Medical voice case automatic generation method and system based on cloud large model

Abstract

This invention discloses a method and system for automatically generating medical voice case records based on a cloud-based large-scale model. It collects doctor's voice data from a medical microphone and performs semantic analysis to identify terminology density. A terminology calibration mapping is constructed through context matching evaluation and ambiguity resolution. Case structure response data from the cloud-based large-scale model is obtained, and priority factors for completion are generated based on structural missingness. Content allocation coefficients are generated based on the cumulative rate of change. Diagnostic conflicts are identified and sorted through the terminology calibration mapping, and key nodes for case generation are established by combining sorting consistency detection. Finally, draft case documents are output based on these key nodes. Standardized case documents are generated by statistical analysis of doctor modification frequency and allocation of key verification points. This effectively solves problems such as insufficient handling of terminology ambiguity, lack of dynamic case structure completion capability, and imperfect error-prone area verification mechanisms in medical voice transcription.

Description

Technical Field

[0001] This invention relates to the field of medical information processing technology, and in particular to a method and system for automatically generating medical voice medical records based on a cloud-based large model. Background Technology

[0002] In clinical practice, doctors need to record a large amount of medical information while conducting consultations, and medical record writing accounts for a significant portion of their daily work. Although voice input reduces the burden of manual writing to some extent, doctors commonly mix professional terminology with colloquial expressions in their oral presentations. Some terms have different clinical meanings in different contexts, making it difficult to directly use the transcribed text to generate standardized case reports.

[0003] Existing speech-based medical record generation solutions typically rely on fixed templates to fill in fields in the speech transcription results, lacking a dynamic evaluation mechanism for the overall structural integrity of the medical record. When key fields are missing due to insufficient recording quality or incomplete statements, the system cannot effectively identify the impact of the missing fields on the overall medical record, nor can it rationally allocate resources for completion based on the priority of diagnostic logic. The generated draft medical records exhibit a regular distribution of errors across different fields, and existing solutions lack targeted verification mechanisms for these regular errors. The final output medical record documents still require doctors to invest considerable effort in manual review, failing to effectively improve doctors' clinical work efficiency. Summary of the Invention

[0004] This invention discloses a method and system for automatically generating medical voice medical records based on a cloud-based large-scale model. It aims to address issues such as insufficient handling of terminological ambiguity in medical voice transcription, lack of dynamic completion capabilities for medical record structures, and an imperfect verification mechanism for error-prone areas in medical record drafts. By performing semantic analysis and terminology calibration on physician voice data, assessing the missing value and allocating content to the medical record structure generated by the cloud-based large-scale model, and establishing key nodes for medical record generation by combining diagnostic priority and ranking consistency detection, a complete automated process from voice acquisition to standardized medical record document output is ultimately achieved.

[0005] The first aspect of this invention proposes a method for automatically generating medical voice medical records based on a large cloud-based model, comprising the following steps: Collect doctors' voice data from medical microphones, and perform semantic analysis on the doctors' voice data to identify term density; The term density is evaluated by context matching to generate a matching degree matrix. Low matching degree segments are identified from the matching degree matrix. Ambiguity disambiguation processing is performed on the low matching degree segments to generate term disambiguation tags. A term calibration mapping is constructed based on the term disambiguation tags. Obtain case structure response data from a large cloud model and perform structural scanning to identify the degree of case structure loss. Based on the degree of case structure loss, perform priority assessment to generate a completion priority factor. Based on the cumulative change rate of the completion priority factor, generate a content allocation coefficient. The terminology calibration mapping is used to analyze case knowledge and identify diagnostic conflicts. Based on the diagnostic conflicts, the case knowledge is analyzed and sorted to generate diagnostic priority tags. The content allocation coefficient and the diagnostic priority tags are sorted in a consistent manner to identify sorting deviation segments. Based on the sorting deviation segments, key nodes for case generation are established. For the key nodes of case generation, output a case draft document, perform doctor modification frequency statistics on the case draft document to generate error-prone area markers, and use the content allocation coefficient to perform key verification and allocation on the error-prone area markers to generate standardized case documents.

[0006] A second aspect of this invention proposes an automatic medical voice medical record generation system based on a cloud-based large model, comprising: The voice acquisition module is used to acquire doctors' voice data from medical microphones and to perform semantic analysis on the doctors' voice data to identify the density of terms. The terminology calibration module is used to perform context matching evaluation on the terminology density to generate a matching degree matrix, identify low matching degree segments from the matching degree matrix, perform ambiguity resolution processing on the low matching degree segments to generate term disambiguation tags, and construct a terminology calibration mapping based on the term disambiguation tags. The structural analysis module is used to acquire case structural response data from a large cloud model and perform structural scanning to identify the degree of missing case structures. Based on the degree of missing case structures, priority assessment is performed to generate a completion priority factor, and a content allocation coefficient is generated based on the cumulative change rate of the completion priority factor. The diagnostic processing module is used to analyze case knowledge and identify diagnostic conflicts through the terminology calibration mapping, generate diagnostic priority tags based on the diagnostic conflicts, perform sorting consistency detection on the content allocation coefficient and the diagnostic priority tags to identify sorting deviation segments, and establish key nodes for case generation based on the sorting deviation segments. The case generation module is used to output a case draft document for key nodes of case generation, perform doctor modification frequency statistics on the case draft document to generate error-prone area markers, and use the content allocation coefficient to perform key verification and allocation of the error-prone area markers to generate standardized case documents.

[0007] The beneficial effects of this invention are reflected in the following points: 1. Semantic analysis of doctors' speech data is performed to identify terminology density. Combined with multi-granularity context window matching evaluation and ambiguity resolution, the speech-transcribed text, which mixes colloquial expressions and professional terms, is transformed into terminology calibration mapping with clear semantic orientation. This effectively solves the problems of non-standard terminology expression in doctors' statements and the same expression pointing to different clinical meanings in different contexts, thus improving the accuracy of speech content conversion to standard case terminology. 2. Missing data scanning and cascading propagation analysis are performed on case structure response data generated by a large cloud model to identify the degree of impact of missing fields on the overall structure. Combined with the importance ranking of elements, a completion priority factor is generated, and the content allocation coefficient is dynamically adjusted based on the cumulative change rate of the priority factor. This allows completion resources to be flexibly allocated according to the current case structure status, improving the problem that fixed template schemes cannot effectively distinguish completion priorities when key fields are missing. 3. Diagnostic conflicts are identified and prioritized through terminology calibration mapping to generate diagnostic priority markers. Consistency testing of the ranking is performed using content allocation coefficients. After correcting deviations in the two ranking systems, key nodes for case generation are established to drive the orderly output of case draft documents. Error-prone areas are identified based on the distribution of doctors' historical modification frequency and the cascading modification relationships between fields. Content allocation coefficients are used to perform differentiated verification and allocation of error-prone areas, specifically addressing the lack of a key verification mechanism for regularly error-prone fields in case drafts, thus improving the overall output quality of standardized case documents. Attached Figure Description

[0008] Figure 1 This is a flowchart of the method for automatically generating medical voice medical records based on a large cloud model, as described in this invention.

[0009] Figure 2 This is a structural block diagram of the medical voice medical record automatic generation system based on a cloud-based large model, which is the subject of this invention. Detailed Implementation

[0010] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0011] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0012] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0013] The technical solutions of the embodiments of this application will be described below.

[0014] like Figure 1 As shown, this embodiment of the invention provides a method for automatically generating medical voice medical records based on a large cloud model, including the following steps S110-S150: Step S110: Collect doctor's voice data from a medical microphone, and perform semantic analysis on the doctor's voice data to identify term density.

[0015] Specifically, the process involves collecting doctors' voice data using medical microphones. The medical microphones continuously collect raw audio signals during the doctor's consultation process. If multiple sound sources cause persistently low signal-to-noise ratio (SNR) in the doctor's voice data, a beamforming algorithm adaptively calibrates the direction of the medical microphone array, focusing the pickup on the main speaking position. After calibration, the SNR of the doctor's voice data typically improves by 8 to 15 dB. If the SNR is still below the usable threshold, the corresponding segment is marked as unreliable, and the transcription process is paused. Valid voice segments are segmented from the continuous audio stream, each segment corresponding to a single complete statement by the doctor. Segment boundaries are determined by energy mutation points. When background noise interference is strong, the boundary calibration threshold is adaptively increased to prevent noise from being misjudged as valid content. Segments in the doctor's voice data with silence duration exceeding a threshold are marked as missing segments. Missing segments are not included in subsequent parsing. When the missing segment ratio exceeds 20% of the entire segment, a collection quality alarm is triggered, prompting a re-recording of the corresponding segment. Doctors' speech data is transcribed into text sequences through acoustic modeling. Segments with abnormally fast speech rate trigger secondary recognition to reduce character error rate. For example, when doctors describe the history of acute illnesses, their speech rate is significantly faster. Secondary recognition combines speech rate parameters to re-acoustically decode the segment and finally outputs the result with higher confidence. The text sequence is divided and stored in units of semantic paragraphs, and the boundaries of each semantic paragraph are aligned with the boundaries of the valid speech segments.

[0016] Semantic analysis was performed on the physician's speech data to identify terminology density. Medical terms were extracted segment by segment from the transcribed text sequence of the physician's speech data. Term localization was based on dual verification using a medical dictionary and contextual semantics. Simply relying on dictionary matching could misclassify common words as technical terms; semantic verification significantly reduced the misclassification rate. When the two verification results were inconsistent, the semantic verification conclusion prevailed. Terminology density was determined by the ratio of the number of times a term appeared within each semantic segment to the total number of words in the segment. A higher ratio indicated a higher concentration of terms in that segment. When physicians focused on stating diagnostic criteria, professional terms such as "pulmonary embolism," "D-dimer," and "deep vein thrombosis" were highly concentrated, resulting in significantly higher terminology density ratios. Transitional segments in the consultation process, characterized by colloquial descriptions, typically had lower terminology density ratios. The transcription quality of the physician's speech data directly affected the completeness of terminology extraction. Segments with a character error rate exceeding 5% were marked as low-quality in terms of terminology density. These low-quality marked segments had reduced weights rather than being directly excluded, in order to retain some of their semantic contribution. Semantic paragraphs with abnormally low terminology density usually correspond to situations where doctors use descriptive language instead of standard terms. For example, when "atrial fibrillation" is described as "irregular heartbeat" or "insulin resistance" is described as "problem with blood sugar regulation," the terminology density ratio of such paragraphs is close to zero, but they actually carry important diagnostic information. The original text sequence of the corresponding paragraph is completely preserved and output along with the terminology density.

