Acute abdominal pain diagnosis system and method based on feature mapping and dynamic game arbitration

Abstract

The application relates to the technical field of medical artificial intelligence and clinical decision support systems, and provides an acute abdominal disease diagnosis system and method based on feature mapping and dynamic game arbitration. The system extracts a standardized global abnormal feature vector by using a multi-modal perception and mapping module, independently reasons corresponding N different acute abdominal disease types by using a multi-agent parallel reasoning engine to obtain output states, and uses a conflict detection and arbitration agent to monitor the N output states, triggers a game arbitration mechanism when a diagnosis conflict exists, discriminates scores of the conflicting output states by using a threshold weighted feature discrimination matrix, and further determines a diagnosis result, effectively alleviates a logical conflict problem in multi-label classification, realizes "white box" and interpretability of a diagnosis process, simultaneously realizes a "gray" decision in line with clinical thinking, the diagnosis process is closer to uncertainty decision requirements of a real medical scene, and the diagnosis result credibility is improved.

Description

Technical Field

[0001] This application relates to the field of medical artificial intelligence and clinical decision support systems, and in particular to a diagnostic system and method for acute abdomen based on feature mapping and dynamic game arbitration. Background Technology

[0002] Acute abdomen has always been a challenge in clinical diagnosis due to its rapid onset, diverse types, and high degree of overlap in characteristics. Current Clinical Decision Support System (CDSS) technologies mainly fall into two categories:

[0003] 1) Rule-based expert systems: These methods rely on predefined logical rules by medical experts (e.g., "If white blood cell count > 10 and right lower quadrant pain, appendicitis is suspected"). While highly interpretable, they struggle to handle the large amounts of unstructured natural language text in Electronic Health Records (EHRs), and the rule base is extremely expensive to maintain, making them unable to cope with complex and varied clinical phenotypes.

[0004] 2) End-to-end deep learning-based methods: In recent years, Large Language Models (LLMs) based on Transformers have been applied to medical diagnosis. While these methods excel at understanding text, they suffer from the "illusion" problem, and the reasoning process is a "black box," failing to provide a chain of evidence that meets the requirements of evidence-based medicine. More importantly, general models output probability distributions, which often lead to logically mutually exclusive "double positive" results when encountering diseases with similar characteristics (such as appendicitis and diverticulitis), lacking an effective conflict resolution mechanism.

[0005] In other words, existing computer-aided diagnostic technologies mainly face the following mathematical and logical challenges: the curse of dimensionality in unstructured data, the mutual exclusion conflict in multi-label classification, and the uninterpretability of the reasoning process.

[0006] Therefore, there is an urgent need for an intelligent diagnostic solution that can transform natural language into formal features and resolve diagnostic conflicts through independent parallel reasoning and dynamic game mechanisms. Summary of the Invention

[0007] In view of this, embodiments of this application provide a diagnostic system and method for acute abdominal pain based on feature mapping and dynamic game arbitration, in order to solve the problems of difficulty in processing unstructured data, diagnostic conflicts caused by overlapping features of multiple diseases, and lack of interpretable evidence chains in end-to-end models in existing auxiliary diagnostic technologies for acute abdominal pain.

[0008] A first aspect of this application provides a diagnostic system for acute abdomen based on feature mapping and dynamic game arbitration, comprising:

[0009] The multimodal perception and mapping module is configured to receive multi-source heterogeneous data and extract features from the multi-source heterogeneous data through a task-adaptive hybrid feature extraction engine to obtain a standardized global anomaly feature vector.

[0010] The multi-agent parallel reasoning engine is configured to use N disease diagnosis agents to independently reason based on a standardized global anomaly feature vector, and obtain N output states containing quadruples; where each disease diagnosis agent corresponds to a type of acute abdominal disease, and each disease diagnosis agent is injected with clinical guideline knowledge constraints for the corresponding disease type; N is a positive integer greater than 1.

[0011] The conflict detection and arbitration agent is configured to monitor N output states. In response to determining that at least two output states have a diagnostic conflict, it triggers a game arbitration mechanism, calls a threshold-weighted feature discrimination matrix to score each output state that has a conflict, and determines the diagnostic result based on the score of each output state.

[0012] A second aspect of this application provides a method for diagnosing acute abdomen based on feature mapping and dynamic game arbitration, including:

[0013] It receives multi-source heterogeneous data, extracts features from the multi-source heterogeneous data through a task-adaptive hybrid feature extraction engine, and obtains a standardized global anomaly feature vector.

[0014] N disease diagnosis agents are used to independently reason based on standardized global anomaly feature vectors, resulting in N output states containing quadruples; each disease diagnosis agent corresponds to a type of acute abdominal disease, and each disease diagnosis agent is injected with clinical guideline knowledge constraints for the corresponding disease type; N is a positive integer greater than 1.

[0015] Monitor N output states. In response to determining that at least two output states have a diagnostic conflict, trigger the game arbitration mechanism, call the threshold-weighted feature discrimination matrix to score each output state that has a conflict, and determine the diagnostic result based on the score of each output state.

[0016] A third aspect of this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method.

[0017] The beneficial effects of the embodiments of this application compared with the prior art are:

[0018] This application embodiment extracts hybrid features from multi-source heterogeneous patient diagnostic data through a multimodal perception and mapping module to obtain a standardized global anomaly feature vector. A multi-agent parallel inference engine then independently infers based on this standardized global anomaly feature vector to derive output states corresponding to N different acute abdominal disease types. A conflict detection and arbitration agent monitors these N output states. When a diagnostic conflict exists, a game arbitration mechanism is triggered. A threshold-weighted feature discrimination matrix is ​​used to score each conflicting output state, thereby determining the diagnostic result. This effectively alleviates the logical conflict problem in multi-label classification, achieving "white-box" and interpretability in the diagnostic process. Simultaneously, it enables "gray-scale" decision-making that aligns with clinical thinking, making the diagnostic process closer to the uncertain decision-making needs of real medical scenarios and improving the credibility of diagnostic results. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the structure of an acute abdominal pain diagnosis system based on feature mapping and dynamic game arbitration provided in an embodiment of this application.

