Hierarchical early warning method and system for dizziness diagnosis

Abstract

The present application relates to the technical field of medical information processing and intelligent risk assessment, and discloses a hierarchical early warning method and system for dizziness diagnosis, which comprises the following steps: S1, receiving multi-modal evidence data related to a test subject, wherein the multi-modal evidence data comprises diagnosis text information and examination-related data, and the examination-related data comprises at least one of the following: eye movement tracking data, nystagmus examination data, hearing examination data and vestibular function detection data; S2, processing the multi-modal evidence data respectively to obtain structured feature vector information of the multi-modal evidence data, and generating quality indicators corresponding to the multi-modal evidence data; the method and system organize the multi-modal evidence into evidence event sequences and update them one by one to form multi-stage risk assessment results, and then identify uncertainty types and give stage interpretation information, thereby improving the explainability and clinical adaptability of the early warning results.

Description

Technical Field

[0001] This invention relates to the field of medical information processing and intelligent risk assessment technology, specifically to a graded early warning method and system for vertigo diagnosis. Background Technology

[0002] Vertigo is a common and complex clinical symptom involving a variety of possible causes, including peripheral vertigo (such as benign paroxysmal positional vertigo, vestibular neuritis, etc.) and central vertigo (such as stroke, brain tumor, etc.). Clinically, doctors need to conduct a comprehensive analysis of the patient's medical history and a series of objective examinations (such as eye movement / nystagmus video, audiograms, vestibular evoked myogenic potential data, etc.) to diagnose the specific type and cause of vertigo.

[0003] However, the diverse and nonspecific symptoms of vertigo make the diagnostic process complex and susceptible to interference from the patient's subjective experience. Existing assessment methods often rely on one-off fusion or static assessments, which output a single risk value, confidence level, or consistency score based on all or part of the evidence obtained. This approach typically fails to reflect the varying impacts of different evidence combinations on the assessment conclusions and struggles to provide identifiable evidence when assessment results change, thus affecting the interpretability and safety of risk warnings.

[0004] Therefore, there is a need for a technical solution that can analyze changes in risk assessment results as evidence accumulates and output explanatory information related to these changes, in order to improve the interpretability and clinical usability of the vertigo assessment process. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a graded early warning method and system for vertigo diagnosis. This method and system organizes multimodal evidence into a sequence of evidence events and merges and updates them one by one to form a multi-stage risk assessment result. It then identifies the type of uncertainty and provides stage explanation information, thereby improving the interpretability and clinical applicability of the early warning result.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A graded early warning method for vertigo diagnosis includes: S1, receiving multimodal evidence data related to the examinee, the multimodal evidence data including consultation text information and examination-related data, the examination-related data including at least one of the following: eye-tracking data, nystagmus test data, hearing test data, and vestibular function test data; S2, processing the multimodal evidence data to obtain structured feature vector information of the multimodal evidence data, and generating quality indicators corresponding to the multimodal evidence data; S3, forming an evidence event sequence by merging the structured feature vectors corresponding to the multimodal evidence data according to the access order or a preset evidence access strategy; S4, performing fusion and update one by one according to the order of the evidence event sequence, in each fusion... S5. After updating, output the risk assessment results for the corresponding stage to obtain the risk assessment results for each stage; S6. Calculate the overall uncertainty index and the evidence conflict index based on the risk assessment results for each stage, and determine the uncertainty type based on the overall uncertainty index and the evidence conflict index. The uncertainty type includes at least insufficient evidence uncertainty and evidence conflict uncertainty; S7. Determine the risk change based on the risk assessment results for each stage, and then determine the key evidence stage that leads to the change in the risk assessment results, and generate stage explanation information; S8. Map the risk assessment results, uncertainty types, and stage explanation information to a graded early warning level, and output risk warning information and / or risk handling suggestions corresponding to the graded early warning level.

[0007] In this invention, preferably, step S2, processing the multimodal evidence data includes: S21, inputting the consultation text information into a natural language processing model for word segmentation, entity recognition, semantic understanding, and relation extraction, extracting text semantic features, and obtaining structured feature vector information of the consultation text; S22, inputting the examination-related data into at least one artificial intelligence model for signal, image, or sequence analysis, extracting examination data features, and obtaining structured feature vector information corresponding to the examination-related data.

[0008] In this invention, preferably, it further includes: S55, determining the type of evidence relationship based on the overall uncertainty index, the evidence conflict index, and the risk assessment results of each stage, wherein the type of evidence relationship includes insufficient evidence, corroborating evidence, and conflicting evidence; wherein, when the risk assessment results of adjacent stages meet at least one of the following, the evidence relationship is determined to be corroborating evidence: the support of the risk assessment results of adjacent stages continuously increases in the same risk assessment direction; the overall uncertainty index decreases in the later stage of adjacent stages and the evidence conflict index does not increase.

