Medical data association rule confidence coefficient dynamic updating system

Abstract

The invention discloses a medical data association rule confidence degree dynamic updating system, and relates to the field of machine learning. The system comprises a rule monitoring module, a confidence coefficient updating module and a rule base management module, wherein the rule monitoring module obtains a symptom set and a labeling result and retrieves a matched target association rule in a rule base; the confidence coefficient updating module determines a time sequence attenuation parameter according to the disease category, and performs attenuation calculation on the pre-stored confidence coefficient in combination with a time interval from the last verification timestamp to obtain a basic confidence coefficient; if the labeling result is consistent with the rule, judging that the evidence is a forward evidence, updating the timestamp, and improving the confidence coefficient by using a forward enhancement strategy; if not, negative evidence is judged, counting is accumulated, and when the counting reaches a threshold value in a preset time window, the confidence coefficient is reduced based on a negative punishment strategy; and the rule base management module writes the updated new confidence into the rule base. By implementing the technical scheme provided by the invention, the accuracy of medical diagnosis is improved.

Description

Technical Field

[0001] This application relates to the field of machine learning, and more particularly to a dynamic update system for the confidence of association rules in medical data. Background Technology

[0002] Currently, with the continuous improvement of medical informatization, medical institutions have accumulated massive amounts of data, including electronic medical records, medical images, and treatment logs. Utilizing data mining techniques to deeply analyze this multi-source, heterogeneous medical big data, uncovering potential correlations between symptoms and diseases, and building intelligent clinical auxiliary diagnostic systems has become an important means to improve the efficiency of medical services and reduce the rate of misdiagnosis and missed diagnosis.

[0003] In related technologies, association rule mining algorithms are typically used to construct disease diagnosis models. First, historical medical record data is preprocessed, extracting symptom sets and diagnosed diseases as transaction items. Next, association rules are calculated by statistically analyzing the frequency of co-occurrence of different itemsets in the database. Finally, rules that meet a minimum threshold are stored in a static knowledge base. When a new case is input, the rule base is matched to the patient's symptoms, and possible diagnostic results are ranked and recommended based on pre-calculated confidence scores. The rule base is updated periodically through full re-mining or incremental cumulative counting.

[0004] However, these technologies have limitations when dealing with disease scenarios where symptoms are similar but incidence rates differ significantly. Common diseases have accumulated massive sample sizes in historical databases, resulting in highly stable rule confidence levels. When similar symptoms appear in a specific period but belong to a different type of disease, even if the disease has recently appeared in concentrated outbreaks, the sample size is negligible compared to the historically accumulated sample size of common diseases. This leads to a tendency to output historically high-probability common diseases when calculating recommendation rankings, resulting in insufficient accuracy in assisted diagnosis. Summary of the Invention

[0005] This application provides a dynamic update system for the confidence of medical data association rules, which can be used to improve the accuracy of medical diagnosis.

[0006] The first aspect of this application provides a dynamic update system for the confidence level of medical data association rules, including a rule monitoring module, a confidence update module, and a rule base management module, wherein: The rule monitoring module acquires the symptom set and the identified annotation results, and retrieves the target association rule matching the symptom set from the medical diagnosis association rule base. The confidence update module acquires the disease category corresponding to the target association rule, determines the corresponding time decay parameter based on the disease category, calculates the time interval between the current time and the last verification timestamp of the target association rule record, and uses the time decay parameter and time interval to perform time decay calculation on the pre-stored confidence of the target association rule to obtain the base confidence. The confidence update module also determines whether the identified annotation result is consistent with the disease corresponding to the target association rule; if so, it is considered positive evidence. If the evidence is positive, it is considered negative evidence. The confidence update module is also used to update the last verification timestamp to the current time when the evidence is considered positive, and to increase the value of the basic confidence based on the positive reinforcement strategy to obtain the updated new confidence. The confidence update module is also used to perform an incremental operation on the negative evidence counter associated with the target association rule when the evidence is considered negative, and to reduce the value of the basic confidence based on the negative penalty strategy when the count value of the negative evidence counter reaches the preset trigger threshold within the preset time window to obtain the updated new confidence. The rule base management module is used to write the new confidence into the medical diagnosis association rule base.

[0007] In the above embodiments, by using a time-series decay mechanism and a dynamic reward and punishment strategy based on clinical evidence, the weight of long-term unverified rules is reduced, and the confidence of rules is adjusted in real time according to actual diagnosis and treatment results. This enables the medical diagnosis association rule base to have the ability to self-evolve and weed out the inferior, promptly eliminating outdated or erroneous rules, and ensuring that the rule base always keeps up with the evolution of diseases and the development of medical practice, thereby improving the accuracy of medical diagnosis.

[0008] In conjunction with some embodiments of the first aspect, in some embodiments, the confidence update module specifically includes an error trend analysis unit and a dynamic penalty execution unit, wherein: The error trend analysis unit is used to acquire the timestamp sequence of all negative evidence within a preset time window when the count value of the negative evidence counter reaches a preset trigger threshold, and calculate the average time interval between adjacent negative evidence, generating an error density index based on the average time interval. The error trend analysis unit is also used to calculate the difference ratio between the current error density index and the historical error density index at the time of the last trigger penalty, in order to determine the error acceleration parameter of negative evidence generation. The dynamic penalty execution unit is used to nonlinearly amplify the preset benchmark penalty coefficient based on the error acceleration parameter to obtain the amplified penalty coefficient. The dynamic penalty execution unit is also used to use the amplified penalty coefficient to perform weighted deduction on the basic confidence level when the error acceleration parameter is positive and exceeds the preset warning threshold, to obtain a new confidence level.

[0009] In the above embodiments, the dynamic trend of rule failure is deeply quantified by the error acceleration parameter, the risk of a sharp deterioration in rule quality is identified, strong suppression is implemented in the early stage of error outbreak, the elimination process of seriously failed rules is accelerated, the interference window of erroneous knowledge on clinical decision-making is shortened, thereby ensuring that the rules are validated at high frequency and improving the accuracy of assisted diagnosis.

[0010] In conjunction with some embodiments of the first aspect, in some embodiments, the confidence update module further includes a resilience recovery and circuit breaker protection unit, wherein: The elastic recovery and circuit breaker protection unit is used to start the observation window for the target association rule before the dynamic penalty execution unit uses the amplified penalty coefficient to perform weighted deduction on the basic confidence score. The elastic recovery and circuit breaker protection unit is also used to determine that negative evidence is short-term fluctuation interference if the number of positive evidence received continuously within the observation window exceeds the preset recovery threshold. The elastic recovery and circuit breaker protection unit is also used to generate a new confidence score by using the weighted average of the basic confidence score and the pre-stored confidence score when it is determined to be short-term fluctuation interference, and then terminate the observation window.

[0011] In the above embodiments, by introducing an observation window and a circuit breaker mechanism before imposing severe penalties, a distinction is made between substantial rule failures and short-term random fluctuations at the data level. When a rule is detected to quickly recover its positive performance after fluctuations, excessive penalties are intercepted and a smooth recovery is performed. This avoids high-value rules being incorrectly downgraded or eliminated due to occasional noisy data, enhances the system's anti-interference capability when facing complex medical data, ensures the stability and reliability of the diagnostic rule base, and thus improves the accuracy of the final assisted diagnostic results.

[0012] In conjunction with some embodiments of the first aspect, in some embodiments, the system further includes an evidence source distribution analysis module and a local context isolation module, wherein: The evidence source distribution analysis module is used to extract the source metadata of all negative evidence that triggered the preset trigger threshold when the count value of the negative evidence counter reaches the preset trigger threshold. This module also calculates the distribution dispersion index of the source metadata in the preset distribution model, generates a negative evidence spatial concentration index, and transmits the spatial concentration index to the local context isolation module. The local context isolation module generates a global update interception instruction when the spatial concentration index is higher than a preset isolation threshold. In response to the global update interception instruction, the local context isolation module blocks the confidence update module's reduction of the basic confidence level, maintaining the global confidence value of the corresponding target association rule in the medical diagnosis association rule base unchanged. Finally, the local context isolation module generates a context suppression mask for the region pointed to by the source metadata and associates the context suppression mask with the target association rule.

