Medical quality control scene-based mode intelligent mining and screening method, system, electronic device and storage medium

Abstract

The application discloses a medical quality control scene-based mode intelligent mining and screening method and system, an electronic device and a storage medium. The method comprises the following steps: constructing a time sequence event, assigning an event weight, performing weighted PrefixSpan mining, calculating time sequence semantic novelty, calculating a risk correlation degree, evaluating knowledge graph operability, and comprehensively screening. The application is used for accurately mining a medical quality control mode.

Description

Technical Field

[0001] This application belongs to the technical field of medical information processing, specifically relating to a pattern intelligent mining and screening method, system, electronic device, and storage medium based on medical quality control scenarios. Background Technology

[0002] Medical quality control relies on the timely identification of abnormal patterns and other quality control patterns during patient diagnosis and treatment. A valuable medical quality control pattern typically manifests as an ordered combination of multiple clinical events over time. For example, a rapid rise in body temperature followed by an accelerated increase in white blood cell count and then an increase in procalcitonin may indicate a risk of sepsis; a sudden increase in drainage volume followed by a sustained decrease in hemoglobin and then an increased heart rate may suggest postoperative bleeding. Automatically identifying these high-frequency temporal patterns with clinical early warning value from massive amounts of patient time-series diagnostic and treatment data is fundamental to building medical quality control rules, early warning systems, and clinical decision support.

[0003] Existing technologies for discovering patterns in medical quality control mainly employ three types of approaches. The first type is traditional sequence pattern mining algorithms, which discover high-frequency subsequences by scanning sequence databases, recursively constructing projection databases, or generating candidate sets layer by layer. The frequency of a time-series pattern in the sequence database is used as a support metric, and high-frequency patterns are filtered by setting a minimum support threshold. In general transaction sequence analysis, this approach has the advantage of requiring no domain knowledge and being able to discover any high-frequency combinations. The second type is anomaly detection methods based on traditional machine learning. These typically extract statistical features or embedding vectors from time-series data and use classification or clustering models to identify abnormal patterns. For example, time-series anomaly detection methods based on Long Short-Term Memory networks identify abnormal visit patterns by encoding the time series of multiple patient visits. The third type is reasoning methods based on medical knowledge graphs. These methods construct medical entities such as diseases, symptoms, examinations, treatments, and drugs into a graph structure and identify known violations or abnormal association patterns through graph reasoning.

[0004] However, these existing technologies have several technical shortcomings when it comes to pattern mining and screening in medical quality control scenarios. First, equal-weighted high-frequency pattern mining is not suitable for medical scenarios. Traditional sequence pattern mining algorithms treat every event in the sequence equally, using frequency as the sole metric. In medical scenarios, critical events and routine nursing events have drastically different clinical significance. Equal-weighted mining leads to a large number of clinically meaningless routine event patterns drowning out truly valuable abnormal combination patterns, resulting in an excessively high proportion of high-frequency but low-value routine nursing procedures in the output. Second, it does not consider the temporal proximity between events and their outcomes. The temporal pattern value of medical events is closely related to the time interval between them and the adverse event outcome. Predictive patterns of precursor events occurring several hours before the outcome have a much higher predictive value than those occurring tens of hours before the outcome. Traditional sequence pattern mining does not distinguish between the position of an event in the sequence and its temporal relationship with the outcome, and therefore cannot effectively mine precursor patterns with early warning value. Furthermore, the mining results lack a clinical value assessment mechanism. High-frequency temporal pattern mining typically generates a large number of output patterns, including many clinically known common-sense patterns, accidental co-occurrence patterns, and purely descriptive patterns that cannot be intervened. Current technologies lack a mechanism for systematically assessing and screening the mining results, making them unsuitable for direct application in quality control rule construction and requiring extensive manual screening. Next, novelty assessment lacks temporal semantics. Existing novelty assessments of temporal patterns often employ similarity calculation methods based on event sets, ignoring the temporal order relationships between medical events. Temporal patterns involving the same type of medical events but with different temporal orders may have completely different clinical meanings, and set similarity alone cannot accurately determine whether a pattern has been covered by existing rules. Finally, operability judgment lacks knowledge graph support. Determining whether a temporal pattern has interventional value requires knowing whether there are corresponding clinical intervention measures for the abnormal events at the end of the pattern. Current technologies rely on manual annotation, which is inefficient and lacks standardized criteria, making it impossible to automatically identify high-value patterns that can be intervened upon discovery. Summary of the Invention

[0005] To address one or more of the aforementioned problems, this application provides a first aspect of a pattern intelligent mining and screening method based on a medical quality control scenario, a second aspect of a pattern intelligent mining and screening system based on a medical quality control scenario, a third aspect of an electronic device, and a fourth aspect of a storage medium, for improving the accuracy of pattern mining in medical quality control.

[0006] The technical solution of this application is as follows.

[0007] The first aspect is a pattern intelligent mining and screening method based on medical quality control scenarios, including: The system receives patient diagnosis and treatment data from multiple medical business systems, performs structured event extraction and event type standardization, uses preset adverse outcome events as anchor points, arranges event instances in ascending order of time within a preset observation window, and constructs a patient event sequence database. Each event type is assigned an event severity weight based on its clinical severity. The time of occurrence of the adverse outcome event is used as the anchor point, and the time decay weight of each event instance is calculated based on the time interval between each event instance and the anchor point. The event severity weight is multiplied by the time decay weight to obtain the comprehensive weight of each event instance. Set minimum weighted support threshold and minimum weighted confidence threshold, and then execute the weighted PrefixSpan algorithm. During the recursive construction of the projection database, propagate the weighted product of prefix pattern matching, calculate the weighted support and weighted confidence of candidate patterns, and retain candidate patterns that satisfy the condition that the weighted support is greater than or equal to the minimum weighted support threshold and the weighted confidence is greater than or equal to the minimum weighted confidence threshold as the weighted frequent time series pattern set. For each candidate pattern in the weighted frequent temporal pattern set, calculate the event set Jaccard similarity and temporal edit distance similarity between the candidate pattern and each rule in the quality control rule base. The event set Jaccard similarity and temporal edit distance similarity are weighted and fused to obtain the comprehensive semantic similarity. The novelty score is calculated by taking the maximum comprehensive semantic similarity between the candidate pattern and all rules. For each candidate pattern, calculate the conditional probability of the adverse outcome event occurring within a preset time window after the candidate pattern appears, and the average time lead between the time of the first complete match of the candidate pattern and the time of the adverse outcome event. The conditional probability and the average time lead are weighted and fused to obtain the risk correlation score. For each candidate pattern, the last event type of that candidate pattern is extracted, and an intervention query path search is performed based on a pre-built medical knowledge graph. An operability score is assigned based on the results of the intervention query path search. For each candidate pattern, the novelty score, risk correlation score, and operability score are weighted and fused to calculate the comprehensive value score. Candidate patterns with a comprehensive value score greater than or equal to a preset value threshold are retained. After clustering and deduplication, they are sorted in descending order of comprehensive value score, and a quality control pattern list is output.

