A medical interrogation decision optimization system and method based on negative evidence driving

Abstract

The application discloses a medical inquiry decision optimization system and method based on negative evidence driving, and relates to the technical field of intelligent medical treatment.The technical solution points are as follows: a data processing module is used to construct a positive and negative symptom set;a medical knowledge modeling module is used to construct a knowledge graph containing a symptom cluster;a negative evidence driven candidate space pruning module is used to perform multi-level exclusion and probability attenuation based on negative symptoms and the symptom cluster;a conflict fusion and hierarchical candidate management module is used to calculate positive and negative support and divide a main candidate and a differential diagnosis set;an inquiry decision optimization module is used to select an optimal question by using reinforcement learning; a safety constraint control module is used to monitor high-risk symptoms and enforce intervention; and a diagnosis output module is used to output results.The application uses negative evidence to quickly narrow down the candidate disease space, avoids redundant punishment through a symptom cluster, combines conflict processing and a safety mechanism, significantly improves diagnosis accuracy, inquiry efficiency and system robustness, and effectively controls medical risks.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical technology, and more specifically, to a medical consultation decision optimization system and method based on negative evidence. Background Technology

[0002] With the rapid development of artificial intelligence technology, intelligent medical consultation systems are being increasingly widely used in fields such as assisted diagnosis, triage and guidance, and health management. These systems simulate the doctor's consultation process, engaging in multiple rounds of interaction with patients to collect symptom information, thereby inferring possible diseases and providing corresponding medical advice. Given the uneven distribution of medical resources and the lack of experience among primary care physicians, efficient and accurate intelligent consultation systems are of great significance in improving the accessibility of medical services and reducing the workload of doctors.

[0003] Currently, mainstream intelligent medical consultation decision-making methods mainly include symptom-matching scoring and ranking methods: These methods calculate a matching score for each candidate disease based on the positive symptoms provided by the patient (i.e., symptoms the patient clearly exhibits). The system then ranks the diseases according to their scores and prioritizes asking about symptoms related to high-scoring diseases. While simple to implement, these methods have significant shortcomings in practical applications: first, they do not fully utilize negative symptoms denied by the patient, which often have strong disease exclusion capabilities; second, when faced with multiple symptoms related to the same disease, they lack modeling of the correlation between symptoms, easily leading to duplicate scoring or over-penalization. Knowledge graph-based disease reasoning methods also exist: these methods construct a knowledge graph containing diseases, symptoms, and their relationships, and perform disease reasoning through graph traversal or graph neural networks. These methods can utilize structured information from knowledge, but current technologies primarily focus on forward reasoning, insufficiently utilizing negative symptoms, and lack effective handling of symptom redundancy within the same symptom cluster.

[0004] In summary, existing intelligent medical consultation and decision-making technologies still have the following problems:

[0005] (1) Insufficient use of negative symptom information: Existing methods generally focus on positive matching of positive symptoms, and do not systematically utilize negative symptoms as strong evidence for actively trimming the candidate disease space, resulting in an excessively large disease search space and low consultation efficiency.

[0006] (2) Symptom redundancy leads to probability estimation bias: The same disease is often associated with multiple symptoms with high clinical relevance. When patients deny these symptoms one after another, existing methods often treat them as multiple independent pieces of negative evidence, repeatedly attenuating the probability of the disease, which leads to the disease being over-punished or even wrongly excluded.

[0007] (3) In real clinical scenarios, patients’ complaints are often incomplete, ambiguous or even incorrect. Existing methods have low tolerance for input noise and missing information, resulting in large fluctuations in diagnostic results and poor stability of candidate disease sets.

[0008] Therefore, the present invention aims to provide a medical consultation decision optimization system and method based on negative evidence to solve the above problems. Summary of the Invention

[0009] The purpose of this invention is to provide a medical consultation decision optimization system and method based on negative evidence. This invention utilizes negative evidence to quickly narrow down the candidate disease space, avoids redundant penalties through symptom clusters, and combines conflict handling and security mechanisms to significantly improve diagnostic accuracy, consultation efficiency, and system robustness, thereby achieving effective management of medical risks.

[0010] The above-mentioned technical objective of the present invention is achieved through the following technical solution: a medical consultation decision optimization system and method based on negative evidence, including a data processing module, a medical knowledge modeling module, a negative evidence-driven candidate space trimming module, a conflict fusion and hierarchical candidate management module, a consultation decision optimization module, a security constraint control module, and a diagnostic output module;

[0011] The data processing module is used to receive and parse the consultation information input by the user, extract and standardize the symptoms, and construct a set of positive symptoms and a set of negative symptoms.

[0012] The medical knowledge modeling module is used to construct and store a medical knowledge graph, which includes a disease set, a symptom set, disease-symptom association probabilities, and several predefined symptom clusters, each of which contains one or more clinically relevant symptoms.

[0013] The candidate space pruning module driven by negative evidence is communicatively connected to the data processing module and the medical knowledge modeling module, respectively. It is used to receive the negative symptom set and the symptom cluster, perform multi-level exclusion on the candidate disease set, calculate the comprehensive exclusion score based on the symptom cluster to update the disease probability, and output the pruned candidate disease set and its normalized probability distribution.

[0014] The conflict fusion and hierarchical candidate management module is communicatively connected to the negative evidence-driven candidate space trimming module. It is used to receive the trimmed disease probability distribution, and combine the positive symptom set and the negative symptom set to calculate the positive support and negative rejection of each candidate disease. Based on the calculation results, the candidate diseases are divided into the main candidate set and the differential diagnosis set.

[0015] The consultation decision optimization module is communicatively connected to the conflict fusion and hierarchical candidate management module, and is used to receive the current state, which includes at least the main candidate set and the differential diagnosis set, and select the optimal consultation question through a reinforcement learning model and output it to the user terminal.

[0016] The safety constraint control module is communicatively connected to the consultation decision optimization module, and is used to monitor unconfirmed high-risk symptoms in real time, and send a mandatory inquiry command to the consultation decision optimization module or an early termination command to the diagnosis output module when a risk is detected.

[0017] The diagnostic output module is communicatively connected to the conflict fusion and hierarchical candidate management module and the safety constraint control module, respectively, and is used to output the diagnostic results, differential diagnosis list, risk level prompts and medical suggestions in a formatted manner when the preset termination conditions are met.

[0018] The present invention is further configured such that the candidate space pruning module driven by negation evidence is specifically used for:

[0019] Strong exclusion is performed: when a disease exists, and the probability of the occurrence of a certain symptom is greater than a preset high threshold, and the symptom belongs to the set of negative symptoms, the disease is directly excluded from the candidate disease set.

[0020] Perform soft exclusion: Based on the comprehensive rejection score of symptom clusters, probability decay is performed. For each disease and each symptom cluster, the comprehensive rejection score of the cluster is calculated as follows: take the maximum value of the conditional probabilities of all symptoms denied by the patient for the disease within the symptom cluster; then, multiply the current probability of the disease by the rejection factor corresponding to each symptom cluster, where each rejection factor is equal to 1 minus the comprehensive rejection score of the cluster, thereby updating the probability of the disease.

