An automated assisted MDT consultation method based on LLM

Abstract

The present invention belongs to the field of medical auxiliary consultation technology, and specifically relates to an MDT automated auxiliary consultation method based on LLM. By establishing a patient personal case encryption system, obtaining case data to generate prompt words, and performing anthropomorphic and non-anthropomorphic multidisciplinary auxiliary diagnosis and treatment through the LLM model, an optimized auxiliary diagnosis and treatment plan is obtained. The LLM model and the doctor's diagnosis and treatment report are compared and analyzed to evaluate the diagnostic accuracy and rationality of the treatment plan, and the best LLM auxiliary diagnosis and treatment model is screened; the current patient case data is obtained and input into the best LLM auxiliary diagnosis and treatment model to generate an auxiliary diagnosis and treatment plan. The diagnostic accuracy and rationality of the treatment plan of the joint diagnosis and treatment of multiple departments are improved. By developing data integration technology, the data barriers between departments are broken down and the standardized processing of multi-source data is achieved. At the same time, the transparency and credibility of LLM in complex medical scenarios are enhanced, a collaboration mechanism between LLM and human doctors is established, and the best collaboration model is explored.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical auxiliary consultation, and in particular relates to an LLM-based MDT automated auxiliary consultation method. Background Art

[0002] With the rapid development of large language model (LLM) technology, its application in the medical field has gradually become a research hotspot, especially in multi-disciplinary diagnosis and treatment models. Multi-disciplinary diagnosis and treatment (MDT) refers to a model that integrates insights from various disciplines to develop optimal treatment plans for patients with complex diseases. In recent years, LLM has demonstrated high accuracy in single-department diagnoses. For example, in imaging diagnosis of lung cancer and breast cancer, LLM can quickly identify lesion characteristics and provide diagnostic recommendations. LLM also shows potential in medical record management, using natural language processing technology to automatically organize and summarize medical records, improving the efficiency of clerical work.

[0003] However, despite certain progress, the application of LLM in multi-departmental joint diagnosis and treatment still faces many challenges. First, the data types generated by different departments are diverse and complex, and the data formats and standards are inconsistent, which makes it difficult for LLM to integrate multi-source data. Second, LLM's auxiliary diagnostic capabilities in handling complex cases still need to be further verified and improved, especially in multi-department collaboration, where its interdisciplinary adaptability and flexibility are insufficient. In addition, the collaborative mechanism between LLM and human doctors is not yet mature, and how to effectively apply LLM in actual clinical practice still needs to be explored.

[0004] Existing technologies used in multidisciplinary collaborative diagnosis and treatment primarily focus on simple data aggregation and preliminary analysis, lacking the ability to deeply integrate and analyze multi-source data. For example, when processing multidisciplinary data, LLM struggles to ensure data accuracy and consistency, limiting its diagnostic capabilities in complex cases. Furthermore, existing research has largely focused on the application of a single LLM model, lacking research on combinations of different LLM models, making it difficult to fully realize the potential of LLM. These technical bottlenecks limit the practical application of LLM in multidisciplinary collaboration and urgently need to be addressed.

[0005] To address the shortcomings of existing technologies, this paper proposes a method for optimizing LLM model combinations, aiming to improve diagnostic accuracy and the rationality of treatment plans in multi-disciplinary collaborative diagnosis and treatment. By developing data integration technology, this method breaks down data barriers between departments and achieves standardized processing of multi-source data. Furthermore, it introduces model interpretation technology to enhance the transparency and credibility of LLM in complex medical scenarios. Furthermore, a collaborative mechanism between LLM and human physicians is established to explore optimal collaborative models. Summary of the Invention

[0006] The purpose of this invention is to provide an automated, assisted MDT consultation method based on LLM to improve diagnostic accuracy and the rationality of treatment plans in multi-departmental collaborative diagnosis and treatment. By developing data integration technology, this method breaks down data barriers between departments and achieves standardized processing of multi-source data. Furthermore, model interpretation technology is introduced to enhance the transparency and credibility of LLM in complex medical scenarios. Furthermore, a collaborative mechanism between LLM and human doctors is established to explore optimal collaborative models.

