Abnormal information identification method and system for physical examination report

Abstract

The application discloses a physical examination report abnormal information identification method and system, acquires report data carrying field mode and result text and completes hierarchical expression normalization; according to the structure characteristic weight cumulative value of the field mode, text form and item name semantics, the report is distributed to a test value channel or an image semantic channel; in the image semantic channel, a semantic condition tree taking a semantic condition object as a node and a root-to-leaf path as a rule is used to generate a candidate and a character position interval, and then through non-detachable term protection, continuity screening, inclusion deduplication, punctuation symbol sentence isolation processing and combined negative filtering, a standardized abnormal information label is output. The application can adapt to different report structures without training samples, and reduces error filtering caused by mixed positive description and negative description in the same paragraph.

Description

Technical Field

[0001] This invention relates to the field of medical text data processing and medical information system data standardization technology, and in particular to a method and system for identifying abnormal information in physical examination reports applied to the scenario of physical examination report data processing. Background Technology

[0002] The report data received by the physical examination information system typically comes from laboratory information systems, medical image storage and transmission systems, and internal business systems from different vendors. Laboratory data is usually represented by fields such as item name, result value, unit, and reference range, while imaging data is often represented by the examination item name and longer natural language text results. To facilitate subsequent statistics, retrieval, and information display, it is necessary to convert abnormal information from different sources and with different expressions into unified data tags.

[0003] Existing text matching methods typically maintain a keyword list, independent regular expressions, or character-level prefix indexes for each tag. With the increase in hierarchical systems, synonyms, and organ descriptions, the number of repetitive rules rises with the number of participating institutions, making it impossible to uniformly reuse common semantic prefixes. Another method uses a fixed-word-count window to identify negative descriptions, but the negative scope in Chinese reports is often separated by commas, semicolons, or periods, making it difficult for a fixed window to simultaneously accommodate adjacent clauses and long modifiers.

[0004] In existing technologies, a set of rules is first selected based on the project name, then the target word is searched in the entire result text, and whether to retain it is determined by whether a negative word appears near the target word. This approach relies on pre-labeled report categories or fixed department names, and uses character distance as the negative scope, which cannot utilize the stable information carried by the report field structure and punctuation boundaries; at the same time, each rule is stored independently, and modifications to common conditions still need to be synchronized one by one.

[0005] Therefore, existing technologies present at least two interrelated technical problems: First, how to leverage the inherent stability of the report data structure to automatically distribute reports across sources and reduce redundant storage and calculation of common matching conditions as rules continue to expand; Second, how to ensure that negative filtering applies to the correct clause when positive and negative descriptions are located in the same result text, while protecting technical terms containing words such as "none" or "low" but used as a whole, thereby reducing the accidental deletion and omission of candidate information. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a method and system for identifying abnormal information in physical examination reports, so as to separate test reports and imaging reports without training a classification model, reuse common semantic conditions of different rules, and constrain the scope of negation semantics based on the clause boundaries formed by punctuation.

[0007] To address the aforementioned technical problems, this invention provides a method for identifying abnormal information in physical examination reports, applicable to physical examination report data processing scenarios, including:

[0008] Obtain the medical examination report data to be processed, which includes field patterns and result text; normalize the grading symbols and grading system names in the result text to obtain standardized report data.

[0009] The report structure features are constructed based on the field pattern, the text form of the result text, and the semantics of the inspection item name. The report type is determined by the cumulative weight value of the report structure features, and the standardized report data is allocated to the inspection value channel or the image semantic channel.

[0010] In the numerical inspection channel, candidate anomalies are obtained based on the result value and the reference range; in the semantic image channel, the result text is matched using a semantic condition tree, where the nodes of the semantic condition tree are semantic condition objects and the root-to-leaf path is the matching rule, to obtain candidate image anomalies and character position ranges.

[0011] Based on the character position range, the image anomaly candidates are subjected to indivisible term protection, alphanumeric continuity filtering, inclusion relationship deduplication, punctuation sentence isolation processing, and combined negation filtering to obtain retained image anomaly candidates;

[0012] The test anomaly candidates and the retained image anomaly candidates are summarized, and standardized anomaly information labels are output.

[0013] Further, the process of constructing report structure features based on the field pattern, the text morphology of the result text, and the semantics of the inspection item name, determining the report type using the cumulative weight value of the report structure features, and allocating the standardized report data to the inspection value channel or the image semantic channel includes: the existence status of the detection unit field, the reference range field, and the anomaly indication field to obtain field existence features; extracting the character length, number format status, and positive mark status of the result text to obtain text morphology features; matching the inspection item name with a preset image item vocabulary to obtain item semantic features; multiplying and summing the field existence features, the text morphology features, and the item semantic features with their corresponding directional weights to obtain a report type score; comparing the report type score with a diversion threshold, allocating the report type score to the image semantic channel when it is located on the image side of the diversion threshold, and allocating it to the inspection value channel when it is located on the inspection side of the diversion threshold.

[0014] Further, the step of matching the result text using a semantic condition tree in the image semantic channel, where the nodes of the semantic condition tree are semantic condition objects and the root-to-leaf path is the matching rule, to obtain image anomaly candidates and character position ranges, includes: reading standardized anomaly information tags and corresponding ordered condition object sequences from a rule database; configuring the condition objects in the ordered condition object sequence as keyword condition objects, set condition objects, or regular expression condition objects, and encapsulating unified matching processing steps for different types of condition objects, wherein the unified matching processing steps are matching processing steps that take standardized result text as input and satisfy status, hit content, and character position set as output; inserting nodes level by level from the root node of the semantic condition tree according to the ordered condition object sequence, merging common prefix nodes with the same content and type, and binding the standardized anomaly information tags to the last-level nodes; traversing the semantic condition tree, calling the unified matching processing steps for the child nodes of the current node, and pruning subtrees that do not satisfy the corresponding condition objects; when reaching a leaf node bound with the standardized anomaly information tags, recording the character position hit by the semantic condition path corresponding to the leaf node, to obtain the image anomaly candidates and corresponding character position ranges.

[0015] Further, the traversal of the semantic condition tree, calling the unified matching processing step on the child nodes of the current node, and pruning subtrees that do not meet the corresponding condition objects, includes: establishing a path state containing the current node identifier, the met condition objects, and the matched character positions; performing a depth-first search on each child node of the current node according to the child node priority; recording the character position of the string when the string corresponding to the keyword condition object is contained in the result text; recording the matched candidate string and its corresponding character position when at least one candidate string contained in the set condition object is contained in the result text; recording the matching interval when the regular expression corresponding to the regular expression condition object matches the result text; appending the corresponding match information to the path state and entering the next level node; restoring the path state before entering the current path and searching for sibling nodes when the current path does not meet any child node; and determining the leaf nodes to be retained according to at least one of the conflict resolution conditions: leaf node priority, path length, or label priority.

[0016] Further, based on the character position interval, the process of performing indivisible term protection, alphanumeric continuity filtering, inclusion relationship deduplication, punctuation sentence isolation processing, and combined negation filtering on the image anomaly candidates to obtain retained image anomaly candidates includes: locating the term position interval of the preset indivisible term in the result text, deleting image anomaly candidates whose character position interval partially overlaps with the term position interval and does not completely cover the term; when the image anomaly candidate hits a string of alphanumeric characters, calculating the difference between adjacent hit character positions, and deleting image anomaly candidates whose difference is not one; comparing the inclusion relationship of the character position sets of any two image anomaly candidates, and deleting image anomaly candidates whose character position set is contained in another character position set; dividing the result text into clauses based on punctuation positions, independently detecting negation trigger words within each clause, and deleting image anomaly candidates associated with clauses containing negation trigger words; combining the set of negation words with the set of positive descriptive words to generate a combined negation pattern, and deleting image anomaly candidates that hit the combined negation pattern to obtain the retained image anomaly candidates.

[0017] Further, the step of dividing the result text into clauses based on punctuation positions, independently detecting negative trigger words within each clause, and deleting image anomaly candidates associated with clauses containing negative trigger words includes: identifying Chinese commas, semicolons, and periods in the result text to form a punctuation boundary sequence arranged in ascending order of character position; forming corresponding clauses based on the character intervals between two adjacent punctuation boundaries and establishing a mapping between clause identifiers and character intervals; comparing the character position interval of each image anomaly candidate with the mapping to determine at least one clause identifier associated with the image anomaly candidate; matching preset negative trigger words within each clause and excluding negative characters located within indivisible terms; deleting the image anomaly candidate when it is associated with only one clause and the clause contains a valid negative trigger word; deleting the image anomaly candidate when it spans multiple clauses and any associated clause contains a valid negative trigger word; and retaining image anomaly candidates whose associated clauses do not contain valid negative trigger words.

[0018] Furthermore, before and after matching the result text using the semantic condition tree, a progressive cascaded matching strategy is configured. This progressive cascaded matching strategy includes: pre-normalizing Roman numerals, hierarchical letters, and hierarchical system names in the result text; performing exact matching between the pre-normalized check item names and the alias field in the search index; using the semantic condition tree for matching when the exact match fails; performing full-text search on the result text when the semantic condition tree fails, and performing double filtering based on full-text search score and edit distance similarity; performing word-by-word sequential matching when the double filtering fails; performing hierarchical entity-specific recognition on result text containing preset hierarchical system names; and determining the image anomaly candidates according to the first valid hit or a preset priority level.

[0019] Further, the step of performing word-by-word sequential matching when the double screening fails includes: splitting the target keyword into a target character sequence that maintains its original order; establishing a set of occurrence positions of each target character in the target character sequence in the result text; terminating the word-by-word sequential matching of the corresponding target keyword when the occurrence count of any target character exceeds a threshold; recursively selecting character positions from each set of occurrence positions to form candidate position paths; deleting candidate position paths that do not maintain the original order of the target character sequence; performing character boundary verification on the last character when the target keyword ends with a letter or number; and determining the candidate position path that passes the character boundary verification and covers all target characters as the word-by-word sequential matching result.

