An old injury multi-modal auditing method and system for insurance claim settlement

Abstract

The application discloses an old injury multi-modal auditing method and system for insurance claim settlement, and aims at the technical difficulties of existing old injury auditing, such as dependence on manual work, low efficiency, strong subjectivity, information isolation, inability to mine implicit clues of 'unmentioned' injury conditions, and single processing mode. The application extracts and structures non-structured data such as medical reports by using OCR and medical fine-tuning large models; generates injury condition records by time-series correlation data with examination sites as dimensions; adopts mixed reasoning of deterministic rules and Bayesian models, and analyzes injury conditions by combining semantic consistency quantification comparison; and aggregates risk score grading to output explainable auditing conclusions. The application realizes automation and standardization of old injury auditing, improves auditing efficiency and accuracy, strengthens anti-fraud capability, and is suitable for multiple insurance claim settlement auditing scenes such as automobile insurance and accidental insurance.

Description

Technical Field

[0001] This invention relates to the fields of insurance claims, medical review, and artificial intelligence technology, specifically to a multimodal review method and system for old injuries in insurance claims. Background Technology

[0002] In the fields of insurance claims, forensic identification, and medical diagnosis, accurately distinguishing between new and old injuries is a crucial step. It directly affects the determination of accident liability, the assessment of insurance compensation amounts, and the fairness of claims results, and is both a key focus and a challenge in claims review.

[0003] Currently, the mainstream method for reviewing old injuries relies on manual work. Medical experts or claims reviewers manually read medical imaging reports such as X-rays, CT scans, and MRIs, as well as disability assessment reports, sorting out information and cross-referencing them one by one. This model has many drawbacks: First, it is inefficient and highly subjective, with the review heavily reliant on individual professional experience. Judgment standards are difficult to standardize, and oversights can easily occur due to fatigue when handling a large number of cases, resulting in inconsistent review quality. Second, the information is isolated and lacks correlation. Existing information systems only achieve electronic storage of reports and cannot automatically extract key information such as examination sites, injury descriptions, and examination times. Reviewers need to manually compare reports, which is time-consuming and labor-intensive. Third, it lacks automated analysis of temporal logic. There is no dedicated rule system to interpret the examination results of the same site at different times, especially ignoring the potential meaning of injuries "not mentioned" in the report, making it difficult to uncover hidden fraud or misdiagnosis clues. Fourth, the processing modality is singular. Most automated solutions are only adapted to structured text and cannot effectively extract unstructured text information from images and scanned documents, making it difficult to achieve end-to-end automated review.

[0004] In summary, existing technologies cannot meet the needs of reviewing massive numbers of claims cases. The manual mode is inefficient and prone to errors, while the automation solution is insufficient. The industry urgently needs an automatic, fast, objective, and accurate multimodal intelligent review technology for old injuries to improve review efficiency, reduce claims risks, and ensure the fairness of claims settlement. Summary of the Invention

[0005] To address these issues, this invention provides a multimodal review method and system for old injuries in insurance claims, solving the problems of existing old injury review methods that rely on manual labor, resulting in low efficiency and strong subjectivity; isolated information that cannot be deeply correlated with time-series data; difficulty in uncovering hidden clues such as "unmentioned" injuries; and single processing modality, resulting in low review accuracy and difficulty in adapting to the review needs of massive claims cases.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a multimodal review method for old injuries in insurance claims, comprising:

[0007] S1. Multimodal data acquisition: Receive multimodal unstructured data uploaded by users;

[0008] S2. Information Extraction and Structuring: Text information in the multimodal unstructured data is extracted using OCR technology, and semantic parsing of the text information is performed using a large language model fine-tuned with medical report data to output structured key information;

[0009] S3. Data Association and Organization: Using the examination site as the key dimension, the structured key information is sorted chronologically according to the examination date, and a corresponding ordered list of injury records is generated for each examination site.

[0010] S4. Intelligent rule engine reasoning: It adopts a hybrid reasoning architecture of deterministic rules and probabilistic models to analyze the ordered injury record list and output the judgment result of suspected old injury with confidence score.

[0011] S5. Comprehensive Risk Decision and Output: Aggregate the output results of the intelligent rule engine, calculate the comprehensive old injury risk score of the case, make hierarchical decisions based on the comprehensive old injury risk score, and generate an interpretable review conclusion containing key evidence chains.

[0012] As a preferred solution for multimodal review of old injuries for insurance claims, in S1, the multimodal unstructured data includes images of medical examination reports and images of disability assessment reports.

