Emergency pre-examination grading system and method based on collaborative decision-making of large language model and tree model

Abstract

The invention discloses an emergency pre-examination grading system and method based on collaborative decision of a large language model and a tree model. The method comprises the following steps: constructing a data set according to patient information and screening data; training an initial random forest model based on the data set, generating a basic decision rule, mining and extracting high-frequency features based on association rules, and combining to obtain a candidate rule pool; constructing a cue word structure adaptive to the field, driving LLM to complete rule correction, and forming a correction rule set; performing multiple rounds of rule random division and rule combination generation based on the correction rule set, and then screening out an optimal rule; and on the basis of a specific scene, the expert rule and the corrected optimal rule are fused, matching is carried out on patients, and emergency pre-examination grading is realized. According to the method, the interpretability is improved while the grading accuracy is improved, and the problem of poor cross-courtyard generalization of a machine learning model is effectively solved. The generalization ability of the rule is remarkably improved, and the method is adaptive to a multi-center combined diagnosis and treatment scene.

Description

Technical Field

[0001] This invention relates to the technical field of emergency triage, and more particularly to an emergency triage and classification system and method based on collaborative decision-making using a large language model and a tree model. Background Technology

[0002] Emergency triage is a crucial step in emergency medicine, aiming to rapidly classify patients based on the urgency of their condition to ensure critically ill patients receive priority treatment. Traditional triage relies primarily on experienced nurses' human assessments, but this method is highly subjective, inefficient, and susceptible to human influence. With the development of machine learning technology, data-driven approaches have become an inevitable trend to overcome the bottlenecks of traditional triage: the structured data (vital signs, laboratory indicators) and unstructured data (chief complaints, medical history texts) accumulated in emergency scenarios provide a foundation for rule optimization, while machine learning has the advantage of processing multidimensional data. Models such as gradient boosting and random forests significantly outperform traditional scales when used for ICU admission prediction. The rise of large language models (LLMs) has further opened up new avenues: they can not only parse implicit information in medical records but also integrate clinical knowledge and error cases through logical reasoning, providing technical support for generating interpretable rules.

[0003] Currently, to optimize the accuracy of emergency triage, many machine learning-based methods have been proposed, achieving good triage accuracy. Several LLM-based techniques for judging emergency triage have also been proposed, optimizing the emergency triage model. Among these, the existing technical solution most similar to this invention is: ① A prediction model developed based on the XGBoost algorithm (Agius, S., Cassar, V., Magri, C. et al. Predicting Emergency Severity Index (ESI) level, hospital admission, and admitting ward in an emergency department using data-driven machine learning. BMC Med Inform Decis Mak 25, 281 (2025).), used for ESI level, hospital admission probability, and inpatient ward classification of emergency department patients. The study employs a two-stage approach: first, preliminary predictions are made using basic demographic and symptom data; then, key blood test results are integrated to optimize prediction accuracy. ② An LLM-based fine-tuning approach (Shen Q, Zhang X, Ren H, et al. Knowledge-embedded large language models for emergency triage, Knowledge-Based Systems 318,113431(2025).). This approach introduces a Sequential Domain and Task Adaptation (SDTA) framework to improve the accuracy and consistency of emergency triage through a large language model. The approach includes a two-stage adaptation: first, continued pre-training on MIMIC-IV clinical notes to learn medical domain knowledge; then, using cue-based fine-tuning techniques on the MIETIC dataset, specifically fine-tuning the model for the ESI triage task.