[0021] Figure 2 This is a data flow logic diagram of the engine provided in the embodiments of this application.

[0022] Figure 3 This is a schematic diagram of the algorithm flow of the game arbitration mechanism provided in the embodiments of this application.

[0023] Figure 4 This is a schematic diagram of the self-evolutionary closed-loop process provided in the embodiments of this application.

[0024] Figure 5 This is a flowchart illustrating an intelligent diagnostic method for acute abdominal pain based on feature mapping of multi-source heterogeneous data and dynamic game arbitration provided in an embodiment of this application.

[0025] Figure 6 This is a schematic diagram of the electronic device provided in the embodiments of this application. Detailed Implementation

[0026] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0027] The following will describe in detail, with reference to the accompanying drawings, an acute abdominal diagnosis system and method based on feature mapping and dynamic game arbitration according to an embodiment of this application.

[0028] As mentioned above, existing computer-aided diagnostic technologies mainly face the following mathematical and logical challenges: the curse of dimensionality in unstructured data, the mutual exclusion conflict in multi-label classification, and the uninterpretability of the reasoning process.

[0029] The curse of dimensionality in unstructured data refers to the fact that EHRs are high-dimensional unstructured texts, and traditional Natural Language Processing (NLP) methods struggle to reduce their dimensionality to computable feature vectors while preserving medical semantics.

[0030] The mutual exclusion conflict in multi-label classification refers to the fact that existing deep learning models usually output probability distributions of multiple labels. However, in the case of acute abdominal pain, multiple diseases (such as appendicitis and diverticulitis) have highly overlapping features in the feature space, which often leads to the model outputting logically mutually exclusive results and lacks an automated conflict resolution mechanism.

[0031] The inexplicability of the reasoning process refers to the lack of explicit expression of intermediate states in the end-to-end model, which fails to meet the clinical requirement for traceability of the chain of evidence.

[0032] In view of this, in order to solve the aforementioned mathematical and logical challenges, this application provides an intelligent diagnostic system for acute abdominal pain based on feature mapping and dynamic game arbitration. Through a neural symbolic reasoning architecture, it realizes transparent reasoning from natural language medical records to accurate diagnosis, and effectively solves the logical mutual exclusion problem in complex cases by using a game arbitration mechanism. This addresses the problems existing in current auxiliary diagnostic technologies for acute abdominal pain, such as the difficulty in processing unstructured data, diagnostic conflicts caused by overlapping features of multiple diseases, and the lack of interpretable evidence chains in end-to-end models.

[0033] Figure 1 This is a schematic diagram of the structure of an acute abdominal diagnosis system based on feature mapping and dynamic game arbitration provided in an embodiment of this application. Figure 1 As shown, the system includes a multimodal perception and mapping module, a multi-agent parallel inference engine, and a conflict detection and arbitration agent.

[0034] The multimodal perception and mapping module is configured to receive multi-source heterogeneous data and extract features from the multi-source heterogeneous data through a task-adaptive hybrid feature extraction engine to obtain a standardized global anomaly feature vector.

[0035] The multi-agent parallel reasoning engine is configured to use N disease diagnosis agents to independently reason based on a standardized global anomaly feature vector, resulting in N output states containing quadruples; where each disease diagnosis agent corresponds to a type of acute abdominal pain, and each disease diagnosis agent is injected with clinical guideline knowledge constraints for the corresponding disease type; N is a positive integer greater than 1.

[0036] The conflict detection and arbitration agent is configured to monitor N output states. In response to determining that at least two output states have a diagnostic conflict, it triggers a game arbitration mechanism, calls a threshold-weighted feature discrimination matrix to score each output state that has a conflict, and determines the diagnostic result based on the score of each output state.

[0037] According to the technical solution provided in the embodiments of this application, a multimodal perception and mapping module is used to extract mixed features from multi-source heterogeneous patient diagnostic data to obtain a standardized global abnormal feature vector. A multi-agent parallel reasoning engine is used to independently reason based on this standardized global abnormal feature vector to obtain output states corresponding to N different acute abdominal disease types. A conflict detection and arbitration agent is used to monitor the N output states. When a diagnostic conflict exists, a game arbitration mechanism is triggered. A threshold-weighted feature identification matrix is ​​used to identify and score each conflicting output state, thereby determining the diagnostic result. This effectively alleviates the logical conflict problem in multi-label classification, realizes the "white-box" and interpretability of the diagnostic process, and realizes "gray-scale" decision-making that conforms to clinical thinking. The diagnostic process is closer to the uncertainty decision-making needs of real medical scenarios, and improves the credibility of diagnostic results.

[0038] In some embodiments of this application, multi-source heterogeneous data includes at least the patient's current medical history data, physical examination data, laboratory test data, and imaging report data.

[0039] Feature extraction from multi-source heterogeneous data using a task-adaptive hybrid feature extraction engine can include: preprocessing and structuring the multi-source heterogeneous data; and performing semantic-to-symbolic feature mapping on the structured multi-source heterogeneous data to obtain a standardized global anomaly feature vector.

[0040] Furthermore, preprocessing and structuring the multi-source heterogeneous data can include: converting each data point into a string; removing redundancy and desensitizing each string using regular expressions; locating the index position of each anchor word in the processed string by calling a pre-set clinical paragraph anchor word library, and using a sliding window mechanism to extract the text between two adjacent anchor words as a sub-data block of the modality corresponding to the first anchor word; wherein, the first anchor word is the preceding anchor word among two adjacent anchor words; the pre-set clinical paragraph anchor word library includes at least the anchor words present medical history, physical examination, laboratory tests, and imaging; marking the missing content of modalities; and combining the sub-data blocks of the modalities corresponding to each anchor word to obtain the structured multi-source heterogeneous data.