[0013] In the above embodiments, spatial distribution analysis is used to distinguish between general rule failures and specific anomalies caused by local environments. When a high concentration of negative evidence is detected, a local isolation mechanism is triggered. This blocks erroneous penalties on the global confidence level while only shielding specific risk areas. This prevents data bias in a single area from contaminating the entire rule base, ensuring the stability of the overall service while achieving local risk control, thereby ensuring the overall reliability and accuracy of the assisted diagnostic recommendations.

[0014] In conjunction with some embodiments of the first aspect, in some embodiments, the local context isolation module specifically includes a referral link prediction unit and an association blocking execution unit, wherein: The referral link prediction unit is used to obtain historical referral records of the region pointed to by the source metadata and construct a flow probability map. The referral link prediction unit is also used to identify downstream associated regions where the probability of receiving referred patients from the region pointed to by the source metadata exceeds a preset risk threshold based on the flow probability map. The association blocking execution unit is used to generate a temporary warning mask for the corresponding downstream associated region and associate the temporary warning mask with the target association rule. The association blocking execution unit is also used to block the output matching result of the target association rule based on the temporary warning mask when a diagnosis request from a downstream associated region is detected and has a referral tag from the region pointed to by the source metadata.

[0015] In the above embodiments, by constructing a flow probability map to predict the direction of risk propagation, the referral cases from abnormal sources can be identified and intercepted before misdiagnosis occurs in downstream areas. This cuts off the hidden error propagation chain caused by patients moving across regions and avoids secondary misdiagnosis caused by the reuse of invalid rules in different places, thereby improving the rigor and accuracy of diagnosis in cross-regional hierarchical diagnosis and treatment scenarios.

[0016] In conjunction with some embodiments of the first aspect, in some embodiments, the system further includes a detection and masking module, wherein: The detection and mask removal module is used to generate a penetration instruction during the period when the context suppression mask is in effect, and output the output matching result of the target association rule in the area pointed to by the source metadata according to the penetration instruction; the detection and mask removal module is also used to obtain the verification data fed back to the output matching result, and calculate the proportion of positive evidence in the verification data within a preset period; the detection and mask removal module is also used to delete the context suppression mask associated with the target association rule when the proportion of positive evidence exceeds a preset positive evidence threshold.

[0017] In the above embodiments, by introducing active detection and closed-loop verification during risk isolation, the static shielding state is transformed into a dynamic sensing process. Through small-sample penetration testing, signals of improvement in the local medical environment are captured, and the suppression is automatically lifted after the validity of the verification rules is restored. This avoids effective rules being idle for a long time due to outdated risk assessments, ensuring that the availability of the diagnostic knowledge base is always synchronized with current clinical practice, thereby guaranteeing the continuous accuracy of auxiliary diagnostic services.

[0018] In conjunction with some embodiments of the first aspect, in some embodiments, the confidence update module specifically includes an interference feature analysis unit and a penalty decay execution unit, wherein: The interference feature analysis unit is used to extract the remaining symptom items in the symptom set that are not covered by the target association rule when they are determined to be negative evidence, and to calculate the association explanatory power between the remaining symptom items and the determined annotation results; the penalty decay execution unit is used to generate a penalty decay factor that is negatively correlated with the association explanatory power based on the association explanatory power; the penalty decay execution unit is also used to reduce and correct the preset penalty magnitude in the negative penalty strategy using the penalty decay factor.

[0019] In the above embodiments, by introducing an interference attribution mechanism based on residual symptoms, it is possible to distinguish whether the failure of a rule stems from its own logical errors or from interference by other concurrent symptoms. When external factors are found to have high interpretability of the diagnostic results, the penalty magnitude is automatically reduced, avoiding excessive reduction of the confidence of high-quality rules due to noisy data in complex cases. This protects the stability of core rules in the face of multiple coexisting diseases, ensures the objectivity and fairness of confidence scoring, and thus guarantees the overall robustness and accuracy of the medical auxiliary diagnostic system.

[0020] In conjunction with some embodiments of the first aspect, in some embodiments, the system further includes a rule fission and evolution module, wherein: The rule fission and evolution module is used to extract high-frequency differential features from the negative evidence that triggers the negative penalty strategy, distinguishing them from the symptom set of the target association rule. The rule fission and evolution module is also used to combine the high-frequency differential features with the symptom set of the target association rule to generate a new antecedent, and use the definite labeling results in the negative evidence as a new consequent to construct a derived association rule. The rule fission and evolution module is also used to write the derived association rule into the medical diagnosis association rule library.

[0021] In the above embodiments, by constructing a knowledge self-growth mechanism based on negative feedback, erroneous data that causes the original rules to fail is transformed into nutrients for constructing new rules. By automatically extracting differentiated features and generating more refined derived rules, automatic completion and coverage of disease subtypes or complex symptoms are achieved, solving the problem of insufficient applicability of general rules in specific complex scenarios and improving the accuracy and comprehensiveness of medical diagnosis.

[0022] In conjunction with some embodiments of the first aspect, in some embodiments, the confidence update module specifically includes a competition suppression unit, wherein: The competition suppression unit is used to retrieve the winning competition rule from the medical diagnosis association rule base based on the symptom set and the determined labeling results when the evidence is judged as negative. The competition suppression unit is also used to obtain the winning confidence of the winning competition rule and calculate the exclusive suppression coefficient that is positively correlated with the winning confidence. The competition suppression unit is also used to introduce the exclusive suppression coefficient into the negative penalty strategy.

[0023] In the above embodiments, a dynamic inhibition mechanism based on competitive success is introduced, using the confidence level of the correct answer to inversely weight the penalty for incorrect answers. When a competing rule with high confidence exists, the invalidation of the current rule is determined to be more exclusive, thereby imposing a more severe penalty to accelerate weight decay. By rapidly widening the confidence gap between superior and inferior rules, ambiguity in diagnostic decision-making is eliminated, thereby improving the identification and accuracy of medical diagnosis.

[0024] In conjunction with some embodiments of the first aspect, in some embodiments, the confidence update module specifically includes an organizational hierarchy weighting unit, wherein: The institution-level weighting unit is used to extract the level identifier of the medical institution that generated the marked result when obtaining the marked result; the institution-level weighting unit is also used to map the level identifier to the corresponding evidence confidence weight based on the preset institution authority mapping table; the institution-level weighting unit is also used to introduce the evidence confidence weight into the positive reinforcement strategy and the negative penalty strategy, so as to correct the new confidence level with the evidence confidence weight.

[0025] In the above embodiments, by introducing a multi-dimensional weighting mechanism based on the authority of medical institutions, the credibility of evidence can be intelligently distinguished according to the level of the diagnostic source. By assigning higher weight to diagnostic data from high-level medical institutions, the interference of misdiagnosis noise from primary or low-qualification institutions on the rule base is reduced, ensuring that the update of rule confidence is mainly driven by high-quality, high-authority clinical data, thereby improving the professionalism and accuracy of the auxiliary diagnostic system.

[0026] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: 1. This application reduces the weight of long-term unvalidated rules through a time-decay mechanism and a dynamic reward and punishment strategy based on clinical evidence, and adjusts the confidence of rules in real time according to actual diagnosis and treatment results. This enables the medical diagnosis association rule base to have the ability to self-evolve and weed out the inferior, promptly remove outdated or erroneous rules, and ensure that the rule base always keeps up with the evolution of diseases and the development of medical practice, thereby improving the accuracy of medical diagnosis.