[0007] In order to solve the above technical problems, the technical solutions adopted by the present invention are as follows:

[0008] An LLM-based MDT automated assisted diagnosis method comprises the following steps:

[0009] S1: Based on medical big data, obtain historical patient case data throughout their life cycle, and integrate it into the patient case integration encryption system after preprocessing;

[0010] S2: Obtain patient case data, establish a patient personal case encryption system, and classify patient cases;

[0011] S3: Retrieving the patient's medical records from the patient's personal medical record encryption system to generate at least two prompt words, the two prompt words including a first prompt word group and a second prompt word group, entering the first prompt word group into a non-anthropomorphic LLM group, and entering the second prompt word group into an anthropomorphic LLM group;

[0012] S4: An optimized auxiliary diagnosis and treatment plan is obtained through anthropomorphic and non-anthropomorphic multidisciplinary auxiliary diagnosis and treatment using the LLM model;

[0013] S5: Compare and analyze the LLM model with the doctor's diagnosis and treatment report, evaluate the diagnostic accuracy and rationality of the treatment plan, and screen the best LLM-assisted diagnosis and treatment model;

[0014] S6: Obtain the current patient case data and input it into the optimal LLM auxiliary diagnosis and treatment model to generate an auxiliary diagnosis and treatment plan.

[0015] Preferably, the specific process of preprocessing the historical patient's life cycle case data in step S1 is as follows:

[0016] S11: Detect duplicate values ​​in the full life cycle case data using a preset similarity function, where the similarity function formula is: , where A and B are the feature sets of two records;

[0017] S13: establishing time series data of the full life cycle case data based on the time sequence, and performing smoothing processing on the time series data;

[0018] S12: Identify missing values ​​for the full life cycle case data and fill in the missing values. The filling formula is: ,in x fill is the fill value, x i is a valid value, n is the number of valid values;

[0019] S13: Normalize the data after filling missing values. The formula is: x norm =( xx min ) / ( x max -x min ),in, x norm is the normalized value, x min is the minimum value, x max is the maximum value, x The data after filling missing values.

[0020] Preferably, step S1 further includes encrypting the pre-processed full life cycle case data by the patient case integration encryption system. The specific process is as follows:

[0021] S14: randomly generating an original key of a specified number of bits, performing key expansion on the original key, and dividing the original key into a 4×8 matrix;

[0022] S15: performing byte replacement, row shift, column mixing, and round constant addition operations on the matrix to generate a specified number of subkeys, each subkey being used for one encryption round;

[0023] S16: inputting plaintext data of a specified number of bits and the first subkey, dividing the plaintext data into a 4×4 matrix, and performing a specified round of encryption operation based on the 4×4 matrix.

[0024] Preferably, the specific process in step S15 is as follows:

[0025] S151: replacing each byte in the matrix according to a specified nonlinear transformation rule;

[0026] S152: Leave the first row of the matrix unchanged, shift the second row left by 1 byte, shift the third row left by 2 bytes, and shift the fourth row left by 3 bytes;

[0027] S153: Perform a linear transformation on the columns in the matrix, transforming the 4 bytes of each column through a specified matrix multiplication. The specific formula is as follows:

[0028] ;

[0029] in, a 0- a 3 are the elements in the original vector before the operation, b 0- b 3 are the elements of the output column vector obtained after the operation.

[0030] Preferably, obtaining the target patient case data in step S2 further includes decrypting the target patient case data, and the specific process is as follows:

[0031] S21: Perform reverse round constant addition on the target patient case data using the reversed extended key;

[0032] S22: Use the reversed extended key to perform a reverse column mixing operation on the target patient case data. The specific formula is as follows:

[0033] ;

[0034] in, a 0- a 3 are the elements in the original vector before the operation, b 0- b 3 are the elements of the output column vector obtained after the operation.

[0035] S23: Perform inverse column shift and inverse byte transformation operations, and finally perform an XOR operation with the first subkey through the inverse operation of the initial round to restore the plaintext. The specific formula is as follows:

[0036] ;

[0037] Among them, 0 x 1 B is an irreducible polynomial in the finite field GF(28) x 8 + x 4 + x 3 + x +1 means, a For elements, a ≪1: Yes a Perform a bitwise shift left by 1 bit. It is an exclusive OR operation.

[0038] Preferably, the specific process of obtaining patient case data and classifying patient cases in step S2 is as follows:

[0039] S24: establishing a real-time data interface, and obtaining patient case data based on the real-time data interface;

[0040] S25: Obtain the patient's disease type and current stage based on the patient's case data;

[0041] S26: Monitor changes in data sources in real time through API polling or message queues, push newly generated data to the patient's personal medical record encryption system through the API interface, and update the corresponding fields or records;

[0042] S27: Encrypt the newly generated data through steps S14-S16.

[0043] Preferably, the specific process of retrieving the patient's medical records from the patient's personal medical record encryption system and generating at least two prompt words in step S3 is as follows:

[0044] S31: Retrieving patient case data from the patient's personal case encryption system, including the patient's medical history, symptoms, and various examination data;

[0045] S32: Extract features from the patient case data and generate two prompt words based on the extracted features, including a first prompt word group and a second prompt word group. The first prompt word group covers the patient's medical history, chief complaint and examination information, and the second prompt word group is detailed information for a specified disease or complex case, including symptoms, examination information and existing diagnostic data.