[0041] In other words, the diagnostic system can first receive multi-source heterogeneous original medical record data from patients, and then perform text denoising and segmentation cleaning on the original medical record data, decoupling it into four independent sub-modal text blocks. At this time, the data still retains its natural language form in order to preserve complete semantic details.

[0042] In some examples, the raw medical record data received by the system can be recorded as a string. Then, use regular expressions to remove meaningless control characters (such as HTML tags). (e.g., , etc.) and redundant whitespace characters. Next, a privacy desensitization operation can be performed, replacing patient names, doctor names, etc., with preset characters, such as double underscore placeholders, according to preset entity recognition rules, so that the desensitized data meets data compliance requirements.

[0043] Further, the anonymized data can be anchored and segmented. In some examples, the system can have a built-in database of clinical paragraph anchor terms. The system iterates through the anonymized data to locate the index positions of the aforementioned anchor words. Using a sliding window mechanism, two adjacent anchor words are linked together. and The text content extracted between "present medical history" and "physical examination" is defined as a sub-data block of that modality. For example, the text between "present medical history" and "physical examination" can be extracted and assigned to... .

[0044] Then, missing value and anomaly detection can be performed. During detection, if a certain modality (such as...) If the extracted content is empty, the system automatically marks that dimension as NULL to prevent the feature extractor from producing illusion output on empty text during subsequent feature extraction.

[0045] The final structured output can be represented as ;in, For the present medical history text (e.g., "migratory right lower quadrant abdominal pain for 3 hours"), For physical examination text (e.g., "McBurney's point tenderness (+)"), For laboratory testing data, This is an imaging examination report.

[0046] To enable the inference engine to perform computable inference, the diagnostic system provided in this application embodiment can also perform a "semantic-symbolic" transformation on the above-mentioned structured output to obtain a standardized global anomaly feature vector.

[0047] In some embodiments of this application, performing semantic-to-symbolic feature mapping on structured multi-source heterogeneous data may include: calling a rule-based deterministic extractor to extract features from a first type of sub-data block in the structured multi-source heterogeneous data using a first feature extraction method to obtain a first feature; the first feature extraction method includes a hard logic matching method based on regular expressions; the first type of sub-data block includes a sub-data block of the modality corresponding to the anchor word laboratory; calling an entity recognition model to extract features from a second type of sub-data block in the structured multi-source heterogeneous data using a second feature extraction method to obtain a second feature; the second feature extraction... The extraction methods include using a BERT-CRF entity extraction model fine-tuned from a medical corpus to extract key signs; the second type of sub-data block includes sub-data blocks of anchor words corresponding to the modalities of the body search; the generative semantic understander is invoked to use the third feature extraction method to extract features from the third type of sub-data blocks in the structured multi-source heterogeneous data to obtain the third feature; the third feature extraction method includes using a large language model fine-tuned with instructions for semantic structuring; the third type of sub-data block includes sub-data blocks of anchor words corresponding to the current medical history and imaging modalities; the first feature, the second feature and the third feature are combined to obtain a standardized global anomaly feature vector.

[0048] The BERT-CRF entity extraction model is a deep learning model that combines a pre-trained language model (Bidirectional Encoder Representations from Transformers, BERT) and a Conditional Random Field (CRF). The entity recognition model can be a dedicated entity recognition model, specifically trained for the application scenario of diagnosing acute abdominal pain.

[0049] In other words, this application embodiment designs a task-adaptive hybrid feature extraction engine. Unlike the traditional single-model logic that processes all text, the diagnostic system provided in this application embodiment dynamically routes to three heterogeneous extractors based on the differences in input modalities:

[0050] Rule deterministic extractor (for) For strongly structured data such as laboratory indicators, hard logic matching based on regular expressions (Regex) is used to ensure zero error in numerical extraction.

[0051] Entity recognition model (for) For key physical signs in the examination text, a BERT-CRF entity extraction model finely tuned with medical corpus is used to extract key signs to ensure the precise boundaries of medical terms.

[0052] Generative semantic understanders (for) and For complex temporal and causal descriptions in present medical history and imaging reports, a large language model with instruction tuning is used for semantic structuring to capture implicit logic.

[0053] This design, which decouples deterministic logic from generative logic, ensures both the accuracy of key indicators and the semantic richness of natural language. The extraction process can be represented by a formula. ;in, Indicates the source of input information The corresponding feature extractor.

[0054] Next, the mapping function is standardized. By mapping natural language to the SNOMED-CT standard terminology set, a standardized global anomaly feature vector can be generated. As factual basis for subsequent reasoning: Among them, U j This is the union operator, which merges the feature sets extracted from the four submodals: present medical history, physical examination, laboratory tests, and imaging. These are standardized clinical features (or feature elements), representing clinically significant atomized diagnostic facts mapped from raw medical data.

[0055] The engine's data flow logic is as follows: Figure 2 As shown. Reference Figure 2 If the raw data includes "white blood cells" The symptoms "right lower quadrant tenderness," "persistent abdominal pain for 2 days," and "appendiceal dilation" can be analyzed using a rule-based deterministic extractor based on regular expressions for the data "white blood cells." Feature extraction was performed to obtain the feature "white blood cell count: high". An entity recognition model was used to extract the feature "right lower quadrant tenderness" from the data "right lower quadrant tenderness". Furthermore, a generative semantic understander was used based on a large language model to extract the features "persistent abdominal pain for 2 days" and "appendix dilatation", resulting in the feature "appendix dilatation, abdominal pain for 2 days". Indicates the unit liter.