[0008] As one implementation of the first aspect, the time decay weight is calculated using a piecewise exponential decay method, including: For event instances where the time elapsed since the occurrence of the adverse outcome event does not exceed a first preset time, the time decay weight of the event instance is calculated using a first decay rate. For event instances where the time elapsed between the occurrence of an adverse outcome event and the occurrence of an adverse outcome event exceeds a first preset time but does not exceed a second preset time, the baseline weight value is first calculated up to the first preset time according to the first decay rate, and then the time decay weight of the event instance is calculated again starting from the baseline weight value and using the second decay rate. Among them, the first preset duration is less than the second preset duration, and the first attenuation rate is less than the second attenuation rate; For patient event sequences where no adverse outcome events occurred, the time decay weights of all corresponding event instances were set to the preset baseline weight values.

[0009] As one implementation of the first aspect, calculating the weighted support of candidate patterns includes: For any candidate pattern, iterate through each patient event sequence in the patient event sequence database and determine whether the patient event sequence contains the corresponding candidate pattern. If it does, obtain the comprehensive weight of each event instance in the patient event sequence that matches the corresponding candidate pattern. Multiply the comprehensive weights of each event instance to obtain the sequence matching weight of the corresponding candidate pattern in the patient event sequence. Sum the sequence matching weights of all patient event sequences that contain the corresponding candidate pattern and divide the sum by the total number of sequences in the patient event sequence database to obtain the weighted support of the candidate pattern. During the recursive construction of the projected database, the weighted product of prefix pattern matching is propagated, including: When constructing the projection database for the current prefix pattern, for each patient event sequence containing the prefix pattern, the last matching position of the prefix pattern in the patient event sequence is located. All events after the matching position constitute the projection sequence corresponding to the patient event sequence, and the corresponding sequence matching weight is stored as the projection weight associated with the projection sequence. When expanding candidate events in the projection database, for each expanded candidate event in the projection sequence, the projection weight is multiplied by the comprehensive weight of the event instance to which the expanded candidate event belongs, to obtain the weighted support contribution value of the expanded candidate pattern.

[0010] As one implementation of the first aspect, calculating the temporal edit distance similarity includes: Align the sequences of candidate patterns and corresponding event types, and calculate the minimum editing cost required to convert the candidate pattern into a rule, taking the insertion, deletion, and replacement operations of the event types as editing operations. Among them, the unit cost of the insertion operation and the deletion operation are equal, and the unit cost of the replacement operation is set to different cost values ​​according to whether the replaced event type belongs to the same clinical category. The replacement cost within the same clinical category is less than the replacement cost across clinical categories. In the process of calculating the minimum editing cost, if there are cases where the candidate pattern and the rule have the same event type but the time sequence is reversed, a preset reversal penalty cost will be added on the basis of the minimum editing cost. The temporal edit distance similarity is calculated based on the larger of the minimum edit cost, the number of corresponding event instances in the candidate pattern and the rule; Calculating comprehensive semantic similarity includes: The comprehensive semantic similarity is obtained by summing the event set Jaccard similarity and the temporal edit distance similarity according to preset weights.

[0011] As one implementation of the first aspect, calculating the average time lead includes: In each patient event sequence containing the corresponding candidate pattern and where an adverse outcome event has occurred, the matching timestamp of the last event of the corresponding candidate pattern is obtained, and the difference between the occurrence timestamp and the matching timestamp of the adverse outcome event in each patient event sequence where an adverse outcome event has occurred is calculated to obtain the individual lead time. The average time lead is obtained by summing up the individual lead times of all patient event sequences that resulted in adverse outcomes and dividing by the total number of patient event sequences that resulted in adverse outcomes. The conditional probability and the average time lead are weighted and fused, specifically including: Calculate the quotient of the average lead time divided by the preset maximum effective lead time window, and then take the smaller value between the average lead time and the quotient as the normalized lead time score. The conditional probability is multiplied by the preset outcome probability weighting coefficient, the normalized lead time score is multiplied by the difference between the value and the outcome probability weighting coefficient, and the two products are added together to obtain the risk correlation score.

[0012] As one implementation of the first aspect, the medical knowledge graph includes at least abnormal event nodes, disease nodes, intervention measure nodes, examination item nodes, a first relationship where abnormal events indicate diseases, a second relationship where diseases correspond to intervention measures, a third relationship where abnormal events require examination items, and a fourth relationship where abnormal events directly correspond to intervention measures. The search path for implementing intervention measures includes: The first query path checks whether the last event type has an intervention node directly connected through the fourth relationship; The second query path queries whether the last event type has a disease node connected through the first relationship, and whether the disease node has an intervention measure node connected through the second relationship. The third query path checks whether the last event type has a check item node connected through a third relationship; Assigning operability scores, including: If the first query path is hit, the first operability score is assigned; If the first query path is not hit but the second query path is hit, then the second operability score is assigned. If only the third query path is hit, then the third operability score is assigned; If no query path is hit, a fourth operability score is assigned; Among them, the first operability score is greater than the second operability score, the second operability score is greater than the third operability score, and the third operability score is greater than the fourth operability score.

[0013] As one implementation method of the first aspect, after clustering and deduplication, the data are sorted in descending order of comprehensive value score, specifically including: Hierarchical clustering is performed on all remaining candidate patterns based on temporal edit distance, and candidate patterns whose temporal edit distance between any two candidate patterns is less than a preset deduplication threshold are classified into the same pattern category. Within each pattern category, only the candidate pattern with the highest overall value score is retained as the representative pattern; Within the pattern category, representative patterns are arranged in descending order of their comprehensive value scores to form a quality control pattern list.

[0014] Secondly, a pattern intelligent mining and screening system based on medical quality control scenarios includes: The time-series event construction module is used to receive patient diagnosis and treatment data from multi-source medical business systems, extract structured events and standardize event types. Using preset adverse outcome events as anchor points, it arranges event instances in ascending order of time within a preset observation window to construct a patient event sequence database. The event weight assignment module is used to assign an event severity weight to each event type based on the clinical severity of the event type. Using the occurrence time of the adverse outcome event as the anchor point, the time decay weight of each event instance is calculated based on the time interval between each event instance and the anchor point. The event severity weight is multiplied by the time decay weight to obtain the comprehensive weight of each event instance. The weighted prefixSpan mining module is used to set the minimum weighted support threshold and the minimum weighted confidence threshold, and then execute the weighted prefixSpan algorithm. During the recursive construction of the projection database, the weighted product of prefix pattern matching is propagated, the weighted support and weighted confidence of candidate patterns are calculated, and candidate patterns that satisfy the condition that the weighted support is greater than or equal to the minimum weighted support threshold and the weighted confidence is greater than or equal to the minimum weighted confidence threshold are retained as the weighted frequent time series pattern set. The temporal semantic novelty module is used to calculate the event set Jaccard similarity and temporal edit distance similarity between each candidate pattern in the weighted frequent temporal pattern set and each rule in the quality control rule base. The event set Jaccard similarity and temporal edit distance similarity are weighted and fused to obtain the comprehensive semantic similarity. The novelty score is calculated by taking the maximum comprehensive semantic similarity between the candidate pattern and all rules. The risk correlation module is used to calculate the conditional probability of an adverse outcome event occurring within a preset time window after the appearance of each candidate pattern, as well as the average time lead between the first complete match of the candidate pattern and the occurrence of the adverse outcome event. The risk correlation score is obtained by weighting and fusing the conditional probability and the average time lead. The knowledge graph operability module is used to extract the last event type of each candidate pattern, perform an intervention query path search based on the pre-built medical knowledge graph, and assign an operability score according to the results of the intervention query path search. The comprehensive screening module is used to calculate the comprehensive value score for each candidate mode by weighting and fusing the novelty score, risk correlation score, and operability score. Candidate modes with a comprehensive value score greater than or equal to a preset value threshold are retained. After clustering and deduplication, the modes are sorted in descending order of comprehensive value score and the quality control mode list is output.