[0046] Preferably, the specific process of performing anthropomorphic and non-anthropomorphic multidisciplinary auxiliary diagnosis and treatment by the LLM model in step S4 to obtain an optimized auxiliary diagnosis and treatment plan is as follows:

[0047] S41: The LLM model simulates the roles of doctors in different departments based on the input first prompt phrase group and second prompt phrase group, and generates a comprehensive auxiliary diagnosis and treatment plan from multiple angles;

[0048] S42: The LLM model analyzes the information of the prompt words in the first prompt word group and the second prompt word group, retrieves the built-in medical knowledge base, optimizes the auxiliary diagnosis and treatment plan, and generates an optimized auxiliary diagnosis and treatment plan.

[0049] Preferably, conflict detection and resolution methods are provided in both steps S41 and S42, and the specific process is as follows:

[0050] Detect data conflicts based on version control or timestamps;

[0051] The detected data conflicts are resolved according to preset rules, wherein the preset rules include last write priority or authoritative data source priority.

[0052] S5 also includes an optimization method for the LLM model, including the following process:

[0053] S51: establishing a target optimization function for evaluating the performance of the model combination, and defining variables of the target optimization function;

[0054] S52: Based on the variables of the target optimization function, a population of size N is selected, where N represents the number of individuals in the initial population;

[0055] S53: Randomly generate M model combinations, each combination is a binary vector of length n;

[0056] S54: Create a fitness function and calculate the fitness of each model combination. The specific formula of the fitness function is as follows:

[0057] ;

[0058] in, w 1, w 2, w 3 is the weight coefficient, which is adjusted according to actual needs. Diagnostic accuracy is an indicator that measures the proportion of correct diagnostic results in the medical diagnosis process. Treatment plan rationality is an evaluation of whether the treatment plan formulated for the patient's condition is scientific, effective, and suitable for the patient's actual situation. Information integration efficiency represents the efficiency of collecting, organizing, integrating, and utilizing various types of information in the medical process.

[0059] S55: Select individuals based on fitness ratio. The fitness ratio formula for each individual is as follows:

[0060] ;

[0061] in, f (x i ) is the i Individual x i The fitness function value of N is the total number of individuals in the population, f (x j ) is the j Individual x j The fitness function value of

[0062] S56: Randomly select individuals to enter the next generation according to the fitness ratio, randomly select two parent individuals x1 and x2, and exchange some of their genes at random positions to generate two offspring individuals;

[0063] ;

[0064] ;

[0065] in,c is a randomly selected crossover position used to determine the split point for exchanging genes between parent individuals. n is the number of locations;

[0066] S57: Randomly select an individual x , randomly select a position i ,Will x i The value changes from 0 to 1 or from 1 to 0;

[0067] S58: Generate a new population, and iteratively execute steps S54-S57 until the maximum number of iterations is reached or a preset stopping condition is met.

[0068] The beneficial effects of the present invention include:

[0069] The present invention provides an automated MDT assisted consultation method based on LLM. By establishing a patient personal case encryption system, obtaining case data to generate prompt words, and performing anthropomorphic and non-anthropomorphic multidisciplinary assisted diagnosis and treatment through the LLM model, an optimized auxiliary diagnosis and treatment plan is obtained. The LLM model is compared and analyzed with the doctor's diagnosis and treatment report to evaluate the diagnostic accuracy and rationality of the treatment plan, and the optimal LLM assisted diagnosis and treatment model is screened. The current patient's case data is obtained and input into the optimal LLM assisted diagnosis and treatment model to generate an auxiliary diagnosis and treatment plan. This method improves the diagnostic accuracy and rationality of treatment plans in multi-department joint diagnosis and treatment. By developing data integration technology, the data barriers between departments are broken down and standardized processing of multi-source data is achieved. At the same time, the transparency and credibility of LLM in complex medical scenarios are enhanced, a collaborative mechanism between LLM and human doctors is established, and the optimal collaboration model is explored. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 Schematic diagram of the process of the LLM-based MDT automated assisted diagnosis method of the present invention.

[0071] Figure 2 This is a flow chart of the present invention for obtaining patient case data and classifying patient cases.

