Method, system and application for improving reliability of medical decision

Abstract

The invention provides a method and system for improving reliability of medical decision and application. The method comprises the following steps: constructing a data set covering any disease according to medical data; training the medical vertical large model according to the data set to obtain a medical decision set; performing priority ranking on any decision in the medical decision set by constructing a multi-dimensional evaluation system to obtain a primary medical decision ranking result; and dynamically adjusting a weight parameter and a decision rule of any medical decision in the primary medical decision sorting result by taking the feedback suggestions and the actual diagnosis and treatment effects as reward signals to obtain a final medical decision sorting. According to the method, the problems of data fragmentation and illusion faced by the current large medical model and the industrial pain points of insufficient application suitability are effectively solved, and the clinical diagnosis and treatment efficiency and quality are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present disclosure belongs to the field of biological computers, and particularly relates to a method, system and application for improving the reliability of medical decision-making, aiming to solve data problems through a cross-professional knowledge fusion system, introduce a dynamic reordering algorithm to suppress model "hallucinations", and promote the application of mathematical models in medicine through multi-scenario optimization and multi-dimensional verification. BACKGROUND

[0002] In recent years, artificial intelligence technology has been increasingly widely applied in the field of medical health. Medical vertical large models, with their powerful data processing and analysis capabilities, have gradually become key technical supports for assisting clinical decision-making and optimizing diagnosis and treatment processes. In the aspects of disease diagnosis, treatment plan formulation, and prognosis evaluation, medical vertical large models have shown great application potential, bringing new opportunities to improve the quality and efficiency of medical services.

[0003] However, the current medical vertical large models face many industry pain points that need to be solved in the actual landing process. At the data level, medical professional data is scattered in different institutions and systems, and the data annotation cost is high. Moreover, it is difficult to efficiently convert massive heterogeneous clinical data into a knowledge system that can be understood and utilized by models, resulting in semantic understanding deviations of the models. At the technical level, model "hallucination" problems occur frequently, and traditional reordering techniques cannot flexibly adapt to the complex and variable needs of medical scenarios, affecting the reliability of model output results. At the application level, existing models lack depth optimization for medical scenarios, the decision logic lacks explainability, and knowledge updates lag behind clinical practice, which seriously restricts their large-scale application in the clinic.

[0004] Therefore, in order to solve the above problems, a method for promoting the safe and reliable application of medical vertical large models in specific medical scenarios is urgently needed. SUMMARY

[0005] The present disclosure aims to provide a method, system and application for improving the reliability of medical decision-making, which at least solves one technical problem in the prior art.

[0006] The technical solution of the present disclosure is:

[0007] A method for improving the reliability of medical decision-making, based on dynamic sorting and medical vertical large model fusion, comprising:

[0008] Collecting medical data;

[0009] Constructing a diagnosis data set covering any disease according to the medical data;

[0010] Constructing a medical vertical large model; and training the medical vertical large model according to the data set to obtain a medical decision set;

[0011] With real-time scene features and clinical evidence intensity as the core calibration basis, a multi-dimensional evaluation system is constructed to prioritize any decision in the medical decision set, obtaining a primary medical decision ranking result;

[0012] Obtain feedback from clinicians on the primary medical decision ranking result and the actual diagnosis and treatment effect of any medical decision in the primary medical decision ranking result;

[0013] The feedback and actual diagnosis and treatment effect are used as reward signals to dynamically adjust the weight parameters and decision rules of any medical decision in the primary medical decision ranking result, obtaining the final medical decision ranking;

[0014] Select the medical decision ranked first or second in the medical decision ranking as the final medical decision.

[0015] The medical data is collected; the medical data constructs a diagnosis data set covering any disease, including:

[0016] Collect a preset number of electronic medical record data and corresponding medical images;

[0017] Use natural language processing technology to structure and analyze the electronic medical record data, extract key entities and relationships;

[0018] Use image recognition algorithms to extract features from the medical images and identify features of the lesion area;

[0019] Use data cleaning algorithms to process the key entities and relationships, and the features of the lesion area, remove duplicate values and incorrect data, and obtain the data set.

[0020] The medical vertical large model is constructed; the medical vertical large model is trained according to the data set, obtaining a medical decision set, including:

[0021] Use QWEN-2.5-VL-32B multi-modal model as the basic architecture;

[0022] Use deep learning algorithms to pre-train based on general medical data, enabling the basic architecture to master basic semantics and image feature representation in the medical field, obtaining a basic model;

[0023] Fine-tune the basic model with data, deeply integrate specialist knowledge base data, include the clinical indicators of the patient to be analyzed in the training process, construct a preliminary model, and give the preliminary model the ability to analyze medical decisions from the perspective of disease prognosis;

[0024] Use a multi-objective optimization algorithm to adjust the parameters of the preliminary model, obtaining the medical vertical large model;

[0025] Obtaining the medical decision set according to the data set by using the medical vertical large model.