[0021] Perform probability normalization: Divide the updated probabilities of all candidate diseases by the sum of the probabilities of all candidate diseases to obtain the normalized probability distribution.

[0022] The present invention is further configured such that the symptom cluster generation logic in the medical knowledge modeling module includes: constructing a symptom co-occurrence matrix based on historical medical record data, calculating the correlation strength between symptoms, and using a clustering algorithm to classify symptoms with a correlation strength greater than a preset threshold into the same symptom cluster; wherein, symptoms within the same symptom cluster are defined as having clinical substitutability, and only their maximum rejection contribution value is taken when calculating the comprehensive rejection score.

[0023] The present invention is further configured such that the conflict fusion and hierarchical candidate management module is specifically used for:

[0024] Calculate the positive support for each disease: for each symptom in the set of positive symptoms, multiply it by its conditional probability under that disease, and sum the results in a weighted manner;

[0025] Calculate the negative exclusion degree for each disease: for each symptom in the set of negative symptoms, multiply it by its conditional probability under that disease, and sum the results in a weighted manner;

[0026] The overall score for each disease is defined as the product of positive support minus negative rejection and a preset coefficient;

[0027] Diseases with an absolute value of a comprehensive score less than a preset threshold are classified into the differential diagnosis set to indicate diseases with conflicting evidence.

[0028] For diseases in the differential diagnosis set, a delayed decision mechanism is enabled to preserve their probability and prioritize the generation of discriminative consultation questions for them in subsequent rounds.

[0029] The present invention is further configured such that: the reward function of the reinforcement learning model in the consultation decision optimization module is a multi-objective weighted function, specifically including: a diagnostic accuracy reward, a consultation efficiency reward, and a risk penalty; wherein, the diagnostic accuracy reward gives a positive high score when the final diagnosis is correct, the consultation efficiency reward is inversely proportional to the number of consultation rounds to encourage a reduction in the number of rounds, and the risk penalty gives a large negative score when the model misses high-risk symptoms.

[0030] The present invention is further configured such that the safety constraint control module is specifically used for:

[0031] Real-time detection of any unconfirmed high-risk symptoms, i.e., symptoms that are neither in the positive symptom set nor the negative symptom set;

[0032] When an unconfirmed high-risk symptom is detected, a mandatory inquiry instruction is sent to the consultation decision optimization module to force priority inquiry into the high-risk symptom.

[0033] When the preset safety termination conditions are met, an early termination command is sent to the diagnostic output module to terminate the consultation and output medical advice.

[0034] This invention also provides a medical consultation decision optimization method based on negative evidence, comprising the following steps:

[0035] S1. Obtain the patient's symptom information through the data processing module, and construct a set of positive symptoms and a set of negative symptoms;

[0036] S2. Through the candidate space pruning module driven by the negative evidence, based on the negative symptom set and the symptom cluster, the candidate disease set is subjected to multi-level exclusion and probability decay to obtain the pruned candidate disease set and its normalized probability distribution.

[0037] S3. Through the conflict fusion and hierarchical candidate management module, based on the positive symptom set, negative symptom set and the cropped disease probability, calculate the positive support and negative rejection of each candidate disease, and divide the candidate diseases into the main candidate set and the differential diagnosis set.

[0038] S4. The consultation decision optimization module selects the optimal consultation question based on the current state and outputs it to the user terminal.

[0039] S5. Obtain user feedback on the consultation questions through the data processing module, and update the positive symptom set and the negative symptom set;

[0040] S6. Repeat steps S2 to S5 until the preset termination condition is met, and output the diagnostic result through the diagnostic output module.

[0041] The present invention is further configured such that: step S2 involves multi-level exclusion and probability attenuation of the candidate disease set, specifically including:

[0042] S21. Strong exclusion: For any disease, if any symptom exists, and the probability of the symptom occurring under the disease is greater than a preset high threshold and the symptom belongs to the set of negative symptoms, then the disease is directly excluded.

[0043] S22, Soft exclusion: For each disease and each symptom cluster, first calculate the maximum value of the conditional probability of all denied symptoms in the cluster for the disease, as the comprehensive exclusion score of the cluster; then multiply the current probability of the disease by the exclusion factor corresponding to each symptom cluster, where each exclusion factor is equal to 1 minus the comprehensive exclusion score of the cluster, to obtain the updated disease probability.

[0044] S23. Probability Normalization: Divide the updated probabilities of all candidate diseases by the sum of the probabilities of all candidate diseases.

[0045] The present invention is further configured such that: in step S3, the candidate diseases are divided into a primary candidate set and a differential diagnosis set, specifically including:

[0046] S31. Calculate the positive support and negative rejection for each disease;

[0047] S32. Diseases whose absolute value of the difference between positive support and negative rejection is less than a preset threshold are classified into the differential diagnosis set;

[0048] S33. For diseases in the differential diagnosis set, a delayed decision mechanism is enabled, and questions that can distinguish different diseases in the set are generated first in subsequent rounds.

[0049] The present invention also provides a medical consultation decision optimization device based on negative evidence, comprising at least one processor; and a memory communicatively connected to at least one of the processors; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to implement a medical consultation decision optimization method based on negative evidence.

[0050] In summary, the present invention has the following beneficial effects:

[0051] 1. This invention utilizes the patient's clear negative symptoms as high-weight evidence, combined with the redundancy elimination logic of symptom clusters, to quickly and accurately exclude a large number of irrelevant diseases, thereby significantly narrowing the candidate disease space. By performing differential diagnosis within this constrained space, interference from irrelevant diseases can be effectively avoided, greatly improving the accuracy of diagnosis. Moreover, verified by real clinical data, the accuracy and coverage of this invention are significantly better than traditional positive scoring ranking methods and reinforcement learning consultation systems without negative pruning.

[0052] 2. This invention, through a multi-level exclusion mechanism driven by negative evidence, can quickly eliminate a large number of impossible diseases in the early stages of diagnosis, thereby concentrating subsequent diagnosis resources on the identification of a few core candidate diseases. This step directly reduces the number of rounds required in the entire diagnosis process. At the same time, by introducing an efficiency reward term into the reinforcement learning reward function, the model is further prompted to prioritize the most discriminative questions, thus optimizing diagnostic efficiency.

[0053] 3. This invention introduces a symptom cluster mechanism and a cluster-based comprehensive rejection score calculation method. Specifically, for multiple negative symptoms within the same symptom cluster, only the maximum conditional probability is taken as the rejection contribution of that cluster. This step fundamentally solves the defect of the independence assumption between negative evidence and effectively prevents excessive probability decay caused by repeated denial due to similar symptoms. Through noise resistance experiments, it is demonstrated that in a noisy environment with random mislabeling, missing symptoms, and ambiguous expression, the accuracy reduction, diagnostic result volatility, and candidate set volatility of this invention are significantly better than the control system, showing extremely strong robustness.

[0054] 4. In the conflict fusion and hierarchical candidate management module, this invention simultaneously calculates the positive support and negative rejection of candidate diseases and classifies diseases with smaller differences into a differential diagnosis set for special management. This allows diseases in this set to be not directly deleted, but instead a delayed decision-making mechanism to prioritize the generation of the most discriminative questions that can distinguish these diseases in subsequent consultation rounds. This operation effectively avoids the accidental deletion of real diseases when information is insufficient, and is especially suitable for complex cases with atypical symptoms or multiple overlapping diseases, greatly improving the clinical applicability of the system.

