Outpatient automatic triage method based on multi-modal data fusion

Abstract

The invention relates to the technical field of intelligent triage, in particular to an outpatient service automatic triage method based on multi-modal data fusion, which comprises the following steps of: acquiring the number of reception people, queue length and registration increment through a platform, inputting K-means clustering to evaluate reception load, calculating reception efficiency per unit time, extracting a chief complaint feature vector, and inputting the chief complaint feature vector into an SVM (Support Vector Machine) matching department. A matching degree and queue length weighted sorting path is combined, a dynamic threshold value and diversity constraints are introduced, the reception pressure and the emergency level are fused to adjust the priority, and a triage result is output. According to the method, the matching degree and balance are improved and the path flexibility and fault tolerance are optimized by fusing the three-dimensional pressure level and the sliding window calculation efficiency, extracting the symptom vector to optimize the matching precision, fusing the matching degree and the queue dynamic weight, introducing the diversity factor to expand the path and constructing the multi-dimensional evaluation model dynamic response; and the waiting stability, the resource utilization efficiency and the service quality are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent triage, and in particular to an automatic outpatient triage method based on multimodal data fusion. Background Art

[0002] The field of intelligent triage technology involves the use of information technology to guide patients in seeking medical treatment, conduct preliminary diagnosis and prioritization during the medical service process, aiming to improve the efficiency of medical treatment and the rationality of resource allocation. The core issues in this technical field include the collection of basic patient information, symptom identification and analysis, recommendation of medical departments, initial disease screening and urgency assessment, and are widely used in scenarios such as hospital outpatient management systems, intelligent consultation platforms, and mobile medical applications. The technological development in this field shows a trend of shifting from manual judgment to intelligent-assisted judgment, gradually forming a technical system based on the combination of multi-source information fusion and rule models, which puts higher requirements on the automation and precision of the triage process.

[0003] Among them, the traditional outpatient automatic triage method refers to the method of classifying and judging the patient's symptoms and giving a recommendation for a department or preliminary disease type through keyword matching, symptom classification rules or a fixed diagnosis and treatment pathway library after the patient arrives at the hospital or enters the chief complaint information at the front end of the online consultation. This method completes automatic triage based on a structured chief complaint template, a disease symptom dictionary, and a symptom-department mapping table. The technical issues it targets are department recommendations and disease type classification based on the patient's chief complaint information. The traditional method uses methods such as standardization of the chief complaint field, establishment of a static correspondence between symptoms and departments, and rule-based matching procedures to achieve chief complaint content analysis and triage recommendations.

[0004] Traditional methods rely on static symptom-department mapping tables and fixed rule bases. Department recommendations do not incorporate real-time reception pressure data, resulting in a disconnect between resource allocation and real-time load. The keyword matching mechanism cannot effectively analyze the semantic associations and complex symptom characteristics of the chief complaint text. The symptom classification rules lack the ability to integrate multi-dimensional features. The rigid triage path causes uneven distribution of waiting queues. The single urgency assessment dimension leads to priority sorting deviations. Static threshold settings are difficult to match the dynamic changes in the number of visits. The adjustment space for triage results is limited, which can easily lead to local overload or idleness of department resources. Patient waiting time fluctuates significantly due to the coverage of the rule base. Summary of the Invention

[0005] In order to solve the technical problems that the traditional method relies on static symptom department mapping tables and fixed rule bases, department recommendations are not included in real-time reception pressure data, resulting in a disconnect between resource allocation and real-time load, the keyword matching mechanism cannot effectively analyze the semantic association and complex symptom characteristics of the main complaint text, the symptom classification rules lack the ability to integrate multi-dimensional features, the triage path is solidified, resulting in uneven distribution of waiting queues, the single urgency assessment dimension leads to priority sorting deviation, the static threshold setting is difficult to match the dynamic changes in the number of visits, the adjustment space for triage results is limited, and it is easy to cause local overload or idleness of department resources, and the patient waiting time is significantly fluctuated due to the coverage of the rule base, the embodiment of the present invention provides an outpatient automatic triage method based on multimodal data fusion. The technical solution is as follows:

[0006] On the one hand, a method for automatic outpatient triage based on multimodal data fusion is provided, the method comprising:

[0007] S1: The hospital information platform is used to obtain the real-time number of patients, waiting queue length, and registration increment per unit time. These three parameters are input into the K-means clustering algorithm to perform three-dimensional spatial pressure level classification and generate the department's patient load assessment value.

[0008] S2: Based on the department's reception load assessment value, the registration increment per unit time and the real-time number of patients are called, and the number of patients completed per unit time is calculated using a sliding window mechanism to generate the department's real-time reception efficiency parameter;

[0009] S3: Call the real-time reception efficiency parameter of the department, perform natural language processing on the patient's chief complaint information, extract the patient's symptom feature vector, input it into the support vector machine model for kernel function mapping and logical matching, and generate a department matching score;

[0010] S4: Call the department matching score and the waiting queue length for weighted summation, introduce a dynamic score threshold adjustment mechanism and a path diversity constraint factor, combine the benchmark score ratio for priority sorting, and output a triage path candidate set;

[0011] S5: Based on the triage path candidate set, extract the candidate department reception pressure status and patient urgency level corresponding to the path, perform priority evaluation and dynamic adjustment through the weight fusion algorithm, and output the triage execution result.

[0012] As a further solution of the present invention, the K-means clustering algorithm divides the department's reception load and combines the patient's symptom feature vector to achieve a comprehensive assessment of the department's triage pressure and reception efficiency, thereby optimizing the selection of triage paths;

[0013] The department's reception load assessment value includes a real-time reception number threshold, a waiting queue warning interval, and a dynamic baseline for registration increments; the department's real-time reception efficiency parameters include the standard deviation of the number of patients received per unit time, resource occupancy rate, and response time change rate; the department matching score includes symptom vector similarity, keyword matching weight, and department professional matching; the triage path candidate set includes a priority weighted value, a dynamic threshold adjustment amount, and a path mutual exclusion identifier; the triage execution result specifically refers to the target department allocation code, the urgency grading coefficient, and the path selection confidence.

[0014] As a further solution of the present invention, the specific steps of S1 include:

[0015] S101: Obtain the number of patients, queue length, and registration increment data through the hospital information platform, uniformly process them according to time tags, and fit them into three-dimensional numerical vectors to obtain a three-dimensional vector dataset;

[0016] S102: Based on the three-dimensional vector data set, calling the K-means clustering algorithm to input three-dimensional vector samples, setting an initial center, classifying the samples according to the Euclidean distance and updating the center position, iterating until the center is stable, and obtaining the result of the patient load clustering;

[0017] S103: Based on the patient load clustering result and the corresponding three-dimensional features, the department cluster number is marked in the original vector and remapped into a grade value to generate a department patient load assessment value;

[0018] The department's reception load assessment value is obtained by quantifying three-dimensional features such as the number of patients received, the length of the waiting queue, and the increase in registrations into discrete levels through the K-means clustering algorithm, reflecting the department's reception pressure level in real time and assisting hospital management decisions.

