High-risk chest pain patient triage identification system based on machine learning

Abstract

The invention discloses a high-risk chest pain patient triage identification system based on machine learning, and belongs to the technical field of medical triage, and the system specifically comprises the steps: storing chest pain related keywords and triage indexes in a keyword library, and coding the chest pain related keywords and triage indexes into text feature vectors; the text recognition module collects medical records, traverses text information, marks subjective keywords, and calculates an initial triage level; the image acquisition module collects and preprocesses patient condition images; the text and graph matching module encodes an image into a feature vector, retrieves a nearest text feature vector, and marks a matching keyword; and the triage correction module compares the subjective keyword set with the matching keyword set, determines a final triage grade according to different set relationships, takes an initial grade when the matching keyword set is a subset of the subjective keyword set, takes a matching grade when the matching keyword set is a subset of the subjective keyword set, and comprehensively judges when the matching keyword set and the matching keyword set are intersected or mutually excluded, so that accurate triage of the chest pain patient is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of medical triage, and in particular to a high-risk chest pain patient triage and identification system based on machine learning. Background Art

[0002] In the medical emergency system, rapid and accurate triage of high-risk chest pain patients is a key link in ensuring the safety of patients' lives. The complexity and diversity of chest pain symptoms make it a very challenging task to accurately judge the severity of the patient's condition and conduct reasonable triage. Currently, in the triage process of high-risk chest pain patients, existing technologies have obvious limitations.

[0003] In most cases, medical institutions mainly rely on the medical records filled out by patients themselves as an important basis for triage. However, the medical records filled out by patients are highly subjective. On the one hand, due to the lack of medical knowledge, patients may not describe their symptoms accurately and comprehensively. For example, patients may not be able to clearly and accurately describe the specific location of chest pain, the nature of pain (such as tingling, dull pain, squeezing pain, etc.), the duration of pain, and whether there are other related symptoms (such as dyspnea, sweating, dizziness, etc.), resulting in the lack of key condition information. On the other hand, when filling out medical records, patients may be affected by psychological factors, such as anxiety and tension, which leads to deviations in the description. This highly subjective medical record information is difficult to fully and objectively reflect the patient's true condition.

[0004] Triage and sorting based solely on such subjective medical records can easily lead to inaccurate triage results. For high-risk chest pain patients, incorrect triage may lead to serious consequences, such as misdiagnosing truly high-risk patients as low-risk, making them unable to receive emergency treatment in time, delaying the best treatment opportunity, and increasing the patient's life risk; or misdiagnosing low-risk patients as high-risk, resulting in unnecessary waste of medical resources, and may also affect the efficiency of treatment for other patients who really need treatment. Therefore, the existing method of triage that relies solely on the patient's subjective medical records cannot achieve accurate and efficient triage of high-risk chest pain patients, and an innovative triage identification system is urgently needed to solve this problem. Summary of the invention

[0005] The purpose of the present invention is to provide a high-risk chest pain patient triage and identification system based on machine learning to solve the following technical problems: Triage and sorting based solely on subjective medical records can easily lead to inaccurate triage results.

[0006] The purpose of the present invention can be achieved by the following technical solutions: A high-risk chest pain patient triage and identification system based on machine learning, including: A keyword library is used to store keywords related to chest pain, each keyword is annotated with a corresponding triage index; and based on a pre-trained text recognition model, the keywords are encoded into a vector space and a text feature vector is output; A text recognition module is used to collect the medical records filled in by the patient, identify the text information in the medical records, traverse the text information based on the keyword library, and when the text information contains a keyword in the keyword library, mark the contained keyword as a subjective keyword, and calculate the initial triage level of the patient according to the triage index of the subjective keyword; An image acquisition module, used to collect patient condition images based on a camera and pre-process the condition images, wherein the condition images include facial images and body movement images; A text-image matching module, for encoding the disease image into a vector space and outputting an image feature vector based on a preset image recognition model, searching the keyword library for a text feature vector closest to the image feature vector, marking it as a matching vector, and marking a keyword corresponding to the matching vector as a matching keyword; The triage correction module is used to compare the set of matching keywords with the set of subjective keywords. When the matching keyword set is a subset of the subjective keyword set, the patient's final triage level is the initial triage level; when the subjective keyword set is a subset of the matching keyword set, the patient's final triage level is the triage level corresponding to the matching keyword set; when the subjective keyword set and the matching keyword set are in an intersection relationship or a mutually exclusive relationship, the patient's final triage level is comprehensively judged based on the two sets.