[0055] 5. This invention monitors the confirmation status of high-risk symptoms in real time in the safety constraint control module, enabling proactive detection of high-risk symptoms that have not been asked about or confirmed during the consultation process, and has the ability to force intervention. That is, when a risk is detected, a forced inquiry instruction can be sent to the consultation decision module to ensure that high-risk symptoms are prioritized for investigation; in extreme risk situations, the consultation can also be terminated in advance and medical advice can be output.

[0056] 6. The consultation decision-making process of this invention follows the clinical diagnostic logic of first eliminating the dangerous and impossible, and then identifying subtle differences. Through spatial trimming driven by negative evidence, the system can clearly show users the reasons for excluding certain diseases, which enhances the interpretability of the system and the doctor's trust. In the clinical adaptability verification with the participation of doctors, this invention has received high scores in multiple dimensions such as the rationality of consultation logic, professionalism, clinical consistency and risk control capabilities, proving that it is highly consistent with the actual clinical decision-making process. Attached Figure Description

[0057] Figure 1 This is a schematic diagram of the module structure of a medical consultation decision optimization system based on negative evidence in Embodiment 1 of the present invention;

[0058] Figure 2 This is a schematic diagram of the service architecture of a medical consultation decision optimization system based on negative evidence in Embodiment 1 of the present invention;

[0059] Figure 3 This is a schematic diagram of the environmental boundary of a medical consultation decision optimization system based on negative evidence in Embodiment 1 of the present invention;

[0060] Figure 4 This is a schematic diagram of the interaction sequence of a medical consultation decision optimization system based on negative evidence in Embodiment 1 of the present invention;

[0061] Figure 5 This is a schematic diagram of the workflow of a medical consultation decision optimization system based on negative evidence in Embodiment 1 of the present invention;

[0062] Figure 6 This is a schematic diagram of the framework structure of a medical consultation decision optimization device based on negative evidence in Embodiment 7 of the present invention;

[0063] Figure 7 This is a flowchart illustrating the steps of a medical consultation decision optimization method based on negative evidence in Embodiment 3 of the present invention. Detailed Implementation

[0064] The following is in conjunction with the appendix Figures 1-7 The present invention will be described in further detail below.

[0065] Example 1: A medical consultation decision optimization system based on negative evidence, including a data processing module, a medical knowledge modeling module, a negative evidence-driven candidate space trimming module, a conflict fusion and hierarchical candidate management module, a consultation decision optimization module, a security constraint control module, and a diagnosis output module.

[0066] In this embodiment, the data processing module includes a text parsing unit, a standardization coding unit, and a set construction unit. The text parsing unit receives user input text and outputs the identified original symptoms and negative markers to the standardization coding unit. The standardization coding unit maps the original symptoms to standard symptom codes and outputs them to the set construction unit. The set construction unit classifies the symptoms into positive symptom sets and negative symptom sets according to the negative markers.

[0067] The data output by this module to the negative evidence-driven candidate space trimming module and the conflict fusion and hierarchical candidate management module is in structured object format. The structured object contains a list of positive symptoms and a list of negative symptoms. Each symptom entry contains a standard code, name and confidence level fields. Data transmission is carried out through inter-module function calls or message queues to ensure real-time performance.

[0068] In this embodiment, the medical knowledge modeling module includes a knowledge graph storage unit, a symptom cluster generation unit, and an attribute definition unit. The symptom cluster generation unit calculates symptom clusters offline based on historical medical record data and stores the results in the knowledge graph storage unit. The attribute definition unit defines necessary symptoms and high-risk symptoms and also stores them in the knowledge graph storage unit. During system operation, the knowledge graph storage unit responds to query requests from the negative evidence-driven candidate space pruning module and the safety constraint control module.

[0069] The data output by this module to the negative evidence-driven candidate space pruning module includes: a disease-symptom conditional probability matrix (stored in sparse matrix form) and a list of symptom cluster members (each cluster is represented by a set of symptom codes). The data output to the safety constraint control module includes a list of high-risk symptom codes.

[0070] In this embodiment, the candidate space pruning module driven by negative evidence includes an initialization unit, a multi-level exclusion unit, and a probability normalization unit. The initialization unit receives the full set of diseases from the medical knowledge modeling module, establishes an initial candidate disease probability distribution, and passes the results to the multi-level exclusion unit.

[0071] The multi-level exclusion unit receives the set of negative symptoms from the data processing module and symptom cluster data from the medical knowledge modeling module. Internally, this unit sequentially executes a strong exclusion subunit, a soft exclusion subunit, and a path-level pruning subunit. The strong exclusion subunit outputs a list of directly excluded diseases; the soft exclusion subunit calculates a comprehensive exclusion score based on symptom clusters and iteratively updates the disease probabilities; the path-level pruning subunit removes disease sub-paths based on the overall knowledge graph structure. The processing results of each subunit are passed sequentially, ultimately outputting the excluded candidate disease set and its unnormalized probability distribution to the probability normalization unit. The probability normalization unit receives the above data, performs normalization calculations, and outputs the pruned candidate disease set and its normalized probability distribution.

[0072] The data output by this module is sent to the conflict fusion and hierarchical candidate management module in the form of a list of key-value pairs. Each key-value pair contains a disease identifier and a normalized probability value. To ensure accuracy, the probability value is represented as a double-precision floating-point number.

[0073] In this embodiment, the conflict fusion and hierarchical candidate management module includes a scoring calculation unit, a hierarchical division unit, and a delayed decision management unit. The scoring calculation unit receives the positive symptom set and the negative symptom set from the data processing module, as well as the candidate disease probability distribution from the negative evidence-driven candidate space pruning module. This unit calculates the positive support and negative rejection of each candidate disease and transmits the calculation results (including the comprehensive score of each disease) to the hierarchical division unit.

[0074] The hierarchical division unit divides candidate diseases into a main candidate set, a differential diagnosis set, and an exclusion set based on the comprehensive score and preset thresholds. The main candidate set is arranged in descending order of comprehensive score. The differential diagnosis set consists of diseases whose absolute difference between positive support and negative rejection is less than the threshold. The delayed decision management unit records the "pending confirmation" status of diseases in the differential diagnosis set and provides priority processing marks to the consultation decision optimization module in subsequent rounds.

[0075] The state vector sent by this module to the consultation decision optimization module is an encapsulated object containing the current set of positive symptoms, the set of negative symptoms, the list of primary candidate diseases (including scores), and the list of differential diagnoses (including confirmation markers). The final data sent to the diagnosis output module includes a stratified list of all diseases and the comprehensive scores and positive / negative score details for each disease.

[0076] In this embodiment, the consultation decision optimization module includes a state encoding unit, an action space generation unit, a reinforcement learning strategy unit, and an action output unit.