[0019] As a further solution of the present invention, the specific steps of S2 include:

[0020] S201: Based on the department's patient load assessment value, the increment of registrations per unit time and the real-time number of patients are called, the sliding window boundary is delineated at fixed time intervals, the registration and patient data corresponding to the start and end times are recorded within the window, and the total number of registrations within the window is aggregated to generate a summary value of registrations per unit time;

[0021] S202: Calling the unit time registration summary value and the real-time number of patients received in the corresponding time period, performing a ratio conversion between the number of registrations and the number of patients received based on a unified sliding window time point, correcting the rate based on the window span, and generating a department reception rate sequence value;

[0022] The department's admission rate sequence value is a value generated in chronological order, reflecting the change in the ratio of the number of registrations to the number of admissions within a sliding window, and is used to measure the dynamic trend of the department's admission efficiency;

[0023] S203: Based on the department reception rate sequence value, the average of the sliding window reception rate is extracted, the cumulative processing is balanced with the window total, and the weighted rate sequence is converted to obtain the department real-time reception efficiency parameter.

[0024] As a further scheme of the present application, the specific steps of S3 include:

[0025] S301: The department reception efficiency parameter and patient complaint information are called, the keyword group is screened according to the string cosine similarity of the symptom label and the complaint text, the sentence element nesting vector is fitted combining the sentence element structure position in the sentence, the joint coding of the keyword and the structure is performed, and the symptom feature vector value is generated;

[0026] The symptom feature vector value represents the comprehensive symptom feature coding result of the patient complaint in the semantic and structural two layers;

[0027] S302: Based on the symptom feature vector value, the kernel function mapping mechanism of the support vector machine model is called, the inner product result of the vector and the training vector group is calculated, the mapping distance coefficient distribution is generated according to the distance relationship of the feature dimension linear combination value and the model interface;

[0028] The mapping distance coefficient is the distance quantization result of the similarity between the symptom feature vector after the support vector machine model and the label category in the high-dimensional space;

[0029] S303: According to the mapping distance coefficient distribution, the label category of the distance value is screened, the set reception efficiency parameter of the label in the training set is called, and the department matching degree score is generated through the weighted combination calculation result between the parameter and the matching label.

[0030] As a further scheme of the present application, the mapping distance coefficient is calculated, and the formula is adopted:

[0031]

[0032] Wherein, ζ ac represents the matching distance coefficient value between the a th sample of the symptom feature vector and the c th label category, γ ab represents the feature weight coefficient of the b th feature in the a th sample in the support vector projection space, ν cb represents the support vector numerical component of the c th label category in the b th feature dimension, λ cb represents the training model offset correction term of the b th dimension of the c th label category, η ab represents the structure similarity adjustment factor of the a th sample in the b th feature dimension, and ξ acrepresents the average symptom label alignment value of the a-th sample and the c-th label category calculated in the training stage, κ ac represents the label structure cross measure value of the a-th sample and the c-th label category in the current mapping stage, ∈ represents a constant to avoid zero denominator, and h represents the total number of feature dimensions of the symptom feature vector.

[0033] As a further scheme of the present application, the specific steps of S4 include:

[0034] S401: Based on the department matching degree score and the corresponding candidate queuing length, the average waiting time of patients under the department and the matching degree score value are extracted, the total weighted score is calculated according to the fixed weight, the score is sorted according to the size and the index position is recorded, and the weighted sorting priority value is generated;

[0035] The weighted sorting priority value is a comprehensive score calculated according to the fixed weight of the department matching degree, the queuing length and the average waiting time, etc., which is used to determine the priority of the patient in the current queuing system;

[0036] S402: According to the weighted sorting priority value, a dynamic score threshold adjustment mechanism is introduced, the upper and lower limits of the weighted score are judged in the score variation control interval, the adjustment offset is calculated according to the deviation median ratio, the sorting value is corrected, and the regulated sorting sequence is obtained;

[0037] The dynamic score threshold adjustment mechanism is a regulation method for dynamically correcting the sorting value according to the deviation degree of the weighted score and the median, and inhibiting the influence of extreme scores on the overall sorting;

[0038] S403: For the regulated sorting sequence, a path diversity constraint factor is introduced, the path combination independence and distribution uniformity are judged according to the queuing time density distribution and the path repetition distribution coefficient, the sorting is re-ordered according to the benchmark score proportion, and the triage path candidate set is obtained;

[0039] The diversity constraint factor dynamically adjusts the sorting result to realize the rationality of the triage path distribution by measuring the path repetition rate and the resource concentration.

[0040] As a further scheme of the present application, the adjustment offset is calculated by using the formula:

[0041]

[0042] Where ΔS x represents the sorting adjustment offset of the x-th patient, M represents the median of the weighted sorting priority value of the patients in the current sorting queue, n represents the total number of patients in the current sorting queue, W x represents the weighted sorting priority value of the x-th patient, L kThe waiting time density distribution value of the kth candidate path is in minutes / path, R k represents the path duplication coefficient of the kth candidate path, and z represents the total number of paths in the path distribution candidate set.

[0043] As a further solution of the present invention, the specific steps of S5 include:

[0044] S501: Based on the triage path candidate set, extract the candidate department admission data corresponding to each path and the number of patients in the current time period, calculate the ratio of the patient queue volume in the candidate department to the department's admission capacity value, and generate a admission pressure coefficient based on the admission capacity benchmark value;

[0045] The reception pressure coefficient reflects the ratio of the candidate department's current queue load to its reception capacity, and is used to measure its busyness;

[0046] S502: calling the patient reception pressure coefficient, comparing the urgency level of the patients associated with the path, and weighting and fusing the patient reception pressure with the level weight value to generate a path fusion priority value;

[0047] The path fusion priority value is a score derived from the patient's urgency and the path admission pressure, and is used to sort and select the optimal triage path;

[0048] S503: Sort the paths in descending order according to the path fusion priority value, select the top path and match it with the current queue update frequency, perform real-time correction on the path selection, and obtain the triage path execution result.

[0049] As a further solution of the present invention, the path fusion priority value is calculated using the formula:

[0050]

[0051] Among them, P i represents the fusion priority value of the i-th path, λ ij represents the admission pressure coefficient of the jth candidate department in the i-th path, ω ij represents the urgency level weight of the patients admitted by the jth candidate department in the i-th path, m represents the total number of candidate departments included in the path, α i represents the overall adjustment coefficient of the i-th path, Q i represents the average number of patients in the queue of the candidate department of the i-th path, in units of people, C i represents the average reception capacity of the candidate departments of the i-th path, in person / hour, β i is the capacity adjustment compensation value of the i-th path, in person / hour, δ i represents the urgency index value of the patient associated with the i-th path, γi is the fusion interference correction term of the i-th path.

[0052] The beneficial effects brought about by the technical solution provided by the embodiment of the present invention include at least:

[0053] By adopting three-dimensional spatial pressure level division combined with a sliding window mechanism to dynamically calculate department reception efficiency, extracting symptom feature vectors through natural language processing and applying kernel function mapping to optimize matching accuracy, dynamically weighted fusion of department matching and waiting queue is achieved, and a diversity constraint factor is introduced to expand the range of path selection. A multi-dimensional evaluation model is constructed based on real-time load and urgency. Triage decisions have dynamic response capabilities, the matching degree between department resource allocation and patient needs is improved, the distribution balance of waiting queues is enhanced, and it can match fluctuations in treatment pressure during differentiated time periods. The flexibility and fault tolerance of triage paths are optimized, the stability of patient waiting time is improved, and the efficiency of medical resource utilization and service quality are simultaneously improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 Schematic diagram of the workflow of the present invention. DETAILED DESCRIPTION

[0055] The technical solution of the present invention is described below in conjunction with the accompanying drawings.