[0017] Step S120: Context matching evaluation is performed on term density to generate a matching degree matrix. Low matching degree segments are identified from the matching degree matrix. Ambiguity resolution processing is performed on the low matching degree segments to generate term disambiguation tags. Term calibration mapping is constructed based on the term disambiguation tags.

[0018] In some embodiments, the step of performing contextual matching evaluation on the term density to generate a matching degree matrix includes: constructing multi-granularity context windows based on word-level context, sentence-level context, and paragraph-level context for the term density, performing hierarchical matching detection to generate a multi-granularity inconsistency position set; performing cross-granularity cross-validation to filter co-occurring inconsistency positions from the multi-granularity inconsistency position set to generate a misuse intensity distribution; extracting high-frequency misuse contexts from the misuse intensity distribution as high-risk context combinations; and performing matching degree annotation on the high-risk context combinations to generate a matching degree matrix.

[0019] A multi-granularity context window is constructed based on term density at the word, sentence, and paragraph levels for hierarchical matching detection, generating a multi-granularity inconsistency location set. Term density drives multi-granularity context detection for each term within a corresponding semantic paragraph based on its ratio to the paragraph's semantic content. Paragraphs with high term density undergo complete detection at all three granularity levels: word, sentence, and paragraph. Paragraphs with low term density undergo detection only at the word and sentence levels. Paragraphs with zero term density are excluded from the detection process entirely, without generating any inconsistency records. The word-level window checks the semantic collocation compliance of adjacent words with a single term as the center. A word-level inconsistency is triggered when "pneumonia" is paired with "surgical resection," while a collocation of "pneumonia" with "antibiotic treatment" is compliant. The sentence-level context window covers the entire sentence containing the term, checking whether the term's grammatical role is consistent with the standard usage scenario. A sentence-level inconsistency is triggered when "elevated blood sugar" appears in a treatment plan sentence within a diagnostic context. The paragraph-level context window covers the entire semantic paragraph containing the term, checking whether the semantic references of multiple terms within the same paragraph are consistent. A paragraph containing tumor marker-related terms within a paragraph indicating infectious diseases is triggered as a paragraph-level inconsistency. The multi-granularity inconsistency location set labels each inconsistency record at its respective granularity level. When the same term appears repeatedly across granularities in the multi-granularity inconsistency location set, priority weights are superimposed. Entries with a weight exceeding a threshold after superposition are marked as high-risk locations. Semantic paragraphs with a high density of high-risk locations have the highest priority weight for inconsistency records. When the proportion of high-risk locations in the multi-granularity inconsistency location set exceeds 30%, there is a systematic term misuse bias in the current consultation voice.

[0020] Cross-granularity cross-validation is used to filter co-occurrence inconsistency locations across a multi-granularity inconsistency location set, generating a misuse intensity distribution. Cross-validation is based on whether a term triggers inconsistency annotation simultaneously across all executed granularity dimensions. For paragraphs with high term density, simultaneous triggering at the word, sentence, and paragraph levels is required. For paragraphs with low term density, only word and sentence level detection is performed, and the corresponding co-occurrence inconsistency judgment condition is adjusted to simultaneous triggering at two granularities. False alarms only exist at locations where a single granularity trigger occurs. For example, when a doctor explains test results to a patient, they use colloquial language to mention "cardiac enzymes." The adjacent words are non-standard collocations such as "get a blood test," resulting in a low word-level match. However, the overall semantics of the sentence clearly point to cardiac injury marker detection, and the paragraph revolves around the differentiation of acute chest pain. Both sentence and paragraph levels match normally, so this location is filtered out and not included in the co-occurrence inconsistency count. The filtering operation effectively reduces noise interference in the misuse analysis. The spatial density of co-occurring inconsistencies at multiple granularities reflects the overall severity of misuse in the corresponding semantic paragraph. For example, when a doctor describes digestive system diseases by mixing respiratory system terms, this paragraph exhibits a high concentration of co-occurring inconsistencies at multiple granularities, resulting in a significant peak in the misuse intensity distribution at that location. The misuse intensity distribution uses each semantic paragraph as a statistical unit, with the density value determined by dividing the number of co-occurring inconsistencies within the paragraph by the total number of terms in the paragraph. A higher density value indicates a more severe degree of term misuse in that paragraph. A flat density distribution curve suggests that misuse is dispersed throughout the entire consultation process, while a concentrated peak indicates a significant decrease in the standardization of the specific diagnostic statement. Locations with a density exceeding twice the global mean in the misuse intensity distribution are marked as high-intensity misuse areas. The paragraph matching criteria for high-intensity misuse areas are more stringent, and the inconsistency threshold is correspondingly tightened. When the set of multiple granularities of inconsistencies is empty, the entire misuse intensity distribution is assigned a zero value, and the extraction process is skipped.

[0021] High-frequency misuse contexts in the misuse intensity distribution were extracted as high-risk context combinations. The determination of high-frequency misuse contexts was based on the number of times the same contextual pattern triggered peaks in the misuse intensity distribution. A repetition threshold was set at 3 times; peaks below this threshold were considered occasional misuses and not included in the extraction. Occasional misuses are usually related to doctors' temporary verbal expression deviations rather than systemic terminology cognitive biases. The doctor repeatedly mixed "dyspnea" with cardiac function terms in multiple different paragraphs of the chief complaint. This pattern repeatedly triggered peaks in the misuse intensity distribution and reached the threshold, thus being extracted as a member of the high-risk context combination. Each member in the high-risk context combination corresponds to a specific type of high-frequency misuse contextual pattern. A concentration of member types suggests a fixed tendency for terminology bias in the doctor, while a dispersion of member types suggests insufficient overall contextual standardization. The contextual constraint strategies corresponding to these two situations are fundamentally different. Semantic segments with a density consistently below 50% of the mean in the misuse intensity distribution are not included in the extraction. Their corresponding positions are marked as exempted positions in high-risk context combinations. The terminology compliance of these exempted positions is considered verified and does not consume disambiguation resources. When the overall mean of the misuse intensity distribution is below 0.1, high-risk context combinations are considered empty, and the matching degree annotation process is skipped.

[0022] Matching degree annotation is performed on high-risk context combinations to generate a matching degree matrix. The matching degree annotation evaluates the semantic paragraphs covered by each member in the high-risk context combination one by one. The matching degree score M of each term in the current context is determined by the semantic distance d_s and the context relevance r_c, M=r_c×(1-d_s / d_max), where d_max is the standardized maximum semantic distance, r_c ranges from 0 to 1, and d_s ranges from 0 to d_max. When d_s exceeds d_max, d_s is truncated to d_max, and the corresponding M value is zero. In the context of coronary heart disease diagnosis, "chest tightness" has a high semantic distance d_s and r_c close to 1.0 with the standard term "squeezing pain in the precordial region", corresponding to an M value of about 0.35, which is lower than the qualified threshold of 0.6, triggering low matching degree annotation. The corresponding standard expression should be "angina pectoris" or "squeezing pain in the precordial region". The matching degree matrix is ​​structured with semantic paragraphs as rows and contextual levels as columns. Each element is filled with the M value of the corresponding paragraph at the corresponding contextual level. The mean value of the row corresponding to paragraphs covered by high-risk context combinations is generally lower than that of uncovered paragraphs. When the difference in matching degree between the two types of paragraphs exceeds 0.3, the high-risk context combination is confirmed as valid. When the difference is less than 0.1, the rationality of the context pattern grouping is re-examined. Elements corresponding to exemption positions in the matching degree matrix are assigned a full score of 1.0 and are not included in the low score statistics. When the proportion of low-matching segments in the overall matching degree matrix exceeds 40%, the case quality is marked as needing optimization. When the proportion is less than 10%, the speech standardization is good.

[0023] Low-match segments are identified from the matching matrix. The M-value of each element in the matching matrix is ​​screened one by one according to a preset low-match threshold. The low-match threshold is consistent with the qualified threshold in the matching labeling stage, set to 0.6. Elements with an M-value lower than this threshold are determined to be low-match positions. When the number of low-match positions in a row corresponding to a semantic paragraph exceeds 50% of the total number of elements in that row, the entire semantic paragraph is marked as a low-match segment. When it is less than 50%, only the low-match positions within the paragraph are recorded, but the entire row is not included in the low-match segment range. The boundary of a low-match segment is determined by the start and end line numbers of consecutive rows where the density of low-match positions continuously exceeds the threshold. Two adjacent rows with dense low-match positions are merged into the same low-match segment when the interval between them is no more than one row; otherwise, they are marked as separate low-match segments. In the matching degree matrix, rows with all columns having a full score of 1.0 correspond to exemption positions. Exempt rows do not participate in the boundary determination of low matching degree segments. Low matching degree segments form a structured index with start and end row numbers and corresponding context level annotations. Each segment in the index records the corresponding semantic paragraph range and average M value. The lower the average M value, the more concentrated the term context matching deviation is within that segment.

[0024] In some embodiments, the step of generating term disambiguation tags based on the low-matching segment includes: identifying ambiguity sources in the low-matching segment to obtain a candidate term distribution; performing conflict accumulation analysis on the candidate term distribution to generate conflict accumulation features; dividing the conflict accumulation features into ambiguity-dense regions and ambiguity-sparse regions; and assigning disambiguation priorities to the ambiguity-dense regions and the ambiguity-sparse regions to generate term disambiguation tags.

[0025] For low-match segments, ambiguity sources are identified to obtain the distribution of candidate terms. All candidate terms corresponding to each position in the low-match segment are extracted. Candidates are derived from a set of standard terms in a medical dictionary that are semantically similar to the original expression. The degree of similarity is jointly measured by glyph similarity and semantic vector distance. Candidates with low scores in both metrics are directly excluded. Candidates that are only glyphically similar but semantically different are also excluded to avoid incorrect mapping. "Liver injury" and "impaired liver function" appear as candidates in adjacent positions within the low-match segment. These two terms indicate different severity levels in clinical context. The overlap of such candidates suggests that the segment contains highly ambiguous terminology. Semantic differences between candidate terms at nodes are marked as high-ambiguity nodes. The candidate term distribution uses each position within the low-match segment as a node, associating nodes with corresponding candidate terms and matching confidence. Segments with high node density have concentrated ambiguity sources, while segments with low density have candidate terms pointing to a single direction, making disambiguation relatively easier. When a low-match segment spans multiple semantic paragraphs, the candidate term distributions of each paragraph are processed independently. The semantic coherence of candidate terms across paragraphs is checked separately. If the coherence is insufficient, the two paragraphs are not merged to prevent cross-segment candidates from interfering with each other. Candidates with a confidence score below 0.4 in the candidate term distribution are removed. If a node has no valid candidates after removal, that position is marked as unable to be automatically disambiguated and triggers manual intervention. When the proportion of manually intervened positions in the candidate term distribution exceeds 25%, the overall disambiguation process is paused and awaits manual completion.

