A method and system for collaborative diagnosis and treatment of spinal disorders

Abstract

The application discloses a kind of spinal disease collaborative diagnosis and treatment method and system, symptom lack item map is formed to spinal image data and diagnosis and treatment record sparse annotation, extract lack feature and form staging clue mark, jointly labeled generation diagnosis and treatment node chain;According to the initial diagnosis and treatment opinion set obtained by the diagnosis and treatment node chain through de-anchor scheduling, time series confidence analysis forms confidence weight graph;Confidence weight graph is used to exclude false consensus interference, and difference index map is formed by analyzing the difference of department opinion, extract difference mode to form complementary atlas, and the initial diagnosis and treatment opinion set is fused to establish diagnosis and treatment knowledge graph;Conflict items and imbalance items in cumulative bias are identified, blind area mark and balance mark are rearranged to generate deployment instruction;Based on the deployment instruction, the collaborative diagnosis and treatment path graph is generated by optimizing the diagnosis and treatment knowledge graph, the residual difference is extracted to form the expansion node set to revise the path, and the spinal collaborative diagnosis and treatment result is output, to realize the joint elimination of multi-department opinion anchor interference and false consensus suppression.

Description

Technical Field

[0001] This invention relates to the field of medical decision support technology, and in particular to a collaborative diagnosis and treatment method and system for spinal diseases. Background Technology

[0002] Spinal diseases have long courses and complex symptoms, and the quality of their diagnosis and treatment highly depends on the completeness and objectivity of multidisciplinary joint assessments. In actual consultation processes, departments submit their opinions in a fixed order. Subsequent departments tend to revise their opinions based on those of earlier departments rather than conducting independent assessments. The problem of homogeneity of opinions deepens with each round of scheduling, and the final diagnostic direction is dominated by the anchoring effect of a few early-submitting departments. Departments with truly high diagnostic accuracy, due to their later submission order or lower confidence weights, fail to have their opinions adequately reflected in the integration process.

[0003] Furthermore, existing solutions lack the ability to identify pseudo-consensus arising from overlapping professional responsibilities when dealing with the degree of departmental opinion concentration. They count highly similar opinions from adjacent departments as multiple independent supports, causing genuine specialist disagreements to be masked by false consensus. Regarding the utilization of symptom information, missing annotations between imaging data and medical records fail to establish a correlation with disease staging characteristics, and staging clues carried by the distribution patterns of missing data are ignored. In terms of integrated conclusion output, residual opinion conflicts in highly central symptom dimensions lack an explicit presentation mechanism, and the transparency of collaborative treatment conclusions regarding long-term structural disagreements is insufficient. Summary of the Invention

[0004] This invention discloses a collaborative diagnosis and treatment method and system for spinal diseases, aiming to solve the problems in multi-department collaborative diagnosis and treatment, such as the anchoring effect of scheduling order compressing the independence of opinions, the pseudo-consensus caused by overlapping professional responsibilities masking real disagreements, the systematic neglect of the judgment basis of low-weight departments, and the lack of explicit presentation of structural residual disagreements in the integrated conclusions. Through the extraction of staging clues driven by missing features, de-anchoring scheduling and confidence weight differential assessment, pseudo-consensus correction clustering and complementary map fusion, blind zone equilibrium joint rearrangement, and residual disagreement path expansion, the invention achieves precise collaborative diagnosis and treatment of spinal diseases.

[0005] The first aspect of this invention proposes a collaborative diagnosis and treatment method for spinal diseases, comprising the following steps:

[0006] Collect spinal imaging data and medical records of patients, perform sparse annotation on the imaging data and medical records to form a symptom missing item map, and extract missing item features based on the symptom missing item map to form staging clue identifiers;

[0007] By combining the symptom missing item map with the staging clue identifier, a joint annotation is performed to form a diagnosis and treatment node chain. According to the diagnosis and treatment node chain, cross-departmental opinions are scheduled to obtain an initial diagnosis and treatment opinion set. Time-series confidence analysis is performed on the initial diagnosis and treatment opinion set to form a confidence weight map.

[0008] The confidence weight map is used to analyze the departmental opinion differences to form a divergence index map. The divergence patterns are extracted from the divergence index map to form a complementary map. The initial set of medical opinions is fused with the complementary map to establish a medical knowledge graph.

[0009] For the cumulative bias in the initial medical opinion set, identify conflict and imbalance items, extract the conflict intensity of the conflict items to form blind zone markers, perform specialty lineage matching between the imbalance items and the divergence index map to determine balance markers, and perform opinion priority rearrangement to generate allocation instructions according to the blind zone markers and the balance markers.

[0010] Based on the allocation instructions, the diagnostic and treatment knowledge graph is globally optimized to generate a collaborative diagnostic and treatment path graph. Residual divergence features are extracted from the collaborative diagnostic and treatment path graph to form an extended node set. Based on the extended node set, the collaborative diagnostic and treatment path graph is revised and the spinal collaborative diagnostic and treatment results are output.

[0011] A second aspect of this invention provides a collaborative diagnosis and treatment system for spinal diseases, comprising:

[0012] The data annotation unit is used to collect patient spinal imaging data and medical records, perform sparse annotation on the imaging data and medical records to form a symptom missing item map, and extract missing item features based on the symptom missing item map to form staging clue identifiers;

[0013] The opinion scheduling unit is used to combine the symptom missing item map with the staging clue identifier to perform joint annotation to form a diagnosis and treatment node chain, schedule cross-departmental opinions according to the diagnosis and treatment node chain to obtain an initial diagnosis and treatment opinion set, and perform time-series confidence analysis on the initial diagnosis and treatment opinion set to form a confidence weight map.

[0014] The graph construction unit is used to use the confidence weight graph to analyze the departmental opinion differences to form a divergence index graph, extract divergence patterns from the divergence index graph to form a complementary graph, and establish a diagnosis and treatment knowledge graph by fusing the initial diagnosis and treatment opinion set with the complementary graph.

[0015] The instruction generation unit is used to identify conflict items and imbalance items in the cumulative deviation of the initial medical opinion set, extract the conflict intensity of the conflict items to form blind zone markers, perform specialty lineage matching between the imbalance items and the divergence index map to determine equilibrium markers, and perform opinion priority rearrangement to generate allocation instructions according to the blind zone markers and equilibrium markers.

[0016] The result output unit is used to globally optimize the diagnosis and treatment knowledge graph based on the allocation instruction to generate a collaborative diagnosis and treatment path graph, extract residual divergence features from the collaborative diagnosis and treatment path graph to form an extended node set, and revise the path of the collaborative diagnosis and treatment path graph based on the extended node set to output the spinal collaborative diagnosis and treatment result.

[0017] The beneficial effects of this invention are reflected in the following points: First, the spatial distribution pattern of missing symptoms itself carries regular clues corresponding to disease staging, transforming this missing information discarded by traditional schemes into staging clue identifiers that drive differentiated configuration of the diagnosis and treatment node chain; at the same time, the frequency of in-depth analysis and the accuracy of counter-mainstream revisions in the revision behavior of various departments are transformed into dynamic evaluation criteria for confidence weights, and the anchoring effect is eliminated by scheduling order perturbation, extending the scope of information utilization in the opinion collection stage from explicit judgment content to two implicit dimensions: missing pattern and revision behavior. Second, the high convergence of departmental opinions is directly regarded as diagnostic consistency in existing schemes, and the homogeneous judgments caused by overlapping professional responsibilities are difficult to distinguish from the true independent verification consensus in appearance. By using the dual thresholds of professional overlap and opinion direction to identify pseudo-consensus and apply merging compression, and tracking the switching events between explicit dominance and implicit penetration influence to identify complementary boundaries, the extraction of divergent structures is based on the true specialist opinion pattern after excluding the contamination of homogeneous judgments, thus solving the problem of systematic compression of opinion diversity by pseudo-consensus. Finally, the persistently low level of conflict intensity is equated with opinion convergence in the existing scheme, while the difference between the superficial consistency driven by inertia and the true convergence in terms of temporal pattern is not identified. The deviation of historical divergence patterns is used as the basis for determining blind spots, and departments with high historical opinion adoption rates but low current confidence weights are used as the objects of balance compensation. The priority of opinions is rearranged based on behavioral history rather than static weights. For residual divergences with high centrality, the opposing evidence is explicitly presented through branch paths, so that the collaborative diagnosis and treatment conclusions can provide the mainstream judgment direction while retaining the complete visibility of structural divergences. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating a collaborative diagnosis and treatment method for spinal diseases according to the present invention.

[0019] Figure 2 This is a structural block diagram of a collaborative diagnosis and treatment system for spinal diseases according to the present invention. Detailed Implementation

[0020] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, circuits, and methods are omitted so as not to obscure the description of this application with unnecessary detail.

[0021] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0022] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0023] The technical solutions of the embodiments of this application are described below.

[0024] like Figure 1 As shown, this embodiment of the invention provides a collaborative diagnosis and treatment method for spinal diseases, including the following steps S11-S15:

[0025] Step S11: Collect the patient's spinal imaging data and medical records, perform sparse annotation on the imaging data and medical records to form a symptom missing item map, and extract missing item features based on the symptom missing item map to form staging clue identifiers.

[0026] Specifically, spinal imaging data and medical records were collected from patients. The spinal imaging data sources covered three modalities: plain X-rays, CT scans, and MRI sequences. During acquisition, the scanning position, slice thickness, and sequence parameters were recorded for each modality. The imaging resolution and tissue contrast differed among the different modalities. Plain X-rays reflected the overall arrangement of bony structures, CT scans revealed details of vertebral bone, and MRI sequences revealed the state of intervertebral discs and soft tissues. These three modalities described the spatial distribution of spinal lesions from different perspectives. The medical records collected included outpatient records, inpatient progress notes, and previous imaging reports. During acquisition, the patient's unique identifier was used as the association key, and medical records from different time points were arranged according to the visit timestamp to ensure a correspondence between spinal imaging data and medical records on the timeline. In spinal imaging data, there are instances where image quality is substandard due to patient positional shifts or motion artifacts. Such images are filtered out using image noise assessment metrics. Images with noise assessment metrics exceeding the preset upper bounds for each modality are marked as low-quality images. Low-quality images are appended with quality labels to the acquisition results, but the original data is retained. Whether to include these images in the subsequent sparse annotation stage will be determined later. Missing periods in medical records due to referrals or cross-hospital visits are marked as missing places to maintain the integrity of the timeline and prevent direct splicing on the timeline from obscuring the impact of historical treatment intervals on the assessment of disease progression. After spinal imaging data and medical records are acquired, they are separately processed and stored, aligned using the visit timestamp as the reference point, awaiting sparse annotation processing.

[0027] Sparse annotation was applied to imaging data and medical records to create a symptom missing item map. The sparse annotation framework used a spinal disease symptom system, encompassing five main symptom dimensions: intervertebral disc degeneration, spinal stenosis, vertebral slippage, scoliosis, and nerve compression. Each dimension included several observation indicators derived from both imaging features and clinical descriptions in the medical records. Imaging data was retrieved modally by category, with corresponding imaging features for each observation indicator. Retrieval results were categorized into three states: clearly visible, suspected presence, and no record found. Clearly visible indicators were marked as observed, suspected presence indicators as pending confirmation, and no record found indicators as missing items. Each modality's retrieval result was then written into the corresponding indicator position within the symptom missing item map. Each clinical description corresponding to an observed indicator in the medical record is checked line by line. Indicators explicitly mentioned in the clinical description are marked as observed in the symptom missing item chart. Indicators not mentioned are supplemented based on the image annotation results. Indicators that are marked as observed in the image but not mentioned in the medical record are judged as omissions in the record. Indicators that are missing in both the image annotation and the medical record are judged as full-dimensional missing items. The causes of the two types of missing items are different. Omissions in the record reflect that the description of the image findings in the medical record is incomplete. Full-dimensional missing items reflect that the symptom evidence for the indicator at the current stage of the consultation is insufficient. The symptom missing item chart uses symptom dimensions and observed indicators as a dual-axis index. The position of each indicator stores the corresponding annotation status. The spatial distribution of missing item markers directly reflects the completeness of the current patient's spinal symptom information and the symptom dimensions where the missing items are concentrated.