[0020] Further, the step of performing hierarchical entity-specific recognition on the result text containing the name of the preset hierarchical system includes: reading hierarchical rule records containing standardized anomaly information tags, keyword sequences, order attributes, and hierarchical values; when the order attribute is ordered, converting the keyword sequence into an ordered regular expression pattern connected in order, wherein the ordered regular expression pattern is a regular expression connecting adjacent keywords with a non-greedy wildcard constrained by the maximum character span; when the order attribute is unordered, converting the keyword sequence into an unordered regular expression pattern containing multiple zero-width assertions, wherein the unordered regular expression pattern is where each keyword is connected by a non-greedy wildcard constrained by the maximum character span. A positive zero-width assertion form is combined with a regular expression within the same clause range; forward and backward numeric boundaries are added to the hierarchical value, the forward numeric boundary is used to exclude cases where the value is immediately adjacent to the preceding number, and the backward numeric boundary is used to exclude cases where the value is immediately adjacent to the following number or forms a numerical range with the following number, thus obtaining a hierarchical boundary pattern; the ordered or unordered regular expression pattern is combined with the hierarchical boundary pattern to obtain a hierarchical entity matching pattern; the hierarchical entity matching pattern is used to process the result text, and the standardized anomaly information label corresponding to the hit record is determined as the hierarchical entity recognition result.

[0021] This invention also provides a system for identifying abnormal information in physical examination reports, comprising: a report normalization module, used to acquire physical examination report data to be processed, containing field patterns and result text, and normalize the grading symbols and grading system names in the result text to obtain standardized report data; a report distribution module, used to construct report structure features based on the field patterns, the text form of the result text, and the semantics of the examination item names, determine the report type using the cumulative weight value of the report structure features, and allocate the standardized report data to the test value channel or the image semantic channel; and a candidate generation module, used to generate candidates based on the result value and... The reference range is used to obtain candidate images for inspection; the semantic condition tree is used to match the result text in the image semantic channel, where the nodes of the semantic condition tree are semantic condition objects and the root-to-leaf path is the matching rule, to obtain candidate images for inspection and character position ranges; the candidate filtering module is used to perform non-separable term protection, alphanumeric continuity filtering, inclusion relation deduplication, punctuation sentence isolation processing and combined negation filtering on the candidate images for inspection based on the character position range, to obtain retained candidate images for inspection; the label output module is used to summarize the candidate images for inspection and the retained candidate images for inspection, and output standardized anomaly information labels.

[0022] Since the report type is determined by field existence, text format, and project name semantics, the traffic splitting is based on the report data pattern itself, without relying on a classification model retrained according to the access institution, thus reducing the adaptation cost of new data sources. Because the semantic condition tree merges common semantic conditions into shared prefixes and prunes the entire subtree when a prefix is ​​not matched, the number of repeated storage and invalid matches required for rule expansion is reduced. Since negation filtering first establishes a mapping between punctuation clauses and candidate character positions, and then identifies valid negation trigger words within the clauses, negation semantics will not be discovered across adjacent characters due to a fixed character window; combined with indivisible terminology protection, it can reduce the false deletion of candidates caused by single words in technical terms being mistakenly used as negation trigger words. Attached Figure Description

[0023] Figure 1 A flowchart illustrating the method for identifying abnormal information in a physical examination report provided in an embodiment of the present invention;

[0024] Figure 2 A flowchart illustrating the report structure feature routing steps provided in an embodiment of the present invention;

[0025] Figure 3 A schematic diagram illustrating the principle of semantic condition tree construction and matching steps provided in an embodiment of the present invention;

[0026] Figure 4 A schematic diagram of the five-layer negative semantic filtering pipeline provided in an embodiment of the present invention;

[0027] Figure 5 A flowchart illustrating the progressive cascading matching strategy provided in an embodiment of the present invention;

[0028] Figure 6 This is a module structure diagram of the abnormal information recognition system for physical examination reports provided in an embodiment of the present invention. Detailed Implementation

[0029] The present invention will be further described below with reference to the accompanying drawings and embodiments. The following embodiments are used to illustrate the technical solution of the present invention and do not constitute a limitation on the scope of protection. The physical examination report data to be processed can be provided by interface messages, database records or batch files; the rule database, retrieval index and thesaurus can be deployed on the same server, or they can be deployed separately and accessed through internal interfaces.

[0030] Example 1

[0031] like Figure 1 As shown, this embodiment provides a method for identifying abnormal information in a physical examination report, which is applied to the scenario of physical examination report data processing, and includes the following steps.

[0032] S1. Obtain the medical examination report data to be processed, which includes field patterns and result text, and normalize the grading symbols and grading system names in the result text to obtain standardized report data;

[0033] The field pattern records at least the existence status of the check item name field, result text field, result value field, unit field, reference range field, and exception message field. During normalization, a character substitution table and a protection mode table are created, converting full-width characters to half-width characters, converting Unicode Roman numerals and ASCII Roman numerals that satisfy the hierarchical context to Arabic numerals, and unifying spaces, hyphens, and case differences in hierarchical system names to a preset format. The protection mode table is used to maintain sub-level expressions composed of numbers and letters and configure them as non-convertible item abbreviations. The normalization module also saves the character offset mapping from the original text to the normalized text so that subsequent matching results can return to the original text position.

[0034] Specifically, each conversion rule in the character replacement table includes the source character pattern, target character pattern, forward context condition, backward context condition, and rule priority. For Roman numerals that may represent both hierarchical values ​​and item abbreviations, the system first checks for the presence of "class," "level," hierarchical system name, or sub-level letters before and after the numeral. If the hierarchical context is satisfied, the conversion is performed; otherwise, the original character is preserved. For expressions like "4A" and "4B" where numbers and letters are adjacent, the numeric and alphanumeric parts are treated as a single protected unit to prevent case normalization or numeric boundary handling from separating them. Repeated execution of the same normalization operation should yield the same result to ensure that message retries and batch reprocessing do not alter character positions.

[0035] In this embodiment of the invention, character offset mapping can be represented by a segment mapping table. Each segment records the start and end positions of the original text, the start and end positions of the normalized text, and the conversion rule identifier. Single-character equal-length replacements directly reuse the position; space deletion, character merging, or character expansion results in a new segment record. After obtaining the character position range in the normalized text, it is converted back to the original text evidence range based on the segment mapping table. If the input contains unparseable control characters, they can be replaced with uniform placeholders and a data quality mark recorded; the data quality mark is only used for log processing and review and does not participate in report structure feature scoring.

[0036] S2. Construct report structure features based on the field pattern, the text form of the result text, and the semantics of the inspection item name; determine the report type using the cumulative weight value of the report structure features; and allocate the standardized report data to the inspection value channel or the image semantic channel.

[0037] The report's structural features consist of field existence features, text morphology features, and item semantic features. Field existence features reflect whether the data source simultaneously provides units, reference ranges, and anomaly warnings; text morphology features reflect whether the result text is a short numerical value, whether it reaches the preset length, and whether it contains structured tags; item semantic features reflect whether the item name contains image examination category terms. Each feature uses a unified binary or segmented value encoding. The positive direction of the directional weight is defined as the image side, and the negative direction is defined as the examination side. The reverse can also be defined, as long as the scoring table and the triage comparison rules remain consistent.

[0038] The report structure features are a set of deterministic features directly extracted from the report data patterns, without the need to train a classifier based on the report text first. Field presence status corresponds to the data structure layer, text length, number format status, and positive marker status correspond to the data morphology layer, and the matching status of the checked item name with the preset image item vocabulary corresponds to the name semantic layer. These three layers of features complement each other: when a data source omits the unit field, the reference range field, the result text number format, and the item semantics can still jointly provide a basis for traffic allocation; when the item name uses a custom format defined by the accessing institution, the field presence status and text morphology can still maintain the stability of the traffic allocation results.

[0039] Report type rating can be expressed as:

[0040] Equation (1)

[0041] in, This indicates the report type score for the currently pending medical examination report data; This indicates the number of report structural features involved in the scoring; Indicates the first The directional weights corresponding to the structural features of each report; Indicates the first The values ​​of the report structure features; The symbol in Equation (1) represents the report structure feature number. The symbols in Equation (1) are only used for report type scoring. Report type scoring does not require the use of positive or negative zero as a fixed threshold. During deployment, the triage threshold can be calibrated based on known interface samples, but the calibration process does not train the text classification model.

[0042] In one example, the unit field, reference range field, and anomaly message field are given negative weights, as are purely numeric result features. Result text exceeding fifty characters and item names matching the image terminology are given positive weights. If a report contains complete units and reference ranges and the result value is a short number, its cumulative value falls on the inspection side; if a report lacks units and reference ranges, the result text is eighty-seven characters long, and the item name matches the image terminology, its cumulative value falls on the image side. Using structural features rather than institution names for data routing ensures that new access interfaces can still enter the corresponding processing channel according to the data pattern even if the department name changes.

[0043] Optionally, the scoring table can be established using a small number of known interface samples for rule calibration. However, the calibration only applies to directional weights and splitting thresholds, without involving parameter training of the text classification model. During calibration, the proportion of each report's structural features in the test and image samples is statistically analyzed. Features that stably point to a certain channel are assigned larger absolute weights, while features that may appear in both types of reports are assigned smaller absolute weights. After calibration, the scoring table version is frozen, and each splitting result is bound to the version number. Thus, even if the scoring table is updated subsequently, the splitting process can still be reproduced based on the historical version.

[0044] S3. In the test numerical channel, test anomaly candidates are obtained based on the result value and the reference range; in the image semantic channel, the result text is matched using a semantic condition tree, where the nodes of the semantic condition tree are semantic condition objects and the root-to-leaf path is the matching rule, to obtain image anomaly candidates and character position ranges.

[0045] In the numerical verification channel, the result value is parsed into a numerical value or a range, and the reference range is parsed into a lower limit, upper limit, and open / closed interval identifier. When the result value falls outside the reference range or the anomaly indication field meets the preset anomaly flag, a verification anomaly candidate carrying the item identifier, deviation direction, and original position is generated. In the image semantic channel, the semantic condition tree resides in memory. The matching engine only accesses the unified matching processing steps and does not depend on which implementation—string inclusion, set selection, or regular expression boundary matching—is used within the node. Leaf node labels are generated only when all semantic condition paths are satisfied, and the set of character positions matched by each condition object is recorded.