[0009] In this invention, preferably, the consultation text information includes at least the chief complaint, onset characteristics, triggering factors, duration, accompanying symptoms, and past medical history; the structured feature vector of the consultation text includes at least symptom features, time features, and triggering relationships.

[0010] In this invention, preferably, in step S2, generating quality indicators corresponding to the multimodal evidence data includes: for consultation text information, generating quality indicators based on the completeness of the text content, the missing status of key symptom fields, and the consistency of the descriptive content; for examination-related data, generating quality indicators based on at least one of the following: data clarity, effective data ratio, identifiability of the target area, and whether the collection time meets preset requirements.

[0011] In this invention, preferably, in step S5, the overall uncertainty index is used to characterize the stability or convergence of the comprehensive risk assessment results, including the fluctuation range of the risk assessment results in multiple stages and / or the concentration of the risk assessment results in the final stage; the evidence conflict index is used to characterize the degree of disagreement in risk judgment between different stages or different evidence, the degree of disagreement including the distribution distance of the current stage risk assessment results relative to the previous stage risk assessment results and / or the degree of deviation between the risk tendency corresponding to the current evidence event and the historical fusion results.

[0012] In this invention, preferably, in step S6, determining the amount of risk change includes: calculating the degree of difference in the distribution of risk assessment results between adjacent stages; determining the key evidence stage that leads to the change in risk assessment results includes: identifying the evidence stage with the largest amount of risk change as the key evidence stage.

[0013] In this invention, preferably, in step S6, the stage explanation information includes at least a key evidence stage identifier and evidence event index information corresponding to the key evidence stage; the evidence event index information includes at least one of the following: text fragment index information of consultation text information, timestamp index information corresponding to examination-related data, image frame index information, and curve node index information.

[0014] In this invention, preferably, in step S4, the sequential execution of the fusion update includes: S41, maintaining the historical fusion state, which includes at least the comprehensive risk assessment distribution of the previous stage, the uncertainty index and conflict index of the previous stage, and the index set of accessed evidence events; S42, for the currently accessed evidence event, determining the evidence quality weight based on its corresponding quality index; S43, weighting and synthesizing the risk assessment distribution vector of the current evidence event with the comprehensive risk assessment distribution in the historical fusion state according to the evidence quality weight to obtain the comprehensive risk assessment distribution of the current stage, which serves as the risk assessment result of the current stage.

[0015] A graded early warning system for vertigo diagnosis includes: a multimodal data acquisition module for acquiring multimodal evidence data related to the subject, the multimodal evidence data including at least consultation text information and at least one examination-related data; one or more processors communicatively connected to the multimodal data acquisition module; a memory; and one or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the one or more processors, the one or more applications being configured to perform the method as described in any of the preceding claims.

[0016] Compared with the prior art, the beneficial effects of the present invention are: 1. By adopting a dynamic fusion and update approach, multimodal evidence is organized into a sequence and fused sequentially, which can reflect the dynamic changes in risk assessment results as evidence accumulates, thus solving the problem that static assessment methods cannot reflect the evolutionary process.

[0017] 2. By calculating uncertainty and conflict indices, the type of uncertainty can be accurately identified, key evidence stages can be located, and interpretable stage information can be generated, thereby improving the interpretability and clinical credibility of risk warnings.

[0018] 3. Adjusting the fusion weights based on evidence quality indicators reduces the interference of low-quality evidence on the assessment results and improves the robustness and accuracy of risk assessment.

[0019] 4. It outputs graded early warning levels and targeted treatment suggestions, which can provide clear diagnosis and treatment guidance for clinicians, assist doctors in quickly formulating the next treatment plan, and improve the efficiency and accuracy of diagnosing the cause of vertigo. Attached Figure Description

[0020] Figure 1 This is a flowchart of a graded early warning method for vertigo diagnosis according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating the processing of multimodal evidence data in step S2 of a graded early warning method for vertigo diagnosis according to an embodiment of the present invention. Figure 3 This is a flowchart illustrating the sequential execution and fusion update of step S4 in a graded early warning method for vertigo diagnosis according to an embodiment of the present invention. Figure 4 This is a flowchart of a graded early warning method for vertigo diagnosis according to another embodiment of the present invention; Figure 5 This is a structural block diagram of a graded early warning system for vertigo diagnosis according to an embodiment of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0023] Please see Figure 1 A preferred embodiment of the present invention provides a graded early warning method for vertigo diagnosis, comprising: S1, receiving multimodal evidence data related to the subject of examination, the multimodal evidence data including consultation text information and examination-related data, the examination-related data including at least one of the following: eye-tracking data, nystagmus test data, hearing test data, vestibular function test data.

[0024] This multimodal evidence data includes at least the patient's medical history and at least one type of examination-related data. The medical history reflects subjective information such as the subject's chief complaint, symptoms, triggering factors, duration, accompanying symptoms, and past medical history. The examination-related data reflects the subject's objective examination results, which may include, but are not limited to, eye movement or nystagmus test data, hearing test data, and vestibular function-related test data. Examples include eye movement or nystagmus test videos, hearing test graphs, or vestibular evoked myogenic potential data.

