Acute stomachache cause differential diagnosis system based on deep learning

Abstract

The invention discloses an acute abdominal pain cause differential diagnosis system based on deep learning, and relates to the technical field of medical artificial intelligence, comprising: a data acquisition module acquires initial self-described text data of a patient; the standardization processing module generates a standardized symptom set through standardization processing; the cause reasoning module performs reasoning according to the standardized symptom set and the medical knowledge graph to obtain a suspected disease set; the contradiction detection module is used for detecting the contradiction between the symptom set and the self-contained text and calculating the contradiction intensity and disease cause relevancy; the priority calculation module ranks contradiction priorities according to contradiction intensity and disease cause relevancy; the contradiction processing module judges the contradiction significance score, if the contradiction significance score is lower than a threshold value, a standardized symptom set is output, and otherwise, an optimal clarification action is generated to solve the contradiction; an iteration updating module updates a symptom set according to clarification action feedback, and causes are inferred again until a termination condition is met; the method can ensure that the pathogenesis reasoning process gradually approaches the real pathogenesis, and improves the diagnosis reliability.

Description

Technical Field

[0001] This invention relates to the field of medical artificial intelligence technology, specifically to a deep learning-based system for differential diagnosis of the causes of acute abdominal pain. Background Technology

[0002] Acute abdominal pain is one of the most common clinical manifestations in the emergency department. Its causes are complex, ranging from benign gastrointestinal dysfunctions to life-threatening acute abdominal conditions (such as appendicitis and aortic dissection). Rapid and accurate identification of the cause is crucial for subsequent treatment decisions and patient prognosis.

[0003] In recent years, with the in-depth application of artificial intelligence technology in the medical field, especially the progress of deep learning in natural language processing and medical knowledge representation, many auxiliary diagnostic systems have emerged. These systems aim to provide doctors with diagnostic suggestions by analyzing text data such as patient complaints and electronic medical records, thereby alleviating the pressure on emergency rooms and reducing the risk of misdiagnosis.

[0004] Existing technical solutions mostly focus on identifying and matching symptoms described by patients, and generating a set of suspected causes based on pre-constructed disease-symptom association maps. However, these methods often assume that the symptom information provided by patients is consistent and reliable in medical logic, without fully considering the potential for ambiguity, chronological inconsistencies, or logical conflicts in patient self-reports. In practical applications, due to differences in patients' expressive abilities, subjective biases, or the anxiety of acute conditions, their self-reports often contain semantic ambiguity, contradictions, or descriptions that do not conform to common medical knowledge (such as the simultaneous presence of right lower quadrant tenderness and a soft, non-tender abdomen). Existing systems lack mechanisms to detect and handle such contradictions, which directly leads to unreasonable etiological assumptions, and consequently, makes subsequent diagnostic reasoning based on unreliable evidence.

[0005] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a deep learning-based system for differential diagnosis of the causes of acute abdominal pain.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows:

[0008] In a first aspect, the present invention discloses a deep learning-based system for differential diagnosis of the etiology of acute abdominal pain, comprising:

[0009] The data acquisition module is used to acquire the patient's initial self-report text data;

[0010] The standardization processing module is used to obtain a standardized symptom set based on the initial self-reported text data through medical entity recognition and standardization processing.

[0011] The etiology reasoning module is used to obtain a set of suspected diseases by reasoning based on a standardized set of symptoms and a pre-built medical knowledge graph.

[0012] The contradiction detection module is used to obtain a contradiction set based on the context of the standardized symptom set and the initial self-report text data through contradiction detection processing, and to calculate the contradiction intensity of each contradiction in the contradiction set and its correlation with the etiology of the suspected disease set; the contradiction set represents two or more propositions in the patient's description that are medically logically conflicting.

[0013] The priority calculation module is used to perform weighted calculations based on the intensity of the contradiction and the correlation of the cause, obtain the significance score of the contradiction, and sort them in descending order to obtain the contradiction priority queue.

[0014] The conflict handling module is used to determine whether the conflict significance score of the first conflict in the conflict priority queue is lower than the preset significance threshold. If it is, a standardized symptom set is output; otherwise, based on the first conflict, the suspected disease set, and the medical knowledge graph, the optimal clarification action is generated through knowledge graph traversal and benefit calculation. The optimal clarification action is executed to verify or resolve the current conflict.

[0015] The iterative update module is used to update the standardized symptom set based on the feedback data after performing the optimal clarification action, and to re-execute the severe etiology reasoning and contradiction significance judgment until the termination condition is met.

[0016] Secondly, the present invention discloses a method for differential diagnosis of the etiology of acute abdominal pain based on deep learning, including: obtaining the patient's initial self-reported text data;

[0017] Based on the initial self-reported text data, a standardized symptom set is obtained through medical entity recognition and standardization processing;

[0018] Based on the standardized symptom set and the pre-constructed medical knowledge graph, a set of suspected diseases is obtained through etiological reasoning.

[0019] Based on the context of the standardized symptom set and the initial self-reported text data, a contradiction set is obtained through contradiction detection, and the contradiction intensity of each contradiction in the contradiction set and its correlation with the etiology of the suspected disease set are calculated; the contradiction set represents two or more propositions in the patient's description that are medically logically conflicting.

[0020] The significance score of the contradiction is obtained by weighting the contradiction intensity and the correlation of the cause, and then sorting them in descending order to obtain the contradiction priority queue.

[0021] Determine whether the significance score of the first contradiction in the contradiction priority queue is lower than the preset significance threshold. If so, output the standardized symptom set; otherwise, based on the first contradiction, the suspected disease set, and the medical knowledge graph, generate the optimal clarification action through knowledge graph traversal and benefit calculation. The optimal clarification action is executed to verify or resolve the current contradiction.

[0022] The standardized symptom set is updated based on the feedback data after the optimal clarification action is performed, and the serious etiology reasoning and contradiction significance judgment are re-executed until the termination condition is met.

