Surgical clinical aid decision-making method and system based on machine learning

Abstract

The invention provides a surgical clinical auxiliary decision-making method and system based on machine learning, and relates to the technical field of intelligent medical treatment, and the method comprises the steps: firstly, collecting a multi-source clinical information set generated by a patient in the whole process of surgical diagnosis and treatment, constructing a surgical decision-making scene association link based on the multi-source clinical information set, and determining a corresponding relationship between each operation decision demand node and the related information. Then, decision elements are extracted from the association link and integrated according to the surgical decision logic, a surgical decision adaptation factor set is generated, and each factor corresponds to a specific surgical decision node demand. Then, by analyzing multi-source information relevance, dynamically adjusting an adaptation factor set, ensuring accurate matching with operation decision node requirements, finally, generating a surgical operation decision scheme and scheme adaptation description for a specific patient, and pushing the scheme and the scheme adaptation description to a surgical clinical decision terminal, so that precision, individuation and dynamics of surgical clinical decisions are realized, and the accuracy, the individuation and the dynamics of the surgical clinical decisions are improved. And the quality and efficiency of surgical diagnosis and treatment are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of smart healthcare technology, and more specifically, to a surgical clinical decision support method and system based on machine learning. Background Technology

[0002] In surgical clinical diagnosis and treatment, accurate and scientific surgical decisions play a decisive role in the patient's treatment outcome and recovery process. Traditional surgical clinical decision-making methods mainly rely on the physician's personal experience, professional knowledge, and limited clinical examination results. When making surgical decisions, physicians need to comprehensively analyze various aspects of the patient's information; however, this information is often scattered across different records and reports, lacking systematic integration and correlation.

[0003] For example, a patient's basic health records, disease progression records, surgical imaging descriptions, and treatment intervention records typically exist in independent forms. Doctors need to spend a significant amount of time and effort to sort through and connect this information to form a comprehensive basis for decision-making. Moreover, the correlations between different types of information are difficult to fully explore and utilize, potentially leading to the overlooking of some important underlying factors during the decision-making process. Furthermore, traditional decision-making methods lack dynamic adjustment mechanisms, making it impossible to optimize decision-making plans in a timely manner based on real-time changes in the patient's condition, thus failing to meet the demands of modern surgical clinics for precise and personalized diagnosis and treatment. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a surgical clinical decision support method based on machine learning, the method comprising: The system collects a set of multi-source clinical information generated by the patient throughout the entire surgical diagnosis and treatment process. This set of multi-source clinical information includes the patient's basic health record, disease progression record, surgical-related image description, and diagnosis and treatment intervention record. The patient's basic health record includes descriptions of physical condition, past medical history, and genetic background. The disease progression record includes descriptions of symptom changes, changes in physical signs, and changes in examination results. The surgical-related image description includes descriptions of lesion location, tissue anatomical relationships, and physiological structure morphology. The diagnosis and treatment intervention record includes descriptions of medication use, nursing measures, and the effects of previous treatments. Based on the multi-source clinical information set, a surgical decision-making scenario association link is constructed. The surgical decision-making scenario association link takes the specific surgical decision-making needs as nodes, and associates and binds the information elements related to each node in the multi-source clinical information set to generate a corresponding relationship chain between nodes and information elements. The decision elements corresponding to each node are extracted from the surgical decision scenario association link. The decision elements corresponding to each node extracted from the surgical decision scenario association link are integrated according to the surgical decision logic to generate a surgical decision adaptation factor set. Each surgical decision adaptation factor in the surgical decision adaptation factor set corresponds to the needs of a specific surgical decision node. By analyzing the correlation between different types of information in the multi-source clinical information set, the surgical decision adaptation factor set is dynamically adjusted so that each surgical decision adaptation factor matches the needs of the corresponding surgical decision node. Based on the adjusted set of surgical decision adaptation factors, a surgical decision plan and plan adaptation instructions are generated for specific patients, and the surgical decision plan and plan adaptation instructions are pushed to the surgical clinical decision terminal.

[0005] Furthermore, embodiments of the present invention also provide a surgical clinical decision support system based on machine learning, characterized in that it includes: A processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to perform the above-described machine learning-based surgical clinical decision support method by executing the machine-executable instructions.

[0006] In another aspect, embodiments of the present invention also provide a computer program product, the computer program product including machine-executable instructions stored in a computer-readable storage medium, a processor of a machine learning-based surgical clinical auxiliary decision-making system reading the machine-executable instructions from the computer-readable storage medium, the processor executing the machine-executable instructions, causing the machine learning-based surgical clinical auxiliary decision-making system to perform the aforementioned machine learning-based surgical clinical auxiliary decision-making method.

[0007] Based on the above, by collecting multi-source clinical information sets of patients throughout the entire surgical diagnosis and treatment process, covering aspects such as the patient's basic health, disease progression, surgical imaging, and treatment interventions, a surgical decision-making scenario association link is constructed based on multi-source clinical information. Relevant information elements are precisely associated and bound with specific surgical decision-making needs, generating corresponding relationship chains. Decision elements are extracted from the association links and integrated to generate a set of surgical decision-making adaptation factors. Each adaptation factor corresponds to a specific surgical decision-making node requirement, improving the pertinence of decision elements. Through the analysis of the correlation of multi-source information, the set of adaptation factors is dynamically adjusted to ensure that each adaptation factor always matches the surgical decision-making node requirement, enhancing the adaptability and accuracy of the decision plan. Finally, a surgical decision plan and plan adaptation instructions for specific patients are generated and pushed to the decision-making terminal, realizing the precision, personalization, and dynamism of surgical clinical decision-making, effectively improving the quality and efficiency of surgical diagnosis and treatment. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of the execution flow of the machine learning-based surgical clinical decision support method provided in an embodiment of the present invention.

[0009] Figure 2 This is a schematic diagram of exemplary hardware and software components of a machine learning-based surgical clinical decision support system provided in an embodiment of the present invention. Detailed Implementation

[0010] Figure 1 This is a flowchart illustrating a machine learning-based surgical clinical decision support method according to an embodiment of the present invention, which will be described in detail below.

[0011] Step S110: Collect a set of multi-source clinical information generated by the patient throughout the entire surgical diagnosis and treatment process. The set of multi-source clinical information includes the patient's basic health record, disease progression record, surgical-related image description, and diagnosis and treatment intervention record. The patient's basic health record includes a description of physical condition, past medical history, and genetic background. The disease progression record includes a description of symptom changes, an evolution of signs, and an evolution of examination results. The surgical-related image description includes a description of lesion location, tissue anatomical relationship, and physiological structure morphology. The diagnosis and treatment intervention record includes a description of medication use, nursing measures, and the effects of previous treatments.

[0012] This embodiment uses a patient with chronic cholecystitis scheduled for laparoscopic cholecystectomy as an example. When collecting multi-source clinical information, the patient's electronic medical record data is retrieved through a hospital information system integration platform. A data exchange protocol based on the HL7 FHIR standard is used to achieve multi-system data integration. The patient's basic health record is obtained through a structured data acquisition module. The description of physical condition includes body shape characteristics (such as height H, weight W, and the calculation of body mass index BMI = W / H). 2 Nutritional status assessment (using the Subjective Comprehensive Assessment Scale (SGA) classification), basic cardiopulmonary function indicators (such as left ventricular ejection fraction (LVEF) and forced expiratory volume in one second (FEV1), and American Society of Anesthesiologists (ASA) classification; description of past medical history integrated medical records within the past ten years through a timeline sorting module, including history of extracorporeal shock wave lithotripsy for gallstones (number of treatments N1, lithotripsy energy parameter E1), history of hypertension (diagnosis time T1, medication adherence score C1); description of genetic background was obtained through a family history questionnaire, including the spectrum of diseases in first-degree relatives (such as age of onset of gallbladder polyps in mother A1, duration of diabetes in father Y1).

[0013] The disease progression record was processed using a time series analysis engine to analyze outpatient follow-up data from the past three years. Symptom changes were described using the Numerical Rating Scale for Pain (NRS) to record the frequency of right upper quadrant pain attacks (increasing from N2 times per year to N3 times per month) and duration (extending from t1 minutes to t2 minutes). The evolution of physical signs was recorded using the dynamic monitoring module to record Murphy's sign (changing from negative to weakly positive) and the area of ​​tenderness in the right upper quadrant (expanding from a diameter of d1 cm to d2 cm). The evolution of examination results was described using the laboratory indicator trend analysis module to generate gallbladder wall thickness (gradually increasing from the baseline value B0 to the current value B1), gallbladder contraction function (decreasing from the normal range F0 to moderately impaired F1), and liver and kidney function indicators (ALT, Cr, etc. fluctuating within the reference range ±K%).

[0014] Surgical-related image descriptions were processed using a medical imaging structured reporting system. Abdominal ultrasound, CT, and MRI pancreaticobiliary imaging data were parsed in DICOM format, and the location description of the lesion was extracted (the gallbladder is located in the right upper quadrant below the costal margin at coordinates X1, Y1, and the distance from the pylorus to the location of the impacted gallbladder stone in the neck is D1). The description of tissue anatomical relationships was generated using three-dimensional reconstruction technology to obtain the adhesion area between the gallbladder and the surrounding omentum S1, the length of the cystic duct L1 and its diameter D2, the diameter of the common bile duct D3 and its course angle θ. The description of physiological structure and morphology included the clarity score of the three layers of the gallbladder wall (from clear C0 to blurry C1), the mucosal surface roughness index R1, the number of strong echogenic masses in the gallbladder lumen N4, and the maximum cross-sectional area S2.

[0015] The diagnosis and treatment intervention records integrate nearly six months of data through the treatment plan traceability module. The medication description includes ursodeoxycholic acid capsules (dose D4mg / time, frequency N5 times daily, treatment duration T2 days) and nonsteroidal anti-inflammatory drugs (type M1, number of days of use T3). The nursing intervention description is extracted from the nursing plan execution record, including the dietary control plan (low-fat diet compliance score C2, regular meal execution rate R2) and pain management methods (heating temperature T4℃, body position adjustment frequency N6 times / day). The description of the initial treatment effect uses the treatment response assessment scale, which shows that after drug treatment, the duration of pain relief shortened from t3 minutes to t4 minutes, the interval between attacks lengthened from d3 days to d4 days, and the ultrasound examination showed the gallbladder wall thickness change rate ΔB=(B1-B0) / T2.

[0016] Step S120: Construct a surgical decision-making scenario association link based on the multi-source clinical information set. The surgical decision-making scenario association link takes specific surgical decision-making needs as nodes, associates and binds information elements related to each node in the multi-source clinical information set, and generates a corresponding relationship chain between nodes and information elements.

[0017] When constructing the correlation links in the surgical decision-making scenario, the decision-making process of laparoscopic cholecystectomy is first deconstructed using a clinical decision node segmentation algorithm to identify four core decision nodes: surgical indication judgment node, surgical path planning node, surgical risk control node, and postoperative recovery plan node. Each node defines its decision direction through decision requirement metadata. For example, the surgical indication judgment node includes three decision dimensions: symptom severity assessment (weight W1), gallbladder function status assessment (weight W2), and drug treatment effect assessment (weight W3), satisfying W1+W2+W3=1. A BERT-based semantic association algorithm is used to match elements in the multi-source clinical information set with each node. The algorithm input is a word vector matrix of clinical information elements (dimension M×V, where M is the number of elements and V is the word vector dimension), and the output is a correlation score matrix (dimension 4×M). A correlation threshold Th1 is set, and elements with scores higher than Th1 are bound to their corresponding nodes. For example, the correlation between gallbladder wall thickness (B1) and the surgical indication judgment node is S1=0.85 (>Th1=0.7), and the correlation between patient body size characteristics (BMI value) and the surgical pathway planning node is S2=0.92 (>Th1=0.7). The cosine similarity method is used to quantify the correlation strength between information elements and node requirements. The specific calculation process is as follows: the information element text is converted into a TF-IDF vector, and an inner product operation is performed with the node requirement description vector. Then, the result is divided by the product of the magnitudes of the two vectors to obtain the correlation score. After generating the preliminary node-element correspondence, it is validated using a clinical expert rule base to correct abnormal correlations (such as adjusting the ASA classification from the surgical pathway planning node to the surgical risk control node).

[0018] Step S121: The multi-source clinical information set is divided into a basic information subset, a disease course information subset, an image description subset, and an intervention record subset according to information type. Each subset contains all clinical information description content of the corresponding type.