[0028] Based on the symptom deficiency map, deficiency features are extracted to form staging clues. The deficiency markers in the symptom deficiency map are statistically analyzed dimension by dimension for deficiency density. Deficiency density is calculated as the proportion of the number of observed indicators with deficiency in that dimension to the total number of observed indicators in that dimension. Symptom dimensions with high deficiency density correspond to the most incomplete symptom information for the current patient in that dimension, while symptom dimensions with low deficiency density correspond to relatively complete symptom information. The deficiency density of each symptom dimension is arranged on the five main symptom dimensions to form a deficiency density distribution vector. The peak dimension of the density distribution vector corresponds to the direction of the most concentrated symptom information deficiency. The distribution pattern of deficiency in symptom dimensions shows regular differences in different stages of spinal diseases. In the early stage, the deficiency in the nerve compression dimension is dominant; in the middle stage, the deficiency in the spinal stenosis and vertebral slippage dimensions alternates; and in the late stage, the deficiency in the scoliosis dimension is a typical feature. Comparing the deficiency density distribution vector with the typical deficiency distribution patterns of each stage forms a preliminary staging direction judgment. In the symptom missing item map, the observation indicators for the pending state are extracted separately. The distribution density of the pending indicators in the symptom dimension is used as the uncertainty distribution vector. A symptom dimension with high uncertainty indicates that the symptom information in that dimension is ambiguous, and that the imaging features and clinical descriptions do not form a consistent judgment in that dimension. The staging clue identifier consists of three components: preliminary staging direction judgment, missing item density distribution vector, and uncertainty distribution vector. In the staging clue identifier, the preliminary staging direction judgment gives the most likely staging interval corresponding to the current symptom missing item pattern. The missing item density distribution vector indicates the completeness of information in each dimension, and the uncertainty distribution vector indicates the reliability of information in each dimension. The three components of the staging clue identifier together describe the supporting ability and limitations of the current symptom missing item map in staging judgment.

[0029] Step S12: Combine the symptom missing item map with the staging clue identifier to form a diagnosis and treatment node chain. According to the diagnosis and treatment node chain, schedule cross-departmental opinions to obtain an initial diagnosis and treatment opinion set. Perform time-series confidence analysis on the initial diagnosis and treatment opinion set to form a confidence weight map.

[0030] Specifically, a diagnostic node chain is formed by combining the symptom missing item map with staging clue identifiers. This chain uses symptom dimensions as nodes and clinical dependencies in symptom evolution as edges, creating an ordered node structure. Each node is categorized into three types: information completion nodes, specialist confirmation nodes, and routine diagnostic nodes. The node type in the chain is adaptively determined based on the missing item density and uncertainty of each dimension in the symptom missing item map. Symptom dimensions with high missing item density in the symptom missing item map are designated as information completion nodes in the diagnostic node chain, where incomplete symptom information requires priority completion during scheduling. Symptom dimensions with high uncertainty are designated as specialist confirmation nodes, where inconsistent imaging features and clinical descriptions require specialist verification. Dimensions with low missing item density and low uncertainty are designated as routine diagnostic nodes, requiring no additional completion or confirmation intervention. The staging process indicated by the preliminary staging direction judgment in the staging clue identifier determines the node arrangement order of the treatment node chain. Nodes corresponding to the symptom dimension of the early stage are arranged first, and nodes corresponding to the late stage are arranged later. The weight of the edges between nodes is quantified by the frequency of synchronous occurrence of the observed indicators of two adjacent dimensions in the symptom missing item diagram. The adjacent dimensions with a high frequency of synchronous occurrence have a higher weight, reflecting the close relationship between the two in clinical evolution. The values ​​of the observed indicators of each dimension in the symptom missing item diagram are compared with the typical characteristics of each stage corresponding to the preliminary staging direction judgment in the staging clue identifier. Observed indicators that deviate from the typical characteristics of each stage are marked as staging discrepancies. The nodes in the treatment node chain of the dimension where the staging discrepancies are located are marked with discrepancy labels. Discrepancy-marked nodes trigger additional specialist verification tasks in cross-departmental scheduling. The uncertainty distribution vector and the missing item density distribution vector in the staging clue identifier jointly constrain the density distribution of each node type in the treatment node chain. The treatment node chain is composed of all nodes, edge weights, and discrepancy labels.

[0031] In some embodiments, obtaining an initial set of treatment opinions by scheduling cross-departmental opinions according to the treatment node chain includes: generating a sequence reference table by identifying the routine scheduling order of each department based on the treatment node chain; performing an anchoring risk assessment on the sequence reference table to identify high-anchoring-impact department pairs; implementing sequence perturbation processing on the high-anchoring-impact department pairs to generate a de-anchoring scheduling sequence; and issuing opinion collection tasks to each department based on the de-anchoring scheduling sequence to form an initial set of treatment opinions.

[0032] Based on the diagnosis and treatment node chain, a sequence reference table is generated to identify the routine scheduling order of each department. The node arrangement order of the diagnosis and treatment node chain determines the scheduling relationship of each department in the current patient's diagnosis and treatment path. Based on the node arrangement order of the diagnosis and treatment node chain, the corresponding department identifier is extracted node by node. When the same department is scheduled in multiple nodes, the first appearance position is taken as the department's position in the sequence reference table. The sequence reference table is formed by arranging the departments in ascending order of their positions. The lower the position, the earlier the department submits its opinion in the corresponding path of the diagnosis and treatment node chain. For parallel scheduling departments with the same position, the department with the shorter historical response time is given priority to avoid position conflicts caused by response delays in parallel scheduling. The similarity of historical opinions between adjacent departments in the sequential reference table is quantified by the mean cosine similarity of the two departments' opinions on the same type of spinal disease in the past. Adjacent departments with high historical opinion similarity suggest that their judgments on this symptom dimension tend to converge over a long period. Subsequent departments tend to make minor adjustments based on the opinions of their predecessors rather than making independent assessments. For two adjacent departments that have consistently presented similar conclusions on the intervertebral disc degeneration dimension, the independent judgment space of their subsequent departments is significantly compressed by the opinions of their predecessors under routine scheduling. Departments with adjacent positions and high historical opinion similarity have the highest anchoring risk. For departments with widely separated positions, even if the historical opinion similarity is high, the anchoring effect is diluted because opinions from multiple intermediate departments have already been involved, and the anchoring risk is relatively reduced.

[0033] Anchoring risk assessment was performed on the sequence reference table to identify high-anchoring-impact department pairs. The anchoring risk value of each adjacent department pair in the sequence reference table was defined as R = S_hist × P_adj, where S_hist is the mean similarity of historical opinions between the two departments (dimensionless, ranging from 0 to 1), P_adj is the reciprocal of the difference in rank between adjacent departments (dimensionless), with P_adj reaching its maximum value of 1 when the rank difference is 1, and the larger the rank difference, the smaller P_adj is. R is a dimensionless value; a larger R indicates a higher anchoring risk for the department pair under the standard scheduling sequence. Department pairs with an R value exceeding 1.5 times the mean R value of all adjacent department pairs in the sequence reference table were identified as high-anchoring-impact department pairs. The threshold of 1.5 times the mean strikes a balance between screening out truly high-risk department pairs and retaining reasonable opinion associations. If the threshold is too low, a large number of department pairs will be marked, and the sequence perturbation range will be too wide, causing the scheduling sequence to deviate from the clinical dependence structure; if the threshold is too high, department pairs with medium anchoring risk will be missed. The average R-value of all adjacent department pairs in the sequence reference table is 0.32, and the threshold of 1.5 is 0.48. A certain adjacent department pair has a historical similarity of 0.85 and a position difference of 1, with an R-value of 0.85, which far exceeds the threshold, and is therefore included in the set of high-anchoring-influence department pairs. After extracting the current positions of the preceding and following departments in the high-anchoring-influence department pairs, the target spacing of the sequence perturbation is determined by the difference between their positions and the R-value. The higher the R-value, the larger the target spacing. The target spacing represents the minimum number of other departments that need to be inserted between two departments in the de-anchoring scheduling sequence. By increasing the positional spacing between high-anchoring-influence department pairs, the immediate visibility of the opinions of the preceding departments when the following departments receive the task is reduced.

[0034] For departments with high anchoring influence, a sequence perturbation process is implemented to generate a de-anchoring scheduling sequence. The sequence perturbation starts from the sequence reference table and adjusts the departmental order while satisfying the target spacing constraints for all departments with high anchoring influence. The adjustment primarily involves shifting the order of subsequent departments to the next position, with the shift amount being the difference between the target spacing and the current actual spacing. Departments with high anchoring influence that already meet the target spacing are not subject to position adjustments, ensuring the principle of minimizing adjustment operations. When the position adjustments of multiple departments with high anchoring influence affect each other, the department with the highest R-value is adjusted first. After adjustment, the actual spacing of the remaining departments is reassessed. Departments that still do not meet the target spacing are shifted to the next position. This process is iterated until all constraints are satisfied. The number of iterations is bounded by the total number of departments; if there are still departments that do not meet the constraints beyond the upper bound, the current optimal solution is adopted. When a department's ranking shifts later than the suggested position in the treatment node chain, the lower bound constraint is the node arrangement order. A department's ranking cannot be earlier than the position of its first scheduled node in the treatment node chain. For example, if a primary responsible department shifts to the eighth position due to a sequence disturbance, but its first scheduled node in the treatment node chain is the sixth, then its lower bound is the sixth position. The earliest scheduling timing stipulated by clinical dependencies is not disrupted. The unanchored scheduling sequence is formed by arranging all departments in ascending order of their adjusted rankings. The difference between the unanchored scheduling sequence and the order reference table is only reflected in the shifting of the ranking of subsequent departments in high-anchored influence department pairs. The scheduling order of non-high-anchored influence department pairs remains unchanged in the unanchored scheduling sequence.

[0035] An initial set of treatment opinions is formed by issuing opinion collection tasks to each department based on a de-anchored scheduling sequence. In this sequence, departments receive opinion collection tasks sequentially from smallest to largest position. Each task includes a symptom gap map of the current patient and the type label of the corresponding node in the treatment node chain for that department. The symptom gap map is fully accessible to all departments. Opinions submitted by preceding departments with high anchoring influence are masked from the current department if the scheduling interval does not reach the target distance. The scheduling interval is quantified by the difference between the time the preceding department submitted its opinion and the time the current department receives the task. In a scenario, two adjacent responsible departments constitute a high-anchoring influence pair. In the de-anchored scheduling sequence, the target distance between them is three scheduling positions. When a subsequent department receives its task, it is masked if the preceding opinion does not exceed the three-position interval. The subsequent department then independently submits its opinion based solely on the symptom gap map and its own specialty knowledge, entering the initial set of treatment opinions. This masking mechanism significantly increases the independence of the opinions from the two departments compared to conventional scheduling. When the scheduling interval exceeds the target spacing, the shielding is lifted. Departments in the subsequent scheduling sequence can refer to previous opinions when submitting their feedback. In this case, the previous opinions have been diluted by the independent opinions of multiple intermediate departments, and their reference value does not constitute an anchoring threat. The opinion collection task has a response timeout threshold. Departments that fail to submit within the threshold are replaced with the average opinion of similar historical cases and marked as delayed submissions. The initial treatment opinion set is formed by summarizing the opinions of each department in order of submission timestamp. The independence of subsequent department opinions from departments with high anchoring influence in the initial treatment opinion set is significantly improved compared to conventional scheduling.

[0036] In some embodiments, the step of performing time-series confidence analysis to form a confidence weight graph on the initial medical opinion set includes: extracting the submission timestamp and opinion content of each opinion in the initial medical opinion set to generate a time-series opinion sequence; identifying the revision direction of the opinion sequence to generate a revision direction record; identifying counter-mainstream revision events for the revision direction record to generate an independent judgment marker; and assigning differentiated weights to the opinions of each department in the initial medical opinion set based on the independent judgment marker to form a confidence weight graph.