[0023] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0024] 1. By introducing a contradiction detection mechanism, the system can identify logical conflicts and semantic ambiguities in the patient's self-report, avoid making etiological inferences based on erroneous or inconsistent input data, thereby reducing the risk of misdiagnosis and improving the accuracy of etiological identification of acute abdominal pain;

[0025] 2. By generating optimal clarification actions and combining them with a medical knowledge graph to dynamically correct suspected diseases, the system can proactively verify and update data after discovering contradictions, ensuring that the etiology reasoning process gradually approaches the true etiology, significantly improving the real-time performance and reliability of diagnosis. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 This is an overall block diagram of the system architecture according to Embodiment 1 of the present invention;

[0028] Figure 2 This is a flowchart illustrating the execution process of the system according to Embodiment 1 of the present invention;

[0029] Figure 3 This is a flowchart of the feedback data update process in the system of Embodiment 1 of the present invention;

[0030] Figure 4 This is an overall block diagram of the method in Embodiment 2 of the present invention. Detailed Implementation

[0031] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] Application Overview: In traditional natural language processing-based acute abdominal pain auxiliary diagnostic systems, patient self-reported text is directly input into the disease classification model after medical entity recognition. The system assumes that all extracted symptoms are true and consistent medical facts. Due to the lack of a mechanism to detect semantic contradictions and logical conflicts within the text, when the patient's description contains subjective expression biases or medical common sense errors, the erroneous symptom data will directly contaminate the knowledge graph matching process, causing the disease inference results to deviate from the true cause. At the same time, the system cannot generate dynamic interactive instructions to correct the quality of the input data.

[0033] For example, an emergency patient's complaint text contains two entities: severe abdominal pain lasting 3 hours and a pain score of 3. The existing system treats these two as independent symptoms and matches them with their corresponding weights in the knowledge base. When there is a logical contradiction between the patient's actual pain intensity and self-reported pain score, the system cannot recognize the value of this contradiction in differentiating between acute appendicitis and irritable bowel syndrome, and directly outputs a suspected set containing both diseases. At this point, the system does not trigger any clarification mechanism, leading to the subsequent calculation of disease probabilities incorrectly using a low pain score as the basis for excluding acute abdominal pain.

[0034] If the above problems are not addressed, contradictory symptoms will cause the disease reasoning path to deviate from the clinical reality, and the confidence of the etiology set output by the system will decrease significantly. In complex cases, undetected semantic conflicts may cause abnormal disease probability distribution, resulting in key etiologies being incorrectly excluded from the suspected set. At the same time, the system cannot simulate the decision-making process of doctors verifying contradictions through interaction, and its diagnostic results lack dynamic correction capabilities, making it difficult to meet the dual requirements of diagnostic reliability and real-time performance in emergency scenarios.

[0035] Faced with the above problems, this application first recognizes that potential logical contradictions in patients' self-reported texts can contaminate the disease reasoning process, and existing technologies lack dynamic verification mechanisms. In response, this application reconstructs the diagnostic process into a closed-loop system that includes contradiction detection and feedback correction. By constructing a contradiction intensity quantification model and a clarification action generation mechanism, the system can proactively identify semantic conflicts and trigger interactive verification, thereby purifying the input data in the iteration process.

[0036] Example 1:

[0037] like Figure 1-3 As shown, the deep learning-based differential diagnosis system for acute abdominal pain includes:

[0038] The data acquisition module is used to acquire the patient's initial self-report text data;

[0039] The standardization processing module is used to obtain a standardized symptom set based on the initial self-narrated text data through medical entity recognition and standardization processing.

[0040] The etiology reasoning module is used to obtain a set of suspected diseases by reasoning based on the standardized symptom set and the pre-constructed medical knowledge graph.

[0041] The contradiction detection module is used to obtain a contradiction set through contradiction detection processing based on the context of the standardized symptom set and the initial self-reported text data, and to calculate the contradiction intensity of each contradiction in the contradiction set and its correlation with the etiology of the suspected disease set; the contradiction set represents two or more propositions in the patient's description that are medically logically conflicting.

[0042] The priority calculation module is used to perform weighted calculations based on the intensity of the contradiction and the correlation of the etiology to obtain a contradiction significance score, and sort them in descending order to obtain a contradiction priority queue.

[0043] The conflict handling module is used to determine whether the conflict significance score of the first conflict in the conflict priority queue is lower than a preset significance threshold. If so, the standardized symptom set is output; otherwise, based on the first conflict, the suspected disease set, and the medical knowledge graph, the optimal clarification action is generated through knowledge graph traversal and benefit calculation. The optimal clarification action is executed to verify or resolve the current conflict.

[0044] The iterative update module is used to update the standardized symptom set based on the feedback data after performing the optimal clarification action, and to re-execute the severe etiology reasoning process and contradiction significance judgment until the termination condition is met.

[0045] The data acquisition module refers to the interface component used to collect the original text information of the patient's self-report. Specifically, it can be implemented using a text input interface in natural language processing. Its role is to ensure that the system can receive the initial description of the patient's complaint as the basis for subsequent processing.

[0046] The standardization processing module refers to the functional module that transforms the unstructured text of patient descriptions into a unified set of medical terms. Specifically, it can be implemented by using medical entity recognition algorithms and standardized terminology database matching technology. Its role is to eliminate semantic ambiguity in patient descriptions and provide structured input for subsequent reasoning.

[0047] The etiology reasoning module refers to the computational unit that infers the probability of a disease based on a standardized set of symptoms. Specifically, it can be implemented using knowledge graph embedding and graph neural network technology. Its role is to quickly generate a set of potential etiologies that match the current symptoms.

[0048] The contradiction detection module refers to the detection mechanism that identifies logical conflicts in patient descriptions. Specifically, it can be implemented using a rule engine based on medical logic and contextual semantic analysis technology. Its role is to proactively discover inconsistencies in input data and avoid reasoning based on incorrect premises.

[0049] The priority calculation module is a quantitative assessment unit that ranks the importance of detected contradictions. Specifically, it can be implemented using a weighted scoring algorithm and a similarity calculation model. Its role is to determine which contradictions pose the greatest threat to the current diagnosis and should be dealt with first.