[0027] 2. This application uses error acceleration parameters to deeply quantify the dynamic trend of rule failure, identify the risk of a sharp deterioration in rule quality, implement strong suppression in the early stage of error outbreak, accelerate the elimination process of severely failed rules, shorten the interference window of erroneous knowledge on clinical decision-making, thereby ensuring that rules are validated at high frequency and improving the accuracy of auxiliary diagnosis.

[0028] 3. This application distinguishes between substantial rule failure and short-term random fluctuations at the data level by introducing an observation window and a circuit breaker mechanism before imposing severe penalties. When a rule is detected to quickly recover its positive performance after fluctuations, excessive penalties are intercepted and a smooth recovery is performed. This avoids high-value rules being incorrectly downgraded or eliminated due to occasional noisy data, enhances the system's anti-interference ability when facing complex medical data, ensures the stability and reliability of the diagnostic rule base, and thus improves the accuracy of the final assisted diagnostic results. Attached Figure Description

[0029] Figure 1 A schematic diagram of the structure of the medical data association rule confidence dynamic update system provided in the embodiments of this application; Figure 2 A schematic diagram of a confidence update module provided in an embodiment of this application; Figure 3 Another structural diagram of a medical data association rule confidence dynamic update system provided in this application embodiment; Figure 4 A schematic diagram of a local context isolation module provided in an embodiment of this application; Figure 5 This is a schematic diagram of an exemplary hardware structure of the medical data association rule confidence dynamic update system in the embodiments of this application. Detailed Implementation

[0030] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification and appended claims of this application, the singular expressions “a,” “an,” “the,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to and includes any or all possible combinations of one or more of the listed items.

[0031] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0032] In related technologies, association rule mining algorithms are commonly used to construct disease diagnosis models. Association rules are calculated by statistically analyzing the frequency of co-occurrence of symptom sets and diagnosed diseases in historical medical records. Rules that meet the criteria are stored in a static knowledge base. When a new case is input, the diagnosis is ranked and recommended based on pre-calculated confidence levels. However, because common diseases have accumulated a large number of samples in historical databases, the corresponding rule confidence levels exhibit extremely high statistical inertia. When similar symptoms of a different disease occur frequently during a specific period, even if that disease has recently become more prevalent, the number of new samples is negligible compared to the historically accumulated samples of common diseases. This leads to a tendency to output historically high-probability common diseases when calculating the recommendation ranking, making it difficult to sensitively capture real-time changes in the disease spectrum, thus resulting in insufficient accuracy in assisted diagnosis.

[0033] In this application embodiment, in order to solve the above problems, a dynamic update system for the confidence of medical data association rules is provided, which enables the medical diagnosis association rule base to have the ability to self-evolve and weed out the inferior, break the constraint of the number of historical samples, promptly remove outdated rules and quickly respond to the current disease trend, thereby improving the accuracy of medical diagnosis.

[0034] This application provides a system for dynamically updating the confidence level of medical data association rules. (See attached document.) Figure 1 .

[0035] Figure 1 This application provides a medical data association rule confidence dynamic update system, which includes: a rule monitoring module 110, a confidence update module 120, and a rule base management module 130.

[0036] The rule monitoring module 110 is used to acquire the symptom set and determine the annotation results, and to retrieve the target association rule that matches the symptom set in the medical diagnosis association rule base; Among them, the symptom set refers to a set of feature data extracted from the medical records or electronic medical records to be processed, describing the patient's current clinical manifestations, such as fever and cough; the definitive labeling result refers to the final diagnostic conclusion obtained after diagnosis or pathological examination, which serves as the truth value to verify the accuracy of the rule; the medical diagnosis association rule base refers to a pre-built database that stores a large number of logical implications of symptom and disease forms and corresponding metadata such as confidence and support; the target association rule refers to a rule entry in the rule base that has an antecedent that is completely consistent with the currently obtained symptom set or satisfies a specific inclusion relationship.

[0037] Specifically, the rule monitoring module 110 first receives structured medical record data from the medical information system through a data interface, parses the symptom list as a symptom set, and extracts the diagnosis or confirmed result as the definitive labeling result. Then, the rule monitoring module 110 initiates a search request in the medical diagnosis association rule base using the symptom set as the query key. The search process typically employs hash matching or inverted index technology to quickly locate the rule that matches the antecedent with the symptom set. If multiple matching rules exist (e.g., combinations of symptoms with different fine-grained granularities), the rule monitoring module 110 will filter out the most relevant one as the target association rule according to a preset matching strategy (e.g., the maximum matching principle), and temporarily store the rule's unique identifier (ID) and related attributes.

[0038] In some embodiments, the process of retrieving target association rules can be implemented in several ways: Optionally, an index structure based on a Trie tree (prefix tree) can be constructed to map symptom combinations in the rule base to paths in the tree; the obtained symptom set can be sorted lexicographically; a traversal search can be performed in the Trie tree, and if the traversal ends and the node is marked as the end of a rule, the rule corresponding to that node can be output as the target association rule. It is understood that a similarity retrieval method based on a vector space model can also be used to achieve fuzzy matching, which is not limited here.

[0039] The confidence update module 120 is used to obtain the disease category corresponding to the target association rule, determine the corresponding time decay parameter based on the disease category, calculate the time interval between the current time and the last verification timestamp of the target association rule record, and use the time decay parameter and time interval to perform time decay calculation on the pre-stored confidence of the target association rule to obtain the basic confidence. Among them, disease category refers to the classification of the disease referred to by the rule according to the International Classification of Diseases, such as "acute infectious disease" and "chronic non-communicable disease"; time decay parameter is a coefficient used to quantify the rate at which the effectiveness of diagnostic rules for different disease categories decreases over time. This parameter is preset and is usually derived based on statistical analysis of large-scale historical medical data and empirical judgment of the evolution patterns of different diseases (for example, influenza rules have short timeliness and large decay parameter); last verification timestamp refers to the system time when the rule was last verified as correct by clinical data; pre-stored confidence refers to the confidence value currently stored in the rule base that has not yet been updated.

[0040] Specifically, the confidence update module 120 first reads the metadata of the target association rule to identify its disease category. Next, the confidence update module 120 queries a pre-configured parameter table to match the time-series decay factor 'a' corresponding to that category. Simultaneously, the confidence update module 120 obtains the current system time and the last verification time of the rule record, calculating the difference 't' between the two. Subsequently, the confidence update module 120 performs time-series decay calculations, typically using an exponential decay model, i.e., base confidence = pre-stored confidence × e. -at Alternatively, a linear decay model can be used: base confidence level = pre-stored confidence level - at (a lower limit needs to be set). This step reduces the weight of outdated rules that have not been validated by clinical data for a long time, thus reflecting timeliness.

[0041] In some embodiments, time-series decay calculation can be implemented in several ways: Optionally, a piecewise linear decay method can be used, setting multiple time thresholds; determining the interval into which the time interval falls; and obtaining the base confidence score by directly subtracting the corresponding value from the pre-stored confidence score based on the fixed deduction value corresponding to the interval. It is understood that a calculation method based on half-life can also be used, and this is not limited here.

[0042] In some embodiments, when the target association rule belongs to a disease category with significant epidemiological characteristics (such as seasonal influenza), a seasonal periodic factor can be introduced to dynamically correct the time-series decay parameter. Specifically, when determining the time-series decay parameter, the confidence update module further determines whether the current time is within the peak season window of the disease; if so, the value of the decay parameter is temporarily reduced to maintain the high confidence of the rule during the epidemic season; if it is outside the epidemic season, the decay parameter is restored or increased to accelerate the cooling of the rule's popularity. In this way, the system can adapt to the periodic patterns of disease occurrence and improve the dynamic adaptability of diagnostic assistance.

[0043] The confidence update module 120 is also used to determine whether the identified labeling result is consistent with the disease corresponding to the target association rule. If so, it is judged as positive evidence; otherwise, it is judged as negative evidence. Consistency refers to the fact that the disease entity represented by the labeling result is the same as the disease entity in the consequent of the target association rule in terms of medical semantics, or there is a clear hierarchical concept inclusion relationship (such as labeling as viral pneumonia and pneumonia in the rule); positive evidence refers to the factual basis of the current clinical case that supports the establishment of the association rule; negative evidence refers to the factual basis of the current clinical case that refutes or does not support the establishment of the association rule.