[0015] Thirdly, an electronic device includes a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor, when executing the computer program, implements the pattern intelligent mining and screening method based on medical quality control scenarios of the first aspect.

[0016] Fourthly, a storage medium is a computer-readable storage medium storing computer instructions for enabling a computer to implement the pattern intelligent mining and screening method based on the medical quality control scenario in the first aspect.

[0017] Compared with the prior art, the advantages of this application are as follows: Improve the accuracy of quality control pattern mining, avoid repeatedly discovering known patterns that are covered by existing rules, and realize automatic judgment of the operability of quality control patterns. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating one of the pattern intelligent mining and screening methods based on a medical quality control scenario in this application.

[0019] Figure 2 This is a schematic diagram of the structural framework of one of the pattern intelligent mining and screening systems based on medical quality control scenarios in this application.

[0020] Figure 3 This is a schematic diagram of the frame of one of the electronic devices in this application. Detailed Implementation

[0021] Referring to the illustrations, the principles of this application are illustrated by way of example implementation in a suitable operating environment. The following description is based on the illustrative specific embodiments of this application and should not be construed as limiting other specific embodiments not detailed herein, such as corresponding adjustments to the order of steps of the method of this application based on the principles of the technical concept.

[0022] like Figure 1 As shown in the figure, this application discloses a pattern intelligent mining and screening method based on medical quality control scenarios, including the following steps.

[0023] S1. Receive patient diagnosis and treatment data from multiple medical business systems, extract structured events and standardize event types. Using preset adverse outcome events as anchor points, arrange event instances in ascending order of time within a preset observation window to construct a patient event sequence database.

[0024] Optionally in this embodiment, each event instance formed after structured event extraction is represented as a six-tuple data structure, which includes: a unique event identifier, a patient identifier, an event type, an event occurrence timestamp, a set of event attributes, and an event severity level. The event type is uniformly assigned using standardized terminology coding, the event attribute set records the abnormal direction and specific value of the event, and the event severity level is used to identify the clinical urgency of the event instance.

[0025] Optionally in this embodiment, event type standardization includes mapping heterogeneous event descriptions from different medical business systems to a unified event type system; for example, mapping “WBC”, “white blood cell count” and “white blood cells” described separately in the testing system to the standard event type LAB_WBC, mapping abnormal body temperatures of 38.6℃ and 39.0℃ to the event type VITALS_TEMP_HIGH, and retaining the original specific values ​​in the event attribute set.

[0026] In this embodiment, the adverse outcome events may include, but are not limited to: diagnosis of sepsis, septic shock, postoperative bleeding, unplanned return to the operating room, ICU transfer, resuscitation, death, acute kidney injury, venous thromboembolism, adverse drug reaction, and hospital-acquired infection; the occurrence time of the adverse outcome event is recorded as the anchor time of the corresponding patient event sequence.

[0027] Optionally in this embodiment, the patient event sequence database also includes control sequences in which no adverse outcome events occurred. For patients who did not experience adverse outcome events, a continuous time window of equal duration to the preset observation window is randomly selected during the patient's hospitalization. All events within the continuous time window are arranged in ascending order of time to construct a control sequence. The control sequence can be used to reduce the support of accidental co-occurrence patterns without outcome association in weighted PrefixSpan sequence pattern mining.

[0028] S2. Assign an event severity weight to each event type based on the clinical severity of the event type. Using the occurrence time of the adverse outcome event as the anchor point, calculate the time decay weight of each event instance based on the time interval between each event instance and the anchor point. Multiply the event severity weight by the time decay weight to obtain the comprehensive weight of each event instance.

[0029] In this embodiment, the time decay weight is calculated using a piecewise exponential decay method, including: For event instances where the time elapsed since the occurrence of the adverse outcome event does not exceed a first preset time, the time decay weight of the event instance is calculated using a first decay rate. For event instances where the time elapsed between the occurrence of an adverse outcome event and the occurrence of an adverse outcome event exceeds a first preset time but does not exceed a second preset time, the baseline weight value is first calculated up to the first preset time according to the first decay rate, and then the time decay weight of the event instance is calculated again starting from the baseline weight value and using the second decay rate. Among them, the first preset duration is less than the second preset duration, and the first attenuation rate is less than the second attenuation rate; For patient event sequences where no adverse outcome events occurred, the time decay weights of all corresponding event instances were set to the preset baseline weight values.

[0030] Optionally in this embodiment, the allocation of event severity weights is performed according to a preset clinical severity level. The correspondence between clinical severity level and event severity weight is as follows: (a) The first severity level is the critical value level, and the corresponding event severity weight is 1.0; the critical value level includes at least one event type among cardiac arrest, severe anaphylactic reaction, shock index greater than 1, and serum potassium greater than 6.5 mmol / L; (b) The second severity level is the high-abnormal level, with a corresponding event severity weight of 0.8; the high-abnormal level includes at least one of the following event types: lactate > 4 mmol / L, systolic blood pressure < 90 mmHg, blood oxygen saturation < 90%, and procalcitonin > 10 ng / mL. (c) The third severity level is the moderate abnormality level, with a corresponding event severity weight of 0.5; the moderate abnormality level includes body temperature greater than 38.5℃ and white blood cell count greater than 15×10.9 At least one of the following event types: / L, hemoglobin less than 70g / L, blood glucose greater than 16.7mmol / L; (d) The fourth severity level is the mild abnormality level, with a corresponding event severity weight of 0.3; the mild abnormality level includes a body temperature of 37.5℃ to 38.5℃ and a white blood cell count of 10 to 15×10. 9 At least one of the following event types: mild elevation of L and C-reactive protein; (e) The fifth severity level is the normal level, with a corresponding event severity weight of 0.1; the normal level includes at least one event type among routine examination completion, medical order issuance, nursing record, and normal vital signs.

[0031] In this embodiment, optionally, in the segmented exponential decay method for calculating the time decay weight, for event instances where the time between the occurrence of the adverse outcome event and the event does not exceed a first preset duration, the time decay weight is calculated using a first exponential function with a natural constant as the base and the negative of the product of the first decay rate and the time interval as the exponent; for event instances where the time between the occurrence of the adverse outcome event exceeds the first preset duration but does not exceed a second preset duration, the time decay weight is first calculated using the first exponential function up to the first preset duration, and then, starting from the benchmark weight value, the time decay weight of the event instance is further calculated using a second exponential function with a natural constant as the base and the negative of the product of the second decay rate and the duration exceeding the first preset duration as the exponent; wherein, the first preset duration is 24 hours, the second preset duration is 72 hours, the first decay rate is 0.03 per hour, the second decay rate is 0.08 per hour, and for patient event sequences where no adverse outcome event has occurred, the time decay weight of all corresponding event instances is set to a preset baseline weight value of 0.1.