[0037] The submission timestamps and content of each opinion in the initial treatment opinion set are extracted to generate a time-series opinion sequence. Opinions in the initial treatment opinion set are arranged from earliest to latest submission timestamp. When multiple opinions exist at the same timestamp, the department's ranking in the de-anchoring scheduling sequence is used as the secondary sorting criterion, with the lower-ranked opinion appearing first, ensuring uniqueness in the ranking. The content of each opinion in the initial treatment opinion set is used as a symptom dimension to extract a key judgment vector. Each component of the key judgment vector corresponds to the opinion's tendency judgment in each symptom dimension, with values ​​of support, opposition, and neutral. Support indicates that the department believes the patient has a positive symptom in that dimension; opposition indicates that it does not; and neutral indicates no clear judgment. The extraction of tendency judgments is based on explicit affirmative or negative statements in the opinion text. Ambiguous or conditional statements are categorized as neutral to avoid ambiguous text being forcibly categorized as support or opposition, leading to misjudgments. Placeholder opinions marked as delayed submissions are uniformly categorized as neutral during key judgment vector extraction to avoid bias in the current judgment distribution caused by historical average substitute values. The chronological opinion sequence is composed of the submission timestamps of each opinion and the key judgment vectors arranged chronologically. The arrangement of the chronological opinion sequence on the submission time axis completely preserves the dynamic evolution of the opinion tendencies of each department in the initial treatment opinion set. In a certain treatment scheduling process, the tendencies of the opinions of a certain primary responsible department in the early stage and the opinions of another department in the middle stage reversed in a certain symptom dimension. The switching of the support and opposition states in this dimension in the chronological opinion sequence is clear and traceable.

[0038] For example, the step of identifying the revision direction of the time-series opinion sequence and generating a revision direction record includes: comparing adjacent opinions in the time-series opinion sequence to identify opinion change points; extracting the revision time interval of the opinion change points to generate a revision interval sequence; identifying deep revision events with ultra-long intervals in the revision interval sequence to generate a deep analysis marker; and performing confidence-weighted adjustments on the opinion change points based on the deep analysis marker to generate a revision direction record.

[0039] The system identifies opinion change points by comparing adjacent opinions in a time-series opinion sequence. For adjacent opinion pairs within the same department, arranged by submission timestamp, key judgment vectors are extracted. These vectors are then compared component-by-component across each symptom dimension. A change in the value of a component in a symptom dimension generates an opinion change point for that dimension. Each opinion change point records three pieces of information: the judgment direction before the change, the judgment direction after the change, and the time of the change. When multiple symptom dimensions change simultaneously within the same adjacent opinion pair, each dimension generates an independent opinion change point. Simultaneous changes in multiple dimensions indicate that the department acquired new information affecting the judgments of multiple symptom dimensions between two submissions. In a specific diagnostic scenario, a department simultaneously revised the judgments of two symptom dimensions after receiving supplementary examination results. The two opinion change points share the same timestamp, forming a cluster of simultaneous change events in the time-series opinion sequence. This indicates the presence of external information triggering an event that systematically impacts the department's current judgment at that moment. The direction of change at each opinion change point is categorized as follows: a shift from support to opposition is denoted as a negative change; a shift from opposition to support is denoted as a positive change; and a shift from support or opposition to neutrality is denoted as a weakening change. The distribution of these three types of change directions is statistically analyzed according to the symptom dimension. The larger the absolute value of the difference between the number of positive and negative changes for a certain symptom dimension, the more consistent the direction of opinion evolution in that dimension during the scheduling process. A strengthening shift from neutrality to support or opposition is separately marked as a re-judgment change. Re-judgment changes usually correspond to the department forming a clear position after obtaining key evidence. All opinion change points identified in the time-series opinion sequence are summarized in chronological order for later use.

[0040] The revision interval sequence is generated by extracting the revision time intervals of opinion change points. The revision time interval for each opinion change point is quantified by the difference between the time of the change and the time of the last submission by the same department. A longer interval indicates that the department has undergone a longer period of independent thinking or information gathering between the two submissions; a shorter interval indicates that the department triggered the revision within a shorter timeframe. The revision interval sequence is formed by arranging the revision time intervals of each opinion change point by department and time sequence. The mean and standard deviation of the intervals for each department in the revision interval sequence serve as benchmark statistics for the department's revision rhythm. The mean interval reflects the typical revision cycle of the department, and the standard deviation reflects the degree of fluctuation in the revision rhythm. Departments that have submitted fewer than three times have their median of the benchmark mean filled in because the sample size of opinion change points is insufficient, thus avoiding systematic bias caused by low sample statistics in subsequent determination of excessively long intervals. The bimodal distribution of revision interval values ​​within the same department in the revision interval sequence suggests two different types of revision behavior within that department. Short-interval peaks correspond to immediate revisions triggered by information, while long-interval peaks correspond to proactive revisions after independent consideration. In one scenario, a department's revision interval sequence exhibited a clear bimodal distribution: one type of interval concentrated within 30 minutes, corresponding to immediate responses, and the other concentrated over several hours, corresponding to proactive revisions after in-depth consideration. These two types of revisions were assigned different confidence levels in the revision direction record. The department-specific setting of the baseline statistic ensures that the criteria for determining excessively long intervals adaptively adjusts according to differences in departmental revision rhythms, preventing departments with inherently slower revision rhythms from misjudging their normal intervals as excessively long intervals.

[0041] A deep analysis marker is generated to identify deep revision events with extremely long intervals in the revision interval sequence. The revision time interval of each opinion change point in the revision interval sequence is compared with the average interval of the corresponding department. Change points with a revision time interval exceeding twice the average interval of the department are initially marked as candidates for extremely long interval changes. Further verification of the change magnitude of the key judgment vectors corresponding to each candidate for extremely long interval changes is performed. The change magnitude is quantified by the Hamming distance between the key judgment vectors before and after the change. The Hamming distance is used to count the number of dimensions whose judgment values ​​of each symptom dimension have changed before and after the change. Candidates whose Hamming distance exceeds 30% of the total number of symptom dimensions are judged as deep revisions. Candidates that simultaneously meet the criteria of extremely long interval and deep revision are confirmed as extremely long interval deep revision events. The 30% threshold is set as a proportion of the total number of symptom dimensions to ensure that a single dimension change is insufficient to trigger a deep revision judgment. Deep revision requires multiple dimensions to undergo systematic adjustments simultaneously; single-dimensional revisions are local adjustments and are usually triggered by supplementary information for that dimension rather than a deep comprehensive assessment. In a certain clinical setting, a department simultaneously revised its assessments of multiple symptom dimensions after an interval twice the average time. The Hamming distance exceeding the 30% threshold confirmed this as an ultra-long-interval deep revision event, indicating that the department had systematically reassessed multiple dimensions over a prolonged period of independent reflection. Ultra-long-interval deep revision events confirmed in the revision interval sequence are registered by department in the deep analysis markers. Each entry records the corresponding department identifier and the symptom dimensions involved. The more events each department records in the deep analysis markers, the more stable its knowledge-driven revision behavior pattern.

[0042] Based on in-depth analysis markers, confidence-weighted revision direction records are generated by increasing the weight of opinion change points. Opinion change points corresponding to in-depth analysis events in each department are located one by one from the in-depth analysis markers. These located opinion change points are then given confidence-weighted increases. The weighting coefficient is determined by the ratio of the number of in-depth analysis events in that department to the total number of opinion change points. A larger weighting coefficient indicates a higher proportion of in-depth analysis events among the department's change points, and a higher degree of knowledge-driven behavior in the overall revision. The weighting coefficient is multiplied by the base confidence level to obtain the post-weighted confidence level of the opinion change point. This weighting operation gives higher credibility weights to revisions from departments with a high degree of knowledge-driven behavior. For example, if in-depth analysis events account for 60% of a department's total change points, and the weighting coefficient is 1.6, the post-weighted confidence level increases to 0.8 when the base confidence level is 0.5. This department receives a significant weighting in the relevant entries in the revision direction record. For opinion change points not covered by in-depth analysis markers, the confidence level remains unchanged at the base confidence level. The base confidence level is the initial mean of the confidence levels of all change points. The difference between the weighted confidence level and the base confidence level reflects the extent to which the in-depth analysis markers increase the confidence level of that change point. The larger the difference, the more prominent the relative credibility of that change point among all change points. When an in-depth analysis event involves multiple symptom dimensions, opinion change points corresponding to each involved dimension are weighted. The weighting coefficient applies consistently to all involved dimensions and is not adjusted separately due to differences in the magnitude of change in different dimensions, ensuring consistent confidence enhancement operations for change points across all dimensions for the same in-depth analysis event. The weighted confidence levels of all opinion change points, along with the corresponding change direction, department identifier, and symptom dimension label, are arranged chronologically to form a revision direction record. The symptom dimensions with a high concentration of high-confidence change points in the revision direction record correspond to the most active revision directions in the current scheduling after in-depth independent thinking by each department.

[0043] Independent judgment markers are generated by identifying counter-mainstream revision events in the revision direction record. The revised judgment direction at each opinion change point in the revision direction record is compared with the majority tendency of opinions submitted during the same period within the same symptom dimension. The majority tendency is marked by the higher number of supporting and opposing votes for the corresponding symptom dimension during that period. Opinion change points whose revised direction is opposite to the majority tendency are confirmed as counter-mainstream revision events, and the weighted confidence level attached to the counter-mainstream revision event is used as the counter-mainstream strength of the event. The ratio of the number of counter-mainstream revision events in each department to the total number of opinion change points in that department is used as the counter-mainstream revision rate. A high counter-mainstream revision rate indicates that the department's judgment is less influenced by collective tendencies and has strong independence. Departments with a counter-mainstream revision rate exceeding the average of all departments are classified as highly independent departments. The counter-mainstream revision events of highly independent departments in the revision direction record are further verified for consistency between the revised judgment direction and the confirmed judgment for that symptom dimension in the final treatment node chain. Consistent directions are considered valid counter-mainstream revision events, indicating that the department took the lead in revising when the mainstream trend deviated and the revised direction was ultimately adopted. In a certain scenario, most departments tend to oppose a particular symptom dimension, while a certain department, based on its own specialized knowledge, reverses the mainstream view to support it. The final confirmed result aligns with the department's revision direction, and this is determined to be a valid reverse-mainstream revision event. The independent judgment marker is composed of the department to which the valid reverse-mainstream revision event belongs and the corresponding symptom dimension. The strength of the independent judgment marker is indicated by the ratio of the number of valid reverse-mainstream revision events to the total number of reverse-mainstream revision events in that department. Departments with higher strength in the independent judgment marker exhibit more stable accuracy in their reverse-mainstream revisions.

[0044] Based on independent judgment markers, differentiated weights are assigned to the opinions of each department in the initial treatment opinion set to form a confidence weight map. The marker strength of each department in the independent judgment markers serves as the basis for adjusting the base weights. Departments with higher marker strengths have their base weights increased. The base weight is set to 1, and the increase is determined by multiplying the independent judgment marker strength by an adjustment coefficient. The adjustment coefficient is the reciprocal of the mean of the independent judgment marker strengths of all departments, ensuring that the overall distribution center of the weights of each department in the initial treatment opinion set does not systematically rise due to the higher marker strengths of some departments after adjustment. After adjusting the basic weights, the results are further corrected based on the degree of agreement between the opinions of each department in the initial diagnosis and treatment opinion set and the preliminary staging direction judgment in the staging clue identifier. The degree of agreement is quantified by the cosine similarity between the key judgment vector of the department and the standard symptom dimension vector corresponding to the preliminary staging direction judgment. Departments with high degree of agreement have a correction coefficient greater than 1, and departments with low degree of agreement have a correction coefficient less than 1. The final confidence weight is defined as W = W_base × cos_sim, where W_base is the value after adjusting the basic weights (dimensionless), cos_sim is the cosine similarity (dimensionless, ranging from 0 to 1), and W is a dimensionless value. The larger W is, the higher the department's opinion is, the higher its independent judgment accuracy and the higher its agreement with the staging direction. The final confidence weights of each department's opinions are determined under the dual adjustment of independent judgment markers and staging consistency. They are arranged in rows by symptom dimension and columns by department, along with the corresponding department identifiers and symptom dimension labels. The confidence weight map is composed of the confidence weight values ​​of all grid cells. The distribution of differences in the weights of each department within the same symptom dimension in the confidence weight map reflects the gradient structure of the reference priority of opinions in that dimension in the initial diagnosis and treatment opinion set. The distribution pattern of high-weight departments in different symptom dimensions in the confidence weight map reflects the differences in the specialist opinion structure of each dimension.