[0026] The real-time scene characteristics and the clinical evidence strength are used as the core calibration basis, a multi-dimensional evaluation system is constructed, and any decision in the medical decision set is prioritized to obtain a primary medical decision sorting result, including:

[0027] Based on the initial weight setting in the multi-dimensional evaluation system, the medical decision set output by the medical vertical large model is sorted to obtain a primary medical decision sorting result;

[0028] The sorting score in the primary medical decision sorting result is obtained by:

[0029] Score initial，i =w E ·E i +w P ·P i +w S ·S i ;

[0030] Wherein, w E , w P , w S are the initial weights of the clinical evidence strength, the patient individual characteristics and the real-time scene characteristics respectively; E i is the clinical evidence strength feature vector, P i is the patient individual characteristics vector, and S i is the real-time scene characteristics vector.

[0031] The multi-dimensional system includes:

[0032] Extracting clinical evidence related features from the data set;

[0033] The medical data of the patient individual is converted into quantifiable numerical features by a data standardization method, and the diagnosis and treatment scene information is classified and coded to be converted into a scene feature vector recognizable by the algorithm, including:

[0034] For the i-th output result, the multi-dimensional feature vector is represented as: X i =[E i , P i , S i ].

[0035] According to the reward signal received by the reinforcement learning mechanism, the primary medical decision sorting result is dynamically calibrated and adjusted to obtain the final medical decision sorting.

[0036] The feedback and actual diagnosis and treatment effect are taken as reward signals, and the weight parameter and decision rule of any medical decision in the primary medical decision sorting result are dynamically adjusted to obtain a final medical decision sorting, including:

[0037] The feedback and actual diagnosis and treatment effect are taken as reward signals;

[0038] According to the strength and direction of the reward signal, the weight values of factors in a multi-dimensional evaluation system are adjusted, the primary medical decision sorting result is recalculated and sorted, and dynamic weight updating is achieved.

[0039] The dynamic weight updating formula is:

[0040]

[0041] Wherein, w t is a weight vector at t-1, and a is a learning rate, is a reward function gradient calculated based on the weight at t-1.

[0042] The primary medical decision sorting result is dynamically calibrated and adjusted to obtain a final medical decision sorting, including:

[0043] The sorting score of the final medical decision sorting after dynamic calibration is:

[0044] Score initial,i =w t,E ·E i +w t，P ·P i +w i,S ·S i ;

[0045] Wherein, E i is a clinical evidence strength feature vector, P i is a patient individual feature vector, S i is a real-time scene feature vector; W t，E is the weight of the clinical evidence strength feature at t, W t，P is the weight of the patient individual feature at t, and W t，S is the weight of the real-time scene feature at t.

[0046] A system for improving the reliability of medical decisions, comprising:

[0047] A collection module for collecting medical data;

[0048] A data processing module connected with the collection module, configured to construct a diagnosis data set covering any disease according to the medical data;

[0049] The construction module is connected with the data processing module, and is used for constructing a medical vertical large model; and the medical vertical large model is trained according to the data set, and a medical decision set is output.

[0050] The preliminary sorting module is connected with the construction module, and is used for performing priority sorting on any decision in the medical decision set by constructing a multi-dimensional evaluation system, taking real-time scene features and clinical evidence strength as core calibration bases, to obtain a primary medical decision sorting result.

[0051] The dynamic sorting module is connected with the preliminary sorting module, and is used for obtaining feedback opinions of a clinician on the primary medical decision sorting result and actual diagnosis and treatment effects of any medical decision in the primary medical decision sorting result; and the feedback opinions and the actual diagnosis and treatment effects are taken as reward signals to dynamically adjust weight parameters and decision rules of any medical decision in the primary medical decision sorting result, to obtain a final medical decision sorting.

[0052] The output module is connected with the dynamic sorting module, and is used for selecting a medical decision ranked first or second in the medical decision sorting as a final medical decision.

[0053] An electronic device for improving reliability of medical decisions, comprising:

[0054] A storage medium for storing a computer program,

[0055] A processing unit in data exchange with the storage medium, for performing the computer program by the processing unit when making medical decisions, to perform steps of the method for improving reliability of medical decisions according to any one of claims 1-7.