[0004] Similar to ①, existing technologies suffer from the dual problems of impractical real-time application and low predictive reliability. First, the first stage relies on basic demographic and chief symptom data reported by patients or recorded by nurses. This type of data is highly subjective and prone to bias due to exaggeration or omission in patients' symptom descriptions, directly affecting the objectivity of the model. Second, the key blood test results that need to be integrated in the second stage need to wait for laboratory output in real emergency scenarios, while triage decisions should be completed quickly upon patient arrival. This reliance on later data will severely delay prediction, which is completely contrary to the actual emergency workflow. However, the assumption that the data is immediately available is obviously not in line with clinical reality, ultimately leading to low efficiency in the clinical deployment of the model. Meanwhile, because this technology integrates subjective data such as chief complaints and symptoms, it relies too heavily on the initial assessments of triage nurses and other personnel, further amplifying decision-making variability. The triage process itself is influenced by "subjective judgment," and it is difficult to completely eliminate subjective bias when building a model on this basis. Over-triage or under-triage may occur due to miscoding or misunderstanding of chief complaint data, directly affecting the allocation of medical resources and patient safety. Even if the data is balanced by synthetic minority oversampling techniques, the inherent limitations of subjective data still make it difficult for the model to maintain consistent performance in diverse emergency environments.

[0005] While existing technologies similar to ② improve triage accuracy by fine-tuning large language models using the SDTA framework, they still suffer from significant drawbacks such as excessive resource consumption and insufficient generalization. On the one hand, the continuous pre-training phase requires processing on large-scale databases such as MIMIC-IV clinical notes, involving trillions of tokens. Although 4-bit quantization and low-rank adaptation techniques can reduce memory consumption, high-performance GPUs and long training times are still required, resulting in extremely high computational resource demands. This not only leads to high hardware and time costs but also sets a very high deployment threshold, making it almost impossible for resource-constrained primary healthcare institutions to implement. Furthermore, the maintenance requirements of large-scale data infrastructure further limit the large-scale application of this technology. On the other hand, although the MIETIC dataset improved the interpretability of the model during the instruction fine-tuning stage, the model still suffers from problems such as hallucination output and insufficient decision transparency. Furthermore, the model training is highly dependent on single-region datasets like MIMIC-IV, lacking adaptation to different regions and emergency room environments of different levels of hospitals. The generalization performance is unstable when applied across scenarios. In addition, the model needs to be continuously monitored and updated to avoid possible biases such as gender and region in the training data, which further increases the complexity and cost of the technology application. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by proposing an emergency triage system and method based on collaborative decision-making using a large language model and a tree model.

[0007] The objective of this invention is achieved through the following technical solution: an emergency triage system based on collaborative decision-making using a large language model and a tree model, the system comprising:

[0008] The data construction module builds a dataset based on patient information and filters the data before inputting it into the LLM-driven decision tree rule refining module.

[0009] The LLM-driven decision tree rule refinement module trains an initial random forest model based on the dataset, generates basic decision rules, and extracts high-frequency features based on association rules, combining them to obtain a candidate rule pool; it constructs a domain-adaptive prompt word structure to drive LLM to complete rule correction, forming a set of corrected rules;

[0010] The rule selection module performs multiple rounds of random rule partitioning and rule combination generation based on the modified rule set, and then selects the optimal rule;

[0011] The rule application and evaluation module integrates expert rules with the revised optimal rules based on specific scenarios to match patients and achieve emergency triage.

[0012] Furthermore, patient information includes demographic information, vital signs information, and nurse triage labels.

[0013] Furthermore, a hierarchical random sampling method is used to divide the training set and the validation set to train the initial random forest model. After training, a decision path traversal algorithm is used to generate basic decision rules from each decision tree of the random forest. Error cases are screened from the validation set, and candidate rules corresponding to high-frequency feature combinations are extracted from the validation set simultaneously using the Apriori association rule mining algorithm.

[0014] Furthermore, the prompt word structure includes the rule context, error case details, correction targets, and inference constraints in the candidate rule pool.

[0015] Furthermore, driving the LLM to complete rule modification specifically involves: guiding the LLM to deconstruct each hierarchical rule in the initial rule pool one by one through zero-sample hints; adding binding conditions for false positive error rules; and removing overly restrictive condition terms for false negative error rules, thereby completing rule modification.

[0016] Furthermore, for each round of segmentation, based on the modified rule set, a specific rule subset representing the patient's critical condition label is generated, and the recall rate of the corresponding rule at each level is calculated to select the optimal rule. The optimal rule set is then compared with the initial rule set to perform significance verification.