[0044] Specifically, the confidence update module 120 extracts the disease code (such as ICD code) of the confirmed annotation result and the disease code of the consequent of the target association rule. The confidence update module 120 first performs string comparison; if the codes are completely identical, they are directly determined to be consistent. If the codes are different, the confidence update module 120 further calls the medical knowledge graph or thesaurus to calculate the semantic similarity between the two disease concepts. If the similarity is higher than a preset judgment threshold, or if the two have a direct relationship in the knowledge graph, they are still determined to be consistent, and a positive evidence signal is output; otherwise, they are determined to be inconsistent, and a negative evidence signal is output.

[0045] In some embodiments, consistency judgment can be achieved in several ways: Optionally, a pre-trained consistency judgment model can be used to convert two disease names into vector representations; the cosine similarity between the two vectors can be calculated; if the similarity is greater than a preset similarity threshold, it is considered positive evidence; otherwise, it is considered negative evidence. It is understood that rule-based expert systems can also be used for logical reasoning judgment, which is not limited here.

[0046] The confidence update module 120 is also used to update the last verification timestamp to the current time when it is determined to be positive evidence, and to increase the value of the base confidence based on the positive reinforcement strategy to obtain the updated new confidence. Among them, the positive reinforcement strategy refers to a predefined numerical growth algorithm, which aims to increase the confidence of a rule by a specific step size or proportion when the rule is verified again by clinical facts. It usually includes control logic for the growth rate to prevent numerical overflow (i.e., exceeding 1). The new confidence value refers to the final confidence value that is ready to be written into the rule base after being adjusted by both time decay and positive reinforcement.

[0047] Specifically, upon receiving a positive evidence signal, the confidence update module 120 first locks the record of the target association rule and modifies the value of the last verification timestamp field to the current system time. This operation resets the freshness of the rule. Subsequently, the confidence update module 120 performs positive reinforcement calculation. The calculation logic is typically designed to add an increment to the base confidence level. This increment can be a fixed value or a dynamic value related to the current confidence level (e.g., (1 - base confidence level) × b, where b is the learning rate). This dynamic design ensures that the confidence level increases more slowly as it approaches 1, always approaching but never exceeding 1. After the calculation is complete, the confidence update module 120 outputs this value as the updated confidence level.

[0048] In some embodiments, the positive reinforcement strategy can be implemented in several ways: optionally, a Bayesian update method can be used, treating the base confidence level as the prior probability; the likelihood can be calculated using the current positive evidence; and the posterior probability can be calculated as the new confidence level according to Bayes' formula. It is understood that a nonlinear mapping growth method based on the Sigmoid function can also be used, which is not limited here.

[0049] The confidence update module 120 is also used to perform an incremental operation on the negative evidence counter associated with the target association rule when it is determined to be negative evidence, and when the count value of the negative evidence counter reaches the preset trigger threshold within the preset time window, reduce the value of the base confidence based on the negative penalty strategy to obtain the updated new confidence. Among them, the negative evidence counter is a variable maintained independently for each rule, used to count the number of times the rule is falsified within a specific time period; the preset time window is a sliding time range used to limit the validity of statistical data; the preset trigger threshold is the minimum number of errors required to trigger the confidence reduction operation. This threshold is preset and is usually determined by experimental testing based on the system's tolerance to noisy data, in order to filter out occasional false diagnosis data; the negative penalty strategy is a numerical reduction algorithm used to significantly reduce the confidence of a confirmed rule when its reliability decreases.

[0050] Specifically, upon receiving a negative evidence signal, the confidence update module 120 first identifies the target association rule ID and locates the corresponding counter in the memory database. The confidence update module 120 increments the counter by 1 and records the time of the negative evidence occurrence. Next, the confidence update module 120 checks the cumulative value of the counter within the current preset time window. If the count value does not reach the trigger threshold, the confidence update module 120 retains only the base confidence as the new confidence (i.e., no penalty is applied temporarily, as it may be data noise). If the count value reaches or exceeds the trigger threshold, the confidence update module 120 activates the negative penalty strategy. The penalty calculation can be a multiplicative penalty (e.g., new confidence = base confidence × preset penalty weight) or a subtractive penalty, thereby reducing the weight of the rule and reflecting its recent low reliability.

[0051] In some embodiments, the negative penalty logic can be implemented in several ways: Optionally, a sliding window queue can be used to record the timestamps of negative events; each time new negative evidence is added, records exceeding the time window at the head of the queue are removed; the queue length is calculated, and if the length is greater than a threshold, then the new confidence score = base confidence score - c (where c is the penalty coefficient) is applied. It is understood that a trend detection method based on exponentially weighted moving averages can also be used to trigger the penalty, which is not limited here.

[0052] In some embodiments, when evidence is determined to be negative and a penalty threshold is triggered, a weighted penalty mechanism based on the authority of the evidence source can be introduced. Specifically, before executing the negative penalty strategy, the confidence update module 120 further obtains the level of the medical institution or the title of the doctor providing the negative evidence; if the authority of the evidence source is extremely high, the penalty coefficient is dynamically increased (e.g., the penalty intensity is doubled), causing the confidence level to drop more significantly; if the authority of the evidence source is average, the standard penalty intensity is maintained. In this way, high-quality feedback information can be responded to more sensitively, erroneous rules can be eliminated more quickly, and overall credibility and diagnostic accuracy can be improved.

[0053] In some embodiments, where more granular and dynamic control over rule confidence is required, such as... Figure 2 As shown, the confidence update module 120 may also include an error trend analysis unit 210, a dynamic penalty execution unit 220, and an elastic recovery and circuit breaker protection unit 230 to achieve adaptive penalty for the density of negative evidence and intelligent fault tolerance for short-term fluctuations, thereby improving the robustness and accuracy of the system.

[0054] Specifically, the error trend analysis unit 210 first intervenes. When the count value of the negative evidence counter reaches the preset trigger threshold within a preset time window, it indicates that the rule has been frequently erroneous recently. At this time, the error trend analysis unit 210 extracts the timestamp sequence of all negative evidence within the time window and calculates the average time interval between two adjacent pieces of negative evidence. The smaller the average time interval, the higher the frequency of errors. An error density index is generated based on the reciprocal of this average or other inverse proportional relationships. The error density index reflects the current density of errors. To capture the dynamic trend of error development, the error trend analysis unit 210 further retrieves the historical error density index recorded at the last time the penalty was triggered, calculates the difference ratio between the current index and the historical index, and thus determines the error acceleration parameter. If the error acceleration parameter is positive and large, it indicates that errors are accelerating and the reliability of the rule is deteriorating rapidly. Subsequently, the dynamic penalty execution unit 220 nonlinearly amplifies (e.g., exponentially amplifies) the preset baseline penalty coefficient (which is usually set based on historical experience) based on the error acceleration parameter to obtain the amplified penalty coefficient. When the error acceleration parameter is positive and exceeds the preset warning threshold (which is usually set according to the system’s sensitivity to risk), the dynamic penalty execution unit 220 uses the amplified penalty coefficient to perform weighted deduction on the basic confidence level, thereby imposing a more severe penalty when the rule quality deteriorates rapidly, and accelerating the elimination of inferior rules.

[0055] Meanwhile, to prevent false positives due to data noise or short-term anomalies, the resilient recovery and circuit breaker protection unit 230 initiates an observation window before executing penalties. During this window, the resilient recovery and circuit breaker protection unit 230 continuously monitors newly arriving evidence. If the number of consecutively received positive evidence exceeds a preset recovery threshold (set based on statistical significance), the previous clustering of negative evidence is determined to be short-term fluctuation interference, rather than a fundamental error in the rule itself. In this case, the resilient recovery and circuit breaker protection unit 230 triggers the circuit breaker protection mechanism, terminating the original severe penalty and instead using a weighted average of the base confidence level and the pre-stored confidence level to generate a new confidence level. This smoothing approach retains some of the time-series decay effect while avoiding a precipitous drop in confidence level due to short-term fluctuations, ensuring the system's stability and adaptability when facing complex medical data.