[0077] The state encoding unit receives the state vector from the conflict fusion and hierarchical candidate management module, converts it into a tensor or feature vector that can be processed by the reinforcement learning model, and passes it to the reinforcement learning policy unit. The action space generation unit obtains the symptoms associated with the current main candidate set and differential diagnosis set from the medical knowledge modeling module, generates a set of candidate actions that have not yet been asked, and also passes it to the reinforcement learning policy unit. The reinforcement learning policy unit (pre-trained DQN or policy network) calculates the Q-value or probability distribution of each candidate action based on the current state encoding, and passes the index of the optimal action to the action output unit. The action output unit converts the action index into a natural language question and outputs it to the user terminal.

[0078] This module reserves an interrupt input port to receive mandatory query commands from the safety constraint control module. When a mandatory command is received, the calculation results of the reinforcement learning policy unit are overwritten, and the action output unit directly outputs the question corresponding to the mandatory query.

[0079] In this embodiment, the security constraint control module includes a risk detection unit, a policy execution unit, and an instruction sending unit.

[0080] The risk detection unit obtains the current positive and negative symptom sets in real time from the conflict fusion and hierarchical candidate management module, and the high-risk symptom set from the medical knowledge modeling module. This unit checks each high-risk symptom individually to see if it has been confirmed (in the positive symptom set) or denied (in the negative symptom set). If it is not in either set, it is marked as "unconfirmed risk," and the risk list is passed to the strategy execution unit. The strategy execution unit determines which strategy to execute based on the risk level and system configuration: if any unconfirmed high-risk symptom exists, a mandatory inquiry instruction is generated; if the number of unconfirmed high-risk symptoms exceeds a preset threshold or a specific critical symptom exists, an early termination instruction is generated. The instruction sending unit sends the mandatory inquiry instruction to the interrupt port of the consultation decision optimization module and the early termination instruction to the diagnostic output module. This module transmits instructions in enumeration type or short control code form, without carrying large amounts of data, to ensure minimal response latency.

[0081] In this embodiment, the diagnostic output module includes a termination condition detection unit, a result formatting unit, and an output interface unit. The termination condition detection unit continuously monitors the candidate disease status from the conflict fusion and hierarchical candidate management module and the early termination command from the safety constraint control module. When the main candidate set converges to a single disease, the highest confidence level exceeds a threshold, the maximum number of consultation rounds is reached, or an early termination command is received, the result output process is triggered. The result formatting unit obtains the final main candidate set, differential diagnosis set, and various scores from the conflict fusion and hierarchical candidate management module, and obtains risk records from the safety constraint control module, organizing these data into a structured diagnostic report. The output interface unit converts the formatted report into user-readable natural language text and sends it to the user terminal for display.

[0082] The diagnostic report in this module uses a hierarchical JSON structure or a plain text template, and includes at least: the most likely disease name and confidence level, a list of differential diagnoses, risk level (high / medium / low), and medical advice text; for high-risk cases, the text includes warning messages such as "seek medical attention immediately".

[0083] In this embodiment, the data flow sequence during the complete consultation cycle is as follows:

[0084] Initial trigger: User input triggers the consultation process; data flow: user terminal → data processing module.

[0085] Symptom analysis: After the data processing module completes the analysis, it sends the data in parallel to the negative evidence-driven candidate space pruning module (sending the negative symptom set) and the conflict fusion and hierarchical candidate management module (sending the positive symptom set and the negative symptom set).

[0086] Knowledge Query: The negative evidence-driven candidate space pruning module reads disease-symptom probability and symptom cluster data from the medical knowledge modeling module; the safety constraint control module reads a list of high-risk symptoms from the medical knowledge modeling module.

[0087] Spatial trimming: After the candidate spatial trimming module driven by negative evidence completes the elimination and probability decay, it sends the trimmed candidate disease probability distribution to the conflict fusion and hierarchical candidate management module.

[0088] Hierarchical Management: The conflict fusion and hierarchical candidate management module integrates symptom data and probability distribution, completes scoring and stratification, and sends the state vector to the consultation decision optimization module.

[0089] Safety Detection: The safety constraint control module obtains the latest symptom set from the conflict fusion and hierarchical candidate management module, performs risk detection, and sends control instructions to the consultation decision optimization module based on the detection results.

[0090] Decision generation: The consultation decision optimization module combines state vectors and safety instructions (if there are no mandatory instructions, the decision is made freely; if there are, the decision is overridden) to generate the optimal question and output it to the user.

[0091] User feedback: When a user answers a question, the data flows back from the user's end into the data processing module, starting the next cycle.

[0092] Termination and Output: When the candidate status in the Conflict Fusion and Hierarchical Candidate Management Module meets the termination conditions, or when the Security Constraint Control Module issues an early termination command, the Diagnostic Output Module reads the final data from the Conflict Fusion and Hierarchical Candidate Management Module and the Security Constraint Control Module, formats it, and outputs it to the user terminal.

[0093] Example 2: A medical consultation decision optimization method based on negative evidence, comprising the following steps:

[0094] S1. Obtain the patient's symptom information through the data processing module, and construct a set of positive symptoms and a set of negative symptoms.

[0095] The system receives user input through a human-computer interaction interface. The input can be in the form of text, voice (converted by voice recognition), or structured questionnaire options.

[0096] The data processing module parses the input content: using natural language processing techniques (such as a deep learning-based named entity recognition model) to identify all mentioned symptom names; simultaneously detecting negative modifiers for each symptom (such as "none," "no," "absent," "deny," etc.) to distinguish whether the symptom is present (positive) or explicitly denied (negative); mapping the identified symptoms to a standard medical terminology database (such as SNOMED CT or ICD coding) to obtain standardized symptom codes; and constructing two sets based on the negation detection results:

[0097] Positive symptom set S + This includes all standardized symptoms that the patient explicitly reports.

[0098] Negative symptom set S - This includes all standardized symptoms that the patient explicitly states do not have.

[0099] If a user does not mention certain symptoms in their initial input, they will not be included in any set and will be left for confirmation in a subsequent question.

[0100] S2. Using a candidate space pruning module driven by negative evidence, based on the set of negative symptoms and the symptom clusters, the candidate disease set is subjected to multi-level exclusion and probability decay to obtain the pruned candidate disease set and its normalized probability distribution.

[0101] This step is further subdivided into three sub-steps: S21, S22, and S23, which are executed in sequence.

[0102] S21, Strong Exclusion.

[0103] Input: The complete disease set D0 (pre-provided by the medical knowledge modeling module) and the negative symptom set S - .

[0104] For each disease in D0 , and S - Each symptom Query pre-stored disease-symptom conditional probabilities .

[0105] If a certain symptom exists its conditional probability Greater than the preset high threshold θ hard (θ) hard If the value is 0.85, then the symptom is determined to be a disease. The characteristic or necessary symptoms.

[0106] Because the patient explicitly denied the symptom ( ∈S - ),disease The probability of its existence was determined to be extremely low, therefore... It is directly moved to the Excluded Set and will not participate in subsequent calculations.

[0107] Output: The set of candidate diseases remaining after strong exclusion, D1=D0\Excluded Set, and the current probability of each disease (the initial probability can be set based on the prior incidence of the disease, or set equally).

[0108] S22, Soft exclusion (based on probability decay of symptom clusters).