[0045] Step S13: Use the confidence weight map to analyze the differences in departmental opinions to form a divergence index map, extract divergence patterns from the divergence index map to form a complementary map, and fuse the initial set of treatment opinions with the complementary map to establish a treatment knowledge graph.

[0046] In some embodiments, the step of using the confidence weight map to analyze departmental opinion differences and form a divergence index map includes: extracting weighted vectors of departmental opinions from the confidence weight map to generate an opinion vector set; calculating the departmental professional overlap of the opinion vector set to generate a professional overlap map; identifying pseudo-consensus clustering events in the professional overlap map to generate pseudo-consensus markers; and generating a divergence index map based on the difference intensity weighting of the pseudo-consensus markers and the opinion vector set.

[0047] Opinion vector sets are generated by extracting the weighted opinion vectors of each department from the confidence weight map. The confidence weight value of each department under each symptom dimension in the confidence weight map is multiplied by the corresponding component of the key judgment vector of that department's opinion to obtain the weighted opinion contribution value. The opinion contribution values ​​of each department within each symptom dimension in the confidence weight map, arranged in departmental order, constitute the opinion weighted vector for that dimension. The opinion weighted vectors of each symptom dimension are aggregated across all dimensions to form the opinion vector set. Positive components in the opinion weighted vector correspond to the department's support tendency in that dimension, while negative components correspond to opposition tendency. Dimensions with similar ratios of positive and negative components indicate highly split opinions, while dimensions with significantly dominant positive or negative components indicate concentrated opinion tendencies. In a certain scenario, within a symptom dimension, most departments have high confidence weight values ​​and consistent opinion directions, resulting in a high proportion of unidirectional components in the opinion weighted vector for that dimension, indicating a low risk of disagreement. Conversely, within another symptom dimension, the confidence weights of each department differ significantly, and the judgment directions are dispersed, with similar ratios of positive and negative components, suggesting substantial disagreement in that dimension. In the opinion vector set, the magnitude of the weighted opinion vector for each symptom dimension is quantified by the mean of the absolute values ​​of each component. Dimensions with large magnitudes have a sufficient total contribution of opinions, while dimensions with small magnitudes generally have low confidence weights across departments, resulting in a limited total contribution of opinions. Even if the variance of low-magnitude dimensions is high, the final difference intensity in the divergence index plot after weighting is correspondingly reduced, avoiding the amplification of opinion noise in dimensions with low reference value. Dimensions with significant differences in the absolute values ​​of their components in the opinion vector set correspond to a high concentration of opinion reference priority in a few departments, while dimensions with smaller differences in the absolute values ​​of their components show a relatively balanced contribution of opinions across departments. These two distribution patterns correspond to different judgments of the causes of divergence during the difference intensity weighting stage. Dimensions with high-weight departments concentrated in the confidence weight plot show significantly larger magnitudes in a few components in the opinion vector set.

[0048] A professional overlap map is generated by calculating the departmental overlap of the opinion vector set. The professional overlap between any two departments in each symptom dimension is quantified by the ratio of the smaller to the larger number of opinion submissions involving the same symptom dimension in similar historical cases. The professional overlap values ​​of each department pair in each symptom dimension are arranged in rows by department pair and columns by symptom dimension to form the professional overlap map. Grid cells in the professional overlap map that are more than 1.5 times the average of all department pairs are marked as high overlap items. A dense distribution of high overlap items in the symptom dimension in the professional overlap map indicates that the departmental professional boundaries in that dimension are blurred. If two departments have similar submission frequencies for the same symptom dimension in historical cases, they assume similar diagnostic responsibilities. The convergence of opinions cannot determine whether they come from independent specialist assessments. In a scenario where two departments with similar responsibilities have highly balanced historical submission frequencies across multiple symptom dimensions, and multiple dimensions in the professional overlap map simultaneously show high overlap markers, the risk of multi-dimensional pseudo-consensus needs to be checked during the pseudo-consensus aggregation event identification phase. When a pair of departments within the same department exhibits high overlap across multiple symptom dimensions in a professional overlap diagram, it indicates that the overlap in their professional responsibilities is not limited to a single dimension. Their functional positioning within the treatment pathway exhibits systematic overlap. A single-dimensional pseudo-consensus risk assessment is insufficient to reflect the true scope of the pseudo-consensus threat between them; a holistic assessment of cross-dimensional pseudo-consensus risk is necessary, on a departmental basis. Departmental pairs with lower mean overlap values ​​in the professional overlap diagram are more reliable when opinions converge, and their consistency is more likely to stem from genuine verification after independent evaluation. The overall distribution pattern of the professional overlap diagram reflects the clarity of the departmental professional structure in the current treatment scenario. Dimensions with a high proportion of homogeneous components in the opinion vector concentration but low professional overlap between corresponding departmental pairs have the highest reliability.

[0049] This study aims to identify pseudo-consensus clustering events and generate pseudo-consensus markers for professional overlap graphs. For department pairs corresponding to high overlap entries in the professional overlap graph, their respective weighted opinion vectors are extracted from the opinion vector set. A candidate for pseudo-consensus is defined as one whose cosine similarity and professional overlap both exceed 1.5 times the mean of all department pairs. Pseudo-consensus markers are used for dimensions where the number of pseudo-consensus candidates exceeds 30% of the total number of department pairs in that dimension. Each pseudo-consensus marker entry records the corresponding dimension, participating department pairs, and the strength of the pseudo-consensus. In the professional overlap graph, department pairs with high overlap but significantly different opinion directions do not constitute pseudo-consensus; their responsibilities overlap but they still generate independent disagreements, reflecting potential for professional complementarity. In the opinion vector set, department pairs with similar opinion directions but low professional overlap represent independently verified true consistency, providing effective support for the credibility of current symptom dimension judgments. The strength of pseudo-consensus is calculated as F = N_pseudo / N_total, where N_pseudo is the number of pseudo-consensus candidate department pairs in this dimension (dimensionless), N_total is the total number of department pairs in this dimension (dimensionless), and F is dimensionless. A larger F indicates that the degree of opinion concentration in this dimension is amplified by pseudo-consensus. In a certain scenario, three out of six department pairs within a certain symptom dimension constitute pseudo-consensus candidates, with an F value of 0.5. This suggests that the convergence of opinions in more than half of the department pairs in this dimension stems from professional overlap rather than independent verification, and the true divergence is severely underestimated. The distribution of F values ​​for each item in the pseudo-consensus marker reflects the differences in pseudo-consensus penetration across different symptom dimensions. Several dimensions with a high concentration of F values ​​in the pseudo-consensus marker will exhibit a systematically low difference intensity in the divergence index diagram if not corrected.

[0050] A divergence index map is generated based on a weighted average of the difference intensity using pseudo-consensus markers and opinion vector sets. In the pseudo-consensus markers, the weighted opinion vectors of the pseudo-consensus departments involved in each entry for the corresponding symptom dimension are merged into a single representative vector. After merging, the number of valid opinion sources within that dimension decreases, with each pseudo-consensus department counted as only one valid opinion source. The variance of all valid vectors within each symptom dimension of the opinion vector set is calculated as the squared mean of the differences between each valid vector and the mean of the valid vectors for that dimension, serving as the original difference intensity. The original difference intensity is further weighted by the amplitude values ​​of the corresponding dimensions in the opinion vector set. The weighting coefficient is the ratio of the amplitude value of that dimension to the mean amplitude of all dimensions. Dimensions with high F-values ​​in the pseudo-consensus markers are subject to a correction coefficient of 1-F / 2 before weighting, further reducing the representativeness of the valid vectors in dimensions with high pseudo-consensus penetration. In a certain scenario, the F-value for a certain symptom dimension is 0.6, and the correction coefficient is 0.7. The representativeness of the valid vectors in this dimension in the difference intensity calculation is reduced to 70% of the original, ensuring that the false suppression of difference intensity by the pseudo-consensus is fully compensated. The correction coefficient is designed as 1-F / 2 instead of 1-F because full reduction would cause the effective vector representativeness of dimensions with F values ​​close to 1 to approach zero, and the calculation of difference intensity would lose its statistical basis. Half reduction strikes a balance between eliminating the influence of spurious consensus and retaining limited statistical representativeness. Covered and uncovered dimensions marked by spurious consensus are weighted using different paths. The difference intensity of uncovered dimensions is calculated directly using the amplitude weighting coefficient without applying F-value correction. Each dimension entry in the divergence index diagram stores the corrected difference intensity value for the corresponding symptom dimension. The true degree of divergence for dimensions with high difference intensity in the divergence index diagram is fully presented after eliminating the interference of spurious consensus. The difference intensity of some dimensions in the divergence index diagram increases significantly after correction compared to before correction; the larger the increase, the more severely the original concentration of opinions in that dimension is amplified by spurious consensus.

[0051] In some embodiments, the step of extracting divergence patterns from the divergence index map to form a complementary graph includes: tracking departmental opinion dominance in the divergence index map to generate a dominance sequence; identifying cross-departmental dominance rotation events for the dominance sequence to generate a rotation node set; performing disease association analysis on the rotation node set to identify complementary boundary diseases; and constructing departmental knowledge complementarity relationships based on the complementary boundary diseases to form a complementary graph.

[0052] The departmental opinion dominance is tracked in the divergence index chart to generate a dominance sequence. Within the high-divergence dimension of the divergence index chart, the weighted vectors of departmental opinions are arranged from largest to smallest amplitude. The department with the largest amplitude is identified as the dominant department for that dimension at that moment. Each time a department submits a new opinion, the amplitude of the weighted vectors of each department's opinion is recalculated, and the department with the largest amplitude is updated as the current dominant department. The continuous update nodes of the dominant departments for each symptom dimension of the divergence index chart are arranged along a timeline to form a dominance time sequence record. Each dimension entry in the dominance sequence records the complete switching time sequence of the dominant department for the corresponding symptom dimension. The duration of continuous dominance by the leading department is quantified by the difference between two adjacent switching timestamps. Dimensions with a large standard deviation of dominance duration exhibit both short-term and long-term dominance. Short-term dominance corresponds to a department briefly achieving the highest weighted magnitude after acquiring new evidence, only to be overtaken again. Long-term dominance corresponds to a department consistently maintaining the highest weighted magnitude over a longer period. In a certain scenario, a symptom dimension frequently experiences switching of the leading department during the scheduling process. A department briefly rises to the leading position after acquiring supplementary test results, and then another department regains dominance based on more sufficient historical case support. This alternating pattern of short-term and long-term dominance forms a period of dense switching in the dominance sequence, suggesting that there is a stable departmental knowledge complementarity structure in this dimension. The mean duration of continuous dominance by the same department in the dominance sequence reflects the stability of the department's judgment in the corresponding symptom dimension. Departments with a long mean duration of continuous dominance indicate that their judgment authority in this dimension is continuously recognized. Departments with frequent short-term dominance indicate that their opinion influence is repeatedly surpassed by other departments as new evidence is introduced. The difference in dominance stability between the two types of departments corresponds to different complementary edge weight distributions in the complementary graph construction. In the divergence index graph, the dominance sequence switching density in the high difference intensity dimension is significantly higher than that in the low difference intensity dimension.

[0053] For example, the step of identifying cross-departmental leadership rotation events and generating a rotation node set for the leadership sequence includes: performing opinion fusion and trace frequency statistics on the leadership sequence to generate a latent influence baseline; identifying silent leadership events based on the latent influence baseline to generate a latent leadership event set; performing explicit and implicit leadership switching pattern analysis on the latent leadership event set to identify explicit and implicit leadership rotation pairs; and locating rotation time sequence nodes based on the explicit and implicit leadership rotation pairs to generate a rotation node set.