[0056] In the embodiments of the present invention, words such as "exemplarily" and "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as an "exemplary" in the present invention should not be interpreted as being preferred or advantageous over other embodiments or designs. Rather, the use of the word "exemplary" is intended to present concepts in a concrete manner. Furthermore, in the embodiments of the present invention, "and / or" can mean both or either of the two.

[0057] In the embodiments of the present invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, when the distinction between them is not emphasized, the meanings they convey are the same. The terms "of," "corresponding," and "corresponding" may sometimes be used interchangeably. It should be noted that, when the distinction between them is not emphasized, the meanings they convey are the same.

[0058] In the embodiments of the present invention, sometimes a subscript such as W1 may be written as a non-subscript such as W1. When the difference is not emphasized, the meanings to be expressed are the same.

[0059] In order to make the technical problems, technical solutions and advantages to be solved by the present invention clearer, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0060] See also Figure 1 The embodiment of the present invention provides an outpatient automatic triage method based on multimodal data fusion. The processing flow of the method may include the following steps:

[0061] S1: The hospital information platform is used to obtain the real-time number of patients, waiting queue length, and registration increment per unit time. These three parameters are input into the K-means clustering algorithm to perform three-dimensional spatial pressure level classification and generate the department's patient load assessment value.

[0062] S2: Based on the department's reception load assessment value, the registration increment per unit time and the real-time number of patients are called, and a sliding window mechanism is used to calculate the number of patients completed per unit time to generate the department's real-time reception efficiency parameter;

[0063] S3: Call the department's real-time reception efficiency parameters, perform natural language processing on the patient's chief complaint information, extract the patient's symptom feature vector, input it into the support vector machine model for kernel function mapping and logical matching, and generate a department matching score;

[0064] S4: Call the department matching score and the waiting queue length for weighted summation, introduce a dynamic score threshold adjustment mechanism and path diversity constraint factor, combine the benchmark score ratio to perform priority sorting, and output the triage path candidate set;

[0065] S5: Based on the triage path candidate set, extract the candidate department reception pressure status and patient urgency level corresponding to the path, perform priority evaluation and dynamic adjustment through the weight fusion algorithm, and output the triage execution result.

[0066] The K-means clustering algorithm divides the department's reception load and combines it with the patient's symptom feature vector to achieve a comprehensive assessment of the department's triage pressure and reception efficiency, thereby optimizing the selection of triage paths;

[0067] The department's reception load assessment value includes the real-time reception number threshold, the waiting queue warning interval and the dynamic baseline of registration increment. The department's real-time reception efficiency parameters include the standard deviation of the number of patients received per unit time, resource occupancy rate and response time change rate. The department matching score includes symptom vector similarity, keyword matching weight and department professional matching. The triage path candidate set includes priority weighted value, dynamic threshold adjustment amount and path mutual exclusion identifier. The triage execution result specifically refers to the target department allocation code, urgency classification coefficient and path selection confidence.

[0068] Specifically, the steps of S1 are:

[0069] S101: Obtain the number of patients, queue length, and registration increment data through the hospital information platform, uniformly process them according to time tags, and fit them into three-dimensional numerical vectors to obtain a three-dimensional vector dataset;

[0070] The number of patients, queue length and registration increment data are obtained through the hospital information platform, which is integrated in the hospital HIS. First, the data is synchronized in time nodes, that is, the whole hour is set as a unified time label, and the three original data of each department are extracted at this time point. The number of patients is the number of patients treated by the doctor in the hour, such as 43 patients treated by the Department of Neurology from 10:00 to 11:00. The queue length is the number of patients registered in the waiting system at 10:55 but not called yet, such as 12 people waiting at that time. The registration increment refers to the increment of the number of registered patients between the previous hour and the current hour, for example, 45 people registered from 9:00 to 10:00, and 63 people registered from 10:00 to 11:00, so the registration increment is 18. These three data form a three-dimensional vector 【43, 12, 18】 according to the hour label. In the construction process, a set of three-dimensional vector data is generated every hour for each department. If you need to process the data of multiple departments for a whole day, for example, 8 departments and 10 hours, you will generate 80 sets of vectors. Combine the vectors into a three-dimensional vector data set, and normalize each set of data uniformly to map the three indicators to the [0, 1] interval. For example, if the maximum value of the registration increment is 40 and the minimum value is 0, the normalized value of the registration increment 18 is 18-0 / 40-0=0.45. After processing the three data, a standard three-dimensional vector is obtained, and a unified format data set is finally formed for subsequent clustering analysis.

[0071] Table 1: Outpatient data sample vector table

[0072]

[0073] As shown in Table 1, the number of patients, queue length, registration increment and other data are extracted according to the hour dimension and converted into three-dimensional vectors, and are uniformly normalized. The normalization uses linear interpolation method, and each indicator is calculated according to its maximum value and minimum value in the real-time range of the data set. For example, the maximum value of the normalized number of patients is 80, and the minimum value is 20. The normalized value of the number of patients is 43-20 / 80-20=0.3833. To simplify the explanation, the normalization interval of the normalized number of patients in the table is [0, 80], and other indicators such as queue length (maximum 20) and registration increment (maximum 40) are also processed accordingly. In real-time deployment, data is automatically extracted and written into an intermediate cache table every hour, and after uniform processing, it is written into the analysis module for subsequent K-means clustering step. The data set formed by this step has a standard format and continuity, providing a complete input data set for the next stage.

[0074] S102: Based on the three-dimensional vector data set, call the K-means clustering algorithm to input the three-dimensional vector samples, set the initial center, classify the samples according to the Euclidean distance and update the center position, iterate until the center is stable, and obtain the result of the patient load clustering;

[0075] Based on the three-dimensional vector data set formed after normalization, the vectors in the data set are first passed into the clustering process as input samples, and the predetermined number of initial center points k = 4 is set, which means that all vectors are divided into four categories. Then, four vectors are randomly selected from the data set as the initial center points, which are recorded as C1 = [0.5, 0.5, 0.5], C2 = [0.3, 0.2, 0.6], C3 = [0.8, 0.7, 0.7], and C4 = [0.6, 0.3, 0.2]. Then, the Euclidean distance calculation operation is performed on each vector sample. The calculation method of this operation is: for the vector X = [x1, x2, x3] and a certain center point C i =[c1, c2, c3], whose distance d i Calculated as:

[0076]

[0077] Taking the sample vector X = [0.5375, 0.6, 0.45] as an example, the distances between it and the four center points are calculated as follows:

[0078]

[0079] Therefore, it is determined that the sample vector belongs to cluster 1 with the smallest distance. The same operation is performed on the samples in sequence and the cluster numbers are assigned to them. After the first round of cluster assignment is completed, the average value of the vectors in each cluster is calculated to update the center point position. For example, there are three groups of vectors in cluster 1: [0.5375, 0.6, 0.45], [0.625, 0.75, 0.625], and [0.475, 0.45, 0.35]. The new cluster center is calculated as:

[0080]

[0081]

[0082] Then enter the next round of iteration, calculate the Euclidean distance of the sample vector based on the new center point again and reclassify it. After each round of execution, the center point is updated, and the above operation is repeated until the center point converges, that is, the difference between the new and old center points is lower than the set threshold. For example, if the threshold is set to 0.001, then when the change in the center of the cluster after update is less than this value, the clustering is terminated, and finally the clustering information of each group of samples is obtained, forming the cluster division result of the consultation load in this stage.