[0046] For a path containing three semantic condition objects, its path hit state can be represented as:

[0047] Equation (2)

[0048] in, Indicates the first The hit status of semantic conditional paths; Indicates the first The number of nodes contained in a semantic conditional path; Indicates the first The first path The matching function for each semantic condition object outputs one if the condition is met, and zero if the condition is not met. Indicates sending to the first The normalized result text of the semantic conditional path; Indicates the semantic conditional path sequence number; Indicates the node number within the path. Only... Only paths with a value of 1 will have their leaf node labels output. Since outputting zero for any prefix condition will result in a zero product, the search process can be pruned immediately without needing to visit all subsequent conditions under that prefix.

[0049] S4. Based on the character position range, perform indivisible term protection, alphanumeric continuity filtering, inclusion relationship deduplication, punctuation sentence isolation processing, and combined negation filtering on the image anomaly candidates to obtain retained image anomaly candidates;

[0050] Candidate filtering is performed in a fixed order. Indivisible term protection first determines the overall boundary of technical terms to prevent subsequent layers from interpreting characters within terms individually; alphanumeric continuity filtering excludes abbreviations composed of scattered positions; inclusion relationships remove duplicates, retaining candidates with more complete coverage information; punctuation-based clause isolation establishes a one-to-one or one-to-many attribution between candidate positions and clauses; combined negation filtering handles "negative word—intermediate modifying text—positive descriptive word" structures not separated by punctuation. Each layer receives the retained set from the previous layer and outputs a new retained set, allowing each layer to be replaced independently without altering the candidate data interface.

[0051] When the preceding clause of the same result text contains a positive description and the following clause contains a negative trigger word, the fixed window may cross the comma and delete candidates from the preceding clause. This method first calculates the clause containing the candidate character position, and then only detects valid negative trigger words in the related clauses; the indivisible term list is also used to mask the words "none" or "low" located inside complete technical terms. Thus, the writing structure carried by punctuation is transformed into a computable positional constraint, reducing accidental deletion across clauses.

[0052] S5. Summarize the candidate abnormalities from the inspection and the candidate abnormalities from the retained images, and output standardized abnormality information labels.

[0053] Before outputting tags, duplicates can be removed based on project identifier, standard tag code, and character position. Each tag is then appended with its source channel, hit rule version, and original text evidence range. Standardized anomaly information tags are the information processing results of the report text and can be used in statistical, retrieval, and review interfaces or subsequent information systems. This embodiment does not use the standardized anomaly information tags to directly form individual-based subsequent handling conclusions.

[0054] Example 2

[0055] like Figure 2 As shown, this embodiment further explains the report structure feature diversion.

[0056] The process of constructing report structure features based on the field pattern, the text form of the result text, and the semantics of the inspection item name, determining the report type using the cumulative weight value of the report structure features, and allocating the standardized report data to the inspection numerical channel or the image semantic channel includes:

[0057] S2.1. Detect the existence status of the unit field, reference range field, and exception message field to obtain the field existence characteristics;

[0058] Field checks are primarily performed based on the key-name mapping table of the data interface. When different systems use different names for the same field, the original field name is first mapped to a unified field identifier before determining whether the field exists and whether its value is valid. Empty strings, strings containing only placeholders, and explicit null values ​​are all considered as non-existent fields.

[0059] Specifically, the key-name mapping table can be grouped by data source identifier and interface version. Each group is configured with candidate keys for unit fields, reference range fields, and exception message fields. Field existence characteristics not only distinguish between "existence" and "non-existence," but also between three states: "field exists and its value is valid," "field exists but its value is empty," and "field does not exist," each encoded with a different value. The validity of reference range fields is determined by whether at least one boundary value can be parsed; the validity of unit fields is determined by whether content remains after removing whitespace; and the validity of exception message fields is determined by whether a controlled message flag is hit. This multi-state encoding prevents reserved empty fields from being mistakenly interpreted as valid structural evidence.

[0060] S2.2. Extract the character length, number format status, and positive marker status of the resulting text to obtain text morphological features;

[0061] Character length can be calculated based on Unicode code points; the numeric format status is determined by a regular expression that allows plus or minus signs, decimal points, and scientific notation; the positive flag status is determined by a controlled flag vocabulary. Character length can be segmented using multiple segment thresholds of ten, twenty, and fifty characters, with the thresholds adjustable based on the interface language and average text length.

[0062] Furthermore, text morphological feature extraction is performed after removing leading and trailing whitespace, but punctuation within the resulting text is not deleted to avoid disrupting subsequent punctuation-based sentence segmentation. Numerical format states can include four types: pure numeric, numeric with comparison operators, numeric range, and non-numeric text. For expressions like "<5.0" or "3.5-6.0", the parser preserves the comparison direction or range boundaries, respectively. Positive marker vocabulary and negative word sets are maintained separately; the former is used only for report routing, while the latter is used only for candidate filtering, avoiding the same term playing a confusing semantic role at different processing stages.

[0063] S2.3. Match the names of the inspection items with a preset image item vocabulary to obtain the semantic features of the items;

[0064] A pre-defined image item vocabulary stores general inspection category terms and their standardized spellings. Before matching, the item names are normalized for capitalization, spaces, and hyphens. This can be implemented using string inclusion or prefix matching; in other implementations, a configurable finite state automaton can also be used, but the output will still be the semantic features of the item.

[0065] Each record in the preset image item vocabulary can contain standard words, acceptable variations, excluded words, and priorities. The item name is checked first against excluded words, then against standard words and acceptable variations to prevent a short abbreviation from generating semantic features when it is part of another long word. For letter abbreviations, character boundaries are checked during matching; for Chinese category words, the longest string is prioritized. The vocabulary version and the scoring table version are recorded separately, so that updating the item semantic vocabulary does not require changing existing directional weights simultaneously.

[0066] S2.4. Multiply the field existence feature, the text morphology feature, and the project semantic feature by their corresponding directional weights and sum them up to obtain the report type score;

[0067] Directional weights are stored in a versioned scoring table. Each record includes a feature identifier, feature value range, weight value, and effective version. The simultaneous presence of all three fields allows for the setting of independent combined weights to express a stronger indication of the test case data. Older versions are retained after changes to the scoring table, enabling the reproduction of historical processing results.

[0068] It is worth noting that a single report structure feature does not directly determine the report type. The scorer first finds the corresponding weight for each valid feature, and then combines the feature weights; missing features are not automatically assigned weights in the opposite direction unless a "missing status" weight is explicitly configured in the scoring table. This processing method can distinguish between the two different meanings of "a field is not provided" and "it has been confirmed that it does not belong to a certain structure," avoiding overly strong scoring due to interface field pruning. During the scoring process, a contribution detail is generated simultaneously. This contribution detail records the feature identifier, feature value, and contribution score, which is used to explain the formation process of the report type score in the review interface.

[0069] S2.5. Compare the report type score with the triage threshold, and assign the report type score to the image semantic channel when it is located on the image side of the triage threshold, and assign the report type score to the test value channel when it is located on the test side of the triage threshold.

[0070] When the score equals the triage threshold, a deterministic decision rule can be set based on whether both the unit field and the reference range field are valid simultaneously. If both are valid, the data will be prioritized for the verification numerical channel; otherwise, it will be prioritized for the image semantic channel. This rule is only used for boundary value disambiguation and does not change the main triage mechanism weighted by structural features.

[0071] In a more specific example, the first report contains a unit field, a reference range field, and an anomaly message field. The result text is a numerical string of length four, and the check item name does not match the preset image item thesaurus. The absolute value of the negative contribution of the first report is greater than the positive contribution, so it enters the test value channel. The second report only contains the check item name and the result text, with a result text length of 126 characters. The check item name matches the preset image item thesaurus. The positive contribution of the second report is greater than the negative contribution, so it enters the image semantic channel. The third report retains the unit field but the field value is empty, and the result text length is 68 characters. The scorer encodes the unit field as "exists but invalid" instead of "validally exists," and combines this with other features to determine the channel, thereby avoiding incorrect routing simply because of a reserved empty field.

[0072] Through the aforementioned multi-layered structural features and deterministic contribution details, this embodiment transforms report type discrimination into configurable and traceable data structure operations. The technical function of this operation is not to understand the specific anomaly meaning expressed in the result text, but rather to select a parsing method adapted to the input structure for the two subsequent different data processing channels. The numerical channel verifies the parsable result values ​​and reference ranges received, while the image semantic channel receives result text requiring tree-based semantic matching. Therefore, it can reduce invalid calculations and erroneous candidates caused by mistakenly sending long texts to the numerical comparator or short values ​​to the semantic matcher.

[0073] Example 3

[0074] like Figure 3 As shown, this embodiment further illustrates the construction and matching of semantic condition trees.

[0075] The process involves matching the resulting text using a semantic condition tree in the image semantic channel. The nodes of the semantic condition tree are semantic condition objects, and the root-to-leaf path represents the matching rules. This yields image anomaly candidates and character position ranges, including:

[0076] S3.1. Read standardized anomaly information tags and corresponding ordered condition object sequences from the rule database;

[0077] Each record in the rule database contains at least a rule identifier, standardized exception information labels, an ordered sequence of condition objects, a priority, and a version number. The ordered sequence of condition objects is arranged in order from common conditions to distinguishing conditions. The loader performs integrity checks on missing types, null condition values, and duplicate labels. Records that fail the checks are not included in the memory tree.

[0078] Specifically, the ordered sequence of condition objects can be stored using a structured object array. Each condition object includes a condition type, condition content, matching options, and position return method; the condition content of a keyword condition object is a single string, the condition content of a set condition object is an array of candidate strings, and the condition content of a regular expression condition object is a regular expression string. When loading a rule, the string is first normalized to the same case and hyphens as the resulting text, and then the rule content summary is calculated. Only when the rule identifier, version number, and rule content summary all pass the consistency check is the rule record added to the tree-building set.

[0079] Rule updates can be implemented using a post-build switching approach. The system loads the new version of the rule and constructs a new semantic condition tree in an independent memory area that does not affect current matching requests. After verifying the number of nodes, leaf node labels, and sampling paths, it switches the current tree version using atomic references. Requests already in the old version of the semantic condition tree continue to use the old version for matching, while new requests use the new version. This avoids a mixed state where half of the nodes come from the old rules and half from the new rules during the update process. Candidate records carry the tree version used, facilitating result reproducibility.