[0025] By receiving medical records and examination data, the system can comprehensively utilize both subjective and objective evidence to provide a foundational data source for subsequent assessments.

[0026] S2 processes the multimodal evidence data to obtain the structured feature vector information of the multimodal evidence data and generates quality indicators corresponding to the multimodal evidence data.

[0027] The system processes the received multimodal evidence data separately, converting the raw data into a structured representation suitable for computation and fusion. Specifically, the system can perform text processing and semantic analysis on consultation text information to extract text features reflecting symptom characteristics, temporal characteristics, and precipitating relationships; and perform signal, image, or sequence processing on examination-related data to extract examination features reflecting objective examination results. Preferably, the structured feature vector extraction in S2 can be achieved through algorithms or artificial intelligence models. In particular, the structured feature extraction of consultation text can be achieved through language models, while the structured feature extraction of nystagmus video can be achieved through neural networks or temporal feature extraction models for video analysis.

[0028] like Figure 2 As shown, preferably, in S2, the processing of multimodal evidence data includes: S21, input the consultation text information into the natural language processing model for word segmentation, entity recognition, semantic understanding and relation extraction, extract the semantic features of the text, and obtain the structured feature vector information of the consultation text; S22, the relevant data is input into at least one artificial intelligence model for signal, image or sequence analysis, the features of the data are extracted, and the structured feature vector information corresponding to the relevant data is obtained.

[0029] This artificial intelligence model is used to extract discriminative feature representations from input examination data and transform these feature representations into structured feature vectors with a unified dimension or that can be fused. The at least one artificial intelligence model can be a single model structure capable of uniformly modeling and comprehensively processing different types of examination-related data to output corresponding structured feature vectors; alternatively, the at least one artificial intelligence model can include multiple models, each performing feature extraction for different types of examination-related data, such as independently analyzing eye movement or nystagmus data, hearing test data, and vestibular function test data.

[0030] In the multi-model implementation, the feature representations output by each model can be converted into corresponding structured feature vectors. In the single-model implementation, data from different modalities can be jointly extracted using a shared coding structure or a multi-branch network structure. Through these methods, automated feature encoding of inspection data from different modalities is achieved, providing a unified format of structured feature vector information for subsequent fusion and update steps.

[0031] In one implementation, the artificial intelligence model may include a convolutional neural network, a recurrent neural network, a Transformer model, or a combination thereof, for feature encoding of eye movement or nystagmus data, hearing test data, or vestibular function test data. The feature representation output by the model, after necessary mapping or standardization, forms corresponding structured feature vector information for use in subsequent fusion and update steps.

[0032] The artificial intelligence model is a trained feature extraction model used to learn data representations from relevant inspection data of the corresponding type. The model can be trained using supervised learning, self-supervised learning, or unsupervised learning methods to obtain effective feature encoding capabilities for inspection data.

[0033] Furthermore, the system can generate quality indicators for various types of multimodal evidence data, which characterize the quality level of the evidence in terms of acquisition conditions, completeness, clarity, or credibility. These quality indicators can serve as reference information in subsequent fusion and update processes, applying differentiated weights to different types of evidence to reduce the impact of low-quality evidence on risk assessment results and improve the robustness of the overall assessment.

[0034] Quality indicators are generated based on the completeness of the text content, the presence of missing key symptom fields, and the consistency of the description. If the text fully records the onset time, triggering factors, and accompanying symptoms, the quality indicator is high; if key information is missing, the quality indicator is medium or low.

[0035] Quality indicators are generated based on the clarity of the inspection data, the proportion of effective data, the identifiability of the target area, and the collection time.

[0036] For nystagmus video data, quality metrics can be generated based on any one or more of the following: video clarity, effective frame rate, identifiability of the eye region, and whether the video acquisition duration meets preset requirements. For example, if the nystagmus video has high clarity, an effective frame rate of 90%, and the eye region is clearly identifiable, then the quality metric is high.

[0037] S3, the structured feature vectors corresponding to the multimodal evidence data are arranged into an evidence event sequence according to the access order or a preset evidence access strategy.

[0038] In one embodiment of the present invention, the system organizes the structured feature vector obtained in S2 according to the access order of multimodal evidence data or a preset evidence access strategy to form an evidence event sequence.

[0039] Each evidence event corresponds to a set of structured feature vectors, which are used to represent the evidence information obtained at a certain moment or stage.

[0040] In some implementations, multimodal evidence data can be accessed in batches along with the diagnosis and treatment process; in other implementations, multimodal evidence data can be obtained all at once. Regardless of the method of evidence acquisition, the system can organize it into a sequence of evidence events and perform fusion and updates sequentially to form risk assessment results at each stage.