[0083] S103: Based on the patient load clustering result and the corresponding three-dimensional features, the department cluster number is marked in the original vector and remapped into a grade value to generate a department patient load assessment value;

[0084] The department's patient load assessment value is obtained by quantifying three-dimensional features such as the number of patients received, waiting queue length, and registration increment into discrete levels through the K-means clustering algorithm. This reflects the department's patient load level in real time and assists hospital management decision-making.

[0085] According to the clustering results corresponding to each sample vector of the above-mentioned completed clustering, the original data index is traced back and numbered. The department and time point corresponding to each vector are clearly defined. For example, the sample vector [0.5375, 0.6, 0.45] belongs to the 10:00 period of the Department of Neurology and belongs to cluster number 1. The load level of the Department of Neurology during this period is marked as 1. The cluster number is then remapped to the reception load level. The mapping method uses numerical level rearrangement. For example, the reception load of the four cluster centers is calculated as follows: the weighted average of the three indicators in the cluster center is used as the load indicator, and the weights are set as w1=0.5, w2=0.3, and w3=0.2, corresponding to the number of patients, queue length, and registration increment. Then a certain center point C i =[x1, x2, x3] is calculated as:

[0086] L i =w1·x1+w2·x2+w3·x3;

[0087] For example, if the center C1 = [0.5458, 0.6, 0.475], its load value is:

[0088] L1=0.5·0.5458+0.3·0.6+0.2·0.475=0.2729+0.18+0.095=0.5479;

[0089] The load values ​​of the cluster centers are calculated accordingly and sorted by size. The smallest load value is mapped to level 1, and the largest load value is mapped to level 4. For example, the center load value is sorted as follows:

[0090] L3=0.435 <L2=0.496<L1=0.5479<L4=0.589;

[0091] They are mapped to levels 1 to 4 respectively. The records in the sample that originally belong to cluster 3 are marked as level 1, and the records that belong to cluster 4 are marked as level 4. In this way, the level value remapping of the entire sample is completed, and the department reception load assessment value of each group of data is obtained. After the level value of each record is updated, it is written into the structured output result, which serves as the input basis for subsequent load prediction and triage strategy adjustment.

[0092] Specifically, the steps of S2 are:

[0093] S201: Based on the department's patient load assessment value, the registration increment per unit time and the real-time number of patients are called, the sliding window boundary is delineated at fixed time intervals, the registration and patient data corresponding to the start and end times are recorded within the window, and the total number of registrations within the window is aggregated to generate a registration summary value per unit time;

[0094] Based on the evaluation value of the department's reception load, the registration increment at each time point and the real-time number of patients in the corresponding time period are called separately, and the sliding window boundary is set according to a fixed time interval. The sliding window length is set to 30 minutes, and the sliding step is 10 minutes each time. In this way, multiple sliding window sequences are formed within a day, and each window covers the data interval of a specific time period. In each window, the registration data index position corresponding to the start time and the end time is first determined, and the registration records in this time period are extracted and the number of registrations is counted by minutes. The total number of registrations in this period is summed up to obtain the total number of registrations in the window. For example, in the time window from 08:00 to 08:30, the registration data is [5, 4, 6, 5, 7, 5] (recorded every 5 minutes). Record), then the total registration value in the window is 5+4+6+5+7+5=32 people. For the registration data collection process, the front-end registration system needs to record the incremental change value once a minute. This value is the registration increment per unit time. The window-level summary value is obtained by adding the incremental data in consecutive time periods. In addition, the real-time number of patients received is dynamically updated data collected once a minute. The number of patients who have entered the reception process after registration in multiple time periods recorded by the hospital HIS system can be obtained through interface calls. For example, the number of patients received per minute in a certain window is recorded as [2, 3, 2, 4, 3, 3], then the total number of patients received in the window is 2+3+2+4+3+3=17 people. Align the registration and reception data to the same window and enter them into the table below simultaneously.

[0095] Table 2: Department registration and reception data table

[0096]

[0097] As shown in Table 2, each window fully records the registration increment sequence and the number of patients received. The total number of registrations at the window level is obtained through summary calculation, providing the original input data for the subsequent calculation of the department's reception rate.

[0098] S202: Call the total number of registrations per unit time and the real-time number of patients received in the corresponding time period, perform a ratio conversion between the number of registrations and the number of patients received based on a unified sliding window time point, modify the rate based on the window span, and generate a series value of the department's patient reception rate;

[0099] The department reception rate sequence value is a value generated in time sequence, reflecting the change of the ratio of the number of registrations to the number of receptions in the sliding window, and is used to measure the dynamic change trend of the reception efficiency of the department;

[0100] The unit time registration summary value and the real-time reception number data in the corresponding time period are paired one by one according to the unified sliding window time point, and a numerical division calculation operation is performed to obtain the conversion ratio between the number of registrations and the number of receptions in the window. In this process, the correction of the window span ratio value needs to be considered. For example, if the window time span is 30 minutes, the converted ratio needs to be standardized per hour, that is, the ratio is multiplied by 2 to convert to the hourly rate value. For example, if the total number of registrations in a window is 32 and the total number of receptions is 17, the original ratio is 32 / 17≈1.882, and after multiplying by the correction coefficient 2, the standardized result is 3.764. This value is the standardized reception rate value in the 08:00-08:30 window. In this way, the total number of registrations and the total number of receptions in the time window are divided one by one to form a rate value sequence after being multiplied by the standardized correction coefficient. The sequence is arranged in time sequence, reflecting the change process of the reception efficiency per unit time in the period. In use, new registrations and reception data need to be collected every minute, the content of the sliding window is updated in real time, and the rate is recalculated. The sequence formed by the above two window results is [3.764, 3.529], which shows that the reception rate is gradually decreasing. The dynamic change trend of the reception pressure can be judged through this sequence.

[0101] S203: Based on the department reception rate sequence value, the average of the reception rate in the sliding window is extracted, which is converted to a weighted rate sequence through accumulation and window total balancing to obtain the real-time reception efficiency parameter of the department;

[0102] Based on the department's reception rate sequence value, the window rate value under the sliding window is continuously added up to calculate the total cumulative reception rate. The total is divided by the number of windows to form a weighted average rate value as the real-time reception efficiency parameter of the current time period. In the specific implementation, the reception rate sequence value in the window is aggregated with a fixed time span and smoothed by the weighted average of the number. For example, at the current time 09:00, the rate sequences of the last three sliding windows are [3.764, 3.529, 3.875], and the total sum is 3.764+3.529+3.875. =11.168, the weighted average value is 11.168 / 3≈3.723, which is the weighted reception efficiency value corresponding to the current moment. This value will be compared with the original average value to determine whether the current reception efficiency is normal, high pressure or overloaded. Different level intervals need to be set in implementation for classification. For example, the efficiency value below 2.5 is the low efficiency interval, 2.5 to 3.5 is the normal interval, 3.5 to 4.5 is the tense interval, and more than 4.5 is the overload interval. The current 3.723 falls into the tense interval, so the response strategy can be updated in real time according to the reception efficiency value.

