Family doctor service recommendation method and system based on user portrait

Abstract

The invention relates to the technical field of intelligent recommendation, in particular to a family doctor service recommendation method and system based on a user portrait, and the method comprises the following steps: collecting the multi-modal health monitoring data and historical treatment evaluation records of residents, and obtaining the basic implementation information of family doctors, a dynamic user portrait is constructed by using a natural language processing algorithm and a time decay function, a doctor ability model for evaluating professional skilled fields and service satisfaction of doctors in multiple dimensions is established, the Euclidean distance between a resident health demand feature vector and a doctor service ability feature vector is calculated, and a family doctor recommendation list is generated according to the distance. According to the method, the multi-dimensional user portrait containing the health features and the behavior preferences is constructed, the portrait weight is dynamically updated by using the time decay factor, the real-time medical demands of residents are accurately captured, and depth similarity calculation is performed in combination with vectorization representation of the doctor service capability, so that efficient and accurate matching of the medical supply and demand parties is realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent recommendation technology, and in particular to a method and system for recommending family doctor services based on user profiles. Background Technology

[0002] The field of intelligent recommendation technology involves the digital processing of analyzing user data using computer algorithms and providing personalized suggestions. Its core lies in mining user behavior patterns and potential needs to filter content or services from massive amounts of information that highly match the user's current intentions. Traditional methods of recommending family doctor services primarily rely on manual screening or simple geographic distance matching for doctor allocation. This typically involves calculating latitude and longitude between residents' addresses and doctors' registered locations, or community staff manually checking doctor schedules and residents' health records to directly provide residents with the contact information of the nearest or available doctor.

[0003] Coarse matching based on geographical location or basic department classification ignores the dynamic changes in residents' health status and in-depth medical needs. It does not conduct in-depth quantitative analysis of the suitability of doctors' past treatment effects with residents' current conditions, resulting in biased recommendation results. It fails to cover residents' personalized medical service needs, causing uneven allocation of medical resources, leading to service delays. It lacks continuous tracking and utilization of long-term health data, reduces the conversion rate of family doctor contract services, and makes it difficult to form a long-term health management closed loop for patients with chronic diseases or special populations. This restricts the improvement of the intelligence level of the primary medical service system and hinders the process of optimizing the allocation of medical resources under the hierarchical medical system. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a family doctor service recommendation method and system based on user profiles.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a family doctor service recommendation method based on user profiles, comprising the following steps: S1: Collect multimodal health monitoring data and historical medical evaluation records of residents within the target jurisdiction, and obtain basic practice information of family doctors through the medical institution management terminal; S2: Perform outlier cleaning and standardization on the multimodal health monitoring data, use natural language processing algorithms to perform disease entity recognition and sentiment analysis on the historical medical evaluation records, and construct a dynamic user profile including physiological indicator features and medical service demand features by combining time decay function. S3: Based on the basic practice information of the family doctor, establish a multi-dimensional evaluation model of the doctor's professional expertise, average daily number of visits, and past patient satisfaction, and configure a quantitative weight matrix corresponding to multiple practice indicators in the doctor's competency model. S4: Map the dynamic user profile to the feature space of the doctor's ability model, calculate the multidimensional weighted Euclidean distance between the resident health demand feature vector and the doctor service ability feature vector, determine the degree of supply and demand matching based on the distance, and generate a family doctor recommendation list in descending order of matching degree.

[0006] As a further aspect of the present invention, step S1 specifically comprises: S11: Periodically extract blood pressure, blood sugar, heart rate, and step count data from smart wearable devices and home medical terminals worn by residents in the jurisdiction through the Internet of Things interface to obtain the multimodal health monitoring data; S12: Access the user-authorized historical medical records database through an encrypted transmission protocol, retrieve the resident's past outpatient medical records, prescription records, and satisfaction ratings for past medical services, and obtain the historical medical evaluation records. S13: Through the data interface of the medical institution management terminal, synchronously retrieve the registered department, professional title level, work schedule and professional skills certificate information of the practicing doctors to obtain the basic practice information of the family doctors.

[0007] As a further aspect of the present invention, step S2 specifically comprises: S21: Use the interquartile range algorithm to identify and remove outliers in the multimodal health monitoring data, and use the maximum-minimum normalization method to map the physiological index data to the standard numerical range to generate a physiological health status vector. S22: Semantic features are extracted from the historical medical evaluation records using a bidirectional long short-term memory network model to identify disease entities and emotional keywords, and time decay coefficients are calculated by combining the timestamp information of residents' medical visits to generate a medical service demand vector. S23: Perform feature concatenation and dimensionality reduction on the physiological health status vector and the medical service demand vector to establish the dynamic user profile.

[0008] As a further aspect of the present invention, step S3 specifically comprises: S31: Perform keyword aggregation analysis on the descriptions of diseases of expertise in the basic practice information of family doctors, generate professional expertise tags for doctors, and collect data on the number of patients and follow-up visits within a specified period to calculate the average daily patient load index. S32: Perform sentiment polarity analysis on past patients' evaluation texts, extract the proportion of positive evaluations and keywords related to service attitude, and combine them with the medical institution's annual assessment scores to generate a past patient satisfaction index. S33: Construct a multi-dimensional competency evaluation system by combining the professional expertise tags, the average daily patient load index, and the past patient satisfaction index. Calculate the information entropy value and coefficient of variation of multiple evaluation indicators using the entropy weight method. Establish objective weight values ​​for multiple professional competency indicators based on the degree of variation, and generate the quantitative weight matrix.