[0041] For example, the preset evidence access strategy is clinical priority ranking, in the following order: consultation text information → nystagmus video data → cranial magnetic resonance imaging data. Therefore, the resulting sequence of evidence events is: first evidence event (feature vector of consultation text), second evidence event (feature vector of nystagmus video), and third evidence event (feature vector of magnetic resonance imaging).

[0042] S4 executes the fusion update one by one according to the sequence of evidence events, and outputs the risk assessment results of the corresponding stage after each fusion update, thereby obtaining the risk assessment results of each stage.

[0043] During the fusion update process, the system combines or updates the structured feature vector corresponding to the current evidence event with the fusion result formed by previous evidence, thereby forming the fusion result for the current stage and generating the corresponding risk assessment result. By performing fusion updates on the evidence event sequence one by one, the system generates multiple staged risk assessment results during the assessment process, reflecting the changes in risk judgment as evidence is gradually fused.

[0044] For example, when evaluating solely based on structured feature vector information from the consultation text, the system outputs an initial risk assessment result. Once structured feature vectors from eye movement or nystagmus examination videos are integrated, the system combines these newly integrated feature vectors with the risk assessment result from the previous stage to generate the current stage's risk assessment result. When further examination data is integrated, the system continues to update the latest stage's risk assessment result based on the current fusion status. The risk assessment result can be used to characterize the subject's vertigo-related risk level under the current evidence conditions.

[0045] like Figure 3 As shown, the specific steps of performing the merge update one by one include: S41, Maintain the historical fusion status, which includes at least the distribution of the comprehensive risk assessment of the previous stage, the uncertainty indicators and conflict indicators of the previous stage, and the set of indexes of evidence events that have been connected. S42, For the currently accessed evidence event, determine the evidence quality weight based on its corresponding quality index; S43. The risk assessment distribution vector of the current evidence event is weighted and synthesized with the comprehensive risk assessment distribution in the historical fusion state according to the evidence quality weight to obtain the comprehensive risk assessment distribution of the current stage, which is used as the risk assessment result of the current stage.

[0046] S5. Calculate the overall uncertainty index and the evidence conflict index based on the risk assessment results of each stage, and determine the uncertainty type based on the overall uncertainty index and the evidence conflict index. The uncertainty type includes at least the uncertainty of insufficient evidence and the uncertainty of conflicting evidence.

[0047] The overall uncertainty index characterizes the stability or convergence of the comprehensive risk assessment results, including the volatility of the risk assessment results across multiple stages and / or the concentration of the risk assessment results in the final stage. The evidence conflict index characterizes the degree of disagreement in risk judgment between different stages or different pieces of evidence. The degree of disagreement includes the distributional distance of the current stage's risk assessment results relative to the previous stage's risk assessment results and / or the deviation of the risk tendency corresponding to the current evidence event from the historical fusion results. Based on the overall uncertainty index and the evidence conflict index, the system identifies the sources of uncertainty in the risk assessment results, enabling it to distinguish whether the instability in the risk assessment is due to insufficient quantity or quality of evidence or to conflicts between pieces of evidence, thereby improving the interpretability of the assessment results.

[0048] S6. Based on the risk assessment results at each stage, determine the amount of risk change, then identify the key evidence stage that led to the change in the risk assessment results, and generate stage explanation information.

[0049] The system compares the risk assessment results of adjacent or multiple stages to determine the amount of risk change. By analyzing the distribution of risk change, the system can identify the evidence stage that has a significant impact on the risk assessment results and designate that stage as the key evidence stage.

[0050] Determining the magnitude of risk change includes calculating the degree of difference in the distribution of risk assessment results between adjacent stages.

[0051] The key evidence phase for determining changes in risk assessment results includes identifying the phase with the greatest change in risk as the key evidence phase.

[0052] The system further generates stage explanation information to characterize the position of the key evidence stage throughout the assessment process and its impact on changes in the risk assessment results. The stage explanation information includes at least: (1) Key evidence stage identifier (target stage identifier that triggers uncertainty type); (2) Evidence index information corresponding to the target stage.

[0053] The evidence event index information includes at least one of the following: text fragment index information of consultation text information, timestamp index information corresponding to examination-related data, image frame index information, and curve node index information.

[0054] This step enables traceable analysis of the sources of risk changes, helping to pinpoint key evidence leading to changes in risk assessment.

[0055] S7 maps the risk assessment results, uncertainty types, and stage explanation information to graded early warning levels, and outputs risk warning information and / or risk handling suggestions corresponding to the graded early warning levels.

[0056] In one implementation, the risk warning information output by the system includes at least: the current risk level, the stability of the risk assessment, the source of uncertainty, and the key evidence stage; the handling suggestions output by the system include at least one of the following: further inspection suggestions, manual review suggestions, or follow-up suggestions.