[0019] The collected multi-source clinical information sets were structurally split using an information classifier. This classifier employed a multi-label classification model based on random forests. The input was a feature vector of the clinical information text (including part-of-speech features, medical entity features, and semantic role features), and the output was a membership vector (dimension 1×4) of four subsets. When the membership value of a certain dimension was greater than a threshold Th2, the information was assigned to the corresponding subset. The basic information subset was stored in a relational database table structure, including a physical condition table (fields: height H, weight W, BMI, SGA classification, LVEF, FEV1, ASA classification), a past medical history table (fields: disease code D1, diagnosis time T1, treatment method M1, treatment effect score E1), and a genetic background table (fields: kinship R1, disease type D2, age of onset A1, disease course Y1). The disease course information subset was stored in a time-series database, categorized by symptom changes (sampling frequency F1 times / day), physical sign evolution (sampling frequency F2 times / day), and examination result evolution (sampling frequency F3 times / week) to establish a time series. The image description subset is stored in a NoSQL database, indexed by examination type (ultrasound, CT, MRI) and examination timestamp T. Each record contains the DICOM file path P and structured report JSON data. The intervention record subset is stored in a transactional database, with associated tables created based on medication (fields: drug code M2, dosage D4, frequency N5, route of administration R2, start time T5, end time T6), nursing interventions (fields: intervention code C3, execution frequency N6, duration t5, implementer ID), and preliminary treatment effects (fields: assessment scale Z1, score V1, assessment time T7). Each subset contains metadata fields: information description content, collection timestamp, data source system ID, and reliability score (range 0-1, calculated based on data acquisition equipment accuracy and operator qualifications).

[0020] Step S122: Extract scenario description elements related to surgical decision-making from each information subset. The scenario description elements are specific descriptions of the conditions related to surgical decision-making. Extract basic condition description elements that affect surgical tolerance from the basic information subset. Extract disease course description elements that reflect the trend of disease development from the disease course information subset. Extract anatomical structure description elements that affect the choice of surgical path from the image description subset. Extract early intervention description elements that affect the adjustment of surgical plan from the intervention record subset.

[0021] For the basic information subset, a medical entity recognition model based on BiLSTM-CRF is used to extract scene description elements. This medical entity recognition model includes an embedding layer (converting input text into word vectors), a bidirectional LSTM layer (capturing contextual features), and a CRF layer (optimizing entity label sequences). The output is a text sequence labeled with entity types. Factors influencing surgical tolerance were extracted from the description of physical condition, including age stratification (A∈[18, 44] young adults, [45, 59] middle-aged adults, [60, +∞] elderly adults), body mass index (BMI<18.5 underweight, 18.5≤BMI<24 normal, 24≤BMI<28 overweight, BMI≥28 obese), cardiopulmonary function indicators (LVEF≥50% normal, <50% decreased; FEV1 percentage of predicted value≥80% normal, 60%-79% mildly decreased, 40%-59% moderately decreased, <40% severely decreased), liver and kidney function status (ALT within the reference range [L1, U1] normal, >U1 abnormal; Cr within the reference range [L2, U2] normal, >U2 abnormal), and comorbidities (hypertension grade H1-H4, diabetes type T1-T2, etc.).

[0022] Within the disease course information subset, the time series trend analysis module extracts elements related to disease progression. The sliding window method is used to calculate the rate of change in symptom frequency (ΔF=(N3-N2) / T, where T is the observation time span), the time of onset of positive physical signs (e.g., the first positive Murphy's sign at time T8), and the dynamic change curve of examination indicators (calculated using linear regression to determine the slope of gallbladder wall thickness change k=(B1-B0) / T2). For example, when ΔF > 0.1 times / month and k > 0.1 mm / month, the disease progression trend is considered significant.

[0023] In the image descriptive subset, anatomical descriptive elements are extracted using an image structured report template. The template includes the gallbladder's location (coordinates X1, Y1), size (long axis L2 × short axis D5), shape (regular / irregular), cystic duct length L1 and diameter D2 (compared to the reference value D0; D2 > D0 × 1.5 indicates dilation), common bile duct diameter D3 (normal reference value Dn; D3 > Dn × 1.2 indicates dilation), surrounding tissue relationships (adhesion area between the gallbladder and surrounding omentum S1, adhesion grade G1-G4), and vascular distribution (cystic artery course type T3-T5, distance from the gallbladder wall D6).

[0024] In the intervention record subset, the previous intervention description elements are extracted through the drug-effect association analysis module. The treatment response index of ursodeoxycholic acid is calculated using the dose-effect model (RI = ΔB / D4 / T2, where ΔB is the change value of the gallbladder wall thickness), the effective duration of non-surgical treatment (t6 = the extended value of the attack interval d4 - d3), and the drug allergy history (allergy drug list M3, allergy reaction type R3). When RI < 0 and t6 < T0 (the minimum effective duration threshold), it is determined that the non-surgical treatment effect is不佳.

[0025] Step S1221: Read the clinical information description content in each information subset one by one, and split each clinical information description content into independent description statements according to the sentence structure.

[0026] The clinical information description content in the information subset is split through the sentence splitting algorithm in natural language processing. First, the text is preprocessed (removing special characters and standardizing spaces), and then sentence splitting is performed based on punctuation marks (periods, question marks, exclamation marks, semicolons) and semantic pauses (such as the logical pause of "," in long sentences). For example, the sentence "The patient has a medium build, good blood pressure control, a five-year history of hypertension, and takes calcium channel blockers" in the basic information subset is split into four independent description statements after processing: "The patient has a medium build", "Good blood pressure control", "Has a five-year history of hypertension", "Takes calcium channel blockers". Each description statement stores the following metadata: the original document ID, the start / end character positions in the original document, and the sentence splitting confidence (calculated based on the punctuation mark type and context semantics, range 0 - 1). For ambiguous sentence splitting (such as "Right upper abdominal pain accompanied by nausea, vomiting, and fever"), rule-based post-processing is used (preferably splitting according to conjunctions such as "accompanied by", "and"), ensuring that each statement expresses a single clinical fact.

[0027] Step S1222: Perform semantic analysis on each description statement to identify the core expression content in the description statement that reflects the patient's physical state, disease changes, anatomical structure, and treatment interventions.

[0028] A BERT-based semantic role labeling model is used to analyze each descriptive statement. This model adds a semantic role classification layer to the pre-trained BERT model. The input is a sequence of word vectors for the descriptive statement, and the output is a semantic role label for each word (including core roles A0-A5, and modifier roles AM-ADV / LOC / TMP, etc.). Taking "gallbladder wall thickness increased compared to before, accompanied by mucosal roughness" as an example, the model first segments the statement to obtain the word sequence: [gallbladder wall, thickness, compared to before, increased, accompanied by, mucosal surface, roughness]. Then, through contextual feature extraction, the core expression is identified as "gallbladder wall thickness increased" (A0=gallbladder wall, A1=thickness, V=increased) and "mucosal surface roughness" (A0=mucosal surface, A1=roughness). At the same time, a medical ontology knowledge base (such as UMLS) is used to standardize the identified core expression, mapping "thickness increase" to the standard term "thickening" and "roughness" to "mucosal abnormality". For expressions containing numerical values ​​(such as "gallbladder wall thickness 5mm"), the numerical features and their units are extracted simultaneously and stored as (attribute: gallbladder wall thickness, value: 5, unit: mm).

[0029] Step S1223: Compare the core statements with the direction of surgical decision-making, and select the core statements that are related to surgical implementation, risk control, and outcome prediction.

[0030] A surgical decision-making lexicon was constructed, comprising terms across three dimensions: surgical implementation (e.g., surgical indications, contraindications, operational conditions), risk control (e.g., bleeding risk, infection risk, organ damage risk), and outcome prediction (e.g., postoperative recovery time, complication rate, quality of life improvement). A cosine similarity algorithm was used to calculate the semantic similarity between core expressions and terms in each dimension. Expressions with a similarity value greater than the threshold Th3 were considered relevant. For example, the similarity S between "increased gallbladder wall thickness" and the surgical implementation term "surgical indication - gallbladder wall thickening" was 0.89 (>Th3=0.75), so this expression was retained. The expression "patient's occupation is teacher" had a similarity <0.3 with all three dimensions and was excluded. During the comparison process, expressions containing negative words (e.g., "no history of hypertension") were marked with a negation detection algorithm to indicate their semantic polarity, ensuring that the screening results accurately reflected clinical facts.

[0031] Step S1224: Classify the screened core statements according to their clinical significance, distinguishing between statements reflecting the current situation, statements reflecting changes, and statements reflecting relationships.

[0032] The core content is categorized based on temporal and relational features: Descriptions reflecting the current state do not contain time-varying words (e.g., "gallbladder diameter is D7" or "blood pressure is BP1") and are identified using a current state marker vocabulary (e.g., "is," "exists"). Descriptions reflecting change contain time-dimensional words (e.g., "increased compared to before," "gradually decreased," "suddenly increased") and are identified using a time adverb dictionary (e.g., "compared to before," "recently," "within three months") and a change verb dictionary (e.g., "increased," "decreased," "worsened"). Descriptions reflecting relationships contain causal relationship words (e.g., "led to," "caused," "due to") or accompanying relationship words (e.g., "accompanied," "simultaneously," "merged") and are identified using a relational vocabulary. For example, "gallbladder diameter is D7" is categorized as a current state description, "gallbladder wall thickness increased by ΔB compared to before" is categorized as a change description, and "gallbladder stones caused cystic duct obstruction" is categorized as a relational relationship description. The classification results are stored with classification confidence (calculated based on the number and position of feature words).

[0033] Step S1225: Extract specific features from the statements reflecting the current situation, including specific information on body parts, symptom characteristics, structural characteristics, and intervention measures; perform trend analysis on the statements reflecting changes, including information on the direction, magnitude, and time span of state changes; and perform association analysis on the statements reflecting relationships, including information on causal relationships, accompanying relationships, and restrictive relationships between different clinical elements.

[0034] For the description reflecting the current situation, "Multiple strong echogenic foci are visible in the gallbladder, with a maximum diameter of D8," an entity attribute extraction algorithm is used to extract the body part (gallbladder) and structural features (strong echogenic foci, maximum diameter D8). The structural features include the feature name (strong echogenic foci), quantity (multiple), quantified value (D8), and unit (mm). For the description reflecting the change, "The frequency of pain attacks increased from N2 times per month to N3 times per month," a trend analysis algorithm is used to extract the direction of change (increase), the magnitude of change (ΔN = N3 - N2 times / month), and the time span of change (T = current time - first recording time). The rate of change v = ΔN / T is calculated through linear fitting. For the description reflecting the correlation, "A long-term high-fat diet leads to gallstone formation," a relation extraction model is used to extract the causal relationship (cause: long-term high-fat diet; effect: gallstone formation). The model is based on bidirectional LSTM and an attention mechanism, identifying relational words such as "leads to" to determine the relation type, and identifying the causal subject through entity recognition.

[0035] For example, step S1225-1: Extract all clinical elements in the description content that reflects the relationship, where each clinical element is a specific physical state, symptom, structure, or intervention involved in the description content.

[0036] A clinical element extraction method combining rule-based and machine learning was adopted. First, preliminary matching was performed using a medical entity dictionary (including categories such as disease, symptom, sign, anatomical structure, drug, and procedure). Then, a CRF model was used to identify unmatched entities. Taking the statement "poorly controlled diabetes leading to delayed wound healing" as an example, the extracted clinical elements were: physical condition (poorly controlled diabetes) and symptom (delayed wound healing). During the element extraction process, fuzzy expressions (such as "high blood sugar") were standardized and mapped to "poorly controlled blood sugar," and composite elements (such as "hypertension combined with diabetes") were split (splitting into "hypertension" and "diabetes"). Each extracted clinical element includes element type (physical condition / symptom / structure / intervention), standardized terminology, original expression, and confidence score (calculated based on matching score and contextual features).

[0037] Step S1225-2: Classify the extracted clinical elements, distinguishing between main elements and related elements. Main elements are the clinical objects described in the core content, while related elements are other clinical objects that are related to the main elements.

[0038] Syntactic dependency analysis was used to determine the grammatical relationships between clinical elements. The Stanford CoreNLP toolkit was used to extract subject-verb-object structures, identifying the agent (subject) of the action as the principal element and the receiver (object) as the related element. In the expression "gallstones caused an attack of cholecystitis," syntactic analysis showed that "gallstones" was the subject of "caused," and "attack of cholecystitis" was the object; therefore, the principal element was "gallstones," and the related element was "attack of cholecystitis." For expressions without explicit action words (such as "abdominal pain accompanied by fever"), co-occurrence relationships and semantic weights were used to determine the principal elements (elements that typically appear frequently or at the beginning of the sentence). The classification results were stored with element role labels (principal / related) and syntactic relationship types (subject-verb / verb-object / attributive, etc.).

[0039] Step S1225-3: Analyze the positional relationship and grammatical connection relationship between the main element and each related element in the content of the expression, and identify the related words that connect the two. Related words include cause, accompany, influence, and limit.