[0056] The above technical solution, by introducing error acceleration and dynamic penalty mechanisms, can identify and severely punish rules whose quality deteriorates rapidly, thus purifying the rule base. At the same time, by combining elastic recovery and circuit breaker protection mechanisms, it can avoid drastic fluctuations in confidence when facing occasional data fluctuations, thereby improving stability and anti-interference ability while ensuring the accuracy of diagnostic rules.

[0057] In some embodiments, where negative evidence may stem from multiple coexisting diseases or complex symptom interference, the confidence update module 120 further includes an interference feature analysis unit and a penalty decay execution unit to distinguish between the rule's own defects and external interference, thereby achieving precise protection of the rule's confidence.

[0058] Specifically, when the current case is determined to be negative evidence, the interference feature analysis unit first performs a set difference operation on the current symptom set and the antecedent of the target association rule to extract the remaining symptom items not covered by the rule. These remaining symptom items may contain the true reasons leading to the actual diagnosis (i.e., the definitive labeling result). Subsequently, the interference feature analysis unit uses a pre-constructed medical knowledge graph or a Bayesian probability model to calculate the probability value of the definitive labeling result derived from these remaining symptom items, i.e., the association explanatory power. This indicator quantifies the contribution of interfering symptoms to the final diagnosis.

[0059] Next, the penalty attenuation execution unit generates a penalty attenuation factor based on the association's explanatory power. The two are negatively correlated; that is, the higher the association's explanatory power (indicating that the misdiagnosis is mainly caused by interfering symptoms rather than errors in the rule itself), the smaller the generated attenuation factor (e.g., close to 0). Finally, the penalty attenuation execution unit uses this attenuation factor to reduce and correct the preset penalty magnitude in the negative penalty strategy (e.g., multiplying the preset magnitude by the attenuation factor). The preset penalty magnitude is based on historical experimental data of the system's error tolerance or a standard deduction value set by domain experts. Through this process, false negative evidence caused by patient comorbidities or complex conditions is identified, and the penalty to the target rule is reduced accordingly, preventing the confidence of high-quality rules from being erroneously reduced due to external interference factors.

[0060] The above technical steps, by introducing interference feature analysis, can remove interfering factors in complex cases and quantify prediction biases caused by non-rule-specific reasons; by dynamically attenuating the penalty magnitude, the robustness and fairness of the system in handling complex clinical complication scenarios are improved while ensuring the purity of the rule base.

[0061] In some embodiments, when multiple rules compete for the same symptom set and the target rule fails, the confidence update module 120 further includes a competition suppression unit, which accelerates the elimination of inferior rules by introducing the suppression effect of the winning rule, thereby improving the convergence speed and diagnostic accuracy of the rule base.

[0062] Specifically, when a target association rule is judged as negative evidence, it means that the rule's prediction has failed. At this point, the competition suppression unit is activated. Using the current symptom set as the antecedent and the actual confirmed labeling result as the consequent, it searches the medical diagnosis association rule base in reverse to find the rule that can correctly deduce the diagnosis result, i.e., the winning rule. This step aims to find the correct answer and authority in the current scenario. Subsequently, the competition suppression unit reads the current confidence value of the winning rule (i.e., the winning confidence) and calculates the exclusive suppression coefficient based on a preset mapping function (usually constructed based on a sigmoid or linear function, with parameters fitted from historical data). Since the exclusive suppression coefficient is positively correlated with the winning confidence, the higher the confidence of the winning rule, the larger the generated suppression coefficient, representing a stronger rejection force of the correct answer against the incorrect answer. Finally, the competition suppression unit introduces this exclusive suppression coefficient into the negative penalty strategy, usually as a multiplier factor applied to the base penalty magnitude (e.g., final deduction value = base deduction value × exclusive suppression coefficient).

[0063] The above technical steps introduce a competition suppression mechanism, which uses the high confidence of correct rules to dynamically increase the penalty for incorrect rules, simulating the process of survival of the fittest. This not only accelerates the decline of the confidence of low-quality or outdated rules and their exit from the decision sequence, but also solves the problem of long-term coexistence of redundant rules in the rule base, thus improving the speed at which the system converges to the optimal rule set.

[0064] In some embodiments, where medical data sources are diverse and of varying quality, the confidence update module 120 also includes an institution-level weighting unit, which differentiates the confidence update magnitude by varying the authority of different medical institutions, thereby improving the reliability of the rule base construction.

[0065] Specifically, the institutional-level weighting unit intervenes simultaneously with the system acquiring the labeled results, parsing and extracting the grade identifier (e.g., tertiary-A, secondary-B, or community health service center) of the medical institution that generated the diagnosis from the metadata. To quantify the diagnoses of different institutions, the institutional-level weighting unit calls a pre-built institutional authority mapping table, which is constructed based on the hospital accreditation standards issued by the National Health Commission and industry-recognized medical level assessment data. By looking up the table, the institutional-level weighting unit maps the extracted grade identifiers to specific numerical evidence confidence weights. Subsequently, the institutional-level weighting unit deeply integrates this weight factor into the calculation formulas of positive reinforcement and negative penalty strategies. In positive reinforcement, high-weighted evidence leads to a greater increase in confidence, meaning that the endorsement of authoritative institutions can gain trust in the rules more quickly; in negative penalty, high-weighted counterexamples lead to a more drastic decrease in confidence, meaning that the denial by authoritative institutions has a veto effect.

[0066] The above technical steps introduce an institutional hierarchy weighting mechanism, so that the update of confidence no longer depends solely on the quantity of evidence, but focuses more on the quality and authority of the evidence. This not only reduces the pollution of the rule base by misdiagnosis data that may exist in primary or low-level medical institutions, but also enhances the sensitivity to high-quality medical knowledge, thereby improving the medical authority of the final diagnostic rule.

[0067] The rule base management module 130 is used to write new confidence levels into the medical diagnosis association rule base.

[0068] Here, the new confidence level refers to the final value output by the confidence level update module 120 after dynamic adjustment; writing refers to the database update operation.

[0069] Specifically, after receiving the new confidence level and the corresponding rule ID, the rule base management module 130 constructs a database update instruction. To ensure data consistency and integrity, the rule base management module 130 typically initiates a database transaction. Within the transaction, the rule base management module 130 locks the corresponding record row in the rule base, modifies the value of the confidence level field to the new confidence level, and may also update audit fields such as the last modification time. After confirming that the write operation is correct, the transaction is committed, completing the persistence. Furthermore, if the new confidence level is lower than a preset eviction threshold, the rule base management module 130 can also perform logical or physical deletion operations to remove the rule from the active database.

[0070] In some embodiments, write operations can be implemented in several ways: optionally, an asynchronous write mode can be used, sending update requests to a message queue; a background consumer process can read messages in batches; multiple update operations can be merged and written to the database at once to reduce database pressure. It is understood that a master-database write method under a read-write separation architecture can also be used, but this is not limited here.

[0071] In the above embodiments, by using a time-series decay mechanism and a dynamic reward and punishment strategy based on clinical evidence, the weight of long-term unverified rules is reduced, and the confidence of rules is adjusted in real time according to actual diagnosis and treatment results. This enables the medical diagnosis association rule base to have the ability to self-evolve and weed out the inferior, promptly eliminating outdated or erroneous rules, and ensuring that the rule base always keeps up with the evolution of diseases and the development of medical practice, thereby improving the accuracy of medical diagnosis.

[0072] like Figure 3 The diagram shown is another structural schematic of the medical data association rule confidence dynamic update system provided in this application embodiment, including: rule monitoring module 310, confidence update module 320, rule base management module 330, evidence source distribution analysis module 340, local context isolation module 350, and detection and mask removal module 360.