[0026] Conflict accumulation analysis is performed on the candidate term distribution to generate conflict accumulation features. Conflict determination is based on whether the candidate term sets of adjacent nodes belong to different system disease diagnostic categories. When two adjacent node candidates come from the circulatory system and digestive system categories respectively, it is determined as a node conflict. When the candidate term distribution contains three consecutive candidate groups of "heart failure", "chronic obstructive pulmonary disease" and "cirrhosis", the three belong to different system diseases, the number of adjacent node conflicts accumulates rapidly, and the conflict accumulation feature shows a high accumulation peak in this segment. The conflict accumulation amount C_i is accumulated node by node according to the position sequence, C_i = C_{i-1} + w_c × conflict_i, and the initial value of the sequence C_0 is 0, where conflict_i is the conflict determination value between the i-th node and the predecessor node (conflict is 1, no conflict is 0), w_c is the conflict weight, w_c is 1.5 when adjacent node candidates come from three or more different diagnostic categories, and 1.0 when they come from two categories. The conflict accumulation feature is composed of the C_i sequence of all nodes. The larger the sequence slope, the more concentrated the ambiguity outbreak in this area. In segments with high candidate term density, the slope of the C_i sequence is generally larger. Nodes C_i without candidate terms remain unchanged and do not affect the overall trend judgment. A continuously flat conflict accumulation feature curve indicates that the semantic direction of candidate terms within a segment is generally consistent. The location of a sudden increase in the curve slope corresponds to the segment boundary where ambiguity sources erupt in concentrated bursts.

[0027] Based on the conflict accumulation characteristics, ambiguity-dense regions and ambiguity-sparse regions are defined. The slope of the conflict accumulation feature curve is determined by dividing the difference in C_i between adjacent nodes by the node spacing. Regions with slopes exceeding the mean of the entire curve are classified as ambiguity-dense regions, while those with slopes below the mean are classified as ambiguity-sparse regions. The mean serves as an adaptive threshold, dynamically adjusted according to the overall complexity of the consultation content. In multi-disease consultation scenarios, the mean is typically higher than in single-disease consultation scenarios. A high overall slope of the conflict accumulation feature indicates severe cross-system terminology mixing in the consultation voice, while a low overall slope indicates that terminology usage is concentrated within a single system. Ambiguity-dense regions typically correspond to paragraphs where doctors intensively use multi-system terminology during the chief complaint statement or differential diagnosis reasoning stage. When multi-disease statements are involved, the scope of ambiguity-dense regions is usually broad. Ambiguity-sparse regions typically correspond to medical history inquiries and transitional statements. When single-system disease terms appear, the candidate set is small, and the automatic disambiguation success rate is generally higher in ambiguity-sparse regions, with significantly lower processing time compared to ambiguity-dense regions. When the slope of the conflict accumulation feature curve repeatedly crosses the mean threshold within a short distance, the corresponding interval is alternately classified into the ambiguity-dense region and the ambiguity-sparse region. This pattern suggests that ambiguity is intermittent rather than continuously concentrated. When the area of ​​the ambiguity-dense region exceeds 60%, the disambiguation pressure is greater, and when the area of ​​the ambiguity-sparse region exceeds 80%, the overall disambiguation difficulty is lower.

[0028] Disambiguation priorities are assigned to densely ambiguous and sparsely ambiguous regions to generate term disambiguation tags. Densely ambiguous regions are assigned high disambiguation priority, and nodes within these regions are processed sequentially according to the local slope of their conflict accumulation features, from highest to lowest. When "atrial fibrillation" and "ventricular premature contractions" both appear at nodes with high slope, disambiguation is prioritized. After all nodes in the densely ambiguous region have been processed, the process moves to the sparsely ambiguous region. Sparsely ambiguous regions are assigned low disambiguation priority, and the competition among candidate terms within these regions is low. "Cough" is a unique candidate in a single upper respiratory tract infection context with a confidence level close to 1.0. Disambiguation of such nodes in the sparsely ambiguous region requires almost no additional resources, and the overall processing speed is much faster than in the densely ambiguous region. The term disambiguation markers are generated by summarizing the disambiguation results of each node. Each marker corresponds to the node position and the standard term used. The confidence level of the term disambiguation markers corresponding to nodes in the ambiguity-dense region is usually lower than that of nodes in the ambiguity-sparse region. Term disambiguation markers with a confidence level lower than 0.5 are subject to additional manual verification. The dense section of verification markers indicates that the overall confidence level of the disambiguation algorithm at that position is insufficient. The verification markers contributed by the ambiguity-dense region usually account for more than 80% of all verification markers, which collectively reflects the concentrated difficulty of disambiguation of nodes in the ambiguity-dense region.

[0029] A terminology calibration map is constructed based on terminology disambiguation markers. The standard term at each node of the terminology disambiguation marker and the corresponding original expression form a mapping entry. Mapping entries are grouped according to the clinical category to which the term belongs. Each category, such as cardiovascular, respiratory, and digestive systems, maintains its own independent mapping segment. Mapping entries within the same category share contextual constraints to avoid contradictory mapping directions for similar terms in different contexts. Terminology disambiguation markers with a confidence level higher than 0.8 directly generate definitive mapping entries. Terminology disambiguation markers with a confidence level between 0.5 and 0.8 generate mapping entries awaiting confirmation and are marked with a manual review flag. Terminology disambiguation markers with a confidence level lower than 0.5 do not generate entries at their corresponding positions initially; these entries are merged into the terminology calibration map after manual completion. Deterministic mapping entries and pending mapping entries are managed hierarchically according to contextual tags to manage ambiguous mapping relationships. "Chest pain radiating to the back" maps to "tearing chest pain" in the context of aortic dissection, and "dizziness and blurred vision" maps to "orthostatic hypotension" in the context of hypotension. The same original expression can point to different standard terms in different contexts. When the same original expression is associated with multiple contextual tags, the corresponding mapping entry is dynamically selected according to the contextual tag matching priority. The wider the coverage of the term disambiguation markers, the higher the entry coverage of the term calibration mapping. When the entry coverage is less than 60%, manually entered supplementary entries are merged with automatically generated entries. After merging, the coverage of the term calibration mapping is reassessed. Once the coverage reaches the target, the merging process terminates.

[0030] Step S130: Obtain case structure response data from the cloud-based large model and perform structure scanning to identify the degree of case structure missingness. Based on the degree of case structure missingness, perform priority assessment to generate a completion priority factor. Based on the cumulative change rate of the completion priority factor, generate a content allocation coefficient.

[0031] Specifically, the system acquires case structure response data from a cloud-based large-scale model and performs structural scanning to identify the degree of structural deficiencies in the cases. The case structure response data is jointly generated by the cloud-based large-scale model based on semantically parsed text sequences and terminology calibration mappings. The response content is arranged according to the field structure of the standard case template, covering core fields such as chief complaint, present illness, past medical history, physical examination, auxiliary examinations, diagnostic impression, and treatment plan. The filling status of each field is checked one by one during the structural scanning stage. Fields with complete content are marked as filled, while those with only placeholders or missing content are marked as pending completion. Partially filled fields are converted to intermediate states according to the filling ratio. The case structure missingness D is determined by the weighted sum of the missing states of each field, D = Σ(w_i × m_i) / Σ(w_i). This summation iterates through all n case template fields, where w_i is the importance weight of the i-th field, and m_i is the missing state value of the i-th field. m_i is 1 when a field is completely missing, 0 when it is fully filled, and determined by subtracting the filling ratio from 1 when it is partially filled. The weight w_i is determined by the importance of each field in the case completeness assessment, with the chief complaint and diagnostic impression fields having the highest weights. A significantly higher case structure missingness occurs when the diagnostic impression is completely missing. When multiple high-weight fields are simultaneously missing in the case structure response data, the case structure missingness increases rapidly. In a case structure response data session where both past medical history and auxiliary examinations are empty (both fields belong to the high-weight category), the corresponding case structure missingness value increases significantly. When only low-weight fields are missing in the case structure response data, the case structure missingness is usually within a low range.

[0032] In some embodiments, the step of prioritizing the assessment based on the degree of structural missingness of the case to generate a completion priority factor includes: performing cascading propagation analysis on the degree of structural missingness of the case to identify cascading triggering features; performing structural stability assessment based on the cascading triggering features to generate a structural missing severity; performing element importance analysis on the structural missing severity to extract high-priority elements and low-priority elements; and establishing a priority ranking based on the high-priority elements and the low-priority elements to generate a completion priority factor.

[0033] For example, the step of performing cascading propagation analysis on the case structure missing degree to identify cascading trigger features includes: performing field dependency analysis based on the case structure missing degree to identify primary related fields; expanding from the primary related fields layer by layer along the dependency direction to generate multi-level related field chains; performing influence cutoff filtering on the multi-level related field chains according to the level depth to generate a set of effective influential fields; and extracting fields from the set of effective influential fields where multiple related field chains intersect to form cascading trigger features.

[0034] Field dependency analysis is performed based on case structure missingness to identify primary related fields. Dependency analysis starts with each missing field in the case structure missingness and searches the field dependency graph of the case template for downstream fields directly connected to each missing field. A direct connection is defined as a dependency path length of 1, meaning the downstream field's content directly references the output of the missing field. Indirectly dependent fields, due to path lengths exceeding 1, are not included in the current identification scope. When the "Physical Examination" field is missing in the case structure missingness, the "Diagnostic Impression" field, which directly depends on its result, is directly connected to it, forming a primary related field for that missing field. Although "Treatment Plan" has an indirect dependency on "Physical Examination," its path length is 2, so it is not included in the primary related field scope. The influence of this field will be incorporated into deeper association chains during the layer-by-layer expansion stage. The same missing field may correspond to multiple primary related fields. When multiple missing fields share the same downstream field, that field is labeled multiple times in the primary related fields. The number of labelings reflects how many upstream missing fields influence the field; a higher number of labelings indicates a more concentrated influence from upstream missing fields. The size of the set of primary associated fields increases with the number of missing fields in the case structure. Missing fields with high out-degree in the field dependency graph correspond to more primary associated fields. Missing fields with an out-degree of 1 are usually associated with only a single downstream field, and their propagation range is relatively limited.