[0040] A word association identification method based on lexical co-occurrence and dependency syntax is adopted. First, words in the word association table are searched within the text range (usually no more than 5 words) between the subject element and the associated element. Then, dependency syntax analysis is used to verify whether the word truly connects the subject and the associated element. For example, in "cystic duct obstruction leads to bile stasis," the word "leads to" exists between the subject element "cystic duct obstruction" and the associated element "bile stasis," and syntactic analysis shows that the subject of "leads to" is "cystic duct obstruction," and the object is "bile stasis." Therefore, the associated word is determined to be "leads to." For implicit associations (without explicit associated words), such as "abdominal pain and fever," default associated words (such as "accompanied") are supplemented using a medical knowledge graph.

[0041] Step S1225-4: Based on the semantic meaning of the related words, determine the type of association between the main element and the related elements. When the related words are "cause" or "initiate", it is determined to be a causal relationship; when the related words are "accompany" or "occur simultaneously", it is determined to be an accompanying relationship; when the related words are "limit", "constrain", or "influence", it is determined to be a restrictive relationship.

[0042] Construct a word association-type mapping table: Causal relationships include "lead to," "initiate," "cause," and "result"; accompanying relationships include "accompany," "simultaneously," "merge," and "concurrently"; and constraint relationships include "restrict," "constrain," "influence," and "determine." When an associated word is identified, the association type is directly determined based on the mapping table. For example, "lead to" corresponds to a causal relationship, "accompany" to an accompanying relationship, and "influence" to a constraint relationship. For expressions with multiple associated words (such as "long-term smoking leads to decreased lung function, affecting surgical tolerance"), the association type is determined separately for each clause (the first half is causal, the second half is constraint).

[0043] Step S1225-5: For causal relationships, set the main element as the cause and the related element as the effect, and analyze the details of the transmission path of causal action. The details of the transmission path are the ways in which the main element triggers the generation or change of the related element.

[0044] The causal transmission pathway is analyzed based on a medical pathophysiology knowledge base, which includes a disease mechanism atlas (e.g., "gallstone impaction → cholestasis → bacterial infection → inflammatory response"). Taking "gallstone impaction leading to cystic duct obstruction" as an example, the transmission pathway details are obtained by searching the knowledge base: "gallstone moves and becomes impacted in the cystic duct (mechanical obstruction) → bile outflow is blocked → increased pressure in the gallbladder → dilation of the cystic duct → cystic duct obstruction." During the pathway analysis, intermediate steps (e.g., "bile outflow is blocked") and mechanisms of action (e.g., "mechanical obstruction") are added to ensure the integrity of the transmission pathway. The transmission pathway details are stored as a directed graph structure, where nodes represent pathophysiological events and edges represent the causal relationships between events.

[0045] Step S1225-6: For the concomitant relationship, analyze the details of the temporal sequence relationship between the main element and the associated element. The details of the temporal sequence relationship include simultaneous occurrence, sequential occurrence, alternating occurrence, and the corresponding relationship of occurrence frequencies.

[0046] Establish a time series of element occurrences through the timestamp information in the病程记录 (medical record). Calculate the time difference between the main element and the associated element Δt = |T_main - T_associated|. When Δt < T_threshold (usually 24 hours), it is determined as simultaneous occurrence; when T_main < T_associated and Δt < T_threshold, it is determined that the main element appears first; when T_main > T_associated and Δt < T_threshold, it is determined that the associated element appears first; when the occurrence times of the two alternate and the periods are similar, it is determined as alternating occurrence. The corresponding relationship of occurrence frequencies is calculated by the ratio of the number of occurrences of the two within a unit time R = N_main / N_associated. When R ≈ 1, it is determined that the frequencies are consistent. For example, for the temporal sequence details of "right upper abdominal pain accompanied by fever": the pain occurrence time is T1, the fever occurrence time is T2, Δt = T2 - T1 = 2 hours (sequential occurrence), the number of pain occurrences within 24 hours is N1 = 3, the number of fever occurrences is N2 = 3, and R = 1 (frequency correspondence).

[0047] Step S1225-7: For the restrictive relationship, analyze the restrictive direction and the details of the restrictive scope of the main element on the associated element. The restrictive direction includes promoting the development of the associated element and inhibiting the development of the associated element. The details of the restrictive scope are the range limiting the change of the associated element.

[0048] Determine the restrictive direction based on the nature of the main element and medical knowledge: when the main element is a positive factor (such as "good nutritional status"), the restrictive direction on the associated element (such as "wound healing") is promotion; when the main element is a negative factor (such as "coagulation dysfunction"), the restrictive direction is inhibition. The details of the restrictive scope are determined by clinical guidelines. For example, in "coagulation dysfunction restricts the surgical blood loss", the restrictive scope is "the surgical blood loss needs to be controlled within Q milliliters" (Q is the safety threshold calculated based on the prothrombin time PT and the international normalized ratio INR, Q = Q0×(1 - INR / INR_normal), where Q0 is the allowable blood loss under normal coagulation function).

[0049] Step S1225-8: Structurally record the specific analysis results of each associated relationship, generate associated relationship description entries. Each associated relationship description entry includes the main element, the associated element, the associated type, and the associated details.

[0050] The description entries of the relationships are stored in JSON format, with specific fields including: main element (name, type, standardized terminology), related element (name, type, standardized terminology), relationship type (causal / accompanying / constraining), and relationship details (the transmission path of the causal relationship, the temporal details of the accompanying relationship, and the direction and scope of the constraining relationship). For example: {"main element":{"name":"gallbladder stones","type":"structure","standardized terminology":"gallbladder disease"},"related element":{"name":"cholecystitis","type":"symptoms","standardized terminology":"acute cholecystitis"},"relation type":"causal relationship","relation details":{"transmission path":"gallbladder stone impaction → cystic duct obstruction → bile stasis → bacterial infection → inflammatory response"}}.

[0051] Step S1225-9: Cross-validate all relationship description entries, compare relationship descriptions involving the same set of clinical elements in different expressions, adjust inconsistent descriptions, and summarize all verified relationship description entries to form the relationship analysis results corresponding to the expressions that reflect the relationship.

[0052] A conflict detection mechanism for associations is established. When the same set of clinical elements (subject + association) exhibits different association types in different descriptions (e.g., "gallstones lead to cholecystitis" vs. "cholecystitis causes gallstones"), verification is performed by consulting authoritative medical literature (e.g., UpToDate, clinical guidelines), retaining descriptions that conform to medical consensus. For temporal inconsistencies (e.g., A precedes B vs. B precedes A), corrections are made by comparing the original timestamp data. After verification, all association description entries are grouped by subject element, and an association resolution result report is generated, including association type statistics, main association paths, and conflict correction records for each group of elements.

[0053] Step S123: Divide the node types of surgical decision-making. The node types include surgical indication judgment node, surgical path planning node, surgical risk control node, and postoperative recovery plan node. Each node type has a corresponding surgical decision-making direction.

[0054] Based on the clinical decision-making process and the Delphi method to determine the node types, N surgical experts (N≥10) were invited to nominate and score the decision nodes, and nodes with a score ≥4 points (out of 5) were selected as core nodes. The decision-making directions for the surgical indication judgment node included symptom assessment (pain frequency, duration, severity), gallbladder function assessment (contraction rate, wall thickness), and non-surgical treatment effect assessment (drug response index, changes in attack interval); the decision-making directions for the surgical pathway planning node included trocar placement selection (based on body type and lesion location), instrument selection (conventional laparoscopy / 3D laparoscopy / robotic surgery), and surgical approach design (antegrade / retrograde resection); the decision-making directions for the surgical risk control node included bleeding risk assessment (coagulation function, vascular anatomy), adjacent organ damage risk (anatomical variations, degree of adhesion), and conversion to open surgery risk (degree of inflammation, surgeon's experience); the decision-making directions for the postoperative recovery plan node included pain management (analgesia regimen selection), dietary guidance (liquid / semi-liquid / regular diet transition), and complication prevention (infection, bile leakage monitoring). Each decision direction includes 3-5 specific evaluation indicators, forming a three-level structure of node-direction-indicator.

[0055] Step S124: Match the corresponding scene description elements for each node type, match the basic condition description elements with the surgical indication judgment node, match the disease course description elements with the surgical risk control node, match the anatomical structure description elements with the surgical path planning node, and match the early intervention description elements with the postoperative recovery plan node to generate the initial correspondence between nodes and elements.

[0056] Construct a node-element matching matrix, where rows represent node types (4 types), columns represent scenario description element types (basic condition / disease course / anatomical structure / early intervention), and matrix elements are matching weights (0-1). Initial weights are set according to clinical decision-making logic: the matching weight W1 between the surgical indication judgment node and the basic condition description element is 0.8, and its weight is <0.3 compared to other elements; the matching weight W2 between the surgical risk control node and the disease course description element is 0.75; the matching weight W3 between the surgical pathway planning node and the anatomical structure description element is 0.9; and the matching weight W4 between the postoperative recovery plan node and the early intervention description element is 0.85. The matching degree between each scenario description element and node is calculated by weighted summation. An initial correspondence is established when the matching degree >Th4 (e.g., 0.6). For example, the matching degree between "ASA classification" in the basic condition description element and the surgical indication judgment node = 0.8 × ASA weight = 0.8 × 0.9 = 0.72 > 0.6, therefore a matching relationship is established.

[0057] Step S125: Perform semantic association analysis on the scene description elements corresponding to each node type, explore the intrinsic association between scene description elements in different information subsets, explore the mutual influence relationship between anatomical structure description elements and basic condition description elements, and explore the causal relationship between disease course description elements and early intervention description elements.

[0058] A graph neural network (GNN)-based association mining model is employed, representing scene description elements as graph nodes and co-occurrence relationships between elements as edges. The GNN learns node embedding vectors, and vector similarity is calculated to measure association strength. For anatomical structure description elements (e.g., "gallbladder adhesion to surrounding tissues") and baseline condition description elements (e.g., "BMI value"), the association strength S is calculated as the number of co-occurrences divided by the total number of samples by analyzing the co-occurrence frequency of N patient data. A mutual influence relationship is determined when S > Th5 (e.g., 0.3) (e.g., the incidence of adhesion in patients with a BMI > 28 is K times higher than in patients with a BMI < 24). For disease course description elements (e.g., "symptom exacerbation") and early intervention description elements (e.g., "insufficient drug dosage"), causal associations are mined using time-series causal inference algorithms (e.g., Granger causality test). A causal relationship is determined when the p-value < 0.05.

[0059] Step S126: Based on the discovered inherent relationships, optimize the initial correspondence between nodes and elements, integrate the scene description elements with inherent relationships into the same node type, and generate the correspondence between nodes and related element groups.

[0060] The initial correspondence is adjusted based on the correlation strength S. When the correlation strength S between anatomical structure description element A and basic condition description element B is greater than Th6 (e.g., 0.5), B is moved from its original node (e.g., the surgical indication judgment node) to the node where A is located (e.g., the surgical path planning node). For example, "coagulation function indicators" originally belonged to the basic condition description element, and its correlation strength S with the disease course description element "bleeding risk" of the surgical risk control node is 0.7 > 0.5, so it is integrated into the surgical risk control node. The adjusted scene description elements are sorted according to their correlation strength to form related element groups. Each related element group contains core elements (with the highest correlation strength) and auxiliary elements (with the next highest correlation strength).

[0061] Step S127: Arrange the link nodes in the surgical decision-making scenario association link according to the sequence of surgical decision-making processes, based on the node type as the link node in the surgical decision-making scenario association link, to generate the initial decision-making scenario link structure.

[0062] Based on the clinical decision-making process for laparoscopic cholecystectomy, the sequence of nodes is determined as follows: surgical indication assessment → surgical pathway planning → surgical risk control → postoperative recovery plan. A directed graph structure is used to represent the initial decision-making scenario chain, where nodes represent decision node types, directed edges represent the direction of the decision process, and decision dependencies are labeled on the edges (e.g., surgical pathway planning depends on the result of the surgical indication assessment). Each node contains attributes: node ID, a list of decision directions, and associated element group IDs. Edges contain attributes: dependency type (strong dependency / weak dependency) and triggering conditions (e.g., surgical pathway planning is triggered when the surgical indication assessment result is "suitable for surgery").

[0063] Step S128: Sort the associated element groups corresponding to the link nodes in each surgical decision-making scenario association link according to the degree of association with the decision-making needs of the link nodes in the surgical decision-making scenario association link, and generate ordered association element groups. The scenario description elements in each ordered association element group are arranged in order of priority.