[0013] As a preferred solution for the multimodal review method of old injuries for insurance claims, in S2, the training process of the large language model includes: collecting and labeling medical report text fragments, constructing a training set containing the medical report text fragments and corresponding standard labels, and using the training set to perform targeted training and fine-tuning of the large language model;

[0014] The standard label includes the examination site, injury type, and judgment criteria;

[0015] The structured key information includes the inspection date, time of the incident, inspection site, injury type, and judgment basis;

[0016] The injury categories include old injuries, new injuries, no injuries, and injuries not mentioned.

[0017] As a preferred solution for the multimodal review method of old injuries for insurance claims, in S4, the deterministic rules include immediate high-risk rules and multi-report association analysis rules, and the probabilistic model is a Bayesian inference model.

[0018] As a preferred solution for the multimodal review method of old injuries for insurance claims, the immediate high-risk rule is as follows: the earliest examination result of the corresponding examination site within 15 days after the incident is judged. If "old injury" is explicitly mentioned, it is marked as a high-risk clue, and the initial confidence value is set to the highest.

[0019] As a preferred approach for multimodal review of old injuries in insurance claims, the multi-report correlation analysis rules include the following scenarios:

[0020] Scenario A: When all the ordered injury records for the same examination site are "not mentioned", the semantic consistency of the judgment basis corresponding to each record is judged. If the semantics are consistent, the Bayesian inference model is activated to calculate the posterior probability that there is an old injury in the corresponding examination site.

[0021] Scenario B: When the orderly injury records of the same examination site are mixed with "not mentioned" and "old injury", if the record with the earlier time is "not mentioned" and the record with the later time is "old injury", then the semantic consistency of the judgment basis corresponding to the two is judged. If the semantics are consistent, the conclusion of "old injury" is supported.

[0022] If the "old injury" is recorded first, then subsequent "unmentioned" records will not affect the existing conclusion;

[0023] Scenario C: When the ordered injury records for the same examination site contain "new injuries" or are all "old injuries", the corresponding judgment conclusion is output directly.

[0024] As a preferred solution for the multimodal review method of old injuries for insurance claims, the calculation process of the Bayesian inference model is as follows: Let event H be "old injury exists in the examined area", and event E be "N consecutive reports conclude 'not mentioned' and the judgment basis is highly consistent";

[0025] Calculating posterior probability based on Bayes' theorem :

[0026]

[0027] In the formula, For prior probability, Let be the likelihood probability. This represents the total probability.

[0028] The prior probability is obtained based on the statistical analysis of historical claims data, and the likelihood probability is obtained based on the statistical analysis of the probability of observing N consecutive "not mentioned" reports when there is an old injury.

[0029] When the posterior probability exceeds a preset risk threshold, it is determined to be a highly suspected old injury, and an alert is triggered.

[0030] As a preferred solution for multimodal review of old injuries for insurance claims, the semantic consistency judgment method is as follows: A Sentence-BERT pre-trained model is used to convert the judgment basis text into a fixed-length semantic vector. The cosine similarity between the two semantic vectors is calculated and mapped to the [0,1] interval. When the semantic similarity value... If the semantic consistency threshold τ is greater than or equal to the preset semantic consistency threshold, it is determined to be semantically consistent.

[0031] The semantic similarity The calculation formula is:

[0032]

[0033] In the formula, , These are the semantic vectors corresponding to the text used for the judgment.

[0034] As a preferred solution for multimodal review of old injuries for insurance claims, S5 describes the comprehensive old injury risk score. The calculation is performed using a linear weighted aggregation model, and the formula is as follows:

[0035]

[0036] In the formula, This is an indicator value indicating whether a high-risk rule has been triggered immediately. The posterior probability output by the Bayesian inference model; The semantic similarity value of the key comparison points; , , These are weighting coefficients calibrated using historical data;

[0037] Based on the comprehensive old injury risk score, cases are divided into low-risk, medium-risk, and high-risk categories. Low-risk cases are automatically approved, medium-risk cases are recommended for manual review, and high-risk cases are marked as cases requiring priority review.

[0038] This invention also provides a multimodal review system for old injuries in insurance claims, employing the aforementioned multimodal review method for old injuries in insurance claims, including:

[0039] Multimodal data acquisition unit: used to receive multimodal unstructured data uploaded by users;

[0040] Information extraction and structuring unit: used to extract text information from the multimodal unstructured data using OCR technology, perform semantic parsing on the text information using a large language model fine-tuned with medical report data, and output structured key information;

[0041] Data association and organization unit: used to sort the structured key information in chronological order according to the examination date, with the examination site as the key dimension, and generate a corresponding ordered list of injury records for each examination site;

[0042] Intelligent rule engine reasoning unit: Used to analyze the ordered injury record list using a hybrid reasoning architecture of deterministic rules and probabilistic models, and output the judgment result of suspected old injury with confidence score;

[0043] Comprehensive Risk Decision and Output Unit: This unit aggregates the output of the intelligent rule engine, calculates the comprehensive risk score of old injuries in a case, makes a graded decision based on the comprehensive risk score of old injuries, and generates an interpretable audit conclusion containing a key chain of evidence.