[0057] In a preferred embodiment of the present invention, such as Figure 4 As shown, a graded early warning method for vertigo diagnosis also includes: S55, based on the overall uncertainty index, the evidence conflict index, and the risk assessment results at each stage, determines the type of evidence relationship, which includes insufficient evidence, corroborating evidence, and conflicting evidence.

[0058] Among them, the evidentiary relationship is determined to be of the corroborative type when the risk assessment results of adjacent stages meet at least one of the following conditions: The support for risk assessment results in adjacent phases increases continuously in the same risk assessment direction; The overall uncertainty index decreases in the subsequent stage of an adjacent stage, while the evidence conflict index does not increase.

[0059] In one embodiment of the present invention, taking the received multimodal evidence data including consultation text information, nystagmus video (nystagmus examination data) and imaging examination data as an example, by constructing an evidence event sequence and performing fusion updates on each of them, it can adapt to complex clinical scenarios where multiple evidence events are gradually accessed, and realize the dynamic evolution and stage interpretation of risk assessment results, uncertainty types and key evidence stages.

[0060] (1) Handling of the first evidence event (medical interview transcript) and initial risk assessment First, the system receives the patient's medical history text. This text contains the patient's chief complaint: "I suddenly felt dizzy when I woke up this morning." The system processes this text, extracting the symptom triggering time, inducing conditions, and subjective experience features to generate a corresponding structured feature vector and a quality index.

[0061] For the consultation text information, the system uses a natural language processing (NLP) model to perform semantic parsing and structured processing. The NLP model can be a deep learning-based text understanding model, which can perform word segmentation, entity recognition, semantic understanding, and relation extraction on the input consultation text, thereby extracting key semantic features related to vertigo assessment.

[0062] In this embodiment, the structured feature vector of the consultation text includes at least one or more of the following sub-features: symptom description features, time and onset features, triggering and remission features, and accompanying symptom features.

[0063] Based on the structured feature vector of the consultation text, the system identified that the chief complaint was related to changes in body position, and the symptoms conformed to the typical descriptive features of common peripheral vertigo. The system used the structured feature vector of the consultation text as the first evidence event in the evidence event sequence, performed a fusion update, and obtained the first-stage risk assessment results.

[0064] At this stage, the risk assessment results generated by the system tend to be related to peripheral vertigo risk. However, since the assessment is based solely on the chief complaint information, the system retains a certain degree of uncertainty for further verification by subsequent evidence.

[0065] (2) Fusion update after accessing the second evidence event (nystagmus video) Subsequently, the system receives video data of the subject's eye movements as relevant data for nystagmus examination. The system processes the nystagmus video data, extracts examination features reflecting the direction, timing, and duration of nystagmus, generates corresponding structured feature vectors for nystagmus examination, and generates corresponding quality indicators.

[0066] For example, in this embodiment, the system extracts the following feature information from the nystagmus video: Directional characteristics of nystagmus: downbeat nystagmus; Nystagmus occurred during the following time period: 00:15-00:21 on the video timeline; Nystagmus characteristics: It has typical manifestations related to central nervous system lesions.

[0067] For video data related to nystagmus examination, the system employs an artificial intelligence model, different from the text processing model used in medical consultations, to analyze and process the nystagmus videos. This AI model can be a neural network-based video analysis model or a temporal feature extraction model, capable of performing frame-by-frame or sequence analysis on the input nystagmus video to extract structured information reflecting eye movement characteristics.

[0068] In this embodiment, the structured feature vector of nystagmus video includes at least one or more of the following sub-features: nystagmus direction features, nystagmus dynamic features (such as nystagmus amplitude), evoked related features, and temporal features.

[0069] The system uses the structured feature vector of nystagmus examination as the next evidence event in the evidence event sequence, performs fusion update according to the order of the evidence event sequence, and obtains the risk assessment result of the second stage. This result points to a risk judgment with central lesion characteristics and is significantly inconsistent with the risk assessment result of the first stage. Based on this, the system determines that the uncertainty type is evidence conflict uncertainty and identifies the nystagmus examination stage as the key evidence stage.

[0070] (3) Access and integration of third evidence events (imaging examinations) During further diagnosis and treatment, the system receives cranial imaging data from the examinee, such as magnetic resonance imaging results.

[0071] The system processes the imaging examination data, generates corresponding structured feature vectors and quality indicators, and uses them as the third evidence event in the evidence event sequence to perform fusion updates, thereby obtaining the risk assessment results of the third stage.

[0072] If the imaging results are consistent with the direction of the second stage, the system can further converge the risk assessment results and reduce the overall uncertainty index.

[0073] (4) Uncertainty Evolution and Type Adjustment under Multi-Evidence Events Based on the risk assessment results from the first to the third phase, the system continuously calculates the overall uncertainty index and the evidence conflict index. As third-party evidence events are added, the system finds that the risk assessment results gradually converge in the same direction, the evidence conflict index gradually decreases, and the overall uncertainty index decreases accordingly.