[0007] As a further solution of the present invention: in the triage correction module, when the subjective keyword set and the matching keyword set are in an intersection relationship or a mutually exclusive relationship, the specific process of comprehensively judging the final triage level of the patient based on the two sets is: The corresponding subjective triage level and matching triage level are calculated according to the keywords in the subjective keyword set and the matching keyword set respectively. When the subjective triage level is greater than the matching triage level, the final triage level is the average of the two. When the matching triage level is greater than the subjective triage level, the final triage level is the matching triage level.

[0008] As a further solution of the present invention: the specific process of calculating the triage level according to the keywords is: Get the triage index of each keyword in the set in turn. For the i-th keyword, the triage index X i , then the corresponding weight ω is assigned to the keyword i , , then the weighted value of the triage index is , where e is a natural number, α is a preset parameter, and α>0, Xmin represents the minimum triage index in the set, n represents the number of keywords in the set, and the triage level corresponding to the weighted value is obtained according to the preset weighted value-triage level correspondence table.

[0009] As a further solution of the present invention: in the text recognition module, the specific process of traversing the text information is as follows: Based on OCR technology and natural language processing technology, the text information in the medical record is recognized, the font, size and layout of the text are corrected and processed, and the text information is segmented. Based on the constructed keyword library, a string matching algorithm is used to compare each segmented word with the keywords in the keyword library. When a segmented word is detected that semantically matches the keyword in the keyword library, the keyword that matches the segmented word is marked as a subjective keyword, and the corresponding triage numerical attribute is called.

[0010] As a further solution of the present invention: in the image acquisition module, the process of preprocessing the disease image is: Gaussian filtering and median filtering are used to suppress random noise in the disease image, and adaptive histogram equalization is used to optimize the grayscale distribution of the disease image. Use the image scaling algorithm to adjust the disease image to the preset standard size, and convert the format of the disease image to convert the bitmap format into a tensor format that meets the processing requirements of the deep learning model.

[0011] As a further solution of the present invention: in the text-image matching module, the specific process of searching the keyword library for the text feature vector closest to the image feature vector and marking it as a matching vector is as follows: The vectors in the keyword library are divided to build a tree structure. Starting from the root node of the keyword library, according to the positional relationship between the image feature vector and the node partition hyperplane, choose to enter the left subtree or the right subtree for search. In the subtree, repeatedly choose to enter the left subtree or the right subtree of the next level for search until reaching the leaf node. At the leaf node, the distance between the image feature vector and the text feature vector stored in the leaf node is calculated based on the Euclidean distance, and the current closest text feature vector and the corresponding distance value are recorded; then backtrack to the parent node to check whether there are other subtrees containing closer vectors. If so, search the subtree, and update the nearest text feature vector and distance value. Repeat the above process until the spatial index structure corresponding to the entire keyword library is traversed, and finally determine the text feature vector closest to the image feature vector and mark it as a matching vector.

[0012] As a further solution of the present invention: in the triage correction module, when the matching keyword set is completely identical to the subjective keyword set, the initial triage level is output as the final triage level.

[0013] As a further solution of the present invention: in the triage correction module, the triage order of the patients is sorted according to the final triage level of the patients.