[0044] The present invention has the following advantages:

[0045] First, this invention is compatible with multimodal unstructured data such as medical examination report images, scanned copies, and disability assessment reports. It automatically extracts and structures key information through OCR and fine-tuning of a large language model, replacing the manual reading, manual input, and cross-comparison of each report. It can process a large number of claims in batches, increasing the review speed several times, significantly reducing manual input, and lowering the operating costs of insurance institutions.

[0046] Secondly, this invention abandons the subjective judgment method that relies on the personal experience of auditors and constructs a standardized and quantifiable audit rule system; it adopts deterministic rules and Bayesian probability model collaborative reasoning, combined with semantic consistency quantitative comparison technology, to accurately interpret various injury information, avoid misjudgments caused by human fatigue and experience differences, and ensure the objectivity and consistency of audit results.

[0047] Third, this invention takes the examination site as the main dimension, and integrates the scattered data from multiple reports according to the examination date and time sequence to generate an ordered injury record list for each site; it automatically completes the comparison and logical deduction of key information across reports, effectively explores the inherent relationship between the examination results of the same site at different times, and solves the pain points of traditional technology in terms of information fragmentation and inability to conduct in-depth correlation analysis.

[0048] Fourth, this invention innovatively defines rules for in-depth interpretation of "unmentioned" injury status in reports, and uses semantic consistency judgment to infer the suspected risk of old injuries, breaking through the limitation that it is difficult for humans to identify hidden fraud clues; it accurately captures abnormal information, promptly discovers false claims, excessive claims and other behaviors, effectively prevents the risk of claims fraud, and protects the legitimate rights and interests of insurance institutions and policyholders.

[0049] Fifth, this invention generates an audit report that includes triggering rules, confidence scores, semantic similarity, and a complete chain of key evidence, with transparent and traceable conclusions; it automatically classifies cases into three levels—low, medium, and high—based on comprehensive risk scores, and implements automatic approval, manual review, and key examination strategies respectively, balancing audit efficiency and risk control accuracy, and providing a reliable and implementable basis for claims decisions. Attached Figure Description

[0050] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0051] The structures, proportions, sizes, etc. illustrated in this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed herein, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0052] Figure 1 This is a schematic diagram of a multimodal review method for old injuries in insurance claims provided in Embodiment 1 of the present invention;

[0053] Figure 2 This is a technical roadmap for a multimodal review method for old injuries in insurance claims provided in Embodiment 1 of the present invention;

[0054] Figure 3 This is a schematic diagram of the architecture of a multimodal review system for old injuries for insurance claims provided in Embodiment 1 of the present invention. Detailed Implementation

[0055] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0056] Example 1

[0057] See Figure 1 and Figure 2 This invention provides a multimodal review method for old injuries in insurance claims, comprising:

[0058] S1. Receive multimodal unstructured data uploaded by users;

[0059] Specifically, in insurance claims scenarios, user-submitted review materials are mostly in unstructured image format, rather than standardized electronic text. This step, by supporting multimodal data input, breaks through the limitation of traditional automated systems that can only process structured text, covering all types of original materials in claims review and achieving full data access from the source. Unified acquisition of multimodal data is the foundation for subsequent automated processing, avoiding the tediousness and errors of manual data entry.

[0060] S2. Extract text information from the multimodal unstructured data using OCR technology, and perform semantic parsing on the text information using a large language model fine-tuned with medical report data to output structured key information;

[0061] Specifically, the first step is to use Optical Character Recognition (OCR) technology to convert pixel information in images into editable text characters, solving the problem of textualizing unstructured data. However, the text output by general OCR is unstructured natural language and cannot be directly used for logical reasoning. Therefore, a large language model, specifically fine-tuned for medical reports, is introduced. By learning from massive amounts of labeled medical report data, it masters medical terminology, report writing logic, and injury description standards. This model can accurately identify and extract key fields crucial for assessing old injuries, transforming unstructured text into structured data that computers can understand and compute, providing standardized input for subsequent rule-based reasoning.

[0062] S3. Using the examination site as the key dimension, sort the structured key information according to the examination date in chronological order, and generate a corresponding ordered injury record list for each examination site.

[0063] Specifically, chronic injuries are "previous injuries over time," and their assessment heavily relies on comparing changes in the condition of the same body part at different points in time. Current technology simply stores reports as independent documents, failing to establish connections between data points. This step uses "examination site" as the association key to aggregate information about the same site scattered across multiple reports and sorts it chronologically by examination time, constructing a complete timeline of injury for each site. This temporal data organization transforms isolated information points into a logically connected sequence, enabling subsequent temporal logical reasoning.