[0009] As a further aspect of the present invention, step S4 specifically comprises: S41: Extract the demand feature dimension corresponding to the doctor's practice indicators from the dynamic user profile, and convert it into a resident health demand feature vector that is consistent with the doctor's competence model dimension; S42: Call the quantized weight matrix to calculate the difference between the resident health demand feature vector and the service capacity feature vector of each doctor in the candidate doctor database, and generate the multidimensional weighted Euclidean distance. S43: Sort the multidimensional weighted Euclidean distances in ascending order of value, select candidate doctors whose distance values ​​are less than a preset recommendation threshold, and establish a family doctor recommendation list.

[0010] As a further aspect of the present invention, the process of correcting the intensity of emotional keywords by the time decay function includes: Obtain the timestamps and original emotional intensity values ​​corresponding to multiple entries in the historical medical evaluation record, and calculate the time difference between the current system time and the timestamps. Substitute the time difference into a preset exponential decay model, calculate the weighting coefficients for the differentiated historical periods, and use the weighting coefficients to weight and correct the original emotional intensity value to generate a real-time emotional feature value that can reflect the residents' recent psychological needs. The calculation formula for the exponential decay model is as follows: ; in, This represents the real-time weighting coefficients after time decay correction. Represents the initial weight baseline value. The time sensitivity constant representing the rate of decay. The timestamp value represents the current system time. The timestamp value representing when the historical evaluation record was generated.

[0011] As a further aspect of the present invention, the process of calculating the objective weight value includes: Build including Doctors and The original data matrix of the professional practice indicators is used to nonnegate the elements in the matrix, and the first indicator is calculated. The first item under the indicator The characteristic weighting value of each doctor; Calculate the first based on the aforementioned characteristic weight value The information entropy of each indicator is used to obtain the information utility value of the indicator. After normalizing the information utility value, an objective weight value is generated. The quantitative weight matrix is ​​established by combining the objective weight values ​​corresponding to all professional performance indicators.

[0012] As a further aspect of the present invention, the calculation process of the multidimensional weighted Euclidean distance includes: Obtain the resident health needs feature vector and the target doctor's service capacity feature vector, and calculate the numerical difference between the two in multiple feature dimensions; By introducing the weight coefficients in the quantization weight matrix, a weighted summation operation is performed on the square terms of the numerical differences, and the square root of the summation result is performed to generate a matching metric value. The formula for calculating the multidimensional weighted Euclidean distance is as follows: ; in, Representing residential users With family doctor The multidimensional weighted Euclidean distance values ​​between them Represents the total number of feature dimensions. Index representing feature dimension, Represents the quantization weight matrix of the th Weight coefficients for each feature dimension The feature vector representing the residents' health needs is in the th... Normalized values ​​of dimensions The feature vector representing the doctor's service capability is in the th... Normalized values ​​of dimensions.

[0013] As a further aspect of the present invention, the process of generating the family doctor recommendation list includes: The multidimensional weighted Euclidean distance is converted into a percentage matching score, and the candidate doctors are initially screened based on the matching score, retaining the preferred set of doctors whose scores are in the top 30%. Obtain the current number of signed contracts and service radius for each doctor in the preferred doctor set, remove doctors whose current number of signed contracts has reached the limit or whose service radius fails to cover the residents' residences, and generate a candidate set of doctors who can be signed contracts; The candidate set of doctors eligible for contracting is arranged in descending order of matching scores. The practice records and recommendation reasons of the top five doctors are extracted to generate the family doctor recommendation list.

[0014] A family doctor service recommendation system based on user profiles, the system being used to implement the aforementioned family doctor service recommendation method based on user profiles, the system comprising: The multi-source data acquisition module is used to collect multimodal health monitoring data and historical medical evaluation records of residents within the target jurisdiction, and to obtain basic practice information of family doctors through the medical institution management terminal; The profile building and analysis module is used to clean and standardize outliers in the multimodal health monitoring data, use natural language processing algorithms to identify disease entities and analyze sentiment tendencies in the historical medical evaluation records, and build dynamic user profiles by combining time decay functions. The doctor model building module is used to build a multi-dimensional evaluation model of the doctor's professional expertise, average daily number of patients seen, and past patient satisfaction based on the basic practice information of the family doctor, and to configure a quantitative weight matrix corresponding to multiple practice indicators in the doctor's competency model. The intelligent recommendation calculation module is used to map the dynamic user profile to the feature space of the doctor's ability model, calculate the multidimensional weighted Euclidean distance between the resident health demand feature vector and the doctor service ability feature vector, determine the degree of supply and demand matching based on the distance, and generate a family doctor recommendation list in descending order of matching degree.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: This invention collects multi-source heterogeneous health data and constructs dynamic user profiles using a time decay algorithm. It captures the evolution characteristics of residents' health status in real time, establishes a doctor competency model based on multi-dimensional professional attribute weights, and uses a vector space distance algorithm to calculate the similarity of features between supply and demand sides. This enables personalized and accurate recommendations, solves the shortcomings of traditional methods such as coarse matching and neglect of dynamic needs, optimizes the allocation structure of medical resources, improves the response speed and contract matching degree of family doctor services, enhances the targeting of health management for chronic diseases and special populations, promotes the transformation of primary healthcare services towards intelligence and refinement, and facilitates the deep integration of doctor-patient relationships under the hierarchical medical system. Attached Figure Description

[0016] Figure 1 This is a flowchart of the family doctor service recommendation method based on user profiles according to the present invention; Figure 2 This is a flowchart of the data collection and professional information acquisition process of this invention; Figure 3 This is a flowchart illustrating the dynamic user profile construction and sentiment correction process of this invention. Figure 4 This is a flowchart illustrating the physician competency model establishment and weight calculation process of this invention. Figure 5 This is a flowchart of the supply and demand matching calculation and doctor recommendation process of this invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the software-based technical solution is described in detail below with reference to system architecture diagrams and embodiments. It should be understood that the specific embodiments described herein are only for explaining the technical solutions of this invention and do not constitute a limitation on the scope of protection.