[0050] The conflict resolution module is a decision-making unit that generates and executes clarification actions. Specifically, it can be implemented using knowledge graph traversal and information entropy optimization algorithms. Its role is to dynamically select the interactive actions that best reduce diagnostic uncertainty.

[0051] The iterative update module refers to the loop control unit that dynamically corrects the symptom set based on clarification feedback. Specifically, it can be implemented using a state snapshot saving and incremental update mechanism. Its role is to achieve self-correction and incremental optimization of the diagnostic process.

[0052] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0053] The data acquisition module obtains the patient's initial self-reported text data through the hospital's electronic medical record system or intelligent consultation system. For example, the patient describes experiencing severe abdominal pain last night, which lasted for 3 hours, but the pain score is only 3 points.

[0054] The standardization processing module uses a pre-trained medical named entity recognition model to process the initial self-report text, identify key medical entities such as severe abdominal pain, lasting for 3 hours, and pain score of 3, and map these entities to standardized medical terms to form a standardized symptom set.

[0055] The etiology reasoning module utilizes a pre-built medical knowledge graph to match symptoms in a standardized symptom set with disease nodes in the knowledge graph. Through graph traversal and similarity calculation, it obtains a set of suspected diseases with high relevance to the symptom set, such as acute appendicitis and gastrointestinal influenza.

[0056] The contradiction detection module analyzes the logical relationships between symptoms in a standardized symptom set and between symptoms and the initial text context, based on predefined medical logic rules. In this example, a logical contradiction was detected between severe abdominal pain and a pain score of 3, because severe pain usually corresponds to a higher pain score.

[0057] The priority calculation module quantifies and evaluates detected contradictions. It calculates the contradiction strength, such as determining the strength value based on the semantic distance between severity and a score of 3. Simultaneously, it calculates the correlation between the contradiction and the suspected disease set; for example, the degree of pain involved in the contradiction has a significant impact on the identification of acute appendicitis. Combining these two factors, a significance score is calculated for the contradiction, and the contradiction sets are ranked accordingly.

[0058] The conflict resolution module first determines whether the salience score of the first conflict in the queue exceeds a preset threshold. If it does, a clarification action needs to be generated. The module performs a breadth-first search in the knowledge graph, starting with pain intensity and pain score, to find medical operation nodes connecting these two concepts, such as asking for detailed pain characteristics. Combining this with a set of suspected diseases, the module generates the optimal clarification action, such as asking the patient to describe their specific pain experience in detail and reassessing the pain score.

[0059] The iterative update module performs clarification actions and obtains patient feedback. For example, if the patient clarifies that the pain is actually severe and the score should be 8, the module updates the standardized symptom set accordingly, correcting the pain score from 3 to 8. Then, the etiological reasoning and contradiction detection process is repeated until no significant contradictions exist or the preset iteration limit is reached.

[0060] Through the aforementioned approach, this application can proactively identify semantic contradictions in patient descriptions and verify and resolve these contradictions by dynamically generating and executing clarification actions. This iteratively optimized diagnostic process significantly improves the quality and consistency of input data, thereby enhancing the accuracy and reliability of disease inference results. Simultaneously, the system's interactive diagnostic process more closely resembles the diagnostic thinking of human physicians, improving the system's applicability and credibility in real-world clinical environments. Furthermore, by prioritizing contradictions highly correlated with suspected diseases, the system can more efficiently focus on key diagnostic information, accelerating the differential diagnosis process. This is of great significance for emergency scenarios requiring rapid decision-making, such as acute abdominal pain.

[0061] This application further proposes a process for obtaining a suspected disease set, which includes: calculating a weighted score for each symptom in the standardized symptom set and each disease in the knowledge base based on the association weights between symptoms and diseases stored in the medical knowledge graph; and selecting diseases whose scores exceed a preset disease score threshold or whose scores rank among the top predetermined number to form a suspected disease set.

[0062] The association weights between symptoms and diseases are quantified using predefined causal strength values ​​in a medical knowledge graph, derived from clinical research statistics or expert consensus. Weighted scoring employs a linear weighting model, summing the association weights of each symptom in the standardized symptom set to generate a total score for each disease. Disease scoring thresholds are determined based on the score distribution of real cases in historical diagnostic data, dynamically adjusted by setting percentile values. The predetermined number of differential diagnoses is set to 3 to 5, a range based on the number of differential diagnoses typically considered by physicians in emergency situations.

[0063] Specifically, when the standardized symptom set includes abdominal pain, fever, and elevated white blood cell count, the association weight between abdominal pain and appendicitis in the medical knowledge graph is 0.8, and the association weight with gastroenteritis is 0.6; the association weight between fever and appendicitis is 0.7, and the association weight with gastroenteritis is 0.5. After weighted scoring, the score for appendicitis is 0.8 + 0.7 = 1.5, and the score for gastroenteritis is 0.6 + 0.5 = 1.1. If the preset disease scoring threshold is 1.2, appendicitis is included in the suspected disease set; if the predetermined number is set to 2, the top two scoring diseases are selected. This method effectively focuses on high-probability diseases and reduces interference from low-relevance diseases through a quantitative screening mechanism.

[0064] Through the above technical solution, this application can quickly screen out the most relevant suspected diseases based on the patient's standardized symptom set. This improves the targeting and efficiency of subsequent etiological reasoning and contradiction detection, avoiding ineffective calculations for a large number of irrelevant diseases. At the same time, by setting scoring thresholds or quantity limits, the size of the suspected disease set is ensured to be moderate, including the main candidate causes without affecting system performance due to an excessive number of candidates.

[0065] This application further proposes a method for calculating the significance score of contradiction using the following formula:

[0066]

[0067] in, This represents the conflict intensity value determined according to predefined rules. Indicates the degree of correlation with the cause. , These are preset weighting coefficients;

[0068] Etiological correlation is determined by calculating the Jaccard similarity between the union of the disease sets associated with each of the two contradictory propositions and the suspected disease set.