[0080] S3.2. Configure the condition objects in the ordered condition object sequence as keyword condition objects, set condition objects or regular expression condition objects, and encapsulate unified matching processing steps for different types of condition objects;

[0081] The keyword condition object stores a single string and returns all matching positions of that string; the set condition object stores multiple candidate strings, and the condition is satisfied as long as at least one candidate string matches, while also returning the matching members and their positions; the regular expression condition object stores a regular expression that has undergone safety checks and returns the matching range. The input to the unified matching processing step is the normalized result text, and the output is the satisfaction status, the matched content, and the set of character positions, thus making the tree traversal logic unaware of the specific matching type.

[0082] The unified matching process can also return a matching confidence level and a failure reason code. The keyword condition object returns a certain level when the complete string is matched consecutively; the set condition object records the actual matched members, preventing situations where only the set is satisfied without being able to locate evidence; the regular expression condition object returns the position corresponding to the named capture group, facilitating the separate location of graded values ​​and category words. The failure reason code can distinguish between string not found, numeric boundary not satisfied, regular expression execution timeout, and empty input. The traversal engine only determines whether to enter a child node based on the satisfaction status; other returned fields are written to the path status and output with the candidates.

[0083] S3.3. Insert nodes level by level from the root node of the semantic condition tree according to the ordered condition object sequence, merge common prefix nodes with the same content and type, and bind the standardized abnormal information label to the last level node;

[0084] During insertion, a node key is generated using the condition type, normalized condition content, and matching options. If the same node key already exists at the current level, the node is reused; otherwise, a new child node is created. Therefore, common prefixes are stored only once, and modifications to common conditions can be uniformly applied through tree reloading. Last-level nodes can be bound to one or more labels, and the output order of nodes along the same path is determined by label priority.

[0085] The node key includes the condition type to prevent nodes with the same keyword string and regular expression text but different matching semantics from being incorrectly merged; the node key includes matching options to configure features such as case sensitivity, character boundary sensitivity, and allowing synonym variants. Common prefix merging is only performed when the node keys are exactly the same and the parent nodes are the same. Therefore, common prefix sharing does not change the order of conditions in any original rule, nor does it force the merging of conditions with the same name from different scopes.

[0086] In one example, all three rules use category and target description conditions as their first two levels of conditions, differing only in the final level's hierarchical system or hierarchical value conditions. After tree construction, each of the first two levels retains one instance, and subsequent levels branch according to specific conditions. If a fourth rule with the same first two levels of conditions is added, simply add a corresponding successor node under the existing second-level node. During traversal, text that fails to meet the category condition is pruned at the first level, while text that meets the category condition but fails to meet the target description condition is pruned at the second level. Only text that simultaneously meets the common conditions enters the hierarchical condition branch.

[0087] S3.4. Traverse the semantic condition tree, call the unified matching processing step on the child nodes of the current node, and prune the subtrees of objects that do not meet the corresponding conditions;

[0088] The traversal employs a depth-first search and maintains the path state. Unsatisfied child nodes do not visit their descendants; for satisfied child nodes, the path state is appended with the hit position, and the search continues downwards. Assuming the root node has one hundred common prefixes for organs or inspection categories, and the current text only matches one of these prefixes, the remaining ninety-nine subtrees are excluded at the first level, and subsequent searches only involve the matched branches. Sharing common prefixes reduces rule storage and runtime redundant comparisons.

[0089] To further reduce duplicate matching, the return results of a unified matching process can be cached by node key within the same request. When multiple paths reference the same node key through shared nodes or equivalent conditions, the satisfied states and character position sets in the cache are directly reused. The cache's lifecycle is limited to the processing of a single result text and is not saved across requests to prevent cross-contamination between different result texts. During concurrent processing, the semantic condition tree is shared as a read-only object, and each request maintains its own path stack and matching cache; therefore, node states are not modified between requests.

[0090] The search cost of a semantic condition tree is primarily determined by the number of actually accessed nodes, rather than by the number of Cartesian combinations of rule paths. Placing keywords with high discriminative power and low computational cost at higher levels allows for the pruning of irrelevant branches at an earlier stage; complex regular expression conditions with higher execution costs can be configured at lower levels that have already passed multiple semantic prefixes. The order of these nodes is also constrained by the semantic order of the rules; adjustments are only allowed for equivalent sorting that does not alter the conditional logic.

[0091] S3.5. When the leaf node bound to the standardized anomaly information label is reached, the character position hit by the semantic condition path corresponding to the leaf node is recorded to obtain the image anomaly candidate and the corresponding character position range.

[0092] The character position range can be the minimum start and maximum end points of all hit positions in the path, or it can simultaneously retain a set of discrete positions for continuous alphanumeric filtering. Candidate records can also carry rule identifiers, path depths, label priorities, and original text offset mappings for subsequent filtering and verification.

[0093] The data structure for the image anomaly candidates may further include candidate identifiers, standardized anomaly information labels, semantic condition path identifiers, a hit segment array, character position ranges, a discrete character position set, a tree version, and candidate generation time. Each element in the hit segment array corresponds to a semantic condition object, storing the condition type, hit text, and start and end positions. The character position range is used to quickly determine clause affiliation, and the discrete character position set is used for continuity and inclusion relationship judgment. Storing both simultaneously avoids duplicate scanning of result text in subsequent filtering stages.

[0094] From a technical perspective, the semantic condition tree does not produce the final output, but rather high-recall candidates with clearly defined evidence locations. Sharing common prefixes solves the problems of rule redundancy and repeated evaluation, while a unified matching process addresses the coordination issues of multiple matching granularities within the same path. Character position records provide necessary input for subsequent punctuation clause isolation processing. If only leaf node labels are output without their positions, subsequent filtering cannot determine which clause the label originates from; if only tiling regular expressions are used without sharing common prefixes, the same check category conditions will be repeatedly stored and executed during rule expansion. Therefore, the tree structure, unified interface, and positional output together constitute an inseparable data processing mechanism.

[0095] Example 4

[0096] This embodiment further illustrates the depth-first search and backtracking of the semantic condition tree.

[0097] The step of traversing the semantic condition tree and calling the unified matching processing step on the child nodes of the current node to prune subtrees that do not meet the corresponding conditions includes:

[0098] S3.4.1. Establish a path state that includes the current node identifier, objects that have met the conditions, and the positions of the matched characters;

[0099] The path state can be saved using stack frames. Each stack frame records the length of the hit set before entering the node, so that it can be recovered during backtracking without copying the entire set.

[0100] Specifically, the stack frame can also record the parent node identifier, the sequence number of the sibling node to be visited, and the cumulative priority of the current path. To prevent erroneous rules from forming excessively deep paths, a maximum tree depth can be set during the loading phase, and rules exceeding the maximum tree depth can be rejected. During the running phase, a maximum number of nodes accessed in a single request can be set; when the limit is reached, low-priority branches are terminated and the reason for exceeding the limit is recorded. The above restrictions are used to constrain resource consumption and do not change the depth-first search results when the limit is not triggered.

[0101] S3.4.2. Perform a depth-first search on each child node of the current node according to the child node priority;

[0102] The priority of child nodes can be determined in order of regularity boundary conditions, keyword length, and rule configuration priority, so that conditions with stronger discriminative power are validated first.

[0103] S3.4.3. Record the character position of the string when the string corresponding to the keyword condition object is contained in the result text;

[0104] String search returns all occurrences instead of just the first occurrence, to support positional attribution when the same word appears repeatedly in different clauses.

[0105] S3.4.4. When at least one candidate string contained in the set condition object is contained in the result text, record the hit candidate string and the corresponding character position;

[0106] When multiple members in the same set are matched, they can be retained first according to string length, or all can be recorded and processed in the leaf node conflict resolution stage.

[0107] S3.4.5. Record the matching interval when the regular expression corresponding to the regular expression condition object matches the result text;

[0108] Regular expressions are pre-compiled during the loading phase, and maximum execution time or expression complexity limits are set to prevent abnormal rules from causing processing threads to be occupied for a long time.

[0109] S3.4.6. Append the corresponding hit information to the path state and enter the next level node. If the current path does not satisfy any child node, restore the path state before entering the current path and search for sibling nodes.

[0110] The restore operation undoes the hit information added to the current branch, but retains the common conditions already satisfied by the parent path, thereby enabling one-time evaluation of shared prefix conditions and reuse of multiple sub-branches.

[0111] The backtracking process uses write point markers instead of reconstructing the path state. Before entering a child node, the lengths of the hit segment array and the discrete character position set are recorded; when a sub-branch fails, the two containers are truncated to the length corresponding to the write point marker, and then the sibling node is pushed onto the search stack. Since the parent path data remains unchanged before the truncation position, the backtracking will neither leave behind the character positions of failed branches nor repeat the execution of already satisfied common prefix conditions.

[0112] S3.4.7. Determine the leaf nodes to be retained based on at least one of the following conflict resolution conditions: leaf node priority, path length, or label priority.

[0113] The priority can be configured to first compare rule priorities, then compare the number of objects that meet the conditions, and finally compare label priorities; if they are still the same, multiple candidates are retained and deduplicated using the inclusion relation. This deterministic order ensures reproducible results for the same rule version and the same input.

[0114] It should be noted that conflict resolution only handles the candidate ordering when multiple leaf nodes simultaneously satisfy the conditions; it does not change the satisfaction status of any semantic condition object. A longer path usually indicates more constraints, but path length cannot unconditionally override explicitly set rule priorities. Therefore, explicit rule priority comparison can be performed first, followed by path length comparison within the same priority level. If two candidates have the same label but different character positions, the two positional evidences are retained and merged into a single candidate with the same label. If the labels are different and the priorities are exactly the same, both candidates are retained and enter the five-layer filtering pipeline, allowing subsequent inclusion relationships and clause scopes to continue participating in the evaluation.

[0115] Example 5

[0116] like Figure 4 As shown in the figure, this embodiment further illustrates the five-layer negative semantic filtering pipeline.