[0018] In the description of this invention, the system architecture relationships or data processing flows indicated by terms such as "layer," "module," "interface," "data flow," "client," and "server" are all defined based on the architecture diagram or flowchart corresponding to the embodiments. This way of describing is only used to clearly illustrate the logical relationships between the elements in the technical solution, and not to limit the physical deployment form. The term "multiple" includes two or more technical units, including but not limited to multiple data nodes, processing threads, service instances, or functional components and other scalable elements. The specific number is determined according to the actual business scenario and needs to be specifically specified.

[0019] Please see Figure 1 and Figure 2 This invention provides a technical solution: a family doctor service recommendation method based on user profiles, comprising the following steps: S1: Collect multimodal health monitoring data and historical medical evaluation records of residents within the target jurisdiction, and obtain basic practice information of family doctors through the medical institution management terminal.

[0020] S11: Regularly extract blood pressure, blood sugar, heart rate, and step count data from smart wearable devices and home medical terminals worn by residents in the jurisdiction through the Internet of Things interface to obtain multimodal health monitoring data.

[0021] S12: Access the user-authorized historical medical records database through an encrypted transmission protocol to retrieve residents' past outpatient medical records, prescription records, and satisfaction ratings for past medical services, and obtain historical medical evaluation records.

[0022] S13: Through the data interface of the medical institution management terminal, synchronously retrieve the registered departments, professional titles, work schedules and professional skills certificates of practicing doctors to obtain basic practice information of family doctors.

[0023] Step S1 initiates a comprehensive data collection process for residents' health data within the target area. This involves sending parallel data request commands to IoT gateways and medical institution management servers within the area via distributed data collection service nodes configured in the data center. First, a long-term communication channel is established with smart wearable devices and home medical terminals. The MQTT protocol is used to subscribe to a data stream with the topic "Health / Monitor / #", receiving JSON-formatted data packets in real-time from devices such as wristbands, smart blood pressure monitors, and blood glucose meters. The data packet payload is parsed to extract key physiological indicator fields, including systolic blood pressure, diastolic blood pressure, fasting blood glucose, 2-hour postprandial blood glucose, resting heart rate, and daily steps. During data extraction, for continuous heart rate monitoring data sampled every 10 minutes, data integrity is checked. If more than three consecutive time steps of packet loss are detected, the data for that time period is marked as invalid, and a retransmission command is triggered. Simultaneously, an SQL query request is initiated to the historical medical database authorized by the user's digital certificate through a configured SSL / TLS 1.3 encrypted transmission channel. For each target resident, a multi-table joint query operation is performed, linking electronic medical records, prescription records, and patient satisfaction surveys. Outpatient medical record text data from the past 36 months is retrieved and downloaded, covering chief complaint, present illness, past medical history, and diagnostic conclusions. Simultaneously, information on drug names, dosages, and frequency of use is extracted from prescription records, along with users' subjective ratings (out of 1-10) and unstructured text evaluations of each medical service.

[0024] While collecting resident-side data, the system concurrently calls the RESTful API interface of the medical institution's management terminal to access the human resources management subsystem. By sending HTTP GET requests, it retrieves the complete practice records of registered family doctors within the jurisdiction. The returned XML or JSON response body is parsed to extract the doctor's unique identifier, registration department code (standardized to ICD-10 departmental classification), professional title (e.g., chief physician, attending physician, mapped to numerical codes 1-4), current shift schedule (accurate to service status every half-day), and a list of professional skill certificates held. This multi-source heterogeneous data is then uniformly aggregated into the temporary storage area of ​​the Hadoop Distributed File System and metadata is tagged, recording the data source IP, collection timestamp, and data type, thus constructing a raw data lake for subsequent data cleaning and processing. For example, for resident "Zhang San" (user ID: U_440304_001) living in XX Street, at 08:00 on May 20, 2024, the heart rate data sequence of the previous 24 hours was successfully extracted from his smart bracelet. The average resting heart rate was 72 beats / min, and the average morning blood pressure of the past week was obtained from the home blood pressure monitor cloud platform, which showed an average of 138 / 88 mmHg. At the same time, his medical record at the community health service center in November 2023 was successfully retrieved. The medical record text shows "Main diagnosis: primary hypertension; Chief complaint: dizziness for one week", and the prescription record includes "amlodipine besylate tablets 5mg qd". The satisfaction rating for this visit was 8 points, and the evaluation text was "the doctor's attitude was kind and the explanation was detailed". On the doctor's side, the file of Dr. Li Si, who is in charge of this grid, was successfully retrieved, showing that he is a "Deputy Chief Physician of General Practice", sees patients all day Monday to Friday, and holds a "Chronic Disease Management Specialist" certificate. This series of actions ensured the comprehensiveness and timeliness of data input, providing a solid data foundation for building high-precision user profiles.