[0103] Specifically, the steps of S3 are:

[0104] S301: Recall department reception efficiency parameters and patient chief complaint information, select keyword groups based on the cosine similarity between symptom labels and the chief complaint text, fit sentence element nesting vectors based on the sentence element structure position in the sentence, perform joint encoding of keywords and structures, and generate symptom feature vector values;

[0105] The symptom feature vector value represents the comprehensive symptom feature coding result of the patient's complaint at both the semantic and structural levels;

[0106] The department reception efficiency parameter and the patient complaint information are called. First, the "department reception efficiency parameter" is disassembled, and its source is the current time period reception efficiency value calculated according to the sliding window reception rate sequence in the foregoing. For example, when the current time is 09:00, the three window reception rates are 3.764, 3.529, and 3.875, and the weighted average is (3.764+3.529+3.875) / 3=3.723, which is the current department reception efficiency parameter. Then, the "patient complaint information" is refined, which is specifically composed of the chief complaint text input by the patient when registering, such as "cough for three days with fever" and "chest tightness with shortness of breath and sore throat". First, the text is segmented, keywords are extracted, and the position index of each keyword in the sentence is labeled. For example, "cough", "fever", and "three days" are located at positions 1, 2, and 3, respectively. Next, the cosine similarity between each keyword and the standard symptom label is calculated. The value ranges from 0 to 1. The closer the cosine value is to 1, the more similar the complaint is to the label. For example, the similarity between the complaint "cough for three days with fever" and the label "fever" is 0.89, and the similarity with the label "sore throat" is only 0.12. Therefore, "fever" is one of the important labels in the keyword group ["cough", "fever"]. If the keyword group is ["cough", "fever"], then the position structure index is normalized with the original complaint text length to construct a sentence position nested vector. The length of the complaint text is defined as L, and a keyword is located at the i-th word, so the position in the nested vector is i / L. For example, the complaint length is 10, and "cough" is at the 2nd position, which represents a vector position of 0.2. The keyword vector and the position vector are concatenated into the final structure encoding to form a complaint structure nested matrix. Each keyword nested matrix is then linearly mapped and combined to generate a fixed-dimensional symptom feature vector. For example, if the feature dimension is set to 6, the output is [0.88, 0.2, 0.76, 0.4, 0.12, 0.3], which includes the joint encoding results of semantic similarity and structure position. Finally, the vector is used as the symptom feature expression of the current patient complaint in the multi-dimensional space for subsequent classification mapping operations.

[0107] Table 3: Symptom feature vector generation example table

[0108]

[0109] As shown in Table 3, different complaint texts can generate structured symptom feature vectors in standard format after keyword screening and position information normalization.

[0110] S302: Based on the symptom feature vector value, the kernel function mapping mechanism of the support vector machine model is called to calculate the inner product result of the vector and the training vector group. According to the distance relationship between the linear combination value and the model interface, a mapping distance coefficient distribution is generated.

[0111] The mapping distance coefficient is the quantitative result of the distance between the symptom feature vector and the label category in high-dimensional space after the symptom feature vector passes through the support vector machine model;

[0112] Based on the generated symptom feature vector value, the support vector machine (SVM) model is first called to determine the degree of match between the vector and multiple label categories through its internal Gaussian radial basis kernel function mechanism. In this process, each training sample and its corresponding label category are traversed, and the current symptom vector is compared and matched with the sample vectors in the training set in terms of feature dimensions. The mapping distance coefficient is calculated using the formula:

[0113]

[0114] Among them, ac Represents the matching distance coefficient value between the a-th sample of the symptom feature vector and the c-th label category, γ ab represents the feature weight coefficient of the bth feature in the ath sample in the support vector projection space, ν cb represents the numerical component of the support vector of the cth label category on the bth feature dimension, λ cb represents the training model offset correction term under the bth dimension of the cth label category, η ab represents the structural similarity adjustment factor of the a-th sample on the b-th feature dimension, ξ ac represents the average symptom label alignment value calculated between the a-th sample and the c-th label category during the training phase, κ ac It represents the label structure cross-measure value between the a-th sample and the c-th label category in the current mapping stage, ∈ represents a constant to avoid the denominator being zero, and h represents the total number of feature dimensions of the symptom feature vector.

[0115] Specifically, the difference in each dimension of the feature vector between the current vector and each labeled sample in the training set is calculated to measure their proximity in high-dimensional space. To more accurately assess the match, an adjustment factor representing structural differences is introduced to strengthen the sensitive dimension in feature matching while also providing a certain degree of tolerance when the vector structures are not completely consistent. Furthermore, the structural alignment parameter formed during the original training phase is introduced to further determine the stability of the match between symptoms and labels by comparing the current input with the original distribution.

[0116] For example, let's assume the sample dimension is 3. Now calculate the matching distance coefficient between the first sample and the second label. Assume the participation values ​​are as follows:

[0117] γ a1 =0.29,γ a2 =0.47,γ a3 =0.10;

[0118] ν c1 =0.30,ν c2 =0.45,ν c3 =0.12;

[0119] λ c1 =0.01,λ c2 =0.02,λ c3 =0.01;

[0120] η a1 =0.93,η a2 =0.88,η a3 =0.90;

[0121] ξ ac =0.92,κ ac =0.89,∈=0.01;

[0122] The numerator is calculated as follows:

[0123] (0.29 0.30-0.01)+(0.47 0.45-0.02)+(0.10 0.12-0.01)=(0.087-0.01)+(0.2115-0.02)+(0.012-0.01)=0.2705;

[0124] The first term in the denominator:

[0125]

[0126] The second term in the denominator: |ξ ac -κ ac |+∈=|0.92-0.89|+0.01=0.04;

[0127] Final denominator:

[0128] 0.5337+0.04=0.5737;

[0129] The matching distance coefficient value is calculated as:

[0130]

[0131] This value represents the distance coefficient between the first symptom feature vector sample and the second label category in the high-dimensional support vector machine kernel space. The smaller the value, the closer the mapping space distance and the closer the proximity. When the value is greater than 1.0, it indicates that the main complaint symptom is far away from the support boundary constructed by the label category in the training set. The innovation in the formula is to use the structural similarity factor η ab The square of the label component The joint participation in the denominator construction effectively regulates the interference degree of structural differences on label matching calculations, and at the same time introduces two independent stages of structural alignment and the current structure cross term ξ ac , κ ac , which can enhance the response sensitivity of structural stability assessment and ultimately form a measurable label matching distance result through composite weighted processing between multiple parameters.