[0117] Based on the character position range, the image anomaly candidates are subjected to indivisible term protection, alphanumeric continuity filtering, inclusion relation deduplication, punctuation sentence isolation processing, and combined negation filtering to obtain retained image anomaly candidates, including:

[0118] S4.1. Locate the term position range of the preset indivisible term in the result text, and delete image anomaly candidates that partially overlap with the term position range and do not cover the complete term;

[0119] An indivisible glossary stores specialized descriptive phrases, standardized phrases, and version numbers. For each term, all occurrence intervals are searched, and a fast retrieval index is built using an interval tree. If a candidate interval only covers a portion of the term, it means the candidate crosses or truncates the overall term boundary; if the candidate completely covers the term, it is not deleted at that level. Term protection is performed before negation detection, ensuring that negative words within a term are protected first.

[0120] Specifically, the indivisible terminology list can also record term categories, allowed prefixes and suffixes, and the longest match priority. When multiple term position intervals contain each other, the complete term interval with more characters is retained first; when two term intervals partially overlap but do not contain each other, each term interval is retained separately, and the overlapping part is marked as a common protected position. Negative words are only excluded as negative trigger words if they are completely inside a protected interval; negative words located outside a protected interval but adjacent to a term are still treated as ordinary negative trigger words. This boundary rule avoids extending "term protection" to unbounded exemptions for adjacent words.

[0121] The protection of indivisible terms examines three relationships between candidate intervals and term intervals. If a candidate completely contains the term interval, it is retained; if the candidate interval is entirely within the term but the labeling rule only hits a portion of the term, the candidate is deleted; if the candidate and term interval only partially overlap, they are also deleted. This is because the latter two relationships indicate that the candidate evidence severs the entire term, while the first relationship at least fully preserves the term's semantics.

[0122] S4.2. When the text of the image anomaly candidate is an alphanumeric string, calculate the difference between the positions of adjacent hit characters and delete image anomaly candidates with a difference of not one.

[0123] This layer uses a discrete set of character positions instead of just the beginning and end intervals. The difference between adjacent positions in consecutive abbreviations is always one; a character sequence concatenated from different positions in the text must have at least one difference greater than one. For standard abbreviations containing hyphens, the hyphen can be configured as an allowed separator character in the vocabulary.

[0124] S4.3. Compare the inclusion relationships of the character position sets of any two image anomaly candidates, and delete image anomaly candidates whose character position set is included by the other character position set;

[0125] Candidates are first sorted in descending order of location set size, and then the smaller set is compared with the retained set, which reduces the number of pairwise comparisons. When two sets are identical but have different labels, they are retained according to rule priority, or both labels are retained for subsequent business rule selection. In other implementations, the same deduplication function can also be achieved using a range-containing index.

[0126] After comparing the inclusion relationships, the positions of suggestion words and negation words can be combined to check the differences between the two candidates. First, locate the starting position of the suggestion word and the position of the effective negation word within the current clause, then calculate the difference between the two candidate character position sets; if all the unique character positions of a candidate fall between the suggestion word and the effective negation word, and the other candidate has more complete evidence, then the former candidate is deleted. This process is still based on the relative relationship of the candidate character position sets, used to exclude shorter candidates made up of scattered characters in semantically ambiguous areas.

[0127] S4.4. Divide the resulting text into clauses based on punctuation positions, independently detect negative trigger words within each clause, and delete image anomaly candidates associated with clauses containing negative trigger words;

[0128] Punctuation boundaries are not empirical word count windows, but rather structural signals that can be directly located in the original text. Clause boundaries are first generated, and then candidate positions and negation trigger word positions are projected onto the same clause identifier space. Only when a candidate and a valid negation trigger word share a clause, or when any clause spanned by a candidate contains a valid negation trigger word, is the candidate deleted at that layer. Because adjacent clauses have different identifiers, negation words from one clause will not propagate to the other.

[0129] The clauses can be represented using left-closed, right-open intervals, with the starting position belonging to the current clause and the ending punctuation position not belonging to the current clause; the right boundary of the clause at the end of the text is the text length. For consecutive punctuation marks, only the position following the last consecutive punctuation mark is used as the starting point of the next clause to avoid generating clauses without content. Commas inside parentheses or quotation marks can be selected as ordinary characters or clause boundaries according to the data source writing specifications; when parenthesis level recognition is enabled, only commas with a parenthesis level of zero are added to the punctuation boundary sequence to avoid mistakenly splitting supplementary descriptions within parentheses into independent discoveries.

[0130] In one example, the resulting text consists of two comma-separated clauses. The target description in the first clause is located at character positions five through fourteen, and the negation trigger word in the second clause is located at character positions twenty-three through twenty-four. Another target description in the second clause is located at character positions twenty-seven through thirty. The position range of the first candidate is only associated with the first clause, therefore the negation trigger word in the second clause does not affect the first candidate; the second candidate is associated with the negation trigger word in the same clause and is therefore deleted. If a fixed window covering thirty characters is used, the negation trigger word may simultaneously cover both the first and second candidates, but clause identifiers based on punctuation boundaries can separate them.

[0131] S4.5. Combine the set of negative words with the set of positive descriptors to generate a combined negative pattern, delete the image anomaly candidates that match the combined negative pattern, and obtain the retained image anomaly candidates.

[0132] A combined negation pattern consists of a negation word, a middle character pattern restricted to the current clause, and a positive descriptor. Adding a new negation word only requires updating the negation word set, and adding a new positive descriptor only requires updating the positive descriptor set; the pattern generator will automatically construct the corresponding combination. If a clause contains independent state words such as "normal" or "negative" and there are no words that inversely modify those state words, it can also be configured to directly remove the associated candidates for that clause. After the five-layer pipeline is split according to responsibility, term protection, positional continuity, set deduplication, scope isolation, and combined patterns each handle different types of false positives, without needing to stack all exceptions in a single regular expression.

[0133] Before concatenating negative and positive descriptors, the pattern generator escapes special characters in the terms, limiting the intermediate character pattern to not crossing Chinese commas, semicolons, periods, and newlines. The generated combined negative patterns are cached according to the versions of the negative and positive descriptor lists. If either list version changes, the old cache is invalidated and recompiled. To avoid overly broad matching, an upper limit can be set on the number of intermediate characters; this upper limit only constrains combined patterns in single sentences without punctuation and does not replace the clause scope determination in step S4.4.

[0134] The five-layer negation semantic filtering pipeline is essentially a priority-based candidate set transformation process. The first layer protects the boundaries of technical terms, the second layer verifies the continuity of character evidence, the third layer eliminates duplicate candidates whose information is fully contained, the fourth layer establishes the scope of punctuation clauses, and the fifth layer processes combined negation structures without explicit delimiters within the same clause. This execution order is technically dependent: if negation words are detected first and then term protection is performed, characters within the term may have already led to candidate deletion; if combined negation is performed first without establishing clause boundaries, combined patterns may be detected across adjacent clauses. Candidates are reduced layer by layer in this order, ensuring that each layer only consumes data that has passed the pre-constraints.

[0135] Example 6

[0136] This embodiment further explains the punctuation mark-based sentence isolation process.

[0137] The process of dividing the resulting text into clauses based on punctuation positions, independently detecting negation trigger words within each clause, and deleting image anomaly candidates associated with clauses containing negation trigger words includes:

[0138] S4.4.1. Identify Chinese commas, semicolons, and periods in the resulting text, forming a punctuation boundary sequence arranged in ascending order of character position;

[0139] Add the text start position at the beginning of the boundary sequence and the text end position at the end. If the system only processes data that uses Chinese commas, only Chinese commas can be enabled; when the data source contains semicolons or periods as separators, enable the corresponding boundary type.

[0140] Punctuation recognition is performed on the normalized result text, preserving the original text position through character offset mapping. Each element in the boundary sequence contains the punctuation character, the normalized text position, the original text position, and the bracket level. For newlines, a strong boundary equivalent to a period can be configured depending on the data source; for colons, a boundary can be configured when followed by an enumeration of check results, and a normal character is maintained when representing numerical proportions or time.

[0141] S4.4.2. Form corresponding clauses using the character intervals between two adjacent punctuation marks, and establish a mapping between clause identifiers and character intervals;

[0142] A mapping can be represented as an array of intervals sorted by their starting positions. A binary search is used to determine the clause to which any character position belongs. Empty clauses and clauses containing only whitespace characters are ignored when establishing the mapping.

[0143] Each clause identifier can also be associated with the clause start and end positions, the set of valid negation trigger word positions, the set of term protection intervals, and the set of candidate identifiers. When establishing the mapping, all clause intervals are completed first, and then the negation trigger words, protection intervals, and candidate intervals are projected onto the clause identifiers respectively, avoiding repeated scanning of the entire text for each candidate. For single result texts containing many clauses, the interval array remains ordered by position, allowing all objects to be assigned using a single linear scan.

[0144] S4.4.3. Compare the character position range of each image anomaly candidate with the mapping to determine at least one clause identifier associated with the image anomaly candidate;

[0145] A single association is established when the start and end points of a candidate fall in the same clause; when they fall in different clauses, all clause identifiers covered from the start to the end point are enumerated, and the candidate is marked as a cross-clause candidate.

[0146] When candidates consist of a set of discrete character positions, in addition to comparing the first and last positions, the clause affiliation of each discrete character position is checked one by one. If the set of discrete positions involves more than two clauses, even if the first and last intervals of the candidates appear continuous due to punctuation removal, they are still marked as cross-clause candidates. This processing is compatible with the data structure of the word-by-word sequential matching output and can identify candidates formed by combinations of scattered characters in different clauses.

[0147] S4.4.4. Match the preset negation trigger word within each clause and exclude negation words located inside indivisible terms;

[0148] Negation triggers can be fixed strings or regular expressions with contextual constraints. After a match, it is checked whether the position of the negation trigger is contained within a term protection range; if it is contained, no valid negation tag is generated.

[0149] Negation trigger word records can include the trigger word text, matching pattern, direction of action, and exception pattern. Direction of action distinguishes between trigger words that typically modify the following content and status words that typically modify the preceding content; exception patterns exclude non-negative meanings caused by fixed collocations. If the same character position simultaneously hits a regular negation trigger word and an indivisible term protection zone, the term protection result takes precedence; if the negation trigger word crosses the boundary of a protection zone, it is only excluded if the entire negation trigger word is located within the protection zone.