[0064] The Analytic Hierarchy Process (AHP) was used to calculate the weights of the correlation between related elements and the decision-making needs of nodes. Surgical experts were invited to compare the importance of each element pairwise to construct a judgment matrix, and the weight vector was calculated using the eigenvalue method. For example, in the related element group of the surgical indication judgment node, the weights of "gallbladder wall thickness" (W1=0.35), "symptom frequency" (W2=0.3), "drug treatment effect" (W3=0.25), and "comorbidities" (W4=0.1) were sorted from high to low to generate an ordered related element group: [gallbladder wall thickness, symptom frequency, drug treatment effect, comorbidities].

[0065] Step S129: Bind the ordered associated element group to the corresponding link node in the surgical decision-making scenario association link, set the calling order of the ordered associated element group under the link node in each surgical decision-making scenario association link, and generate the preliminary surgical decision-making scenario association link.

[0066] A node-feature binding protocol establishes a mapping relationship between ordered related feature groups and link nodes. Each node contains a pointer to the related feature group and a call order table. The call order is set to call the features sequentially according to their priority. During the decision-making process, the feature with the highest weight is processed first, followed by the next highest weight feature. For example, the call order of the surgical indication judgment node is as follows: first process "gallbladder wall thickness"; if its value is greater than the B threshold (e.g., 3mm), then continue processing "symptom frequency"; if "gallbladder wall thickness" is less than or equal to the B threshold, then directly output the decision suggestion "not to perform surgery for now". The binding relationship and call order are stored in the decision process configuration file, allowing clinical experts to adjust them through a visual interface.

[0067] Step S1210: Based on the preliminary surgical decision-making scenario association links, optimize the links, merge the link nodes in the surgical decision-making scenario association links with duplicate functions, supplement the missing key decision nodes, and generate surgical decision-making scenario association links.

[0068] Duplicate nodes are identified through functional similarity calculations. When the overlap in decision-making directions between two nodes is greater than Th7 (e.g., 0.8), they are considered functionally duplicated and merged (the node with higher priority is retained, and the related element groups of the other node are integrated). For example, the functional overlap between the "surgical risk assessment node" and the "surgical risk control node" is 0.85 (>0.8), so they are merged into the "surgical risk control node." Missing nodes are supplemented through a clinical decision completeness check. If a "preoperative preparation assessment node" is missing, it is added according to guidelines, defining its decision-making direction (e.g., fasting time confirmation, preoperative medication verification) and related element groups (e.g., "fasting status" in the baseline description element, "preoperative medication history" in the intervention record element). The optimized process is modeled and validated using Petri nets to ensure deadlock-free and reachable decision-making processes.

[0069] Step S130: Extract the decision elements corresponding to each node from the surgical decision scenario association link, integrate the decision elements corresponding to each node extracted from the surgical decision scenario association link according to the surgical decision logic, and generate a surgical decision adaptation factor set. Each surgical decision adaptation factor in the surgical decision adaptation factor set corresponds to the needs of a specific surgical decision node.

[0070] The algorithm iterates through each node in the surgical decision-making scenario's interconnected links, extracting scenario description elements from the ordered sets of associated elements bound to each node. Each scenario description element is semantically decomposed to obtain basic semantic units (e.g., "gallbladder wall," "thickness," "increase"), and units that directly influence the decision outcome are selected (e.g., irrelevant units like "patient name" are removed). These basic semantic units are aggregated by attribute category (condition / structure / intervention / risk) to form attribute-semantic aggregates. These aggregates are then chained together according to the decision-making process, undergoing semantic fusion (e.g., "increased gallbladder wall thickness" and "increased bleeding risk" are merged into "increased gallbladder wall thickness leads to increased bleeding risk"), and assigned decision-related identifiers (e.g., "surgical risk control - bleeding risk"), ultimately generating a set of surgical decision-adaptive factors. Each adaptive factor contains a list of basic semantic units, a set of attribute categories, and a decision-related identifier, forming a many-to-one mapping relationship with the link nodes.

[0071] Step S131: Traverse each link node in the surgical decision-making scenario association link and obtain the ordered association element group bound to each link node in the surgical decision-making scenario association link.

[0072] A depth-first search (DFS) algorithm is used to traverse the directed graph structure of the surgical decision-making scenario's interconnected links. Starting from the initial node (the surgical indication judgment node), each node and its adjacent nodes are visited sequentially. For each node, the associated feature group index table is queried using the node ID to obtain the bound ordered associated feature group. For example, when visiting the surgical path planning node, the index table returns the ordered associated feature group ID of that node as G2, and then the feature list corresponding to G2 is retrieved from the feature group library: [gallbladder location, adhesion degree, body shape characteristics, surgeon's experience]. During the traversal, the node visit order and the number of feature group calls are recorded for subsequent decision-making process optimization.

[0073] Step S132: Semantically decompose the scene description elements in each ordered associated element group, and decompose each scene description element into multiple indivisible semantic basic units, each of which carries specific clinical information meaning.

[0074] Step S1321: Select the first scene description element in each ordered related element group, decompose the sentence structure of the scene description element, and separate the grammatical components of subject, predicate, object, attributive, and adverbial.

[0075] A component decomposition method based on dependency parsing is adopted, and the Stanford Parser tool is used to perform syntactic analysis on the scene description element statements to generate a syntactic tree. Taking the first element "increased gallbladder wall thickness" in the ordered associative element group as an example, the syntactic tree shows: subject (gallbladder wall), modifier (thickness), and predicate (increase). For complex sentence structures (such as "the patient suffers right upper quadrant pain due to gallstones"), the decomposition is performed to extract the subject (patient), adverbial (due to gallstones), predicate (cause), and object (right upper quadrant pain). The decomposition results are stored in the form of a grammatical component-word mapping table, such as {"subject":["gallbladder wall"], "modifier":["thickness"], "predicate":["increase"]}.

[0076] Step S1322: Extract the words or phrases that carry the core semantics in each grammatical component, and perform independent semantic judgment on the extracted core words or phrases to determine whether each core word or phrase can independently express a clinical information meaning.

[0077] The information gain algorithm is used to calculate the semantic information content of each word. A word is considered a core word when its information content is greater than Th8 (e.g., 0.6). Core words are extracted from grammatical components: the subject "gallbladder wall" has an information content of 0.85 > 0.6, the modifier "thickness" has an information content of 0.7 > 0.6, and the predicate "increase" has an information content of 0.75 > 0.6. Independent semantic judgment is achieved by querying a medical terminology database. If a word has an independent entry in the database (e.g., "gallbladder wall" corresponds to an anatomical entry), it is determined that it can independently express clinical information; if it is a modifier (e.g., "mild"), it is determined that it cannot express clinical information independently.

[0078] Step S1323: For core words or phrases that can independently express the meaning of complete clinical information, directly identify the core words or phrases as semantic basic units; for core words or phrases that cannot independently express the meaning of complete clinical information, analyze the semantic relationship between the core words or phrases and other core words or phrases, combine the core words or phrases with semantic relationship to generate combined units that independently express the meaning of complete clinical information, and identify the combined units as semantic basic units.

[0079] "Gallbladder wall," "thickness," and "increase" can all independently express complete clinical information, and therefore are identified as basic semantic units. For the core term "mild," which cannot be expressed independently, its semantic association (modifier-head relationship) with "thickening" is analyzed, and "mild thickening" is generated as a basic semantic unit. The combination process uses semantic role labeling to determine the modification relationship, ensuring that the combined unit expresses complete meaning (e.g., "mild" modifies "thickening," not "gallbladder wall").

[0080] Step S1324: Semantically decompose all scene description elements in the ordered associated element group to generate a set of semantic basic units corresponding to each scene description element. Perform uniqueness screening on the semantic basic units in each set of semantic basic units and remove duplicate semantic basic units.

[0081] For each scene description element in the ordered associated element group, repeat the semantic decomposition process of steps S1321-S1323. For example, "gallbladder stone size is D8" is decomposed into "gallbladder stone", "size", and "D8"; "gallbladder duct diameter dilation" is decomposed into "gallbladder duct", "diameter", and "dilation". Deduplication is performed on the semantic basic unit set of each element. A hash table is used to store the unit content. When a new unit completely matches an existing unit in the table (including terms and attributes), it is determined to be a duplicate and removed. For example, if "gallbladder wall" appears multiple times in the same element set, only one is retained.

[0082] Step S1325: Label the semantic role of each semantic basic unit in the original scene description elements. The semantic role includes the description object, description attribute, description action, and description relationship.

[0083] Semantic role labeling models (such as PropBank) are used to label basic semantic units with roles: describing objects (representing clinical entities, such as "gallbladder wall" or "gallbladder stones"), describing attributes (representing entity features, such as "thickness," "size," or "diameter"), describing actions (representing state changes, such as "increase," "expansion," or "embedding"), and describing relationships (representing connections between elements, such as "leading to" or "accompanying"). During the labeling process, general semantic roles are adjusted in conjunction with medical domain knowledge. For example, "thickness" is explicitly defined as a "descriptive attribute" in the medical context, rather than the general role of "topic."

[0084] Step S1326: Arrange the labeled semantic basic units according to the logical order of the semantic basic units in the original scene description elements to generate an ordered semantic basic unit sequence corresponding to each scene description element. Summarize the ordered semantic basic unit sequences corresponding to all scene description elements to generate a set of semantic basic units corresponding to the entire ordered associated element group.

[0085] The labeled semantic basic units are arranged according to the sentence order of the original scene description elements. For example, the ordered sequence of "gallbladder wall thickness increases" is [(gallbladder wall, description object), (thickness, description attribute), (increase, description action)]. All ordered sequences of elements are summarized to form a set of semantic basic units, represented using set theory as U={u1, u2, ..., un}, where ui is the semantic basic unit and its role label. The set is stored while retaining the unit source information (original element ID, position in the sequence) for easy traceability.

[0086] Step S133: Filter the semantic basic units in each ordered associated element group that directly affect the decision results of the link nodes in the corresponding surgical decision-making scenario association link, remove the semantic basic units that do not affect the decision direction, and generate a set of semantic basic units corresponding to the link nodes in each surgical decision-making scenario association link.

[0087] A decision impact assessment model is constructed, with semantic basic units and node decision directions as inputs and an impact score (0-1) as output. The model is trained based on historical decision data, and the contribution of each unit to the decision result is calculated through logistic regression. For example, the impact score of "gallbladder wall thickness" on the surgical indication judgment node is 0.85 > Th9 (0.5), so it is retained; the impact score of "patient occupation" is 0.1 < 0.5, so it is removed. The filtered semantic basic units are sorted by impact score to form a set of semantic basic units corresponding to each node, which includes unit ID, impact score, and role label.

[0088] Step S134: Perform attribute labeling on the semantic basic units in each semantic basic unit set. According to the type of clinical information carried by the semantic basic unit, label its attribute category. The attribute category includes disease description attribute, structural description attribute, intervention description attribute, and risk description attribute.

[0089] Establish a clinical information type classification system: Disease description attributes (representing disease state, such as "pain," "inflammation," "thickening"), structural description attributes (representing anatomical features, such as "gallbladder wall," "cystic duct," "diameter"), intervention description attributes (representing treatment measures, such as "medications," "surgery," "nursing care"), and risk description attributes (representing potential hazards, such as "bleeding," "infection," "obstruction"). Using medical ontology (such as SNOMEDCT), map basic semantic units to the classification system; for example, "gallbladder wall" maps to a structural description attribute, "thickening" maps to a disease description attribute, and "bleeding risk" maps to a risk description attribute. Attribute labeling results are stored as unit-attribute key-value pairs.

[0090] Step S135: Aggregate semantic basic units of the same attribute category to generate attribute semantic aggregates corresponding to the link nodes in each surgical decision-making scenario association link. Each attribute semantic aggregate contains all semantic basic units of the same attribute category.

[0091] Using attribute categories as aggregation keys, the set of basic semantic units is grouped to form attribute semantic aggregates. For example, in the set of basic semantic units for the surgical indication judgment node, the structural description attribute aggregate includes {gallbladder wall, cystic duct, gallstone}, and the condition description attribute aggregate includes {thickening, dilation, embedding}. Each aggregate contains attribute category labels, a list of units, and relationships between units (such as the relationship between "gallbladder wall" and "thickening"). During the aggregation process, co-occurrence analysis is used to supplement the semantic relationships between units, constructing a semantic network within the attribute.

[0092] Step S136: Following the logical order of surgical decision-making, connect the attribute semantic aggregates corresponding to the link nodes in all surgical decision-making scenario association links, so that the attribute semantic aggregates of the link nodes in the previous surgical decision-making scenario association link and the attribute semantic aggregates of the link nodes in the next surgical decision-making scenario association link generate a logical association.

