower limit is still displayed in the recommendation list, it will not appear in the first few pages of recommendations, thus reducing the interference of low-quality evidence on physicians' decisions. The quality verification unit ensures that the recommendation list always contains a certain proportion of high-quality evidence, providing physicians with reliable evidence-based knowledge support.

[0040] The output unit generates a final, sorted, personalized evidence recommendation list and sends it to the doctor's end via a push interface. The output unit generates the recommendation list according to a preset list length and display format. Each recommended piece of evidence includes key information such as the evidence title, evidence summary, source, quality rating, and publication date. The output unit also generates supplementary information related to the recommendation list, including explanations of the recommendation rationale and recommendations for similar evidence, enhancing the doctor's user experience.

[0041] The dynamic profile update mechanism ensures that the multi-dimensional doctor behavior profile can reflect the latest changes in doctors' interests in real time. When a doctor generates new query or browsing behavior, the profile building layer adjusts the weight values ​​of the corresponding features in the doctor profile vector according to the preset profile update strategy. The profile update strategy includes a time decay mechanism, a click feedback mechanism, and a dwell time feedback mechanism.

[0042] The time decay mechanism assigns a weight value to historical query keywords that gradually decreases according to a time decay function. The time decay function can be an exponential decay function or a linear decay function. The decay rate is set individually based on the doctor's activity level. When the doctor's recent behavior is active, a faster decay rate is used to reflect short-term interests first; when the doctor's recent behavior is inactive, a slower decay rate is used to maintain the stability of long-term interests.

[0043] The click feedback mechanism increases the weight of the relevant topic in the interest preference feature when a doctor clicks on a candidate piece of evidence. The mechanism first identifies the distribution of topic tags for the clicked evidence, then increases the weight of the corresponding topic dimension in the interest preference feature vector according to a preset increment. The weight increment is adjusted based on the contextual information at the time of the click, with a larger weight increment given when the click occurs on high-quality evidence.

[0044] The dwell time feedback mechanism determines a doctor's high interest in evidence content when the dwell time exceeds a preset threshold, thus increasing the weight of the corresponding topic in the interest preference feature. The mechanism first calculates the actual dwell time on the evidence details page, then calculates the weight increment based on the ratio of dwell time to the preset threshold. When the dwell time significantly exceeds the preset threshold, the weight increment increases non-linearly to reflect the importance of the doctor's in-depth reading behavior.

[0045] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the scope of the invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0046] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for personalized evidence-based knowledge delivery based on physician behavioral profiles, characterized in that, include: Continuously collect doctors' department attribute information, historical query records, disease spectrum data, and browsing behavior data, and construct multi-dimensional doctor behavior profiles based on the data; When a doctor initiates a query request through the evidence-based knowledge retrieval system, the system obtains the search keywords of the query request and retrieves an initial set of candidate evidence that matches the search keywords from a preset evidence-based database. The candidate evidence in the initial evidence candidate set is matched with the multidimensional doctor behavior profile to obtain the correlation score between each candidate evidence and the doctor profile. Based on the relevance score, the initial evidence candidate set is reordered to generate a personalized evidence recommendation list that conforms to the characteristics of the doctor's specialty field and historical topics of interest, and the personalized evidence recommendation list is pushed to the doctor's terminal for display.

2. The method according to claim 1, characterized in that, The system continuously collects doctors' departmental attribute information, historical query records, disease spectrum data, and browsing behavior data, and constructs a multi-dimensional doctor behavior profile based on the data, including: Collect doctors' department identifiers, professional titles, and professional field tags as static profile features; Collect doctors’ search keywords, search timestamps, and search result click records submitted within a preset time period as dynamic behavioral characteristics; The disease types, research directions of interest, and frequently accessed evidence categories that doctors actively collect or mark in the evidence-based knowledge retrieval system are used as interest and preference features. The duration of doctors' browsing of evidence summaries, scrolling speed, and number of repeated visits were collected as behavioral pattern features. The static profile features, dynamic behavior features, interest preference features, and behavior pattern features are fused to obtain a fused multidimensional doctor behavior profile vector.