[0025] The aforementioned MQTT protocol refers to Message Queuing Telemetry Transport Protocol, a "lightweight" communication protocol based on the publish / subscribe model. This protocol is built on top of the TCP / IP protocol and is designed to provide message transmission services for remote devices with low hardware performance and network environments with poor network conditions.

[0026] Please see Figure 1 and Figure 3 S2: Perform outlier cleaning and standardization on multimodal health monitoring data, use natural language processing algorithms to identify disease entities and analyze sentiment tendencies on historical medical evaluation records, and construct dynamic user profiles that include physiological indicator features and medical service demand features by combining time decay functions.

[0027] S21: The interquartile range algorithm is used to identify and remove outliers in multimodal health monitoring data. The maximum-minimum normalization method is used to map the physiological index data to a standard numerical range to generate a physiological health status vector.

[0028] S22: Semantic features are extracted from historical medical evaluation records using a bidirectional long short-term memory network model to identify disease entities and emotional keywords. The time decay coefficient is calculated by combining the timestamp information of residents' medical visits to generate a medical service demand vector.

[0029] S23: Perform feature concatenation and dimensionality reduction on the physiological health status vector and the medical service demand vector to establish a dynamic user profile.

[0030] The process of using the time decay function to correct the intensity of sentiment keywords includes: Obtain the timestamps and original emotional intensity values ​​corresponding to multiple entries in the historical medical evaluation records, and calculate the time difference between the current system time and the timestamps; Substitute the time difference into the preset exponential decay model, calculate the weight coefficient for the differentiated historical period, and use the weight coefficient to weight and correct the original emotional intensity value to generate a real-time emotional feature value that can reflect the residents' recent psychological needs. The calculation formula for the exponential decay model is: ; in, This represents the real-time weighting coefficients after time decay correction. Represents the initial weight baseline value. The time sensitivity constant representing the rate of decay. The timestamp value represents the current system time. The timestamp value representing when the historical evaluation record was generated.

[0031] Step S2 starts the data cleaning and standardization engine, first reading the raw multimodal health monitoring data from the temporary storage area. For physiological indicator data, the Python Pandas library is used to load the data frame, and an outlier detection algorithm based on interquartile range is executed independently for each type of indicator (such as systolic blood pressure). The first quartile of the data sequence is calculated. With the third quartile This leads to the interquartile range. Set the outlier detection threshold to: The process iterates through all sampling points, marking data points outside the closed interval as noise and removing them. For the cleaned data, a min-max normalization method is used to map each physiological indicator to... The standard range. Specifically, for a specific indicator value. Obtain the minimum value of this indicator within the medical reference range. With the maximum value (For example, systolic blood pressure is set to 90-180 mmHg), calculate .like If the value exceeds the reference range, it is truncated to ensure that the output value is strictly within the unit interval, thereby generating a standardized physiological health state vector. .

[0032] Next, the natural language processing pipeline is initiated to process historical medical record evaluations. A pre-trained BERT model is loaded to segment and embed the outpatient medical record text into words. The embedded vectors are then input into a bidirectional long short-term memory network model for feature extraction. The Bi-LSTM model consists of two independent LSTM layers, one forward and one backward, with each layer having 128 hidden units. At time step... Input gate Controlling the inflow of new information, the forget gate Determines the state of old cells The degree of retention, output gate Control the current hidden state The output is processed by capturing symptom descriptions (such as "headache" and "palpitation") in the text as symptom entities through forward propagation, capturing modifiers in the context through backpropagation, and using a conditional random field layer to decode the labels of the output sequence to accurately identify the symptom entities. The sentiment analysis task for the evaluation text is processed simultaneously. The same Bi-LSTM architecture is used to connect fully connected layers and a Softmax classifier to classify the sentiment polarity and predict the intensity of "satisfaction rating text". During model training, the cross-entropy loss function is used to calculate the difference between the predicted probability distribution and the true sentiment labels (such as positive and negative). The Adam optimizer is used to update the parameters, with an initial learning rate set to 0.001 and a batch size set to 32. Iterative training is performed for 50 epochs until the validation set accuracy converges to over 92%.

[0033] After extracting the disease entity and emotional keywords, a time dimension is introduced to dynamically adjust the feature strength. For each historical record, its generation timestamp is obtained. and read the current system timestamp Calculate the time difference (Units converted to days). Substituting the emotional intensity value into the exponential decay model Weighting is applied. The calculation formula for the exponential decay model is: .in, This represents the real-time weighting coefficients after time decay correction, which are used for subsequent weighting of user profile features. This represents the initial weight baseline value, which is usually set to 1.0; The time sensitivity constant representing the rate of decay is set to 0.02 here, based on the Ebbinghaus forgetting curve and the timeliness experiment of medical service experience. The timestamp value represents the current system time; This represents the timestamp value when the historical evaluation record was generated. Verified through A / B testing, when… When set to 0.02, this refers to an extremely satisfactory review from one year ago (365 days ago) (initial strength). ),calculate Its impact on the current profile is negligible; however, the weight of the evaluation from one week ago (7 days ago) is... It still retains a high level of influence. Through this step, real-time emotional feature values ​​that reflect residents' recent psychological needs are generated, and combined with the characteristics of the symptoms, a vector of medical service needs is constructed. Finally, an autoencoder was used to... and After feature concatenation and dimensionality reduction, the high-dimensional sparse vector is compressed into a 64-dimensional dense vector, thus formally establishing a dynamic user profile that includes both physiological and psychological dimensions.