[0072] Figure 3 This is a logic diagram of the LLM-based MDT automated assisted diagnosis method of the present invention. DETAILED DESCRIPTION

[0073] The following is combined with Figures 1 to 3 The present invention is described in further detail:

[0074] Example 1

[0075] See also Figure 1 and Figure 3 As shown, an LLM-based MDT automated assisted diagnosis method includes the following steps:

[0076] S1: Based on medical big data, historical patients' full life cycle case data is obtained, and after pre-processing, it is integrated into the patient case integration encryption system. The historical patients' full life cycle case data obtained include medical history, symptoms, physical examinations, laboratory tests, diagnosis and treatment records of patients with various diseases, and are pre-processed by removing duplicate data, smoothing, and filling missing values.

[0077] S2: Obtain patient case data, establish a patient personal case encryption system, classify patient cases, and use encryption technology to store and control access to various patient data to ensure privacy and security.

[0078] S3: Retrieve the patient's case from the patient's personal case encryption system to generate at least two prompt words, the two prompt words include a first prompt phrase group and a second prompt phrase group, enter the first prompt phrase group into the non-anthropomorphic LLM group, and enter the second prompt phrase group into the anthropomorphic LLM group. The first prompt phrase group is a general prompt phrase group, which needs to cover basic information such as medical history, chief complaint, and physical examination results. The second prompt phrase group is a precision prompt phrase group, which is for detailed information on specific diseases or complex cases. In the process of generating prompt phrases, natural language processing technology is used to segment, denoise and structure the medical record text. The non-anthropomorphic LLM group includes Claude-3.5, and the anthropomorphic LLM group includes GPT-4o.

[0079] S4: Use the LLM model to perform anthropomorphic and non-anthropomorphic multidisciplinary auxiliary diagnosis and treatment to obtain optimized auxiliary diagnosis and treatment plans. Parameter setting: Set the parameters of the LLM model, such as temperature parameters, to control the diversity and accuracy of the generated content, the maximum output length, etc. Input the prompt words into the LLM model through the API interface, call the model to generate auxiliary diagnosis and treatment plans, and optimize the generated auxiliary diagnosis and treatment plans.

[0080] S5: Compare and analyze the LLM model with the physician's diagnosis and treatment report to evaluate the diagnostic accuracy and rationality of the treatment plan, and select the optimal LLM-assisted diagnosis and treatment model. This evaluation will be conducted through simulation comparisons and statistical analysis to verify the performance of the LLM model.

[0081] S6: Obtain the current patient's case data and input it into the optimal LLM-assisted diagnosis and treatment model to generate an auxiliary diagnosis and treatment plan. The system also predicts the patient's possible diseases based on their symptoms and test results. Association rule mining: Discovers frequently occurring association rules in the data. For example, it discovers the association between certain symptoms and specific diseases. Cluster analysis: Divides patient data into clusters to identify similar patient groups. For example, patients can be divided into high-risk and low-risk groups based on their disease severity and treatment response.

[0082] Example 2

[0083] Based on Example 1, the specific process of preprocessing the historical patient's life cycle case data in step S1 is as follows:

[0084] S11: Detect duplicate values ​​in the full life cycle case data using a preset similarity function, where the similarity function formula is: , where A and B are the feature sets of two records;

[0085] S13: establishing time series data of the full life cycle case data based on the time sequence, and performing smoothing processing on the time series data;

[0086] S12: Identify missing values ​​for the full life cycle case data and fill in the missing values. The filling formula is: ,in x fill is the fill value, x i is a valid value, n is the number of valid values;

[0087] S13: Normalize the data after filling missing values. The formula is: x norm =( xx min ) / ( x max -x min ),in, x norm is the normalized value, x min is the minimum value, x max is the maximum value, x The data after filling missing values.

[0088] In this embodiment, step S1 also includes the patient case integration encryption system encrypting the pre-processed full life cycle case data. The specific process is as follows:

[0089] S14: randomly generating an original key of a specified number of bits, performing key expansion on the original key, and dividing the original key into a 4×8 matrix;

[0090] S15: performing byte replacement, row shift, column mixing, and round constant addition operations on the matrix to generate a specified number of subkeys, each subkey being used for one encryption round;

[0091] S16: inputting plaintext data of a specified number of bits and the first subkey, dividing the plaintext data into a 4×4 matrix, and performing a specified round of encryption operation based on the 4×4 matrix.

[0092] The specific process in step S15 is as follows:

[0093] S151: replacing each byte in the matrix according to a specified nonlinear transformation rule;

[0094] S152: Leave the first row of the matrix unchanged, shift the second row left by 1 byte, shift the third row left by 2 bytes, and shift the fourth row left by 3 bytes;

[0095] S153: Perform a linear transformation on the columns in the matrix, transforming the 4 bytes of each column through a specified matrix multiplication. The specific formula is as follows:

[0096] ;

[0097] in, a 0- a 3 are the elements in the original vector before the operation, b 0- b 3 are the elements of the output column vector obtained after the operation.