[0014] Beneficial effects of the present invention: The present invention uses a keyword library with a pre-annotated triage index, combined with OCR and natural language processing technology, to accurately extract key symptom descriptions in medical records, then collect the patient's facial expressions and body movements to provide objective physiological state information, and then match the image feature vector with the text feature vector to verify whether the keywords in the subjective medical record are consistent with the objective condition image, and play a role in supplementing the patient's condition details that are not clearly described. Based on the triage index and the keyword analysis corresponding to the image and text, the present invention can achieve accurate identification and efficient triage of high-risk patients, sort the patient queue in real time according to the final triage level, automatically put high-risk patients in the front, shorten the waiting time for treatment, and optimize the allocation of medical resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The present invention will be further described below in conjunction with the accompanying drawings.

[0016] Figure 1 It is a module schematic diagram of the present invention. DETAILED DESCRIPTION

[0017] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0018] See also Figure 1 As shown, the present invention is a high-risk chest pain patient triage and identification system based on machine learning, comprising: The keyword library, as one of the core components of this high-risk chest pain patient triage and identification system based on machine learning, undertakes crucial information storage and preprocessing tasks. During its construction, a professional medical team and data scientists have extensively collected a large number of clinical cases, medical literature, and authoritative chest pain diagnosis guidelines, and conducted in-depth mining and screening of various types of information related to chest pain, and carefully extracted a series of keywords closely related to chest pain. These keywords cover multiple dimensions, including chest pain symptom descriptions (such as "compressive pain", "stabbing pain", "dull pain"), pain sites (such as "precordial area", "behind the sternum", "left chest wall"), accompanying symptoms (such as "dyspnea", "palpitations", "sweating"), and types of diseases that may cause chest pain (such as "myocardial infarction", "angina pectoris", "pneumothorax").

[0019] For each extracted keyword, professional medical personnel strictly mark the corresponding triage index based on their rich clinical experience and relevant medical standards. The triage index is a key quantitative indicator to measure the severity of the condition represented by the keyword, and provides an important basis for judgment for the subsequent triage process. For example, keywords related to "myocardial infarction" may be assigned a higher triage index because the condition they represent is extremely critical; while some keywords with relatively mild symptoms, such as "occasional mild chest pain", are marked with a lower triage index.

[0020] After completing the annotation of keywords and triage index, the system further encodes the keywords in the keyword library based on the pre-trained high-performance text recognition model. The text recognition model usually adopts the neural network architecture in deep learning, such as recurrent neural network (RNN) and its variant long short-term memory network (LSTM), or models based on Transformer architecture. These models are trained on a large amount of text data and have strong semantic understanding and feature extraction capabilities. In the encoding process, the model regards each keyword as a text sequence, and converts the keyword from the original text form to a feature vector in a high-dimensional vector space by analyzing its internal semantic structure. In the vector space, the semantic similarity between keywords can be reflected by geometric relationships such as the distance or angle between vectors. For example, the semantically similar keywords "compressive pain" and "constrictive pain" have text feature vectors that are close in position and short in distance in the vector space; while the vectors of keywords with large semantic differences, such as "chest pain" and "headache", are relatively far apart in space. This operation of encoding keywords into vector space and outputting text feature vectors not only provides an efficient data structure foundation for the subsequent text recognition module to quickly and accurately match the text information in the medical records, but also lays a solid technical support for the text-image matching module to achieve effective comparison and association between image feature vectors and text feature vectors when fusing cross-modal information, thereby helping the entire high-risk chest pain patient triage and identification system to achieve accurate and efficient triage functions in all aspects.

[0021] The text recognition module, as the key front-end link of the entire high-risk chest pain patient triage identification system based on machine learning, plays an indispensable role in the collection and processing of patient medical records. The core task of this module is to efficiently and accurately extract key information related to chest pain from the medical records filled out by the patient, and preliminarily determine the patient's triage level based on this information, providing an important basis for the subsequent triage process.