[0056] An application of the method for improving reliability of medical decisions as described above to esophageal cancer medical decisions.

[0057] The present disclosure has at least the following beneficial effects:

[0058] The method for improving reliability of medical decisions according to the present disclosure, through the fusion of dynamic reordering technology and medical vertical large model, constructs a “data-technology-application” trinity solution, effectively solving the current industry pain points of data fragmentation, “illusion” problem and insufficient application adaptability faced by medical large models. In terms of clinical application value, it can reduce the occurrence of misdiagnosis and missed diagnosis, provide standardized and personalized diagnosis and treatment suggestions for clinicians, and significantly improve the efficiency and quality of clinical diagnosis and treatment. At the same time, through the method according to the present disclosure, medical resources can be reasonably arranged, unnecessary examinations and treatments can be avoided, and the utilization rate of medical resources can be improved. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 is a flowchart of the method for improving the reliability of medical decision-making described in the present disclosure;

[0060] Figure 2 is a flowchart of the medical vertical large model construction process;

[0061] Figure 3 is a flowchart of deep learning;

[0062] Figure 4 is a block diagram of the system for improving the reliability of medical decision-making described in the present disclosure. DETAILED DESCRIPTION

[0063] The present disclosure will be further described below with reference to the accompanying drawings.

[0064] The existing medical large model has a high frequency of "hallucination" phenomenon in the diagnosis of complex conditions, and the output diagnosis and treatment suggestions deviate greatly from the actual clinical situation. The multi-modal data fusion technology is not mature enough, and when integrating text, image, test and other types of data, it cannot fully tap the potential correlation between data, resulting in the model performance failing to meet the expectations, and the model accuracy greatly decreasing.

[0065] In view of the problems of the prior art, the present disclosure provides a method for improving the reliability of medical decision-making. Specific embodiment 1:

[0067] The present disclosure provides an embodiment:

[0068] As Figure 1 A method for improving the reliability of medical decision-making, based on dynamic sorting and medical vertical large model fusion, comprising: collecting medical data; constructing a diagnosis data set covering any disease according to the medical data; constructing a medical vertical large model; and training the medical vertical large model according to the data set to obtain a medical decision-making set; taking real-time scene features and clinical evidence strength as the core calibration basis, constructing a multi-dimensional evaluation system to prioritize any decision in the medical decision-making set, obtaining a primary medical decision-making sorting result; obtaining feedback opinions of clinicians on the primary medical decision-making sorting result and actual diagnosis and treatment effects of any medical decision in the primary medical decision-making sorting result; taking the feedback opinions and actual diagnosis and treatment effects as reward signals, dynamically adjusting the weight parameters and decision rules of any medical decision in the primary medical decision-making sorting result to obtain the final medical decision-making sorting; selecting the medical decision ranked first or second in the medical decision-making sorting as the final medical decision.

[0069] Specifically, the medical data is collected; the medical data constructs a diagnosis data set covering any disease, including: collecting a preset number of electronic medical record data and corresponding medical images; using natural language processing technology to structure the electronic medical record data, extract key entities and relationships; using image recognition algorithms to extract features from the medical images, identify features of the lesion area; using data cleaning algorithms to process the key entities and relationships, features of the lesion area, remove duplicate values and incorrect data, and obtain the data set.

[0070] This embodiment takes esophageal cancer as an example to explain the disclosed method in detail.

[0071] For the diagnosis and treatment of esophageal cancer, multi-channel collection of multi-source heterogeneous data is carried out, specifically covering the following types and sizes: a total of 16052 electronic medical records, including patient basic information such as age, gender, place of origin, chief complaint, history of present illness, past history, physical examination records, diagnosis conclusion, etc. Text information, including 9009 initial diagnosis medical records and 7043 follow-up medical records; 38525 high-definition esophageal image data, including 27289 gastroscopy images and 11236 esophagography images, with a resolution of 1920x1080 pixels, covering different lesion stages, including early, medium and late stages, and different lesion types such as squamous cell carcinoma and adenocarcinoma; 2568 pathological examination reports, including morphological description of pathological biopsy tissue, immunohistochemical indicators such as CK19 and P53, test results, and pathological staging; 24078 laboratory examination data, including blood routine, biochemical indicators, tumor markers such as CEA and SCC, and test results; 4816 follow-up data, recording postoperative recovery, recurrence and metastasis information, survival time and other prognosis-related data.