[0109] Input: The set of candidate diseases after strong exclusion, D1; the set of negative symptoms, S. - 1. Predefined symptom clusters C (each cluster contains a set of symptoms that are highly clinically relevant).

[0110] Symptom grouping: S - The symptoms are grouped according to their respective symptom clusters. For example, if S - The sample includes "cough" and "fever". "Cough" belongs to the respiratory symptom cluster C1, and "fever" belongs to the fever-related symptom cluster C2. Therefore, the grouping result is that group C1 contains "cough" and group C2 contains "fever".

[0111] Calculate the overall rejection score for each symptom cluster: for each disease and each symptom cluster Ck Calculate the impact of all symptoms denied by the patient within this cluster. The maximum value of the conditional probability. Formulaically expressed as: That is, if multiple symptoms are denied within a cluster, only the one with the highest diagnostic significance (highest conditional probability) is taken as the exclusion contribution for that cluster.

[0112] Iteratively update disease probabilities: for each disease Multiply its current probability by the product of the rejection factors corresponding to all symptom clusters, where each rejection factor equals 1 minus the overall rejection score of that cluster; that is: Explanation of the principle: This chain multiplication operation achieves the effect that "each symptom cluster can only produce at most one decay", avoiding excessive probability decay caused by the repeated denial of multiple similar symptoms associated with the same disease (i.e., the symptom redundancy problem). At the same time, for clusters in which no negative symptoms fall, their Score=0 and rejection factor=1, which does not affect the probability.

[0113] Output: Updated probabilities of each disease after soft exclusion. And the current candidate disease set D2=D1 (diseases that were not excluded after strong exclusion will not be directly deleted in this step, only their probability will be reduced).

[0114] S23, Probability Normalization.

[0115] Input: Unnormalized probabilities of each disease after soft exclusion , and the current candidate disease set D2.

[0116] Calculate the sum of the probabilities of all candidate diseases Then, for each disease d, calculate the normalized probability: This ensures that the sum of the probabilities of all candidate diseases is 1, facilitating subsequent fusion calculations with other evidence (such as positive symptoms).

[0117] Output: The pruned set of candidate diseases D'=D2, and the normalized probability distribution P'(d) for each disease.

[0118] S3. Through the conflict fusion and hierarchical candidate management module, based on the positive symptom set, negative symptom set and the cropped disease probability, calculate the positive support and negative rejection of each candidate disease, and divide the candidate diseases into the main candidate set and the differential diagnosis set.

[0119] This step is divided into three sub-steps: S31, S32, and S33.

[0120] S31. Calculate the positive support and negative rejection for each disease.

[0121] Input: Positive symptom set S+ negative symptom set S - The cropped candidate disease set D' and normalized probability P'(d) (optional, or the score can be calculated directly without relying on prior probability).

[0122] Positive Support Score + (d): Traversing S + For each symptom s in the query, the conditional probability P(s|d) is calculated and multiplied by the preset weight w for that symptom. i (The weight is usually 1.0, but can be adjusted according to the importance of symptoms); then sum all the products: This value reflects the strength of support for disease d from the current positive symptoms.

[0123] Negative Repulsion Score - (d): Traversing S - For each symptom s in the query, the conditional probability P(s|d) is multiplied by a preset weight w. j (Usually 0.5 or 1.0, adjusted according to the actual situation), then sum: This value reflects the strength of rejection of disease d by negative symptoms.

[0124] Output: Score for each disease d + (d) and Score - (d) value.

[0125] S32. Diseases whose absolute value of the difference between positive support and negative rejection is less than a preset threshold are classified into the differential diagnosis set.

[0126] Input: Score for each disease + (d) and Score - (d) Preset threshold ε (ε=0.3).

[0127] For each disease d, calculate its degree of conflict of evidence. .

[0128] If Δ(d) < ε, the disease is considered to be in a state of conflicting evidence: it is neither sufficiently proven nor sufficiently falsified, and there is uncertainty. These diseases are classified into the Differential Set. The remaining diseases are classified into the Primary Set and classified according to the comprehensive score. (λ is a preset coefficient, usually taken as 1.0) Sort in descending order.

[0129] S33. For diseases in the differential diagnosis set, activate the delayed decision-making mechanism.

[0130] Each disease in the differential diagnosis set is marked as "pending confirmation".

[0131] In the subsequent step S4 (consultation decision optimization), the system will prioritize asking questions based on symptoms that can distinguish the differences between these diseases.

[0132] The diseases in this set will not be excluded, nor will they be determined as a final diagnosis; instead, a decision will be made after more evidence is available.

[0133] Output: The updated state vector, containing at least the Primary Set, Differential Set, and S. + S - .

[0134] S4: Through the consultation decision optimization module, the optimal consultation question is selected based on the current state and output to the user.

[0135] Input: Current state st=(S + ,S - (Primary Set, Differential Set), and optional intervention commands from the safety constraint control module.

[0136] Action space generation: From all symptoms associated with the Primary Set and Differential Set, filter those that have not yet been asked (i.e., not in S). + ∪S - The symptoms of (in the middle) constitute the candidate action set A. t .

[0137] Safety intervention judgment: If the safety constraint control module sends a mandatory inquiry instruction (the instruction specifies a specific high-risk symptom), the reinforcement learning decision in this step is skipped, and the high-risk symptom is directly selected as the action.

[0138] Differentiation-first strategy: If the Differential Set is not empty, prioritize selecting from symptoms that best differentiate disease pairs in the set; specifically, for each disease pair (d i ,d j )∈Differential Set, calculate the conditional probability difference |P(s|d) for each symptom. i )-P(s|d j Choose the symptom that maximizes the difference. If multiple symptoms have similar distinguishing abilities, choose the one with the highest overall information gain.

[0139] Reinforcement learning model decision-making: If the Differential Set is empty or the discriminative policy fails to select a unique action, a pre-trained reinforcement learning model (such as a deep Q-network) is invoked. The model takes the current state encoding as input and outputs in the action space A.t Based on the Q-value distribution, the action with the highest Q-value is selected as the optimal consultation question.

[0140] Question output: Convert the selected action (symptom) into a natural language question, such as "Do you have [symptom name]?", and display it to the patient through the user interface.

[0141] Output: The consultation questions sent to the user.

[0142] S5. Obtain user feedback on the consultation questions through the data processing module, and update the positive symptom set and the negative symptom set.

[0143] The user answers the question output in step S4 (usually "yes", "no" or "unsure").

[0144] The user's answer is sent to the data processing module, which performs the following operations:

[0145] Analyze the answers to identify the corresponding symptoms (i.e., the symptoms previously asked about) and the user's affirmative or negative intent.

[0146] If the answer is "yes", then add the symptom to the set of positive symptoms S. + .

[0147] If the answer is "no", then add the symptom to the negative symptom set S. - .

[0148] If the answer is "uncertain", it will not be added to any set for the time being, pending further evidence or a fuzzy processing strategy.

[0149] For those already existing in S + or S - Do not repeat the symptoms mentioned above.

[0150] Status Update: Updated S + and S - This will be used as input for the next iteration, and steps S2 to S5 will be executed again.

[0151] S6. Repeat steps S2 to S5 until the preset termination condition is met, and output the diagnostic results through the diagnostic output module.