[0132] S303: Based on the distribution of the mapped distance coefficients, the label categories of the distance values ​​are screened, the reception efficiency parameters set for the labels in the training set are called, and the weighted combination calculation results between the parameters and the matching labels are used to generate the department matching score;

[0133] According to the matching distance coefficient ζ generated above ac , the distance values ​​corresponding to the label categories are used to form a set of coefficient distribution lists, and then the label categories are sorted according to the value. The label categories with a distance value less than 0.65 are selected to enter the subsequent screening process, and the label matching score is calculated in combination with the reception efficiency parameter. First, the reception efficiency parameter value recorded by the label in the training set is extracted. The current time point corresponds to "cardiology" as 3.723, "respiratory department" as 2.918, and "neurology" as 2.311. When the distance value between the chief complaint and the "chest pain" label is 0.4713, and the distance from the "shortness of breath" label is 0.628, the matching score calculation process is established for the cardiology department and the respiratory department respectively, and the scoring function is called. First, the inverse proportion of the distance value is taken as the label matching strength, for example, matching strength = 1-ζ ac , the intensity of "chest pain" = 1 - 0.4713 = 0.5287, and the intensity of "shortness of breath" = 1 - 0.628 = 0.372. This intensity is then multiplied by the corresponding label's department efficiency parameter to obtain the initial matching score. The score for chest pain matching the cardiology department is 3.723 × 0.5287 ≈ 1.968, and the score for shortness of breath matching the respiratory department is 2.918 × 0.372 ≈ 1.086. These scores are normalized to form a final matching score list. The highest-scoring candidate is selected as the target department label for this round of recommendations. In this example, cardiology is recommended for this symptom because its score of 1.968 is higher than other candidates. In the above calculation, the "distance threshold of 0.65" is set based on the distance distribution density of at least 90% of valid symptom label matches in the training set. That is, when the distance exceeds 0.65, most labels cannot be accurately matched. Therefore, this value serves as a screening threshold and can be fine-tuned based on sample training results. This matching process implements a fusion scoring mechanism between symptom vector features and the reception efficiency parameters of the label category, providing a quantitative basis for the output of the final department label.

[0134] Specifically, the steps of S4 are:

[0135] S401: Based on the department matching degree score and the corresponding queue length, the average waiting time and the matching degree score value of the patients under the department are extracted, the total weighted score is calculated according to the fixed weight, the score is sorted and the index position is recorded, and the weighted sorting priority value is generated;

[0136] The weighted sorting priority value is a comprehensive score calculated by the department matching degree, the length of the waiting queue and the average waiting time according to the fixed weight, which is used to determine the priority of the patient in the current queuing system;

[0137] Based on the department matching degree score and the corresponding queue length, the matching degree score value of each candidate department in the current time period is first called, which is calculated from the product of the patient's complaint vector and the label category in the foregoing matching process, for example, the patient's complaint is "chest pain", and the matching degree score value is 1.968. At the same time, the length of the waiting queue information of each candidate department is extracted, for example, the number of people waiting in the cardiology department is 12, the number of people waiting in the respiratory department is 9, and the number of people waiting in the neurology department is 15. The average waiting time of patients generated according to the records of the past thirty minutes is extracted, which is 8.2 minutes, 10.5 minutes and 13.6 minutes for the cardiology department, the respiratory department and the neurology department respectively. After calling the above three types of data, the weighted score calculation operation is performed in turn, wherein the matching degree score value, the length of the waiting queue and the average waiting time are respectively set to fixed weight coefficients of 0.5, 0.3 and 0.2. The multiple numerical values are normalized to the interval of 0-1, for example, the maximum matching degree is 1.968, and the minimum is 1.012. The chest pain matching degree normalized value is (1.968-1.012) / (1.968-1.012)=1.0. If the matching degree of the neurology department is 1.450, the normalized value is (1.450-1.012) / 0.956≈0.458. The maximum number of people waiting is 15, and the minimum is 9. The cardiology department is (15-12) / 6=0.5, and the neurology department is (15-15) / 6=0. The maximum average waiting time is 13.6, and the minimum is 8.2. The cardiology department waiting time is normalized to (13.6-8.2) / 5.4≈1.0, and the neurology department is (13.6-13.6) / 5.4=0. After normalization, the weighted score is calculated by multiplying the corresponding weight coefficients, for example, the weighted score of the cardiology department is 1.0×0.5+0.5×0.3+1.0×0.2=0.5+0.15+0.2=0.85, and the weighted score of the neurology department is 0.458×0.5+0×0.3+0×0.2=0.229. The candidate departments are sorted according to the weighted score, and the sorting index value is recorded to form the sorting priority value list.

[0138] Table 4: Weighted sorting calculation parameter table

[0139]

[0140]

[0141] As shown in Table 4 , cardiology department ranked first, respiratory department ranked second, and neurology department ranked third for subsequent patient queue processing.

[0142] S402: Based on the weighted ranking priority value, a dynamic scoring threshold adjustment mechanism is introduced to determine the upper and lower limits of the weighted score within the score change control range. The adjustment offset is calculated based on the deviation ratio from the median value, and the ranking value is corrected to obtain the adjusted ranking sequence;

[0143] According to the weighted sorting priority value list, a dynamic scoring threshold adjustment mechanism is introduced to measure the offset between each patient's weighted score and the median value of the current sorting sequence, and calculate the adjustment offset using the formula:

[0144]

[0145] Where, ΔS x represents the sort adjustment offset of the xth patient, M represents the median of the weighted sort priority values ​​of the patients in the current sorting queue, n represents the total number of patients in the current sorting queue, and the unit is person, W x represents the weighted ranking priority value of the xth patient, L k The waiting time density distribution value of the kth candidate path is in minutes / path, R k represents the path duplication coefficient of the kth candidate path, and z represents the total number of paths in the path distribution candidate set.

[0146] First, determine the current queue weighted score value set. For example, if the ranking priority values ​​of 5 patients are [0.85, 0.729, 0.589, 0.423, 0.229], then the median M is the third value after sorting, that is, M = 0.589. Calculate the square of the difference between each patient's weighted value and the median, which is (0.85-0.589) 2 =0.068, (0.729-0.589) 2 =0.0196, (0.423-0.589) 2 =0.0276, (0.229-0.589) 2= 0.1296, and the average value thereof is (0.068 + 0.0196 + 0.0276 + 0.1296) / 4 = 0.0617, and the current standard deviation is 0.2484 obtained by taking the square root thereof, and the weighted ranking value of the first patient is 0.85, and the offset amplitude value = |0.85-0.589| / 0.2484 = 1.05; and then, according to the path density and the repetition distribution factor set, for example, there are 3 paths in the candidate path, the queuing time density is [2.4, 1.8, 2.1] minutes / path, and the path repetition distribution coefficient is [0.8, 0.5, 0.9], and each product is 1.92, 0.9, 1.89, and the total sum is 4.71, and the average value is 4.71 / 3 = 1.57, and the final offset adjustment amount AS is calculated x = 1.05 x 1.57 = 1.6485, and the adjusted ranking value is the original ranking value ± AS, and the higher value is adjusted downward and the lower value is adjusted upward according to the ranking direction, and the values are uniformly close to the median value, for example, the first patient is adjusted downward from 0.85 to 0.85-0.1648 = 0.685, and the new ranking sequence is generated by re-ranking the ranking values, thereby ensuring the balanced distribution of the ranking results.

[0147] S403: For the adjusted ranking sequence, a path diversity constraint factor is introduced, the path combination independence and distribution balance are determined according to the queuing time density distribution and the path repetition distribution coefficient, the candidate set of triage paths is obtained by re-ranking according to the benchmark score proportion.