[0150] S4.4.5. Delete the image anomaly candidate when it is associated with only one clause and the clause contains a valid negative trigger word;

[0151] When multiple candidates appear in a clause, more granular rules can be configured according to the dependency direction between the candidate and the negation trigger word; if no configuration is made, a clause-level conservative deletion rule is used.

[0152] Optionally, dependency direction determination is implemented using character position. Negation triggers with backward action only process candidates whose starting point is after the end position of the negation trigger, while state words with forward action only process candidates whose ending point is before the start position of the state word; bidirectional action words process all candidates within the same clause. When no action direction is configured or the positional relationship cannot be determined, clause-level conservative deletion is performed, thus maintaining determinism without relying on an additional language model.

[0153] S4.4.6. Delete the image anomaly candidate when it spans multiple clauses and any related clause contains a valid negative trigger word;

[0154] Cross-clause candidates typically indicate that rule conditions come from multiple logical segments. Any related clause containing a negative trigger word is deleted, which avoids concatenating conditions that do not belong to the same discovery into a label.

[0155] For candidates spanning more than three clauses, they can be deleted directly based on the maximum number of clauses spanned before negation detection. This maximum number of clauses spanned is part of the candidate evidence integrity check, based on the principle that a single semantic conditional path should typically be completed within a single consecutive discovery or a small number of adjacent supplementary clauses; discrete evidence spanning too many clauses is more likely to come from random combinations of word-for-word sequential matching. If the deployment scenario allows for clause-spanning descriptions, the restriction can be appropriately relaxed, but the deletion rule will still apply if any associated clause contains a valid negation trigger word.

[0156] S4.4.7. Retain image anomaly candidates in which all associated clauses do not contain valid negation trigger words.

[0157] The record retains clause identifiers, candidate positions, and rule versions, facilitating the highlighting of original text evidence in the review interface. Using this position mapping mechanism, the negation scope is determined by text punctuation, rather than by a manually set fixed number of characters.

[0158] From a technical perspective, clause identifiers convert punctuation marks in natural language into discrete scope numbers, enabling comparison of candidate positions, negation trigger word positions, and term protection positions within the same coordinate system. Fixed windows only indicate "distance proximity," not whether two characters are separated by logical delimiters; this embodiment utilizes both positional distance and boundary attribution, ensuring that even if a negation trigger word and a candidate character are very close, cross-clause deletion will not occur as long as they are in different clauses. Conversely, even if a long modifier exists between a negation trigger word and a candidate character, valid negation can still be applied as long as they belong to the same clause.

[0159] In one application example, the normalized result text contains three clauses. The first clause contains a complete technical term and a first candidate; the second clause contains a negation trigger word and a second candidate; and the third clause contains an independent state word. Term protection first marks specific words in the first clause as non-negation positions; clause mapping then associates the first candidate with the first clause and the second candidate with the second clause. The first candidate is retained because the associated clause does not contain a valid negation trigger word; the second candidate is deleted because it shares a clause with a valid negation trigger word; and the third clause does not generate a candidate due to the independent state word rule. Each deletion or retention result in the entire process carries a reason code and evidence position, allowing for repeated execution by the same rule version to obtain consistent results.

[0160] Example 7

[0161] like Figure 5 As shown, this embodiment configures a progressive cascading matching strategy before and after the semantic condition tree.

[0162] Before and after matching the resulting text using the semantic condition tree, a progressive cascading matching strategy is configured, the progressive cascading matching strategy including:

[0163] A1. Perform prenormalization on the Roman numerals, hierarchical letters, and hierarchical system names in the resulting text;

[0164] Pre-normalization shares the transformation table with step S1, but retains the original value, target value, and character offset for each transformation in this step for boundary checks of hierarchical entities.

[0165] Pre-normalization also generates a standardized token sequence, which includes at least consecutive Chinese characters, consecutive alphanumeric characters, hierarchical system names, and punctuation. Subsequent precise matching uses the normalized check item name, full-text search uses the normalized result text, and character-by-character sequential matching uses a normalized string that preserves character order. Multiple levels are provided with consistent input from the same pre-normalization, avoiding character position drift caused by separate conversions at each level.

[0166] A2. Match the pre-normalized check item names precisely with the alias field in the search index;

[0167] Exact matching uses the alias field without tokenization, and directly retrieves the bound tag upon a match. It is suitable for handling fully indexed project names and stable aliases.

[0168] Each document in the index can contain standardized anomaly information tags, standard names, an array of aliases, inspection categories, and rule versions. The alias field uses an exact index and has undergone the same normalization process as the inspection item name during writing. Queries include both report type and inspection category as filters to prevent the same short alias from pointing to irrelevant tags in different categories. When multiple documents are precisely matched, inspection category consistency is compared first, followed by alias length and document priority.

[0169] For scenarios requiring the injection of domain weights, script scoring or equivalent function scoring can be used during the exact match and full-text search stages. The script reads the inspection category field from candidate documents. When the inspection category matches the category extracted from the current inspection item name, the base search score is multiplied by a preset domain weight coefficient; otherwise, the base score is maintained or a smaller coefficient is applied. The domain weight coefficient is configured in the search strategy table with a version number and is not written to the fixed text of the indexed documents, thus enabling adjustments to the ranking strategy without rebuilding the entire index.

[0170] A3. When an exact match is not found, the semantic condition tree is used for matching;

[0171] The semantic condition tree is responsible for processing rules that are jointly limited by multiple conditions and have a common prefix. Its output includes labels and character positions, providing positional basis for subsequent negative filtering.

[0172] If an exact match has been found but its binding rule requires validation of the result text conditions, the exact match result can be used as the entry branch constraint of the semantic condition tree, instead of directly outputting the label. This entry branch constraint narrows the set of child nodes to be searched from the root node, while still validating the result text using the complete semantic condition path. This utilizes the precise index of the item name to reduce the tree search scope without omitting key conditions from the result text.

[0173] A4. When the semantic condition tree is not matched, perform a full-text search on the result text and perform dual filtering based on the full-text search score and edit distance similarity;

[0174] Full-text search first generates a candidate list, then excludes candidates that only receive high search scores due to the prevalence of common local words based on edit distance similarity. Edit distance similarity can be expressed as:

[0175] Equation (3)

[0176] in, This represents the edit distance similarity of the current full-text search candidates; Indicates the edit distance function; This refers to the normalized result text or a comparison segment extracted from the normalized result text; This indicates the standard phrase corresponding to the full-text search candidate; express The number of characters; express The number of characters. The symbols in equation (3) are not used in other formulas. Only when the full-text search score and Candidates are retained when their respective thresholds are reached. The similarity threshold can be selected between 0.70 and 0.95. A higher threshold is recommended for short texts, while a lower threshold can be used for long texts when segments are extracted.

[0177] Specifically, the comparison segment can be extracted based on the highlighted positions returned by the full-text search. Using the minimum and maximum starting and ending points of the highlighted words as the center, a preset number of characters are extended forward and backward, and the segment is truncated at the nearest punctuation mark boundary, resulting in... This local segment is closer in length to the candidate standard phrase, which avoids irrelevant characters in the entire long text excessively reducing the edit distance similarity. The full-text search scoring threshold and the edit distance similarity threshold employ a dual threshold; if either threshold is not met, no candidate is retained. The dual thresholds respectively constrain search relevance and character sequence similarity.

[0178] When the full-text search service is temporarily unavailable or a query times out, the cascading controller records the failure reason for the current level and directly proceeds to word-by-word sequential matching and hierarchical entity-specific identification according to the configuration, without interpreting the service anomaly as "no anomaly information". After the search service recovers, reports with failure reason codes can be reprocessed. This degradation mechanism ensures that the tag output process can continue, while retaining the report data and policy versions required to re-execute the deterministic path.

[0179] A5. Perform word-by-word sequential matching if the double filtering fails;

[0180] Word-by-word sequential matching allows other characters to exist between target characters, but requires all target characters to appear in their original order. It also sets protection for high-frequency characters and protection for last alphanumeric boundaries to reduce the chance of random concatenation of short keywords in long texts.

[0181] A6. Perform hierarchical entity recognition on result text containing names of preset hierarchical systems;

[0182] The hierarchical entity identification process can be performed in parallel with the aforementioned steps and is suitable for texts where there are multiple writing orders for the hierarchical system name, target category term, and hierarchical value. Numerical boundary conditions are used for the hierarchical values ​​to avoid mismatching a single level number to a longer number or level range.

[0183] A7. Determine the image anomaly candidates according to the first valid hit or the preset priority level.

[0184] The cascading mode can be configured to stop upon the first valid match, or it can allow multiple levels to return candidates and select them in a reliable order: exact match, semantic condition tree, full-text search, and word-by-word sequential match. Hierarchical entity-specific recognition results can be configured to take precedence over ordinary full-text search results according to rules. Different levels are connected through a unified candidate data structure, thus allowing for independent upgrades or replacements.

[0185] A unified candidate data structure should include at least standardized anomaly information labels, candidate source level, base score, secondary screening score, character position, rule or index version, and filtering status. In the first-time valid hit mode, "valid" means that the candidate is not only generated from the current level but has also passed character boundary checks and negative semantic filtering applicable to that candidate. The multi-level aggregation mode first merges identical labels and overlapping evidence positions, then sorts them according to source level priority, rule priority, and score order to prevent low-confidence fallback candidates from covering high-confidence accurate candidates.

[0186] From an overall technical perspective, the progressive cascaded matching strategy combines matching methods with different recall capabilities and computational costs in a predetermined order. Exact matching handles stable aliases, semantic condition trees handle multi-condition combinations, full-text search and edit distance handle approximate expressions, word-by-word sequential matching handles expressions with a few inserted separator characters, and hierarchical entity identification handles combinations of system names and level values. When the previous level obtains a valid result, subsequent calculations can be reduced; when the previous level fails, the next level expands the recall scope. Each level is connected to a unified candidate data structure and a five-layer filtering pipeline, thus expanding recall does not circumvent negative semantic constraints.