[0032] Optionally in this embodiment, the comprehensive weight of each event instance is obtained by multiplying the event severity weight by the time decay weight. Specifically, for any event instance, the event severity weight corresponding to the event type to which the event instance belongs is multiplied by the time decay weight corresponding to the event instance, and the result of the multiplication is used as the comprehensive weight of the event instance. If the event instance belongs to the normal level and is located in the control sequence, the comprehensive weight of the event instance is the product of 0.1 and 0.1.

[0033] S3. Set the minimum weighted support threshold and the minimum weighted confidence threshold, and then execute the weighted PrefixSpan algorithm. During the recursive construction of the projection database, propagate the weighted product of the prefix pattern matching, calculate the weighted support and weighted confidence of the candidate patterns, and retain the candidate patterns that satisfy the condition that the weighted support is greater than or equal to the minimum weighted support threshold and the weighted confidence is greater than or equal to the minimum weighted confidence threshold as the weighted frequent time series pattern set.

[0034] In this embodiment, calculating the weighted support of candidate patterns includes: For any candidate pattern, iterate through each patient event sequence in the patient event sequence database and determine whether the patient event sequence contains the corresponding candidate pattern. If it does, obtain the comprehensive weight of each event instance in the patient event sequence that matches the corresponding candidate pattern, and multiply the comprehensive weights of each event instance to obtain the sequence matching weight of the corresponding candidate pattern in the patient event sequence. Sum the sequence matching weights of all patient event sequences that contain the corresponding candidate pattern, and divide the sum by the total number of sequences in the patient event sequence database to obtain the weighted support of the candidate pattern.

[0035] In this embodiment, the propagation of the weighted product of prefix pattern matching during the recursive construction of the projection database includes: When constructing the projection database for the current prefix pattern, for each patient event sequence containing the prefix pattern, the last matching position of the prefix pattern in the patient event sequence is located. All events after the matching position constitute the projection sequence corresponding to the patient event sequence, and the corresponding sequence matching weight is stored as the projection weight associated with the projection sequence. When expanding candidate events in the projection database, for each expanded candidate event in the projection sequence, the projection weight is multiplied by the comprehensive weight of the event instance to which the expanded candidate event belongs, to obtain the weighted support contribution value of the expanded candidate pattern.

[0036] Optionally in this embodiment, the calculation of weighted confidence includes: for any candidate pattern, dividing the weighted support of the candidate pattern by the weighted support of the prefix pattern formed after removing the last event from the candidate pattern to obtain the weighted confidence of the candidate pattern; the weighted confidence is used to measure the weighted conditional probability of the extended event occurring under the condition that the prefix pattern has occurred.

[0037] In this embodiment, optionally, retaining candidate patterns that satisfy the conditions of weighted support greater than or equal to the minimum weighted support threshold and weighted confidence greater than or equal to the minimum weighted confidence threshold includes: if the weighted support of a candidate pattern is less than the minimum weighted support threshold, or the weighted confidence of the candidate pattern is less than the minimum weighted confidence threshold, then pruning and removing the candidate pattern and its subsequent extended branches.

[0038] In this application, weighted PrefixSpan sequence pattern mining is performed, including the following steps: (1) Initial scanning steps: Scan the patient event sequence database, calculate the weighted support of each event type with a length of 1, and retain the event types with a weighted support greater than or equal to the minimum weighted support threshold as the initial frequent prefix set; (2) Projection database construction steps: For the current prefix pattern, construct its projection database; for each patient event sequence containing the prefix pattern, locate the last matching position of the prefix pattern in the patient event sequence, construct all events after the matching position as the projection sequence corresponding to the patient event sequence, and store the sequence matching weight of the patient event sequence for the prefix pattern as the projection weight associated with the projection sequence. (3) Recursive mining steps: Scan all suffix events in the projection database. For the event type of each extended candidate event, multiply the projection weight by the comprehensive weight of the event instance to which the event type belongs to obtain the weighted support contribution value of the extended candidate pattern. Sum the weighted support contribution values ​​of each projection sequence and divide by the total number of sequences in the patient event sequence database to obtain the weighted support of the extended candidate pattern. (4) Pruning steps: If the weighted support of the expanded candidate pattern is less than the minimum weighted support threshold, then prune the expanded branch; otherwise, add the expanded candidate pattern to the frequent pattern set and continue recursively with the expanded candidate pattern as the new prefix. (5) Termination step: When the projection database is empty, or the weighted support of all expanded candidate patterns is lower than the minimum weighted support threshold, terminate the current recursive branch.

[0039] The weighted prefixSpan sequence pattern mining method in this application differs fundamentally from the traditional prefixSpan algorithm. First, traditional prefixSpan uses the number of sequences containing the pattern as support, while the weighted prefixSpan mining method in this application uses the sum of the products of the comprehensive weights of each event instance in the matching event as the weighted support. Second, the projection database of traditional prefixSpan only maintains suffix events, while this application propagates the weighted product of prefix matching in the projection database, so that the weighted support of the extended pattern is calculated based on the complete prefix weights for any pattern length. Third, traditional prefixSpan cannot distinguish between critical event combinations and normal event combinations, while this application uses a dual weighting of event severity weight and time decay weight to automatically give higher weighted support to anomalous event combinations with high severity and close to adverse outcomes.

[0040] In this embodiment, the combined weights of each event instance are multiplied to obtain the sequence matching weight of the patient event sequence for the corresponding candidate pattern. Through the multiplication mechanism, if an event instance at any position in the candidate pattern is a low-weight routine event, the combined weight of the low-weight event instance will lower the sequence matching weight of the entire candidate pattern, thereby inhibiting the clinically insignificant combination pattern containing routine events from obtaining high weighted support.

[0041] S4. For each candidate pattern in the weighted frequent temporal pattern set, calculate the event set Jaccard similarity and temporal edit distance similarity between the candidate pattern and each rule in the quality control rule base. Weight the event set Jaccard similarity and temporal edit distance similarity to obtain the comprehensive semantic similarity. Take the maximum comprehensive semantic similarity between the candidate pattern and all rules to calculate the novelty score.

[0042] In this embodiment, calculating the temporal edit distance similarity includes: Align the sequences of candidate patterns and corresponding event types, and calculate the minimum editing cost required to convert the candidate pattern into a rule, taking the insertion, deletion, and replacement operations of the event types as editing operations. Among them, the unit cost of the insertion operation and the deletion operation are equal, and the unit cost of the replacement operation is set to different cost values ​​according to whether the replaced event type belongs to the same clinical category. The replacement cost within the same clinical category is less than the replacement cost across clinical categories. In the process of calculating the minimum editing cost, if there are cases where the candidate pattern and the rule have the same event type but the time sequence is reversed, a preset reversal penalty cost will be added on the basis of the minimum editing cost. The temporal edit distance similarity is calculated based on the larger of the minimum edit cost, the number of corresponding event instances in the candidate pattern and the rule.

[0043] In this embodiment, calculating the comprehensive semantic similarity includes: The comprehensive semantic similarity is obtained by summing the event set Jaccard similarity and the temporal edit distance similarity according to preset weights.