[0148] On the basis of the adjusted ranking sequence, the path diversity constraint factor is introduced to determine the queuing path combination, first, the independence of each path in the candidate path set is determined, the path independence score is calculated according to the queuing time density distribution value and the path repetition distribution coefficient, for example, the queuing time density of path A is 2.1, and the repetition distribution coefficient is 0.6, and the path independence score is 2.1 x (1-0.6) = 0.84, and the score is calculated by traversing the path, for example, the path independence score of path A is 0.84, the path independence score of path B is 1.5 x (1-0.5) = 0.75, and the path independence score of path C is 1.9 x (1-0.8) = 0.38, and the average value of the path independence score is the benchmark score proportion, for example, the average value of this time is (0.84 + 0.75 + 0.38) / 3 = 0.6567, and then the score of each path is compared with the benchmark value, if the score is higher than the value, it is determined as a high independence path, and is marked as a priority candidate path, and then the number of high independence paths included in each candidate path combination is weighted and ranked, for example, path combination 1 includes paths A and B, and the average score is (0.84 + 0.75) / 2 = 0.795, path combination 2 includes paths B and C, and the average is (0.75 + 0.38) / 2 = 0.565, and path combination 3 only contains path A, and the score is 0.84, and the path combination scores are ranked from high to low, and finally the ranking result of the candidate set of triage paths is formed.

[0149] Specifically, the steps of S5 are as follows:

[0150] S501: Based on the candidate triage path set, the candidate department reception data corresponding to each path and the number of patients in the current time period are extracted, the ratio of the patient queue volume to the department reception capacity value in the candidate department is calculated, and the reception pressure coefficient is generated according to the reception capacity reference value;

[0151] The reception pressure coefficient reflects the ratio of the current queue load to the reception capacity of the candidate department, and is used to measure the degree of busyness thereof;

[0152] Based on the candidate triage path set, the candidate departments associated with each path are first extracted, and the reception data of the departments in the current time period are obtained in turn, including the current queue number and the reception capacity per unit time of each department. The queue number can be obtained in real time through the outpatient registration system, for example, the number of patients waiting for reception in A department is 25, in B department is 18, and in C department is 33. The reception capacity can be obtained through original statistical data, for example, the daily reception capacity of A department is 120 people, which is converted to 15 people / hour according to an 8-hour working day, the reception capacity of B department is 10 people / hour, and the reception capacity of C department is 20 people / hour. Then, the queue ratio of each department is calculated, i.e. the current queue number divided by the reception capacity, for example, the queue ratio of A department is 25÷15≈1.67, the queue ratio of B department is 1.80, and the queue ratio of C department is 1.65. Then, the reception pressure coefficient of the department is generated by taking the reception capacity reference value (set to 1.50, representing a medium load state) as a comparison standard, for example, the absolute value of the difference between the ratio and the reference value is taken as the basic quantity of the pressure factor, and then multiplied by a calibration coefficient μ (μ takes 1.2) for conversion, for example, the pressure coefficient of A department is |1.67-1.50|×1.2≈0.204, the pressure coefficient of B department is 0.36, and the pressure coefficient of C department is 0.18. The larger the value is, the higher the reception pressure of the department is. If the pressure coefficient interval is divided into: low pressure zone (<0.2), medium pressure zone (0.2-0.4), and high pressure zone (>0.4), then A and C are in the medium pressure zone, and B is in the high pressure zone. The above operation is repeated for each department in the candidate path and the results are recorded, and finally the reception pressure coefficient of the candidate department in each path is generated.

[0153] S502: Call the reception pressure coefficient, compare the emergency level of the patients associated with the path, and fuse the reception pressure and the grade weight value to generate a path fusion priority value;

[0154] The path fusion priority value is the score obtained by the patient emergency level and the path reception pressure, and is used to sort and select the optimal triage path;

[0155] Call the aforementioned array of patient pressure coefficients. For each candidate path, first call its corresponding patient pressure coefficient λ according to the department sequence. ij , combined with the urgency level of each patient in the pathway, the urgency level adopts a five-level scoring system, with level 1 being the most urgent and level 5 being the least urgent, and weights ω are assigned to each. ij For example, path 1 includes three departments, and the emergency levels of the patients received are 2, 3, and 1, respectively. The pressure coefficients of the patients received are 0.20, 0.36, and 0.18, respectively. The sum of this part of the path is: 0.20×4+0.36×3+0.18×5=0.80+1.08+0.90=2.78. Then, according to the overall parameters of the path, the queue volume Q of the path is obtained. i , is the average number of patients queuing in the department within the path, assuming there are 27 people, and the admission capacity value C i is the average reception capacity of the department within the path, assuming 17 people / hour, and the capacity compensation value β i Assuming 3 people / hour (taken from the medical resource deviation compensation standard), the patient's comprehensive emergency value δ i It is obtained by weighted average of the disease score. For example, the average score of patients associated with the current path is 2.2, the corresponding emergency value is 3.6, and the path correction coefficient α i Set to 0.8, fusion interference correction term γ i Set to 0.5.

[0156] Calculate the path fusion priority value using the formula:

[0157]

[0158] Among them, P i represents the fusion priority value of the i-th path, λ ij represents the admission pressure coefficient of the jth candidate department in the i-th path, ω ij represents the urgency level weight of the patients admitted by the jth candidate department in the i-th path, m represents the total number of candidate departments included in the path, α i represents the overall adjustment coefficient of the i-th path, Q i represents the average number of patients in the queue of the candidate department of the i-th path, in units of people, C i represents the average reception capacity of the candidate departments of the i-th path, in person / hour, β i is the capacity adjustment compensation value of the i-th path, in person / hour, δ i represents the urgency index value of the patient associated with the i-th path, γ i is the fusion interference correction term of the i-th path.

[0159] The calculation process of the path fusion priority value is as follows:

[0160]

[0161] Table 5: Path fusion priority calculation parameter table

[0162]

[0163] As shown in Table 5, the calculation of the path fusion priority value involves cross-referencing and calculation of multiple quantitative parameters, and the specific value selection needs to be based on the original load data of the department and the real-time evaluation grade of the patient.

[0164] S503: According to the path fusion priority value, the paths are sorted in descending order, the first path is selected and matched with the current queuing update frequency, the real-time correction of path selection is performed, and the triage path execution result is obtained.

[0165] According to the path fusion priority value P calculated in the above steps i , the candidate paths are sorted in descending order, for example, the priority value of path 1 is 2.65, the priority value of path 2 is 2.48, and the priority value of path 3 is 2.20, then the sorting result is path 1> path 2> path 3, then the first path 1 is taken as the current triage path initial option, then the set queuing information update time interval parameter is called, which is set to 5 minutes, it is judged whether the current data update time exceeds the threshold value, for example, the current time is 11:36, the last update time is 11:30, and the interval is 6 minutes, which exceeds the preset threshold value, then the queuing number and the reception capacity data of the department included in path 1 are refreshed in real time, and the calculation process of the path fusion priority value is re-executed, if the priority value fluctuates after updating, for example, the new value decreases to 2.30, which is lower than the priority value of path 2, which is 2.48, then path 2 rises to the first option after re-sorting, and finally the latest sorting result is used to determine the path execution option, forming the final execution result of the triage path.

[0166] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. An automatic outpatient triage method based on multimodal data fusion, characterized in that: The following steps are involved: S1: The hospital information platform is used to obtain the real-time number of patients, waiting queue length, and registration increment per unit time. These three parameters are input into the K-means clustering algorithm to perform three-dimensional spatial pressure level classification and generate the department's patient load assessment value. S2: Based on the department's reception load assessment value, the registration increment per unit time and the real-time number of patients are called, and the number of patients completed per unit time is calculated using a sliding window mechanism to generate the department's real-time reception efficiency parameter; S3: Call the real-time reception efficiency parameter of the department, perform natural language processing on the patient's chief complaint information, extract the patient's symptom feature vector, input it into the support vector machine model for kernel function mapping and logical matching, and generate a department matching score; S4: Call the department matching score and the waiting queue length for weighted summation, introduce a dynamic score threshold adjustment mechanism and a path diversity constraint factor, combine the benchmark score ratio for priority sorting, and output a triage path candidate set; S5: Based on the triage path candidate set, extract the candidate department reception pressure status and patient urgency level corresponding to the path, perform priority evaluation and dynamic adjustment through the weight fusion algorithm, and output the triage execution result.