[0074] (5) Dynamic stage interpretation update of key evidence stage As the system gradually integrates multiple evidence events, it continuously updates the judgment results of the key evidence stage based on the changes in the risk assessment results at each stage.

[0075] In this embodiment, the system identifies the second stage (nystagmus examination evidence event) as the key evidence stage that triggers a change in the direction of risk assessment, while also recording the corroborating role of the third stage evidence event in the stabilization process of the risk assessment results. The stage explanation information generated by the system can reflect the complete evolution path of the risk assessment results from initial judgment, conflict, and gradual convergence.

[0076] (6) Tiered early warning output under multi-evidence events The system integrates the risk assessment results, uncertainty types, and stage explanations of the final stage to output corresponding graded early warning levels and risk alerts. Through the stepwise fusion and evaluation of multiple evidentiary events, the system can provide doctors with comprehensive early warning information that includes the evidence evolution process, key evidence stages, and the current level of risk stability.

[0077] Furthermore, this invention provides an example of the implementation of the fusion update: (1) Single evidence risk assessment output (providing unified input for fusion) In order to unify and integrate evidence from different modalities, the system processes each evidence event at the feature layer and outputs a set of phased risk assessment distributions, which are essentially "the degree to which the evidence supports different risk assessment directions".

[0078] For any evidence event e, the system outputs: R_e: Risk assessment distribution vector (e.g., corresponding to multiple risk levels or multiple risk assessment directions), satisfying that each component is non-negative and the sum is 1; Q_e: Evidence quality weight (calculated from quality features), used to represent the credibility of evidence.

[0079] R_e is used for subsequent fusion updates, and Q_e is used to control the intensity of the fusion impact.

[0080] (2) Historical integration status and integration update The system maintains historical fusion states S_{t-1}, which include at least: The distribution of the comprehensive risk assessment in the previous stage is R_{t-1}; The uncertainty and conflict indicators of the previous stage (used for comparison and stage interpretation). The evidence event index set has been accessed.

[0081] When subsequent evidence event e_t is added, the incremental fusion update after the new evidence is added. When a new evidence event e_t is added, the system, based on the evidence quality weight Q_{e_t} and the historical comprehensive risk distribution R_{t-1}, introduces the risk assessment distribution R_{e_t} of the new evidence to obtain an updated comprehensive risk distribution R_t, which serves as the risk assessment result for the current stage. The updated result maintains the distribution form (non-negative and normalized). One feasible update method is to use W_{e_t} as the weight to weight and synthesize R_{e_t} and R_{t-1} into R_t.

[0082] When subsequent evidence events are added, the system repeats the above process using R_t as the new historical fusion state as the basis to achieve continuous updates.

[0083] Furthermore, this invention provides an example of the calculation of the fused index and the determination of uncertainty: After obtaining R_t, the system calculates at least the following two types of metrics: (1) Overall uncertainty index U_t: used to characterize the concentration of the distribution of comprehensive risk assessment.

[0084] When R_t exhibits similar probabilities in multiple directions and has a relatively flat distribution, U_t is higher; when R_t is significantly biased towards a certain direction, U_t is lower.

[0085] (2) Evidence conflict index C_t: used to characterize the degree of divergence between new evidence and existing evidence / historical integration status in terms of risk orientation.

[0086] C_t can be determined based on the degree of difference between the newly added evidence distribution R_{e_t} and the historical distribution R_{t-1} or the degree of difference between multiple evidence distributions.

[0087] (3) Stage change Δ_t: used to characterize the change of R_t relative to R_{t-1}, and used for key evidence stage identification and stage interpretation.

[0088] The system can identify the sources of uncertainty based on U_t, C_t, and Δ_t: (A) Insufficient Evidence If the overall risk assessment distribution is still scattered (U_t is high) and there is no obvious evidence discrepancy (C_t is low), it is judged as an insufficient evidence state.

[0089] (B) Conflict of Evidence When new evidence significantly increases the degree of evidence disagreement (C_t increases), or causes a significant reversal / jump in the overall risk tendency (Δ_t is significant), it is judged as an evidence conflict state.

[0090] Furthermore, based on the risk assessment results, uncertainty types, and stage-specific explanations at each stage, tiered early warning outputs are provided. Several examples are given below: (1) Example of graded early warning output for cases of insufficient evidence After performing a step-by-step fusion and update of the evidence event sequence, the system found that: the risk assessment results at each stage were scattered; the risk assessment results changed only slightly with the access of evidence; no obvious evidence conflicts were detected; and the overall uncertainty index was at a high level. Under these circumstances, the system determined that the current uncertainty type was insufficient evidence uncertainty.

[0091] Based on this, the system determines the warning level to be a "prompt" warning and outputs corresponding risk warning information, such as: the current evidence is insufficient to form a stable risk judgment; it is recommended to supplement further objective examination data or improve the consultation information. In this case, the treatment suggestions are mainly "supplement evidence / supplement examinations / improve information", and it may indicate that the current assessment results need to be further confirmed with subsequent evidence.