[0044] In this embodiment, optionally, the unit cost of insertion and deletion operations is set to 1; the unit cost of replacement operation is set according to whether the replaced event types belong to the same clinical category. If the two replaced event types belong to the same clinical category, the unit cost of replacement operation is set to 0.5; if the two replaced event types belong to different clinical categories, the unit cost of replacement operation is set to 1; the preset reversal penalty cost is set to 0.5; during the calculation of the minimum editing cost, if there are cases where the candidate patterns and rules have the same event types but reversed temporal order, then 0.5 is added to the minimum editing cost as a temporal reversal penalty.

[0045] Optionally in this embodiment, the event set Jaccard similarity and the temporal edit distance similarity are weighted and summed according to a preset weight, including: multiplying the event set Jaccard similarity by a first preset fusion weight of 0.4, multiplying the temporal edit distance similarity by a second preset fusion weight of 0.6, and adding the two products to obtain the comprehensive semantic similarity.

[0046] Optionally in this embodiment, the novelty score is calculated by taking the maximum value of the comprehensive semantic similarity between the corresponding candidate pattern and all rules, including: subtracting the maximum value of the comprehensive semantic similarity from the value 1 to obtain the novelty score of the candidate pattern; if the novelty score of the candidate pattern is greater than 0.6, the candidate pattern is determined to have significant novelty.

[0047] S5. For each candidate pattern, calculate the conditional probability of the adverse outcome event occurring within a preset time window after the candidate pattern appears, and the average time lead between the time of the first complete match of the candidate pattern and the time of the adverse outcome event. The conditional probability and the average time lead are weighted and fused to obtain the risk correlation score.

[0048] In this embodiment, calculating the average time lead includes: In each patient event sequence containing the corresponding candidate pattern and where an adverse outcome event has occurred, the matching timestamp of the last event of the corresponding candidate pattern is obtained, and the difference between the occurrence timestamp and the matching timestamp of the adverse outcome event in each patient event sequence where an adverse outcome event has occurred is calculated to obtain the individual lead time. The average time advance is obtained by summing up the individual advances of all patient event sequences that resulted in adverse outcomes and dividing by the total number of patient event sequences that resulted in adverse outcomes.

[0049] In this embodiment, the conditional probability and the average time lead are further weighted and fused, specifically including: Calculate the quotient of the average lead time divided by the preset maximum effective lead time window, and then take the smaller value between the average lead time and the quotient as the normalized lead time score. The conditional probability is multiplied by the preset outcome probability weighting coefficient, the normalized lead time score is multiplied by the difference between the value and the outcome probability weighting coefficient, and the two products are added together to obtain the risk correlation score.

[0050] In this embodiment, optionally, the preset outcome probability weighting coefficient is set to 0.7, and the preset maximum effective advance time window is set to 24 hours. The quotient value obtained by dividing the average advance time by the preset maximum effective advance time window is: the normalized quotient value is obtained by dividing the average advance time by 24 hours. The conditional probability of an adverse outcome event occurring within the preset time window after the appearance of the corresponding candidate pattern is calculated as follows: the number of sequences containing the candidate pattern and whose corresponding adverse outcome event occurs within 24 hours is counted, and the result is divided by the total number of sequences containing the candidate pattern to obtain the conditional probability.

[0051] S6. For each candidate pattern, extract the last event type of the candidate pattern, perform an intervention query path search based on the pre-built medical knowledge graph, and assign an operability score according to the results of the intervention query path search.

[0052] In this embodiment, the medical knowledge graph includes at least abnormal event nodes, disease nodes, intervention measure nodes, examination item nodes, a first relationship indicating a disease from an abnormal event, a second relationship indicating an intervention measure corresponding to a disease, a third relationship indicating an examination item required for an abnormal event, and a fourth relationship indicating an intervention measure directly corresponding to an abnormal event.

[0053] Optionally, in this embodiment, the medical knowledge graph may also include drug nodes; the relationship types in the medical knowledge graph may also include a fifth relationship in which the intervention is prohibited under certain conditions; wherein, the fifth relationship connects the intervention node and the condition node, and is used to identify the contraindication information of the corresponding intervention under specific clinical conditions.

[0054] In this embodiment, the further step of performing the intervention measure query path search specifically includes: The first query path checks whether the last event type has an intervention node directly connected through the fourth relationship; The second query path queries whether the last event type has a disease node connected through the first relationship, and whether the disease node has an intervention measure node connected through the second relationship. The third query path checks whether the last event type has a check item node connected through a third relationship; Assigning operability scores, including: If the first query path is hit, the first operability score is assigned; If the first query path is not hit but the second query path is hit, then the second operability score is assigned. If only the third query path is hit, then the third operability score is assigned; If no query path is hit, a fourth operability score is assigned; Among them, the first operability score is greater than the second operability score, the second operability score is greater than the third operability score, and the third operability score is greater than the fourth operability score.

[0055] In this embodiment, optionally, the first operability score is set to 1.0, the second operability score is set to 0.7, the third operability score is set to 0.4, and the fourth operability score is set to 0.1; if the last event type of the candidate pattern belongs to the normal level, then regardless of the knowledge graph query result, the operability score of the candidate pattern is directly determined to be 0.

[0056] S7. For each candidate mode, the novelty score, risk correlation score, and operability score are weighted and fused to calculate the comprehensive value score. Candidate modes with a comprehensive value score greater than or equal to the preset value threshold are retained. After clustering and deduplication, they are sorted in descending order of comprehensive value score, and the quality control mode list is output.

[0057] In this embodiment, after clustering and deduplication, the data are sorted in descending order of comprehensive value score, specifically including: Hierarchical clustering is performed on all remaining candidate patterns based on temporal edit distance, and candidate patterns whose temporal edit distance between any two candidate patterns is less than a preset deduplication threshold are classified into the same pattern category. Within each pattern category, only the candidate pattern with the highest overall value score is retained as the representative pattern; Within the pattern category, representative patterns are arranged in descending order of their comprehensive value scores to form a quality control pattern list.

[0058] Optionally, in this embodiment, the weighted fusion of novelty score, risk correlation score, and operability score is performed as follows: the novelty score is multiplied by a first comprehensive weight of 0.3, the risk correlation score is multiplied by a second comprehensive weight of 0.5, and the operability score is multiplied by a third comprehensive weight of 0.2. The three products are then added together to obtain the comprehensive value score. The sum of the first, second, and third comprehensive weights is 1, and the weight of the risk correlation score is greater than the weights of the novelty score and the operability score. Specifically, the candidate patterns whose comprehensive value score is greater than or equal to a preset value threshold are retained as follows: the preset value threshold is set to 0.6. Candidate patterns with a comprehensive value score greater than or equal to 0.6 are retained. Before retaining candidate patterns whose comprehensive value score is greater than or equal to the preset value threshold, candidate patterns whose last event type belongs to the normal level and candidate patterns with a conditional probability less than 0.2 are removed.

[0059] Optionally, in this embodiment, the preset deduplication threshold is set to 0.2; candidate modes with a temporal edit distance of less than 0.2 between any two candidate modes are classified into the same mode category.