[0022] In actual operation, the text recognition module first uses advanced optical character recognition (OCR) technology to perform preliminary extraction of text information from the medical records submitted by the patient. OCR technology uses pixel analysis of medical record images and complex pattern recognition algorithms to convert the text in the image into a text format that the computer can understand. However, due to the diverse sources of patient medical records, there may be problems such as different fonts, different font sizes, and chaotic typesetting formats, which will seriously affect the recognition accuracy of OCR technology. Therefore, the text recognition module then uses a series of preprocessing methods in natural language processing (NLP) technology to comprehensively correct and process the recognized text information. In terms of font correction, the system uses a deep learning model to learn and classify the features of different fonts, and can accurately identify and convert various special fonts into standard fonts to ensure the consistency of text information. For the font size problem, normalization is performed by analyzing the pixel ratio of the text in the image and the relative size relationship with the surrounding text to eliminate the interference of font size differences on subsequent analysis. In terms of typesetting, the text recognition module uses an algorithm based on layout analysis to identify information such as paragraph structure, title hierarchy, and text flow in the medical records, and reorganizes the chaotic typesetting into an orderly text structure to facilitate subsequent word segmentation and keyword matching operations.

[0023] After completing the above preprocessing steps, the text recognition module uses the word segmentation algorithm in NLP technology to perform detailed word segmentation on the sorted text information. Common word segmentation algorithms include rule-based word segmentation methods, statistical model-based word segmentation methods, and deep learning-driven word segmentation models. For example, the hidden Markov model (HMM) word segmentation algorithm based on statistical models, by learning a large amount of text data, constructs the probability of each word appearing and the transition probability model between words, so that the continuous text sequence can be accurately divided into independent words. The result after word segmentation forms a set containing many words, which provides a basis for subsequent comparison with the keyword library.

[0024] Based on the established keyword library, the text recognition module uses an efficient string matching algorithm to compare each word in the word set with the keywords in the keyword library one by one. There are many types of string matching algorithms, such as the classic KMP (Knuth-Morris-Pratt) algorithm and the BM (Boyer-Moore) algorithm. These algorithms can quickly locate matching items in a large number of keywords by cleverly utilizing the characteristics of the string, greatly improving the matching efficiency. In the matching process, the system not only focuses on the complete matching of the word segmentation and the keyword at the character level, but also uses semantic understanding technology to identify words with similar semantics. For example, when the word segmentation is "severe chest pain", even if there is no exactly the same expression in the keyword library, through semantic analysis, the system can associate it with semantically matching keywords such as "severe chest pain". Once a word segmentation that matches the semantics of a keyword in the keyword library is detected, the text recognition module will immediately mark the keyword that matches the word segmentation as a subjective keyword, and quickly call the pre-annotated triage numerical attribute of the keyword to prepare for calculating the initial triage level of the patient.

[0025] When calculating the initial triage level of the patient, the text recognition module processes each keyword in the subjective keyword set in turn according to a specific weighted calculation method. For the i-th keyword in the set, its corresponding triage index is X i In order to more accurately reflect the importance of each keyword in the diagnosis of the disease, the system will assign a corresponding weight ω to the keyword i . Weight ω i The calculation uses an algorithm based on the relative size of the keyword triage index, namely: ; After calculating the weighted value of each keyword, the weighted values ​​of all keywords are accumulated to obtain the weighted value of the triage index of the entire subjective keyword set. The formula is: ; Where e is a natural constant, α is a parameter pre-set based on a large amount of clinical data and expert experience, and α>0, X min represents the smallest triage index in the subjective keyword set, and n represents the number of keywords in the set. In this way, the importance of keywords with higher triage indexes in the overall calculation can be highlighted, while avoiding excessive influence caused by excessively high triage indexes of individual keywords.

[0026] Finally, based on the pre-established weighted value-triage level correspondence table, which is determined through statistical analysis of a large amount of clinical case data and expert evaluation, the system can quickly and accurately find the triage level corresponding to the weighted value, thereby completing the calculation of the patient's initial triage level.

[0027] The image acquisition module is used to collect the patient's condition image based on the camera, the condition image includes facial image and limb movement image, and pre-process the condition image, suppress the random noise in the condition image by using Gaussian filtering and median filtering, and optimize the grayscale distribution of the condition image by using adaptive histogram equalization; use the image scaling algorithm to adjust the condition image to a preset standard size, and convert the format of the condition image, converting the bitmap format into a tensor format that conforms to the processing of the deep learning model.