[0056] Next, the features extracted by the three extractors can be mapped and merged through the medical terminology mapping module to obtain a standardized global feature vector. .

[0057] The task-adaptive hybrid feature extraction mechanism provided in this application overcomes the limitations of a single Neuro-Linguistic Programming (NLP) model in processing complex medical records. By combining the accuracy of the rule engine with the semantic understanding capabilities of a large model, it effectively solves the problems of "curse of dimensionality" and key information extraction bias in electronic medical records.

[0058] In some embodiments of this application, the diagnostic system may define a set of logically mutually exclusive candidate diseases that require differential diagnosis. ,For example, =Acute appendicitis. =Acute cholecystitis, etc. (For the candidate disease set) Each type of disease The diagnostic system can call upon independent diagnostic agents. To reason. Among them, It is a positive integer less than or equal to N.

[0059] Each of them All were infused with clinical guideline knowledge specific to the corresponding disease. In one example, It can be structured (Prompt word) Template, which embeds the diagnostic scoring criteria for the disease (such as the Alvarado scoring rules for appendicitis).

[0060] The target disease diagnostic agent obtains its output state, comprising a four-tuple, as follows: Based on standardized global anomaly feature vectors, the agent determines the original posterior probability of the corresponding disease type and saves the inference chain; the original posterior probability is mapped to the corresponding confidence level according to a preset clinical threshold range; the mapped confidence level is one of the following: exclusion confidence level, suspected confidence level, very likely confidence level, and confirmed confidence level; the disease type, inference chain, confidence level, and other information are combined to obtain the four-tuple of the target disease diagnostic agent, thus yielding the output state of the target disease diagnostic agent. The target disease diagnostic agent can be any disease diagnostic agent.

[0061] In other words, for each type of disease Diagnostic intelligent agents Output the raw posterior probability Based on a preset clinical threshold range, it is mapped to four discrete confidence levels. :

[0062] ;

[0063] The final output is a quadruple. ;in: Indicates the target disease or possible subtype of the disease for diagnosis; For the generated natural language reasoning chain (e.g., "The patient presents with metastatic right lower quadrant abdominal pain and elevated white blood cell count, consistent with an Alvarado score of 7, therefore the judgment is positive"), Information that other agents deem necessary to mention (such as clinical clues pointing to other diseases).

[0064] The multi-agent parallel reasoning engine provided in this application not only outputs results but also generates structured natural language reasoning chains based on injected guidance knowledge, enabling doctors to fully trace the basis of diagnosis and meet clinical requirements for the traceability of evidence chains. Furthermore, unlike traditional binary classification (0 or 1) which is too coarse-grained and difficult to adapt to the uncertainty of early-stage symptoms, and considering that in early-stage acute abdominal pain, symptoms are often atypical, and doctors typically do not directly determine "yes" or "no," but rather say "possible" or "highly probable," the segmented confidence model provided in this application is closer to the uncertainty decision-making needs of real-world medical scenarios.

[0065] In some embodiments of this application, the conflict detection and arbitration agent determines that a diagnostic conflict has occurred in the following manner: constructing a set to be identified; the set to be identified includes output states in a quadruple with confidence levels of suspected confidence level, very likely confidence level, and confirmed confidence level; in response to determining that the number of output states in the set to be identified is greater than or equal to 2, a diagnostic conflict is determined to have occurred.

[0066] Figure 3 This is a schematic diagram of the algorithm flow of the game arbitration mechanism provided in the embodiments of this application. Figure 3 As shown, after the game arbitration algorithm starts, it can first receive data from each diagnostic agent. Output status Then construct the review set. The filtering rules are as follows: ,Right now This includes states with a confidence level of "suspected" or higher.

[0067] Next, the system calculates the set. cardinality :

[0068] like The output of the "primary diagnosis" is "no obvious signs of acute abdomen", and the output of complete process information, including feature extraction results and agent reasoning process.

[0069] like Output the "primary diagnosis" as the single disease, along with complete process information.

[0070] That is, if the judgment If the condition is not met, a complete and traceable diagnostic report can be output, and the game arbitration algorithm will end.

[0071] Conversely, if (i.e., there are two or more diseases at the suspected or higher level): This is determined as "diagnostic conflict exists". The diagnostic system automatically triggers a dynamic game arbitration mechanism to conduct differential diagnosis.

[0072] Among them, the base Represents a set The number of elements contained therein, i.e., the set to be identified. The number of disease output statuses with a medium confidence level of suspected or higher.

[0073] In some embodiments of this application, the game arbitration mechanism uses the following method to score the discrimination of each output state: calling the threshold-weighted feature discrimination matrix W, where the elements of matrix W are the discrimination weights. Characteristic features For disease Discrimination weights; Discrimination weights As a dynamically adjustable parameter, its initial value is set based on clinical guidelines and dynamically updated based on clinical feedback by a self-evolving agent; for each output state where a diagnostic conflict occurs, the formula is used... Calculate the score for the type of disease corresponding to this output state; where, For disease The score, To standardize the global anomaly feature vector, As an indicator function, if the patient possesses the following characteristics but ,otherwise ; For disease The systemic risk coefficient, such as the risk coefficient for acute pancreatitis, is set at 1.5 to increase the diagnostic priority of high-risk diseases; This is a risk adjustment coefficient used to prioritize warnings for high-risk disease types when scores are similar.

[0074] Still referencing Figure 3 After the dynamic game arbitration mechanism is triggered, it can call the threshold-weighted feature discrimination matrix W to score each output state and determine the primary diagnostic result based on the score of each output state. Then, a complete and traceable diagnostic report is output, and the game arbitration algorithm ends.

[0075] In other words, when arbitration is triggered, the arbitration agent can invoke an external authentication knowledge base. Mathematically, this process is represented by the calculation of a "weighted feature matching score".