[0034] The BERT model mentioned above refers to the Transformers bidirectional encoder representation model, a pre-trained language model based on the Transformer architecture, which aims to pre-train deep bidirectional representations by jointly adjusting the left and right contexts in all layers.

[0035] Please see Figure 1 and Figure 4 S3: Based on the basic practice information of family doctors, establish a multi-dimensional physician competency model to evaluate the physician's professional expertise, average daily number of patients seen, and past patient satisfaction, and configure a quantitative weight matrix corresponding to multiple practice indicators in the physician competency model.

[0036] S31: Perform keyword aggregation analysis on the descriptions of diseases of expertise in the basic practice information of family doctors, generate professional expertise tags for doctors, and collect data on the number of patients and follow-up visits within a specified period to calculate the average daily patient load index.

[0037] S32: Perform sentiment polarity analysis on past patient evaluation texts, extract the proportion of positive evaluations and keywords related to service attitude, and combine them with the medical institution's annual assessment scores to generate a past patient satisfaction index.

[0038] S33: Construct a multi-dimensional competency evaluation system by combining professional expertise tags, daily average patient load index, and past patient satisfaction index. Calculate the information entropy value and coefficient of variation of multiple evaluation indicators using the entropy weight method. Establish objective weight values ​​for multiple professional performance indicators based on the degree of variation, and generate a quantitative weight matrix.

[0039] The process of calculating objective weight values ​​includes: Build including Doctors and The original data matrix of the professional practice indicators is used to nonnegate the elements in the matrix, and the first indicator is calculated. The first item under the indicator The characteristic weighting value of each doctor; Calculate the first based on the characteristic weight value The information entropy of each indicator is used to obtain the information utility value of the indicator. After normalizing the information utility value, an objective weight value is generated. A quantitative weight matrix is ​​established by combining the objective weight values ​​corresponding to all professional performance indicators.

[0040] Step S3 constructs a multi-dimensional physician competency assessment model based on basic family doctor practice information. First, keywords are extracted from the physician's personal profile and descriptions of their areas of expertise. The TF-IDF algorithm is used to calculate the weight of each medical term in a specific physician's document, and the top 5 keywords with the highest weights are selected as the physician's "professional expertise tags" (e.g., "diabetes management," "pediatric asthma," "postpartum rehabilitation"). Simultaneously, the system accesses the medical institution's operational database to track the total number of patients seen by each physician over the past 90 calendar days. Number of follow-up visits Define the daily average patient load index. This indicator directly reflects the busyness of doctors and the difficulty of making appointments; it defines the follow-up visit rate. This serves as an auxiliary indicator for measuring doctor-patient bonding. Next, we delve deeper into past patient reviews. We perform fine-grained sentiment analysis on all reviews attributed to that doctor, extracting the number of positive reviews. Compared with the total number of evaluations Calculate the positive evaluation ratio Simultaneously, word cloud analysis technology was used to identify service attitude-related keywords (such as "patient," "meticulous," "hurried," and "indifferent"), which were then combined with the hospital's annual performance evaluation of medical ethics and conduct (out of 100) to create a weighted index of past patient satisfaction. .

[0041] To ensure the objectivity and scientific rigor of the evaluation system, the entropy weight method is used to determine the weights of each professional performance indicator. A system is constructed that includes... Doctors and Original data matrix of professional practice indicators Non-negative transformation is performed on the elements in the matrix: for positive indicators (the higher the value, the better, such as satisfaction), For negative indicators (the lower the value, the better, such as patient load); To avoid zero values ​​in logarithmic operations, here, for all... Add a tiny offset of 0.001. Then, calculate the... The first item under the indicator The characteristic proportion of each doctor Calculate the first based on the characteristic proportion value. Information entropy of the indicator , where constant This leads to the information utility value of the indicator. After normalizing the information utility values, objective weight values ​​for each indicator are generated. and all Combined to form a quantized weight matrix .

[0042] Table 1. Example of data for family doctor practice indicators and entropy weight method calculation. Table 1 shows a sample of professional performance indicators for some family doctors in a community healthcare institution. The "load index (normalized inverse)" is calculated based on the average daily number of patients seen. The calculation sets a maximum daily capacity of 60 patients and a minimum of 0 patients for doctors within the jurisdiction. The formula is as follows: Taking Doc_01 as an example, it receives an average of 45 patients per day, calculated as follows: This indicates a low level of idle time (high load). Through actual calculation, for the above dataset, after entropy weighting, the weight of "good at label matching" is obtained. The weight of "average daily patient load" The weight of "patient satisfaction" This result indicates that, within the current distribution of physicians, the degree of matching in professional competence is the most significant factor in differentiating physician suitability. This quantified weight matrix will be directly used in subsequent distance calculations to ensure that the recommendation results are not based on a single dimension, but rather on the optimal solution resulting from the combined effect of multiple indicators.

[0043] The aforementioned entropy weighting method refers to an objective weighting method that judges the degree of dispersion of an indicator by calculating its information entropy. The greater the degree of dispersion of an indicator, the greater its impact (i.e., weight) on the comprehensive evaluation, thereby avoiding the human bias brought about by subjective weighting methods.