[0073] The rule monitoring module 310, the confidence update module 320, and the rule base management module 330 are together with Figure 1 The illustrated embodiment includes a rule monitoring module 110, a confidence update module 120, and a rule base management module 130. Similarly, see the descriptions of the rule monitoring module 110, confidence update module 120, and rule base management module 130; they will not be repeated here.

[0074] The evidence source distribution analysis module 340 is used to extract the source metadata of all negative evidence that triggers the preset trigger threshold when the count value of the negative evidence counter reaches the preset trigger threshold. Among them, source metadata refers to attribute information describing the background of negative evidence generation, which usually includes the geographical location of the medical institution (such as province and city), department type (such as respiratory department and emergency department), equipment model or data acquisition terminal ID, etc.; preset trigger threshold refers to the critical number of errors set by the system to start the in-depth analysis process. This threshold is preset and is usually determined by ROC curve analysis of historical operating data based on the balance between system load capacity and sensitivity to abnormal events.

[0075] Specifically, when the negative evidence counter in the evidence source distribution analysis module 340 accumulates to a threshold, a trigger signal is sent to the evidence source distribution analysis module 340. In response to this signal, the evidence source distribution analysis module 340 accesses the system log or evidence repository, and retrieves all relevant negative evidence records based on the target association rule ID and the current time window. Subsequently, the evidence source distribution analysis module 340 parses the specific source fields from these records, constructing a metadata set containing multiple source information entries.

[0076] In some embodiments, source metadata can be extracted in several ways: optionally, by querying a Structured Query Language (SQL) database and executing a Select statement to filter error logs within a specific time period; parsing JSON-formatted log fields to extract Location and Department attributes; and encapsulating the extraction results into a List object. It is understood that a streaming computing framework can also be used to capture metadata in real time, which is not limited here.

[0077] The evidence source distribution analysis module 340 is also used to calculate the distribution dispersion index of source metadata in the preset distribution model, generate the negative evidence spatial concentration index, and transmit the spatial concentration index to the local context isolation module. Among them, the preset distribution model refers to the mathematical model used to describe the distribution pattern of data in spatial or logical dimensions, such as the Gaussian distribution model or the spatial grid model based on the geographic information system; the distribution dispersion index refers to the numerical value that quantifies the degree of dispersion of data points in the model space, such as the standard deviation, Gini coefficient or spatial autocorrelation coefficient; the negative evidence spatial concentration index refers to the numerical value that reflects whether the error is concentrated in a specific area, which is calculated in reverse based on the dispersion index (the higher the concentration, the more likely the error is to be localized).

[0078] Specifically, the evidence source distribution analysis module 340 first maps the extracted source metadata to the coordinate system of a preset distribution model (e.g., mapping hospital addresses to latitude and longitude coordinates). Next, the evidence source distribution analysis module 340 calculates the centroid of these coordinate points and the average Euclidean or standard distance from all points to the centroid, using this as an indicator of distribution dispersion. To obtain a more intuitive sense of concentration, the evidence source distribution analysis module 340 typically employs normalization, for example, using the formula S=1 / (1+D) or S=e -D (Where D represents dispersion and S represents concentration), the spatial concentration index of negative evidence is calculated. The closer the index value is to 1, the more spatially concentrated the negative evidence is (e.g., concentrated in a certain hospital or city); the closer the value is to 0, the more randomly distributed the errors are nationwide. After the calculation is completed, the evidence source distribution analysis module 340 sends the index to the local context isolation module 350 through the internal message bus.

[0079] In some embodiments, the concentration index can be generated in several ways: Optionally, a grid density-based algorithm can be used to divide the map into several grids; the number of pieces of evidence falling into each grid can be counted; the variance of the grid density can be calculated, with a larger variance indicating greater concentration. Optionally, the K-Means clustering algorithm can be used, setting the number of cluster centers K=1; the sum of squared distances from all sample points to the cluster centers can be calculated; and the reciprocal of the sum of squared distances can be used as the concentration index. It is understood that the Moran index can also be used for spatial autocorrelation analysis, which is not limited here.

[0080] The local context isolation module 350 is used to generate a global update interception command when the spatial concentration index is higher than the preset isolation threshold. Among them, the preset isolation threshold refers to the numerical limit for determining whether negative evidence has significant local features. This threshold is preset and is usually set based on confidence intervals in statistics or business experience. The global update interception instruction is a system control signal used to notify the confidence update module to suspend the modification operation of the global attributes of the rule.

[0081] Specifically, after receiving the spatial concentration index, the local context isolation module 350 compares it with the internally stored isolation threshold. If the index value is greater than the threshold, it indicates that the current negative evidence is highly concentrated in a specific area (for example, a unique viral mutation may have occurred in that area, or there may be a systematic deviation in the parameters of medical equipment in that area), rather than a general failure of the rule itself. Therefore, the local context isolation module 350 determines that the global confidence of the rule should not be penalized and generates a global update interception instruction. This instruction typically includes the target rule ID, the interception reason code (such as a local anomaly), and the instruction's validity period.

[0082] The local context isolation module 350 is also used to respond to the global update interception command, block the confidence update module from reducing the basic confidence, and keep the global confidence value of the corresponding target association rule in the medical diagnosis association rule base unchanged. Here, blocking refers to intercepting or canceling an ongoing process flow; the global confidence score refers to the general confidence level of a rule that is effective in all regions by default.

[0083] Specifically, the local context isolation module 350 sends the generated global update interception command to the control interface of the confidence update module 320. Before executing the negative penalty strategy, the confidence update module 320 checks whether there is an interception command for the current rule. Once a valid command is detected, the confidence update module 320 immediately terminates the originally planned numerical deduction operation and directly retains the current base confidence or pre-stored confidence as the final result. This prevents the effectiveness of the rule on a global scale from being mistakenly affected by special circumstances in local areas (such as differences in local epidemics).

[0084] In some embodiments, blocking operations can be implemented in several ways: Optionally, an interceptor can be set before the confidence update method using aspect-oriented programming; the interception instruction can be checked; if it exists, a specific exception can be thrown or the process can be returned directly, skipping the execution of the original method. It is understood that blocking can also be achieved through database transaction rollback, but this is not limited here.

[0085] The local context isolation module 350 is also used to generate a context suppression mask for the region pointed to by the source metadata and associate the context suppression mask with the target association rule.

[0086] Among them, context suppression mask refers to a data structure or logical label used to define the scope of a rule's failure or degradation in a specific context; association refers to establishing a mapping relationship between rule IDs and mask objects.

[0087] Specifically, the local context isolation module 350 parses the source metadata to identify specific regions in the negative evidence set. Then, the local context isolation module 350 creates a context suppression mask object, which details the conditions under which suppression takes effect. Next, the local context isolation module 350 calls the rule base management interface to write the mask into the extended attribute table or association table of the target association rule. Once the mask takes effect, when the system subsequently processes diagnostic requests from that specific region, it will recognize the mask, thereby temporarily blocking or reducing the recommendation weight of that rule in that region; while for requests from other regions, since the mask conditions are not matched, the rule continues to function normally with its original high confidence.

[0088] In some embodiments, mask generation and association can be implemented in several ways: optionally, a JSON object describing the mask rules can be constructed; the JSON string can be stored in a rule document in a NoSQL database; or the mask can be loaded synchronously when the inference engine loads the rules. It is understood that bitmap indexes can also be used to mark region states, but this is not limited here.

[0089] In some embodiments, when changes in medical equipment or testing standards in a localized area lead to systematic negative evidence, the local context isolation module 350 can also introduce a device parameter calibration mechanism. Specifically, when generating a context suppression mask, the local context isolation module 350 further analyzes the device model information in the source metadata; if it finds that negative evidence is concentrated on a specific device model, the generated mask includes not only regional information but also device model limitations; simultaneously, the local context isolation module 350 sends a device calibration recommendation notification to the system administrator, which can precisely limit the suppression range to specific devices, avoiding impact on medical institutions using other normal equipment in the area, thus achieving more granular and precise isolation.