[0187] In one example, the first input, the item name, is exactly the same as the alias field, and the exact match level directly generates candidates and passes the filtering. The second input, the item name, is not an exact match, but the resulting text simultaneously satisfies the common category conditions, target description conditions, and hierarchical boundary conditions in the semantic condition tree, and candidates are generated by the semantic condition tree. The third input differs from the standard phrase by only a few characters, and is returned by full-text search and then passed through the edit distance similarity threshold. The target characters in the fourth input maintain their order but are separated by modifier characters, and candidates are generated by word-by-word sequential matching. All inputs are ultimately converted into candidate records with the same field structure, and subsequent filtering and label output do not need to know which internal matching implementation the candidates come from.

[0188] Example 8

[0189] This embodiment provides a further explanation of word-by-word sequential matching.

[0190] The step of performing word-by-word sequential matching when the double filtering fails includes:

[0191] A5.1. Break down the target keyword into a sequence of target characters that retains their original order;

[0192] During the splitting process, the boundary attributes of consecutive alphanumeric segments are preserved, and the length of the target keyword is recorded.

[0193] A5.2. Establish a set of occurrence positions of each target character in the target character sequence in the result text;

[0194] The position set is stored in ascending order. When the same character appears multiple times in the target character sequence, the position set of that character is referenced respectively. However, when selecting recursively, the later position must be greater than the previous position.

[0195] A5.3. Terminate word-by-word sequential matching of the corresponding target keyword when the number of occurrences of any target character exceeds the threshold.

[0196] The threshold for the number of occurrences can be set to ten, or it can be set in segments based on the length of the resulting text. This protection prevents the number of combined paths from increasing rapidly due to high-frequency single characters.

[0197] A5.4. Recursively select character positions from each set of occurrence positions to form candidate position paths;

[0198] During recursion, the nearest successor position is prioritized, and backtracking is performed early if the remaining text length is insufficient to accommodate the remaining target characters.

[0199] A5.5. Delete candidate position paths that do not maintain the original order of the target character sequence;

[0200] The positions of adjacent target characters must strictly increase; optionally, a maximum span can be set to exclude paths that cross too much irrelevant text.

[0201] A5.6. When the target keyword ends with a letter or a number, perform character boundary validation on the last character;

[0202] Character boundary checks require that the last character is not followed by any of the same type of letters or numbers, or that the last character is located at the end of the text, to prevent the target keyword from becoming part of a longer number.

[0203] A5.7. The candidate position paths that pass the character boundary verification and cover all target characters are determined as the character-by-character sequential matching results.

[0204] If multiple results exist, the one with the smaller total span is selected first; if the spans are the same, the one with the earlier starting position is selected. The resulting set of discrete positions then undergoes alphanumeric continuity filtering and punctuation sentence segmentation.

[0205] Example 9

[0206] This embodiment further explains the hierarchical entity identification.

[0207] The step of performing hierarchical entity recognition on the result text containing the name of the preset hierarchical system includes:

[0208] A6.1. Read the hierarchical rule records containing standardized anomaly information tags, keyword sequences, ordinal attributes, and hierarchical values;

[0209] The hierarchical rule record can also include a hierarchical system identifier, allowed letter suffixes, priority, and version number to support parallel configuration of different systems.

[0210] A6.2. When the order attribute is ordered, the keyword sequence is converted into an ordered regular expression pattern that is connected in order. The ordered regular expression pattern is a regular expression that connects adjacent keywords with a non-greedy wildcard constrained by the maximum character span.

[0211] Adjacent keywords are connected using a non-greedy character pattern constrained by the maximum length to avoid absorbing irrelevant content across the entire text.

[0212] A6.3. When the order attribute is unordered, the keyword sequence is converted into an unordered regular expression pattern containing multiple zero-width assertions. The unordered regular expression pattern is a regular expression in which each keyword is combined in the same clause range in the form of positive zero-width assertions.

[0213] Each keyword corresponds to a positive zero-width assertion. The rule condition is considered satisfied only when all assertions are satisfied within the same clause.

[0214] A6.4. Add forward and backward number boundaries to the graded value. The forward number boundary is used to exclude the case where the number immediately preceding it is a number, and the backward number boundary is used to exclude the case where the number immediately following it is a number or forms a numerical range with the number following it, thus obtaining the graded boundary pattern.

[0215] Forward numeric boundaries exclude cases where the preceding number is immediately adjacent, and backward numeric boundaries exclude cases where the following number is immediately adjacent or where the numbers are connected as a range of values; for hierarchical values ​​with letter suffixes, the letter part is matched without distinguishing between uppercase and lowercase letters.

[0216] A6.5. Combine the ordered regularization pattern or the unordered regularization pattern with the hierarchical boundary pattern to obtain a hierarchical entity matching pattern;

[0217] When combining different keywords, create named capture groups so that the matching results can return the positions of the system name, target category words, and grading values.

[0218] A6.6. Process the result text using the hierarchical entity matching mode, and determine the standardized anomaly information label corresponding to the hit record as the hierarchical entity recognition result.

[0219] The hierarchical entity recognition result carries the hierarchical rule version and capture location. If the same clause hits multiple adjacent hierarchical values ​​at the same time, the output can be determined based on the rule priority or the hierarchical value closest to the target category word.

[0220] Example 10

[0221] like Figure 6 As shown, this embodiment provides a system for identifying abnormal information in a physical examination report, including:

[0222] M1. Report Normalization Module, used to obtain the medical examination report data to be processed, which includes field patterns and result text, and normalize the grading symbols and grading system names in the result text to obtain standardized report data;

[0223] The report normalization module may include an interface adaptation unit, a character conversion unit, and an offset mapping unit, which respectively complete field mapping, text normalization, and position conversion between the original text and the normalized text.

[0224] The interface adaptation unit can maintain independent field mapping configurations for different data sources, but the converted normalized report data adopts a unified internal structure. The character conversion unit and the offset mapping unit are completed within the same processing transaction, so that any normalized text has a corresponding original text position mapping; if the conversion fails, the original input, failure rule identifier and data source version are retained, and incomplete normalized report data is not passed to subsequent modules.

[0225] M2. Report distribution module, used to construct report structure features based on the field pattern, the text form of the result text and the semantics of the inspection item name, determine the report type by the cumulative weight value of the report structure features, and allocate the normalized report data to the inspection value channel or the image semantic channel;

[0226] The report distribution module reads the versioned scoring table and outputs the channel identifier, which, together with the standardized report data, is transmitted to the candidate generation module.

[0227] To maintain configuration consistency within a single processing task, the report routing module acquires configuration snapshots of the scoring table version, preset image item vocabulary version, and field mapping version at the start of the task. Even if new configurations are released in the background during processing, the current task still uses the aforementioned configuration snapshot to complete, and the next task reads the new version. The configuration snapshot identifier is written to the final processing record to support re-execution using the same version.

[0228] M3. Candidate generation module, used to obtain test anomaly candidates based on the result value and reference range in the test numerical channel; and to match the result text using a semantic condition tree in the image semantic channel, wherein the nodes of the semantic condition tree are semantic condition objects and the root-to-leaf path is the matching rule, to obtain image anomaly candidates and character position ranges.

[0229] The candidate generation module may include a numerical comparison unit, a semantic condition tree matching unit, and a progressive cascade matching unit. Each unit outputs candidate records in a uniform format, ensuring that each candidate record contains at least the tag encoding, source channel, character position, and rule version.

[0230] The candidate generation module can use a thread pool or task queue to process different reports in parallel, but the cascading order within the same result text remains unchanged. The semantic condition tree and retrieval index are shared by multiple tasks as read-only versions of the object, while path status, candidate lists, and matching caches are maintained independently by each task. When the maximum processing time for a single task is reached, the module stops the low-priority fallback branch, outputs the candidates that have been formed and passed the filtering, and appends a timeout reason code for subsequent retries.

[0231] M4. Candidate filtering module, used to perform indivisible term protection, alphanumeric continuity filtering, inclusion relationship deduplication, punctuation sentence isolation processing and combined negation filtering on the image anomaly candidates according to the character position range, to obtain retained image anomaly candidates;

[0232] The candidate filtering module can load five filters using a chain of responsibility approach. Each filter receives a candidate list and a result text, and outputs an updated candidate list and a deletion reason code to verify the filtering path.

[0233] Each filter in the chain of responsibility also outputs the number of input candidates, the number of deleted candidates, and the processing time. When a filter malfunctions, the candidate filtering module does not silently allow the relevant candidates; instead, it marks the corresponding candidates as pending review or terminates the current report processing based on the configuration. The terminology version, negative term set version, and positive descriptor set version are saved with the filtering records, ensuring that the deletion reason includes not only the rule type but also the specific configuration version.

[0234] M5. Tag Output Module, used to summarize the inspection anomaly candidates and the retained image anomaly candidates, and output standardized anomaly information tags.

[0235] The tag output module can write the results to a message queue, database, or interface response, and simultaneously save the evidence range and rule version. The system can be implemented by at least one processor and a memory. When the computer program stored in the memory is executed by the processor, the processor completes the data processing process described in any of Embodiments 1 to 9.

[0236] In this embodiment of the invention, a regression validation set can be constructed using de-identified historical reports. The regression validation set is grouped according to the following criteria: test class structure, image class structure, mixture of positive and negative descriptions within the same text, indivisible terms, hierarchical boundaries, similar text, and text separated by word order. After each update of the scoring table, rule database, or thesaurus, the report type splitting results, candidate labels, candidate character positions, filtering reason codes, and final standardized anomaly information labels are checked to ensure they meet expectations. Regression validation does not train a classification model; its purpose is to verify that updates to deterministic rule versions do not disrupt existing processing paths.

[0237] During system operation, the system can also record the proportion of traffic splitting channels, the hit rate of each level of matching, the deletion rate of each filter layer, the proportion of unlabeled reports, and the average processing time. When the proportion of traffic splitting channels or the proportion of unlabeled reports for a certain data source continuously deviates from the historical baseline, a configuration review prompt is generated to check whether the field mapping or data format has changed. The monitoring only targets the running status of the data processing link and does not change the standardized anomaly information labels of the current report.