[0054] The implicit influence baseline is generated by statistically analyzing the frequency of opinion fusion and record keeping in the dominance sequence. The number of times each department exhibits explicit dominance in each symptom dimension within the dominance sequence is counted based on the number of time nodes in the dominance sequence record where that department is marked as the dominant department. After each department submits its opinion, the cosine similarity between the key judgment vector of subsequent departments and the key judgment vector of the original department is calculated pairwise. The proportion of subsequent submissions with a cosine similarity higher than 0.8 is used as the implicit influence rate of that department. The mean implicit influence rate of all departments within each symptom dimension of the dominance sequence is used to obtain the implicit influence baseline for that dimension. The implicit influence baseline set stores the mean baseline of the implicit influence rate of all departments within the corresponding symptom dimension for each dimension's entries. Departments with high implicit influence rates and low explicit dominance frequency indicate that the transmission power of their opinions exceeds the explicit influence reflected by their weights. In a certain scenario, a department with low confidence weights failed to gain explicit dominance, but after submitting its opinion, the key judgment vectors of multiple subsequent departments were highly similar to its own, and the implicit influence rate was significantly higher than the baseline for that dimension. This indicates that the department's opinion actually dominated the subsequent judgment direction in that dimension, and its influence was manifested in the form of implicit penetration rather than explicit weights. The implicit influence rate threshold is set at 0.8 rather than a higher value because judgment vectors with a cosine similarity of 0.8 or higher are basically consistent in the direction of the support and opposition components, indicating that the judgment positions of subsequent departments are highly similar to that department. Similarity below 0.8 corresponds to a significant deviation in the judgment direction and cannot be considered a manifestation of opinion transmission effect. Dimensions with high baseline values ​​in the implicit influence baseline set correspond to a generally strong opinion transmission effect in that dimension. The opinions submitted by most departments can have a high penetration of the judgment direction of subsequent departments. Identifying silent dominant departments in such dimensions is relatively difficult.

[0055] To identify silent dominant events based on the baseline of latent influence, a set of latent dominant events is generated. The baseline values ​​for each symptom dimension in the latent influence baseline set are compared departmentally with the latent influence rate for each department. Departments with a latent influence rate exceeding 1.5 times the baseline for that dimension but with a visible dominant event frequency lower than the average for all departments are identified as silent dominant departments. Each high latent influence rate record at the corresponding symptom dimension and time point in a silent dominant department is identified as a silent dominant event. The set of latent dominant events is organized using a dual index of department and symptom dimension, including all confirmed silent dominant events. Departments with only a high latent influence rate but also a high visible dominant event frequency are considered to have both latent and visible dominance, but are not considered silent dominant. Both conditions are indispensable to ensure accurate identification of silent dominant departments, distinguishing between genuine latent penetration and normal opinion transmission in visible dominant departments. Implicitly dominant events are clustered together as silent dominant events occurring consecutively within the same department on the same symptom dimension. A larger number of events within a cluster indicates a longer duration of implicit influence by the department on that dimension. In a certain scenario, a department's implicit influence rate on multiple consecutive scheduling nodes on a certain symptom dimension exceeded the baseline by 1.5 times without gaining explicit dominance. During this period, consecutive event clusters formed within the implicitly dominant event set, indicating that the department continuously influenced the judgment direction of subsequent departments through high similarity in opinion content. The contrast between the formation of these event clusters and the department's relatively low confidence weight suggests that the confidence weight map underestimates the department's influence on opinions; its actual judgment ability has formed a stable implicit transmission path in historical scheduling.

[0056] A pattern analysis of the switching between explicit and implicit dominance is performed on the latent dominance event set to identify pairs of explicit and implicit dominance rotation. Each silent dominant department in the latent dominance event set is paired with the corresponding explicit dominant department in the symptom dimension. Periods in the latent dominance event set where both types of dominance appear at least once on the timeline and alternate are defined as periods of explicit-implicit dominance switching. Pairs of explicit-implicit dominance rotation are confirmed by candidate pairs that alternate at least twice within the switching period. Unidirectional explicit suppression or implicit penetration does not constitute a switch. The switching response time τ for the two types of dominance in the explicit-implicit dominance rotation pair is τ = t_hidden - t_explicit, where t_hidden is the time of occurrence of the silent dominant event (in time units), t_explicit is the time of the corresponding explicit dominance switch (in time units), and τ is a unit of time. A smaller τ indicates that implicit dominance penetration occurs faster after explicit dominance switching; a negative τ indicates that implicit dominance penetration occurs before explicit dominance switching. In a certain scenario, after a shift in explicit dominance, the implicit influence rate of a department rebounded to more than 1.5 times the baseline within a very short time, with τ approaching zero. This indicates that the switching between the two is extremely tightly coupled, and the switching of explicit and implicit dominance pairs occurs almost synchronously, suggesting that the department's opinion transmission channels in this dimension are highly solidified. A negative τ value in actual scheduling means that the silent dominance department, while still retaining explicit dominance, has already begun to exert influence on subsequent departments beyond its expected weighted magnitude. This situation indicates that the department had proactively influenced subsequent judgments through the arrangement of opinion content before losing explicit dominance. In this case, the explicit and implicit dominance pairs exhibit predictive characteristics. Departments belonging to event clusters with significant predictive characteristics in implicit dominance events correspond to higher complementary edge weights in complementary graph construction.

[0057] Based on the dominance / latent dominance rotation pairs, a rotation node set is generated by locating the rotation time sequence nodes. The dominance switching time and latent dominance switching time of each pair in the dominance / latent dominance rotation pair are extracted separately. These two types of switching times alternate on the time axis, forming a time sequence switching chain for the rotation pair. The rotation node set is composed of switching time as time sequence nodes, rotation pair department identifier as node attribute, corresponding symptom dimension, and τ-mean as node label. In the dominance / latent dominance rotation pairs, rotation pairs with a τ-mean close to zero correspond to almost synchronous switching of dominance and latent dominance. Rotation pairs with a larger τ-mean correspond to relatively loose switching coupling, with a time lag in the alternation of dominance. The difference in τ-mean is reflected in the rotation node set in the form of node labels, enabling the rotation node set to distinguish the essential differences in the opinion influence structure between tightly coupled and loosely coupled rotation pairs. The period of concentrated node distribution in the rotation node cluster corresponds to the synchronous switching of multiple explicit and implicit dominance pairs. In a certain scenario, multiple symptom dimensions of the scheduling process simultaneously experience explicit and implicit dominance switching within a certain period. The rotation node set forms a node cluster during this period, indicating that a newly submitted opinion or external information simultaneously triggers the reorganization of the multi-dimensional opinion dominance pattern. The greater the temporal overlap between the period of concentrated node clustering in the rotation node cluster and the period of dense switching in the dominance sequence, the closer the linkage between explicit and implicit dominance switching, and the more systematic the reorganization of the multi-dimensional opinion dominance pattern during this period.

[0058] Disease association analysis was performed on the rotation node set to identify complementary boundary diseases. The symptom dimensions involved in each rotation event in the rotation node set corresponded to the disease type in the current patient's medical record on an event-by-event basis. A set of rotation events occurring at multiple different time points within the same rotation department, with the corresponding disease type being consistent, was identified as a disease-related rotation group. Complementary boundary diseases were identified when the number of rotation events exceeded 30% of the total number of rotation events for that department. Disease-related rotation groups indicate that the department has a stable disease-driven tendency for the alternation of dominance under that disease type. In a certain scenario, the dominance rotation events of two departments under a certain spinal degeneration-related disease accounted for 40% of their total rotation events, exceeding the 30% threshold, and were therefore identified as complementary boundary diseases. This indicates that the knowledge coverage boundaries between the two departments in the diagnosis and treatment of this disease type are the most ambiguous, and their complementary needs are the strongest. The 30% threshold ensures that complementary boundary diseases are statistically representative. Disease types with a low percentage of rotation events below this threshold indicate that these disease types are not the primary driver of the alternation of dominance between the two. Including them in the complementary boundary diseases would introduce interference from weakly correlated diseases into the complementary graph. The complementary boundary associations involving multiple departmental pairs within the same disease type in the rotation node set indicate that this disease type has multiple sets of overlapping departmental knowledge boundaries. The more departmental pairs involved, the broader the multi-departmental collaborative knowledge coverage required for this disease type. The complementary edges between departmental pairs form dense local subgraphs in the complementary graph. Higher subgraph density indicates a deeper level of collaborative diagnosis and treatment complexity for this disease type. The disease type corresponding to the period with the highest node density in the rotation node set usually highly overlaps with the complementary boundary diseases.

[0059] A complementary knowledge graph is constructed based on complementary boundary diseases, establishing complementary relationships between departments. The complementary relationships between departments within complementary boundary diseases are represented by complementary edges. The weight of each complementary edge is quantified by the number of rotation events within the disease-related rotation group of the corresponding complementary boundary disease for that department. The complementary graph is an undirected graph with departments as nodes and complementary edges as edges, and each edge in the graph is labeled with its corresponding complementary boundary disease. Complementary edges for the same department under multiple complementary boundary diseases are retained independently, with each edge weight quantified separately and not merged into a single edge. When the edge weights for a department under two complementary boundary diseases differ significantly, it indicates a significant difference in the strength of the complementary relationship between the two under different disease types. Disease types with higher weights correspond to the diagnostic and treatment scenarios where the dominance of the two changes most frequently and the need for complementarity is most urgent. In the complementary graph, departments with high node degrees have complementary relationships with other departments across multiple symptom dimensions, exhibiting broad knowledge boundary coverage. Departments with low node degrees, on the other hand, have complementary needs concentrated in a few symptom dimensions. The two types of departments participate differently in the opinion fusion stage. Departments with high node degrees are jointly activated by multiple departments across multiple dimensions during collaborative fusion, while departments with low node degrees are only activated in collaborative fusion of specific dimensions. Disease labeling allows the complementary graph to filter applicable complementary relationships based on the current patient's primary disease type. In a scenario where the current patient's primary disease is a certain type of spinal degeneration, the complementary graph only activates complementary edges associated with that disease type for collaborative fusion. Edges labeled with other diseases do not participate in this fusion, preventing complementary relationships of unrelated diseases from interfering with the current patient's opinion fusion results.

[0060] A diagnostic knowledge graph is established by fusing opinions from an initial set of treatment opinions with a complementary graph. The key judgment vectors of each department's opinions in the initial set of treatment opinions are aligned with the complementary edge weights of each department in the complementary graph along the symptom dimension. For departments with high complementary edge weights in the complementary graph, a collaborative fusion strategy is adopted for the corresponding symptom dimension, with the fusion result being a weighted average of the weighted vectors of the two departments' opinions. The weighting coefficients are the proportions of each department's complementary edge weight to the sum of their respective weights. For departments with no complementary edges or edge weights below the mean in the initial set of treatment opinions, an independent fusion strategy is adopted, directly averaging the confidence weights of each department in the confidence weight graph without introducing complementary relationship adjustment. The collaborative fusion strategy eliminates the specialist blind spots that might be introduced by a single department's dominance within the complementary boundary dimension. In a scenario where two departments have balanced complementary edge weights in a certain symptom dimension, the fusion judgment vector in that dimension after collaborative fusion simultaneously incorporates the judgment basis from both departments' respective areas of expertise. The difference in contribution of each department's opinions in that dimension from the initial set of treatment opinions to the final fusion judgment vector after collaborative fusion is significantly smaller than that of the independent fusion strategy. The diagnostic knowledge graph uses symptom dimensions as nodes and establishes directed edges based on the condition that both dimensions' fusion judgments are supportive and that there are clinically dependent edges in the diagnostic node chain. The direction of the directed edges is consistent with the direction of clinical dependence in the diagnostic node chain. The weight of the directed edges is quantified by the cosine similarity of the fusion judgment vectors of the two dimensions. The higher the cosine similarity, the more consistent the fusion judgment direction of the two dimensions is in the current patient, and the closer the clinical correlation. The product of the in-degree and out-degree of each node in the diagnostic knowledge graph is used as the graph centrality of that dimension. The diagnostic knowledge graph is composed of all symptom dimension nodes, directed edges, edge weights, and graph centrality. Dimensions with high graph centrality are supported by multiple dimensions and provide diagnostic basis for multiple dimensions, occupying key node positions in the current patient's diagnostic path.