[0035] Multi-level dependency chains are generated by expanding layer by layer along the dependency direction from the first-level dependency fields. Each field in the first-level dependency field becomes a second-level node, and the process continues tracing down the directed edges of the dependency graph. The direct downstream nodes of the second-level nodes form the third level, and the direct downstream nodes of the third-level nodes form the fourth level, and so on, extending layer by layer until a leaf node is reached or the preset maximum number of expansion layers is exceeded. The maximum number of expansion layers is usually set to 5 to prevent infinite expansion when loops exist in the dependency graph. Fields located at the core of the dependency graph in the first-level dependency fields typically have a large number of downstream layers. When "Diagnostic Impression" is a second-level node, its downstream "Disposal Plan" forms the third level. If "Disposal Plan" has no downstream fields, the expansion terminates at that branch. The expansion depth of different branches is determined by the length of their respective dependency chains. Each complete path starting from a missing field, extending layer by layer through the first-level dependency fields to a leaf node constitutes a dependency chain. The length of the chain reflects the propagation depth of the missing field's impact. Multi-level dependency chains are formed by summing all dependency chains triggered by all missing fields. The number of chains increases rapidly with the number of missing fields and the complexity of the dependency graph. In a multi-level related field chain, there are shared nodes between different chains. Shared nodes are traversed by multiple chains at the same time. The higher the frequency of occurrence, the more severe the impact of multiple missing nodes in the upstream is on the field. Fields with high annotation frequency in the first-level related fields are usually also frequently shared nodes in the multi-level related field chain.

[0036] The multi-level related field chain is filtered by influence cutoff according to hierarchical depth to generate a set of effective influential fields. The initial influence value of each missing field is determined by normalizing its weighted missing state value m_i×w_i in the case structure missingness by Σ(w_i). The summation is traversed through all case template fields, with values ​​ranging from 0 to 1. The influence decreases with hierarchical depth. The influence decay coefficient of each node in the multi-level related field chain is set to 0.6 per level. The influence of the missing field itself remains at its original value. The influence of the first-level related field (first level) node is multiplied by 0.6, the second-level node by 0.36, the third-level node by 0.216, and so on. The decay rate can be appropriately adjusted according to the dependency chain density of the case template. In a multi-level associated field chain, nodes whose influence after decay is lower than the cutoff threshold of 0.1 are removed entirely. The cutoff threshold of 0.1 is the absolute value within the normalized range of influence. When a node at a certain level of a chain is removed, all nodes below that level are simultaneously removed. All remaining nodes are deduplicated and merged to form an effective influence field set. The deduplication operation retains the maximum influence value of each node in all associated field chains, ensuring that the degree of damage to each field is based on the worst-case scenario. If the "treatment plan" field appears in three associated field chains with influence values ​​of 0.36, 0.6, and 0.216 respectively, then 0.6 is taken as the representative influence value in the effective influence field set. The size of the effective influence field set is determined by the total number of nodes in the multi-level associated field chain and the decay rate. The smaller the decay coefficient, the smaller the size of the effective influence field set. A higher proportion of shared nodes in the effective influence field set indicates a denser multi-level associated field chain structure. Dense structures usually correspond to complex cross-dependencies between fields in multi-disease comorbidity consultation scenarios. The size of the effective influence field set in such scenarios is usually significantly larger than that in single-disease consultation scenarios.

[0037] Fields that intersect multiple related field chains from the effective influencing field set form cascading trigger features. Fields with more than a preset threshold number of intersections are identified as intersection locations. The number of intersections for each field in the effective influencing field set is determined by the frequency of that field's appearance before deduplication in the multi-level related field chains. The threshold is usually set to 2, meaning that it must be covered by at least two independent related field chains simultaneously. Fields covered by a single chain do not have cascading trigger characteristics due to their single influence path. The field with the highest number of intersections in the effective influencing field set is usually located in the core dependency position of the case template. "Diagnostic Impression" is traversed by the chief complaint reasoning chain, auxiliary examination chain, and present medical history chain in most case structures, with an intersection count of 3 or more, making it a high-priority trigger node in the cascading trigger features. Cascade triggering features consist of all fields that meet the intersection conditions and their corresponding intersection strength values. The intersection strength is determined by the number of intersections and the average influence of each chain at that field. When the number of intersections is the same, fields with higher average influence receive greater intersection strength. The number of nodes in a cascade triggering feature is usually significantly less than the total number of fields in the effective influencing field set. The simplified set of triggering nodes makes the structural stability assessment more focused, concentrating computational resources on key locations that truly have a cascade effect. Fields in the effective influencing field set that do not meet the intersection threshold are not included in the cascade triggering feature. These fields are only covered by a single chain, and their missing impact is relatively independent. The simplification effect of the cascade triggering feature is more significant when the size of the effective influencing field set is large.

[0038] Structural stability assessment is performed based on cascading trigger features to generate the severity of structural defects. The density and intensity distribution of high-trigger nodes together reflect the stability of the current case structure. In acute chest pain cases, when both the "present illness" and "auxiliary examinations" fields are missing, both are core nodes in the diagnostic reasoning chain. In cascading trigger features, high-trigger nodes are highly concentrated, indicating a break in the core logic of the case. When high-trigger nodes are dispersed across different fields such as chief complaint, past medical history, and social history, the structural damage is relatively uniform. The completion strategies corresponding to the two modes are fundamentally different. Structural stability assessment calculates the median centrality of each high-trigger node in the field dependency graph based on cascading trigger features. Nodes with high median centrality are traversed by multiple dependency paths. The "present illness" field usually has a high median centrality, and its absence disrupts both the diagnostic reasoning chain and the treatment plan generation chain. In contrast, the "family history" field usually has a low median centrality, and its absence only affects a limited number of downstream fields. The severity of structural defects is weighted and synthesized by the median centrality of each field and the triggering intensity of cascading triggering features. The synthesis formula is S=Σ(b_i×t_i) / N, which sums over all N high-triggering nodes. Here, b_i is the normalized median centrality of the i-th high-triggering node, ranging from 0 to 1, t_i is the corresponding triggering intensity, determined by normalizing the product of the number of intersections and the mean influence of each chain at that field to the range of 0 to 1, and N is the total number of high-triggering nodes. The larger the structural defect severity S, the more severe the structural damage to the case. When the structural defect severity exceeds 0.75, the overall integrity of the case is fundamentally threatened. The more high-triggering nodes in the cascading triggering features, the higher the structural defect severity usually is. When it is below 0.3, the structural damage is relatively mild, mainly affecting marginal fields, with limited impact on the coherence of the core diagnostic logic.

[0039] Element importance analysis is performed to extract high-priority and low-priority elements based on the severity of structural defects. The contribution value of each field's severity determines its effectiveness in restoring overall structural integrity during the completion operation. Fields with high contribution values ​​have the most significant effect on reducing the overall defect rate after completion and should be prioritized when resources are limited. Element importance analysis ranks the contribution values ​​of each field. The top 30% of fields are extracted as high-priority elements, the bottom 40% as low-priority elements, and the middle 30% are dynamically assigned to high or low priority based on the current level of structural defect severity. This dynamic assignment mechanism allows the completion strategy to be adjusted in real time according to changes in the case's structural status. When the overall structural defect severity is high, middle-priority fields tend to be assigned to high-priority elements to expand the completion scope and ensure that as many key fields as possible are covered in a single round of completion. When the overall structural defect severity is low, middle-priority fields tend to be assigned to low-priority elements to reduce unnecessary resource consumption. High-priority elements typically include fields that strongly support the core logic of the case, such as "diagnostic impression," "present medical history," and "auxiliary examinations." Low-priority elements typically include fields that have limited impact on immediate diagnosis, such as "social history" and "family history." The boundary between the two types of elements is dynamically adjusted according to the severity of structural missing data. When the severity of structural missing data exceeds 0.75, the range of low-priority elements is compressed to less than 20% of the total number of fields.

[0040] Priority factors for case completion are generated by establishing a priority ranking system based on high-priority and low-priority elements. The priority ranking arranges the contribution values ​​of each field in high-priority elements from highest to lowest, while the fields in low-priority elements are arranged sequentially after the high-priority elements. Within both categories, completion is performed in descending order of contribution value, ensuring that the completion operation proceeds from highest to lowest benefit in restoring case integrity, preventing low-benefit fields from preempting completion resources for high-benefit fields. The completion priority factor maps the ranking position of all fields to a continuous value from 0 to 1. The field ranked first corresponds to a completion priority factor close to 1.0, and the field ranked last is close to 0. Fields with equal contribution values ​​are broken up according to the standard order in the case template, ensuring the determinism and reproducibility of the ranking results. Among high-priority elements, the "diagnostic impression" field typically has the highest contribution value in emergency scenarios, corresponding to the first-ranked completion priority factor, followed closely by the "auxiliary examinations" field due to its supporting role in the diagnostic impression. Among low-priority elements, fields such as "social history" and "family history" generally have low contribution values, and their corresponding completion priority factor values ​​are usually below 0.3. The absence of these fields has a limited impact on the immediate use of the case. The greater the difference between the overall completion priority factor of low-priority elements and the mean of high-priority elements, the more obvious the degree of priority differentiation of the missing structural elements in this case.

[0041] In some embodiments, generating content allocation coefficients based on the cumulative rate of change of the completion priority factor includes: extracting the cumulative rate of change fluctuation information of the completion priority factor; performing rate change analysis on the cumulative rate of change fluctuation information to identify rate inflection point characteristics; establishing content weight adjustment rules at the rate inflection point characteristics to generate a weight adjustment table; and constructing content allocation coefficients based on the weight adjustment table.