[0090] In some embodiments, where local abnormalities may spread with patient referrals, such as Figure 4 As shown, the local context isolation module 350 may also include a referral link prediction unit 410 and an association blocking execution unit 420 to achieve pre-control of potential risk areas and prevent erroneous rules from spreading in the associated medical network.

[0091] Specifically, the referral link prediction unit 410 obtains historical referral records for the region (i.e., the confirmed source region of the anomaly) from the medical information exchange platform or regional health information platform. These records contain data on the flow of patients from hospitals in this region to hospitals in other regions. The referral link prediction unit 410 uses this data to construct a flow probability map, where nodes represent different regions or hospitals, edges represent referral relationships, and the weight of the edges represents the referral probability. Next, the referral link prediction unit 410 performs path analysis based on this map to identify downstream associated regions where the probability of receiving referred patients from the source region exceeds a preset risk threshold (this threshold is preset based on an epidemiological transmission model or the distribution law of medical resources). Although these regions have not yet experienced large-scale negative evidence, they are highly likely to face the same anomaly soon.

[0092] Subsequently, the association blocking execution unit 420 generates a temporary warning mask for these identified downstream association regions and associates this mask with the target association rule. Unlike a complete suppression mask, the temporary warning mask has a conditional triggering characteristic. When the system detects a diagnosis request from these downstream regions, the association blocking execution unit 420 further checks whether the request contains a referral tag from the source region. If it does, it indicates that the patient is highly likely to be a direct associate of the abnormal source. In this case, the association blocking execution unit 420 will block the output of the target association rule based on the temporary warning mask to prevent the rule from failing on these high-risk referral cases.

[0093] The above technical steps, by constructing a referral link prediction mechanism, isolate known abnormal areas, predict and control potential risk transmission paths, prevent cross-regional rule failures caused by patient movement, and improve the system's security and forward-looking defense capabilities under a complex hierarchical medical system.

[0094] The detection and mask removal module 360 ​​is used to generate a penetration instruction during the period when the context suppression mask is in effect, and output the target association rule output matching result in the area pointed to by the source metadata according to the penetration instruction; Among them, the penetration command refers to the control command that allows the system to force rule matching and output results even when a suppression mask is present, similar to gray-scale testing; the output matching result refers to the diagnostic suggestions generated by the rule trial run in the suppressed area.

[0095] Specifically, the detection and mask removal module 360 ​​periodically scans the currently active context suppression masks. For each mask, the detection and mask removal module 360 ​​generates a penetration instruction with a low-probability sampling marker. When new case data from the suppressed region enters the system, the rules are normally masked, but if the sampling strategy of the penetration instruction is matched, the system will forcibly activate the rule for inference and mark the generated diagnostic suggestion (output matching result) as a exploratory suggestion output, or simply record it in the background without directly displaying it to the front end, for the purpose of collecting subsequent verification feedback.

[0096] In some embodiments, penetration detection can be implemented in several ways: Optionally, a random number generator can be set in the inference engine; when a request comes from a masked region, a random number between 0 and 1 is generated; if it is less than 0.05 (penetration rate), the mask limit is ignored, and the result is output normally and labeled. It is understood that manual penetration can also be performed by inviting specific doctors to participate in the test, which is not limited here.

[0097] The detection and mask removal module 360 ​​is also used to acquire verification data fed back to the output matching results and calculate the proportion of positive evidence in the verification data within a preset period. Among them, validation data refers to the adoption of the system's probing suggestions by clinicians or the feedback of the final diagnosis results; the preset period refers to the time span for statistical validation data; and the proportion of positive evidence refers to the proportion of the number of validation data that supports the correctness of the rules to the total number of probes.

[0098] Specifically, the detection and masking module 360 ​​continuously tracks the subsequent status of the output matching results generated by the penetration command through the system's feedback collection interface. The module collects whether the doctor adopted the suggestion or whether subsequent pathological examinations confirmed it. After a preset period, the module calculates the total amount of collected verification data and filters out the number of positive evidence points deemed correct or adopted. Finally, the module performs a division operation to calculate the percentage of positive evidence.

[0099] The detection and mask removal module 360 ​​is also used to delete the context suppression mask associated with the target association rule when the proportion of positive evidence exceeds the preset positive evidence threshold.

[0100] Among them, the preset positive evidence threshold refers to the accuracy standard for determining whether a rule can return to normal in a local area. This threshold is preset and is usually set to a level equivalent to the global confidence level. Deletion refers to removing the association between the rule and the mask so that it can return to normal service status.

[0101] Specifically, the detection and mask removal module 360 ​​compares the calculated percentage of positive evidence with a preset positive evidence threshold. If the percentage exceeds the threshold, it indicates that the accuracy of the rule in the previously problematic area has recovered to an acceptable level (this could be due to the end of the virus mutation cycle or the completion of calibration of local medical equipment). At this point, the detection and mask removal module 360 ​​determines that isolation is no longer necessary and immediately performs the removal operation. The detection and mask removal module 360 ​​accesses the rule base or mask storage table, finds the corresponding target association rule and context suppression mask record, and performs deletion operations or invalidates the mask status. After the operation is completed, the service of the rule in the area will be fully restored to normal and no longer affected by suppression.

[0102] In some embodiments, mask deletion can be implemented in several ways: Optionally, the database's Delete statement can be invoked to physically delete mask records based on the rule ID and region ID. It is understood that it can also be achieved by sending a de-isolation event notification to the configuration center, which then pushes the notification to each inference node to refresh the cache; this is not limited here.

[0103] In some embodiments, where negative evidence is not simply noise but implies a new disease subtype or a specific complication pattern, the medical data association rule confidence dynamic update system also includes a rule fission and evolution module to achieve automatic value-added and refined diagnostic granularity, thereby improving the system's coverage and adaptability.

[0104] Specifically, when a target association rule triggers a negative penalty strategy due to frequent negative evidence, the rule fission and evolution module is activated to deeply mine the negative evidence judged as erroneous. First, the module performs a set difference operation between the symptom set in the negative evidence and the antecedent symptoms of the original target association rule, extracting features not covered by the original rule. Next, the module performs statistical frequency analysis on these features, filtering out high-frequency differential features whose frequency exceeds a preset threshold (this threshold is preset and usually determined based on statistical significance tests, such as chi-square tests, to exclude occasional interference). Subsequently, the module executes rule reconstruction logic, combining these high-frequency differential features with the symptom set of the original rule (union operation) to generate a more stringent antecedent, and using the actual confirmed labeling result in the negative evidence as the new consequent, thereby constructing a completely new derived association rule.

[0105] The above technical steps achieve self-learning through rule fission and evolution mechanisms, transforming negative evidence that originally led to a decrease in confidence into nutrients for building new rules; not only automatically enriching the rule base's descriptive dimensions for difficult or complex diseases, but also achieving automatic evolution from coarse rules to fine rules, improving the diagnostic system's coverage and accuracy in the face of complex and ever-changing clinical situations.

[0106] In the above embodiments, by identifying the spatial distribution characteristics of negative evidence, when errors are detected to be concentrated in a specific area, the penalty for global confidence is blocked through a local context isolation mechanism, and a context suppression mask for that area is generated. This distinguishes between the general failure of the rule itself and the specific failure caused by the local environment, and avoids the pollution of the global rule base by local abnormal data. Thus, while ensuring the stability of the global diagnostic service, the system achieves the control and isolation of local risks, and improves the diagnostic accuracy of the system in complex geographical environments.