2. The outpatient automatic triage method based on multimodal data fusion according to claim 1 is characterized in that: The department's reception load assessment value includes a real-time reception number threshold, a waiting queue warning interval, and a dynamic baseline for registration increments; the department's real-time reception efficiency parameters include the standard deviation of the number of patients received per unit time, resource occupancy rate, and response time change rate; the department matching score includes symptom vector similarity, keyword matching weight, and department professional matching; the triage path candidate set includes a priority weighted value, a dynamic threshold adjustment amount, and a path mutual exclusion identifier; the triage execution result specifically refers to the target department allocation code, the urgency grading coefficient, and the path selection confidence.

3. The outpatient automatic triage method based on multimodal data fusion according to claim 1 is characterized in that: The specific steps of S1 include: S101: Obtain the number of patients, queue length, and registration increment data through the hospital information platform, uniformly process them according to time tags, and fit them into three-dimensional numerical vectors to obtain a three-dimensional vector dataset; S102: Based on the three-dimensional vector data set, calling the K-means clustering algorithm to input three-dimensional vector samples, setting an initial center, classifying the samples according to the Euclidean distance and updating the center position, iterating until the center is stable, and obtaining the result of the patient load clustering; S103: According to the reception load cluster division result and the corresponding three-dimensional features, the department cluster number is marked in the original vector and remapped into a grade value to generate a department reception load evaluation value.

4. The outpatient automatic triage method based on multimodal data fusion according to claim 3 is characterized in that: The specific steps of S2 include: S201: Based on the department's patient load assessment value, the increment of registrations per unit time and the real-time number of patients are called, the sliding window boundary is delineated at fixed time intervals, the registration and patient data corresponding to the start and end times are recorded within the window, and the total number of registrations within the window is aggregated to generate a summary value of registrations per unit time; S202: Calling the unit time registration summary value and the real-time number of patients received in the corresponding time period, performing a ratio conversion between the number of registrations and the number of patients received based on a unified sliding window time point, correcting the rate based on the window span, and generating a department reception rate sequence value; S203: Based on the department's reception rate sequence value, perform average extraction on the reception rate under the sliding window, convert it into a weighted rate sequence through accumulation processing and window total balance, and obtain the department's real-time reception efficiency parameter.

5. The outpatient automatic triage method based on multimodal data fusion according to claim 4 is characterized in that: The specific steps of S3 include: S301: Recall the department's reception efficiency parameter and the patient's chief complaint information, select keyword groups based on the cosine similarity between the symptom label and the chief complaint text, fit the sentence element nesting vector based on the sentence element structure position in the sentence, perform joint encoding of the keyword and structure, and generate a symptom feature vector value; S302: Based on the symptom feature vector value, calling the kernel function mapping mechanism of the support vector machine model, calculating the inner product result of the vector and the training vector group, and generating a mapping distance coefficient distribution based on the distance relationship between the linear combination value of the feature dimension and the model interface; S303: According to the mapping distance coefficient distribution, the label category of the distance value is screened, the reception efficiency parameter set for the label in the training set is called, and the weighted combination calculation result between the parameter and the matching label is calculated to generate a department matching score.

6. The outpatient automatic triage method based on multimodal data fusion according to claim 5 is characterized in that: The mapping distance coefficient is calculated using the formula: Among them, ac Represents the matching distance coefficient value between the ath sample of the symptom feature vector and the cth label category, γ ab represents the feature weight coefficient of the bth feature in the ath sample in the support vector projection space, ν cb represents the numerical component of the support vector of the cth label category on the bth feature dimension, λ cb represents the training model offset correction term under the bth dimension of the cth label category, η ab represents the structural similarity adjustment factor of the a-th sample on the b-th feature dimension, ξ ac represents the average symptom label alignment value calculated between the a-th sample and the c-th label category during the training phase, κ ac It represents the label structure cross-measure value between the a-th sample and the c-th label category in the current mapping stage, ∈ represents a constant to avoid the denominator being zero, and h represents the total number of feature dimensions of the symptom feature vector.

7. The outpatient automatic triage method based on multimodal data fusion according to claim 5 is characterized in that: The specific steps of S4 include: S401: Based on the department matching score and the corresponding waiting queue length, extract the average waiting time and matching score of the patients in the department, calculate the total weighted score according to the fixed weight, sort by score and record the index position, and generate a weighted sort priority value; S402: Based on the weighted ranking priority value, a dynamic scoring threshold adjustment mechanism is introduced to determine the upper and lower limits of the weighted score within the score change control interval, and an adjustment offset is calculated based on the deviation ratio from the median value to correct the ranking value to obtain a regulated ranking sequence; S403: For the adjusted sorting sequence, a path diversity constraint factor is introduced. Based on the waiting time density distribution and the path repetition distribution coefficient, the independence and distribution balance of the path combination are determined. The triage paths are re-sorted according to the benchmark score ratio to obtain a candidate set of triage paths.

8. The outpatient automatic triage method based on multimodal data fusion according to claim 7 is characterized in that: The adjustment offset is calculated using the formula: Where, ΔS x represents the sort adjustment offset of the xth patient, M represents the median of the weighted sort priority values ​​of the patients in the current sorting queue, n represents the total number of patients in the current sorting queue, and the unit is person, W x represents the weighted ranking priority value of the xth patient, L k The waiting time density distribution value of the kth candidate path is in minutes / path, R k represents the path duplication coefficient of the kth candidate path, and z represents the total number of paths in the path distribution candidate set.

9. The outpatient automatic triage method based on multimodal data fusion according to claim 7, characterized in that: The specific steps of S5 include: S501: Based on the triage path candidate set, extract the candidate department admission data corresponding to each path and the number of patients in the current time period, calculate the ratio of the patient queue in the candidate department to the department's admission capacity value, and generate a admission pressure coefficient based on the admission capacity benchmark value; S502: calling the patient reception pressure coefficient, comparing the urgency level of the patients associated with the path, and weighting and fusing the patient reception pressure with the level weight value to generate a path fusion priority value; S503: Sort the paths in descending order according to the path fusion priority value, select the top path and match it with the current queue update frequency, perform real-time correction on the path selection, and obtain the triage path execution result.

10. The outpatient automatic triage method based on multimodal data fusion according to claim 9, characterized in that: The path fusion priority value is calculated using the formula: Among them, P i represents the fusion priority value of the i-th path, λ ij represents the admission pressure coefficient of the jth candidate department in the i-th path, ω ij represents the urgency level weight of the patients admitted by the jth candidate department in the i-th path, m represents the total number of candidate departments included in the path, α i represents the overall adjustment coefficient of the i-th path, Q i represents the average number of patients in the queue of the candidate department of the i-th path, in units of people, C i represents the average reception capacity of the candidate departments of the i-th path, in person / hour, β i is the capacity adjustment compensation value of the i-th path, in person / hour, δ i represents the urgency index value of the patient associated with the i-th path, γ i is the fusion interference correction term of the i-th path.

Images