Claims

1. A method for identifying abnormal information in a physical examination report, characterized in that, Applications include: processing of medical examination report data Obtain the medical examination report data to be processed, which includes field patterns and result text; normalize the grading symbols and grading system names in the result text to obtain standardized report data. The report structure features are constructed based on the field pattern, the text form of the result text, and the semantics of the inspection item name. The report type is determined by the cumulative weight value of the report structure features, and the standardized report data is allocated to the inspection value channel or the image semantic channel. In the numerical inspection channel, candidate anomalies are obtained based on the result value and reference range. In the semantic image channel, a progressive cascade matching strategy is configured to obtain candidate image anomalies and character position ranges. The progressive cascade matching strategy includes: matching the pre-normalized inspection item name with the alias field in the retrieval index; when no match is found, the result text is matched using a semantic condition tree, where nodes are semantic condition objects, the root-to-leaf path is the matching rule, and the semantic condition tree is obtained by inserting nodes sequentially from the root node according to an ordered sequence of condition objects and merging common prefix nodes with the same content and type. When traversing the semantic condition tree, a unified matching processing step is called to prune subtrees that do not meet the corresponding condition objects. When a leaf node with a standardized anomaly information label is reached, the character position matched by the corresponding semantic condition path is recorded; when no match is found in the semantic condition tree, a full-text search is performed on the result text, and double filtering is performed based on the full-text search score and edit distance similarity; when no match is found in the double filtering, a character-by-character sequential matching is performed; and the candidate image anomalies is determined according to the first valid hit or a preset priority level. Based on the character position range, the image anomaly candidates are subjected to indivisible term protection, alphanumeric continuity filtering, inclusion relationship deduplication, punctuation sentence isolation processing, and combined negation filtering to obtain retained image anomaly candidates; The test anomaly candidates and the retained image anomaly candidates are summarized, and standardized anomaly information labels are output.

2. The method according to claim 1, characterized in that, The process of constructing report structure features based on the field pattern, the text form of the result text, and the semantics of the inspection item name, determining the report type using the cumulative weight value of the report structure features, and allocating the standardized report data to the inspection numerical channel or the image semantic channel includes: The existence status of the unit field, reference range field, and exception message field is detected to obtain the field existence characteristics; Extract the character length, number format status, and positive marker status of the resulting text to obtain text morphological features; The names of the inspection items are matched with a preset image item vocabulary to obtain the semantic features of the items; The report type score is obtained by multiplying the field existence feature, the text morphology feature, and the project semantic feature by their corresponding directional weights and summing them. The report type score is compared with the triage threshold. When the report type score is on the image side of the triage threshold, it is assigned to the image semantic channel. When the report type score is on the test side of the triage threshold, it is assigned to the test value channel.

3. The method according to claim 1, characterized in that, The method of matching the resulting text using a semantic condition tree includes: Read standardized anomaly information tags and corresponding ordered condition object sequences from the rule database; The condition objects in the ordered condition object sequence are configured as keyword condition objects, set condition objects, or regular expression condition objects, and a unified matching processing step is encapsulated for the different types of condition objects. The unified matching processing step is a matching processing step that takes the normalized result text as input and satisfies the status, hit content, and character position set as output. Insert nodes level by level from the root node of the semantic condition tree according to the ordered condition object sequence, merge common prefix nodes with the same content and type, and bind the standardized abnormal information label to the last level node; Traverse the semantic condition tree, call the unified matching processing step on the child nodes of the current node, and prune the subtrees of objects that do not meet the corresponding conditions; When the leaf node bound to the standardized anomaly information label is reached, the character position hit by the semantic condition path corresponding to the leaf node is recorded to obtain the image anomaly candidate and the corresponding character position range.

4. The method according to claim 3, characterized in that, The step of traversing the semantic condition tree and calling the unified matching processing step on the child nodes of the current node to prune subtrees that do not meet the corresponding conditions includes: Establish a path state that includes the current node identifier, objects that have met the conditions, and the positions of the matched characters; Perform a depth-first search on each of the child nodes of the current node according to their priority; Record the character position of the string when the string corresponding to the keyword condition object is contained in the result text; Record the matched candidate string and its corresponding character position when at least one candidate string contained in the set condition object is contained in the result text; Record the matching interval when the regular expression corresponding to the regular expression condition object matches the result text; The corresponding hit information is appended to the path state and the next level node is entered. If the current path does not satisfy any child node, the path state before entering the current path is restored and sibling nodes are searched. The leaf nodes to be retained are determined based on at least one of the conflict resolution criteria: leaf node priority, path length, or label priority.

5. The method according to claim 1, characterized in that, Based on the character position range, the image anomaly candidates are subjected to indivisible term protection, alphanumeric continuity filtering, inclusion relation deduplication, punctuation sentence isolation processing, and combined negation filtering to obtain retained image anomaly candidates, including: Locate the term position range of the preset indivisible term in the result text, and delete image anomaly candidates whose character position range partially overlaps with the term position range and does not cover the complete term; When the text of the image anomaly candidate is an alphanumeric string, calculate the difference between the positions of adjacent hit characters and delete image anomaly candidates with a difference of not one. Compare the inclusion relationships between any two sets of character positions of image anomaly candidates, and delete image anomaly candidates whose character position set is included by the other set of character positions; The resulting text is divided into clauses based on punctuation positions. Negation trigger words are detected independently within each clause, and image anomaly candidates associated with clauses containing negation trigger words are deleted. The set of negative words and the set of positive descriptors are combined to generate a combined negative pattern. Image anomaly candidates that match the combined negative pattern are deleted to obtain the retained image anomaly candidates.

6. The method according to claim 5, characterized in that, The process of dividing the resulting text into clauses based on punctuation positions, independently detecting negation trigger words within each clause, and deleting image anomaly candidates associated with clauses containing negation trigger words includes: Identify Chinese commas, semicolons, and periods in the resulting text and form a punctuation boundary sequence arranged in ascending order of character position; The corresponding clauses are formed by the character intervals between two adjacent punctuation marks, and a mapping between clause identifiers and character intervals is established; Compare the character position range of each image anomaly candidate with the mapping to determine at least one clause identifier associated with the image anomaly candidate; Match the preset negation trigger word within each clause, and exclude negation words located inside indivisible terms; The image anomaly candidate is deleted when it is associated with only one clause and the clause contains a valid negative trigger word. The image anomaly candidate is deleted when it spans multiple clauses and any associated clause contains a valid negative trigger word. Retain image anomaly candidates in which all associated clauses do not contain valid negation trigger words.

7. The method according to claim 1, characterized in that, The progressive cascading matching strategy also includes: Perform prenormalization on the Roman numerals, hierarchical letters, and hierarchical system names in the resulting text; Perform hierarchical entity recognition on the result text containing the name of the preset hierarchical system.

8. The method according to claim 1, characterized in that, The step of performing word-by-word sequential matching when the double filtering fails includes: Break the target keyword into a sequence of target characters that retains its original order; Establish a set of occurrence positions of each target character in the target character sequence in the result text; The word-by-word sequential matching of the corresponding target keyword will be terminated when the number of occurrences of any target character exceeds the threshold. Recursively select character positions from each set of occurrence positions to form candidate position paths; Delete candidate position paths that do not maintain the original order of the target character sequence; When the target keyword ends with a letter or a number, character boundary validation is performed on the last character; The candidate position paths that pass the character boundary verification and cover all target characters are determined as the character-by-character sequential matching results.

9. The method according to claim 7, characterized in that, The step of performing hierarchical entity-specific recognition on the result text containing the name of the preset hierarchical system includes: Read hierarchical rule records containing standardized anomaly information tags, keyword sequences, ordinal attributes, and hierarchical values; When the order attribute is ordered, the keyword sequence is converted into an ordered regular expression pattern that is connected in order. The ordered regular expression pattern is a regular expression that connects adjacent keywords with a non-greedy wildcard constrained by the maximum character span. When the order attribute is unordered, the keyword sequence is converted into an unordered regular expression pattern containing multiple zero-width assertions. The unordered regular expression pattern is a regular expression in which each keyword is combined in the same clause range in the form of positive zero-width assertions. Add forward and backward number boundaries to the graded value. The forward number boundary is used to exclude the case where the number immediately preceding it is a number, and the backward number boundary is used to exclude the case where the number immediately following it is a number or forms a numerical range with the number following it, thus obtaining the graded boundary pattern. The ordered regularization pattern or the unordered regularization pattern is combined with the hierarchical boundary pattern to obtain the hierarchical entity matching pattern. The resulting text is processed using the hierarchical entity matching mode, and the standardized anomaly information label corresponding to the hit record is determined as the hierarchical entity recognition result.

10. A system for identifying abnormal information in a physical examination report, characterized in that, include: The report normalization module is used to obtain the medical examination report data to be processed, which includes field patterns and result text, and normalize the grading symbols and grading system names in the result text to obtain standardized report data. The report distribution module is used to construct report structure features based on the field pattern, the text form of the result text, and the semantics of the inspection item name, determine the report type by the cumulative weight value of the report structure features, and allocate the standardized report data to the inspection value channel or the image semantic channel. The candidate generation module is used to generate test anomaly candidates based on the result value and the reference range in the test value channel; A progressive cascaded matching strategy is configured in the image semantic channel to obtain image anomaly candidates and character position ranges. The progressive cascaded matching strategy includes: matching the pre-normalized inspection item name with the alias field in the retrieval index; when the match fails, matching the result text using a semantic condition tree, where the nodes of the semantic condition tree are semantic condition objects, the root-to-leaf path is the matching rule, and the semantic condition tree is obtained by inserting nodes level by level from the root node according to an ordered sequence of condition objects and merging common prefix nodes with the same content and type; when traversing the semantic condition tree, a unified matching processing step is called to prune subtrees that do not meet the corresponding condition objects; when a leaf node bound with a standardized anomaly information label is reached, the character position matched by the corresponding semantic condition path is recorded; when the semantic condition tree fails, full-text search is performed on the result text and double-filtering is performed based on full-text search score and edit distance similarity; when the double-filtering fails, character-by-character sequential matching is performed; and the image anomaly candidates are determined according to the first valid hit or a preset priority level. The candidate filtering module is used to perform non-separable term protection, alphanumeric continuity filtering, inclusion relationship deduplication, punctuation sentence isolation processing, and combined negation filtering on the image anomaly candidates based on the character position range, so as to obtain retained image anomaly candidates. The label output module is used to summarize the candidate abnormalities detected and the candidate abnormalities in the retained images, and output standardized abnormality information labels.

Images