[0028] The text-image matching module is used to encode the disease image into a vector space based on a preset image recognition model to output an image feature vector, and then enter the key stage of retrieving the text feature vector closest to the image feature vector.

[0029] First, in order to measure the proximity between two vectors, the system uses common distance measurement methods, such as Euclidean distance, cosine similarity, etc. Taking Euclidean distance as an example, its calculation method is to take the square root of the sum of the squares of the differences in the corresponding dimensions of the two vectors in multidimensional space. The smaller the value, the closer the distance between the two vectors in space.

[0030] After determining the distance measurement method, the system begins to traverse and search the text feature vectors in the keyword library. Given that the number of text feature vectors in the keyword library may be large, spatial index structures such as KD trees and Ball trees are used to improve retrieval efficiency. These data structures can reasonably divide vectors in high-dimensional space and build a tree structure, so that when searching, the area that may contain the nearest neighbor vector can be quickly located, avoiding comparing all vectors one by one.

[0031] The specific retrieval process is as follows: starting from the root node of the keyword library, according to the positional relationship between the image feature vector and the node partition hyperplane, choose to enter the left subtree or the right subtree for further search. After entering the subtree, for the nodes in the subtree, again according to the positional relationship between the image feature vector and the current subtree node partition hyperplane, determine whether to enter the left subtree or the right subtree of the current subtree (the subtree here is relative to the subtree entered). In other words, every time you enter a new subtree level, you must compare the positional relationship between the image feature vector and the hyperplane partitioning the node at that level again, and then make a choice of entering the left subtree or the right subtree based on the comparison result. By repeating this process of comparing and selecting to enter the next level subtree at each subtree level, it continues until the leaf node is reached. At the leaf node, calculate the distance between the image feature vector and the text feature vector stored in the leaf node, and record the current nearest text feature vector and the corresponding distance value. Then, trace back to the parent node to check whether there are other subtrees that may contain closer vectors. If so, search the subtree and update the nearest text feature vector and distance value. This process is repeated until the spatial index structure corresponding to the entire keyword library is traversed, and the text feature vector closest to the image feature vector is finally determined and marked as a matching vector. Through this rigorous and efficient retrieval process, the text feature vector closest to the disease image feature vector can be accurately found from the keyword library, providing a key basis for subsequent triage and correction operations.

[0032] The triage correction module, as a key link in the entire high-risk chest pain patient triage identification system based on machine learning, undertakes the important task of comprehensively analyzing and correcting the information obtained in the early stage through the text recognition module and the text-image matching module, so as to accurately determine the final triage level of the patient and reasonably arrange the triage order. The core work of this module is to conduct an in-depth and detailed comparative analysis of the matching keyword set and the subjective keyword set, and use scientific and reasonable algorithms to determine the final triage level of the patient based on different set relationships.

[0033] When the matching keyword set is a subset of the subjective keyword set, this means that the condition information covered by the subjective keywords extracted from the patient's medical record text information is reflected in the matching keywords obtained through image analysis, and may also contain other text descriptions of the condition that are not reflected in the image. At this time, the system believes that the initial triage level calculated based on the patient's medical record text information has relatively comprehensively reflected the severity of the patient's condition, so the initial triage level is directly determined as the patient's final triage level. This decision not only reflects the recognition of the comprehensiveness of the medical record text information, but also follows the principles of simplicity and efficiency of the triage process.

[0034] On the contrary, when the subjective keyword set is a subset of the matching keyword set, it indicates that the matching keywords obtained through image analysis reflect richer and more comprehensive information about the condition, and the text information of the patient's medical record only covers part of it. In this case, the system determines that the triage level corresponding to the matching keyword set obtained based on image analysis can more accurately reflect the patient's actual condition, so the triage level corresponding to the matching keyword set is determined as the patient's final triage level. This decision-making method makes full use of the important role of image information in supplementing and improving the judgment of the condition, ensuring the accuracy of the triage results.