[0060] In this embodiment, one or more of the following can be optionally selected: a preset observation window, a minimum weighted support threshold, a minimum weighted confidence threshold, a first comprehensive weight, a second comprehensive weight, a third comprehensive weight, and a preset value threshold. These can be adapted and adjusted according to the characteristics of the target medical scenario. For example, in the intensive care unit or emergency department scenario, the minimum weighted support threshold can be reduced to 0.02 and the second comprehensive weight can be increased to 0.6 to improve the sensitivity of discovering rare but high-risk precursor patterns. In the chronic disease management scenario, the observation window can be extended to 7 days and the third comprehensive weight can be increased to 0.3 to prioritize the screening of time-series patterns with long-term intervention value. In the outpatient quality control scenario, the first comprehensive weight can be increased to 0.4 to prioritize the discovery of novel patterns that have not yet been covered by existing outpatient quality control rules.

[0061] In this embodiment, the standardized terminology coding used for event type standardization includes at least one of CTCAE adverse event terminology, LOINC test coding, and ICD-10 diagnostic coding; the medical knowledge graph is constructed based on a public medical knowledge base, which includes at least one of UMLS, SNOMED CT, and Chinese Medical Knowledge Graph CMeKG, and is expanded in combination with the hospital's own clinical pathways and quality control standards.

[0062] To further illustrate the effectiveness of the method described in this application, some examples are provided below to further explain its effects.

[0063] One example is the mining of prodromal patterns of sepsis.

[0064] Data from 12,000 patients in the ICU of a tertiary hospital over the past two years were extracted. Among them, there were 860 adverse outcome events with confirmed sepsis (meeting the Sepsis-3 criteria) and 3,000 control events (non-sepsis). The observation window was set at 72 hours. The event types included 12 symptomatic events, 23 laboratory events, 8 medical order events, and 5 outcome events.

[0065] The event severity weighting is configured as follows: (i) For critical events (lactic acid > 4 mmol / L, shock index > 1), the value is 1.0; (ii) Highly abnormal (PCT > 10 ng / mL, systolic blood pressure < 90 mmHg) was 0.8; (iii) Moderate abnormalities (body temperature > 38.5℃, WBC > 15×10) 9 / L) is 0.5; (iv) Mild abnormalities (body temperature 37.5-38.5℃, WBC 10-15×10) 9 The value of ( / L) is 0.3.

[0066] The first decay rate and the second decay rate in the time decay weight are 0.03 / h (0-24h) and 0.08 / h (24-72h), respectively; the weighted support is 0.03, the weighted confidence is 0.5, and the maximum pattern length used during mining is 5.

[0067] Table 1 shows a comparison of the mining results after performing weighted PrefixSpan mining with those of traditional PrefixSpan: Table 1 Comparison of Mining Results The most frequent pattern output by traditional PrefixSpan mining is "admission, temperature measurement, blood pressure measurement, nursing record recording" (support score 0.89, no quality control value), while the weighted PrefixSpan mining in this application effectively suppresses this common pattern.

[0068] Another example is the mining of prodromal patterns of postoperative bleeding.

[0069] For postoperative surgical patients, sequences of adverse outcomes such as postoperative bleeding were extracted from 320 cases and control sequences from 1500 cases. The observation window was 72 hours. The high-value patterns identified by weighted PrefixSpan analysis were "sudden increase in drainage volume (>100mL / h for 2 hours), decrease in Hb (Δ>15g / L / 4h), and increase in heart rate (>110bpm)".

[0070] Another example is the verification of the distinction between time sequences.

[0071] This study validates the role of temporal edit distance in novelty assessment. The existing rule is "elevated WBC, elevated body temperature, elevated PCT, bacterial infection warning." For candidate patterns P1 ("elevated body temperature, elevated WBC, elevated PCT") and P2 ("elevated WBC, elevated body temperature, elevated PCT") (WBC precedes body temperature, consistent with the rule), based solely on Jaccard similarity, candidate patterns P1 and P2 are judged as complete duplications. However, after introducing temporal edit distance, candidate pattern P1 retains some novelty due to the different temporal order. Clinical evaluation shows that the pattern of "elevated body temperature followed by elevated WBC" is more common in the early stages of viral infection, representing a novel pattern not covered by the existing rule, thus validating the effectiveness of temporal edit distance.

[0072] Compared with the prior art, the advantages of this application are as follows: This application aims to improve the accuracy of quality control pattern mining and achieve automatic focusing on high-value patterns. Through weighted PrefixSpan pattern mining, it can suppress the submergence of abnormal patterns by routine event patterns and enhance the clinical value of the output quality control patterns. This application ensures the reliability of weighted support calculation during recursive mining; it does not lose weight information due to increased recursion depth, ensuring that weighted PrefixSpan can be stably calculated for pattern mining at any pattern length. This application focuses on the time window of high-value prefix events before the outcome occurs, and the mined patterns have early warning attributes rather than post-event descriptive patterns. To avoid repeatedly discovering known patterns covered by existing rules, this application can distinguish pattern differences involving the same event type but different temporal sequences, and accurately identify novel patterns that have not yet been covered by the existing rule base. This application enables automatic judgment of the operability of quality control models. It can automatically determine whether there are corresponding clinical intervention measures for the abnormal events at the end of the model and assign operability scores accordingly. There is no need to manually mark the operability of each model, which improves the efficiency of model screening and ensures that the output quality control model has clear clinical intervention value. This application achieves quantitative ranking and deduplication output of quality control patterns. Through comprehensive screening and clustering deduplication mechanisms, the mining results are upgraded from simple high-frequency patterns to high-value quality control patterns that have undergone quantitative evaluation. The output list of quality control patterns can be directly used for quality control rule construction, reducing the cost of subsequent manual screening.

[0073] like Figure 2 As shown in the figure, this application discloses a pattern intelligent mining and screening system based on a medical quality control scenario, including: The time-series event construction module is used to receive patient diagnosis and treatment data from multi-source medical business systems, extract structured events and standardize event types. Using preset adverse outcome events as anchor points, it arranges event instances in ascending order of time within a preset observation window to construct a patient event sequence database. The event weight assignment module is used to assign an event severity weight to each event type based on the clinical severity of the event type. Using the occurrence time of the adverse outcome event as the anchor point, the time decay weight of each event instance is calculated based on the time interval between each event instance and the anchor point. The event severity weight is multiplied by the time decay weight to obtain the comprehensive weight of each event instance. The weighted prefixSpan mining module is used to set the minimum weighted support threshold and the minimum weighted confidence threshold, and then execute the weighted prefixSpan algorithm. During the recursive construction of the projection database, the weighted product of prefix pattern matching is propagated, the weighted support and weighted confidence of candidate patterns are calculated, and candidate patterns that satisfy the condition that the weighted support is greater than or equal to the minimum weighted support threshold and the weighted confidence is greater than or equal to the minimum weighted confidence threshold are retained as the weighted frequent time series pattern set. The temporal semantic novelty module is used to calculate the event set Jaccard similarity and temporal edit distance similarity between each candidate pattern in the weighted frequent temporal pattern set and each rule in the quality control rule base. The event set Jaccard similarity and temporal edit distance similarity are weighted and fused to obtain the comprehensive semantic similarity. The novelty score is calculated by taking the maximum comprehensive semantic similarity between the candidate pattern and all rules. The risk correlation module is used to calculate the conditional probability of an adverse outcome event occurring within a preset time window after the appearance of each candidate pattern, as well as the average time lead between the first complete match of the candidate pattern and the occurrence of the adverse outcome event. The risk correlation score is obtained by weighting and fusing the conditional probability and the average time lead. The knowledge graph operability module is used to extract the last event type of each candidate pattern, perform an intervention query path search based on the pre-built medical knowledge graph, and assign an operability score according to the results of the intervention query path search. The comprehensive screening module is used to calculate the comprehensive value score for each candidate mode by weighting and fusing the novelty score, risk correlation score, and operability score. Candidate modes with a comprehensive value score greater than or equal to a preset value threshold are retained. After clustering and deduplication, the modes are sorted in descending order of comprehensive value score and the quality control mode list is output.