[0093] The semantic aggregates of attributes are linked according to the decision node sequence (surgical indication judgment → surgical pathway planning → surgical risk control → postoperative recovery plan), and logical connections are established through decision dependencies. For example, the condition description aggregate of the surgical indication judgment node (containing "thickening" and "frequent symptoms") and the structural description aggregate of the surgical pathway planning node (containing "gallbladder location" and "adhesion degree") are logically linked through "surgical necessity": when the condition description aggregate meets the surgical indication, the evaluation of the structural description aggregate is triggered. The association is implemented through a rule engine, with the rule form being "if surgical indication judgment.condition description.thickening = yes and frequency > threshold, then surgical pathway planning.structural description.adhesion degree = needs evaluation".

[0094] Step S137: Perform semantic fusion on the concatenated attribute semantic aggregates, link and integrate the basic semantic units in different attribute semantic aggregates with logical relationships, eliminate semantic redundancy, and generate fused semantic units.

[0095] A semantic fusion model based on an attention mechanism is employed. This model encodes concatenated aggregate sequences and captures logical associations between units through self-attention (e.g., the causal association between "gallbladder wall thickening" and "increased bleeding risk"). Unit pairs with association strength > Th10 (e.g., 0.7) are integrated to generate fused semantic units (e.g., "gallbladder wall thickening leads to increased bleeding risk"). Simultaneously, a redundancy detection algorithm (calculating cosine similarity between units) eliminates duplicate information; for example, "gallbladder stones" and "cholelithiasis" are standardized into the same unit. The fusion process preserves the attribute categories and semantic roles of the original units and adds association type labels (e.g., "causal" and "accompanying").

[0096] Step S138: Assign a decision association identifier to each fusion semantic unit. The decision association identifier is used to indicate the surgical decision-making process and scope of influence corresponding to the fusion semantic unit.

[0097] The decision-related identifier adopts a three-level coding structure of "decision-making stage - assessment item - impact level", such as "surgical risk control - bleeding risk - high". Here, the decision-making stage corresponds to the link node name, the assessment item is the specific decision direction under the node, and the impact level is the degree of influence of the unit on the decision (high / medium / low). The identifier generation rules are based on the attribute category and association type of the fused semantic unit. For example, the fused unit with the risk description attribute corresponds to the "surgical risk control" stage, and the unit with the causal association type has an impact level of "high". The identifier is stored as metadata and bound to the fused semantic unit.

[0098] Step S139: Define the fusion semantic unit with decision association identifier as a surgical decision adaptation factor. Each surgical decision adaptation factor contains a basic semantic unit, an attribute category, and a decision association identifier.

[0099] The structure of the surgical decision-making adaptor factors is as follows: {Semantic basic unit: [u1, u2, ..., un], attribute category: {c1, c2, ..., cm}, decision association identifier: "stage-item-level"}. For example, adaptor factor F1: {Semantic basic unit: [gallbladder wall, thickening, bleeding risk, increased], attribute category: {condition description attribute, risk description attribute}, decision association identifier: "surgical risk control - bleeding risk - high"}. Adaptor factors are identified by unique IDs for easy subsequent adjustment and traceability.

[0100] Step S1310: Collect all surgical decision adaptation factors, arrange them according to the decision-making process sequence corresponding to the surgical decision adaptation factors, and generate a surgical decision adaptation factor set. Each surgical decision adaptation factor in the surgical decision adaptation factor set is associated with a link node in a certain surgical decision scenario association link.

[0101] The process iterates through all decision nodes, collecting surgical decision adaptation factors generated at each node. These factors are then arranged in the order of decision-making stages (surgical indication assessment → surgical pathway planning → surgical risk control → postoperative recovery plan), forming a surgical decision adaptation factor set. The set uses a list data structure, with each element containing an adaptation factor ID, a decision association identifier, and a corresponding node ID. A mapping relationship between adaptation factors and link nodes is established through the node IDs, ensuring that each adaptation factor can be traced back to its corresponding decision node. The set also includes metadata: generation time, version number, and data source, used for version control and auditing.

[0102] Step S140: By analyzing the correlation between different types of information in the multi-source clinical information set, the surgical decision adaptation factor set is dynamically adjusted so that each surgical decision adaptation factor matches the needs of the corresponding surgical decision node.

[0103] This process extracts the change trajectories of various information types from a multi-source clinical information set (such as changes in the status of basic health records and trends in disease progression records), and identifies change nodes (such as the time point when gallbladder wall thickness increases from B0 to B1) through time series analysis. These change nodes are then semantically compared with the semantic basic units in the adaptor factors to determine the association type (direct / indirect / unrelated), the scope of influence (affected semantic basic units), and the mode of influence (supplementation / correction / replacement / addition). Key change elements (change nodes that directly affect and have a high degree of influence) are screened, and the semantic basic units, attribute categories, and decision association identifiers of the adaptor factors are updated. Finally, a logical coherence check (consistency of adaptor factors across different stages) is performed to integrate and generate an adjusted set of adaptor factors.

[0104] Step S141: Extract the information change trajectory of different types of information in the multi-source clinical information set. The information change trajectory covers the status changes of basic health records, the trend changes of disease progression records, the feature changes of surgical related image descriptions, and the effect changes of diagnosis and treatment intervention records.

[0105] A time-series extraction algorithm was used to extract information change trajectories from a multi-source clinical information set. The status change trajectory of basic health records was constructed from regular physical examination data (e.g., quarterly BMI measurement sequences); the trend change trajectory of disease progression records was constructed from symptom diaries and physical sign records (e.g., daily pain score sequences); the characteristic change trajectory of surgical-related imaging descriptions was extracted from historical imaging reports (e.g., bi-weekly gallbladder wall thickness measurement sequences); and the effect change trajectory of diagnostic and treatment intervention records was constructed by comparing indicators before and after treatment (e.g., inflammatory marker sequences before and after medication). Trajectory data was stored in timestamp-value pairs, missing values ​​were filled using linear interpolation, and outliers were detected and corrected using the IQR method.

[0106] Step S142: Bind the information change trajectory to the corresponding link node in the surgical decision-making scenario association link, and label the information change trajectory type and change characteristics corresponding to each link node in the surgical decision-making scenario association link.

[0107] The trajectory-node matching algorithm binds information change trajectories to link nodes. The algorithm calculates the similarity between trajectory features (such as change frequency, amplitude, and trend) and node decision requirements; the node with the highest similarity is the binding target. For example, the trajectory of gallbladder wall thickness change (features: slow upward trend, amplitude ΔB) has the highest similarity to the surgical indication judgment node, therefore it is bound. The trajectory type (state / trend / feature / effect change) and change characteristics (rising / falling / fluctuating, linear / non-linear, periodic / non-periodic), for example, the gallbladder wall thickness trajectory is labeled as "feature change - linear rise".

[0108] Step S143: Through semantic comparison, analyze the association type between each information change trajectory and the surgical decision adaptation factor bound to the link node in the corresponding surgical decision scenario association link. The association type is used to qualitatively describe the degree of semantic association between the two. Based on the association type, determine the scope and mode of influence of the information change trajectory on the surgical decision adaptation factor.

[0109] Step S1431: Extract the change nodes in each information change trajectory. Each change node corresponds to the time point when the information description content changes and the content of the change.

[0110] A change point detection algorithm (such as the cumulative sum control chart method) is used to identify change nodes in the information change trajectory. A change node is identified when the trajectory value exceeds the control limit (mean ± 3 standard deviations). For example, in the gallbladder wall thickness trajectory, if the measured value at time point T exceeds the upper control limit, it is identified as a change node, and the change content is "gallbladder wall thickness increased from the previous value to the current value". Each change node includes a timestamp, the previous value, the current value, the amount of change (current value - previous value), and the rate of change (amount of change / time interval). Step S1432: Extract the semantic basic units from the surgical decision adaptation factors bound to the link nodes in the corresponding surgical decision scenario association link, and extract the clinical information content carried by each semantic basic unit.

[0111] Extract a list of semantic basic units from the surgical decision adaptation factors bound to the link nodes. For example, adaptation factor F1 of the surgical risk control node contains the semantic basic units "gallbladder wall", "thickening", "bleeding risk", and "increase". Map each unit to clinical information content through terminology standardization, such as "thickening" mapping to "increased gallbladder wall thickness" and "increase" mapping to "increased bleeding risk level".

[0112] Step S1433: Perform a semantic comparison between the changed content of the changed node and the clinical information content of the semantic basic unit to determine the association type between the two. The association type includes direct association, indirect association, and no association.

[0113] A BERT-based semantic similarity calculation model was used to convert the change content and clinical information content into vectors, and the cosine similarity S was calculated. When S > Th11 (0.8), it was considered directly related (e.g., the similarity S = 0.92 between "gallbladder wall thickness increased to B1" and "gallbladder wall thickening"); when 0.4 ≤ S ≤ 0.8, it was considered indirectly related (e.g., the similarity S = 0.65 between "increased blood pressure" and "increased bleeding risk"); when S < 0.4, it was considered unrelated (e.g., the similarity S = 0.2 between "occupational change" and "gallbladder wall thickening").

[0114] Step S1434: Based on the determined association type, select the semantic basic units that are associated with the change node. The semantic basic units that are associated with the change node constitute the influence range of the information change trajectory on the surgical decision adaptation factor.

[0115] The scope of influence includes semantic basic units with direct and indirect correlation types. For example, the change node "gallbladder wall thickness increases to B1" includes both the directly related unit "gallbladder wall thickening" and the indirectly related unit "increased bleeding risk". The scope of influence is represented as a set of semantic basic unit IDs, and is tagged with correlation type and similarity value during storage.

[0116] Step S1435: Classify the change types of the change nodes. The change types include supplementary change, corrective change, substitution change, and new change.

[0117] Changes are categorized based on their relationship to the original information: Supplementary changes refer to the addition of new information (e.g., "adding a description of gallbladder polyps"); Corrective changes refer to modifications of existing information (e.g., "changing gallbladder wall thickness from B0 to B1"); Substitute changes refer to replacing existing information (e.g., "changing the diagnosis from cholecystitis to gallstones"); and New changes refer to the appearance of new types of information (e.g., "first appearance of jaundice symptoms"). The change type is determined by comparing the information fingerprints (generated using a hash algorithm) before and after the change. Different fingerprints with partial overlap indicate a corrective change; completely different fingerprints indicate a substitute change; newly added fields indicate supplementary changes; and new types of fields indicate new changes.

[0118] Step S1436: For different types of changes, set the corresponding impact methods. The impact method for supplementary changes is to add new semantic basic units in the surgical decision adaptation factor. The impact method for corrective changes is to modify some of the descriptive content in the semantic basic units. The impact method for alternative changes is to replace the original semantic basic units with new semantic basic units. The impact method for new changes is to add a completely new semantic basic unit category in the surgical decision adaptation factor.

[0119] For supplementary changes, add semantic basic units corresponding to the new information to the adaptation factors (e.g., add the "gallbladder polyp" unit); for corrective changes, modify the attribute values ​​of the units (e.g., change "gallbladder wall thickness B0" to "B1"); for substitution changes, delete the original units and add new units (e.g., delete the "cholecystitis" unit and add the "gallbladder stones" unit); for novel changes, create new attribute categories and add units (e.g., add the "jaundice" unit to the disease description attribute). The mapping relationship between the impact method and the change type is stored in the rule base and can be adjusted through the expert interface.

[0120] Step S1437: Associate and bind the semantic basic units in the scope of influence with the corresponding influence methods, and set the influence type that each semantic basic unit in the scope of influence will be affected by.

[0121] Construct an impact binding table, where the semantic basic unit IDs within the scope of the behavior's impact are listed, columns represent the impact method, and cells represent the binding status (yes / no). For example, the change node "gallbladder wall thickness increases to B1" (a corrective change) is bound to the semantic basic unit "gallbladder wall thickening" with the "corrective" impact method; and to "increased bleeding risk" with the "indirect corrective" impact method. The binding process automatically matches based on the association type and change type; directly related units with corrective changes are bound to the "corrective" method, and indirectly related units are bound to the "indirect corrective" method.

[0122] Step S1438: Construct an influence relation matrix. The row dimension of the influence relation matrix represents the change nodes, the column dimension represents the semantic basic units within the influence range, and the elements of the influence relation matrix represent the corresponding influence methods.

[0123] The influence relationship matrix is ​​a two-dimensional array M, where M[i][j] represents the influence of the i-th change node on the j-th semantic basic unit. For example, if the matrix rows represent the change nodes "gallbladder wall thickening" and "high blood pressure", and the columns represent the semantic basic units "gallbladder wall thickening" and "increased bleeding risk", then M[0][0] = "correction", M[0][1] = "indirect correction", and M[1][1] = "indirect correction". The matrix is ​​stored using a sparse matrix, recording only non-empty influence methods.