[0042] Extract the cumulative rate of change fluctuation information of the completion priority factor. The completion priority factor is dynamically updated as the simulation progresses through multiple simulation completion iterations. Each simulation iteration assumes that the current highest priority missing field has been filled and recalculates the priority factor of the remaining fields accordingly. The simulation iteration does not actually generate field content but evaluates the expected change trajectory of the priority factor as each field is completed sequentially. In a certain simulation iteration, it is assumed that after the "present medical history" field is successfully filled, the missing status of this field changes from pending completion to completed. The completion priority factor values ​​of the remaining fields are adjusted accordingly. The magnitude and direction sequence of the change of the completion priority factor between adjacent iterations together constitute the cumulative rate of change fluctuation information. After a high-priority field is successfully completed, the corresponding completion priority factor decreases significantly, while after a low-priority field is completed, the decrease in the completion priority factor is smaller. The cumulative rate of change fluctuation information shows continuous large negative fluctuations during the high-priority field concentrated completion phase and slow small fluctuations during the low-priority field completion phase. In a certain simulation iteration, assuming that "diagnostic impression" and "present medical history" are filled in simultaneously, the corresponding completion priority factor drops sharply within a single iteration. The cumulative rate of change fluctuation information shows a sharp negative peak at this position, reflecting the structural reorganization of priority brought about by the concentrated completion of high-weight fields. When the completion priority factor remains stable for a long time during the iteration, the cumulative rate of change fluctuation information approaches zero. This state indicates that the current simulation completion progress has stalled, and the remaining uncompleted fields have a limited impact on the overall priority pattern. When the overall amplitude of the cumulative rate of change fluctuation information is small, the inflection point judgment threshold is tightened accordingly to ensure the sensitivity of inflection point detection.

[0043] Rate change analysis is performed on the cumulative rate of change fluctuation information to identify rate inflection point characteristics. Rate change analysis calculates the second difference of the numerical sequence of the cumulative rate of change fluctuation information. The location of a sudden increase in the absolute value of the second difference corresponds to the iteration node with the most dramatic rate change. The weighting of the content before and after this node shifts significantly. In clinical scenarios, this shift usually corresponds to the transition from the core diagnostic information collection stage to the auxiliary information completion stage in the completion process. In emergency chest pain cases, after the core fields such as chief complaint, present illness, and auxiliary examinations are completed, the remaining fields to be completed become auxiliary fields such as past medical history and family history. This transition is represented by a sudden increase in the absolute value of the second difference. After the "auxiliary examinations" field is completed, the previously rapidly decreasing cumulative rate of change fluctuation information suddenly flattens out, and the absolute value of the second difference at this location is significantly higher, marking a rate inflection point characteristic node. The inflection point determination criterion is that the absolute value of the second-order difference exceeds twice the sequence mean. Positions below this multiple are considered normal rate fluctuations and are not included in the rate inflection point feature. Too low a multiple will introduce a large number of false inflection points that interfere with the weight adjustment logic, while too high a multiple will cause the true inflection point to be missed. When the second-order differences of several consecutive iteration positions in the cumulative rate of change fluctuation information are all high, the position with the largest absolute value is taken as the representative inflection point to avoid excessive density of adjacent inflection points and redundancy in the rate inflection point feature. When the number of rate inflection point features is small, it indicates that the overall completion process is stable. When the number is large, it indicates that the priority structure has undergone multiple jumps and reorganizations during the completion process. When the overall variance of the cumulative rate of change fluctuation information is small, the rate inflection point features are usually an empty set.

[0044] Content weight adjustment rules are established at rate inflection point features to generate a weight adjustment table. Each inflection point divides the completion process into several stages. Within each stage of an adjacent inflection point, the completion strategy remains relatively stable. When crossing an inflection point, the strategy switches according to the priority structure. The content weight allocation for each stage is determined independently and does not interfere with each other. The magnitude of the rate change at each rate inflection point determines the adjustment intensity of the content weight at that inflection point. When the rate changes abruptly from a rapid decline to a stable state, it indicates that high-priority fields have been largely completed, and the content weight for the current stage should be tilted towards the remaining low-to-medium priority fields. When the rate changes abruptly from a gradual decline to an accelerated decline, it indicates that a new batch of high-priority fields has entered the completion range, and the weight should be re-concentrated towards high-weight fields. In a certain consultation, the completion of "physical examination" triggered a rate inflection point feature. After the inflection point, the rate slowed down. Correspondingly, at this inflection point in the weight adjustment table, the content weights of "social history" and "family history" were slightly increased, and the overall weight focus shifted appropriately from core diagnostic fields to supplementary information fields. When the rate inflection point feature is an empty set, the weight adjustment table only contains default uniform weight entries. No dynamic adjustment is performed throughout the process. Each field participates in content allocation according to its static importance weight. The number of entries in the weight adjustment table corresponds one-to-one with the number of inflection points in the rate inflection point feature. The denser the inflection points, the more frequent the adjustment instructions.

[0045] Content allocation coefficients are constructed based on a weight adjustment table. Each weight adjustment instruction is executed sequentially according to its iteration number. The weight adjustment table continuously updates the current weight values ​​of each field. After all instructions are executed, the final weight values ​​of each field are normalized to the range of 0 to 1. The normalized results constitute the basic numerical vector of the content allocation coefficients. Instructions with larger adjustment ranges in the weight adjustment table significantly affect the final distribution of the content allocation coefficients, while instructions with smaller adjustment ranges only cause localized minor adjustments. The "Diagnostic Impression" field is assigned high weights in multiple instructions in the weight adjustment table, resulting in a final content allocation coefficient close to 1.0. Conversely, the "Family History" field has relatively low weights at all stages, with corresponding content allocation coefficients typically below 0.2. When the weight adjustment table uses the default uniform weights, the values ​​of each field in the content allocation coefficients tend to be consistent. The differences mainly stem from the static weight differences in field importance. When the dynamic adjustment mechanism is not activated, the content allocation coefficients reflect the inherent priority structure of the case template fields themselves. When the content allocation coefficients are highly concentrated in a few fields, it indicates that the priority differentiation in the current completion stage is obvious. When the distribution is relatively even, it indicates that the urgency of completing each field is similar. The more entries in the weight adjustment table, the greater the deviation between the final distribution of the content allocation coefficients and the initial static weights.

[0046] Step S140: Case knowledge is parsed and diagnostic conflicts are identified through terminology calibration mapping. Based on the diagnostic conflicts, diagnostic priority labels are generated by identification and sorting. The sorting consistency of the content allocation coefficient and diagnostic priority labels is detected to identify sorting deviation segments. Based on the sorting deviation segments, key nodes for case generation are established.

[0047] Specifically, diagnostic conflicts are identified through case knowledge analysis using terminology calibration mapping. Case knowledge analysis matches standard terms from the terminology calibration mapping with diagnostic rules in the case knowledge base, which is a pre-deployed, built-in structured rule library independent of the data flow at each step. It covers co-occurrence relationships, causal chains, and exclusion logic between terms. When "aortic dissection" and "common hypertension" co-occur in the same differential diagnosis context, an exclusion rule check is triggered. If the exclusion rule fails, a diagnostic conflict marker is directly generated. Diagnostic conflicts are triggered by two types of situations: one is when multiple diagnostic hypotheses corresponding to the same symptom cluster have mutually exclusive diagnostic evidence in the knowledge base; the other is when the treatment plan stated by the doctor is logically incompatible with the current diagnostic impression under the rules. For example, if a patient's case simultaneously presents the diagnostic impression of "acute myocardial infarction" and the "warfarin anticoagulation therapy" regimen, it is determined to be a medication diagnostic conflict. Diagnostic conflict markers record the terminology pairs involved in the conflict, the triggering rule number, and the conflict type. Each diagnostic conflict marker is associated with two competing diagnostic hypotheses involved in the conflict. The competing diagnostic hypotheses are directly extracted from the exclusion or incompatibility relationships in the triggering rules. The clarity of the exclusion relationship between the two competing diagnostic hypotheses is determined by the rule confidence of the corresponding triggering rule in the knowledge base. The higher the rule confidence, the clearer the exclusion relationship between the two diagnostic hypotheses. The higher the coverage of the terminology calibration mapping, the more comprehensive the types of diagnostic conflicts that case knowledge parsing can detect. When the coverage is insufficient, some terms will bypass conflict detection because they are not mapped to standard expressions.

[0048] In some embodiments, the step of generating a diagnostic priority label based on the diagnostic conflict identification and ranking includes: evaluating the identification path of the diagnostic conflict to generate an identification difficulty value; selecting the least difficult path pair according to the identification difficulty value to generate an optimal identification path; identifying a diagnostic decision node on the optimal identification path; and generating a diagnostic priority label by performing probability ranking through the diagnostic decision node.

[0049] A differential diagnosis path assessment is performed to generate a differential diagnosis difficulty value for diagnostic conflicts. For each pair of competing diagnostic hypotheses in a conflict, the assessment searches the case knowledge base for a sequence of examination items that can distinguish between them. Each sequence of examination items constitutes a differential diagnosis path; the longer the path, the more examination steps are required to differentiate the two diagnostic hypotheses. In diagnostic conflicts, the differential diagnosis path between "acute myocardial infarction" and "aortic dissection" is usually short, requiring only myocardial enzyme analysis and aortic CTA, resulting in a relatively low differential diagnosis difficulty value. Conversely, the differential diagnosis path between "functional dyspepsia" and "early gastric cancer" is longer, requiring multiple endoscopic and pathological examinations for a definitive diagnosis, resulting in a relatively high differential diagnosis difficulty value. The identification difficulty value D_v is determined by a weighted average of three factors: identification path length, examination accessibility, and diagnostic resolution. The calculation formula is D_v = α_v × L_n + β_v × (1 - A_c) + γ_v × (1 - R_d), where L_n is the normalized path length, ranging from 0 to 1, determined by dividing the actual number of path nodes by the maximum number of conflicting path nodes in the case knowledge base; A_c is the examination accessibility coefficient, ranging from 0 to 1, determined by the equipment configuration information of the medical institution to which the current case belongs. Information is dynamically read; the higher the accessibility, the closer A_c is to 1. R_d is the diagnostic discrimination coefficient, ranging from 0 to 1, determined by the average of the positive and negative predictive values ​​of the examination results at each node in the path. The stronger the discrimination, the closer R_d is to 1. α_v, β_v, and γ_v are the weighting coefficients for each item, with the sum of the three being 1. The default values ​​are 0.3, 0.4, and 0.3, respectively. D_v ranges from 0 to 1; a value close to 0 indicates that the two diagnostic hypotheses are easily distinguishable, while a value close to 1 indicates that differentiation is extremely difficult. The differentiation difficulty values ​​of all diagnostic pairs in diagnostic conflicts are summarized and arranged from low to high to form a differentiation difficulty ranking sequence.