[0072] In this embodiment, natural language processing technology can be used to structure the 16052 electronic medical record texts and extract key entities such as disease names, symptoms, drugs, and relationships such as symptom-disease associations and drug-dose correspondences, with an accuracy of over 92%; image recognition algorithms are used to extract features from 38525 esophageal images, identify the location, size, shape, and other features of the lesion area, and the feature extraction accuracy is over 88%, constructing a data set covering the entire process of esophageal cancer diagnosis, treatment, and prognosis.

[0073] And, a data cleaning algorithm is used to remove repeated values and error data in the data, and the data is standardized based on medical industry standards such as HL7 and DICOM, and the data format and coding rules are unified to eliminate format differences between different data sources. The medical knowledge graph technology is introduced, the processed data is structured and integrated with 3 authoritative esophageal cancer diagnosis and treatment guidelines such as NCCN esophageal cancer clinical practice guidelines, CSCO esophageal cancer diagnosis and treatment guidelines, 1605 kinds of drug information including indications, contraindications, usage and dosage, 963 disease association knowledge such as complications and comorbidities, to build a dynamic updated esophageal cancer specialist knowledge base, and the total data amount of the knowledge base is 25.7GB, of which the structured knowledge accounts for 65% and the unstructured knowledge accounts for 35%, providing precise and comprehensive knowledge reserves for medical vertical large models.

[0074] Specifically, the medical vertical large model is constructed; the medical vertical large model is trained according to the data set to obtain a medical decision set, including: Figure 2 QWEN-2.5-VL-32B multi-modal model is selected as the basic architecture, deep learning algorithm is used, and pre-training is carried out based on massive general medical data to enable the model to master the core capabilities such as basic semantics and image feature representation in the medical field. In the fine-tuning stage, the esophageal cancer specialist knowledge base data is deeply integrated, the survival time and disease progression of esophageal cancer patients are included in the model training, and the medical vertical large model is constructed to enable the model to analyze the diagnosis and treatment scheme from the perspective of disease prognosis.

[0075] Moreover, a multi-objective optimization algorithm is used to improve the diagnosis accuracy, reduce the incidence of hallucinations, and enhance the explainability of decision logic as optimization goals, and the model parameters are finely adjusted to realize performance breakthroughs of the model in esophageal imaging analysis and diagnosis and treatment scheme generation, and to ensure that the model output meets the actual needs of the clinic.

[0076] Specifically, the real-time scene features and the clinical evidence strength are used as the core calibration basis, and a multi-dimensional evaluation system is constructed to prioritize any decision in the medical decision set to obtain a primary medical decision sorting result, including: based on the initial weight setting in the multi-dimensional evaluation system, the medical decision set output by the medical vertical large model is sorted to obtain a primary medical decision sorting result; the sorting score in the primary medical decision sorting result is obtained by:

[0077] Score initial,i =W E ·E i +w P ·P i +w S ·S i ;

[0078] wherein w E , w P , and w S are initial weights of the clinical evidence intensity, the patient individual characteristics, and the real-time scene characteristics, respectively; E i is a clinical evidence intensity feature vector, P i is a patient individual feature vector, and S i is a real-time scene feature vector.

[0079] Preferably, the multi-dimensional system comprises: extracting clinical evidence related features from the data set; converting patient individual medical data into quantifiable numerical features through a data standardization method, and classifying and encoding diagnosis and treatment scene information into a scene feature vector recognizable by an algorithm; and dynamically calibrating and adjusting the primary medical decision ranking result according to the reward signal received by the reinforcement learning mechanism to obtain the final medical decision ranking. For the multi-dimensional system, the i-th output result is represented by a multi-dimensional feature vector X i = [E i , P i , S i ]; and the flowchart of the reinforcement learning mechanism can refer to Figure 3 .

[0080] Specifically, the feedback opinions and actual diagnosis and treatment effects are taken as reward signals to dynamically adjust the weight parameters and decision rules of any medical decision in the primary medical decision ranking result to obtain the final medical decision ranking, which comprises: taking the feedback opinions and actual diagnosis and treatment effects as reward signals; adjusting the weight values of various factors in the multi-dimensional evaluation system according to the strength and direction of the reward signals, recalculating and ranking the primary medical decision ranking result to achieve dynamic weight updating; wherein the dynamic weight updating formula is: wherein w t - is the weight vector at t-1 time, a is the learning rate, is the reward function gradient calculated based on the weight at t-1 time.

[0081] Specifically, the dynamic calibration and adjustment of the primary medical decision ranking result to obtain the final medical decision ranking comprises: the ranking score of the final medical decision ranking after dynamic calibration is Score initial，i = w t，E ·E i + w t，P ·P i + w t，S ·S i ; wherein E i is a clinical evidence intensity feature vector, P i is a patient individual feature vector, and S iis the real-time scene feature vector; W t，E is the weight of the clinical evidence strength feature at time t, W t，P is the weight of the patient individual feature at time t, W t，S is the weight of the real-time scene feature at time t.