[0017] Furthermore, during the integration of expert rules and the revised optimal rules, each scenario corresponds to a unique level, and each level serves as a nurse pre-screening and grading label, indicating the patient's critical condition.

[0018] The fusion rules for Level 1, Level 2 and Level 3 are: a combined set of expert rules at each level and the corresponding correction rules at the same level, with the rule condition items linked by "OR" logic;

[0019] Level 4 rules: Only Level 4 expert rules are retained, with no additional rules added.

[0020] Furthermore, the specific process of matching patients is as follows: based on the rule priority, the level 1 fusion rule, level 2 fusion rule, level 3 fusion rule, and level 4 rule are applied in sequence. A successful match is defined as satisfying all the conditions of at least one rule in the corresponding rule set, ensuring that each patient corresponds to only one graded result.

[0021] Secondly, this invention provides an emergency triage method based on collaborative decision-making using a large language model and a tree model, the method comprising the following steps:

[0022] (1) Construct a dataset based on patient information and filter the data;

[0023] (2) Train an initial random forest model based on the dataset to generate basic decision rules, and extract high-frequency features based on association rules to combine them to obtain a candidate rule pool; construct a domain-adapted prompt word structure to drive LLM to complete rule correction and form a set of corrected rules;

[0024] (3) Based on the modified rule set, perform multiple rounds of random rule partitioning and rule combination generation, and then select the optimal rule;

[0025] (4) Based on specific scenarios, the expert rules and the modified optimal rules are integrated to match patients and realize emergency pre-examination and triage.

[0026] The beneficial effects of this invention are:

[0027] 1. This invention proposes a solution to improve the accuracy of emergency triage by using only objective vital sign data, without requiring the patient's chief complaint, thus achieving the improvement in the accuracy of emergency triage. This innovative approach anchors core emergency triage data in objective vital signs rather than the chief complaints and subsequent blood test data commonly used in existing technologies. Through targeted technical design, it thoroughly addresses the core pain points of current technologies: First, it abandons data types prone to bias, such as patient complaints and nurses' subjective records, using only objective vital signs data such as heart rate, blood pressure, and blood oxygen saturation, which can be collected immediately during the initial emergency examination. This eliminates subjective interference from exaggerated / omitted patient symptom descriptions and differences in nurses' judgments, mitigating the risk of bias in subjective data (such as errors in patient symptom descriptions and differences in nurses' judgments), significantly improving the objectivity and consistency of triage decisions. Second, since vital sign data does not require waiting for laboratory testing, it can be acquired simultaneously upon patient arrival (during the initial emergency examination), providing "real-time, reliable, and unbiased" basic data support for subsequent rule optimization. This fully adapts to the emergency workflow of "rapid triage and priority treatment," avoiding the prediction delays and triage inaccuracies caused by the subjectivity of data or reliance on subsequent blood test results, significantly improving clinical deployment efficiency.

[0028] 2. This invention constructs a collaborative framework of "machine learning initial rule generation - LLM clinical reasoning correction" to achieve complementary technological advantages. On the one hand, it uses the random forest algorithm to train on objective vital sign data, extracts a set of basic rules covering multiple scenarios through decision path traversal, and uses the Apriori association rule mining algorithm to screen high-confidence rules, ensuring the statistical reliability of the initial rules. On the other hand, it breaks through the limitations of the existing machine learning "black box" by introducing LLM. Through customized prompt word engineering, it guides LLM to deconstruct and analyze the basic rules one by one, generating corrected rules. This innovation leverages the advantages of machine learning in processing multi-dimensional data and endows the rules with clear medical logic through the clinical semantic reasoning capabilities of LLM, improving the interpretability of the rules and resolving the core contradiction of existing technologies being "highly performant but poorly interpretable."