[0050] The optimal identification path is generated by selecting the least difficult path pairs based on their identification difficulty values. The selection process starts with ranking the identification difficulty values ​​and sequentially extracts the diagnostic pairs with the lowest difficulty values ​​and their corresponding identification paths. Diagnostic pairs with the lowest difficulty values ​​are prioritized for identification, achieving the highest diagnostic differentiation efficiency with minimal examination costs. A path pair refers to a shared examination path that can simultaneously provide identification information for multiple diagnostic pairs. A path that satisfies the identification needs of two different sets of diagnostic pairs has higher priority than a path that only satisfies a single diagnostic pair. Multiple diagnostic pairs with similar identification difficulty values ​​that share path nodes are merged into a path group for unified processing. Path merging effectively reduces redundant examinations and lowers the overall identification cost. The optimal identification path is selected from a candidate path set, which includes all paths with identification difficulty values ​​below a set threshold of 0.6. Paths above the threshold are considered difficult to execute in the current case scenario and are not included in the selection range for the optimal identification path. If no usable path is available below the identification difficulty threshold, the threshold is raised to 0.8. If the result is still empty after further filtering, it indicates that the current diagnostic conflict cannot be effectively identified under the existing knowledge base rules. The corresponding diagnostic pair triggers a manual expert consultation marker, which does not affect the normal generation of optimal identification paths for other diagnostic pairs. Once the optimal identification path is determined, the path node sequence is fixed. The node order is arranged from high to low information gain of each node. Diagnostic pairs with lower identification difficulty values ​​typically have fewer nodes in their corresponding optimal identification path.

[0051] Diagnostic decision nodes are identified along the optimal differential diagnosis path. A diagnostic decision node is defined as an examination node on the optimal differential diagnosis path that has a decisive influence on the final diagnostic direction. The examination result of this node can directly exclude one or more competing diagnostic hypotheses. Not all nodes in the optimal differential diagnosis path have decision-making capabilities; some nodes only provide auxiliary support information and do not directly change the diagnostic direction. The identification process evaluates each node in the optimal differential diagnosis path one by one, based on the positive and negative predictive values ​​of the node's examination results. A high positive predictive value indicates a strong likelihood of a diagnosis being established when the result is positive, while a high negative predictive value indicates that a diagnosis can be effectively excluded when the result is negative. Myocardial enzyme profiles show high positive and negative predictive values ​​on the optimal differential diagnosis path for acute myocardial infarction, constituting a typical diagnostic decision node. Nodes on the optimal differential diagnosis path with at least one positive or negative predictive value exceeding 0.85 are marked as diagnostic decision nodes; the remaining nodes are marked as auxiliary nodes. Auxiliary nodes are retained in the optimal differential diagnosis path as supplementary evidence. The examination results of auxiliary nodes only provide supporting information and do not independently change the priority ranking of diagnostic hypotheses. The number of diagnostic decision nodes on the optimal identification path is usually 2 to 4, and the order of the diagnostic decision nodes on the optimal identification path is arranged from high to low according to their information gain.

[0052] Diagnostic priority labels are generated by performing probability ranking through diagnostic decision nodes. The posterior probability of each competing diagnostic hypothesis under the current case evidence is calculated using the probability distribution of the examination results from the diagnostic decision nodes. The posterior probability update employs a Bayesian inference framework, calculated as P(D_k|E)=P(E|D_k)×P(D_k) / Σ(P(E|D)×P(D)), where D_k is the k-th diagnostic hypothesis, E is the examination result of the current diagnostic decision node, and P(D_k) is the prior probability of diagnostic hypothesis D_k, determined by the baseline incidence rate of each diagnosis in the case knowledge base. P(E|D_k) represents the likelihood of observing examination result E under the condition that diagnostic hypothesis D_k is true. It is determined by the probability distribution of examination results corresponding to the diagnosis in the knowledge base. In the denominator, Σ(P(E|D)×P(D)) is the sum of all competing diagnostic hypotheses, and P(D_k|E) is the updated posterior probability. The posterior probability is updated once after each diagnostic decision node, and the posterior probability of the previous node is passed on as the prior probability of the next node. The update effects of multiple diagnostic decision nodes are cumulative. In a patient's case, the prior probability of "acute coronary syndrome" is 0.35. After being updated by two diagnostic decision nodes, electrocardiogram and myocardial enzyme spectrum, the posterior probability rises to 0.78, ranking first in the corresponding diagnostic priority label. Meanwhile, the posterior probability of "aortic dissection" drops to 0.08, ranking last in the corresponding diagnostic priority label. The difference between the two posterior probabilities is 0.7, which is a significant difference. Diagnostic priority labels are sorted from high to low posterior probability. Adjacent diagnostic hypotheses with a posterior probability difference of less than 0.05 are labeled as having equal priority. When the coverage of diagnostic decision nodes is insufficient, low confidence labels are added to the diagnostic priority labels to indicate that the sorting results may deviate from the true diagnostic probability distribution due to incomplete evidence.

[0053] A consistency check is performed on the content allocation coefficient and diagnostic priority markers to identify sorting deviations. The consistency check uses the priority ranking of each diagnostic hypothesis in the diagnostic priority markers as a benchmark. Through a pre-built mapping table of diagnostic hypotheses and case field support relationships in the case knowledge base, the priority ranking of each diagnostic hypothesis is converted into the expected weight ranking of the corresponding core supporting fields. This mapping table is constructed based on clinical diagnostic logic. The establishment of each diagnostic hypothesis depends on key evidence provided by specific fields. The diagnosis of "acute myocardial infarction" relies on the accurate recording of ECG and myocardial enzyme spectrum results, as well as typical chest pain symptoms. When this diagnostic hypothesis has the highest priority ranking, the corresponding "auxiliary examination" and "diagnostic impression" fields in the mapping table have the highest expected weight ranking. This process continues to form the expected weight ranking of fields, which is then compared one by one with the actual weight ranking of each case field based on the content allocation coefficient. The ranking deviation segment is defined as a field interval where there is a significant inconsistency between the expected weight ranking of the fields and the actual weight ranking of the content allocation coefficients. The degree of inconsistency is measured by the absolute value of the difference between the rankings of the corresponding fields in the two sets of rankings. A difference exceeding 3 places is considered a deviation. In a certain case, the "diagnostic impression" field ranked first in the expected weight ranking but only fifth in the content allocation coefficient, a difference of 4, triggering the ranking deviation segment labeling. After the ranking deviation segment is identified, the weights of each deviating field in the content allocation coefficient are corrected towards the expected weight ranking. The correction magnitude is determined by the degree of deviation; the larger the deviation difference, the larger the correction magnitude. When the ranking deviation segment is an empty set, the content allocation coefficients and the expected weight ranking of the fields driven by the diagnostic priority label are already highly consistent, and no correction operation is required.

[0054] Key nodes for case generation are established based on ranking deviation segments. After the ranking deviation segments are corrected, fields with a ranking consistency higher than the threshold of 0.8 after correction receive both high content allocation weight and high diagnostic priority. These fields constitute the core members of the key nodes for case generation. Ranking consistency is determined by the difference between the expected weight ranking of the field and the actual weight ranking of the content allocation coefficient. The calculation formula is consistency_i=1-|rank_diff_i| / N_f, where N_f is the total number of fields participating in the ranking, rank_diff_i is the difference in ranking of the i-th field in the two rankings, and consistency_i ranges from 0 to 1. When the ranking difference is zero, the ranking consistency score is 1.0. The larger the difference, the lower the ranking consistency. The greater the correction magnitude of the sorting deviation segment, the more significant the difference between the final composition of the case generation key node and the original composition. In one consultation, the sorting consistency of the "auxiliary examination" field improved from 0.62 to 0.85 after the sorting deviation segment correction, crossing the threshold and being included in the case generation key node. The "diagnostic impression" field thus had complete examination result support, significantly improving the field coverage of the case diagnosis reasoning chain. The field with the highest sorting consistency score in the case generation key node receives the highest ranking. Fields with similar consistency scores break ties based on the absolute value of their content allocation coefficients, ensuring the ranking within the key node is deterministic. When the sorting deviation segment is an empty set, the content allocation coefficient and the diagnostic priority label are highly consistent. The case generation key node is directly composed of two sets of fields that both rank highly, without requiring correction. This situation typically occurs in consultation scenarios where the doctor's speech is highly standardized and the terminology calibration mapping coverage is complete.

[0055] Step S150: Output a draft case document for key nodes in case generation; statistically analyze the frequency of doctor modifications in the draft case document to generate error-prone area markers; and use content allocation coefficients to perform key verification and allocation of error-prone area markers to generate standardized case documents.

[0056] Specifically, a draft case document is output for each key node in the case generation process. Fields are generated sequentially based on their consistency scores, from highest to lowest. The field with the highest score in the key node is filled first, and the generated content prioritizes the diagnostic direction established by the key node, avoiding deviations from the differential ranking results in diagnostic logic. The field filling order in the draft case document strictly corresponds to the internal order of the key nodes. The chief complaint and diagnostic impression fields are typically located early in the key nodes, and their content is generated first, serving as semantic anchors for the remaining fields. The treatment plan field depends on the diagnostic impression; its generation is triggered only after the diagnostic impression field is filled. Non-key node fields in the draft case document are supplemented and generated after all fields in the key nodes are filled, based on weighted coefficients assigned to the remaining content. The supplementary field content is constrained by the content already generated in the key nodes, and semantic contradictions with the key node content are not permitted. In a particular consultation, the "Diagnostic Impression" field was prioritized at the key node of case generation. After the generated content was confirmed as "acute myocardial infarction," the generated content of the "Treatment Plan" field in the case draft document was subsequently anchored to the corresponding antithrombotic and interventional treatment plan, maintaining strict logical consistency with the diagnostic impression. After the case draft document was generated, a full field integrity scan was performed again. If any field content was still missing, a secondary completion was triggered. The secondary completion prioritized calling low-priority fields outside the key node of case generation.

[0057] In some embodiments, the step of generating error-prone area markers by statistically analyzing the frequency of doctor modifications to the draft medical record document includes: generating a modification frequency distribution by statistically analyzing the frequency of doctor modifications to the draft medical record document; constructing a cascading modification correlation graph between fields based on the modification frequency distribution to generate the degree of modification impact; identifying the scope of error impact based on the degree of modification impact; and performing error-prone area localization based on the scope of error impact to generate error-prone area markers.