[0124] Step S1439: Based on the influence relationship matrix, generate an influence analysis report on the impact of each information change trajectory on the corresponding surgical decision adaptation factor. The influence analysis report records the scope of influence, the mode of influence, and the basis for association.

[0125] The impact analysis report includes the following: basic information on the information change trajectory (type, start and end times, and trend), a list of change nodes (timestamp, change content, and change type), the scope of impact (semantic basic unit ID and name), an impact relationship matrix (visualized table), and the basis for association (semantic similarity value and medical knowledge citation). For example, the report states that "the trajectory of gallbladder wall thickness change directly correlates with the 'gallbladder wall thickening' unit (S=0.92), the influence method is correction, and the basis is that ultrasound examination results show the thickness increased from B0 to B1." The report is generated in PDF format and can be downloaded and reviewed by clinical experts.

[0126] Step S144: Based on the scope and mode of influence, select the change elements that have a direct impact on the corresponding surgical decision adaptation factor from the information change trajectory as key change elements. Each key change element corresponds to a specific semantic basic unit in the surgical decision adaptation factor.

[0127] Define the key change element filtering criteria: the association type is directly related (S>0.8), the change type is modified or substitutional, and the impact method is modification or substitution. Select change elements that meet the criteria from the change nodes in the information change trajectory. For example, "gallbladder wall thickness increased from B0 to B1" (directly related, modified, modification method) is selected as a key change element, corresponding to the semantic basic unit "gallbladder wall thickening". Key change elements are stored including the values ​​before and after the change, the change time, and the corresponding unit ID.

[0128] Step S145: Integrate key change elements into the corresponding surgical decision adaptation factors, replace the semantic basic units in the surgical decision adaptation factors that conflict with the change elements, and supplement the new semantic basic units brought about by the change elements.

[0129] The adaptation factors are updated based on how key changes affect the data: For corrective changes, the attribute values ​​of the semantic basic units are updated to the changed values ​​(e.g., the thickness value for "gallbladder wall thickening" is changed from B0 to B1); for substitution changes, the original units are deleted and new units are added (e.g., "cholecystitis" is deleted and "gallbladder stones" is added); for supplementary changes, new units are added to the adaptation factors (e.g., "gallbladder polyps" is added). Version changes are recorded during the update process, and the unit content before modification is retained for auditing purposes.

[0130] Step S146: Adjust the attribute category labeling of the surgical decision adaptation factor after incorporating key change elements to ensure that the attribute category is consistent with the updated semantic basic unit.

[0131] When the clinical information type of a semantic basic unit changes (e.g., "jaundice" changes from no attribute to a condition description attribute), its attribute category label is updated. The unit type is remapped using a medical terminology database to ensure that the attribute category matches the unit content. For example, if a new unit "jaundice" is mapped to a condition description attribute, it is added to the condition description attribute set of the adaptor factor.

[0132] Step S147: Update the decision association identifier of the surgical decision adaptation factor and adjust the decision influence range indicated by the identifier according to the impact of key change factors on the surgical decision-making process.

[0133] If a key change causes a change in the decision impact level (e.g., "increased bleeding risk" changes from "medium" to "high"), then the impact level portion of the decision association identifier is updated (e.g., from "surgical risk control - bleeding risk - medium" to "-high"). The impact level adjustment is based on the clinical risk assessment model, such as calculating the risk level based on the ΔB value: ΔB > 2mm is "high", 1-2mm is "medium", and < 1mm is "low".

[0134] Step S148: Re-match the adjusted surgical decision adaptation factor with the corresponding link node in the surgical decision scenario association link, so that the semantic basic unit of the surgical decision adaptation factor can reflect the decision requirements of the link node in the surgical decision scenario association link.

[0135] The matching degree is calculated by rematching the adaptation factors and nodes based on the matching degree of decision requirements. The matching degree is the sum of the correlation strength between the semantic basic units in the adaptation factor and the decision direction of the node / the total number of units. When the matching degree is greater than Th12 (0.7), the matching is considered successful; otherwise, the unit composition of the adaptation factor is adjusted. For example, the matching degree of the adjusted "bleeding risk" adaptation factor and the surgical risk control node is 0.85 > 0.7, so the matching is successful.

[0136] Step S149: Perform a logical coherence comparison analysis on all adjusted surgical decision adaptation factors, compare the surgical decision adaptation factors of the previous decision stage with the surgical decision adaptation factors of the next decision stage, adjust the semantic basic units in the surgical decision adaptation factors of the next decision stage, integrate all the adjusted and compared surgical decision adaptation factors, and generate an adjusted surgical decision adaptation factor set.

[0137] A logical consistency check algorithm is used to compare the adaptation factors of the preceding and following stages. If the adaptation factor of the preceding stage is "suitable for surgery" (surgical indication judgment), while the adaptation factor of the following stage contains a "no surgery required" related unit, a logical contradiction is determined. The contradiction is resolved by backtracking and adjusting the basic semantic units of the following stage (e.g., deleting the "no surgery required" unit and adding the "surgery required" unit). After the comparison is completed, the adaptation factors are integrated according to the decision-making stage order to generate an adjusted set, which includes a version number and adjustment record.

[0138] Step S150: Based on the adjusted set of surgical decision adaptation factors, generate a surgical decision plan and plan adaptation instructions for a specific patient, and push the surgical decision plan and plan adaptation instructions to the surgical clinical decision terminal.

[0139] Adaptive factors are grouped according to the decision-making stage. Each group is semantically integrated to generate a comprehensive semantic description (e.g., "Patient has moderate gallbladder wall thickening, frequent symptoms, meets surgical indications"). Decision options (e.g., "Laparoscopic cholecystectomy" or "Open cholecystectomy") are generated based on the comprehensive semantic description and a clinical decision rule base. Logical relationships between options are analyzed to generate sequences (e.g., "Laparoscopic surgery → Postoperative analgesia plan → Dietary guidance"). Implementation details (e.g., trocar location, anesthesia method) are supplemented to generate a preliminary plan. The basis for generation (semantic units in the adaptive factors) is extracted to generate an adaptation description fragment. This fragment is integrated and supplemented with logical relationship descriptions to generate a plan adaptation description. The plan is revised based on surgical guidelines. Finally, the plan and description are integrated and encrypted before being pushed to the clinical decision-making terminal (e.g., physician workstation) via the hospital intranet.

[0140] Step S151: According to the surgical decision-making process, the surgical decision-making adaptation factors in the adjusted surgical decision-making adaptation factor set are grouped, with each group corresponding to a surgical decision-making stage.

[0141] The system uses decision-making stage labels to group matching factors, with label values ​​corresponding to link node names (surgical indication judgment / surgical pathway planning / surgical risk control / postoperative recovery plan). It iterates through the adjusted set of matching factors and assigns them to the corresponding group based on the "decision-making stage" field in the decision association identifier of each factor. For example, matching factors with the decision association identifier "surgical indication judgment - symptom assessment - high" are assigned to the surgical indication judgment group. Each group contains a list of matching factors, a group ID, and a decision-making stage name, supporting independent management and viewing by stage.

[0142] Step S152: Semantically integrate the surgical decision adaptation factors in each group, summarize the basic semantic units in the surgical decision adaptation factors of the same decision-making stage, and generate a comprehensive semantic description of each decision-making stage.

[0143] A Transformer-based text generation model was used to semantically integrate the adaptors within the groups. The model input was a sequence of basic semantic units of the adaptors, and the output was a coherent and comprehensive semantic description. For example, the adaptors for the surgical indication judgment group included units such as "gallbladder wall thickening," "frequent symptoms," and "poor response to drug treatment." After integration, a comprehensive semantic description was generated: "The patient's gallbladder wall thickness is B1, right upper quadrant pain occurs N3 times per month, and after taking ursodeoxycholic acid D4mg / time for T2 days, the duration of symptom relief shortened to t4 minutes. Non-surgical treatment was ineffective, meeting the indications for laparoscopic cholecystectomy." During the integration process, the model automatically supplemented conjunctions and clinical terms to ensure that the description conformed to medical documentation standards.

[0144] Step S153: Based on the comprehensive semantic description of each decision-making stage, generate the decision options corresponding to that decision-making stage. Each decision option directly corresponds to the requirements in the comprehensive semantic description.

[0145] For example, step S1531: The comprehensive semantic description of each decision-making step is parsed in layers and divided into a demand layer, a constraint layer, and an adaptation layer. The demand layer contains the clinical needs that the decision-making step must meet, the constraint layer contains the clinical restrictions that must be followed during the implementation of the decision-making step, and the adaptation layer contains the clinical conditions required for the implementation of the decision-making step.

[0146] A hierarchical semantic parsing model is used to divide the comprehensive semantic description into layers. The demand layer is determined by extracting keywords such as "must," "need," and "goal," such as "complete gallbladder removal" and "controlling bleeding risk." The constraint layer is determined by extracting keywords such as "restriction," "cannot," and "must," such as "patient is obese" and "comorbid hypertension." The adaptation layer is determined by extracting keywords such as "possesses," "needs," and "conditions," such as "laparoscopic equipment is available" and "surgeon is experienced." The parsing results are stored in a three-layer structure tree, with each layer containing keywords, content, and priority.

[0147] Step S1532: Extract the requirement statement from the requirement layer and extract the goals that need to be achieved in this decision-making process. The goals include lesion resection, function preservation, and risk reduction.

[0148] Clinical goal terms are identified from the needs layer. Lesion resection goals include "gallbladder removal" and "stone removal"; function preservation goals include "preservation of Oddi sphincter function" and "bile duct protection"; and risk reduction goals include "reducing bleeding" and "avoiding organ damage." Need descriptions are matched using a goal classification dictionary; for example, "complete gallbladder resection" matches the "lesion resection" goal, and "controlling bleeding risk" matches the "risk reduction" goal. Each goal includes a goal name, an importance weight (1-5 points), and assessment indicators.

[0149] Step S1533: Analyze the constraint statements in the constraint layer and extract the boundary conditions that cannot be broken during the implementation of this decision-making process. The boundary conditions include the patient's physical tolerance limit and the surgical operation standard limit.

[0150] Boundary conditions are extracted from the constraint layer. Patient physical tolerance restrictions include "ASA classification ≤ III", "BMI < 35", and "normal liver and kidney function"; surgical procedure restrictions include "contraindications for laparoscopic surgery" and "aseptic technique requirements". The constraint expressions are matched using a clinical constraint rule base. For example, "patient BMI = 32" matches the boundary condition "BMI < 35", and "normal coagulation function" matches the boundary condition "prothrombin time < 15 seconds". Boundary conditions are stored as conditional expressions (e.g., BMI < 35) and consequences of violation.

[0151] Step S1534: Extract the adaptation representation from the adaptation layer and extract the supporting conditions required for the smooth implementation of this decision-making process. The supporting conditions include the patient's physiological indicators and medical equipment conditions.

[0152] Supporting conditions were extracted from the adaptation layer, and the patient's physiological status included "hemoglobin ≥100g / L" and "platelet count ≥100×10⁶". 9 " / L" etc.; medical equipment conditions include "laparoscopic system available" and "ultrasonic scalpel functioning normally," etc. Adaptation descriptions are matched through a resource matching list. For example, "laparoscopic equipment available" matches the condition "laparoscopic system available," and "operator with more than 50 laparoscopic experiences" matches the condition "operator qualifications." Supporting conditions include condition name, current status (met / not met), and priority.

[0153] Step S1535: Based on the stated objective, and in conjunction with the surgical clinical routine decision-making scheme library, select multiple candidate decision-making schemes that can meet the objective.

[0154] The surgical clinical routine decision-making database contains standard surgical protocol templates (such as laparoscopic cholecystectomy and open cholecystectomy). Each template includes the target matching degree (the degree of fit with each clinical goal), implementation steps, and required resources. The total target matching degree of the protocol templates is calculated based on the needs-level goals, and templates with a total score greater than Th13 (e.g., 80 points) are selected as candidate decision-making protocols. For example, "laparoscopic cholecystectomy" has a target matching degree of 90 points for "lesion resection" and 85 points for "risk reduction," with a total score of 175 points, which is greater than 160 points (Th13), and is therefore selected as a candidate protocol.

[0155] Step S1536: Compare each candidate decision scheme with the constraint description of the constraint layer, eliminate candidate decision schemes that violate the constraints, retain candidate decision schemes that meet the constraints, match the retained candidate decision schemes with the adaptation description of the adaptation layer, and select candidate decision schemes that generate corresponding matching relationships with the adaptation conditions.