[0092] (2) Example of graded early warning output for evidence conflict situations In another embodiment, during the process of fusing the evidence event sequence one by one, the system discovers that: newly added evidence events cause a significant change in the direction of the risk assessment results; there are obvious discrepancies between the risk assessment results at different stages; the evidence conflict index increases significantly; and the stage interpretation information indicates that a certain evidence stage is the key stage that triggers the risk change.

[0093] For example, the risk assessment result in the initial consultation text stage points to a risk related to peripheral vertigo, while the risk assessment result in the subsequent nystagmus examination evidence stage points to a risk related to central vertigo. There is a significant inconsistency in risk orientation between the two. The system identifies the current uncertainty type as conflicting evidence uncertainty and determines the corresponding warning level as either a warning level or a high-risk level warning.

[0094] The system's risk warnings may include: significant discrepancies exist between different pieces of evidence; newly acquired evidence significantly impacts the risk assessment; and a focus on reviewing key evidence stages. Further system recommendations may include: prompt further investigation; and manual review or comprehensive assessment by medical personnel.

[0095] In this scenario, the stage interpretation information is used to indicate the target stage identifier that triggered the conflict and the corresponding evidence index information, thereby enabling medical personnel to quickly locate the source of the conflict.

[0096] (3) Example of graded early warning output in evidence-based situations In another embodiment, during the process of merging and updating the evidence event sequence one by one, the system found that: the newly added evidence events are consistent with the existing evidence in the direction of risk assessment; the risk assessment results gradually converge after multi-stage fusion; the overall uncertainty index gradually decreases; and no obvious evidence conflicts are detected.

[0097] For example, in the initial stage, multiple risk assessments are generated based on the patient's medical history. However, after incorporating evidence from nystagmus testing or other objective examinations, the risk assessment results gradually converge on the same risk direction, and this direction is consistently supported across multiple stages. In this case, the system determines the existence of corroborating evidence and outputs a "judgment tends to stabilize / evidence mutually supports" message while maintaining the risk level mapping.

[0098] The system's output recommendations may include: suggesting that subsequent treatment arrangements be made in accordance with the current risk assessment; or suggesting that appropriate measures be taken based on clinical judgment.

[0099] like Figure 5 As shown, another embodiment of the present invention also discloses a graded early warning system 100 for vertigo diagnosis, the system comprising: The multimodal data acquisition module 110 is used to acquire multimodal evidence data related to the subject of examination, wherein the multimodal evidence data includes at least consultation text information and at least one examination-related data; One or more processors 111, wherein the one or more processors 111 are communicatively connected to the multimodal data acquisition module; Memory 112; One or more applications, wherein the one or more applications are stored in the memory 112 and configured to be executed by the one or more processors 111, the one or more applications being configured to perform the graded early warning method for vertigo diagnosis as described in any of the foregoing embodiments.

[0100] Processor 111 may include one or more processing cores. Processor 111 connects to various parts of the graded early warning system 100 for vertigo diagnosis using various interfaces and lines. It executes various functions and processes data within the graded early warning system 100 for vertigo diagnosis by running or executing instructions, programs, code sets, or instruction sets stored in memory 112, and by calling data stored in memory 112. Optionally, processor 111 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). Processor 111 may integrate one or more of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and computer programs; the GPU is responsible for rendering and drawing the displayed content; and the modem handles wireless communication. It is understood that the modem may also not be integrated into processor 111 and may be implemented separately using a communication chip. Specifically, the methods described in the foregoing embodiments can be executed by one or more processors 111.

[0101] In some implementations, memory 112 may include random access memory (RAM) or read-only memory (ROM). Memory 112 can be used to store instructions, programs, code, code sets, or instruction sets. Memory 112 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for implementing at least one function, instructions for implementing the various method embodiments described below, etc. The data storage area may also store data created during use by the graded early warning system 100 for vertigo diagnosis.

[0102] The above description is a detailed description of the preferred embodiments of the present invention. However, the embodiments are not intended to limit the scope of the patent application of the present invention. All equivalent changes or modifications made under the technical spirit of the present invention should fall within the patent scope covered by the present invention.