[0044] Please see Figure 1 and Figure 5S4: Map the dynamic user profile to the feature space of the doctor's ability model, calculate the multidimensional weighted Euclidean distance between the resident health demand feature vector and the doctor service ability feature vector, determine the degree of supply and demand matching based on the distance, and generate a family doctor recommendation list in descending order of matching degree.

[0045] S41: Extract the demand feature dimensions corresponding to the doctor's professional indicators from the dynamic user profile, and convert them into a resident health demand feature vector consistent with the doctor's competency model dimensions.

[0046] S42: Call the quantized weight matrix to calculate the difference between the resident health demand feature vector and the service capacity feature vector of each doctor in the candidate doctor database, and generate a multi-dimensional weighted Euclidean distance.

[0047] S43: Sort the multidimensional weighted Euclidean distances in ascending order of value, select candidate doctors whose distance values ​​are less than the preset recommendation threshold, and build a family doctor recommendation list.

[0048] The calculation process of multidimensional weighted Euclidean distance includes: Obtain the feature vector of residents' health needs and the feature vector of the target doctor's service capabilities, and calculate the numerical difference between the two in multiple feature dimensions; By introducing weight coefficients from the quantization weight matrix, a weighted summation operation is performed on the square terms of the numerical differences, and the square root of the summation result is taken to generate a matching metric. The formula for calculating the multidimensional weighted Euclidean distance is: ; in, Representing residential users With family doctor The multidimensional weighted Euclidean distance between them. Represents the total number of feature dimensions. Index representing feature dimension, Represents the first in the quantization weight matrix Weight coefficients for each feature dimension The characteristic vector representing residents' health needs in the th Normalized values ​​of dimensions The feature vector representing the physician's service capability is in the th... Normalized values ​​of dimensions.

[0049] The process of generating a family doctor referral list includes: The multidimensional weighted Euclidean distance is converted into a percentage matching score, and the candidate doctors are initially screened based on the matching score, retaining the set of preferred doctors whose scores are in the top 30%. Obtain the current number of signed contracts and service radius for each doctor in the preferred doctor set, remove doctors whose current number of signed contracts has reached the limit or whose service radius fails to cover the residents' residences, and generate a candidate set of doctors who can be signed contracts; Arrange the candidate set of doctors eligible for contract signing in descending order of matching score, extract the professional records and recommendation reasons of the top five doctors, and generate a family doctor recommendation list.

[0050] Step S4 executes the core calculation process for supply and demand matching, mapping the dynamic user profile generated in S2 to the feature space of the doctor's competency model constructed in S3. First, it extracts the demand feature dimensions from the user profile that directly correspond to the doctor's professional indicators. For example, if the user profile shows an extremely high risk of "hypertension" and sentiment analysis shows a strong demand for "patient explanation," then the user's health demand feature vector is... Build as .in, The weighted demand value corresponding to hypertension disease management This corresponds to users' tolerance for doctor waiting time (inverse mapping; the more urgent the need, the lower the tolerance, i.e., a preference for doctors with low workload). Expectations regarding customer service attitude. Assuming normalized user expectations... The feature vector is (This means they urgently need specialist treatment, are not too concerned about waiting time, and highly value attitude.) Subsequently, the candidate doctor database is traversed, and a quantitative weighting matrix is ​​invoked. Calculate the relationship between the user and each doctor. The multidimensional weighted Euclidean distance between them.

[0051] The formula for calculating the multidimensional weighted Euclidean distance is: .in, Representing residential users With family doctor The multidimensional weighted Euclidean distance between them; the smaller the value, the higher the matching degree. This represents the total number of feature dimensions, which is 3 in this case. Index representing the feature dimension; Represents the first in the quantization weight matrix Weight coefficients for each feature dimension; The characteristic vector representing residents' health needs in the th Normalized values ​​of dimensions; The feature vector representing the physician's service capability is in the th... Normalized values ​​of dimensions.

[0052] Taking Doc_01 in Table 1 as an example, its capability feature vector (Corresponding to strengths, workload, and satisfaction) after normalization is Substitute the above parameters into the formula for calculation: First, calculate the square of each difference, where the specialty is... The load item is The attitude item is Then perform a weighted summation, the result is: Finally, the square root is used to obtain the distance. Similarly, calculate the distance to other doctors. After calculation, convert the multidimensional weighted Euclidean distance into a percentage-based matching score. The candidate doctor list is then fully sorted according to scores from highest to lowest. A primary screening strategy is implemented, selecting the top 30% of doctors as the preferred set. Based on this, a secondary filtering is performed using real-world constraints. The real-time database of the family doctor contract management system is accessed to query the number of currently contracted residents for each doctor in the preferred set. If a doctor's contracted residents reach the policy-mandated limit (e.g., 2000), they are removed from the candidate set. Simultaneously, a geographic information system API is called to obtain the latitude and longitude coordinates of the doctor's community health service center and the user's residential coordinates, calculating the straight-line distance or route planning distance. If the distance exceeds the doctor's service radius (e.g., a 15-minute walk or a 3-kilometer radius), and the user has not selected the "Accept Remote Services" option, they are also removed.