[0076] Threshold-weighted feature discrimination matrix W: Elements in the matrix This represents the weight of feature v in distinguishing disease d. For example, suppose feature v is "age of onset > 60 years". In the discrimination matrix, the weight of this feature for "appendicitis" is... It can be set to "-0.5" (negative correlation), while the weight for "diverticulitis" is... It can be set to "+0.8" (strong positive correlation).

[0077] Calculate the discrimination score: for conflict sets For each disease in the formula Calculate its score.

[0078] Final decision: The disease with the highest score will be selected as the primary diagnosis. .

[0079] In some embodiments of this application, the diagnostic system can generate a structured clinical support report based on the arbitration result. This report includes not only the final primary diagnosis. It also includes a Critical Decision Path Summary.

[0080] Specifically, the system backtracks the inference logs of the multi-agent parallel inference engine to extract "decisive features." For example, the system's first-choice diagnosis is "acute appendicitis" (confidence level: Established). Key supporting evidence is "migratory right lower quadrant pain" and "Alvarado score > 7"; the key differential criterion is "absence of jaundice," thus ruling out "acute cholecystitis" during the arbitration stage.

[0081] In addition, the report interface also provides an evidence hyperlink function. When a doctor clicks on a reasoning node (such as "elevated white blood cell count"), the system can directly highlight the corresponding data source in the original medical record, realizing full-chain traceability from the conclusion to the original data.

[0082] As can be seen, unlike the "black box" nature of end-to-end models, this application's embodiments employ a neuro-symbolic architecture. This neuro-symbolic architecture is a hybrid processing architecture that integrates large language model semantic neural networks with clinical guideline weighted matrix symbolic computation, comprising two clearly defined processing streams: "neural perception" and "symbolic arbitration."

[0083] In the "neural perception" stage, the powerful unstructured text processing capabilities of deep learning models (such as large language models and BERT) are used as the system's "sensors" to extract facts from the medical records written in the doctor's natural language and transform them into standardized global feature vectors (symbolization).

[0084] In the "symbolic arbitration" phase, addressing the pain points of large language models being prone to "illusions" and logically mutually exclusive (double positive) results, the system blocks and takes over the generative decision-making of the large language model when encountering diagnostic conflicts. Instead, it invokes a fully deterministic "weighted feature discrimination matrix" and mathematical formula (Score_d) for adjudication. The initial logic of this matrix is ​​derived from explicit clinical guidelines, ensuring 100% certainty and white-box interpretability of diagnostic results.

[0085] Furthermore, neural symbolic processing can be understood as a form of explicit rule-based reasoning. For example, if guidelines suggest that "a certain indicator being more than three times the reference value is one of the important criteria for diagnosis, and a diagnosis can be made with two pieces of evidence," then this indicator being greater than 3 is a specific element of neural symbolic reasoning.

[0086] This approach addresses the technical problem of existing deep learning models frequently outputting "double positive" results when dealing with diseases with overlapping features (such as appendicitis and diverticulitis). The embodiments of this application introduce a dynamic game arbitration mechanism into the acute abdominal pain diagnosis system. By introducing a specificity discrimination matrix, the positive and negative contributions of features to different diseases are mathematically quantified, achieving automated conflict resolution based on quantized weights.

[0087] Meanwhile, unlike some related technologies where the initial weights are manually set when using the discrimination matrix for classification, the initial weights in the matrix W of this application are directly mapped from the sensitivity and specificity of the feature for a specific disease in clinical guidelines. For example, highly specific exclusionary features are assigned high negative discrimination weights, thereby ensuring that the arbitration judgment does not deviate from the clinical guidelines and avoids obvious misdiagnosis.

[0088] Furthermore, this application embodiment provides the game arbitration mechanism with a "human-in-the-loop" continuous self-evolution capability: unlike the traditional model update mode that relies on large-scale retraining, this application embodiment designs a parameter fine-tuning mechanism based on semantic feedback. The system can understand the doctor's corrective behavior, automatically analyze inference loopholes through a large language model, and adjust the weight parameters of the discrimination matrix in a targeted manner. This mechanism enables the system to continuously adapt to the disease characteristics of specific hospitals or specific populations during clinical use, achieving low-cost iteration that becomes more accurate with use.

[0089] In some embodiments of this application, a self-evolving agent dynamically updates the identification weights based on clinical feedback. This can include: receiving a corrected diagnostic result, triggering the self-evolving agent to extract the original inference chain and the corrected diagnostic result from the system, and performing a difference analysis on the original inference chain and the corrected diagnostic result; updating the value of the target identification weight in matrix W based on the difference analysis result; wherein the target identification weight is any one or more identification weights in matrix W; performing a virtual regression test in a preset historical case database using a test identification matrix containing the updated target identification weight, and calculating the change in diagnostic accuracy; and determining the test identification matrix as the updated identification matrix W in response to the determination that the decrease in accuracy does not exceed a preset tolerance threshold.

[0090] In other words, when the diagnostic system receives a revised diagnostic result, it can trigger an Evolution Agent to extract the system's original inference chain and the revised diagnostic result, and perform a difference analysis on the original inference chain and the revised diagnostic result. The revised diagnostic result can be the doctor's initial diagnosis obtained from other analysis systems. Conclusions that are not adopted, for example, the diagnostic system recommends "acute appendicitis," but the doctor corrects it to "acute diverticulitis."

[0091] The self-evolutionary closed-loop process is as follows: Figure 4 As shown. Reference Figure 4 A self-evolving intelligent agent can execute the following "reflection-correction" closed loop:

[0092] Root Cause Analysis: The system reasoning chain is read from the complete and traceable diagnostic report output by the self-evolving agent's self-game arbitration algorithm, and the correction results are read to analyze the reasons for the reasoning failure (e.g., the system over-assigned the "elderly" feature to the discrimination matrix W, resulting in a missed diagnosis of appendicitis in the elderly).