[0074] In this embodiment, the optional operating environment for the deployment system is as follows: a solid-state drive with a central processing unit of no less than 16 cores, a memory of no less than 64GB, and a storage capacity of no less than 2TB; an operating system of Ubuntu 22.04; a graph database Neo4j for storing medical knowledge graphs; a relational database PostgreSQL 14 for storing patient event sequence databases; a computing framework Spark for performing massively parallel weighted PrefixSpan sequence pattern mining; a Python scientific computing stack for performing numerical calculations; and a backend SpringBoot and a frontend Vue 3.0 for providing interactive services.

[0075] The system described in the above embodiments is used to execute the corresponding pattern intelligent mining and screening method based on medical quality control scenarios in the foregoing embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0076] like Figure 3 As shown, based on the same application concept, and corresponding to any of the above embodiments, this application also discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the above-mentioned intelligent pattern mining and screening method based on medical quality control scenarios.

[0077] Specifically, the device includes: a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, memory 1020, input / output interface 1030, and communication interface 1040 are interconnected within the device via the bus 1050.

[0078] The processor 1010 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), GPU (Graphics Processing Unit), or one or more integrated circuits, to implement relevant programs and achieve the technical solutions provided in the embodiments of this specification.

[0079] The memory 1020 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1020 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1020 and called by the processor 1010. The input / output interface 1030 is used to connect input / output modules to realize information input and output. Input / output modules can be configured as components in the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices can include keyboards, mice, touch screens, microphones, various sensors, etc., and output devices can include displays, projectors, speakers, vibrators, indicator lights, etc.

[0080] The communication interface 1040 is used to connect a communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB (Universal Serial Bus), network cable, etc.) or wireless means (such as mobile network, WIFI (Wireless Fidelity), Bluetooth, etc.).

[0081] Bus 1050 includes a pathway for transmitting information between various components of the device, such as processor 1010, memory 1020, input / output interface 1030, and communication interface 1040.

[0082] It should be noted that although the above-described device only shows the processor 1010, memory 1020, input / output interface 1030, communication interface 1040, and bus 1050, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments of this specification, and not necessarily all the components shown in the figures.

[0083] The electronic devices described above are used to implement the corresponding intelligent pattern mining and screening method based on medical quality control scenarios in any of the foregoing embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0084] Based on the same application concept, and corresponding to any of the above embodiments, this application also discloses a computer-readable storage medium that stores computer instructions for enabling a computer to implement the above-described pattern intelligent mining and screening method based on a medical quality control scenario.

[0085] The computer-readable storage medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transfer medium, which can be used to store information accessible by a computing device. The computer instructions stored in the storage medium of the above embodiments are used to enable the computer to implement the pattern intelligent mining and screening method based on medical quality control scenarios as described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0086] The above description is merely a preferred embodiment and the technical principles employed in this application. This application is not limited to the specific embodiments or combinations thereof, and various obvious changes, readjustments, and substitutions that can be made by those skilled in the art will not depart from the scope of protection of this application. Therefore, although this application has been described in detail through the above embodiments, this application is not limited to the above embodiments, and may include more other equivalent embodiments without departing from the concept of this application, the scope of which is determined by the scope of the claims.

Claims

1. A pattern intelligent mining and screening method based on medical quality control scenarios, characterized in that, include: The system receives patient diagnosis and treatment data from multiple medical business systems, performs structured event extraction and event type standardization, uses preset adverse outcome events as anchor points, arranges event instances in ascending order of time within a preset observation window, and constructs a patient event sequence database. Each event type is assigned an event severity weight based on its clinical severity. The occurrence time of the adverse outcome event is used as an anchor point, and the time decay weight of each event instance is calculated based on the time interval between each event instance and the anchor point. The event severity weight is multiplied by the time decay weight to obtain the comprehensive weight of each event instance. Set a minimum weighted support threshold and a minimum weighted confidence threshold, and then execute the weighted PrefixSpan algorithm. During the recursive construction of the projection database, propagate the weighted product of prefix pattern matching, calculate the weighted support and weighted confidence of candidate patterns, and retain the candidate patterns that satisfy the condition that the weighted support is greater than or equal to the minimum weighted support threshold and the weighted confidence is greater than or equal to the minimum weighted confidence threshold as the weighted frequent time series pattern set. For each candidate pattern in the weighted frequent temporal pattern set, calculate the event set Jaccard similarity and temporal edit distance similarity between the candidate pattern and each rule in the quality control rule base. Weight the event set Jaccard similarity and the temporal edit distance similarity to obtain the comprehensive semantic similarity. Take the maximum value of the comprehensive semantic similarity between the candidate pattern and all rules to calculate the novelty score. For each candidate pattern, calculate the conditional probability of an adverse outcome event occurring within a preset time window after the candidate pattern appears, and the average time advance between the time of the first complete match of the candidate pattern and the time of occurrence of the adverse outcome event. The conditional probability and the average time advance are weighted and fused to obtain a risk correlation score. For each candidate pattern, the last event type of that candidate pattern is extracted, and an intervention query path search is performed based on a pre-built medical knowledge graph. An operability score is assigned based on the results of the intervention query path search. For each candidate mode, the novelty score, the risk correlation score, and the operability score are weighted and fused to calculate a comprehensive value score. Candidate modes with a comprehensive value score greater than or equal to a preset value threshold are retained. After clustering and deduplication, they are sorted in descending order of comprehensive value score, and a quality control mode list is output.

2. The pattern intelligent mining and screening method based on medical quality control scenarios according to claim 1, characterized in that, The time decay weight is calculated using a piecewise exponential decay method, including: For event instances where the time elapsed since the occurrence of the adverse outcome event does not exceed a first preset time, the time decay weight of the event instance is calculated using a first decay rate. For event instances where the time elapsed since the occurrence of the adverse outcome event exceeds the first preset duration but does not exceed the second preset duration, a baseline weight value up to the first preset duration is first calculated according to the first decay rate, and then the time decay weight of the event instance is calculated again starting from the baseline weight value and using the second decay rate. Wherein, the first preset duration is less than the second preset duration, and the first attenuation rate is less than the second attenuation rate; For patient event sequences where no adverse outcome events occurred, the time decay weights of all corresponding event instances were set to the preset baseline weight values.