[0035] When the subjective keyword set and the matching keyword set are in an intersecting relationship, that is, there are partially overlapping keywords in the two sets, but each contains keywords that the other does not have. At this time, the disease information reflected by the two sets has both overlapping parts and their own unique aspects. In order to comprehensively consider these two parts of information, the triage correction module first calculates the corresponding subjective triage level and matching triage level according to the keywords in the subjective keyword set and the matching keyword set according to the established weighted calculation method. As mentioned above, the triage index and corresponding weight of each keyword will be considered during the calculation process to ensure that the calculation result can accurately reflect the severity of the disease represented by the keyword. When the calculated subjective triage level is greater than the matching triage level, it means that the severity of the disease reflected by the text information of the medical record is relatively high, but the image information also has a certain reference value. Therefore, the system adopts the method of taking the average of the two to determine the final triage level. This method balances the influence of the two information sources to a certain extent, making the final triage level more objective. On the contrary, when the matching triage level is greater than the subjective triage level, it indicates that the condition reflected by the image information is more serious. The system directly determines the matching triage level as the final triage level to give priority to timely treatment of more serious conditions.

[0036] If the subjective keyword set and the matching keyword set are mutually exclusive, that is, the two sets do not have any overlapping keywords, this means that the medical record text information and image information reflect completely different disease characteristics. At this time, the triage correction module also calculates the triage levels corresponding to the two sets respectively, and then determines the final triage level based on the above comparison rules. In this way, even when the two information sources are quite different, the system can comprehensively consider various factors and make relatively reasonable triage decisions.

[0037] In particular, when the matching keyword set is exactly the same as the subjective keyword set, it means that the medical record text information and image information reflect the patient's condition in a highly consistent manner. In this case, the system believes that the initial triage level has accurately reflected the patient's condition, so it directly outputs the initial triage level as the final triage level, which not only ensures the stability of the triage results, but also improves the triage efficiency.

[0038] After determining the final triage level of the patient, the triage correction module also shoulders the important responsibility of sorting the triage order of the patients. The system will sort all patients in order from high to low according to the final triage level. The higher the triage level, the more critical the patient's condition is and the priority treatment is needed. Through this scientific and reasonable sorting method, it can ensure the most reasonable allocation of medical resources, so that high-risk chest pain patients can receive timely and effective treatment in the shortest time, thereby maximizing the safety of patients' lives and improving the operating efficiency of the entire medical emergency system.

[0039] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data for software simulation. The preset parameters and thresholds in the formula are selected by technicians in this field according to actual conditions. The above is only a specific implementation method of the present application, but the protection scope of the present application is not limited to this. Any technician familiar with the technical field can easily think of changes or replacements within the technical scope disclosed in this application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application shall be based on the protection scope of the claims. An embodiment of the present invention is described in detail above, but the content is only a preferred embodiment of the present invention and cannot be considered to limit the scope of implementation of the present invention. All equal changes and improvements made according to the scope of application of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. A high-risk chest pain patient triage and identification system based on machine learning, characterized in that: include: Keyword library, used to store keywords related to chest pain, each keyword is marked with a corresponding triage index; Based on the pre-trained text recognition model, the keywords are encoded into the vector space and the text feature vector is output; A text recognition module is used to collect the medical records filled in by the patient, identify the text information in the medical records, traverse the text information based on the keyword library, and when the text information contains a keyword in the keyword library, mark the contained keyword as a subjective keyword, and calculate the initial triage level of the patient according to the triage index of the subjective keyword; An image acquisition module, used to collect patient condition images based on a camera and pre-process the condition images, wherein the condition images include facial images and body movement images; A text-image matching module, for encoding the disease image into a vector space and outputting an image feature vector based on a preset image recognition model, searching the keyword library for a text feature vector closest to the image feature vector, marking it as a matching vector, and marking a keyword corresponding to the matching vector as a matching keyword; The triage correction module is used to compare the set of matching keywords with the set of subjective keywords. When the matching keyword set is a subset of the subjective keyword set, the patient's final triage level is the initial triage level; when the subjective keyword set is a subset of the matching keyword set, the patient's final triage level is the triage level corresponding to the matching keyword set; when the subjective keyword set and the matching keyword set are in an intersection relationship or a mutually exclusive relationship, the patient's final triage level is comprehensively judged based on the two sets.