[0156] A constraint satisfaction algorithm is used to evaluate candidate decision schemes. The implementation conditions of each scheme are compared with the boundary conditions of the constraint layer. If a violation is found (e.g., "open surgery" requires BMI < 30, but the patient's BMI = 32), the scheme is eliminated. For the retained schemes, the matching degree with the support conditions of the adaptation layer is calculated (number of satisfied conditions / total number of conditions). Schemes with a matching degree > Th14 (e.g., 0.8) proceed to the next stage. For example, "laparoscopic cholecystectomy" satisfies 8 / 10 support conditions, with a matching degree of 0.8 > 0.8, and passes the screening.

[0157] Step S1537: Supplement the candidate decision schemes that generate corresponding matching relationships with details. Combine the specific clinical information in the comprehensive semantic description, supplement the specific operational points and precautions in the implementation process of the scheme, arrange the implementation steps in the candidate decision schemes after supplementing the details according to the logical sequence of surgical operations, supplement the connecting text between the steps, and integrate them into decision options.

[0158] Based on the patient's specific circumstances (such as gallbladder location and degree of adhesion) in the comprehensive semantic description, supplement the operational points, such as "choose three trocar sites: umbilicus, below the xiphoid process, and below the right costal margin," and "use blunt dissection when separating adhesions around the gallbladder." Precautions include "avoid damaging the cystic artery" and "prevent bile leakage." Arrange the implementation steps according to the surgical procedure (disinfection → establishing pneumoperitoneum → inserting a trocar → exploration → resection → hemostasis → closure of the abdomen), and supplement connecting instructions (such as "after establishing pneumoperitoneum, pay attention to the needle insertion angle when inserting the trocar"). The integrated decision options include the protocol name, a list of implementation steps, operational points, and precautions.

[0159] Step S1538: Uniquely identify each decision option, mark the target, constraint compliance status, and adaptation status corresponding to each decision option, and finally generate the set of decision options corresponding to this decision-making stage.

[0160] Assign a unique ID to each decision option (e.g., Option A, Option B), and label its corresponding objective (e.g., "lesion resection + risk reduction"), constraint compliance status (e.g., "meets all constraints"), and fit condition matching status (e.g., "8 / 10 conditions satisfied"). Generate a set of decision options, sorted from highest to lowest matching degree, containing information such as option ID, name, objective, status, and matching degree, to facilitate selection by clinicians.

[0161] Step S154: Analyze the logical relationship between decision options in different decision-making stages. Based on the decision options of the previous decision-making stage, adjust the content of the decision options in the next decision-making stage to generate a coherent sequence of decision options.

[0162] A rule base for decision-making options is constructed to define the influence of options in preceding stages on options in subsequent stages, such as "selecting laparoscopic surgery" (surgical path planning) → "postoperative analgesia using nonsteroidal anti-inflammatory drugs" (postoperative recovery plan). The content of subsequent stage options is adjusted based on the rule base. For example, if "laparoscopic surgery" is selected in the preceding stage, then "epidural analgesia" (commonly used in open surgery) is removed from the subsequent analgesia options, while "intravenous analgesia" is retained. A logical reasoning engine ensures the consistency of the option sequence and avoids contradictions (such as selecting laparoscopy first, then selecting postoperative care after open surgery).

[0163] Step S155: Arrange the decision option sequence according to the order of surgical procedures, supplement the implementation details of each decision option, and generate a preliminary surgical decision plan.

[0164] The decision-making options are arranged in the order of preoperative preparation → surgical procedure → postoperative care. Preoperative preparation includes "fasting and fluid restriction for 8 hours" and "preoperative medication (antibiotics)". Surgical procedure includes "laparoscopic three-port cholecystectomy" and "anesthesia method (general anesthesia)". Postoperative care includes "analgesia plan (intravenous PCA)" and "dietary guidance (liquid diet for 6 hours postoperatively)". Supplementary implementation details include: fasting start time T10, antibiotic name M4 and dosage D9, operation duration t7, and postoperative monitoring indicators (body temperature, blood pressure, abdominal pain score). The preliminary plan is presented in a structured document format, including a plan title, a list of implementation steps, detailed descriptions, and the person in charge.

[0165] Step S156: Extract the surgical decision adaptation factor corresponding to each decision option, and extract the generation basis of each decision option. The generation basis includes the semantic basic unit, attribute category, and decision association identifier in the surgical decision adaptation factor.

[0166] The process of generating decision options traces their corresponding surgical decision-making adaptation factors. For example, the option "laparoscopic cholecystectomy" corresponds to adaptation factors F1 (gallbladder wall thickening) and F2 (mild adhesion). The semantic basic units (gallbladder wall, thickening, adhesion, mild), attribute categories (condition description attributes, structural description attributes), and decision association identifiers (surgical path planning - operational difficulty - medium) of these adaptation factors are extracted as the basis for option generation. This information is stored as a many-to-many mapping between option IDs and adaptation factor IDs.

[0167] Step S157: Associate and bind each decision option with its generation basis to generate an adaptation description fragment for each decision-making stage.

[0168] The decision options and their generation basis are associated via IDs to generate an adaptation description fragment in the format "Select [Option Name], based on [Adaptation Factor 1 Semantic Description] ([Attribute Category], [Decision Association Identifier]), [Adaptation Factor 2 Semantic Description] ([Attribute Category], [Decision Association Identifier])". For example, "Select laparoscopic three-port cholecystectomy, based on the patient's gallbladder wall thickening (condition description attribute, surgical path planning - operation difficulty - medium) and mild adhesion between the gallbladder and surrounding tissues (structural description attribute, surgical path planning - approach selection - easy)". The fragment includes the option ID, description text, and the source of the basis (adaptation factor ID).

[0169] Step S158: Integrate the adaptation description fragments of all decision-making stages according to the decision-making process sequence, supplement the logical relationship description between the adaptation descriptions of different decision-making stages, and generate the solution adaptation description.

[0170] The integrated protocol description is structured according to the decision-making process (surgical indication assessment → surgical pathway planning → surgical risk control → postoperative recovery plan), supplementing the logical connections with additional descriptions. For example, "Due to the patient's gallbladder wall thickening (surgical indication assessment), hemostasis must be carefully considered when choosing laparoscopic surgery (surgical risk control), therefore, enhanced postoperative bleeding monitoring is necessary (postoperative recovery plan)." The integrated protocol description includes explanatory segments for each stage, logical connections, and an overall conclusion, presented in a natural paragraph format to facilitate physician understanding of the decision-making basis.

[0171] Step S159: Retrieve the surgical procedure specification text, check each implementation detail description in the preliminary surgical decision plan one by one, mark the content that is inconsistent with the specification text and correct it; adjust the connection order between decision options according to the logical order of surgical implementation, and generate a surgical decision plan with optimized process.

[0172] The hospital's knowledge base was used to retrieve laparoscopic cholecystectomy procedure guidelines (such as the "Clinical Practice Guidelines - Surgery Section"). A text comparison algorithm was employed to verify the implementation details of the preliminary plan (such as the disinfection area, trocar size, and hemostasis method). Inconsistencies were marked (e.g., the guidelines require disinfection to extend below the xiphoid process, while the plan specifies below the costal margin) and corrected. A workflow optimization algorithm was used to adjust the order of options (e.g., moving "postoperative analgesia" after "operative completion" instead of "return to ward") to ensure adherence to the surgical procedure's logic. The optimized plan includes correction records and a version number.

[0173] Step S1510: Integrate the optimized surgical decision-making scheme and scheme adaptation description to generate the final surgical decision-making scheme and scheme adaptation description. The scheme adaptation description is used to show the correspondence between each decision content and the surgical decision adaptation factor.

[0174] The optimized surgical decision-making plan (structured steps) and the plan adaptation instructions (natural language basis) are integrated into the final document. The plan section includes the surgical name, implementation steps, and detailed requirements; the adaptation instructions section includes the decision-making basis for each step (corresponding adaptation factors). For example, the "puncture site selection" step in the plan corresponds to the description "determine the puncture coordinates based on the gallbladder location (structural description attribute) and body shape characteristics (basic condition attribute)". The final document is in PDF format and supports adding electronic signatures and version control.

[0175] Step S1511: Push the surgical decision-making scheme and scheme adaptation instructions to the surgical clinical decision-making terminal.

[0176] The final document is encrypted and sent to the surgeon's clinical decision-making terminal (such as the electronic medical record system on the doctor's workstation) via the hospital information system's push service. The push process uses the HTTPS protocol and includes document metadata (patient ID, protocol version, generation time) and access control (only surgical team members can view it). Upon receiving the document, the terminal generates a notification, and the doctor can view, download, or print the document online. Viewing and modification logs are recorded for auditing purposes.

[0177] In one exemplary embodiment, a machine learning-based surgical clinical decision support system is provided. This system can be a terminal, server, etc., and its internal structure diagram can be as follows: Figure 2As shown, this machine learning-based surgical clinical decision support system includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, near-field communication, or other technologies. When the computer program is executed by the processor, it implements a machine learning-based surgical clinical decision support method. The display unit generates visually visible images and can be a display screen, projection device, or virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device can be a touch layer covering the display screen, or a button, trackball, or touchpad set on the shell of a machine learning-based surgical clinical decision support system, or an external keyboard, touchpad, or mouse, etc.

[0178] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A surgical clinical decision support method based on machine learning, characterized in that, The method includes: The system collects a set of multi-source clinical information generated by the patient throughout the entire surgical diagnosis and treatment process. This set of multi-source clinical information includes the patient's basic health record, disease progression record, surgical-related image description, and diagnosis and treatment intervention record. The patient's basic health record includes descriptions of physical condition, past medical history, and genetic background. The disease progression record includes descriptions of symptom changes, changes in physical signs, and changes in examination results. The surgical-related image description includes descriptions of lesion location, tissue anatomical relationships, and physiological structure morphology. The diagnosis and treatment intervention record includes descriptions of medication use, nursing measures, and the effects of previous treatments. Based on the multi-source clinical information set, a surgical decision-making scenario association link is constructed. The surgical decision-making scenario association link takes the specific surgical decision-making needs as nodes, and associates and binds the information elements related to each node in the multi-source clinical information set to generate a corresponding relationship chain between nodes and information elements. The decision elements corresponding to each node are extracted from the surgical decision scenario association link. The decision elements corresponding to each node extracted from the surgical decision scenario association link are integrated according to the surgical decision logic to generate a surgical decision adaptation factor set. Each surgical decision adaptation factor in the surgical decision adaptation factor set corresponds to the needs of a specific surgical decision node. By analyzing the correlation between different types of information in the multi-source clinical information set, the surgical decision adaptation factor set is dynamically adjusted so that each surgical decision adaptation factor matches the needs of the corresponding surgical decision node. Based on the adjusted set of surgical decision adaptation factors, a surgical decision plan and plan adaptation instructions are generated for specific patients, and the surgical decision plan and plan adaptation instructions are pushed to the surgical clinical decision terminal.

2. The surgical clinical decision support method based on machine learning according to claim 1, characterized in that, The construction of the surgical decision-making scenario association link based on the multi-source clinical information set includes: The multi-source clinical information set is divided into a basic information subset, a disease course information subset, an image description subset, and an intervention record subset according to information type. Each subset contains all clinical information description content of the corresponding type. From each information subset, scenario description elements related to surgical decision-making are extracted. These scenario description elements are specific descriptions of the conditions related to surgical decision-making. From the basic information subset, basic condition description elements affecting surgical tolerance are extracted. From the disease course information subset, disease course description elements reflecting the trend of disease development are extracted. From the image description subset, anatomical structure description elements affecting the choice of surgical path are extracted. From the intervention record subset, early intervention description elements affecting the adjustment of surgical plan are extracted. The nodes for surgical decision-making are categorized into types, including surgical indication judgment, surgical pathway planning, surgical risk control, and postoperative recovery plan. Each node type has a corresponding surgical decision-making direction. Match the corresponding scenario description elements for each node type, match the basic condition description elements with the surgical indication judgment node, match the disease course description elements with the surgical risk control node, match the anatomical structure description elements with the surgical path planning node, and match the early intervention description elements with the postoperative recovery plan node to generate the initial correspondence between nodes and elements. Semantic association analysis is performed on the scene description elements corresponding to each node type to explore the intrinsic association between scene description elements in different information subsets, explore the mutual influence relationship between anatomical structure description elements and basic condition description elements, and explore the causal relationship between disease course description elements and early intervention description elements. Based on the discovered inherent relationships, the initial correspondence between nodes and elements is optimized, and scene description elements with inherent relationships are integrated into the same node type to generate the correspondence between nodes and related element groups. The link nodes in the surgical decision-making scenario are arranged according to the sequence of surgical decision-making processes, based on node type as the link node in the associated link, to generate the initial decision-making scenario link structure. The associated element groups corresponding to the link nodes in each surgical decision-making scenario are sorted according to their degree of association with the decision-making needs of the link nodes in the surgical decision-making scenario, generating ordered associated element groups. The scenario description elements in each ordered associated element group are arranged in order of priority. Bind ordered association element groups to the corresponding link nodes in the surgical decision-making scenario association link, set the calling order of ordered association element groups under the link nodes in each surgical decision-making scenario association link, and generate a preliminary surgical decision-making scenario association link. Based on the preliminary surgical decision-making scenario association links, link optimization is performed, link nodes in surgical decision-making scenario association links with redundant functions are merged, missing key decision nodes are added, and surgical decision-making scenario association links are generated.