[0152] Iterative loop: Steps S2→S3→S4→S5 constitute a complete consultation cycle. In each round, the system uses newly added negative or positive evidence to further prune the candidate space, update disease probabilities, adjust stratification status, and generate the next question. This loop continues until any of the following termination conditions are met:

[0153] Primary candidate set convergence: Only one disease remains in the Primary Set, and its normalized probability or comprehensive score exceeds a preset threshold (e.g., 0.85).

[0154] Confidence threshold achieved: The overall score or posterior probability of the top-ranked disease exceeds a high threshold (e.g., 0.90), regardless of whether there are other candidate diseases.

[0155] Maximum consultation rounds reached: To prevent infinite loops, a maximum number of rounds (e.g., 15 rounds) is set, and the system will terminate the consultation once this limit is reached.

[0156] Early termination of safety module: If the safety constraint control module detects a high-risk situation that cannot be eliminated, it sends an early termination command.

[0157] User-initiated termination: Users can choose to end the consultation early.

[0158] Output diagnostic results: When the termination condition is met, the diagnostic output module performs the following operations:

[0159] Obtain the final Primary Set, Differential Set, and detailed scores for each disease from the Conflict Fusion and Hierarchical Candidate Management module.

[0160] Obtain risk records from the safety constraint control module (including confirmed high-risk symptoms, unconfirmed but alert symptoms, etc.).

[0161] Generate a structured diagnostic report, which should include at least:

[0162] Most likely disease name and its confidence level (probability or score);

[0163] List of differential diagnoses (if applicable);

[0164] Risk level indication (high, medium, low);

[0165] Tiered medical care recommendations (such as "home observation", "outpatient visit", "go to the emergency room immediately").

[0166] The report is converted into natural language text and displayed to patients through the user's device.

[0167] End of Method: This consultation session ends after the output is complete.

[0168] The key parameters and adjustable items in this embodiment are shown in Table 1:

[0169] Table 1 Key Parameters and Adjustable Items

[0170]

[0171] Example 3: Validation of decision-making accuracy based on real clinical data.

[0172] This embodiment selects anonymized case data from outpatient data of tertiary hospitals, totaling:

[0173] Disease types: Approximately 120 common internal medicine diseases;

[0174] Number of cases: 10,000;

[0175] Each case includes: chief complaint, positive symptom set S⁺, negative symptom set S⁻, and final diagnosis (as the gold standard).

[0176] Two control systems were constructed simultaneously:

[0177] Control Method A: Traditional ranking method based on positive symptom scores;

[0178] Comparative Method B: Reinforcement learning diagnostic system without negative evidence clipping.

[0179] Experimental Procedure: In this embodiment, the method described in Embodiment 2 (executed according to steps S1 to S6) and control methods A and B were used to simulate consultations on 10,000 case data. Each consultation process followed the complete iterative logic of the method of this invention until the termination condition was met, at which point a diagnostic result was output. The diagnostic outputs of each method were recorded and compared with the gold standard.

[0180] The following metrics are used to evaluate system performance:

[0181] Top-1 accuracy: The proportion of cases where the first-line diagnosis is consistent with the gold standard;

[0182] Top-3 coverage: The proportion of cases where the gold standard is included in the top three diagnostic results;

[0183] Average consultation rounds: The average number of interactions required to complete a consultation for each case;

[0184] False deletion rate: The proportion of real diseases that are excluded by the system.

[0185] The experimental results are shown in Table 2 below:

[0186] Table 2 Decision Accuracy Verification Table

[0187]

[0188] Experimental results demonstrate that the Top-1 accuracy (79.4%) and Top-3 coverage (91.2%) of the method of this invention are significantly better than those of control methods A and B. The average number of consultation rounds of the method of this invention (6.1 rounds) is significantly less than that of the two control methods (9.3 rounds and 8.7 rounds), indicating that the negative evidence-driven spatial pruning mechanism effectively improves consultation efficiency. The false deletion rate of the method of this invention (2.3%) is much lower than that of control method B (6.8%), proving that the symptom cluster mechanism and differential diagnosis set management can effectively avoid the incorrect exclusion of real diseases.

[0189] This embodiment demonstrates that the present invention significantly improves diagnostic accuracy through a negative symptom-driven candidate pruning mechanism; the number of consultation rounds is significantly reduced due to the reduction of irrelevant disease interference; and the "differential diagnosis set" mechanism effectively avoids the problem of mistakenly deleting real diseases.

[0190] Example 4: System stability verification (noise immunity) experiment.

[0191] This embodiment is used to verify the robustness of the method of the present invention under conditions of incomplete input information, noise, and ambiguous expression.

[0192] Experimental Design: Based on the original test set, three types of noise data were constructed:

[0193] Random mislabeling: 10% of negative symptoms were randomly mislabeled as positive or not mentioned;

[0194] Random deletion: 15% of positive symptom information was randomly deleted;

[0195] Ambiguity: Simulate non-standard input from real users, introducing ambiguous expressions and synonym replacements.

[0196] The method in Embodiment 2 of the present invention and the control method B are respectively applied to the above-mentioned noise data to perform the consultation and decision-making process.

[0197] The evaluation indicators are as follows:

[0198] Accuracy decrease: the change in accuracy relative to a noise-free environment;

[0199] Diagnostic outcome volatility: The degree of consistency in diagnostic results across multiple runs of the same case;

[0200] Candidate set stability: The rate of change of the candidate disease set in each round.

[0201] The experimental results are shown in Table 3:

[0202] Table 3 Stability Verification Results

[0203]

[0204] Experimental results demonstrate that the accuracy decrease of the method of the present invention in noisy environments is significantly less than that of the control method B; the volatility of the diagnostic results of the method of the present invention is significantly lower than that of the control method B, indicating that the output results are more stable and reliable; the volatility of the candidate set of the method of the present invention is much smaller than that of the control method B, indicating that the hierarchical candidate management strategy effectively avoids the interference of local errors on the overall decision-making.

[0205] This embodiment demonstrates that the present invention, through a soft exclusion mechanism and conflict fusion strategy, exhibits strong robustness against noise; hierarchical candidate management avoids the impact of local errors on overall decision-making; and the system maintains stable output even when information is incomplete.

[0206] Example 5: Clinical fit verification experiment (physician participation in the evaluation).

[0207] This embodiment is used to verify the performance of the method of the present invention in terms of clinical logical rationality, professionalism and risk control capabilities, and to evaluate its degree of conformity with the actual clinical decision-making process.

[0208] Experimental setup: Ten clinicians with more than 5 years of experience were invited to participate in the evaluation. A certain number of typical cases were selected, and the doctors conducted simulated consultations with the system of this invention. Without knowing the internal mechanism of the system, the doctors made subjective evaluations of the consultation process of the system.

[0209] Evaluation dimensions include:

[0210] Logical rationality of the consultation: Does the order of the system's questions conform to clinical reasoning logic?

[0211] Question professionalism: Are the questions posed by the system professional and accurate?

[0212] Clinical consistency: Whether the systematic diagnostic reasoning process is consistent with the physician's own clinical thinking;

[0213] Risk control capability: Whether the system can proactively identify and prioritize the handling of high-risk symptoms.