[0053] Table 2. Example of generating a family doctor recommendation list. As shown in Table 2, although Doc_02 was geographically closest and had a decent score, it was automatically removed because the number of signed contracts had been reached; Doc_06 was removed because it exceeded the service radius. The system ultimately recommended Doc_01, Doc_03, and Doc_05 to the user, which had the best overall matching score and service capabilities (Doc_01's score was calculated based on the aforementioned distance of 0.1071). This ensures the accuracy and feasibility of the recommendations. Finally, the candidate set of doctors eligible for contract signing, after double screening, is again sorted in descending order of matching score. Detailed professional files of the top five doctors are extracted, including photos, biographies, tags indicating areas of expertise, and specific reasons for recommendation. A final family doctor recommendation list is then generated and pushed to users' mobile devices. Experimental data shows that compared to traditional methods based on geographical proximity or random allocation, this approach significantly improves the contract signing success rate and subsequent satisfaction with service delivery in practical applications. In a pilot test involving 5,000 residents, the conversion rate of residents clicking to view details on the recommendation list generated using this method increased by 45%, the first-month follow-up visit rate after signing increased by 22%, and the average quarterly satisfaction score of signed residents for family doctor services rose from 8.2 to 9.4.

[0054] The aforementioned multidimensional weighted Euclidean distance refers to assigning different weight coefficients to each coordinate dimension based on the standard Euclidean distance, thereby reflecting the differentiated importance of different feature dimensions in similarity calculation.

[0055] A family doctor service recommendation system based on user profiles, which is used to execute the aforementioned family doctor service recommendation method based on user profiles, includes: The multi-source data acquisition module is used to collect multimodal health monitoring data and historical medical evaluation records of residents within the target jurisdiction, and to obtain basic practice information of family doctors through the medical institution management terminal; The profile building and analysis module is used to clean and standardize outlier data from multimodal health monitoring data, use natural language processing algorithms to identify disease entities and analyze sentiment tendencies in historical medical evaluation records, and build dynamic user profiles by combining time decay functions. The doctor model building module is used to build a multi-dimensional evaluation model of doctors’ professional expertise, average daily number of patients seen, and past patient satisfaction based on the basic practice information of family doctors, and to configure a quantitative weight matrix corresponding to multiple practice indicators in the doctor’s competency model. The intelligent recommendation calculation module maps dynamic user profiles to the feature space of the doctor's capability model, calculates the multidimensional weighted Euclidean distance between the feature vector of residents' health needs and the feature vector of doctors' service capabilities, determines the degree of supply and demand matching based on the distance, and generates a family doctor recommendation list in descending order of matching degree.

[0056] The above embodiments illustrate preferred embodiments of the present invention. Any equivalent adjustments to the technical solution based on software engineering methods are within the scope of protection, including but not limited to: implementing algorithm logic using different programming languages, refactoring functional modules into services, adjusting data interaction protocols, and optimizing resource scheduling strategies. Any implementation scheme derived from reasonable modifications to the data processing flow, service call chain, or system architecture layer without departing from the core technology of the present invention should be considered within the protection scope defined by the technical solution of the present invention.

Claims

1. A family doctor service recommendation method based on user profiles, characterized in that, Includes the following steps: S1: Collect multimodal health monitoring data and historical medical evaluation records of residents within the target jurisdiction, and obtain basic practice information of family doctors through the medical institution management terminal; S2: Perform outlier cleaning and standardization on the multimodal health monitoring data, use natural language processing algorithms to perform disease entity recognition and sentiment analysis on the historical medical evaluation records, and construct a dynamic user profile including physiological indicator features and medical service demand features by combining time decay function. S3: Based on the basic practice information of the family doctor, establish a multi-dimensional evaluation model of the doctor's professional expertise, average daily number of visits, and past patient satisfaction, and configure a quantitative weight matrix corresponding to multiple practice indicators in the doctor's competency model. S4: Map the dynamic user profile to the feature space of the doctor's ability model, calculate the multidimensional weighted Euclidean distance between the resident health demand feature vector and the doctor service ability feature vector, determine the degree of supply and demand matching based on the distance, and generate a family doctor recommendation list in descending order of matching degree.

2. The family doctor service recommendation method based on user profiles according to claim 1, characterized in that, The specific steps of S1 are as follows: S11: Periodically extract blood pressure, blood sugar, heart rate, and step count data from smart wearable devices and home medical terminals worn by residents in the jurisdiction through the Internet of Things interface to obtain the multimodal health monitoring data; S12: Access the user-authorized historical medical records database through an encrypted transmission protocol, retrieve the resident's past outpatient medical records, prescription records, and satisfaction ratings for past medical services, and obtain the historical medical evaluation records. S13: Through the data interface of the medical institution management terminal, synchronously retrieve the registered department, professional title level, work schedule and professional skills certificate information of the practicing doctors to obtain the basic practice information of the family doctors.

3. The family doctor service recommendation method based on user profiles according to claim 2, characterized in that, The specific steps of S2 are as follows: S21: Use the interquartile range algorithm to identify and remove outliers in the multimodal health monitoring data, and use the maximum-minimum normalization method to map the physiological index data to the standard numerical range to generate a physiological health status vector. S22: Semantic features are extracted from the historical medical evaluation records using a bidirectional long short-term memory network model to identify disease entities and emotional keywords, and time decay coefficients are calculated by combining the timestamp information of residents' medical visits to generate a medical service demand vector. S23: Perform feature concatenation and dimensionality reduction on the physiological health status vector and the medical service demand vector to establish the dynamic user profile.