[0093] Parameter Proposal: Based on the analysis results, the self-evolving agent generates knowledge constraints specific to the discrimination matrix W or the corresponding disease-specific clinical guidelines. Update instructions (Patch). For example, update the weight of advanced age (over 60 years old) at the time of appendicitis diagnosis from -0.5 to -0.2. ); or in appendicitis diagnosis intelligence ( The prompt includes the statement: 'If the patient is pregnant, the Alvarado score weight is reduced.'

[0094] Virtual Validation and Hot Update: The diagnostic system can perform "virtual regression testing" on a new parameter set within a pre-built historical case database. This calculates the decrease in diagnostic accuracy of the new parameter set on the historical validation set. .like If the preset tolerance threshold is not exceeded, it is determined that there is no significant decrease.

[0095] The tolerance threshold can be dynamically set according to the specific clinical risk level (e.g., setting a critical value for statistical significance P-value), and this application does not impose strict restrictions on it. In one example, the tolerance threshold can be set to 2%.

[0096] If the new parameters, while correcting current errors, do not significantly decrease the accuracy of historically correct cases, the updated parameters are formally written into the database, completing a single self-evolutionary iteration of the system. The hot update can be based on game arbitration using an identification matrix or hierarchical probabilistic reasoning based on guideline constraints. Hierarchical probabilistic reasoning refers to the "confidence grading of diagnostic conclusions." Multiple agents work independently on the same dimension for different diseases, without interfering with each other; and each agent outputs not a simple "yes / no," but a confidence level with a medical grayscale space. Therefore, the technical solution provided in this application conforms to the "grayscale decision-making" characteristics of clinical thinking during reasoning, and is more accurate and reasonable than traditional binary classification models.

[0097] Figure 5 This is a flowchart illustrating a method for diagnosing acute abdomen based on feature mapping and dynamic game arbitration, provided in an embodiment of this application. Figure 5 As shown, the method includes the following steps:

[0098] In step S501, multi-source heterogeneous data is received, and features are extracted from the multi-source heterogeneous data through a task-adaptive hybrid feature extraction engine to obtain a standardized global anomaly feature vector.

[0099] In step S502, N disease diagnosis agents are used to independently perform reasoning based on the standardized global anomaly feature vector to obtain N output states containing quadruples.

[0100] Each disease diagnosis agent corresponds to a type of acute abdominal pain, and each disease diagnosis agent is injected with clinical guideline knowledge constraints for the corresponding disease type; N is a positive integer greater than 1.

[0101] In step S503, N output states are monitored. In response to determining that at least two output states have a diagnostic conflict, a game arbitration mechanism is triggered. The threshold-weighted feature discrimination matrix is ​​called to score each output state that has a conflict, and the diagnostic result is determined based on the score of each output state.

[0102] According to the technical solution provided in the embodiments of this application, a multimodal perception and mapping module is used to extract mixed features from multi-source heterogeneous patient diagnostic data to obtain a standardized global abnormal feature vector. A multi-agent parallel reasoning engine is used to independently reason based on this standardized global abnormal feature vector to obtain output states corresponding to N different acute abdominal disease types. A conflict detection and arbitration agent is used to monitor the N output states. When a diagnostic conflict exists, a game arbitration mechanism is triggered. A threshold-weighted feature identification matrix is ​​used to identify and score each conflicting output state, thereby determining the diagnostic result. This effectively alleviates the logical conflict problem in multi-label classification, realizes the "white-box" and interpretability of the diagnostic process, and realizes "gray-scale" decision-making that conforms to clinical thinking. The diagnostic process is closer to the uncertainty decision-making needs of real medical scenarios, and improves the credibility of diagnostic results.

[0103] In some implementations, the multi-source heterogeneous data includes at least the patient's current medical history data, physical examination data, laboratory test data, and imaging report data; feature extraction is performed on the multi-source heterogeneous data using a task-adaptive hybrid feature extraction engine, including: preprocessing and structuring the multi-source heterogeneous data; and performing semantic-to-symbolic feature mapping on the structured multi-source heterogeneous data to obtain a standardized global anomaly feature vector.

[0104] In some implementations, preprocessing and structuring of multi-source heterogeneous data include: converting each data point into a string; removing redundancy and desensitizing each string using regular expressions; locating the index position of each anchor word in the processed string by calling a pre-set clinical paragraph anchor word library, and using a sliding window mechanism to extract the text between two adjacent anchor words as a sub-data block of the modality corresponding to the first anchor word; wherein, the first anchor word is the preceding anchor word among two adjacent anchor words; the pre-set clinical paragraph anchor word library includes at least the anchor words present medical history, physical examination, laboratory tests, and imaging; marking the missing content of modalities; and combining the sub-data blocks of the modalities corresponding to each anchor word to obtain the structured multi-source heterogeneous data.

[0105] In some implementations, feature mapping from semantics to symbols is performed on the structured multi-source heterogeneous data, including: calling a rule-based deterministic extractor to extract features from a first type of sub-data block in the structured multi-source heterogeneous data using a first feature extraction method to obtain a first feature; the first feature extraction method includes a hard logic matching method based on regular expressions; the first type of sub-data block includes sub-data blocks of the modality corresponding to anchor words; calling an entity recognition model to extract features from a second type of sub-data block in the structured multi-source heterogeneous data using a second feature extraction method to obtain a second feature; the second feature extraction method... This includes a method for extracting key signs using a BERT-CRF entity extraction model fine-tuned from a medical corpus; the second type of sub-data block includes sub-data blocks of anchor words corresponding to the modalities of the body search; a generative semantic understander is invoked to use a third feature extraction method to extract features from the third type of sub-data blocks in the structured multi-source heterogeneous data, resulting in a third feature; the third feature extraction method includes a method for semantic structuring using a large language model fine-tuned with instructions; the third type of sub-data block includes sub-data blocks of anchor words corresponding to the current medical history and imaging modalities; and the first, second, and third features are combined to obtain a standardized global anomaly feature vector.