[0098] Example 3

[0099] On the basis of Example 1 or Example 2, obtaining the target patient case data in step S2 further includes decrypting the target patient case data. The specific process is as follows:

[0100] S21: Perform reverse round constant addition on the target patient case data using the reversed extended key;

[0101] S22: Use the reversed extended key to perform a reverse column mixing operation on the target patient case data. The specific formula is as follows:

[0102] ;

[0103] in, a 0- a 3 are the elements in the original vector before the operation, b 0- b 3 are the elements of the output column vector obtained after the operation.

[0104] S23: Perform inverse column shift and inverse byte transformation operations, and finally perform an XOR operation with the first subkey through the inverse operation of the initial round to restore the plaintext. The specific formula is as follows:

[0105] ;

[0106] Among them, 0 x 1 B is an irreducible polynomial in the finite field GF(28) x 8 + x 4 + x 3 + x +1 means, a For elements, a ≪1: Yes a Perform a bitwise shift left by 1 bit. It is an exclusive OR operation.

[0107] The specific process of obtaining patient case data and classifying patient cases in step S2 is as follows:

[0108] S24: establishing a real-time data interface, and obtaining patient case data based on the real-time data interface;

[0109] S25: Obtain the patient's disease type and current stage based on the patient's case data;

[0110] S26: Monitor changes in data sources in real time through API polling or message queues, push newly generated data to the patient's personal medical record encryption system through the API interface, and update the corresponding fields or records;

[0111] S27: Encrypt the newly generated data through steps S14-S16.

[0112] In this embodiment, the specific process of retrieving the patient's medical records from the patient's personal medical record encryption system and generating at least two prompt words in step S3 is as follows:

[0113] S31: Retrieving patient case data from the patient's personal case encryption system, including the patient's medical history, symptoms, and various examination data;

[0114] S32: Extract features from the patient case data and generate two prompt words based on the extracted features, including a first prompt word group and a second prompt word group. The first prompt word group covers the patient's medical history, chief complaint and examination information, and the second prompt word group is detailed information for a specified disease or complex case, including symptoms, examination information and existing diagnostic data.

[0115] The specific process of obtaining an optimized auxiliary diagnosis and treatment plan by performing anthropomorphic and non-anthropomorphic multidisciplinary auxiliary diagnosis and treatment through the LLM model in step S4 is as follows:

[0116] S41: The LLM model simulates the roles of doctors in different departments based on the input first prompt phrase group and second prompt phrase group, and generates a comprehensive auxiliary diagnosis and treatment plan from multiple angles;

[0117] S42: The LLM model analyzes the information of the prompt words in the first prompt word group and the second prompt word group, retrieves the built-in medical knowledge base, optimizes the auxiliary diagnosis and treatment plan, and generates an optimized auxiliary diagnosis and treatment plan.

[0118] Steps S41 and S42 are both provided with conflict detection and resolution methods, and the specific process is as follows:

[0119] Detect data conflicts based on version control or timestamps;

[0120] The detected data conflicts are resolved according to preset rules, wherein the preset rules include last write priority or authoritative data source priority.

[0121] The data from Groups A, B, C, and D were compared and cross-referenced to evaluate the differences and advantages of different models compared to real-world physician treatment plans. This included comparing non-anthropomorphic and anthropomorphic models, the performance of optimal and average models, and the practical application value of the LLM model and physician plans. Ultimately, a clinically meaningful LLM-based automated MDT consultation framework was developed.

[0122] The automated consultation framework's performance evaluation system will be further refined based on clinical feedback. Feedback from physicians and patients will be collected through questionnaires, interviews, and user testing to understand the system's performance and potential issues in real-world applications. Based on this feedback, the system's strengths and weaknesses will be evaluated, and specific areas for improvement, such as diagnostic accuracy, user experience, and data processing speed, will be identified. Furthermore, new evaluation metrics and methods will be designed to more accurately measure system performance. Following this, the automated consultation framework will be upgraded and optimized based on the evaluation results. This may include improving algorithms, adding new features, and optimizing the user interface. After implementing the optimization plan, feedback will continue to be collected and evaluated, and the system will be continuously adjusted and optimized based on new feedback, forming a cycle of continuous improvement.