3. The surgical clinical decision support method based on machine learning according to claim 1, characterized in that, The step involves extracting decision elements corresponding to each node from the surgical decision-making scenario association link, integrating the decision elements corresponding to each node extracted from the surgical decision-making scenario association link according to surgical decision-making logic, and generating a surgical decision adaptation factor set, including: Traverse each link node in the surgical decision-making scenario association link to obtain the ordered association element group bound to each link node in the surgical decision-making scenario association link. Semantic decomposition is performed on the scene description elements in each ordered associative element group, and each scene description element is divided into multiple indivisible basic semantic units, each of which carries specific clinical information meaning. In each ordered associated element group, semantic basic units that directly affect the decision results of the link nodes in the associated link of the corresponding surgical decision scenario are selected, and semantic basic units that do not affect the decision direction are removed, thereby generating a set of semantic basic units corresponding to the link nodes in the associated link of each surgical decision scenario. The semantic basic units in each semantic basic unit set are labeled with attributes. According to the type of clinical information carried by the semantic basic unit, its attribute category is labeled. The attribute category includes disease description attribute, structural description attribute, intervention description attribute, and risk description attribute. The semantic basic units of the same attribute category are aggregated to generate the attribute semantic aggregate corresponding to the link node in each surgical decision scenario association link. Each attribute semantic aggregate contains all the semantic basic units of the same attribute category. Following the logical order of surgical decision-making, the attribute semantic aggregates corresponding to the link nodes in all surgical decision-making scenario association links are linked together, so that the attribute semantic aggregates of the link nodes in the previous surgical decision-making scenario association link and the attribute semantic aggregates of the link nodes in the next surgical decision-making scenario association link generate a logical association. Semantic fusion is performed on the concatenated attribute semantic aggregates to integrate the basic semantic units in different logically related attribute semantic aggregates, eliminate semantic redundancy, and generate fused semantic units. Each fusion semantic unit is assigned a decision association identifier, which is used to indicate the surgical decision-making process and scope of influence corresponding to the fusion semantic unit. The fusion semantic unit with decision association identifier is defined as the surgical decision adaptation factor. Each surgical decision adaptation factor contains a basic semantic unit, attribute category, and decision association identifier. All surgical decision adaptation factors are collected and arranged according to the decision-making process sequence corresponding to the surgical decision adaptation factors to generate a surgical decision adaptation factor set. Each surgical decision adaptation factor in the surgical decision adaptation factor set is associated with a link node in a certain surgical decision scenario association link.

4. The surgical clinical decision support method based on machine learning according to claim 1, characterized in that, The process of dynamically adjusting the surgical decision-making adaptation factor set by analyzing the correlations between different types of information in a multi-source clinical information set, ensuring that each surgical decision-making adaptation factor matches the needs of the corresponding surgical decision node, includes: Extract the information change trajectory of different types of information in the multi-source clinical information set. The information change trajectory covers the status changes of basic health records, the trend changes of disease progression records, the feature changes of surgical related image descriptions, and the effect changes of diagnosis and treatment intervention records. The information change trajectory is bound to the corresponding link node in the surgical decision-making scenario association link, and the information change trajectory type and change characteristics corresponding to each link node in the surgical decision-making scenario association link are marked. By semantic comparison, the association type between each information change trajectory and the surgical decision adaptation factor bound to the link node in the corresponding surgical decision scenario association link is analyzed. The association type is used to qualitatively describe the degree of semantic association between the two. Based on the association type, the influence range and influence mode of the information change trajectory on the surgical decision adaptation factor are determined. Based on the scope and mode of influence, the changing elements that have a direct impact on the corresponding surgical decision adaptation factor are selected from the information change trajectory as key changing elements. Each key changing element corresponds to a specific semantic basic unit in the surgical decision adaptation factor. Key change elements are integrated into the corresponding surgical decision adaptation factors, replacing the semantic basic units in the surgical decision adaptation factors that conflict with the change elements, and supplementing the new semantic basic units brought about by the change elements. Adjust the attribute category labeling of the surgical decision adaptation factors after incorporating key change elements to ensure that the attribute categories are consistent with the updated semantic basic units; Update the decision association identifiers of surgical decision adaptation factors, and adjust the decision impact range indicated by the identifiers according to the impact of key changing factors on the surgical decision-making process; The adjusted surgical decision adaptation factor is re-matched with the corresponding link node in the surgical decision scenario association link, so that the semantic basic unit of the surgical decision adaptation factor can reflect the decision requirements of the link node in the surgical decision scenario association link. A logical coherence comparison analysis is performed on all adjusted surgical decision adaptation factors. The surgical decision adaptation factors of the previous decision stage are compared with those of the subsequent decision stage. The semantic basic units in the surgical decision adaptation factors of the subsequent decision stage are adjusted. All surgical decision adaptation factors that have been adjusted and compared are integrated to generate an adjusted set of surgical decision adaptation factors.

5. The surgical clinical decision support method based on machine learning according to claim 1, characterized in that, The adjusted set of surgical decision adaptation factors generates surgical decision-making plans and plan adaptation instructions for specific patients, including: According to the surgical decision-making process, the surgical decision-making factors in the adjusted surgical decision-making adaptation factor set are grouped, with each group corresponding to a surgical decision-making stage. Semantic integration is performed on the surgical decision adaptation factors in each group, and the basic semantic units in the surgical decision adaptation factors of the same decision-making stage are summarized to generate a comprehensive semantic description of each decision-making stage. Based on the comprehensive semantic description of each decision-making stage, the corresponding decision options are generated, and each decision option directly corresponds to the requirements in the comprehensive semantic description. Analyze the logical relationships between decision options at different decision-making stages, and based on the decision options of the previous decision-making stage, adjust the content of the decision options of the next decision-making stage to generate a coherent sequence of decision options. The decision option sequence is arranged according to the order of surgical procedures, and the implementation details of each decision option are supplemented to generate a preliminary surgical decision plan. Extract the surgical decision adaptation factor corresponding to each decision option, and extract the generation basis of each decision option. The generation basis includes the semantic basic unit, attribute category, and decision association identifier in the surgical decision adaptation factor. Each decision option is associated with and bound to its generation basis, generating an adaptation description fragment for each decision-making stage; The adaptation descriptions of all decision-making stages are integrated in the order of the decision-making process, and the logical relationship between the adaptation descriptions of different decision-making stages is supplemented to generate the solution adaptation descriptions. Retrieve the surgical procedure manual, verify each implementation detail in the preliminary surgical decision-making plan, mark any inconsistencies with the manual and correct them; adjust the order of decision options according to the logical sequence of surgical execution, and generate an optimized surgical decision-making plan. The optimized surgical decision-making scheme and scheme adaptation description are integrated to generate the final surgical decision-making scheme and scheme adaptation description. The scheme adaptation description is used to show the correspondence between each decision content and the surgical decision adaptation factor.

6. The surgical clinical decision support method based on machine learning according to claim 2, characterized in that, The process of extracting scene description elements related to surgical decisions from each subset of information includes: Read the clinical information description content of each information subset one by one, and split each clinical information description content into independent description statements according to the statement structure; Semantic analysis is performed on each descriptive statement to identify the core expressions in the descriptive statement that reflect the patient's physical condition, changes in the condition, anatomical structure, and treatment intervention; By comparing the core statements with the direction of surgical decision-making, the core statements related to surgical implementation, risk control, and outcome prediction were selected. The core statements selected were categorized according to their clinical significance, distinguishing between statements reflecting the current situation, statements reflecting changes, and statements reflecting relationships. Specific features are extracted from statements reflecting the current situation, including details on body parts, symptoms, structural features, and intervention measures. Trend analysis is performed on statements reflecting changes, extracting information on the direction, magnitude, and time span of state changes. Association analysis is performed on statements reflecting relationships, extracting information on causal, accompanying, and restrictive relationships between different clinical elements. The core content, after feature extraction, trend analysis, and correlation analysis, is structured and organized to generate descriptive fragments. Each descriptive fragment contains specific clinical information and the type of relationship it reflects. Each descriptive fragment is defined as a scene description element. Based on the type of information subset, all scene description elements are categorized and summarized to generate a set of scene description elements corresponding to each information subset.

7. The surgical clinical decision support method based on machine learning according to claim 3, characterized in that, The semantic decomposition of scene description elements in each ordered associative element group, dividing each scene description element into multiple indivisible semantic basic units, includes: Select the first scene description element in each ordered associative element group, and decompose the sentence structure of the scene description element to separate the grammatical components of subject, predicate, object, attributive, and adverbial. Extract the words or phrases that carry the core semantics from each grammatical component, and make independent semantic judgments on the extracted core words or phrases to determine whether each core word or phrase can independently express a clinical information meaning. For core words or phrases that can independently express the meaning of complete clinical information, directly identify the core words or phrases as semantic basic units; for core words or phrases that cannot independently express the meaning of complete clinical information, analyze the semantic relationship between the core words or phrases and other core words or phrases, combine the core words or phrases with semantic relationship to generate combined units that can independently express the meaning of complete clinical information, and identify the combined units as semantic basic units. Semantically decompose all scene description elements in the ordered associated element group to generate a set of semantic basic units corresponding to each scene description element. Then, uniquely filter the semantic basic units in each set of semantic basic units and remove duplicate semantic basic units. Each semantic basic unit is labeled with its semantic role in the original scene description elements. The semantic roles include description object, description attribute, description action, and description relationship. The labeled semantic basic units are arranged according to their logical order in the original scene description elements to generate an ordered semantic basic unit sequence for each scene description element. The ordered semantic basic unit sequences for all scene description elements are then summarized to generate a set of semantic basic units for the entire ordered associated element group.

8. The surgical clinical decision support method based on machine learning according to claim 4, characterized in that, The process involves semantic comparison to analyze the association type between each information change trajectory and the surgical decision adaptation factor bound to the link node in the corresponding surgical decision scenario association link. This association type qualitatively describes the degree of semantic association between the two. Based on the association type, the scope and manner of the impact of the information change trajectory on the surgical decision adaptation factor are determined, including: Extract the change nodes in each information change trajectory. Each change node corresponds to the time point when the information description content changed and the content of the change. Extract the basic semantic units from the surgical decision adaptation factors bound to the link nodes in the corresponding surgical decision-making scenario association link, and extract the clinical information content carried by each basic semantic unit. The changes in the nodes are semantically compared with the clinical information content of the basic semantic units to determine the type of association between them. The type of association includes direct association, indirect association, and no association. Based on the determined association type, semantic basic units that are associated with the change node are selected. The semantic basic units that are associated with the change node constitute the influence range of the information change trajectory on the surgical decision adaptation factor. The change types of change nodes are classified, including supplementary changes, corrective changes, substitution changes, and new changes; For different types of changes, corresponding impact methods are set. The impact method for supplementary changes is to add new semantic basic units to the surgical decision adaptation factors. The impact method for corrective changes is to modify some of the descriptive content in the semantic basic units. The impact method for alternative changes is to replace the original semantic basic units with new semantic basic units. The impact method for new changes is to add a completely new semantic basic unit category to the surgical decision adaptation factors. Associate and bind the basic semantic units within the scope of influence with the corresponding influence methods, and set the type of influence that each basic semantic unit within the scope of influence will receive; Construct an influence relation matrix, where the row dimension of the influence relation matrix represents the change nodes, the column dimension represents the basic semantic units within the influence range, and the elements of the influence relation matrix represent the corresponding influence methods. Based on the aforementioned influence relationship matrix, an influence analysis report is generated on the corresponding surgical decision adaptation factor for each information change trajectory. The influence analysis report records the scope of influence, the mode of influence, and the basis for the association.

9. A surgical clinical decision support system based on machine learning, characterized in that, include: processor; A machine-readable storage medium for storing machine-executable instructions of the processor; The processor is configured to execute the machine learning-based surgical clinical decision support method according to any one of claims 1 to 8 by executing the machine executable instructions.

10. A computer program product, characterized in that, The computer program product includes machine-executable instructions stored in a computer-readable storage medium. The processor of the machine-based surgical clinical decision support system reads the machine-executable instructions from the computer-readable storage medium and executes the machine-executable instructions, causing the machine-based surgical clinical decision support system to perform the machine-based surgical clinical decision support method as described in any one of claims 1 to 8.

Images