[0107] In some embodiments, the medical data association rule confidence dynamic update system 500 is a computer device or includes a computer device. The computer device includes a processor, memory, and a network interface connected via a system bus. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the computer device stores data. The network interface of the computer device is used to communicate with other external terminals or servers via a network connection. In some embodiments, the network interface can be a wired network interface; in some embodiments, the network interface can also be a wireless network interface. When the computer program is executed by the processor, it implements the methods in the embodiments of this application.

[0108] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0109] The above description is merely an exemplary embodiment of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure. Other embodiments of this disclosure will be readily apparent to those skilled in the art upon consideration of the specification and the disclosure of practical truth.

[0110] This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described in this disclosure. The specification and embodiments are to be considered exemplary only, and the scope and spirit of this disclosure are defined by the claims.

Claims

1. A system for dynamically updating the confidence level of medical data association rules, characterized in that, It includes a rule monitoring module, a confidence update module, and a rule base management module, among which: The rule monitoring module is used to acquire the symptom set and the labeling results, and to retrieve the target association rule that matches the symptom set in the medical diagnosis association rule base; The confidence update module is used to obtain the disease category corresponding to the target association rule, determine the corresponding time decay parameter based on the disease category, calculate the time interval between the current time and the last verification timestamp of the target association rule record, and use the time decay parameter and the time interval to perform time decay calculation on the pre-stored confidence of the target association rule to obtain the basic confidence. The confidence update module is also used to determine whether the determined labeling result is consistent with the disease corresponding to the target association rule. If so, it is determined as positive evidence; otherwise, it is determined as negative evidence. The confidence update module is further configured to update the last verification timestamp to the current time when the evidence is determined to be positive, and increase the value of the base confidence based on the positive reinforcement strategy to obtain the updated new confidence. The confidence update module is further configured to perform an incremental operation on the negative evidence counter associated with the target association rule when the evidence is determined to be negative evidence, and reduce the value of the base confidence based on the negative penalty strategy when the count value of the negative evidence counter reaches a preset trigger threshold within a preset time window, so as to obtain the updated new confidence. The rule base management module is used to write the new confidence level into the medical diagnosis association rule base.

2. The system according to claim 1, characterized in that, The confidence update module specifically includes an error trend analysis unit and a dynamic penalty execution unit, wherein: The error trend analysis unit is used to obtain the timestamp sequence of all negative evidence within the preset time window when the count value of the negative evidence counter reaches the preset trigger threshold, calculate the average time interval between adjacent negative evidence, and generate an error density index based on the average time interval. The error trend analysis unit is also used to calculate the difference ratio between the current error density index and the historical error density index when the penalty was last triggered, so as to determine the error acceleration parameter for generating negative evidence. The dynamic penalty execution unit is used to nonlinearly amplify the preset benchmark penalty coefficient based on the error acceleration parameter to obtain the amplified penalty coefficient; The dynamic penalty execution unit is further configured to, when the error acceleration parameter is positive and exceeds a preset warning threshold, use the amplified penalty coefficient to perform a weighted deduction on the basic confidence level to obtain a new confidence level.

3. The system according to claim 2, characterized in that, The confidence update module also includes an elastic recovery and circuit breaker protection unit, wherein: The elastic recovery and circuit breaker protection unit is used to start an observation window for the target association rule before the dynamic penalty execution unit uses the amplified penalty coefficient to perform weighted deduction on the basic confidence. The elastic recovery and circuit breaker protection unit is also used to determine that the negative evidence is short-term fluctuation interference if the number of positive evidence received continuously within the observation window exceeds a preset recovery threshold. The elastic recovery and circuit breaker protection unit is also used to generate a new confidence level by using the weighted average of the base confidence level and the pre-stored confidence level when it is determined to be a short-term fluctuation interference, and to terminate the observation window.

4. The system according to claim 1, characterized in that, The system also includes an evidence source distribution analysis module and a local context isolation module, wherein: The evidence source distribution analysis module is used to extract the source metadata of all negative evidence that triggers the preset trigger threshold when the count value of the negative evidence counter reaches the preset trigger threshold. The evidence source distribution analysis module is also used to calculate the distribution dispersion index of the source metadata in the preset distribution model, generate a negative evidence spatial concentration index, and transmit the spatial concentration index to the local context isolation module. The local context isolation module is used to generate a global update interception command when the spatial concentration index is higher than a preset isolation threshold. The local context isolation module is also used to respond to the global update interception command, block the confidence update module from reducing the basic confidence, and keep the global confidence value of the corresponding target association rule in the medical diagnosis association rule base unchanged. The local context isolation module is also used to generate a context suppression mask for the region pointed to by the source metadata, and associate the context suppression mask with the target association rule.

5. The system according to claim 4, characterized in that, The local context isolation module specifically includes a referral link prediction unit and an association blocking execution unit, wherein: The referral link prediction unit is used to obtain historical referral records of the region pointed to by the source metadata and construct a flow probability map; The referral link prediction unit is also used to identify downstream associated regions where the probability of receiving referred patients from the region pointed to by the source metadata exceeds a preset risk threshold, based on the flow probability map. The associated blocking execution unit is used to generate a temporary warning mask corresponding to the downstream associated region, and associate the temporary warning mask with the target association rule; The association blocking execution unit is further configured to, when detecting a diagnostic request from the downstream association region and having a referral tag from the region pointed to by the source metadata, block the output matching result of the target association rule based on the temporary warning mask.

6. The system according to claim 4, characterized in that, The system also includes a detection and masking module, wherein: The detection and mask removal module is used to generate a penetration instruction during the period when the context suppression mask is in effect, and output the output matching result of the target association rule in the region pointed to by the source metadata according to the penetration instruction; The detection and mask removal module is also used to acquire verification data fed back to the output matching result and calculate the proportion of positive evidence in the verification data within a preset period. The detection and mask removal module is also used to delete the context suppression mask associated with the target association rule when the proportion of positive evidence exceeds a preset positive evidence threshold.

7. The system according to claim 1, characterized in that, The confidence update module specifically includes an interference feature analysis unit and a penalty decay execution unit, wherein: The interference feature analysis unit is used to extract the remaining symptom items in the symptom set that are not covered by the target association rule when the evidence is determined to be negative, and to calculate the association explanatory power between the remaining symptom items and the determined labeling result. The penalty decay execution unit is used to generate a penalty decay factor that is negatively correlated with the correlation explanatory power based on the correlation explanatory power. The penalty decay execution unit is also used to reduce and correct the preset penalty magnitude in the negative penalty strategy using the penalty decay factor.

8. The system according to claim 1, characterized in that, The system also includes a rule-based fission and evolution module, wherein: The rule fission and evolution module is used to extract high-frequency differential features from the negative evidence that triggers the negative punishment strategy, which are different from the symptom set of the target association rule. The rule fission and evolution module is also used to combine the high-frequency differential features with the symptom set of the target association rule to generate a new antecedent, and to use the definite labeling result in the negative evidence as a new consequent to construct a derived association rule; The rule fission and evolution module is also used to write the derived association rules into the medical diagnosis association rule library.

9. The system according to claim 1, characterized in that, The confidence update module specifically includes a competition suppression unit, wherein: The competition suppression unit is used to, when determined to be negative evidence, retrieve the winning competition rule from the medical diagnosis association rule base based on the symptom set and the determined labeling result; The competition suppression unit is also used to obtain the winning confidence of the competition winning rule and calculate the exclusive suppression coefficient that is positively correlated with the winning confidence. The competition suppression unit is also used to introduce the exclusive suppression coefficient into the negative penalty strategy.

10. The system according to claim 1, characterized in that, The confidence update module specifically includes an institutional-level weighted unit, wherein: The institution-level weighting unit is used to extract the level identifier of the medical institution that generated the determined labeling result when obtaining the determined labeling result; The institutional hierarchy weighting unit is also used to map the level identifier to the corresponding evidence confidence weight based on a preset institutional authority mapping table. The institutional-level weighting unit is further configured to introduce the evidence confidence weight into the positive reinforcement strategy and the negative penalty strategy, so as to correct the new confidence level with the evidence confidence weight.

Images