[0029] 3. This invention does not use the SDTA framework to fine-tune a large language model, achieving comprehensive advantages in resource consumption, optimization efficiency, clinical adaptability, and operational costs. Existing technologies require continuous pre-training on massive datasets with trillions of tokens. Even with quantization and low-rank adaptation techniques, they still rely on high-performance GPUs and complex data infrastructure, resulting in high deployment barriers and costs. This invention, however, eliminates the need for fine-tuning, proposing an LLM-driven automatic iterative refinement scheme. By constructing an end-to-end automatic optimization closed-loop process of "error case feature annotation - prompt word engineering modeling - LLM logical reasoning - automatic rule correction," the efficiency of a single rule iteration is improved. First, by comparing the model's prediction results with the true grading labels, error cases are automatically identified and error-related features are labeled. Second, based on error case features and clinical knowledge, LLM automatically generates N correction rules. For low-frequency error cases, zero-shot reasoning generates specific correction rules, ensuring rule coverage completeness and achieving efficient and automated rule optimization. Finally, stratified sampling is used to validate the correction rules in multiple rounds, using "grading accuracy - recall - "Generalization stability" is used as the metric for optimal integration, forming the final rule set. This closed loop, through multi-center data adaptation training, ensures that the accuracy drop is ≤3% when rules are generalized across different hospital areas, meeting the needs of multi-center collaborative diagnosis and treatment scenarios and solving the pain points of high annotation costs and poor generalization in existing technologies. Not only are hardware and time costs significantly reduced, but it can also be easily implemented in resource-limited primary healthcare institutions. At the same time, LLM can automatically identify key features in erroneous cases and generate correction conditions based on clinical knowledge, eliminating the need for manual analysis by experts and reducing the cost of human intervention.

[0030] 4. This invention utilizes the natural language reasoning capabilities of LLM to construct a "clinical knowledge fusion - rule semantic generation" mechanism, which improves interpretability while enhancing the accuracy of triage, and effectively optimizes the problem of poor cross-hospital generalization of machine learning models. This invention significantly improves the generalization ability of rules and is suitable for multi-center collaborative diagnosis and treatment scenarios. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 This is a schematic diagram of the overall architecture of the present invention.

[0033] Figure 2 This is a schematic diagram of the LLM-driven decision tree rule refinement module of the present invention.

[0034] Figure 3 This is a schematic diagram of the rule selection module of the present invention.

[0035] Figure 4 This is a schematic diagram of the rule application and evaluation module of the present invention.

[0036] Figure 5 This is a structural diagram of an emergency triage device based on collaborative decision-making using a large language model and a tree model, according to the present invention. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described below with reference to the accompanying drawings and examples. It should be understood that the specific examples described herein are merely illustrative and not intended to limit the invention.

[0038] like Figure 1 As shown, the present invention provides an emergency triage system based on collaborative decision-making of large language model and tree model. The system includes a data construction module, an LLM-driven decision tree rule refinement module, a rule selection module, and a rule application and evaluation module.

[0039] The data construction module is implemented as follows:

[0040] Based on the patient's admission time, demographic information, vital signs information, and nurse triage information are extracted from the emergency department. Data from different hospital areas is aligned, discretized, and standardized to construct a tabular file with patients as rows and various patient information items as columns. According to specific task requirements and data inclusion / exclusion criteria, the data within the dataset is filtered and then input into the LLM-driven decision tree rule refinement module. The input data column for the i-th patient is represented as follows: ,n represents a patient with n pieces of information, the target variable The set is represented as .

[0041] like Figure 2 As shown, the LLM-driven decision tree rule refining module is implemented as follows:

[0042] First, based on the labeled dataset, a stratified random sampling method is used to divide the dataset into a training set (80%), a validation set (10%), and a test set (10%). The initial random forest model is then trained using this method. The model's hyperparameters are optimized using grid search combined with 5-fold cross-validation to ensure a balance between fit and generalization on the training set. After training, a decision path traversal algorithm is used to generate basic decision rules from each decision tree in the random forest. The initial decision rule set R is then extracted and integrated. Simultaneously, erroneous cases are screened (defined as "samples whose predicted class does not match the true class of the sample," and the error type (false positive / false negative) and erroneous association features are labeled). At the same time, the Apriori association rule mining algorithm is used to extract candidate rules corresponding to high-frequency feature combinations from the validation set, forming a candidate rule pool.