3. The pattern intelligent mining and screening method based on medical quality control scenarios according to claim 1, characterized in that, Calculate the weighted support of candidate patterns, including: For any candidate pattern, traverse each patient event sequence in the patient event sequence database and determine whether the patient event sequence contains the corresponding candidate pattern. If it does, obtain the comprehensive weight of each event instance in the patient event sequence that matches the corresponding candidate pattern, and multiply the comprehensive weights of each event instance to obtain the sequence matching weight of the corresponding candidate pattern in the patient event sequence. Sum the sequence matching weights of all patient event sequences that contain the corresponding candidate pattern, and divide the sum by the total number of sequences in the patient event sequence database to obtain the weighted support of the candidate pattern. The propagation of the weighted product of prefix pattern matching during the recursive construction of the projection database includes: When constructing a projection database for the current prefix pattern, for each patient event sequence containing the prefix pattern, the last matching position of the prefix pattern in the patient event sequence is located, and all events after the matching position constitute the projection sequence corresponding to the patient event sequence. The corresponding sequence matching weight is then stored as a projection weight associated with the projection sequence. When expanding candidate events in the projection database, for each expanded candidate event in the projection sequence, the projection weight is multiplied by the comprehensive weight of the event instance to which the expanded candidate event belongs, to obtain the weighted support contribution value of the expanded candidate pattern.

4. The pattern intelligent mining and screening method based on medical quality control scenarios according to claim 1, characterized in that, Calculating temporal edit distance similarity includes: Align the sequence of candidate patterns and the event types corresponding to the rules, and use the insertion, deletion, and replacement operations of the event types as editing operations to calculate the minimum editing cost required to convert the candidate patterns into rules. Among them, the unit cost of the insertion operation and the deletion operation are equal, and the unit cost of the replacement operation is set to different cost values ​​according to whether the replaced event type belongs to the same clinical category. The replacement cost within the same clinical category is less than the replacement cost across clinical categories. In the process of calculating the minimum editing cost, if there is a case where the candidate pattern and the rule have the same event type but the time sequence is reversed, a preset reversal penalty cost is added on the basis of the minimum editing cost. The temporal edit distance similarity is calculated based on the larger of the minimum edit cost, the number of corresponding event instances in the candidate pattern and the rule; Calculating comprehensive semantic similarity includes: The comprehensive semantic similarity is obtained by summing the Jaccard similarity and temporal edit distance similarity of the event set according to preset weights.

5. The pattern intelligent mining and screening method based on medical quality control scenarios according to claim 1, characterized in that, Calculating the average time lead includes: In each patient event sequence containing the corresponding candidate pattern and where an adverse outcome event has occurred, the matching timestamp of the last event of the corresponding candidate pattern is obtained, and the difference between the occurrence timestamp of the adverse outcome event and the matching timestamp is calculated in each patient event sequence where an adverse outcome event has occurred, to obtain the individual lead time. The average time advance is obtained by summing the individual advances of all patient event sequences that resulted in adverse outcomes and dividing by the total number of patient event sequences that resulted in adverse outcomes. The conditional probability and the average time lead are weighted and fused, specifically including: Calculate the quotient of the average lead time divided by the preset maximum effective lead time window, and then take the smaller value between the value and the quotient as the normalized lead time score. The conditional probability is multiplied by a preset outcome probability weighting coefficient, the normalized lead time score is multiplied by the difference between the value and the outcome probability weighting coefficient, and the two products are added together to obtain the risk correlation score.

6. The pattern intelligent mining and screening method based on medical quality control scenarios according to claim 1, characterized in that, The medical knowledge graph includes at least abnormal event nodes, disease nodes, intervention measure nodes, examination item nodes, a first relationship where abnormal events indicate diseases, a second relationship where diseases correspond to intervention measures, a third relationship where abnormal events require examination items, and a fourth relationship where abnormal events directly correspond to intervention measures. The query path search for the intervention measures specifically includes: The first query path checks whether the last event type has an intervention measure node that is directly connected through the fourth relationship; The second query path queries whether the last event type has a disease node connected through the first relationship, and whether the disease node has an intervention measure node connected through the second relationship; The third query path queries whether the last event type exists as the check item node connected through the third relationship; Assigning operability scores, including: If the first query path is hit, then the first operability score is assigned; If the first query path is not hit but the second query path is hit, then a second operability score is assigned. If only the third query path is hit, then a third operability score is assigned; If no query path is hit, a fourth operability score is assigned; Among them, the first operability score is greater than the second operability score, the second operability score is greater than the third operability score, and the third operability score is greater than the fourth operability score.

7. The pattern intelligent mining and screening method based on medical quality control scenarios according to claim 1, characterized in that, The sorting of data after clustering and deduplication, in descending order of comprehensive value score, specifically includes: Hierarchical clustering is performed on all remaining candidate patterns based on temporal edit distance, and candidate patterns whose temporal edit distance between any two candidate patterns is less than a preset deduplication threshold are classified into the same pattern category. In each of the aforementioned pattern categories, only the candidate pattern with the highest comprehensive value score is retained as the representative pattern; The representative models in the model category are arranged in descending order of their comprehensive value scores to form the quality control model list.

8. A pattern intelligent mining and screening system based on medical quality control scenarios, characterized in that, include: The time-series event construction module is used to receive patient diagnosis and treatment data from multi-source medical business systems, extract structured events and standardize event types. Using preset adverse outcome events as anchor points, it arranges event instances in ascending order of time within a preset observation window to construct a patient event sequence database. The event weight assignment module is used to assign an event severity weight to each event type according to the clinical severity of the event type. Taking the occurrence time of the adverse outcome event as the anchor point, the module calculates the time decay weight of each event instance based on the time interval between each event instance and the anchor point. The module multiplies the event severity weight by the time decay weight to obtain the comprehensive weight of each event instance. The weighted PrefixSpan mining module is used to set a minimum weighted support threshold and a minimum weighted confidence threshold, and then execute the weighted PrefixSpan algorithm. During the recursive construction of the projection database, the weighted product of prefix pattern matching is propagated, the weighted support and weighted confidence of candidate patterns are calculated, and the candidate patterns that satisfy the condition that the weighted support is greater than or equal to the minimum weighted support threshold and the weighted confidence is greater than or equal to the minimum weighted confidence threshold are retained as a set of weighted frequent time series patterns. The temporal semantic novelty module is used to calculate the event set Jaccard similarity and temporal edit distance similarity between each candidate pattern in the weighted frequent temporal pattern set and each rule in the quality control rule base. The event set Jaccard similarity and the temporal edit distance similarity are weighted and fused to obtain a comprehensive semantic similarity. The maximum value of the comprehensive semantic similarity between the candidate pattern and all rules is taken to calculate the novelty score. The risk correlation module is used to calculate, for each candidate pattern, the conditional probability of an adverse outcome event occurring within a preset time window after the candidate pattern appears, and the average time advance between the time of the first complete match of the candidate pattern and the time of occurrence of the adverse outcome event. The risk correlation score is obtained by weighting and fusing the conditional probability and the average time advance. The knowledge graph operability module is used to extract the last event type of each candidate pattern, perform an intervention query path search based on a pre-built medical knowledge graph, and assign an operability score according to the result of the intervention query path search. The comprehensive screening module is used to calculate a comprehensive value score for each candidate mode by weighting and fusing the novelty score, the risk correlation score, and the operability score. Candidate modes with a comprehensive value score greater than or equal to a preset value threshold are retained. After clustering and deduplication, the modes are sorted in descending order of comprehensive value score and a quality control mode list is output.

9. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the intelligent pattern mining and screening method based on medical quality control scenarios as described in any one of claims 1-7.

10. A storage medium, which is a computer-readable storage medium, characterized in that, The device stores computer instructions, which are used to enable the computer to implement the intelligent pattern mining and screening method based on medical quality control scenarios as described in any one of claims 1-7.

Images