[0069] The contradiction intensity value is graded according to the severity of the medical logical conflict. For example, when the symptom of no nausea conflicts with the typical symptom of intestinal obstruction, the contradiction intensity value is set to a high level. Etiological relevance is calculated through the following steps: extract the disease sets corresponding to the two propositions involved in the contradiction, calculate the intersection of their union with the current suspected disease set, divide the size of the union by the size of the intersection, and obtain the Jaccard similarity. The weighting coefficients are adjusted by preset values ​​to determine the contribution ratio of contradiction intensity to causal relevance; for example, when α=0.6 and β=0.4, the severity of the contradiction itself is emphasized.

[0070] Specifically, when a patient is detected to have both upper abdominal pain and right lower quadrant rebound tenderness, a higher contradiction strength value is assigned based on the anatomical conflict rules of abdominal signs. Subsequently, the system retrieves disease sets associated with upper abdominal pain (e.g., gastritis, pancreatitis) and disease sets associated with right lower quadrant rebound tenderness (e.g., appendicitis), and calculates the Jaccard similarity between their union and the current suspected disease set. If the current suspected disease set includes appendicitis and the Jaccard similarity is high, the etiological relevance will increase. Finally, a weighted contradiction significance score is calculated; a higher score indicates a greater interference with the current diagnosis and requires priority processing. This method ensures that the system prioritizes resolving contradictions that are highly relevant to the suspected etiology and have significant medical logical conflicts, thereby improving diagnostic iteration efficiency.

[0071] Through the above technical solution, this application achieves a quantitative assessment of contradictions in patient descriptions. The contradiction significance score comprehensively considers the inherent strength of the contradiction and its relevance to current diagnostic reasoning, providing a scientific priority basis for subsequent contradiction handling. This method can effectively identify and highlight key contradictions that have a significant impact on the diagnostic process, improving the efficiency and accuracy of the system in handling contradictions. Simultaneously, by introducing etiological correlation, the system can dynamically adjust the importance of contradictions, making the contradiction handling process more aligned with actual diagnostic needs, thereby improving the overall diagnostic accuracy and reliability.

[0072] This application further proposes a process for generating the optimal clarification action, which includes: analyzing the primary contradiction to obtain two conflicting medical concepts; using these two medical concepts as nodes to perform a breadth-first path search in the medical knowledge graph to obtain a set of public medical operation nodes; calculating the intersection of this set with a set of diagnostic operation nodes that are highly related to the set of suspected diseases to generate a candidate set of clarification actions; calculating the comprehensive benefit value for each action in the candidate set and selecting the one with the highest value as the optimal clarification action.

[0073] The process of resolving contradictions employs natural language processing (NLP) technology to identify conflicting medical entities in the text. For example, severe abdominal pain and a pain score of 3 are resolved as a contradiction in the pain intensity dimension. Breadth-first path search sets a maximum search depth of 3 layers to ensure the path length remains within an operational range. The set of public medical operation nodes includes intermediate diagnostic and treatment steps connecting two conflicting concepts; for example, an abdominal ultrasound examination can be associated with both appendicitis and intestinal obstruction. Intersection calculation uses the Jaccard coefficient to filter medical operation nodes with a correlation greater than 0.7 with the currently suspected disease. The comprehensive benefit value calculation incorporates information entropy theory, quantifying the value of actions by measuring the reduction in diagnostic entropy.

[0074] Specifically, when the system detects a contradiction between vomiting occurring before abdominal pain and the typical symptom sequence of acute appendicitis, it first extracts the vomiting and abdominal pain sequences as conflict nodes. A search of the medical knowledge graph reveals that both gastroscopy and serum amylase testing can connect to these two nodes. Combining this with the current suspected disease set (including pancreatitis and intestinal obstruction), serum amylase testing is selected as having a correlation of 0.85 with the disease set. When calculating the benefit value of each candidate action, serum amylase testing best distinguishes pancreatitis from other diseases, with a 1.2-bit decrease in diagnostic entropy after its execution, and is ultimately selected as the optimal clarification action. This process, through path constraints and disease relevance filtering, effectively eliminates irrelevant examination items (such as electrocardiograms), reducing the number of candidate actions from the original 58 to 3, thus improving decision-making efficiency.

[0075] Through the aforementioned technical solution, this application can intelligently generate optimal clarification actions, effectively resolving contradictory information in patient descriptions. This method not only improves the accuracy and reliability of the diagnostic system but also simulates the critical thinking and proactive interaction capabilities of human doctors, enabling the system to dynamically collect and verify key information, thereby optimizing the differential diagnosis path. Furthermore, by utilizing medical knowledge graphs and suspected disease sets, the system can generate clarification actions highly relevant to the current diagnostic context, avoiding unnecessary examinations and inquiries, and improving diagnostic efficiency.

[0076] In some of the solutions described above in this application, after generating a candidate set of clarifying actions, the optimal action needs to be selected from multiple candidate actions. However, there is a lack of a quantitative evaluation mechanism, which cannot effectively measure the contribution of different actions to reducing diagnostic uncertainty. This leads to blind selection of actions, which may increase unnecessary examination costs or delay the acquisition of key diagnostic information.

[0077] This application further proposes the following formula for calculating the comprehensive benefit value:

[0078]

[0079] in, The probability distribution of each disease in the current suspected disease set can be calculated using the Shannon entropy formula. The composition is based on the possible output results of candidate action 'a' in the medical knowledge graph. For example, when 'a' is an abdominal ultrasound examination, Results may include finding an enlarged appendix without any abnormalities. The occurrence frequency of result 'o' after action 'a' in historical medical data is determined by analyzing the frequency of occurrence of result 'o'. For example, if result 'o' occurs 30 times out of 100 instances of action 'a', then... =0.3. The calculation needs to update the disease probability according to Bayes' theorem, for example, when the result When disease A is supported, the probability of disease A increases, the probabilities of other diseases are adjusted accordingly, and the entropy value is calculated based on the updated distribution.