[0043] Then, a domain-adaptive large language model prompt word engineering framework is constructed to drive the large language model LLM to complete rule correction: the prompt word structure includes "rule context (conditions, conclusions, sample coverage, and accuracy on the validation set of the initial rule R in the candidate rule pool), error case details (sample feature vectors, true labels, and model error prediction results), correction objectives (improving the rule's ability to avoid error cases), and inference constraints (the number of rule condition items should not exceed 10 to ensure interpretability, and conditions that conflict with common sense in the domain should not be introduced)". Through zero-shot prompts, the LLM is guided to deconstruct each hierarchical rule in the initial rule R one by one - for false positive error rules, "constraint condition items" are added; for false negative error rules, "overly constrained condition items" are removed; finally, N refined corrected rules (N is based on the principle of "rule diversity-performance balance", usually 5-10 rules, and rules with no significant performance improvement are initially screened out through the validation set) are generated, forming the corrected rule set R*.

[0044] like Figure 3 As shown, the rule selection module is implemented as follows:

[0045] The first step is to perform multiple rounds of random partitioning and rule combination generation: stratified random sampling is used to partition the original dataset into training set, validation set, and test set K times (the partitioning ratio is fixed at 8:1:1, and the class distribution is kept consistent with the original dataset in each partition). For each round of segmentation, based on the set of modified rules R* output by LLM, multiple sets of specific modified rule subsets for levels 1, 2, and 3 are generated (denoted as R_1, R_2, and R_3, respectively; the rules within each subset are only applied to the corresponding level, and the three levels represent the critical condition of the patient, with level 1 being the most critical, level 2 the next most critical, and level 3 relatively stable, and labeled based on the nurse's pre-screening classification). On the independent test set of the corresponding round, the recall rate of R_1 for level 1 patients (Recall_1), the recall rate of R_2 for level 2 patients (Recall_2), and the recall rate of R_3 for level 3 patients (Recall_3) are calculated respectively. At the same time, the auxiliary indicators (accuracy, precision, F1 score) of each level of rule are recorded as supplementary evaluation criteria.

[0046] The second step is to select the optimal rule by level: For the Level 1 patient identification task, iterate through all Recall_1 values ​​of R_1 in K rounds of testing, and select the subset of rules with the highest recall as the Level 1 candidate optimal rule (denoted as R_opt_1). If there are multiple rounds of R_1 with the same Recall_1 (difference ≤ 0.5%), then the Precision_1 of the corresponding round is used as the tie-breaker (prioritizing the subset with higher precision). Similarly, select the Level 2 candidate optimal rule (R_opt_2) and Level 3 candidate optimal rule (R_opt_3).

[0047] The third step is statistical significance verification: This involves determining the optimal rule set for the final combination. ( The set of rules (R_opt_1∪R_opt_2∪R_opt_3) is compared with the initial rule set R, and the paired t-test is used to verify them respectively. It significantly outperforms R in terms of Level 1, Level 2, and Level 3 recall (p<0.05); simultaneously, it verifies that the overall macro F1 score of the combination rule is ≥5% higher than R, ensuring that the level-based selection not only improves individual metrics but also guarantees overall performance optimization. The final output passes all validations. As the final set of rules.

[0048]

[0049]

[0050]

[0051] Here, Recall represents recall, TP represents true positive, FN represents false negative, FP represents false positive, and Precision represents precision.