3. The method according to claim 2, characterized in that, The feature fusion process of the static profile features, the dynamic behavior features, the interest preference features, and the behavior pattern features includes: Categorical features are quantified using one-hot encoding or label encoding. Numerical features are quantified using standardization or normalization methods; Textual features are semantically encoded using a pre-trained language model to obtain fixed-dimensional semantic vectors; The encoded features are weighted and summed or concatenated according to a preset fusion strategy to generate a fused multidimensional doctor behavior profile vector.

4. The method according to claim 1, characterized in that, The step of performing feature matching calculations between each candidate piece of evidence in the initial candidate evidence set and the multidimensional doctor behavior profile to obtain a correlation score between each candidate piece of evidence and the doctor profile includes: Extract the text content features, evidence source features, evidence quality rating features, and evidence publication time features of each candidate piece of evidence; The feature vectors of candidate evidence are compared with the multidimensional doctor behavior profile vectors to calculate the basic relevance score. Based on the professional field label of the doctor's current department, calculate the field matching degree between each candidate piece of evidence and the doctor's professional field; Based on the distribution of physicians' historical topics of interest, the thematic similarity between each candidate piece of evidence and the physicians' historical topics of interest is calculated. Based on the behavioral patterns of doctors, calculate the behavioral fit between each candidate piece of evidence and the doctors' browsing preferences. The basic relevance score, the domain matching degree, the topic similarity degree, and the behavior fit degree are weighted and fused to obtain the final relevance score.

5. The method according to claim 4, characterized in that, The vector similarity calculation uses the cosine similarity algorithm.

6. The method according to claim 1, characterized in that, The process of reordering the initial evidence candidate set based on the relevance score to generate a personalized evidence recommendation list that conforms to the characteristics of the physician's specialty field and historical topics of interest includes: The candidate evidence is sorted in descending order of relevance score to generate an initial ranking result. A diversity test is performed on the first preset number of candidate evidences in the initial sorting results. When the topic similarity of adjacent candidate evidences exceeds a preset threshold, the sorting position of subsequent candidate evidences is reduced. The quality of evidence is verified on the adjusted ranking results. When the quality rating of candidate evidence is lower than the minimum quality threshold, its ranking position is adjusted to below the preset lower limit. Output the final sorted list of personalized evidence recommendations.

7. The method according to claim 6, characterized in that, The step of performing a diversity test on the first preset number of candidate pieces of evidence in the initial ranking results also includes: Calculate the topic similarity between adjacent candidate evidence; Consider the diversity of evidence sources and avoid making the evidence in the recommended list too concentrated in the same journal or the same database.

8. The method according to claim 1, characterized in that, Also includes: The multidimensional doctor behavior profile is dynamically updated. When a doctor generates new query or browsing behavior, the weight values ​​of the corresponding features in the doctor profile vector are adjusted according to the preset profile update strategy.

9. The method according to claim 4, characterized in that, The feature vectors used in the feature matching calculation are obtained by semantically encoding the doctor's behavior data and evidence content through a pre-trained language model. The pre-trained language model is fine-tuned based on a large-scale corpus in the medical field.

10. A personalized evidence-based knowledge delivery system based on physician behavioral profiles, characterized in that, It includes at least a profile building module, an evidence retrieval module, a feature matching module, and a re-ranking module, among which: The profile building module is used to continuously collect doctors' department attribute information, historical query records, disease spectrum data, and browsing behavior data, and build a multi-dimensional doctor behavior profile based on the above data. The evidence retrieval module is used to obtain the search keywords of the query request when a doctor initiates a query request through the evidence-based knowledge retrieval system, and retrieve an initial set of candidate evidence that matches the search keywords from a preset evidence-based evidence database. The feature matching module is used to perform feature matching calculations between each candidate piece of evidence in the initial evidence candidate set and the multidimensional doctor behavior profile to obtain a correlation score between each candidate piece of evidence and the doctor profile. The re-ranking module is used to re-rank the initial evidence candidate set based on the relevance score, generate a personalized evidence recommendation list that conforms to the characteristics of the doctor's specialty field and historical topics of interest, and push the personalized evidence recommendation list to the doctor's terminal for display.

Images