[0058] The frequency of doctor modifications to draft medical records was statistically analyzed to generate a modification frequency distribution. The statistical process tracked doctor modifications to various fields during the review process, categorizing modifications as content replacement, addition, and deletion. Content replacement indicated a significant deviation between the original content and clinical facts, with a weight of 1.0; content deletion indicated redundant or erroneous content, with a weight of 0.8; and content addition indicated incomplete content, with a weight of 0.5. The weighted modification counts for each field in the draft medical records were accumulated across multiple rounds of historical case data. The modification frequency distribution was constructed by normalizing the accumulated weighted modification counts for each field, ensuring comparability across different case scales. One field was replaced 62 times, added 28 times, and deleted 10 times in 100 historical draft medical records. Its weighted cumulative modification frequency was significantly higher than other fields, and the peak of the modification frequency distribution was located at this field, indicating that this field is a high-error-prone area. The pattern of the modification frequency distribution reflects the distribution pattern of generation quality in the overall field structure of the case draft document. In respiratory disease cases, the "diagnostic impression" and "treatment plan" fields consistently dominate the modification frequency, exhibiting a distinct peak characteristic, indicating that generation quality issues are concentrated in these two fields. When the modification frequencies of all fields tend to be consistent, the pattern is relatively flat, suggesting that generation quality issues are distributed across all fields. When new fields are added to the case draft document, the modification frequency distribution synchronously expands the corresponding dimension. New fields with less than three rounds of historical modification data accumulation are marked as having insufficient data in the modification frequency distribution and are not included in the weight allocation for atlas construction.

[0059] A cascading modification correlation graph is constructed based on the modification frequency distribution to generate the degree of modification impact. The modification correlation between fields originates from the co-occurrence records of two fields being affected by the same round of modification operations in historical case data. In the modification frequency distribution, the co-occurrence modification probability between high-frequency modified fields is usually higher than that between low-frequency fields. After the "diagnostic impression" field is modified, the probability that the "treatment plan" field will be adjusted in the same round of modification exceeds 70%, forming a strong correlation edge between the two. The cascading modification correlation graph uses fields as nodes. The initial weight of each node is determined by the normalized modification frequency of the corresponding field in the modification frequency distribution. The weight of the directed edge between nodes is determined by the co-occurrence modification probability between the two fields. The higher the weight, the greater the possibility that the other field will be modified in conjunction with the modification of one field. The impact of modifications is weighted by the in-degree and out-degree of each field in the graph. A high in-degree indicates that the field is frequently modified due to changes in upstream fields. The "treatment plan" field has consistently been adjusted frequently due to changes in "diagnostic impression" in historical cases, resulting in a high in-degree value among all fields. A high out-degree indicates that the downstream fields are widely affected after the field is modified. When the "diagnostic impression" field is changed from "unstable angina" to "acute myocardial infarction," multiple downstream fields such as "treatment plan" and "interpretation of auxiliary examinations" are usually adjusted simultaneously, resulting in a significantly high out-degree value. The synthesis formula is P_i = α_p × in_i + β_p × out_i, where in_i and out_i are the normalized in-degree and out-degree of the i-th field, respectively. α_p is set to 0.4 and β_p to 0.6 to highlight the diffusion effect of field modifications, and P_i ranges from 0 to 1. Fields with concentrated peak values ​​in the overall frequency distribution of modifications typically form high-density relational subgraphs in the graph. Within these subgraphs, the degree of influence of modifications on each field is generally high. Fields with a modification influence of less than 0.2 have sparse relationships between nodes, and the scope of influence of modification behavior is limited.

[0060] For example, identifying the error impact range based on the degree of modification impact includes: performing diffusion feature detection on the degree of modification impact to identify dense diffusion segments; setting a stable window span in the dense diffusion segments; performing attenuation processing on the degree of modification impact based on the stable window span to generate an attenuated impact degree; and determining the error impact range based on the attenuated impact degree.

[0061] The impact of modifications is analyzed using diffusion feature detection to identify dense diffusion segments. Diffusion feature detection extracts spatial diffusion patterns of the impact of modifications on each field from historical multi-round case modification data. These patterns reflect the rate and range of modification behavior propagating from high-impact fields to adjacent fields. Field groups with fast propagation rates and wide ranges constitute potential candidates for dense diffusion segments. The propagation strength of the impact of modifications between adjacent fields is determined by both the weight of the inter-field association edges and the numerical value of the impact of the modifications. Fields with a propagation strength exceeding a threshold of 0.5 that appear consecutively are identified as diffusion channels. Intervals within these channels where the impact of modifications remains consistently high are defined as dense diffusion segments. The location of a sudden drop in the impact of modifications corresponds to the natural termination boundary of a dense diffusion segment. The criterion for a sudden drop is that the difference in the impact of modifications between adjacent fields exceeds 0.25. When this condition is met, the propagation interrupts at that location, the dense diffusion segment terminates, and the search for the next starting point begins. The fields "Diagnostic Impression," "Interpretation of Auxiliary Examinations," and "Treatment Plan" have consistently maintained a high level of influence from modifications in historical data, and their cross-influence exceeds the threshold. These three fields collectively constitute a densely affected area, where a modification to any one field within this area has a greater than 65% probability of consequently modifying the other two. The number and size of these densely affected areas tend to stabilize with increasing historical case accumulation. When the accumulated number is less than 50 cases, the boundaries of these densely affected areas exhibit significant uncertainty, and low-confidence labels are added to these areas.

[0062] A stable window span is set for densely diffused regions. The stable window span is determined by both the diffusion intensity and the number of fields within the densely diffused region. Densely diffused regions with higher diffusion intensity and more fields have longer stable window spans. The trajectory of changes in the influence of modifications to each field within the stable window span is fully recorded. When the trajectory tends to stabilize, it indicates that the diffusion effect has fully unfolded within the window. If the trend has not converged, the stable window span is adaptively extended, with an extension step of 20% of the initial stable window span, and a maximum extension of 3 times. When the densely diffused region is small, the initial stable window span is set to 3 field units; when the region has more than 6 fields, the initial stable window span is correspondingly expanded to 5 field units, indicating that the diffusion effect typically completes its propagation within this window. The diffusion-intensive segment, comprised of "Diagnostic Impression," "Interpretation of Auxiliary Examinations," and "Treatment Plan," consists of three fields. The initial stable window span is set to three field units. Historical data shows that the diffusion effect in this segment typically converges within 2.5 field units. The initial stable window span provides sufficient coverage, eliminating the need for adaptive extension. The stable window spans for different diffusion-intensive segments are independent of each other. Each segment determines its window length based on its own diffusion characteristics, avoiding over-cutting or insufficient coverage of some segments due to a uniform window.

[0063] The impact of modifications is attenuated based on the stability window span to generate the attenuated impact. The attenuation function adopts an exponential decay form, with the attenuation formula P_decay(d)=P_origin×e^(-λd), where e is a natural constant approximately equal to 2.718, P_origin is the original impact of the field modification, ranging from 0 to 1, d is the distance of the field from the center of the dense diffusion segment, in the number of field units, and λ is determined by the reciprocal of the stability window span T, i.e., λ=1 / T, where T is in the number of field units. λ is greater than 0; the longer the stability window span, the smaller λ becomes, the smoother the attenuation, and more of the attenuated impact of distant fields is preserved. In a certain consultation scenario, "Diagnostic Impression," "Interpretation of Auxiliary Examinations," and "Treatment Plan" constitute a densely affected area. The "Physical Examination" field is located at the edge of this area. Its original modified impact was 0.72. With a stable window span of 3 field units, λ=1 / 3. After attenuation, the impact at a distance of 2 field units from the center is approximately 0.37, still higher than the cutoff threshold. This indicates a substantial connection between the "Physical Examination" field and this densely affected area, and it should be included in the error impact range. Fields with high original modified impact values ​​may still have a higher impact after attenuation than fields with low original values. Attenuation corrects distance deviations rather than erasing essential differences between fields. The λ values ​​of different densely affected areas are independent. Areas with high diffusion intensity typically correspond to shorter stable window spans, with larger λ values ​​and steeper attenuation. Only the core fields maintain a high impact after attenuation, while areas with low diffusion intensity experience gentler attenuation and a relatively wider impact range. Fields outside the stable window span have an impact of zero after attenuation.

[0064] The scope of error impact is determined based on the degree of attenuation. Fields with an attenuation impact exceeding a cutoff threshold of 0.3 are included in the error impact scope, while fields with an attenuation impact below the threshold are excluded even if they are associated with densely diffused segments. The cutoff threshold is set to balance recall and precision; a higher threshold narrows the error impact scope, increasing the risk of missing potential error fields, while a lower threshold expands the error impact scope, excessively diluting validation resources. Fields with an attenuation impact near the cutoff threshold are most sensitive to the determination of error impact scope. Fields with an absolute difference between their attenuation impact and the cutoff threshold of less than 0.05 are marked as boundary fields, and boundary fields are marked with a low-determinism label within the error impact scope. When the attenuation impact of multiple densely diffused segments covers the same field, the maximum value of the attenuation impact across all segments is used as the final determination criterion. This maximum value method ensures that the most unfavorable damage to the field in any segment's diffusion chain is taken into consideration. Within the scope of the error's impact, each field is ranked from highest to lowest according to its attenuated impact. After the scope of an error's impact was determined, the "Diagnostic Impression" field ranked first with an attenuated impact of 0.87, while the "Physical Examination" field ranked last with an attenuated impact of 0.31. Both of these fields exceeded the cutoff threshold and were included in the scope of the error's impact.

[0065] Error-prone area markers are generated based on the scope of error impact. Error-prone area identification starts from the historical modification patterns of each field within the error impact scope, extracting the most common error types for each field. These error types include terminology deviation, logical relationship errors, and numerical unit errors. The dominant error types differ across fields; for example, the dominant error type for the "Auxiliary Inspection" field is numerical unit errors, while for the "Diagnostic Impression" field, it's terminology deviation. The error-prone area markers for these two types of fields carry different error type labels to guide differentiated verification. Error-prone area markers consist of three elements: field identifier, dominant error type, and error-proneness score. The error-proneness score is determined by the modification impact degree of the corresponding field within the error impact scope and its historical modification frequency. Fields with high modification impact degree and high historical modification frequency have an error-proneness score close to 1.0, while fields with both low scores have a score close to 0. Fields within the error impact scope with a modification impact degree exceeding 0.6 and ranking in the top 20% of historical modification frequency automatically receive a high-risk label in the error-prone area markers. The remaining fields within the error impact scope are assigned medium-risk or low-risk error-prone area markers in descending order of their error-proneness scores. After a case was generated, the "Treatment Plan" field scored 0.91 and was marked as a high-risk error-prone area, while the "Family History" field scored only 0.18 and was marked as a low-risk error-prone area. This difference in labeling directly drives the differentiated allocation of validation resources. Fields outside the scope of error impact do not receive error-prone area labels.