2. According to the machine learning-based high-risk chest pain patient triage and identification system of claim 1, it is characterized by: In the triage correction module, when the subjective keyword set and the matching keyword set are in an intersection relationship or a mutually exclusive relationship, the specific process of comprehensively judging the final triage level of the patient based on the two sets is as follows: The corresponding subjective triage level and matching triage level are calculated according to the keywords in the subjective keyword set and the matching keyword set respectively. When the subjective triage level is greater than the matching triage level, the final triage level is the average of the two. When the matching triage level is greater than the subjective triage level, the final triage level is the matching triage level.

3. According to the machine learning-based high-risk chest pain patient triage and identification system of claim 2, it is characterized in that: The specific process of calculating the triage level based on keywords is as follows: Get the triage index of each keyword in the set in turn. For the i-th keyword, the triage index X i , then the corresponding weight ω is assigned to the keyword i , , then the weighted value of the triage index is , where e is a natural number, α is a preset parameter, and α>0, X min represents the minimum triage index in the set, n represents the number of keywords in the set, and the triage level corresponding to the weighted value is obtained according to the preset weighted value-triage level correspondence table.

4. According to the high-risk chest pain patient triage and identification system based on machine learning in claim 1, it is characterized in that: In the text recognition module, the specific process of traversing text information is as follows: Based on OCR technology and natural language processing technology, the text information in the medical record is recognized, the font, size and layout of the text are corrected and processed, and the text information is segmented. Based on the constructed keyword library, a string matching algorithm is used to compare each segmented word with the keywords in the keyword library. When a segmented word is detected that semantically matches the keyword in the keyword library, the keyword that matches the segmented word is marked as a subjective keyword, and the corresponding triage numerical attribute is called.

5. According to the high-risk chest pain patient triage and identification system based on machine learning in claim 1, it is characterized in that: In the image acquisition module, the process of preprocessing the disease image is as follows: Gaussian filtering and median filtering are used to suppress random noise in the disease image, and adaptive histogram equalization is used to optimize the grayscale distribution of the disease image. Use the image scaling algorithm to adjust the disease image to the preset standard size, and convert the format of the disease image to convert the bitmap format into a tensor format that meets the processing requirements of the deep learning model.

6. According to the high-risk chest pain patient triage and identification system based on machine learning in claim 1, it is characterized by: In the text-image matching module, the specific process of searching the keyword library for the text feature vector closest to the image feature vector and marking it as a matching vector is as follows: The vectors in the keyword library are divided to build a tree structure. Starting from the root node of the keyword library, according to the positional relationship between the image feature vector and the node partition hyperplane, choose to enter the left subtree or the right subtree for search. In the subtree, repeatedly choose to enter the left subtree or the right subtree of the next level for search until reaching the leaf node. At the leaf node, the distance between the image feature vector and the text feature vector stored in the leaf node is calculated based on the Euclidean distance, and the current closest text feature vector and the corresponding distance value are recorded; then backtrack to the parent node to check whether there are other subtrees containing closer vectors. If so, search the subtree, and update the nearest text feature vector and distance value. Repeat the above process until the spatial index structure corresponding to the entire keyword library is traversed, and finally determine the text feature vector closest to the image feature vector and mark it as a matching vector.

7. According to the machine learning-based high-risk chest pain patient triage and identification system of claim 1, it is characterized by: In the triage correction module, when the matching keyword set is completely identical to the subjective keyword set, the initial triage grade is output as the final triage grade.

8. The high-risk chest pain patient triage and identification system based on machine learning according to claim 1 is characterized in that: In the triage correction module, the triage order of the patients is sorted according to their final triage levels.

Images