[0080] Specifically, when calculating the overall benefit value, the initial entropy is first calculated based on the current disease probability distribution to reflect the degree of diagnostic uncertainty. For candidate action 'a', all possible output results are iterated. Calculate the probability of each outcome. and the corresponding updated entropy .Will and The summation of the products of the initial entropy and the expected residual entropy yields the expected residual entropy after action a is performed. Ultimately, U(a) equals the difference between the initial entropy and the expected residual entropy; a larger difference indicates a more significant contribution of action a to reducing uncertainty. For example, when candidate actions include a blood routine examination and a CT scan, if U(a) = 1.2 for the former and U(a) = 2.5 for the latter, then the CT scan is preferentially selected as the optimal clarifying action. This calculation mechanism ensures that the selected action can eliminate diagnostic ambiguity to the greatest extent possible, while avoiding inefficient or redundant medical procedures, by quantifying the expected information gain.

[0081] Through the aforementioned technical solution, this application can quantitatively assess the expected benefits of different clarification actions, thereby selecting the action most helpful in reducing diagnostic uncertainty. This information entropy-based decision-making mechanism enables the system to guide the diagnostic process more intelligently, improving diagnostic efficiency and accuracy. Simultaneously, by considering the possible outcomes and probabilities of action execution, the system can balance benefits and costs, avoiding over-examination and optimizing the utilization of medical resources.

[0082] This application further proposes a method to determine whether the optimal clarification action is an immediate feedback action; if so, the structured feedback data generated after the action is executed is collected immediately, and the standardized symptom set is directly updated using the data; otherwise, if it is a delayed feedback action, the current processing flow is paused and a snapshot of the current state is saved. After the delayed feedback data arrives, the state snapshot is loaded and the standardized symptom set is updated using the delayed feedback data.

[0083] Specifically, the action type is determined by querying a predefined action attribute table, which stores the feedback timeliness tags corresponding to each clarification action; structured feedback data is collected in real time through a preset interface protocol, and its fields are mapped to the corresponding attributes of the standardized symptom set; the status snapshot includes the version identifier of the standardized symptom set, the sorting status of the contradiction priority queue, and the context cache of the medical knowledge graph; when loading the snapshot, the consistency between the snapshot and the current system state is verified to ensure the logical coherence of data updates.

[0084] Specifically, after the system generates the optimal clarification action, it first determines the timeliness of the feedback based on a predefined action type library. For immediate feedback actions, such as blood pressure measurement or pain score review, the system immediately calls the data acquisition interface to obtain structured results, for example, mapping systolic blood pressure (_140 mmHg) to the blood pressure attribute of the standardized symptom set. For delayed feedback actions, such as pathological biopsy or blood culture, the system packages the hash value of the current standardized symptom set, the sorting index of the contradiction queue, and the node access records of the knowledge graph into a snapshot file, stores it in a temporary storage area, and suspends the current thread. When delayed feedback data arrives, the system verifies data integrity using the hash value in the snapshot, restores the context cache of the knowledge graph, and updates the inflammation-related symptom field of the standardized symptom set with key indicators from the feedback data, such as abnormal white blood cell counts. This mechanism separates the processing paths for immediate and delayed actions, avoiding system blockage due to waiting for external examination results. Simultaneously, by preserving the intermediate logic of diagnostic reasoning through state snapshots, it ensures that the system can restart the reasoning process based on a consistent context after delayed data updates.

[0085] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0086] After performing the optimal clarification action, the system first determines whether the action is an immediate feedback action. Immediate feedback actions include, but are not limited to, asking the patient for a pain score or requesting the patient to specify the location of the pain. If it is an immediate feedback action, the system immediately collects the structured feedback data generated after performing the action. For example, the patient answers that the pain score is 7, or points out that the pain is located in the right lower abdomen. The system then uses this newly acquired data to directly update the standardized symptom set, such as updating abdominal pain to right lower abdominal pain with a pain score of 7.

[0087] If the optimal clarification action is not an immediate feedback action but a delayed feedback action, such as suggesting a complete blood count or CT scan, the system pauses the current processing flow. The system saves a snapshot of the current state, including but not limited to the current standardized symptom set, suspected disease set, and conflict priority queue. Once the delayed feedback data arrives, such as the complete blood count result or CT scan report uploaded to the system, the system loads the previously saved state snapshot and updates the standardized symptom set using the newly arrived delayed feedback data. For example, abdominal pain is updated to right lower quadrant pain, accompanied by an elevated white blood cell count or right lower quadrant pain, and the CT scan shows an enlarged appendix.

[0088] Through the above technical solution, this application achieves flexible handling of both immediate and delayed feedback scenarios. For immediate feedback, the system can quickly integrate new information, improving diagnostic efficiency. For delayed feedback, the system ensures the continuity and consistency of the diagnostic process through state saving and recovery mechanisms, avoiding information loss or duplicate data collection. This dynamic update mechanism enables the system to continuously optimize the accuracy and completeness of symptom descriptions, thereby providing more reliable basic data for subsequent etiological reasoning and improving overall diagnostic accuracy.

[0089] This application further proposes that after updating the standardized symptom set based on feedback data and before re-executing the etiological reasoning process, it also includes an evaluation of the conflict resolution effect: based on the updated standardized symptom set, the conflict significance score of the original primary conflict is recalculated; it is determined whether the recalculated conflict significance score is lower than a preset significance threshold; if so, the conflict is considered resolved and removed from the conflict priority queue; otherwise, based on the recalculated conflict significance score, the conflict is reinserted into the corresponding sorting position in the conflict priority queue.

[0090] The conflict resolution effectiveness evaluation step comprises three sub-steps. First, the conflict salience score calculation module is invoked, inputting the updated standardized symptom set and the current suspected disease set. The conflict intensity value and etiological correlation of the original primary conflict are recalculated, generating a new conflict salience score. Second, the new score is compared with a preset threshold. If it is lower than the threshold, a conflict removal operation is triggered, deleting the conflict entity from the queue. Finally, if the new score is still higher than the threshold, the conflict entity is reordered according to the latest score and inserted into the queue, ensuring the queue dynamically reflects the current conflict status. This step shares the conflict salience score calculation logic with the priority calculation module but uses updated input data, forming a closed-loop verification mechanism.