[0061] Step S14: Identify conflicting and unbalanced items for the cumulative bias in the initial set of treatment opinions, extract the conflict intensity of conflicting items to form blind zone markers, perform specialty spectrum matching on unbalanced items and divergence index diagram to determine equilibrium markers, and perform opinion priority rearrangement to generate allocation instructions according to blind zone markers and equilibrium markers.

[0062] Specifically, the cumulative bias in the initial treatment opinion set is used to identify conflicting and imbalanced items. Key judgment vectors for each department's opinions in the initial treatment opinion set are extracted dimension-by-dimensionally by symptom. Conflicts are identified when the judgment vector components from different departments within the same symptom dimension exhibit opposite directions and significant amplitudes. Conflicts are quantified by the absolute value of the difference between the mean amplitude of the supporting tendency component and the mean amplitude of the opposing tendency component within that dimension. Symptom dimensions with an absolute value exceeding 50% of the mean amplitude of all components in that dimension are identified as conflicting items, corresponding to substantial opposition between specialist opinions in that dimension within the current treatment scenario. Imbalances are identified using a confidence weight map. When the variance of the confidence weights of each department within a certain symptom dimension exceeds 1.5 times the mean variance of the weights of all symptom dimensions in the confidence weight map, the corresponding dimension is identified as a dimension with an imbalanced weight distribution. Within a dimension with an imbalanced weight distribution, the opinion contribution of high-weight departments far exceeds that of low-weight departments, and the overall opinion result is dominated by a few high-weight departments. The judgment criteria of low-weight departments are not fully reflected in the opinion fusion. All dimensions with an imbalanced weight distribution are included in the imbalanced item set. The identification scopes of conflict items and imbalance items do not overlap. Conflict items are identified based on opposing judgment directions, while imbalance items are identified based on skewed weight distribution. When a certain symptom dimension in the initial treatment opinion set meets both conditions, it is included in the conflict item set and the imbalance item set respectively. The two sets are processed independently and each carries the corresponding symptom dimension label. In a certain scenario, most departments tend to support a certain symptom dimension, but a certain high-weight department strongly opposes it. In the initial treatment opinion set, this dimension meets both the opposing direction and weight skew conditions. The conflict item records the intensity of its opposing direction, while the imbalance item records the weight deviation of the high-weight department. The two types of deviations are handled differently and are not merged into a single deviation type.

[0063] In some embodiments, the step of extracting conflict intensity from the conflict item to form a blind zone marker includes: extracting multidimensional conflict features from the conflict item to generate a conflict intensity distribution; identifying persistent low-conflict intervals based on the conflict intensity distribution to generate a zero-conflict candidate set; performing inertial consensus discrimination on the zero-conflict candidate set to generate a latent blind zone interval; and jointly locating the latent blind zone interval and the conflict intensity distribution to form a blind zone marker.

[0064] Multidimensional conflict feature extraction is performed on conflict items to generate a conflict intensity distribution. The conflict intensity of each symptom dimension in the conflict item set is quantified by the absolute value of the difference between the mean amplitude of the support tendency component and the mean amplitude of the opposition tendency component. The conflict intensity of each dimension of the conflict item is calculated sequentially along the time axis according to the scheduling nodes. After each scheduling node submits a new opinion, the conflict intensity at that node is recalculated, resulting in a time series sequence of the conflict intensity of that symptom dimension. The row structure of the conflict intensity distribution corresponds to each symptom dimension in the conflict item set, and the content of each row is the time series sequence of the conflict intensity of that dimension as the scheduling nodes progress. The peak intensity time of each symptom dimension in the conflict intensity distribution corresponds to the scheduling node where the specialist opinion opposition of that dimension is most intense. The intensity slope before and after the peak time reflects the formation and dissipation speed of opinion opposition. Dimensions with steep slopes correspond to opinions that are rapidly intensified and rapidly dissipated in a short period of time, while dimensions with gentle slopes correspond to opinions that are slowly accumulated or slowly converged. In the distribution of conflict intensity, the intensity sequence of the same symptom dimension exhibits different distribution patterns at different scheduling stages. In the early scheduling stage, opinions from various departments have not been fully exchanged, and the conflict intensity is generally low. In the mid-term scheduling stage, differences in opinions from various departments gradually emerge, and the conflict intensity tends to increase. In the late scheduling stage, opinions from various departments have been revised multiple times, and some dimensions tend to converge, and the conflict intensity declines. The intensity patterns of the three stages are reflected in the conflict intensity distribution in a time-series segmented manner. When identifying the continuous low conflict interval, it is necessary to distinguish between the low conflict generated by normal convergence in the later stage and the low conflict generated by inertial suppression in the middle stage. The low conflict of normal convergence appears after the intensity peak and the rate of decline is stable. The low conflict of inertial suppression appears before the intensity peak and is low without a significant climbing process. The historical intensity mean of each dimension in the conflict item is used as a reference benchmark to judge whether the current intensity is abnormally low.

[0065] To identify persistently low-conflict intervals in the conflict intensity distribution, a zero-conflict candidate set is generated. The intensity sequence of each symptom dimension in the conflict intensity distribution is compared node-by-node with a threshold of 20% of the global mean for that dimension. Consecutive node segments consistently below the threshold are initially marked as low-conflict candidate segments. The minimum length of a low-conflict candidate segment is bounded by three consecutive scheduling nodes; candidate segments shorter than this bound are removed. The low-conflict candidate segments are further verified for their temporal position within the conflict intensity distribution. Candidate segments located after the intensity peak and whose preceding three nodes have an intensity mean higher than the global mean are considered normally converged segments. Candidate segments located before the intensity peak or exhibiting a low position without a significant intensity increase are considered zero-conflict candidate segments. The zero-conflict candidate set uses the symptom dimension and temporal position as positioning coordinates, including all low-conflict candidate segments determined to be abnormally converged. The distinction between normal convergence segments and zero-conflict candidate segments is primarily based on the temporal position before and after the intensity peak, supplemented by the preceding intensity ramp slope as an auxiliary criterion. Preceding segments with ramp slopes lower than the global average slope correspond to opinions that have not been sufficiently intensified before a sudden drop to a low level. In a certain scenario, a certain symptom dimension exhibits consecutive low-conflict segments during the scheduling phase. The preceding node intensity of this segment does not show a significant peak and the ramp slope is close to zero, which is inconsistent with the high conflict intensity pattern commonly observed in similar historical cases. This segment is included in the zero-conflict candidate set and marked with a deviation from historical patterns. Candidate segments with higher deviations in the zero-conflict candidate set are prioritized during the inertial consensus identification phase.

[0066] The zero-conflict candidate set is used to generate hidden blind zone intervals through inertial consensus identification. The opinion submission records of all departments within each candidate segment of the zero-conflict candidate set are verified node by node. Candidate segments with a non-zero number of opinion submissions but with highly convergent key judgment vector directions across departments are identified as inertial consensus candidate segments. Convergence is determined by a cosine similarity exceeding 0.9 for the average cosine similarity of key judgment vectors across all departments within the candidate segment. A cosine similarity exceeding 0.9 indicates that the judgment positions of all departments are almost completely consistent during that period, and normal specialist disagreements disappear during that period. The inertial consensus candidate segments in the zero-conflict candidate set are further compared with the frequency of opinion disagreements in the same symptom dimension in historical similar cases. The historical disagreement frequency is quantified by the proportion of conflicting items in the same dimension in similar cases. Symptom dimensions with a historical disagreement frequency higher than 50% are considered to deviate from historical patterns when inertial consensus candidate segments appear in the current scheduling. These deviating candidate segments are confirmed as hidden blind zone intervals. Symptom dimensions with a historical disagreement frequency lower than 50% do not trigger hidden blind zone determination even if highly convergent opinions appear, because this dimension has historically been predominantly highly consistent, and the current convergence conforms to historical patterns. The start and end boundaries of the hidden blind zone intervals are determined by the timestamps of the first and last nodes of the corresponding inertial consensus candidate segments. Opinions within the hidden blind zone intervals are highly convergent but do not conform to the historical pattern of divergence, suggesting that departmental opinions during this period are suppressed by the inertia of scheduling sequences or previous high-weight opinions, and the actual differences in specialist judgments are not reflected in the opinions. The set of hidden blind zone intervals is composed of all confirmed hidden blind zone intervals arranged according to symptom dimensions and time sequence.

[0067] Blind zone markers are formed by jointly locating hidden blind zone intervals and conflict intensity distributions. The temporal position of each entry in the hidden blind zone interval set is extracted one by one from the mean intensity of the conflict intensity distribution. Entries with a mean intensity lower than 20% of the global mean of the corresponding symptom dimension and a duration exceeding three scheduling nodes are identified as high-risk blind zone entries. Both conditions must be met simultaneously to ensure that blind zone markers are limited to periods where both intensity suppression and deviation from historical patterns are simultaneously established. Entries meeting only one condition are downgraded to blind zone entries to be observed and are not included in the blind zone markers for the time being. In high-risk blind zone entries, the duration of inertial consensus is quantified by the width of the corresponding hidden blind zone interval. The intensity of suppression is quantified by the absolute value of the difference between the average conflict intensity within the interval and the global average conflict intensity distribution. The product of duration and suppression depth is the comprehensive risk value of the blind zone for that entry. A higher comprehensive risk value indicates a deeper degree of concealment of genuine specialist disagreements during that period. In a certain scenario, a hidden blind zone interval for a certain symptom dimension lasted for eight scheduling nodes, and the average conflict intensity within the interval was only 12% of the global average. The comprehensive risk value was the highest among all entries, indicating that genuine opinion opposition was highly suppressed during that period, and the apparent convergence of opinions deviated significantly from the true judgment pattern. Blind zone markers are composed of the comprehensive risk value, symptom dimension labeling, and temporal position of all high-risk blind zone entries. Entries with a comprehensive risk value exceeding 1.5 times the average of all entries are marked as key blind zones. Key blind zones are subject to a greater reduction in priority for departments with inertial consensus during opinion priority reshuffling, ensuring that suppressed genuine specialist disagreements are given priority release in allocation instructions. The key blind spot entries in the blind spot marking also include a list of relevant departments. The key judgment vectors of each department in the list during that period are recorded as suppression state vectors.

[0068] In some embodiments, the step of performing specialty lineage matching between the imbalanced item and the divergence index map to determine the equilibrium marker includes: extracting the weight deviation of each department from the imbalanced item to generate a weight deviation sequence; performing association mapping between the weight deviation sequence and the divergence index map to identify the imbalanced disease type; extracting the opinion adoption rate of the currently low-weight department for the imbalanced disease type to generate an adoption rate ranking table; and performing weight reversal compensation on the imbalanced disease type based on the adoption rate ranking table to determine the equilibrium marker.

[0069] The weight deviations of each department are extracted from the imbalanced items to generate a weight deviation sequence. Confidence weight maps for each symptom dimension in the imbalanced item set are extracted. The weight deviation of each department in that dimension is obtained by subtracting its confidence weight from the mean confidence weight of all departments in that dimension. The weight deviations of the symptom dimensions involved in the imbalanced items are summarized by department and written into the weight deviation sequence. A positive weight deviation indicates that the department's confidence weight in that dimension is higher than the mean, while a negative weight deviation indicates that it is lower than the mean. The weight deviations of each department are recorded item by item in a two-dimensional table consisting of the symptom dimensions and the departments; this table is the weight deviation sequence. The column-directed slices of the weighted bias sequence reflect the dispersion of weight distribution among departments within a symptom dimension. Symptom dimensions with large slice variances correspond to weights highly concentrated in a few departments, with the opinions of other departments contributing almost nothing in the fusion calculation. In a certain scenario, the confidence weight of a certain department within a symptom dimension far exceeds that of all other departments. The column-directed slice for this dimension exhibits a unimodal skewed distribution, with the biases of other departments all being negative and having large absolute values. This indicates that the opinion fusion results for this dimension are almost dominated by a single department, and the judgment criteria of low-weight departments are systematically ignored in this dimension. When the same department in the weighted bias sequence shows large negative biases in multiple symptom dimensions, it indicates that this department is in a low-weight state in multiple dimensions of the treatment pathway. Its specialist judgment is continuously suppressed across dimensions. The absolute value of the negative biases of such departments is generally high, and the scope of weight underestimation is wider than that of departments with biases only in a single dimension. The two types of situations correspond to different compensation priorities in the weight reversal compensation stage, with the comprehensive compensation need for departments with multi-dimensional underestimation being more urgent.