[0052] like Figure 4 As shown, the rule application and evaluation module is implemented as follows:

[0053] This module focuses on the "targeted integration of expert rules and hierarchical correction rules," achieving accurate output of unique hierarchical results for each sample through a four-level matching mechanism. The specific process is as follows:

[0054] The first step is to define the fusion scheme as four targeted matching scenarios, with each test case scenario corresponding to a unique level: ① Level 1 fusion rule: E1∪R1 (a joint set of Level 1 expert rules and Level 1 modified rules, with the rule conditions linked by "OR" logic, meaning that if a sample satisfies any one of E1 or R1, it is determined to be a Level 1 match); ② Level 2 fusion rule: E2∪R2 (similarly, a joint set of Level 2 expert rules and Level 2 modified rules, linked by "OR" logic); ③ Level 3 fusion rule: E3∪R3 (a joint set of Level 3 expert rules and Level 3 modified rules, linked by "OR" logic); ④ Level 4 rule: only E4 is retained (Level 4 expert rules, with no additional modified rules).

[0055] The second step involves performing the following matching process for each sample in the test set to determine the final classification:

[0056] 1. Perform four matching operations in sequence: apply level 1, level 2, level 3, and level 4 fusion rules to the sample feature vector in sequence. A successful match is defined as the sample features satisfying all conditions of at least one rule in the corresponding rule set.

[0057] 2. Unique Classification Determination: Since the four categories are mutually exclusive (the same sample belongs to only one level), the matching priority is set as "Level 1 > Level 2 > Level 3 > Level 4" (i.e., higher levels take precedence to avoid misjudgment by lower-level rules). This mechanism is essentially "targeted level matching + priority voting," which avoids multiple matching conflicts through level priority and ensures that each sample corresponds to only one classification result.

[0058] To verify the effectiveness of the fusion rules, an evaluation was conducted in terms of hierarchical accuracy and generalization:

[0059] Recall, precision, and F1 score were calculated for levels 1-4, with a focus on verifying the performance improvement of fusion rules (E1∪R1, E2∪R2, E3∪R3) compared to expert rules (E1, E2, E3).

[0060] The rules were validated on independent test sets at three external centers, with recall rates decreasing by ≤ 4% at each level, ensuring the suitability of the rules for different clinical scenarios.

[0061] The above evaluation ultimately validated the supplementary value of the tiered fusion rule to the expert rule, achieving a dual optimization of tiered accuracy and clinical applicability. The evaluation results are shown in Tables 1-3:

[0062] Table 1 Results of an external test set 1 at a certain hospital

[0063]

[0064] Table 2 Results of External Test Set 2 from a Certain Hospital

[0065]

[0066] Table 3. Results of External Test Set 3 from a Certain Hospital

[0067]

[0068] Corresponding to the aforementioned embodiment of an emergency triage method based on collaborative decision-making of a large language model and a tree model, the present invention also provides an embodiment of an emergency triage device based on collaborative decision-making of a large language model and a tree model.

[0069] See Figure 5 The present invention provides an emergency triage device based on collaborative decision-making of a large language model and a tree model, comprising a memory and one or more processors. The memory stores executable code, and when the processor executes the executable code, it is used to implement an emergency triage method based on collaborative decision-making of a large language model and a tree model as described in the above embodiment.

[0070] The present invention provides an embodiment of an emergency triage and classification device based on collaborative decision-making using a large language model and a tree model. This embodiment can be applied to any device with data processing capabilities, such as a computer. The device embodiment can be implemented in software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution. From a hardware perspective, such as... Figure 5 The diagram shown is a hardware structure diagram of any data processing-capable device, including an emergency triage and classification device based on collaborative decision-making using a large language model and a tree model, provided by this invention. (Except for...) Figure 5In addition to the processor, memory, network interface, and non-volatile memory shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0071] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0072] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0073] This invention also provides a computer-readable storage medium storing a program that, when executed by a processor, implements an emergency triage method based on collaborative decision-making using a large language model and a tree model, as described in the above embodiments.

[0074] The computer-readable storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device of any data processing device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units and external storage devices of any data processing device. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0075] The present invention also provides a computer program product, including a computer program, which, when executed by a processor, implements the aforementioned emergency triage method based on collaborative decision-making using a large language model and a tree model.

[0076] The above embodiments are used to explain and illustrate the present invention, but not to limit the present invention. Any modifications and changes made to the present invention within the spirit and scope of the claims shall fall within the protection scope of the present invention.