[0123] S5 also includes an optimization method for the LLM model, including the following process:

[0124] S51: establishing a target optimization function for evaluating the performance of the model combination, and defining variables of the target optimization function;

[0125] S52: Based on the variables of the target optimization function, a population of size N is selected, where N represents the number of individuals in the initial population;

[0126] S53: Randomly generate M model combinations, each combination is a binary vector of length n;

[0127] S54: Create a fitness function and calculate the fitness of each model combination. The specific formula of the fitness function is as follows:

[0128] ;

[0129] in, w 1, w 2, w 3 is the weight coefficient, which is adjusted according to actual needs. Diagnostic accuracy is an indicator that measures the proportion of correct diagnostic results in the medical diagnosis process. Treatment plan rationality is an evaluation of whether the treatment plan formulated for the patient's condition is scientific, effective, and suitable for the patient's actual situation. Information integration efficiency represents the efficiency of collecting, organizing, integrating, and utilizing various types of information in the medical process.

[0130] S55: Select individuals based on fitness ratio. The fitness ratio formula for each individual is as follows:

[0131] ;

[0132] in, f (x i ) is the i Individual x i The fitness function value of N is the total number of individuals in the population, f (x j ) is the j Individual x j The fitness function value of

[0133] S56: Randomly select individuals to enter the next generation according to the fitness ratio, randomly select two parent individuals x1 and x2, and exchange some of their genes at random positions to generate two offspring individuals;

[0134] ;

[0135] ;

[0136] in, c is a randomly selected crossover position used to determine the split point for exchanging genes between parent individuals. n is the number of locations;

[0137] S57: Randomly select an individual x , randomly select a position i ,Will x i The value changes from 0 to 1 or from 1 to 0;

[0138] S58: Generate a new population, and iteratively execute steps S54-S57 until the maximum number of iterations is reached or a preset stopping condition is met.

[0139] In another implementation of this embodiment, the evaluation method in step S5 is to perform simulation test comparison and statistical analysis to verify the performance of the LLM model. The specific process is as follows:

[0140] Determine the experimental objectives: Identify the LLM model performance metrics that need to be verified, such as diagnostic accuracy, treatment plan rationality, and information integration efficiency. Select the experimental dataset: Obtain a representative patient case dataset from the hospital's clinical data center, ensuring that the dataset covers a variety of disease types and complexities. Divide the experimental and control groups: Randomly divide the dataset into experimental and control groups. Use the LLM model in the experimental group. Set experimental parameters: Determine the LLM model parameter settings, such as temperature parameters and maximum output length, to ensure consistency in experimental conditions.

[0141] Data preprocessing: Preprocess the data of the experimental group and the control group, including data cleaning and standardization, to ensure data quality.

[0142] Generate treatment recommendations:

[0143] For the experimental group, the LLM model was used to generate diagnosis and treatment recommendations based on the prompt words.

[0144] For the control group, traditional methods (such as diagnostic reports from senior doctors) were used to generate diagnosis and treatment recommendations.

[0145] Record experimental results: Record in detail the diagnosis and treatment recommendations generated for each group, including diagnosis results, treatment plans, drug recommendations, etc.

[0146] Diagnostic accuracy: Calculate the proportion of the LLM model's diagnostic results that are consistent with the actual diagnostic results.

[0147] Rationality of treatment plan: Evaluate whether the treatment plan generated by the LLM model is consistent with clinical guidelines and expert opinions, using quantitative indicators (such as accuracy, recall rate, and F1 value) for evaluation.

[0148] Information integration efficiency: Evaluate the speed and accuracy of the LLM model in integrating multi-source data, and calculate indicators such as data processing time and data consistency.

[0149] User satisfaction: Collect feedback from doctors and patients through questionnaires, interviews, etc. to evaluate the performance of the system in actual use.

[0150] Perform statistical analysis:

[0151] Descriptive statistical analysis: Calculate the mean, standard deviation, median, etc. of various performance indicators of the experimental group and the control group to understand the basic distribution of the data.

[0152] Hypothesis testing: Chi-square test: used to compare whether the differences between the experimental and control groups in indicators such as diagnostic accuracy and rationality of treatment plans are statistically significant. Chi-square test: used to compare the distribution differences of categorical variables (such as whether the diagnosis result is correct or not).

[0153] Correlation analysis: Analyze the correlation between different performance indicators, such as the relationship between diagnostic accuracy and the rationality of treatment plans.

[0154] Regression analysis: Establish a regression model to evaluate the relationship between the performance indicators of the LLM model and influencing factors (such as data quality and model parameters).

[0155] Interpretation and reporting of results

[0156] Interpretation of results: Based on the statistical analysis results, explain the performance of the LLM model on various performance indicators. For example, if the diagnostic accuracy of the LLM model is significantly higher than that of the control group, it indicates that it has an advantage in diagnosis.