[0082] The method described in this embodiment uses the dynamic reordering algorithm as a key intermediate processing module of the medical vertical large model, realizing the deep integration of the two. The medical vertical large model is used to first analyze and process the input medical data, generate preliminary diagnosis and treatment suggestions, diagnosis results, etc. as output results, constitute primary medical decisions, and perform sorting to obtain primary medical decision sorting results. Then, the dynamic reordering algorithm receives the primary medical decision sorting results and real-time collected multi-dimensional feature data, reorders and calibrates the primary medical decision sorting results through the above core mechanisms and implementation steps, and obtains the final medical decision. Finally, the final medical decision is provided as the final output to the clinician, forming a complete "model analysis-result calibration-clinical application" process, and realizing the effective improvement of the dynamic reordering technology on the output quality of the medical vertical large model.

[0083] Verification link:

[0084] This link trains the dynamic reordering algorithm by simulating medical scenes of different complexity. A diverse simulated scenario dataset is constructed using historical clinical data, covering common diseases, rare diseases, complex complications, and other diagnosis and treatment situations. During the training process, algorithm parameters are constantly adjusted, and the multi-dimensional evaluation system and reinforcement learning mechanism are optimized to enable the algorithm to adapt to the needs of different scenarios. At the same time, internal and external verification methods are combined to comprehensively evaluate the algorithm performance. Internal verification uses simulated scenario data in the test set to evaluate the algorithm's performance indicators such as the suppression effect of the model's "hallucination" and the accuracy of the sorting results; external verification verifies the effectiveness and reliability of the algorithm in different medical institutions and different populations through multi-center clinical trials, ensuring that the algorithm can function well in real medical scenarios.

[0085] In this link, the dynamic reordering algorithm is selected in the esophageal cancer diagnosis and treatment scene, which can be connected to the electronic medical record system and the picture archiving and communication system (PACS) in real time. The algorithm automatically extracts the patient's medical history information, image features, and pathology report data, and generates personalized diagnosis and treatment suggestions after calibration, including examination item recommendations, treatment plan selection, and prognosis risk assessment, to assist doctors in developing precise diagnosis and treatment plans.

[0086] To comprehensively and scientifically evaluate the performance and clinical application value of the constructed model, this study adopts a multi-dimensional verification system combining internal verification, external verification, and clinical practicability evaluation, and clearly presents the evaluation indicators, methods, and expected goals at each stage in table form.

[0087] I) Internal validation

[0088] Internal validation is mainly based on the test set data to preliminarily evaluate the performance of the model, and the cross-validation method is used to improve the reliability of the evaluation. The specific evaluation indicators, methods and expected targets are shown in the following table:

[0089] Table 1: Internal validation table:

[0090] Evaluation index Evaluation method Expected target Accuracy Cross-validation (5-fold) Accuracy ≥ 90% Sensitivity Confusion matrix calculation Sensitivity ≥ 85% Specificity Confusion matrix calculation Specificity ≥ 88% AUC value ROC curve analysis AUC value ≥ 0.92 "hallucination" occurrence rate False output proportion statistics "hallucination" occurrence rate ≤ 5% Decision interpretability SHAP value analysis Key decision feature interpretability rate ≥ 90%

[0091] In the internal validation process, as shown in Table 1, the data set is randomly divided into 5 subsets, and 4 of them are selected as the training set and 1 as the validation set in turn. After repeating the experiment 5 times, the average value is taken as the final evaluation result. By analyzing the difference between the output results of the method described in this embodiment and the actual clinical diagnosis, the shortcomings of this embodiment in feature extraction, parameter setting, etc. are located and optimized accordingly.

[0092] Evaluation results: Through 5-fold cross-validation, the accuracy is expected to reach 92%-93%, higher than the expected target of 90%; the sensitivity is about 87%-88%, meeting the expected requirement of 85%; the specificity can reach 89%-90%, exceeding the expected standard of 88%; the AUC value is between 0.93-0.94, better than the expected target of 0.92; the "hallucination" rate can be controlled at 3%-4%, lower than the expected upper limit of 5%; through SHAP value analysis, the key decision-making feature can be explained about 92%-93%, in line with the expected ≥90%. Overall, the internal validation results show that this embodiment performs well in basic performance indicators, laying a solid foundation for subsequent external validation and clinical application.