[0070] The association mapping between the weighted deviation sequence and the divergence index plot was used to identify imbalanced disease types. The mean positive deviation of each symptom dimension in the weighted deviation sequence was compared dimension-by-dimensionally with the intensity of difference for the corresponding symptom dimension in the divergence index plot. Symptom dimensions with both a mean positive deviation and an intensity of difference exceeding the mean of all dimensions were identified as imbalanced disease types. The fact that both exceed the mean indicates that this dimension exhibits both a severe concentration of weights in high-weight departments and genuine disagreement, with opinions from low-weight departments being systematically suppressed in high-disagreement dimensions. The mean positive deviation of all dimensions in the weighted deviation sequence was 0.18, and the mean intensity of difference in the divergence index plot was 0.35. A certain symptom dimension had a mean positive deviation of 0.26 and an intensity of difference of 0.48, both exceeding their respective mean values, confirming this dimension as an imbalanced disease type. The identification scope of imbalanced disease types is constrained by a dual threshold of weighted deviation sequence and divergence index plot. Only symptom dimensions with high weighted deviation but low divergence intensity indicate that the monopoly of high-weighted departments is based on convergent opinions across departments, and the concentration of weights has not led to the suppression of divergence. Only symptom dimensions with high divergence intensity but low weighted deviation indicate that the weight distribution across departments is balanced, and opinions can be expressed under a relatively equal weight structure. Neither of these single-dimensional exceedances constitutes an imbalanced disease type. Each item in the imbalanced disease type set corresponds to a confirmed imbalanced symptom dimension. Items are arranged in descending order of difference intensity, and each item includes two reference data points: the mean of positive deviation and the difference intensity value. The imbalanced disease type with the highest difference intensity corresponds to the symptom direction with the most severe weight suppression and the most significant divergence in current diagnosis and treatment.

[0071] An adoption rate ranking table is generated by extracting the opinion adoption rates of currently low-weight departments for each imbalanced disease type. Departments with negative weight deviations within the symptom dimension corresponding to each imbalanced disease type in the set of imbalanced disease types are identified as currently low-weight departments. The opinion adoption rate for each low-weight department is calculated based on historical similar cases. Historical similar cases are matched with past cases that share the same primary disease type as the current patient and involve the same symptom dimension of the imbalanced disease type. This matching condition ensures that the adoption rate statistics are derived from real and relevant historical treatment scenarios. Cross-disease adoption rate statistics may introduce historical judgment errors unrelated to the current treatment. The adoption rate is the proportion of times each low-weight department submits opinions in matched historical cases and the final treatment conclusion is consistent with its opinion, out of the total number of submissions by that department in the matched cases. The adoption rate statistics are limited to opinions within the symptom dimension corresponding to the imbalanced disease type; opinion adoption records from other dimensions are not included in the statistics, ensuring that the adoption rate only reflects the department's specialist judgment accuracy in the imbalance dimension. The adoption rate ranking table is grouped by the type of imbalanced disease. Each department within a group includes the adoption rate, the corresponding symptom dimension, and the number of historical matching cases. The entries are arranged in descending order of adoption rate. Departments with fewer than five historical matching cases are marked with low confidence in the adoption rate ranking table. The compensation for departments with low confidence is appropriately reduced during the weight reversal compensation phase. The distribution density of low confidence marked entries in the adoption rate ranking table reflects the sufficiency of historical data accumulation for the current imbalanced disease type.

[0072] Based on the adoption rate ranking table, a weight reversal compensation is applied to imbalanced disease types to determine the equilibrium marker. Departments whose adoption rate exceeds the average adoption rate of all low-weight departments in the adoption rate ranking table are identified as underweight departments. The compensation magnitude for underweight departments is determined by multiplying the absolute value of the negative deviation in the weight deviation sequence of that department with its adoption rate. The larger the product, the higher the compensation magnitude. In a certain scenario, a department has a weight deviation of -0.4 and an adoption rate of 0.8 in the symptom dimension corresponding to a certain imbalanced disease type. The product of these two is 0.32. This department receives a large compensation magnitude in the equilibrium marker, and the confidence weight in this dimension converges to the mean by 0.32 units after compensation. The upper bound of compensation is set to the mean of the confidence weights of all departments in this dimension. After compensation, the confidence weight of underweight departments does not exceed the current mean, ensuring that the compensation operation does not cause underweight departments to jump to become new high-weight monopolistic departments in this dimension, and maintaining the relative balance of the weight distribution after compensation. The compensation for departments with low confidence in the adoption rate ranking table is calculated by multiplying the original calculated value by the ratio of the number of historical matching cases to the lower bound of five cases, which serves as a reduction factor. Departments with three cases have a reduction factor of 0.6, reducing the compensation to 60% of the original value. This reduction mechanism ensures that departments with low statistical confidence receive conservative compensation, preventing excessive weight restoration due to a small number of high adoption rate cases. The balance label is composed of the compensation for all departments with underestimated weights, the corresponding imbalanced disease type symptom dimension label, and the ranking in the adoption rate ranking table. Departments with large compensation in the balance label receive a more significant priority increase in the opinion priority re-ranking. Departments with lower adoption rate rankings but above the mean receive conservative compensation. Imbalanced disease types without underestimated weights do not generate balance label entries.

[0073] Based on the blind zone markers and balance markers, the opinions are prioritized and reassignment instructions are generated. The implicit blind zone intervals of each item in the blind zone markers and the balance compensation dimensions of each item in the balance markers are aligned on the time axis. In the opinion priority reassignment, the opinion priority of departments with inertial consensus is reduced for symptom dimensions covered by the blind zone markers. The longer the duration of inertial consensus, the greater the decrease in opinion priority for the corresponding symptom dimension. The priority reduction is determined by the ratio of the duration of inertial consensus to the average duration of all items in the blind zone markers. Departments with a ratio greater than 1 experience a priority reduction exceeding the baseline reduction, while departments with a ratio less than 1 experience a reduction lower than the baseline reduction. Departments with underestimated weights in the symptom dimensions covered by the balance markers receive an upgrade in priority during the opinion priority reassignment. The upgrade is based on the compensation magnitude of the corresponding department in the balance markers. After the upgrade, opinions from departments with higher priority are preferentially included in the fusion input of the diagnostic knowledge graph in the reassignment instructions. When both blind zone markers and balance markers cover the same symptom dimension, the balance markers are adjusted upwards first, followed by the blind zone markers. This ensures that opinions from departments with underestimated weights are prioritized before those from departments with consistent consensus are given priority, preventing the overlap of these two types of operations from doubly suppressing or doubly boosting the priority of certain departments' opinions. After the priority of opinions for each symptom dimension is rearranged, the scheduling order of departmental opinions is arranged from highest to lowest priority. The allocation instruction consists of the scheduling order and the corresponding symptom dimension label. Each item in the allocation instruction includes three parts: symptom dimension, scheduling order, and priority adjustment basis. The priority adjustment basis is distinguished between the downward adjustment operation from the blind zone markers and the upward adjustment operation from the balance markers. The sources of these two types of operations are retained in the allocation instruction in the form of labels.

[0074] Step S15: Based on the allocation instructions, the diagnostic and treatment knowledge graph is globally optimized to generate a collaborative diagnostic and treatment path graph. Residual divergence features are extracted from the collaborative diagnostic and treatment path graph to form an extended node set. Based on the extended node set, the collaborative diagnostic and treatment path graph is revised and the spinal collaborative diagnostic and treatment results are output.

[0075] Specifically, a collaborative diagnosis and treatment path diagram is generated by globally optimizing the diagnosis and treatment knowledge graph based on allocation instructions. The reduction operation for blind zone markers in the allocation instructions uses the ratio of the duration of the corresponding item's inertial consensus to the average value of all items in the blind zone as the reduction coefficient, reducing the magnitude of the opinion weighting vector of the inertial consensus department at the corresponding symptom dimension node according to this coefficient. The upward adjustment operation for balanced markers expands the magnitude of the opinion weighting vector of the underweighted department according to the compensation magnitude, with the upper bound of compensation set as the current average confidence weight of each department in that dimension. When both types of operations are triggered simultaneously in the same symptom dimension, the upward adjustment is performed first, followed by the reduction, to avoid overlapping and causing some departments' opinions to be doubly suppressed. After the opinion weighting vectors of each symptom dimension node in the diagnosis and treatment knowledge graph are adjusted, they are re-normalized. The updated normalized weighted vectors replace the original vectors of the corresponding nodes in the diagnosis and treatment knowledge graph, and the fusion judgment vectors of each node are recalculated accordingly. The weights of directed edges are replaced by the cosine similarity of the updated fusion judgment vectors. Directed edges whose weight changes exceed 1.5 times the average change of all directed edge weights are considered significantly changed edges. After the node fusion judgment vector and directed edge weights of the diagnostic knowledge graph are updated, each node is arranged from high to low according to the graph centrality. The top three nodes with the highest centrality are marked as core diagnostic nodes. The directed paths between core diagnostic nodes are the main paths. The collaborative diagnosis path graph consists of the updated fusion judgment vectors of all nodes, directed edge weights, core diagnostic node labels, and main paths. The main paths in the collaborative diagnosis path graph centrally reflect the substantial impact of the adjustment of opinion priority after the execution of the allocation instruction on the diagnostic direction. The collaborative diagnosis path graph fully records the diagnostic basis and path structure of each dimension of the current patient's spinal symptoms.

[0076] Residual divergence features are extracted from the collaborative diagnosis and treatment path map to form an expanded node set. After extracting the fusion judgment vector of each node in the collaborative diagnosis and treatment path map, the absolute value of the difference between the mean amplitude of the support tendency component and the mean amplitude of the opposition tendency component is calculated for each node. Nodes whose absolute value exceeds 30% of the mean amplitude of all components of that node are identified as residual divergence nodes. After the execution of the allocation instruction for a residual divergence node, the opposition of specialist opinions within that symptom dimension has not been resolved. Residual divergence nodes in the collaborative diagnosis and treatment path map are classified according to their graph centrality. Residual divergence nodes with a centrality higher than the mean centrality of all nodes are identified as high-centrality residual divergence nodes. These nodes are both in a core position in the diagnosis and treatment path and have unresolved disagreements, thus having the most significant impact on the collaborative diagnosis and treatment results. For example, in a certain scenario, after the execution of the allocation instruction, the amplitudes of the support and opposition components in the fusion judgment vector of a certain symptom dimension node are still similar, and the graph centrality of this node ranks among the top of all nodes. It is identified as a high-centrality residual divergence node, indicating that the opposition of opinions in this symptom dimension has not been effectively converged after priority rearrangement. It is necessary to present two types of judgment directions in the path revision stage through branch expansion. Among residual divergent nodes, low-centrality nodes are identified as high-frequency divergent nodes based on a historical divergence frequency exceeding 70%. Although high-frequency divergent nodes have relatively low centrality, their divergence in this dimension has historically been persistent, and the currently unresolved divergences represent structural problems. The extended node set includes high-centrality residual divergent nodes and high-frequency divergent nodes. Each node entry is labeled with its residual divergence intensity and node type. The entire set is sorted in descending order of residual divergence intensity. The size of the extended node set is bounded to 30% of the total number of nodes in the collaborative diagnosis and treatment pathway map. If the size exceeds this bound, the top 30% of nodes with the highest residual divergence intensity are included.