Claims

1. An emergency triage system based on collaborative decision-making using a large language model and a tree model, characterized in that, The system includes: The data construction module builds a dataset based on patient information and filters the data before inputting it into the LLM-driven decision tree rule refining module. The LLM-driven decision tree rule refinement module trains an initial random forest model based on the dataset, generates basic decision rules, and extracts high-frequency features based on association rules, combining them to obtain a candidate rule pool; it constructs a domain-adaptive prompt word structure to drive LLM to complete rule correction, forming a set of corrected rules; The rule selection module performs multiple rounds of random rule partitioning and rule combination generation based on the modified rule set, and then selects the optimal rule; The rule application and evaluation module integrates expert rules with the revised optimal rules based on specific scenarios to match patients and achieve emergency triage.

2. The emergency triage system based on collaborative decision-making using a large language model and a tree model as described in claim 1, characterized in that, Patient information includes demographic information, vital signs information, and nurse triage labels.

3. The emergency triage system based on collaborative decision-making of a large language model and a tree model according to claim 1, characterized in that, The training and validation sets are divided by hierarchical random sampling to train the initial random forest model. After training, the basic decision rules are generated from each decision tree of the random forest by the decision path traversal algorithm. Error cases are screened from the validation set, and candidate rules corresponding to high-frequency feature combinations are extracted from the validation set by the Apriori association rule mining algorithm.

4. The emergency triage system based on collaborative decision-making using a large language model and a tree model as described in claim 1, characterized in that, The prompt word structure includes the rule context, error case details, correction target, and inference constraints from the candidate rule pool.

5. The emergency triage system based on collaborative decision-making of a large language model and a tree model according to claim 1, characterized in that, The specific steps for driving LLM to complete rule correction are as follows: using zero-sample hints to guide LLM to deconstruct each hierarchical rule in the initial rule pool one by one, adding constraint conditions for false positive error rules, and removing overly constrained condition terms for false negative error rules, thus completing rule correction.

6. The emergency triage system based on collaborative decision-making of a large language model and a tree model according to claim 5, characterized in that, For each round of segmentation, a subset of specific rules representing the patient's critical condition is generated based on the modified rule set. The recall rate of each rule level is calculated to select the optimal rule. The optimal rule set is then compared with the initial rules to perform significance verification.

7. The emergency triage system based on collaborative decision-making of a large language model and a tree model according to claim 1, characterized in that, During the integration of expert rules and the revised optimal rules, each scenario corresponds to a unique level, and each level is a nurse pre-screening and grading label, indicating the patient's critical condition; The fusion rules for Level 1, Level 2, and Level 3 are a combined set of expert rules at each level and the corresponding correction rules, with the rule condition items linked by "OR" logic. Level 4 rules: Only Level 4 expert rules are retained, with no additional rules added.

8. An emergency triage system based on collaborative decision-making using a large language model and a tree model, as described in claim 7, is characterized in that... The specific process of matching patients is as follows: based on the rule priority, the level 1 fusion rule, level 2 fusion rule, level 3 fusion rule, and level 4 rule are applied in sequence. A successful match is defined as satisfying all the conditions of at least one rule in the corresponding rule set, ensuring that each patient corresponds to only one grade result.

9. An emergency triage method based on collaborative decision-making using a large language model and a tree model, employing the system described in any one of claims 1-8, characterized in that, The method includes the following steps: (1) Construct a dataset based on patient information and filter the data; (2) Train an initial random forest model based on the dataset to generate basic decision rules, and extract high-frequency features based on association rules to combine them to obtain a candidate rule pool; construct a domain-adapted prompt word structure to drive LLM to complete rule correction and form a set of corrected rules; (3) Based on the modified rule set, perform multiple rounds of random rule partitioning and rule combination generation, and then select the optimal rule; (4) Based on specific scenarios, the expert rules and the modified optimal rules are integrated to match patients and realize emergency pre-examination and triage.

Images