[0157] Model optimization: Based on experimental results, adjust the LLM model's parameters or improve the model structure to enhance its performance. System improvement: Based on issues discovered during the experiment, optimize the system's data processing flow, user interface, and other aspects to improve user experience and system efficiency. Repeat the experiment: After optimization, repeat the above experimental process to verify the effectiveness of the improvement measures and ensure the continuous improvement of the LLM model's performance.

[0158] In summary, the present invention provides an MDT automated assisted consultation method based on LLM. By establishing a patient personal case encryption system, obtaining case data to generate prompt words, performing anthropomorphic and non-anthropomorphic multidisciplinary assisted diagnosis and treatment through the LLM model, an optimized auxiliary diagnosis and treatment plan is obtained, and the LLM model is compared and analyzed with the doctor's diagnosis and treatment report to evaluate the diagnostic accuracy and rationality of the treatment plan, and the best LLM assisted diagnosis and treatment model is screened out; the current patient case data is obtained and input into the best LLM assisted diagnosis and treatment model to generate an auxiliary diagnosis and treatment plan. The diagnostic accuracy and rationality of the treatment plan of the joint diagnosis and treatment of multiple departments are improved, and by developing data integration technology, the data barriers between departments are broken down, and the standardized processing of multi-source data is realized. At the same time, the transparency and credibility of LLM in complex medical scenarios are enhanced, a collaborative mechanism between LLM and human doctors is established, and the best collaborative model is explored.

Claims

1. An LLM-based MDT automated assisted diagnosis method, characterized in that: The following steps are involved: S1: Based on medical big data, obtain historical patient case data throughout their life cycle, and integrate it into the patient case integration encryption system after preprocessing; S2: Obtain patient case data, establish a patient personal case encryption system, and classify patient cases; S3: Retrieving the patient's medical records from the patient's personal medical record encryption system to generate at least two prompt words, the two prompt words including a first prompt word group and a second prompt word group, entering the first prompt word group into a non-anthropomorphic LLM group, and entering the second prompt word group into an anthropomorphic LLM group; S4: An optimized auxiliary diagnosis and treatment plan is obtained through anthropomorphic and non-anthropomorphic multidisciplinary auxiliary diagnosis and treatment using the LLM model; S5: Compare and analyze the LLM model with the doctor's diagnosis and treatment report, evaluate the diagnostic accuracy and rationality of the treatment plan, and screen the best LLM-assisted diagnosis and treatment model; S6: Obtain the current patient case data and input it into the optimal LLM auxiliary diagnosis and treatment model to generate an auxiliary diagnosis and treatment plan; Step S1 also includes the patient case integration encryption system encrypting the pre-processed full life cycle case data. The specific process is as follows: S14: randomly generating an original key of a specified number of bits, performing key expansion on the original key, and dividing the original key into a 4×8 matrix; S15: performing byte replacement, row shift, column mixing, and round constant addition operations on the matrix to generate a specified number of subkeys, each subkey being used for one encryption round; S16: Inputting plaintext data of a specified number of bits and the first subkey, dividing the plaintext data into a 4×4 matrix, and performing a specified round of encryption operation based on the 4×4 matrix; The specific process in step S15 is as follows: S151: replacing each byte in the matrix according to a specified nonlinear transformation rule; S152: Leave the first row of the matrix unchanged, shift the second row left by 1 byte, shift the third row left by 2 bytes, and shift the fourth row left by 3 bytes; S153: Perform a linear transformation on the columns in the matrix, transforming the 4 bytes of each column through a specified matrix multiplication. The specific formula is as follows: ； in, a 0- a 3 are the elements in the original vector before the operation, b 0- b 3 are the elements of the output column vector obtained after the operation; Acquiring the target patient case data in step S2 also includes decrypting the target patient case data. The specific process is as follows: S21: Perform reverse round constant addition on the target patient case data using the reversed extended key; S22: Use the reversed extended key to perform a reverse column mixing operation on the target patient case data. The specific formula is as follows: ； in, a 0- a 3 are the elements in the original vector before the operation, b 0- b 3 are the elements of the output column vector obtained after the operation; S23: Perform inverse column shift and inverse byte transformation operations, and finally perform an XOR operation with the first subkey through the inverse operation of the initial round to restore the plaintext. The specific formula is as follows: ； Among them, 0 x 1 B is an irreducible polynomial in the finite field GF(28) x 8 + x 4 + x 3 + x +1 means, a For elements, a ≪1: Yes a Perform a bitwise shift left by 1 bit. is the exclusive OR operation; The specific process of obtaining an optimized auxiliary diagnosis and treatment plan by performing anthropomorphic and non-anthropomorphic multidisciplinary auxiliary diagnosis and treatment through the LLM model in step S4 is as follows: S41: The LLM model simulates the roles of doctors in different departments based on the input first prompt phrase group and second prompt phrase group, and generates a comprehensive auxiliary diagnosis and treatment plan from multiple angles; S42: The LLM model analyzes the information of the prompt words in the first prompt word group and the second prompt word group, retrieves the built-in medical knowledge base, optimizes the auxiliary diagnosis and treatment plan, and generates an optimized auxiliary diagnosis and treatment plan.