[0066] Standardized case documents are generated by prioritizing the verification of error-prone areas using a content allocation coefficient. This prioritization multiplies the field weights of the content allocation coefficients by the error-prone area marker error severity scores. Fields with higher products receive more verification resources, ensuring resources are concentrated on high-value, high-risk fields. Fields with both high content allocation coefficients and high error-prone area marker error severity scores are considered "double high-risk" fields and require comprehensive verification across three areas: terminology standardization, logical consistency, and numerical accuracy. In acute myocardial infarction cases, the "diagnostic impression" and "treatment plan" fields are often both "double high-risk" fields; both must undergo all three verifications sequentially before standard case document generation. Failure to pass any one triggers regeneration of the field content and re-execution of the comprehensive verification, repeating this process until all three verifications pass. Fields with high content allocation coefficients but low error severity scores only undergo terminology standardization verification, resulting in significantly lower resource consumption compared to "double high-risk" fields. After all high-risk fields in the error-prone area markers pass verification, low-risk fields are verified sequentially based on the residual weights of the content allocation coefficients. The standardized case document is officially output after all fields pass verification. Before outputting standardized case documents, a final integrity check is performed. The check results are written into the quality report attachment. The quality report includes the check round, check pass time and final check conclusion for each field. If there are unresolved low-certainty fields, a manual review suggestion is attached.

[0067] To implement the above-described method embodiments, a cloud-based large-scale model-based method for automatically generating medical voice medical records is provided to achieve the corresponding functionalities and technical effects. See also... Figure 2 , Figure 2 This paper illustrates a structural block diagram of a cloud-based large-scale model-based automatic medical voice case generation system 200 provided in an embodiment of this application, including: The voice acquisition module 201 is used to acquire doctor's voice data from a medical microphone and to perform semantic analysis on the doctor's voice data to identify the density of terms. The terminology calibration module 202 is used to perform context matching evaluation on the terminology density to generate a matching degree matrix, identify low matching degree segments from the matching degree matrix, perform ambiguity resolution processing on the low matching degree segments to generate term disambiguation tags, and construct a terminology calibration mapping based on the term disambiguation tags. The structural analysis module 203 is used to acquire case structural response data from a large cloud model and perform structural scanning to identify the degree of missing case structures. Based on the degree of missing case structures, priority assessment is performed to generate a completion priority factor, and a content allocation coefficient is generated based on the cumulative change rate of the completion priority factor. The diagnostic processing module 204 is used to perform case knowledge parsing and identify diagnostic conflicts through the terminology calibration mapping, perform identification and sorting based on the diagnostic conflicts to generate diagnostic priority tags, perform sorting consistency detection on the content allocation coefficient and the diagnostic priority tags to identify sorting deviation segments, and establish case generation key nodes based on the sorting deviation segments. The case generation module 205 is used to output a case draft document for the key nodes of case generation, perform doctor modification frequency statistics on the case draft document to generate error-prone area markers, and use the content allocation coefficient to perform key verification and allocation on the error-prone area markers to generate a standardized case document.

[0068] The aforementioned cloud-based large-scale model-based automatic medical voice medical record generation system 200 can implement the cloud-based large-scale model-based automatic medical voice medical record generation method described in the above method embodiments. The options in the above method embodiments are also applicable to this embodiment and will not be detailed here. The remaining content of this application embodiment can be referred to the content of the above method embodiments, and will not be repeated in this embodiment.

[0069] The above embodiments are not an exhaustive list based on the present invention, and there may be many other embodiments not listed. Any substitutions and improvements made without departing from the concept of the present invention are within the protection scope of the present invention.

Claims

1. A method for automatically generating medical voice medical records based on a cloud-based large model, characterized in that, include: Collect doctors' voice data from medical microphones, and perform semantic analysis on the doctors' voice data to identify term density; The term density is evaluated by context matching to generate a matching degree matrix. Low matching degree segments are identified from the matching degree matrix. Ambiguity disambiguation processing is performed on the low matching degree segments to generate term disambiguation tags. A term calibration mapping is constructed based on the term disambiguation tags. Obtain case structure response data from a large cloud model and perform structural scanning to identify the degree of case structure loss. Based on the degree of case structure loss, perform priority assessment to generate a completion priority factor. Based on the cumulative change rate of the completion priority factor, generate a content allocation coefficient. The terminology calibration mapping is used to analyze case knowledge and identify diagnostic conflicts. Based on the diagnostic conflicts, the case knowledge is analyzed and sorted to generate diagnostic priority tags. The content allocation coefficient and the diagnostic priority tags are sorted in a consistent manner to identify sorting deviation segments. Based on the sorting deviation segments, key nodes for case generation are established. For the key nodes of case generation, output a case draft document, perform doctor modification frequency statistics on the case draft document to generate error-prone area markers, and use the content allocation coefficient to perform key verification and allocation on the error-prone area markers to generate standardized case documents.

2. The method according to claim 1, characterized in that, The step of performing contextual matching evaluation on the term density to generate a matching degree matrix includes: The term density is analyzed by constructing multi-granularity context windows based on word-level context, sentence-level context, and paragraph-level context, and then performing hierarchical matching detection to generate a multi-granularity inconsistency location set. Cross-granularity cross-validation is performed on the multi-granularity inconsistency location set to filter co-occurring inconsistency locations and generate a misuse intensity distribution; The high-frequency misuse contexts in the misuse intensity distribution are extracted as high-risk context combinations; The high-risk context combinations are labeled with matching degree to generate a matching degree matrix.

3. The method according to claim 1, characterized in that, The process of generating term disambiguation tags based on the low-matching segments includes: The distribution of candidate terms is obtained by identifying ambiguity sources in the low-matching segments. The candidate term distribution is subjected to conflict accumulation analysis to generate conflict accumulation features; Based on the aforementioned conflict accumulation characteristics, divide the region into a dense region of ambiguity and a sparse region of ambiguity; Disambiguation priorities are assigned to the ambiguity-dense regions and the ambiguity-sparse regions to generate term disambiguation tags.

4. The method according to claim 1, characterized in that, The process of generating completion priority factors based on the degree of structural missingness in the case includes: Cascade propagation analysis was performed on the structural missingness of the cases to identify cascade triggering features; Structural stability assessment is performed based on the cascaded triggering features to generate the severity of structural defects; The severity of the structural missing features is analyzed using feature importance analysis to extract high-priority and low-priority features. Priority sorting is established based on the high-priority elements and the low-priority elements to generate a completion priority factor.

5. The method according to claim 1, characterized in that, The generation of content allocation coefficients based on the cumulative change rate of the completion priority factor includes: Extract the cumulative rate of change fluctuation information of the completion priority factor; The cumulative rate of change fluctuation information is analyzed to identify rate inflection point characteristics; At the rate inflection point feature, establish content weight adjustment rules and generate a weight adjustment table; Construct content allocation coefficients based on the weight adjustment table.

6. The method according to claim 1, characterized in that, The step of generating diagnostic priority markers based on the diagnostic conflicts includes: The identification path is evaluated to generate an identification difficulty value for the diagnostic conflict; The optimal identification path is generated by selecting the path pair with the lowest difficulty based on the identification difficulty value. Identify diagnostic decision nodes on the optimal identification path; The diagnostic decision node generates a diagnostic priority label by performing probability sorting.

7. The method according to claim 1, characterized in that, The step of generating error-prone area markers by statistically analyzing the frequency of doctor modifications to the draft case document includes: The doctor's modification frequency distribution was generated by statistically analyzing the doctor's modification frequency in the aforementioned case draft documents. Based on the modification frequency distribution, a cascading modification correlation graph between fields is constructed to generate the degree of modification impact. Identify the scope of the error's impact based on the degree of impact of the modifications; Based on the scope of the error's impact, error-prone location is performed to generate error-prone area markers.

8. The method according to claim 4, characterized in that, The cascading propagation analysis of the structural missingness of the cases to identify cascading trigger features includes: Based on the missing data of the case structure, field dependency analysis is performed to identify primary related fields; A multi-level association field chain is generated by expanding the first-level association field layer by layer along the dependency direction; The multi-level related field chain is filtered by influence cutoff according to hierarchical depth to generate a set of effective influential fields; Multiple related field chains are extracted from the effective influence field set to form cascading trigger features.

9. The method according to claim 7, characterized in that, The step of identifying the scope of error impact based on the degree of influence of the modification includes: The degree of influence of the modification is assessed by diffusion feature detection to identify densely diffused areas; A stable window span is set in the densely diffused section; The degree of influence of the modification is attenuated based on the stable window span to generate the attenuated degree of influence. The scope of the error's impact is determined based on the degree of impact after attenuation.

10. A medical voice medical record automatic generation system based on a cloud-based large model, characterized in that: include: The voice acquisition module is used to acquire doctors' voice data from medical microphones and to perform semantic analysis on the doctors' voice data to identify the density of terms. The terminology calibration module is used to perform context matching evaluation on the terminology density to generate a matching degree matrix, identify low matching degree segments from the matching degree matrix, perform ambiguity resolution processing on the low matching degree segments to generate term disambiguation tags, and construct a terminology calibration mapping based on the term disambiguation tags. The structural analysis module is used to acquire case structural response data from a large cloud model and perform structural scanning to identify the degree of missing case structures. Based on the degree of missing case structures, priority assessment is performed to generate a completion priority factor, and a content allocation coefficient is generated based on the cumulative change rate of the completion priority factor. The diagnostic processing module is used to analyze case knowledge and identify diagnostic conflicts through the terminology calibration mapping, generate diagnostic priority tags based on the diagnostic conflicts, perform sorting consistency detection on the content allocation coefficient and the diagnostic priority tags to identify sorting deviation segments, and establish key nodes for case generation based on the sorting deviation segments. The case generation module is used to output a case draft document for key nodes of case generation, perform doctor modification frequency statistics on the case draft document to generate error-prone area markers, and use the content allocation coefficient to perform key verification and allocation of the error-prone area markers to generate standardized case documents.

Images