[0106] In some implementations, the target disease diagnostic agent obtains its output state containing a four-tuple as follows: Based on standardized global anomaly feature vectors, the agent determines the original posterior probability of the corresponding disease type and saves the inference chain; the original posterior probability is mapped to the corresponding confidence level according to a preset clinical threshold range; wherein the mapped confidence level is one of the following: exclusion confidence level, suspected confidence level, very likely confidence level, and confirmed confidence level; the agent's corresponding disease type, inference chain, confidence level, and other information are combined to obtain the four-tuple of the target disease diagnostic agent, thereby obtaining the output state of the target disease diagnostic agent; wherein the target disease diagnostic agent is any disease diagnostic agent.

[0107] In some implementations, the conflict detection and arbitration agent determines that a diagnostic conflict has occurred by constructing a set to be identified; the set to be identified includes output states in a quadruple with confidence levels of suspected confidence level, very likely confidence level, and confirmed confidence level; and in response to determining that the number of output states in the set to be identified is greater than or equal to 2, a diagnostic conflict is determined to have occurred.

[0108] In some implementations, the game arbitration mechanism uses the following method to score the discrimination of each output state: It calls a threshold-weighted feature discrimination matrix W, where the elements of matrix W... Characteristic features For disease The discrimination weight; where the discrimination weight As a dynamically adjustable parameter, its initial value is set based on clinical guidelines and dynamically updated based on clinical feedback by a self-evolving agent; for each output state where a diagnostic conflict occurs, the formula is used... Calculate the score for the type of disease corresponding to this output state; where, For disease The score, To standardize the global anomaly feature vector, As an indicator function, if the patient possesses the following characteristics but ,otherwise ; For disease The systemic risk coefficient, This is a risk adjustment coefficient used to prioritize warnings for high-risk disease types when scores are similar.

[0109] In some implementations, the identification weights are dynamically updated based on clinical feedback by a self-evolving agent. The process includes: receiving a corrected diagnostic result, triggering the self-evolving agent to extract the original inference chain and the corrected diagnostic result from the system, and performing a difference analysis on the original inference chain and the corrected diagnostic result; updating the value of the target identification weight in matrix W based on the difference analysis result; wherein the target identification weight is any one or more identification weights in matrix W; performing a virtual regression test in a preset historical case database using a test identification matrix containing the updated target identification weight, and calculating the change in diagnostic accuracy; and determining the test identification matrix as the updated identification matrix W in response to the determination that the decrease in accuracy does not exceed a preset tolerance threshold.

[0110] All of the above-mentioned optional technical solutions can be combined in any way to form the optional embodiments of this application, and will not be described in detail here.

[0111] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0112] Figure 6 This is a schematic diagram of the electronic device 6 provided in an embodiment of this application. Figure 6 As shown, the electronic device 6 of this embodiment includes a processor 601, a memory 602, and a computer program 603 stored in the memory 602 and executable on the processor 601. When the processor 601 executes the computer program 603, it implements the steps in the various method embodiments described above. Alternatively, when the processor 601 executes the computer program 603, it implements the functions of each module / unit in the various device embodiments described above.

[0113] Electronic device 6 can be a desktop computer, laptop, handheld computer, cloud server, or other electronic device. Electronic device 6 may include, but is not limited to, processor 601 and memory 602. Those skilled in the art will understand that... Figure 6 This is merely an example of electronic device 6 and does not constitute a limitation on electronic device 6. It may include more or fewer components than shown, or different components.

[0114] The processor 601 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0115] The memory 602 can be an internal storage unit of the electronic device, such as a hard disk or memory of the electronic device 6. The memory 602 can also be an external storage device of the electronic device 6, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, Flash Card, etc., equipped on the electronic device 6. The memory 602 can also include both internal and external storage units of the electronic device 6. The memory 602 is used to store computer programs and other programs and data required by the electronic device.

[0116] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0117] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A diagnostic system for acute abdomen based on feature mapping and dynamic game arbitration, characterized in that, include: The multimodal perception and mapping module is configured to receive multi-source heterogeneous data and extract features from the multi-source heterogeneous data through a task-adaptive hybrid feature extraction engine to obtain a standardized global anomaly feature vector. The multi-agent parallel reasoning engine is configured to use N disease diagnosis agents to independently reason based on the standardized global anomaly feature vector, and obtain N output states containing quadruplets; wherein, each disease diagnosis agent corresponds to a type of acute abdominal disease, and each disease diagnosis agent is injected with clinical guideline knowledge constraints for the corresponding disease type; N is a positive integer greater than 1; The conflict detection and arbitration agent is configured to monitor N output states. In response to determining that at least two output states have a diagnostic conflict, it triggers a game arbitration mechanism, calls a threshold-weighted feature discrimination matrix to score each output state that has a conflict, and determines the diagnostic result based on the score of each output state.

2. The acute abdominal pain diagnosis system based on feature mapping and dynamic game arbitration according to claim 1, characterized in that, The multi-source heterogeneous data includes at least the patient's current medical history, physical examination data, laboratory test data, and imaging report data; Feature extraction is performed on the multi-source heterogeneous data using a task-adaptive hybrid feature extraction engine, including: The multi-source heterogeneous data is preprocessed and structured. The standardized global anomaly feature vector is obtained by performing semantic-to-symbolic feature mapping on the structured multi-source heterogeneous data.