2. The MDT automated assisted diagnosis method based on LLM according to claim 1, characterized in that: The specific process of preprocessing the full life cycle case data of historical patients in step S1 is as follows: S11: Detect duplicate values ​​in the full life cycle case data using a preset similarity function, where the similarity function formula is: , where A and B are the feature sets of two records; S13: establishing time series data of the full life cycle case data based on the time sequence, and performing smoothing processing on the time series data; S12: Identify missing values ​​for the full life cycle case data and fill in the missing values. The filling formula is: ,in x fill is the fill value, x i is a valid value, n is the number of valid values; S13: Normalize the data after filling missing values. The formula is: x norm =( xx min ) / ( x max -x min ),in, x norm is the normalized value, x min is the minimum value, x max is the maximum value, x The data after filling missing values.

3. The MDT automated assisted diagnosis method based on LLM according to claim 1, characterized in that: The specific process of obtaining patient case data and classifying patient cases in step S2 is as follows: S24: establishing a real-time data interface, and obtaining patient case data based on the real-time data interface; S25: Obtain the patient's disease type and current stage based on the patient's case data; S26: Monitor changes in data sources in real time through API polling or message queues, push newly generated data to the patient's personal medical record encryption system through the API interface, and update the corresponding fields or records; S27: Encrypt the newly generated data through steps S14-S16.

4. The LLM-based MDT automated assisted diagnosis method according to claim 1, characterized in that: The specific process of retrieving the patient's medical records from the patient's personal medical record encryption system and generating at least two prompt words in step S3 is as follows: S31: Retrieving patient case data from the patient's personal case encryption system, including the patient's medical history, symptoms, and various examination data; S32: Extract features from the patient case data and generate two prompt words based on the extracted features, including a first prompt word group and a second prompt word group. The first prompt word group covers the patient's medical history, chief complaint and examination information, and the second prompt word group is detailed information for a specified disease or complex case, including symptoms, examination information and existing diagnostic data.

5. The LLM-based MDT automated assisted diagnosis method according to claim 1, characterized in that: Steps S41 and S42 are both provided with conflict detection and resolution methods, and the specific process is as follows: Detect data conflicts based on version control or timestamps; The detected data conflicts are resolved according to preset rules, wherein the preset rules include last write priority or authoritative data source priority.

6. The LLM-based MDT automated assisted diagnosis method according to claim 1, characterized in that: S5 also includes the optimization method for the LLM model. The following processes are included: S51: establishing a target optimization function for evaluating the performance of the model combination, and defining variables of the target optimization function; S52: Based on the variables of the target optimization function, a population of size N is selected, where N represents the number of individuals in the initial population; S53: Randomly generate M model combinations, each combination is a binary vector of length n; S54: Create a fitness function and calculate the fitness of each model combination. The specific formula of the fitness function is as follows: ； in, w 1, w 2, w 3 is the weight coefficient, which is adjusted according to actual needs. Diagnostic accuracy is an indicator that measures the proportion of correct diagnostic results in the medical diagnosis process. Treatment plan rationality is an evaluation of whether the treatment plan formulated for the patient's condition is scientific, effective, and suitable for the patient's actual situation. Information integration efficiency represents the efficiency of collecting, organizing, integrating, and utilizing various types of information in the medical process. S55: Select individuals based on fitness ratio. The fitness ratio formula for each individual is as follows: ； in, f (x i ) is the i Individual x i The fitness function value of N is the total number of individuals in the population, f (x j ) is the j Individual x j The fitness function value of S56: Randomly select individuals to enter the next generation according to the fitness ratio, randomly select two parent individuals x1 and x2, and exchange some of their genes at random positions to generate two offspring individuals; ； ； in, c is a randomly selected crossover position used to determine the split point for exchanging genes between parent individuals. n is the number of locations; S57: Randomly select an individual x , randomly select a position i ,Will x i The value changes from 0 to 1 or from 1 to 0; S58: Generate a new population, and iteratively execute steps S54-S57 until the maximum number of iterations is reached or a preset stopping condition is met.

Images