[0093] II) External validation

[0094] External validation verifies the generalization ability of this embodiment in different medical environments and populations through multi-center clinical trials. The specific content is as follows:

[0095] Table 2: External validation table:

[0096]

[0097]

[0098] As shown in Table 2, external validation will collect independent data sets from at least 3 different regional tertiary hospitals, covering outpatient, inpatient and emergency cases in multiple scenarios, and compare the differences in diagnostic efficiency and accuracy between this embodiment and traditional clinical diagnosis methods, to verify the stability and effectiveness of this embodiment in diversified medical scenarios.

[0099] Evaluation results: In the cross-institutional effectiveness verification, the fluctuation range of accuracy, sensitivity, and specificity of the embodiment is expected to be 3%-4% in external independent case data testing in 3 different tertiary hospitals, which is less than the expected fluctuation range of 5%, indicating that the embodiment has good cross-institutional applicability. In population adaptability verification, for different regional and age subgroup patient data, the AUC value of each subgroup is expected to be between 0.91-0.93, all higher than the expected target of 0.90, indicating that the embodiment has good adaptability to different populations. Compared with traditional methods, the diagnostic efficiency improvement rate of the embodiment is expected to reach 35%-40%, exceeding the expected 30%, which can significantly improve the clinical diagnostic efficiency. In multi-scenario stability verification, in outpatient, inpatient, and emergency scenarios, the scene adaptation accuracy of the embodiment is expected to be 87%-89%, higher than the expected 85%, which reflects the stability performance of the embodiment in multiple scenarios.

[0100] Three), clinical practicability evaluation

[0101] The clinical practicability evaluation focuses on the actual clinical application value of the model, optimizes the model usability through expert evaluation and user feedback, as follows:

[0102] Table 3: Clinical practicability evaluation table:

[0103]

[0104]

[0105] As shown in Table 3, at least 5 physicians with rich clinical experience are invited to conduct blind evaluation on the diagnosis and treatment suggestions of the embodiment, to evaluate the compliance with clinical standards; feedback on the use experience of the embodiment is collected through questionnaire survey of doctors and patients; the stability of the embodiment in continuous operation state and the response speed to the latest clinical guidelines are tested, to comprehensively evaluate the clinical practicability of the embodiment.

[0106] Evaluation results: In terms of diagnosis and treatment recommendation compliance, the average score of blind evaluation by 5 or more clinicians is expected to be 8.7-8.9, higher than the expected target of 8.5, indicating that the diagnosis and treatment recommendations output by this embodiment have high compliance with clinical standards. In the user satisfaction survey, the satisfaction of doctors and patients is expected to reach 92%-94%, exceeding the expected 90%, reflecting the users' recognition of the model's use experience. In terms of operational convenience, the time taken by clinicians to complete a single case operation is expected to be 2.5-2.8 minutes, less than the expected 3 minutes, so this embodiment is convenient and efficient. In system stability testing, the fault-free operation time of this embodiment is expected to reach 1200-1500 hours, much higher than the expected 1000 hours, and the system stability is good. In terms of knowledge updating timeliness, the delay of key knowledge updating for the latest clinical guidelines is expected to be 48-60 hours, lower than the expected 72 hours, and the system can keep up with the updating of clinical knowledge in a timely manner.

[0107] In summary, the method described in the present disclosure can construct a dynamically updated medical knowledge base that integrates cross-professional knowledge, effectively integrate multi-source heterogeneous medical data, solve the problem of data fragmentation, and provide a dynamic reordering algorithm that adapts to the medical scene, significantly suppressing the "hallucination" problem of large models and improving the reliability and explainability of model decision-making. At the same time, the present disclosure forms a complete system for building, verifying, and evaluating medical vertical large models, ensuring that the models meet the requirements in terms of diagnosis accuracy and clinical practicality. Moreover, the method described in the present disclosure improves the efficiency and quality of clinical diagnosis and treatment, providing practical reference for promoting the development of medical artificial intelligence technology and the progress of the medical and health industry. Specific embodiment 2:

[0109] The present disclosure also provides an embodiment:

[0110] As Figure 4A system for improving the reliability of medical decisions, based on specific embodiment 1, the structure of which comprises: a collection module 100, a data processing module 200, a construction module 300, a preliminary sorting module 400, a dynamic sorting module 500, and an output module 600; wherein the collection module 100 is used to collect medical data; the data processing module 200 is connected with the collection module 100, and is used to construct a data set covering the whole process of diagnosis, treatment, and prognosis of any disease according to the medical data; the construction module 300 is connected with the data processing module 200, and is used to construct a medical vertical large model; and the medical vertical large model is trained according to the data set, and a medical decision set is output; the preliminary sorting module 400 is connected with the construction module 300, and is used to take real-time scene characteristics and clinical evidence strength as the core calibration basis, and prioritize any decision in the medical decision set by constructing a multi-dimensional evaluation system to obtain a primary medical decision sorting result; the dynamic sorting module 500 is connected with the preliminary sorting module 400, and is used to obtain feedback opinions of clinicians on the primary medical decision sorting result and actual diagnosis and treatment effects of any medical decision in the primary medical decision sorting result; and the feedback opinions and actual diagnosis and treatment effects are taken as reward signals to dynamically adjust weight parameters and decision rules of any medical decision in the primary medical decision sorting result to obtain a final medical decision sorting; the output module 600 is connected with the dynamic sorting module 500, and is used to select a medical decision ranked first or second in the medical decision sorting as the final medical decision. Specific embodiment 3:

[0112] The present disclosure also provides an embodiment:

[0113] An electronic device comprising: a storage medium and a processing unit; wherein the storage medium is used to store a computer program, the processing unit exchanges data with the storage medium, and is used to execute the computer program by the processing unit when making a medical decision, to perform the steps of the method for improving the reliability of medical decisions as described in specific embodiment 1.

[0114] The present disclosure further provides an embodiment:

[0115] A computer-readable storage medium having a computer program stored therein; when the computer program is executed, the steps of the method for improving the reliability of medical decisions as described in specific embodiment 1 are performed.

[0116] In this disclosure, a computer readable storage medium can be any tangible medium that contains or stores a program for use by or in connection with an instruction execution system, apparatus, or device. In this disclosure, a computer readable signal medium can include a propagated data signal with computer readable program code embodied therein, for use by or in connection with an instruction execution system, apparatus, or device. The propagated data signal can take any number of forms, including, but not limited to, electro-magnetic, optical, or any suitable combination thereof. A computer readable signal medium can be any computer readable medium that is not a computer readable storage medium and that can communicate, propagate or transport program for use by or in connection with an instruction execution system, apparatus, or device. Program code embodied on a computer readable storage medium can be transmitted using any appropriate medium, including but not limited to wireless, wire line, optical fiber cable, RF, etc., or any suitable combination of the foregoing.

[0117] The present disclosure details the method of the cooling effect of the wetland park on the surrounding different local climate zoning, and predicts the building energy consumption of the wetland park surrounding different local climate zoning, so as to optimize the operation scheme of the air conditioning system of the building surrounding the wetland park, and provide strategies for the urban landscape pattern planning.

[0118] The above disclosure is only a few specific implementation scenarios of the present disclosure, but the present disclosure is not limited thereto, and any changes that can be thought of by those skilled in the art shall fall within the protection scope of the present disclosure. The above disclosure serial number is only for description, not representing the advantages and disadvantages of the implementation scenario.

Claims

1. A method for improving the reliability of medical decision-making, based on dynamic ranking and fusion with medical vertical large models, characterized in that, The method comprises the following steps: collecting medical data; constructing a diagnosis data set covering any disease according to the medical data; constructing a medical vertical large model; training the medical vertical large model according to the data set to obtain a medical decision set; taking real-time scene characteristics and clinical evidence strength as core calibration basis, constructing a multi-dimensional evaluation system to prioritize any decision in the medical decision set, and obtaining a primary medical decision ranking result; obtaining feedback opinions of clinical doctors on the primary medical decision ranking result and actual diagnosis and treatment effects of any medical decision in the primary medical decision ranking result; taking the feedback opinions and actual diagnosis and treatment effects as reward signals, dynamically adjusting the weight parameters and decision rules of any medical decision in the primary medical decision ranking result, and obtaining a final medical decision ranking; selecting a medical decision ranked first or second in the medical decision ranking as a final medical decision.

2. The method of claim 1, wherein, The method comprises the following steps: collecting a preset number of electronic medical record data and corresponding medical images; using natural language processing technology to structurally analyze the electronic medical record data, extracting key entities and relationships; using image recognition algorithms to extract features of the medical images and identify features of the lesion area; using data cleaning algorithms to process the key entities and relationships, features of the lesion area, remove duplicate values and incorrect data, and obtain the data set.