[0091] Specifically, once the iterative update module completes the update of the standardized symptom set, the conflict resolution effectiveness evaluation step is immediately initiated. The system uses the original primary conflict as the evaluation object, recalculates the conflict intensity and etiological correlation using the latest data, and generates a new conflict significance score. If the new score is lower than a preset threshold, it indicates that the conflict has lost its significance after the data update, and the system automatically removes it from the queue to avoid subsequent invalid processing. If the new score is still higher than the threshold, the system redetermines the conflict's ranking position in the queue based on the latest score, ensuring that the queue priority strictly corresponds to the current diagnostic status. This mechanism dynamically verifies the conflict resolution effectiveness, eliminates residual conflict interference caused by incomplete data updates, and improves system iteration efficiency by optimizing the clarification action generation strategy through real-time queue sorting.

[0092] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0093] After updating the standardized symptom set based on feedback data and before re-executing the etiological reasoning process, the evaluation of the conflict resolution effectiveness is also included:

[0094] Based on the updated standardized symptom set, the contradiction significance score of the original primary contradiction is recalculated. Specifically, the updated symptom information can be used to re-execute the processing flow of the contradiction detection module to obtain a new contradiction significance score.

[0095] Determine whether the recalculated contradiction significance score is lower than a preset significance threshold. For example, the new contradiction significance score can be compared with a system-preset significance threshold of 0.5.

[0096] If a new contradiction's significance score is below the significance threshold, the contradiction is considered resolved and removed from the contradiction priority queue. Specifically, this can be achieved by deleting the contradiction object from the queue data structure.

[0097] If the new contradiction significance score is still higher than the significance threshold, then based on the recalculated contradiction significance score, the contradiction is re-inserted into its corresponding sorting position in the contradiction priority queue. The insertion position can be quickly located using algorithms such as binary search.

[0098] Through the above technical solution, this application can dynamically evaluate the effectiveness of conflict resolution and adjust the priority of conflict handling in a timely manner. For resolved conflicts, the system can concentrate resources on other unresolved conflicts; for conflicts that are not fully resolved, the system can continue to process them based on their new salience. This mechanism improves the efficiency of the system in handling conflicts, avoids the duplication of processing resolved problems, and ensures that important conflicts receive continuous attention.

[0099] This application further proposes: based on the symptom occurrence time sequence parsed from the initial self-narrated text data, retrieve the typical symptom development time sequence of suspected diseases from the medical knowledge graph, calculate the similarity between the two, and if the similarity is lower than the preset time sequence similarity threshold, generate a time sequence logical contradiction and add it to the contradiction set.

[0100] The symptom occurrence time sequence is constructed by extracting symptom entities and their occurrence times from the patient's written text using natural language processing technology; the typical symptom development time sequence is obtained from the symptom development pattern attributes of disease nodes in the medical knowledge graph; the similarity calculation adopts the dynamic time warping algorithm, which allows two time sequences of unequal length to be aligned and the minimum path distance to be calculated; the preset time similarity threshold is dynamically adjusted according to the disease type, for example, the threshold is set to 0.65 for acute appendicitis and 0.7 for intestinal obstruction.

[0101] Specifically, when a patient describes experiencing upper abdominal pain initially, which shifts to the right lower quadrant two hours later and is accompanied by fever, the system parses the symptoms into a temporal sequence [upper abdominal pain, right lower quadrant pain, fever]. The typical temporal sequence for acute appendicitis retrieved from the medical knowledge graph is [periumbilical pain, right lower quadrant pain, nausea, fever]. Using a dynamic time warping algorithm, the similarity between the two is calculated to be 0.58. Since this value is below the preset threshold of 0.65 for acute appendicitis, the system generates a temporal logic contradiction. This contradiction triggers the subsequent clarification action generation module, which prioritizes performing follow-up questions to confirm the order of pain shift and fever onset. By verifying the patient's actual symptom temporal sequence, the system updates the standardized symptom set, eliminating diagnostic bias caused by temporal misrepresentation, and improving the accuracy of the suspected disease set output by the etiology reasoning module.

[0102] As a preferred embodiment, the solution of this application is implemented as follows: When parsing the symptom occurrence time sequence from the patient's self-reported text, a method combining regular expression matching based on time expressions and an event sequence model is adopted. For example, when a patient describes experiencing upper abdominal pain starting at 6:00 AM, shifting to the right lower abdomen at 8:00 AM, and vomiting at 10:00 AM, the system will extract three time nodes and their corresponding symptoms, constructing a time sequence [upper abdominal pain, right lower abdominal pain, vomiting]. The typical symptom development time sequence of acute appendicitis is retrieved from the medical knowledge graph as [periumbilical pain, right lower abdominal pain, nausea and vomiting], and the similarity score between the two sequences is calculated using a dynamic time warping algorithm. When the similarity is lower than a preset threshold of 0.7, a time sequence logical contradiction object is generated, which contains the deviation position information between the patient's actual time sequence and the standard disease time sequence, and this contradiction is added to the contradiction set.

[0103] Through the above technical solution, this application effectively solves the problem of misdiagnosis caused by the mismatch between the patient's symptom development process and the typical disease course pattern. By quantitatively assessing the medical logical consistency of the time sequence, it can promptly identify contradictory phenomena that violate pathophysiological laws, such as vomiting followed by abdominal pain. This prompts the system to prioritize the handling of such contradictions, avoiding the inclusion of causes that do not conform to the disease development pattern into the suspected disease set, thereby improving the clinical rationality of the differential diagnosis results.

[0104] This application further proposes that after generating the candidate set of clarification actions and before calculating the comprehensive benefit value, the type characteristics of the primary contradiction are analyzed, and based on the type characteristics, a subset of atomic actions matching the type characteristics is selected from a predefined action type-atomic action mapping table; the intersection of the candidate set of clarification actions and the subset of atomic actions is taken to narrow down the range of candidate actions for subsequent comprehensive benefit evaluation.