4. The family doctor service recommendation method based on user profiles according to claim 3, characterized in that, The specific steps in S3 are as follows: S31: Perform keyword aggregation analysis on the descriptions of diseases of expertise in the basic practice information of family doctors, generate professional expertise tags for doctors, and collect data on the number of patients and follow-up visits within a specified period to calculate the average daily patient load index. S32: Perform sentiment polarity analysis on past patients' evaluation texts, extract the proportion of positive evaluations and keywords related to service attitude, and combine them with the medical institution's annual assessment scores to generate a past patient satisfaction index. S33: Construct a multi-dimensional competency evaluation system by combining the professional expertise tags, the average daily patient load index, and the past patient satisfaction index. Calculate the information entropy value and coefficient of variation of multiple evaluation indicators using the entropy weight method. Establish objective weight values ​​for multiple professional competency indicators based on the degree of variation, and generate the quantitative weight matrix.

5. The family doctor service recommendation method based on user profiles according to claim 4, characterized in that, The specific steps of S4 are as follows: S41: Extract the demand feature dimension corresponding to the doctor's practice indicators from the dynamic user profile, and convert it into a resident health demand feature vector that is consistent with the doctor's competence model dimension; S42: Call the quantized weight matrix to calculate the difference between the resident health demand feature vector and the service capacity feature vector of each doctor in the candidate doctor database, and generate the multidimensional weighted Euclidean distance. S43: Sort the multidimensional weighted Euclidean distances in ascending order of value, select candidate doctors whose distance values ​​are less than a preset recommendation threshold, and establish a family doctor recommendation list.

6. The family doctor service recommendation method based on user profiles according to claim 3, characterized in that, The process by which the time decay function corrects the intensity of sentiment keywords includes: Obtain the timestamps and original emotional intensity values ​​corresponding to multiple entries in the historical medical evaluation record, and calculate the time difference between the current system time and the timestamps. Substitute the time difference into a preset exponential decay model, calculate the weighting coefficients for the differentiated historical periods, and use the weighting coefficients to weight and correct the original emotional intensity value to generate a real-time emotional feature value that can reflect the residents' recent psychological needs. The calculation formula for the exponential decay model is as follows: ； in, This represents the real-time weighting coefficients after time decay correction. Represents the initial weight baseline value. The time sensitivity constant representing the rate of decay. The timestamp value represents the current system time. The timestamp value representing when the historical evaluation record was generated.

7. The family doctor service recommendation method based on user profiles according to claim 4, characterized in that, The process of calculating the objective weight value includes: Build including Doctors and The original data matrix of the professional practice indicators is used to nonnegate the elements in the matrix, and the first indicator is calculated. The first item under the indicator The characteristic weighting value of each doctor; Calculate the first based on the aforementioned characteristic weight value The information entropy of each indicator is used to obtain the information utility value of the indicator. After normalizing the information utility value, an objective weight value is generated. The quantitative weight matrix is ​​established by combining the objective weight values ​​corresponding to all professional performance indicators.

8. The family doctor service recommendation method based on user profiles according to claim 5, characterized in that, The calculation process of the multidimensional weighted Euclidean distance includes: Obtain the resident health needs feature vector and the target doctor's service capacity feature vector, and calculate the numerical difference between the two in multiple feature dimensions; By introducing the weight coefficients in the quantization weight matrix, a weighted summation operation is performed on the square terms of the numerical differences, and the square root of the summation result is performed to generate a matching metric value. The formula for calculating the multidimensional weighted Euclidean distance is as follows: ； in, Representing residential users With family doctor The multidimensional weighted Euclidean distance values ​​between them Represents the total number of feature dimensions. Index representing feature dimension, Represents the quantization weight matrix of the th Weight coefficients for each feature dimension The feature vector representing the residents' health needs is in the th... Normalized values ​​of dimensions The feature vector representing the doctor's service capability is in the th... Normalized values ​​of dimensions.

9. The family doctor service recommendation method based on user profiles according to claim 5, characterized in that, The process of generating the family doctor recommendation list includes: The multidimensional weighted Euclidean distance is converted into a percentage matching score, and the candidate doctors are initially screened based on the matching score, retaining the preferred set of doctors whose scores are in the top 30%. Obtain the current number of signed contracts and service radius for each doctor in the preferred doctor set, remove doctors whose current number of signed contracts has reached the limit or whose service radius fails to cover the residents' residences, and generate a candidate set of doctors who can be signed contracts; The candidate set of doctors eligible for contracting is arranged in descending order of matching scores. The practice records and recommendation reasons of the top five doctors are extracted to generate the family doctor recommendation list.

10. A family doctor service recommendation system based on user profiles, characterized in that: The system is used to implement the family doctor service recommendation method based on user profiles as described in any one of claims 1-9, and the system includes: The multi-source data acquisition module is used to collect multimodal health monitoring data and historical medical evaluation records of residents within the target jurisdiction, and to obtain basic practice information of family doctors through the medical institution management terminal; The profile building and analysis module is used to clean and standardize outliers in the multimodal health monitoring data, use natural language processing algorithms to identify disease entities and analyze sentiment tendencies in the historical medical evaluation records, and build dynamic user profiles by combining time decay functions. The doctor model building module is used to build a multi-dimensional evaluation model of the doctor's professional expertise, average daily number of patients seen, and past patient satisfaction based on the basic practice information of the family doctor, and to configure a quantitative weight matrix corresponding to multiple practice indicators in the doctor's competency model. The intelligent recommendation calculation module is used to map the dynamic user profile to the feature space of the doctor's ability model, calculate the multidimensional weighted Euclidean distance between the resident health demand feature vector and the doctor service ability feature vector, determine the degree of supply and demand matching based on the distance, and generate a family doctor recommendation list in descending order of matching degree.

Images