Claims

1. A graded early warning method for vertigo diagnosis, characterized in that, include: S1, receiving multimodal evidence data related to the subject of examination, the multimodal evidence data including consultation text information and examination-related data, the examination-related data including at least one of the following: eye-tracking data, nystagmus test data, hearing test data, vestibular function test data; S2, process the multimodal evidence data respectively to obtain the structured feature vector information of the multimodal evidence data, and generate quality indicators corresponding to the multimodal evidence data; S3, the structured feature vectors corresponding to the multimodal evidence data are arranged into an evidence event sequence according to the access order or a preset evidence access strategy; S4, perform fusion update one by one according to the order of the evidence event sequence, and output the risk assessment result of the corresponding stage after each fusion update, so as to obtain the risk assessment result of each stage; S5. Calculate the overall uncertainty index and the evidence conflict index based on the risk assessment results of each stage, and determine the uncertainty type according to the overall uncertainty index and the evidence conflict index. The uncertainty type includes at least the uncertainty of insufficient evidence and the uncertainty of conflicting evidence. S6. Based on the risk assessment results at each stage, determine the amount of risk change, then identify the key evidence stage that leads to the change in risk assessment results, and generate stage explanation information. S7, map the risk assessment results, uncertainty types and stage explanation information to graded early warning levels, and output risk warning information and / or risk handling suggestions corresponding to the graded early warning levels.

2. The graded early warning method for vertigo diagnosis according to claim 1, characterized in that, In step S2, the processing of the multimodal evidence data includes: S21, the consultation text information is input into a natural language processing model for word segmentation, entity recognition, semantic understanding and relation extraction, and text semantic features are extracted to obtain the structured feature vector information of the consultation text; S22, the inspection-related data is input into at least one artificial intelligence model for signal, image or sequence analysis, the inspection data features are extracted, and the structured feature vector information corresponding to the inspection-related data is obtained.

3. The graded early warning method for vertigo diagnosis according to claim 2, characterized in that, Also includes: S55. Based on the overall uncertainty index, the evidence conflict index, and the risk assessment results at each stage, determine the type of evidence relationship, which includes insufficient evidence, corroborating evidence, and conflicting evidence. Specifically, the evidentiary relationship is considered to be of the corroborative type when the risk assessment results of adjacent stages meet at least one of the following conditions: The support for risk assessment results in adjacent phases increases continuously in the same risk assessment direction; The overall uncertainty index decreases in the subsequent stage of an adjacent stage, while the evidence conflict index does not increase.

4. The graded early warning method for vertigo diagnosis according to claim 2, characterized in that, The consultation text information includes at least the chief complaint, onset characteristics, triggering factors, duration, accompanying symptoms, and past medical history; the structured feature vector of the consultation text includes at least symptom features, time features, and triggering relationships.

5. The graded early warning method for vertigo diagnosis according to claim 2, characterized in that, In step S2, generating quality indicators corresponding to the multimodal evidence data includes: For consultation text information, quality indicators are generated based on the completeness of the text content, the absence of key symptom fields, and the consistency of the description content. For checking relevant data, generate quality indicators based on at least one of the following: data clarity, proportion of effective data, identifiability of target area, and whether the collection time meets preset requirements.

6. The graded early warning method for vertigo diagnosis according to claim 1, characterized in that, In S5, the overall uncertainty index is used to characterize the stability or convergence of the comprehensive risk assessment results, including the fluctuation range of the risk assessment results in multiple stages and / or the degree of concentration of the risk assessment results in the final stage. The evidence conflict index is used to characterize the degree of disagreement in risk judgment between different stages or different evidence. The degree of disagreement includes the distribution distance of the current stage risk assessment result relative to the previous stage risk assessment result and / or the degree of deviation between the risk tendency corresponding to the current evidence event and the historical fusion result.

7. The graded early warning method for vertigo diagnosis according to claim 1, characterized in that, In S6, Determining the amount of risk change includes calculating the degree of difference in the distribution of risk assessment results between adjacent stages; The key evidence phase for determining changes in risk assessment results includes identifying the phase with the greatest change in risk as the key evidence phase.

8. The graded early warning method for vertigo diagnosis according to claim 2, characterized in that, In step S6, the stage explanation information includes at least a key evidence stage identifier and evidence event index information corresponding to the key evidence stage; The evidence event index information includes at least one of the following: text fragment index information of consultation text information, timestamp index information corresponding to examination-related data, image frame index information, and curve node index information.

9. The graded early warning method for vertigo diagnosis according to claim 1 or 2, characterized in that, In S4, performing a merge update one by one includes: S41, Maintain the historical fusion state, which includes at least the comprehensive risk assessment distribution of the previous stage, the uncertainty index and conflict index of the previous stage, and the set of indexes of evidence events that have been accessed. S42, For the currently accessed evidence event, determine the evidence quality weight based on its corresponding quality index; S43, the risk assessment distribution vector of the current evidence event and the comprehensive risk assessment distribution in the historical fusion state are weighted and synthesized according to the evidence quality weight to obtain the comprehensive risk assessment distribution of the current stage, which is used as the risk assessment result of the current stage.

10. A graded early warning system for vertigo diagnosis, characterized in that, include: A multimodal data acquisition module is used to acquire multimodal evidence data related to the examinee, wherein the multimodal evidence data includes at least consultation text information and at least one examination-related data; One or more processors, wherein the one or more processors are communicatively connected to the multimodal data acquisition module; Memory; One or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the one or more processors, the one or more applications being configured to perform the method as described in any one of claims 1-9.

Images