[0214] The rating is based on a 5-point scale, with 5 being the highest score and 1 being the lowest score.

[0215] Table 4 shows the average scores given by doctors to the method of this invention across various evaluation dimensions:

[0216] Table 4 Average Rating Table

[0217]

[0218] Doctors gave the method of this invention high scores across all evaluation dimensions, including:

[0219] Reasonableness of the consultation logic: Most doctors believe that the system's strategy of prioritizing the use of negative symptoms to rule out a large number of diseases, and then gradually focusing on a few candidate diseases, is highly consistent with the clinical approach of "rule by rule and then identify".

[0220] Professionalism: The questions posed by the system are all based on standard medical terminology and have clear differential diagnostic value;

[0221] Clinical consistency: The system's diagnostic reasoning path closely matches the physician's actual clinical decision-making path;

[0222] Risk control capability: Doctors generally believe that the system's safety constraint mechanism can effectively alert to and avoid high-risk situations.

[0223] Taking a patient with "chest pain" as an example, the following is a detailed explanation of the systematic consultation process:

[0224] Patient input: Positive symptom set S + Chest pain, worsening after activity; negative symptom set S - No fever, no cough, and no history of trauma.

[0225] Systemic findings: Rapid exclusion: Using negative symptoms, the system first excludes infectious diseases (such as pneumonia) and trauma-related diseases (such as rib fractures); Candidate retention: Cardiovascular-related diseases (such as coronary heart disease, angina pectoris, myocarditis, etc.) are retained in the main candidate set, and high-risk diseases (such as unstable angina pectoris, acute myocardial infarction) are included in the differential diagnosis set.

[0226] Priority Questions: The system will then prioritize asking about key symptoms relevant to the high-risk differential diagnosis, including: whether sweating is present; whether there is radiating pain; and whether the pain is sudden and severe.

[0227] The assessing physicians unanimously agreed that the consultation path conformed to the classic clinical thinking pattern of "first rule out risks, then subdivide the risks," and that high-risk diseases were not mistakenly ruled out throughout the process.

[0228] This embodiment demonstrates that the consultation path conforms to the classic clinical thinking pattern of "first eliminating risks, then further subdividing"; high-risk diseases were not mistakenly excluded throughout the consultation process; the system can proactively focus on asking questions about the most valuable symptoms for identification; based on comprehensive doctor scores and case analysis, the method of this invention received high scores in terms of the rationality of consultation logic, the professionalism of questions, clinical consistency, and risk control capabilities, proving that it is highly consistent with the actual clinical decision-making process and has good clinical acceptance and interpretability.

[0229] Example 6: Security verification (high-risk missed diagnosis control).

[0230] This embodiment is specifically used to verify the safety performance of the method of the present invention in high-risk disease scenarios and to evaluate the system's ability to control the risk of missed diagnoses.

[0231] Experimental data: A total of 1,500 cases including high-risk diseases were selected, covering the following types:

[0232] Acute cardiovascular events (such as acute myocardial infarction, unstable angina, aortic dissection);

[0233] Acute neurological events (such as cerebral hemorrhage, acute ischemic stroke);

[0234] Acute abdominal conditions (such as acute pancreatitis, intestinal obstruction, and gastrointestinal perforation).

[0235] Evaluation indicators include:

[0236] High-risk disease identification rate: The proportion of cases that correctly include high-risk diseases in the system's final diagnosis or differential diagnosis;

[0237] High-risk symptom trigger rate: The proportion of cases in which the system actively inquires about or identifies high-risk symptoms during the consultation process;

[0238] Early termination rate: The percentage of cases where an early termination command is triggered by the safety constraint module.

[0239] The performance of the method of the present invention on the above indicators is shown in Table 5:

[0240] Table 5 Performance Results of Each Evaluation Indicator

[0241]

[0242] Experimental results show that:

[0243] High-risk disease identification rate (96.8%): The method of this invention can effectively identify the vast majority of high-risk diseases and include them in the main candidate or differential diagnosis set, avoiding incorrect exclusion due to insufficient information.

[0244] High-risk symptom trigger rate (98.2%): The safety constraint control module can proactively detect unconfirmed high-risk symptoms and trigger inquiries during the consultation process to ensure that key risk signals are not missed.

[0245] Safe termination trigger rate (21.5%): In about one-fifth of high-risk cases, the system judged the risk level to be too high, proactively terminated the consultation in advance and provided medical advice, which reflects the system's risk warning and intervention capabilities.

[0246] This embodiment verifies the safety performance of the method of the present invention in high-risk disease scenarios: the safety constraint control module can effectively identify key risk signals; when necessary, the system can proactively terminate the consultation and advise the patient to seek medical treatment immediately; through the forced inquiry and early termination mechanism, the risk of missed diagnosis of high-risk diseases is significantly reduced.

[0247] Example 7: A medical consultation decision optimization device based on negative evidence, comprising at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to implement a medical consultation decision optimization method based on negative evidence, the method comprising: acquiring patient symptom information through the data processing module and constructing a positive symptom set and a negative symptom set; and, through the negative evidence-driven candidate space pruning module, performing multi-level exclusion and probability attenuation on the candidate disease set based on the negative symptom set and the symptom clusters to obtain a pruned candidate disease set. The system combines and normalizes the probability distribution of each candidate disease; through the conflict fusion and hierarchical candidate management module, based on the positive symptom set, negative symptom set, and the pruned disease probability, it calculates the positive support and negative rejection of each candidate disease and divides the candidate diseases into a main candidate set and a differential diagnosis set; through the consultation decision optimization module, it selects the optimal consultation question based on the current state and outputs it to the user; through the data processing module, it obtains the user's feedback on the consultation question and updates the positive symptom set and negative symptom set; it repeats the above steps S2 to S5 until the preset termination condition is met, and outputs the diagnosis result through the diagnosis output module.

[0248] This specific embodiment is merely an explanation of the present invention and is not intended to limit the invention. After reading this specification, those skilled in the art can make modifications to this embodiment without contributing any inventive step, but such modifications are protected by patent law as long as they are within the scope of the claims of the present invention.

Claims

1. A medical consultation decision optimization system based on negative evidence, characterized in that: It includes a data processing module, a medical knowledge modeling module, a negative evidence-driven candidate space trimming module, a conflict fusion and hierarchical candidate management module, a consultation decision optimization module, a safety constraint control module, and a diagnostic output module; The data processing module is used to receive and parse the consultation information input by the user, extract and standardize the symptoms, and construct a set of positive symptoms and a set of negative symptoms. The medical knowledge modeling module is used to construct and store a medical knowledge graph, which includes a disease set, a symptom set, disease-symptom association probabilities, and several predefined symptom clusters, each of which contains one or more clinically relevant symptoms. The candidate space pruning module driven by negative evidence is communicatively connected to the data processing module and the medical knowledge modeling module, respectively. It is used to receive the negative symptom set and the symptom cluster, perform multi-level exclusion on the candidate disease set, calculate the comprehensive exclusion score based on the symptom cluster to update the disease probability, and output the pruned candidate disease set and its normalized probability distribution. The conflict fusion and hierarchical candidate management module is communicatively connected to the negative evidence-driven candidate space trimming module. It is used to receive the trimmed disease probability distribution, and combine the positive symptom set and the negative symptom set to calculate the positive support and negative rejection of each candidate disease. Based on the calculation results, the candidate diseases are divided into the main candidate set and the differential diagnosis set. The consultation decision optimization module is communicatively connected to the conflict fusion and hierarchical candidate management module, and is used to receive the current state, which includes at least the main candidate set and the differential diagnosis set, and select the optimal consultation question through a reinforcement learning model and output it to the user terminal. The safety constraint control module is communicatively connected to the consultation decision optimization module, and is used to monitor unconfirmed high-risk symptoms in real time, and send a mandatory inquiry command to the consultation decision optimization module or an early termination command to the diagnosis output module when a risk is detected. The diagnostic output module is communicatively connected to the conflict fusion and hierarchical candidate management module and the safety constraint control module, respectively, and is used to output diagnostic results, differential diagnosis list, risk level prompts and medical suggestions in a formatted manner when the preset termination conditions are met.