[0105] As a preferred embodiment, the solution of this application is implemented as follows: After searching the medical knowledge graph to obtain a set of public medical operation nodes, a candidate set of clarification actions is generated, including three types of operations: imaging examination, laboratory testing, and targeted inquiry. At this time, the contradiction detection module identifies the first contradiction as belonging to the symptom temporal conflict type, specifically manifested as the patient's description of the pain location shifting order not matching the typical course of acute pancreatitis. According to the predefined action type-atomic action mapping table, the symptom temporal conflict type is associated with the targeted inquiry type of operation, which includes three atomic actions: inquiring about the symptom start location, confirming the symptom onset time, and verifying the symptom duration. By taking the intersection of the candidate set of clarification actions and the subset of atomic actions, the verification of symptom duration is finally selected as the candidate action, excluding the two irrelevant operations of imaging examination and laboratory testing, thereby reducing the calculation scope of the subsequent comprehensive benefit evaluation.

[0106] Through the above technical solution, this application can dynamically adjust the action selection strategy according to the type of contradiction, significantly reducing computational complexity while ensuring the effectiveness of clarification actions. By using a predefined mapping mechanism between type features and atomic actions, it avoids misselecting check-type operations in time-series conflict scenarios, ensuring that the clarification actions generated by the system are highly consistent with the essence of the contradiction, thereby shortening the iterative processing cycle and improving the efficiency of contradiction resolution.

[0107] In some of the solutions mentioned above in this application, the iterative update module optimizes the diagnostic results by repeatedly performing contradiction handling and symptom set updates, but lacks a clear termination mechanism, which may lead to the processing flow falling into an infinite loop or terminating prematurely, affecting the system's operating efficiency and diagnostic reliability.

[0108] This application further proposes termination conditions including any of the following: the contradiction priority queue becomes empty; the contradiction significance score of the first contradiction in the contradiction priority queue is lower than a preset significance threshold; the number of iterations reaches a preset maximum number of iterations threshold.

[0109] Specifically, an empty conflict priority queue indicates that all conflicts have been processed or removed; a conflict significance score below a threshold indicates that the current conflict's impact on the diagnostic result is negligible; and the maximum iteration threshold limits the resource consumption of the processing flow through a preset value. These three conditions form complementary logic, ensuring that the system completes conflict processing within a reasonable range.

[0110] Specifically, when the system performs iterative updates, it monitors the status of the conflict priority queue, the salience score of the first conflict, and the number of iterations in real time. If the queue is empty, the process is immediately terminated and the current standardized symptom set is output. If the score of the first conflict is below a threshold, it is determined that the current conflict does not need to be processed, and the process is terminated. If the number of iterations reaches the upper limit, it is forcibly terminated to avoid wasting resources. For example, if the maximum number of iterations is set to 5, and the system still fails to meet the first two conditions after 5 consecutive rounds of conflict processing, it will automatically terminate. These three conditions work together to avoid invalid loops and prevent key conflicts from being left unprocessed, ensuring the stability and timeliness of the system's output.

[0111] Through the above technical solution, this application can effectively control the iterative processing flow of the system to terminate within a reasonable range, avoiding resource waste caused by infinite loops. When the patient's self-report contains complex contradictions, this solution ensures that the system stops promptly when the contradiction processing reaches a preset safety boundary, thereby achieving a balance between diagnostic efficiency and accuracy. When contradictions cannot be completely resolved through a limited number of clarification actions, this solution prioritizes outputting the current optimal standardized symptom set, providing reliable data support for clinical decision-making.

[0112] Example 2:

[0113] like Figure 4 As shown, the deep learning-based method for differential diagnosis of acute abdominal pain includes:

[0114] Obtain the patient's initial self-report text data;

[0115] Based on the initial self-reported text data, a standardized symptom set is obtained through medical entity recognition and standardization processing;

[0116] Based on the standardized symptom set and the pre-constructed medical knowledge graph, a set of suspected diseases is obtained through etiological reasoning.

[0117] Based on the context of the standardized symptom set and the initial self-reported text data, a contradiction set is obtained through contradiction detection processing, and the contradiction intensity of each contradiction in the contradiction set and its correlation with the etiology of the suspected disease set are calculated; the contradiction set represents two or more propositions in the patient's description that are medically logically conflicting.

[0118] The significance score of the contradiction is obtained by weighting the contradiction intensity and the correlation between the cause and effect, and then sorted in descending order to obtain the contradiction priority queue.

[0119] If the significance score of the first contradiction in the contradiction priority queue is lower than a preset significance threshold, the standardized symptom set is output; otherwise, based on the first contradiction, the suspected disease set, and the medical knowledge graph, an optimal clarification action is generated through knowledge graph traversal and benefit calculation; the optimal clarification action is executed to verify or resolve the current contradiction.

[0120] The standardized symptom set is updated based on the feedback data after the optimal clarification action is performed, and the etiological reasoning and contradiction significance judgment are re-executed until the termination condition is met.

[0121] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

[0122] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0123] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A deep learning-based system for differential diagnosis of the etiology of acute abdominal pain, characterized in that, include: The data acquisition module is used to acquire the patient's initial self-report text data; The standardization processing module is used to obtain a standardized symptom set based on the initial self-narrated text data through medical entity recognition and standardization processing. The etiology reasoning module is used to obtain a set of suspected diseases by reasoning based on the standardized symptom set and the pre-constructed medical knowledge graph. The contradiction detection module is used to obtain a contradiction set through contradiction detection processing based on the context of the standardized symptom set and the initial self-reported text data, and to calculate the contradiction intensity of each contradiction in the contradiction set and its correlation with the etiology of the suspected disease set. The set of contradictions represents two or more propositions in the patient's description that are logically conflicting in medical terms; The priority calculation module is used to perform weighted calculations based on the intensity of the contradiction and the correlation of the etiology to obtain a contradiction significance score, and sort them in descending order to obtain a contradiction priority queue. The conflict handling module is used to determine whether the conflict significance score of the first conflict in the conflict priority queue is lower than a preset significance threshold. If so, the standardized symptom set is output; otherwise, based on the first conflict, the suspected disease set, and the medical knowledge graph, the optimal clarification action is generated through knowledge graph traversal and benefit calculation. The optimal clarification action is executed to verify or resolve the current conflict. The iterative update module is used to update the standardized symptom set based on the feedback data after performing the optimal clarification action, and to re-execute the severe etiology reasoning process and contradiction significance judgment until the termination condition is met.