3. The method of claim 1, wherein the method further comprises: The method comprises the following steps: taking QWEN-2.5-VL-32B multi-modal model as the basic architecture; using deep learning algorithms to pre-train based on general medical data, enabling the basic architecture to master the basic semantics and image feature representation capabilities in the medical field, and obtaining a basic model; data fine-tuning the basic model, deeply integrating specialty knowledge base data, incorporating the clinical indicators of the patient to be analyzed into the training process, constructing a preliminary model, and giving the preliminary model the ability to analyze medical decisions from the perspective of disease prognosis; using multi-objective optimization algorithms to adjust the parameters of the preliminary model to obtain the medical vertical large model; using the medical vertical large model to obtain the medical decision set based on the data set.

4. The method of claim 1, wherein, The method comprises the following steps: based on the initial weight settings in the multi-dimensional evaluation system, sorting the medical decision set output by the medical vertical large model to obtain a primary medical decision ranking result; the sorting score acquisition method in the primary medical decision ranking result is: Score initial,i = w E · E i + w P · P i + w S · S i ; wherein w E , w P , w S are initial weights of the clinical evidence strength, the patient individual characteristics, and the real-time scene characteristics, respectively; E i is a clinical evidence strength feature vector, P i is a patient individual characteristics feature vector, and S i is a real-time scene feature vector.

5. The method of claim 1 or 4, wherein, the multi-dimensional system comprises: extracting clinical evidence-related features from the data set; using data standardization methods to convert the medical data of individual patients into quantifiable numerical features, and classifying and encoding the diagnosis and treatment scene information into scene feature vectors recognizable by algorithms, including: For the i-th output result, its multi-dimensional feature vector is represented as: X i = [E i , P i , S i ]. According to the reward signal received by the reinforcement learning mechanism, the primary medical decision ranking result is dynamically calibrated and adjusted to obtain the final medical decision ranking.

6. The method of improving reliability of medical decisions according to claim 1, characterized in that, The feedback and the actual diagnosis and treatment effect are taken as the reward signal, and the weight parameter and the decision rule of any medical decision in the primary medical decision ranking result are dynamically adjusted to obtain the final medical decision ranking, which comprises: The feedback and the actual diagnosis and treatment effect are taken as the reward signal; According to the strength and direction of the reward signal, the weight value of each factor in the multi-dimensional evaluation system is adjusted, and the primary medical decision ranking result is recalculated and ranked to achieve dynamic weight updating; The dynamic weight updating formula is: where w t is the weight vector at time t, and a is the learning rate, is the reward function gradient based on the weight computation at time t-1.

7. The method of improving reliability of medical decisions according to claim 6, characterized in that, The primary medical decision ranking result is dynamically calibrated and adjusted to obtain the final medical decision ranking, which comprises: The ranking score of the final medical decision ranking after dynamic calibration is: Score initial,i = w t,E · E i + w t,P · P i + w t,S · S i ; wherein E i is the clinical evidence strength feature vector; P i is the patient individual feature vector; S i is the real-time scene feature vector; w t,E is the weight of the clinical evidence strength feature at time t; w t,P is the weight of the patient individual feature at time t; w t,S is the weight of the real-time scene feature at time t.

8. A system for improving reliability of medical decisions, based on the method for improving reliability of medical decisions according to any one of claims 1 to 7, characterized in that, It comprises: A collection module for collecting medical data; A data processing module connected with the collection module for constructing a data set covering the whole process of diagnosis, treatment and prognosis of any disease according to the medical data; A construction module connected with the data processing module for constructing a medical vertical large model; And training the medical vertical large model according to the data set to output a medical decision set; A preliminary ranking module connected with the construction module for prioritizing any decision in the medical decision set by constructing a multi-dimensional evaluation system with real-time scene characteristics and clinical evidence strength as the core calibration basis to obtain a primary medical decision ranking result; A dynamic ranking module connected with the preliminary ranking module for obtaining the feedback of the primary medical decision ranking result from the clinician and the actual diagnosis and treatment effect of any medical decision in the primary medical decision ranking result; and taking the feedback and the actual diagnosis and treatment effect as the reward signal to dynamically adjust the weight parameter and the decision rule of any medical decision in the primary medical decision ranking result to obtain the final medical decision ranking; An output module connected with the dynamic ranking module for selecting the medical decision ranked first or second in the medical decision ranking as the final medical decision.

9. An electronic device for improving the reliability of medical decision-making, characterized in that, It comprises: A storage medium for storing a computer program, A processing unit for data exchange with the storage medium, for executing the computer program through the processing unit when making medical decisions, and performing the steps of the method for improving the reliability of medical decisions according to any one of claims 1-7.

10. The application of the method for improving the reliability of medical decisions according to any one of claims 1-7 in the medical decision of esophageal cancer.

Images