3. The acute abdominal pain diagnosis system based on feature mapping and dynamic game arbitration according to claim 2, characterized in that, The multi-source heterogeneous data undergoes preprocessing and structuring, including: Convert each data item into a string; Use regular expressions to remove redundancy and desensitize each string; The system calls a pre-defined clinical paragraph anchor terminology library to locate the index position of each anchor term in the processed string, and uses a sliding window mechanism to extract the text between two adjacent anchor terms as a sub-data block of the modality corresponding to the first anchor term. The first anchor word is the preceding anchor word among two adjacent anchor words; the pre-set clinical paragraph anchor word library includes at least the anchor words present illness history, physical examination, laboratory tests, and imaging. Mark the missing modalities with empty content; By combining the sub-data blocks of the modal corresponding to each anchor word, the structured multi-source heterogeneous data is obtained.

4. The acute abdominal pain diagnosis system based on feature mapping and dynamic game arbitration according to claim 2, characterized in that, Performing semantic-to-symbolic feature mapping on structured, multi-source heterogeneous data, including: The rule-based deterministic extractor uses a first feature extraction method to extract features from a first type of sub-data block in the structured multi-source heterogeneous data to obtain a first feature; the first feature extraction method includes a hard logic matching method based on regular expressions; the first type of sub-data block includes a sub-data block of the modality corresponding to the anchor word laboratory. The entity recognition model is invoked to use the second feature extraction method to extract features from the second type of sub-data blocks in the structured multi-source heterogeneous data, thereby obtaining the second feature; the second feature extraction method includes a method for extracting key signs using a BERT-CRF entity extraction model fine-tuned with medical corpus; the second type of sub-data block includes sub-data blocks corresponding to the anchor word body search modality; The generative semantic understander is invoked to use a third feature extraction method to extract features from the third type of sub-data blocks in the structured multi-source heterogeneous data, thereby obtaining the third feature; the third feature extraction method includes a method of semantic structuring using a large language model with instruction fine-tuning; the third type of sub-data blocks includes sub-data blocks of anchor words, present medical history, and image-corresponding modalities; The standardized global anomaly feature vector is obtained by combining the first feature, the second feature, and the third feature.

5. The acute abdominal pain diagnosis system based on feature mapping and dynamic game arbitration according to claim 1, characterized in that, The target disease diagnosis agent obtains the output state containing a quadruple as follows: Based on the standardized global anomaly feature vector, the original posterior probability of the corresponding type of disease of this agent is determined and the inference chain is saved; The original posterior probability is mapped to the corresponding confidence level according to a preset clinical threshold range; wherein the confidence level obtained by mapping is one of the following: exclusion confidence level, suspected confidence level, very likely confidence level, and confirmed confidence level; The disease type, inference chain, confidence level and other information corresponding to this agent are combined to obtain the quadruple of the target disease diagnosis agent, and then the output state of the target disease diagnosis agent is obtained. The target disease diagnosis agent can be any disease diagnosis agent.

6. The acute abdominal pain diagnosis system based on feature mapping and dynamic game arbitration according to claim 5, characterized in that, The conflict detection and arbitration agent determines the occurrence of a diagnostic conflict in the following way: Construct a set to be identified; the set to be identified includes output states in the quadruple with confidence levels of suspected confidence level, very likely confidence level, and confirmed confidence level; A diagnostic conflict is determined when the number of output states in the set to be identified is greater than or equal to 2.

7. The acute abdominal pain diagnosis system based on feature mapping and dynamic game arbitration according to claim 1, characterized in that, The game arbitration mechanism uses the following method to identify and score each output state: Call the threshold-weighted feature discrimination matrix W, where the elements of matrix W are the discrimination weights. Characteristic features For disease The identification weight; Discrimination weight It is a dynamically adjustable parameter, with its initial value set based on clinical guidelines and dynamically updated based on clinical feedback by a self-evolving agent. For each output state where a diagnostic conflict occurs, use the formula Calculate the score for the type of disease corresponding to this output state; where, For disease The score, To standardize the global anomaly feature vector, As an indicator function, if the patient possesses the following characteristics but ,otherwise ; For disease The systemic risk coefficient, This is a risk adjustment coefficient used to prioritize warnings for high-risk disease types when scores are similar.

8. The acute abdominal pain diagnosis system based on feature mapping and dynamic game arbitration according to claim 7, characterized in that, The identification weights are dynamically updated based on clinical feedback by a self-evolving intelligent agent. ,include: Upon receiving the corrected diagnostic result, the self-evolving agent extracts the original inference chain of the system and the corrected diagnostic result, and performs a difference analysis on the original inference chain and the corrected diagnostic result. Based on the difference analysis results, update the values ​​of the target identification weights in matrix W; where the target identification weights are any one or more identification weights in matrix W. In a pre-built historical case database, a virtual regression test is performed using a test discrimination matrix that includes updated target discrimination weights to calculate the magnitude of change in diagnostic accuracy. In response to determining that the decrease in accuracy does not exceed a preset tolerance threshold, the test discrimination matrix is ​​determined to be the updated discrimination matrix W.

9. A diagnostic method for acute abdomen based on feature mapping and dynamic game arbitration, characterized in that, include: Receive multi-source heterogeneous data, and extract features from the multi-source heterogeneous data through a task-adaptive hybrid feature extraction engine to obtain a standardized global anomaly feature vector; N disease diagnosis agents are used to independently reason based on the standardized global anomaly feature vector to obtain N output states containing quadruples; where each disease diagnosis agent corresponds to a type of acute abdominal disease, and each disease diagnosis agent is injected with clinical guideline knowledge constraints for the corresponding disease type; N is a positive integer greater than 1. Monitor N output states. In response to determining that at least two output states have a diagnostic conflict, trigger the game arbitration mechanism, call the threshold-weighted feature discrimination matrix to score each output state that has a conflict, and determine the diagnostic result based on the score of each output state.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the acute abdominal diagnosis method based on feature mapping and dynamic game arbitration as described in claim 9.

Images