2. The medical consultation decision optimization system based on negative evidence as described in claim 1, characterized in that: The candidate space pruning module driven by negative evidence is specifically used for: Strong exclusion is performed: when a disease exists, and the probability of the occurrence of a certain symptom is greater than a preset high threshold, and the symptom belongs to the set of negative symptoms, the disease is directly excluded from the candidate disease set. Perform soft exclusion: Based on the comprehensive rejection score of symptom clusters, probability decay is performed. For each disease and each symptom cluster, the comprehensive rejection score of the cluster is calculated as follows: take the maximum value of the conditional probabilities of all symptoms denied by the patient for the disease within the symptom cluster; then, multiply the current probability of the disease by the rejection factor corresponding to each symptom cluster, where each rejection factor is equal to 1 minus the comprehensive rejection score of the cluster, thereby updating the probability of the disease. Perform probability normalization: Divide the updated probabilities of all candidate diseases by the sum of the probabilities of all candidate diseases to obtain the normalized probability distribution.

3. The medical consultation decision optimization system based on negative evidence as described in claim 1, characterized in that: The symptom cluster generation logic in the medical knowledge modeling module includes: constructing a symptom co-occurrence matrix based on historical medical record data, calculating the correlation strength between symptoms, and using a clustering algorithm to classify symptoms with a correlation strength greater than a preset threshold into the same symptom cluster; wherein, symptoms within the same symptom cluster are defined as having clinical substitutability, and only their maximum rejection contribution value is taken when calculating the comprehensive rejection score.

4. The medical consultation decision optimization system based on negative evidence as described in claim 1, characterized in that: The conflict fusion and hierarchical candidate management module is specifically used for: Calculate the positive support for each disease: for each symptom in the set of positive symptoms, multiply it by its conditional probability under that disease, and sum the results in a weighted manner; Calculate the negative exclusion degree for each disease: for each symptom in the set of negative symptoms, multiply it by its conditional probability under that disease, and sum the results in a weighted manner; The overall score for each disease is defined as the product of positive support minus negative rejection and a preset coefficient; Diseases with an absolute value of a comprehensive score less than a preset threshold are classified into the differential diagnosis set to indicate diseases with conflicting evidence. For diseases in the differential diagnosis set, a delayed decision mechanism is enabled to preserve their probability and prioritize the generation of discriminative consultation questions for them in subsequent rounds.

5. The medical consultation decision optimization system based on negative evidence as described in claim 1, characterized in that: The reward function of the reinforcement learning model in the consultation decision optimization module is a multi-objective weighted function, specifically including: a diagnostic accuracy reward, a consultation efficiency reward, and a risk penalty. The diagnostic accuracy reward gives a positive high score when the final diagnosis is correct, the consultation efficiency reward is inversely proportional to the number of consultation rounds to encourage a reduction in the number of rounds, and the risk penalty gives a large negative score when the model misses high-risk symptoms.

6. The medical consultation decision optimization system based on negative evidence as described in claim 1, characterized in that: The safety constraint control module is specifically used for: Real-time detection of any unconfirmed high-risk symptoms, i.e., symptoms that are neither in the positive symptom set nor the negative symptom set; When an unconfirmed high-risk symptom is detected, a mandatory inquiry instruction is sent to the consultation decision optimization module to force priority inquiry into the high-risk symptom. When the preset safety termination conditions are met, an early termination command is sent to the diagnostic output module to terminate the consultation and output medical advice.

7. A medical consultation decision optimization method based on negative evidence, applied to the consultation decision system according to any one of claims 1-6, characterized in that: Includes the following steps: S1. Obtain the patient's symptom information through the data processing module, and construct a set of positive symptoms and a set of negative symptoms; S2. Through the candidate space pruning module driven by the negative evidence, based on the negative symptom set and the symptom cluster, the candidate disease set is subjected to multi-level exclusion and probability decay to obtain the pruned candidate disease set and its normalized probability distribution. S3. Through the conflict fusion and hierarchical candidate management module, based on the positive symptom set, negative symptom set and the cropped disease probability, calculate the positive support and negative rejection of each candidate disease, and divide the candidate diseases into the main candidate set and the differential diagnosis set. S4. The consultation decision optimization module selects the optimal consultation question based on the current state and outputs it to the user terminal. S5. Obtain user feedback on the consultation questions through the data processing module, and update the positive symptom set and the negative symptom set; S6. Repeat steps S2 to S5 until the preset termination condition is met, and output the diagnostic result through the diagnostic output module.

8. The medical consultation decision optimization method based on negative evidence as described in claim 7, characterized in that: Step S2 involves multi-level exclusion and probability decay of the candidate disease set, specifically including: S21. Strong exclusion: For any disease, if any symptom exists, and the probability of the symptom occurring under the disease is greater than a preset high threshold and the symptom belongs to the set of negative symptoms, then the disease is directly excluded. S22, Soft exclusion: For each disease and each symptom cluster, first calculate the maximum value of the conditional probability of all denied symptoms in the cluster for the disease, as the comprehensive exclusion score of the cluster; then multiply the current probability of the disease by the exclusion factor corresponding to each symptom cluster, where each exclusion factor is equal to 1 minus the comprehensive exclusion score of the cluster, to obtain the updated disease probability. S23. Probability Normalization: Divide the updated probabilities of all candidate diseases by the sum of the probabilities of all candidate diseases.

9. The medical consultation decision optimization method based on negative evidence as described in claim 7, characterized in that: Step S3 involves dividing the candidate diseases into a primary candidate set and a differential diagnosis set, specifically including: S31. Calculate the positive support and negative rejection for each disease; S32. Diseases whose absolute value of the difference between positive support and negative rejection is less than a preset threshold are classified into the differential diagnosis set; S33. For diseases in the differential diagnosis set, a delayed decision mechanism is enabled, and questions that can distinguish different diseases in the set are generated first in subsequent rounds.

10. A medical consultation decision optimization device based on negative evidence, characterized in that: The method includes at least one processor; and a memory communicatively connected to at least one of the processors; wherein the memory stores instructions executable by the processor to implement a medical consultation decision optimization method based on negative evidence driven by any one of claims 7-9.

Images