[0077] Based on the expanded node set, the collaborative diagnosis and treatment path graph is revised to output the results of the spinal collaborative diagnosis and treatment. High-centrality residual divergence nodes in the expanded node set are revised using a branch expansion method. Starting from this node, a new branch path is added to the collaborative diagnosis and treatment path graph. The branch path extends along the direction of the opposing opinion with the highest residual divergence intensity until it rejoins the main path or reaches a leaf node. In a certain scenario, a high-centrality residual divergence node for a certain symptom dimension, after revision, simultaneously has a main path support direction and a branch path opposition direction. The two paths respectively gather departmental opinion chains supporting and opposing the positive symptoms of that dimension, forming a clear judgment fork in the collaborative diagnosis and treatment path graph at this node. High-frequency divergence nodes in the expanded node set are revised using a weight recalibration method. Each high-frequency divergence node in the expanded node set applies a historical divergence frequency weighted correction to the fusion judgment vector. The higher the historical divergence frequency of a node, the larger the correction coefficient. After correction, the fusion judgment vector updates the judgment basis of that node in the collaborative diagnosis and treatment path graph. The weights of adjacent directed edges are updated synchronously without propagation to distant nodes, avoiding cascading disturbances to the global path structure caused by local revisions. After the collaborative diagnosis and treatment pathway map is revised, the diagnostic conclusions of each symptom dimension are arranged from high to low according to the atlas centrality. Each conclusion includes the direction of fusion judgment, the weight of the opinions of the participating departments, and the intensity of residual disagreement. The spinal collaborative diagnosis and treatment results will output all symptom dimension diagnostic conclusions in descending order of atlas centrality. In the spinal collaborative diagnosis and treatment results, the branch path entries are presented in parallel as the independent basis for the two types of judgment directions. The residual disagreement intensity label attached to the conclusions of each dimension of the spinal collaborative diagnosis and treatment results makes both the mainstream diagnostic direction and the structural residual disagreement explicit.

[0078] To implement the collaborative diagnosis and treatment method for spinal diseases corresponding to the above method embodiments, and to achieve the corresponding functions and technical effects. See also Figure 2 , Figure 2 This diagram illustrates a structural block diagram of a collaborative diagnosis and treatment system 200 for spinal diseases according to an embodiment of this application. For ease of explanation, only the parts relevant to this embodiment are shown. The collaborative diagnosis and treatment system 200 for spinal diseases provided in this embodiment includes:

[0079] The data annotation unit 201 is used to collect patient spinal imaging data and medical records, perform sparse annotation on the imaging data and medical records to form a symptom missing item map, and extract missing item features based on the symptom missing item map to form staging clue identifiers;

[0080] Opinion scheduling unit 202 is used to combine the symptom missing item map and the staging clue identifier to perform joint annotation to form a diagnosis and treatment node chain, schedule cross-department opinions according to the diagnosis and treatment node chain to obtain an initial diagnosis and treatment opinion set, and perform time-series confidence analysis on the initial diagnosis and treatment opinion set to form a confidence weight map;

[0081] The graph construction unit 203 is used to use the confidence weight graph to perform departmental opinion differentiation analysis to form a divergence index graph, extract divergence patterns from the divergence index graph to form a complementary graph, and establish a diagnosis and treatment knowledge graph by fusing the initial diagnosis and treatment opinion set with the complementary graph.

[0082] The instruction generation unit 204 is used to identify conflict items and imbalance items in the cumulative deviation of the initial medical opinion set, extract the conflict intensity of the conflict items to form blind zone markers, perform specialty lineage matching between the imbalance items and the divergence index map to determine balance markers, and perform opinion priority rearrangement to generate allocation instructions according to the blind zone markers and the balance markers.

[0083] The result output unit 205 is used to globally optimize the diagnosis and treatment knowledge graph based on the allocation instruction to generate a collaborative diagnosis and treatment path graph, extract residual divergence features from the collaborative diagnosis and treatment path graph to form an extended node set, and revise the path of the collaborative diagnosis and treatment path graph based on the extended node set to output the spinal collaborative diagnosis and treatment result.

[0084] The aforementioned spinal disease collaborative diagnosis and treatment system 200 can implement the spinal disease collaborative diagnosis and treatment method of the above method embodiments. The options in the above method embodiments are also applicable to this embodiment, and will not be detailed here. The remaining content of this application embodiment can be referred to the content of the above method embodiments, and will not be repeated in this embodiment.

[0085] The purpose of the above embodiments is to reproduce and derive the technical solution of the present invention by way of example, and to fully describe the technical solution, purpose and effect of the present invention. The purpose is to enable the public to have a more thorough and comprehensive understanding of the disclosure of the present invention, and not to limit the scope of protection of the present invention.

Claims

1. A collaborative diagnosis and treatment method for spinal diseases, characterized in that, include: Collect spinal imaging data and medical records of patients, perform sparse annotation on the imaging data and medical records to form a symptom missing item map, and extract missing item features based on the symptom missing item map to form staging clue identifiers; By combining the symptom missing item map with the staging clue identifier, a joint annotation is performed to form a diagnosis and treatment node chain. According to the diagnosis and treatment node chain, cross-departmental opinions are scheduled to obtain an initial diagnosis and treatment opinion set. Time-series confidence analysis is performed on the initial diagnosis and treatment opinion set to form a confidence weight map. The confidence weight map is used to analyze the departmental opinion differences to form a divergence index map. The divergence patterns are extracted from the divergence index map to form a complementary map. The initial set of medical opinions is fused with the complementary map to establish a medical knowledge graph. For the cumulative bias in the initial medical opinion set, identify conflict and imbalance items, extract the conflict intensity of the conflict items to form blind zone markers, perform specialty lineage matching between the imbalance items and the divergence index map to determine balance markers, and perform opinion priority rearrangement to generate allocation instructions according to the blind zone markers and the balance markers. Based on the allocation instructions, the diagnostic and treatment knowledge graph is globally optimized to generate a collaborative diagnostic and treatment path graph. Residual divergence features are extracted from the collaborative diagnostic and treatment path graph to form an extended node set. Based on the extended node set, the collaborative diagnostic and treatment path graph is revised and the spinal collaborative diagnostic and treatment results are output.

2. The method according to claim 1, characterized in that, The step of performing time-series confidence analysis on the initial set of medical opinions to form a confidence weight graph includes: Extract the submission timestamps and content of each opinion from the initial treatment opinion set to generate a time-series opinion sequence; The time-series opinion sequence is used to identify the direction of opinion revision and generate a revision direction record; For each revision direction, an independent judgment tag is generated to identify counter-mainstream revision events. Based on the independent judgment marker, the opinions of each department in the initial diagnosis and treatment opinion set are assigned differentiated weights to form a confidence weight map.

3. The method according to claim 1, characterized in that, The process of obtaining an initial set of treatment opinions by scheduling cross-departmental opinions according to the treatment node chain includes: Based on the aforementioned diagnosis and treatment node chain, a sequence reference table is generated to identify the routine scheduling order of each department. Anchoring risk assessment was performed on the aforementioned sequence reference table to identify departments with high anchoring impact. For the departments with high anchoring impact, a de-anchoring scheduling sequence is generated by implementing sequential perturbation processing. Based on the aforementioned de-anchoring scheduling sequence, opinion collection tasks are issued to each department to form an initial set of diagnostic and treatment opinions.

4. The method according to claim 1, characterized in that, The process of using the confidence weight diagram to analyze departmental opinion differences and generate a divergence index diagram includes: The opinion vector set is generated by extracting the weighted vectors of opinions from each department from the confidence weight graph. The departmental overlap degree of the opinion vector set is calculated to generate a professional overlap degree map. The pseudo-consensus clustering events are identified and pseudo-consensus markers are generated based on the aforementioned professional overlap graph; A divergence index map is generated by weighting the difference intensity based on the pseudo-consensus markers and the opinion vector set.

5. The method according to claim 1, characterized in that, The step of extracting divergence patterns from the divergence index map to form a complementary map includes: The departmental opinion dominance sequence is generated by tracking the opinion dominance of the aforementioned divergence index graph. For the aforementioned dominance sequence, identify cross-departmental dominance rotation events and generate a rotation node set; Disease association analysis is performed on the rotating node set to identify complementary boundary diseases; Based on the complementary boundary diseases, complementary knowledge relationships between departments are constructed to form a complementary map.

6. The method according to claim 1, characterized in that, The step of extracting the conflict intensity of the conflict item to form a blind zone marker includes: Multidimensional conflict feature extraction is performed on the conflict terms to generate a conflict intensity distribution; Based on the conflict intensity distribution, identify persistently low conflict intervals and generate a zero-conflict candidate set; The zero-conflict candidate set is subjected to inertial consensus discrimination to generate hidden blind zone intervals; Blind spot markers are formed by jointly locating the hidden blind spot intervals and the conflict intensity distribution.

7. The method according to claim 1, characterized in that, The step of performing specialized genealogical matching between the imbalance term and the divergence index diagram to determine the equilibrium marker includes: Extract the weight deviation of each department from the imbalance term to generate a weight deviation sequence; The imbalance disease types are identified by performing an association mapping between the weighted deviation sequence and the divergence index graph; For the aforementioned imbalanced disease types, the opinion adoption rate of the currently low-weight departments is extracted to generate an adoption rate ranking table; Based on the adoption rate ranking table, a weight reversal compensation is performed on the imbalanced disease types to determine the equilibrium marker.

8. The method according to claim 2, characterized in that, The step of identifying the revision direction of the time-series opinion sequence and generating a revision direction record includes: The time-series opinion sequence is compared with adjacent opinions to identify opinion change points; Extract the revision time intervals of the opinion change points to generate a revision interval sequence; For the revision interval sequence, identify ultra-long interval deep revision events and generate deep analysis markers; Based on the aforementioned in-depth analysis markers, the confidence level of the opinion change points is increased to generate revision direction records.

9. The method according to claim 5, characterized in that, The step of generating a set of rotation nodes for identifying cross-departmental leadership rotation events in the leadership sequence includes: The frequency of opinion fusion and retention is statistically analyzed to generate a baseline of latent influence. A set of latent dominant events is generated by identifying silent dominant events based on the aforementioned latent influence baseline; Perform explicit / implicit dominance switching pattern analysis on the implicit dominance event set to identify explicit / implicit dominance rotation pairs; Based on the explicit and implicit dominant rotation pair, a rotation node set is generated for the positioning rotation time sequence node.

10. A collaborative diagnosis and treatment system for spinal diseases, characterized in that, include: The data annotation unit is used to collect patient spinal imaging data and medical records, perform sparse annotation on the imaging data and medical records to form a symptom missing item map, and extract missing item features based on the symptom missing item map to form staging clue identifiers; The opinion scheduling unit is used to combine the symptom missing item map with the staging clue identifier to perform joint annotation to form a diagnosis and treatment node chain, schedule cross-departmental opinions according to the diagnosis and treatment node chain to obtain an initial diagnosis and treatment opinion set, and perform time-series confidence analysis on the initial diagnosis and treatment opinion set to form a confidence weight map. The graph construction unit is used to analyze the departmental opinion differences using the confidence weight graph to form a divergence index graph, extract divergence patterns from the divergence index graph to form a complementary graph, and fuse the initial set of medical opinions with the complementary graph to establish a medical knowledge graph. The instruction generation unit is used to identify conflict items and imbalance items in the cumulative deviation of the initial medical opinion set, extract the conflict intensity of the conflict items to form blind zone markers, perform specialty lineage matching between the imbalance items and the divergence index map to determine equilibrium markers, and perform opinion priority rearrangement to generate allocation instructions according to the blind zone markers and equilibrium markers. The result output unit is used to globally optimize the diagnosis and treatment knowledge graph based on the allocation instruction to generate a collaborative diagnosis and treatment path graph, extract residual divergence features from the collaborative diagnosis and treatment path graph to form an extended node set, and revise the path of the collaborative diagnosis and treatment path graph based on the extended node set to output the spinal collaborative diagnosis and treatment result.

Images