2. The deep learning-based differential diagnosis system for acute abdominal pain according to claim 1, characterized in that: The process of obtaining the suspected disease set includes: Based on the association weights between symptoms and diseases stored in the medical knowledge graph, a weighted score is calculated for each symptom in the standardized symptom set and each disease in the knowledge base. The suspected disease set is formed by selecting diseases whose scores exceed a preset disease score threshold or whose scores rank among the top predetermined number.

3. The deep learning-based differential diagnosis system for acute abdominal pain according to claim 1, characterized in that: The specific method for calculating the significance score of the contradiction is as follows: in, This represents the conflict intensity value determined according to predefined rules. Indicates the degree of correlation with the cause. , These are preset weighting coefficients; The etiological correlation is determined by calculating the Jaccard similarity between the union of the disease sets associated with each of the two contradictory propositions and the suspected disease set.

4. The deep learning-based differential diagnosis system for acute abdominal pain according to claim 1, characterized in that: The process of generating the optimal clarification action includes: Analyzing the first contradiction reveals two conflicting medical concepts it contains; Using the two conflicting medical concepts as nodes, a breadth-first path search is performed in the medical knowledge graph to obtain a set of common medical operation nodes connecting the two medical concepts. Calculate the intersection of the public medical operation node set and the diagnostic operation node set that is highly related to the suspected disease set to generate a clarification action candidate set; For each candidate action in the clarification action candidate set, calculate its comprehensive benefit value, and select the candidate action with the highest comprehensive benefit value as the optimal clarification action.

5. The deep learning-based differential diagnosis system for acute abdominal pain according to claim 4, characterized in that: The formula for calculating the comprehensive benefit value is as follows: in, This represents the overall benefit value of candidate clarification action a; The diagnostic entropy before the action is performed is calculated using the probability distribution of each disease in the current suspected disease set. This represents the set of all possible output results after executing candidate clarification action a; This indicates that the result is observed given that action a has been performed. The likelihood probability is obtained from the pre-stored historical medical data statistical model or the causal association strength defined in the medical knowledge graph; This indicates the result observed after action a is performed. Diagnostic entropy at time, based on results The updated disease probability distribution was calculated.

6. The deep learning-based differential diagnosis system for acute abdominal pain according to claim 1, characterized in that: The step of updating the standardized symptom set based on feedback data after performing the optimal clarification action specifically includes: Determine whether the optimal clarification action is an immediate feedback action; If so, the structured feedback data generated after the action is performed is collected immediately, and the standardized symptom set is updated directly using the data; Otherwise, it is a delayed feedback action. The current processing flow is paused and a snapshot of the current state is saved. After the delayed feedback data arrives, the state snapshot is loaded and the standardized symptom set is updated using the delayed feedback data.

7. The deep learning-based differential diagnosis system for acute abdominal pain according to claim 1, characterized in that: After updating the standardized symptom set based on feedback data and before re-executing the etiological reasoning process, the process also includes an evaluation of the conflict resolution effectiveness: Based on the updated standardized symptom set, the contradiction significance score of the original primary contradiction is recalculated; Determine whether the recalculated contradiction significance score is lower than the preset significance threshold; If so, the contradiction is considered resolved and removed from the contradiction priority queue; Otherwise, based on the recalculated contradiction significance score, the contradiction is re-inserted into the corresponding sorting position in the contradiction priority queue.

8. The deep learning-based differential diagnosis system for acute abdominal pain according to claim 1, characterized in that: The step of obtaining a contradiction set based on the context of the standardized symptom set and the initial self-reported text data through contradiction detection processing based on predefined rules further includes: The symptom occurrence sequence described by the patient was parsed from the initial self-reported text data; Retrieve the time sequence of typical symptom development of the suspected diseases from the medical knowledge graph; Calculate the similarity between the time sequence of symptom onset described by the patient and the time sequence of typical symptom development of the disease; Determine whether the similarity is lower than a preset temporal similarity threshold: if yes, generate a temporal logical contradiction and add it to the contradiction set.

9. The deep learning-based differential diagnosis system for acute abdominal pain according to claim 4, characterized in that: After generating the candidate set of clarification actions but before calculating the overall benefit value, action space optimization is also included: Analyze the type characteristics of the first contradiction, and based on the type characteristics, filter out a subset of atomic actions that match the type characteristics from a predefined action type-atomic action mapping table; The intersection of the candidate set of clarification actions and the subset of atomic actions is taken to narrow down the range of candidate actions for subsequent comprehensive benefit evaluation.

10. A deep learning-based method for differential diagnosis of the etiology of acute abdominal pain, applied to the deep learning-based system for differential diagnosis of the etiology of acute abdominal pain as described in any one of claims 1-9, characterized in that, Includes the following steps: Obtain the patient's initial self-report text data; Based on the initial self-reported text data, a standardized symptom set is obtained through medical entity recognition and standardization processing; Based on the standardized symptom set and the pre-constructed medical knowledge graph, a set of suspected diseases is obtained through etiological reasoning. Based on the context of the standardized symptom set and the initial self-reported text data, a contradiction set is obtained through contradiction detection processing, and the contradiction intensity of each contradiction in the contradiction set and its correlation with the etiology of the suspected disease set are calculated. The set of contradictions represents two or more propositions in the patient's description that are logically conflicting in medical terms; The significance score of the contradiction is obtained by weighting the contradiction intensity and the correlation between the cause and effect, and then sorted in descending order to obtain the contradiction priority queue. If the significance score of the first contradiction in the contradiction priority queue is lower than a preset significance threshold, the standardized symptom set is output; otherwise, based on the first contradiction, the suspected disease set, and the medical knowledge graph, an optimal clarification action is generated through knowledge graph traversal and benefit calculation; the optimal clarification action is executed to verify or resolve the current contradiction. The standardized symptom set is updated based on the feedback data after the optimal clarification action is performed, and the etiological reasoning and contradiction significance judgment are re-executed until the termination condition is met.

Images