[0064] S4. Using a hybrid reasoning framework of deterministic rules and probabilistic models, the ordered injury record list is analyzed, and the result of the suspected old injury determination with confidence score is output.

[0065] Specifically, single deterministic rules cannot handle the ambiguity and uncertainty in medical reports, while pure probabilistic models lack interpretability and fail to meet the compliance requirements of insurance claims. Therefore, this invention adopts a hybrid reasoning architecture: for scenarios with clear medical evidence and industry consensus, deterministic rules are used to quickly reach conclusions, ensuring reasoning efficiency; for complex scenarios with ambiguity and requiring empirical judgment, a probabilistic model is introduced for quantitative evaluation, outputting results with confidence levels. The two work together to ensure system efficiency, improve judgment accuracy in complex scenarios, and preserve the interpretability of the reasoning process.

[0066] S5. Aggregate the output results of the intelligent rule engine, calculate the comprehensive old injury risk score of the case, make a graded decision based on the comprehensive old injury risk score, and generate an interpretable review conclusion containing key evidence chains.

[0067] Specifically, the intelligent rule engine may output multiple independent inference results, such as triggering multiple rules or calculating multiple confidence values. These scattered results need to be aggregated and analyzed to form an overall risk assessment of the entire case. Through a linear weighted aggregation model, risk factors of different dimensions are quantified into a unified comprehensive risk score, achieving standardized risk measurement. Tiered decision-making based on risk scores allows for the rational allocation of review resources: low-risk cases are automatically approved, improving efficiency; medium- and high-risk cases require manual intervention to ensure accuracy. Simultaneously, a complete chain of key evidence is output, making the review conclusions transparent and traceable, meeting the compliance requirements of insurance claims, and reducing claims disputes.

[0068] In one possible embodiment, in S1, the multimodal unstructured data includes images from medical examination reports and images from disability assessment reports.

[0069] Specifically, medical examination reports (X-rays, CT scans, MRIs, etc.) are the primary medical evidence for determining the nature of the injury, directly recording its imaging characteristics; disability assessment reports are legally binding professional evaluation documents, containing the final qualitative conclusion of the injury. These two types of documents are the most important and authoritative materials in the review of old injuries for insurance claims, covering the review needs of the vast majority of claims. Inputting these documents into the data type ensures that the system can make judgments based on the most reliable information sources.

[0070] In one possible embodiment, in S2, the training process of the large language model includes: collecting and labeling medical report text fragments, constructing a training set containing the medical report text fragments and corresponding standard labels, and using the training set to perform targeted training and fine-tuning of the large language model;

[0071] Specifically, general-purpose large language models lack expertise in the medical field and cannot accurately understand the technical terminology and specific expressions in medical reports. Therefore, it is necessary to construct a high-quality training set specifically for medical reports, where medical experts annotate the text fragments of the reports, clearly defining the examination site, injury category, and judgment basis for each fragment. Through fine-tuning, the large language model learns the unique writing style, terminology system, and logical structure of medical reports, significantly improving its semantic understanding ability and information extraction accuracy of medical texts, ensuring the reliability of the structured information output.

[0072] The standard label includes the examination site, injury type, and judgment criteria;

[0073] Specifically, these three types of labels are the main elements for judging old injuries. "Examination site" is the basis for data association; "injury category" is the main object of reasoning. This invention innovatively divides it into four categories: "old injury, new injury, no injury, and not mentioned," covering all possible injury descriptions in medical reports. In particular, treating "not mentioned" as an independent category lays the foundation for subsequent mining of hidden clues; "judgment basis" is used to verify the reliability of injury conclusions and is also the input for semantic consistency analysis.

[0074] The structured key information includes the inspection date, time of the incident, inspection site, injury type, and judgment basis;

[0075] Specifically, "Inspection Date" and "Accident Time" serve as the temporal basis for constructing the chronological logic, used to determine the temporal correlation between the injury and the insured event; "Inspection Location," "Injury Category," and "Judgment Basis" are the main information for injury assessment. These five fields comprehensively cover all the information required for reviewing old injuries, ensuring both the comprehensiveness of the information and avoiding redundant information from interfering with subsequent reasoning.

[0076] The injury categories include old injuries, new injuries, no injuries, and injuries not mentioned.

[0077] Specifically, traditional methods only focus on "old injuries" and "new injuries" explicitly mentioned in the report, ignoring the important condition of "not mentioned." This invention treats "not mentioned" as a separate injury category because in medical practice, doctors typically only describe abnormalities found during the current examination and may not proactively mention stable, old injuries. This classification method allows the system to capture "silent evidence" and uncover hidden risks of old injuries that traditional methods cannot detect.

[0078] In one possible embodiment, in S4, the deterministic rule includes an immediate high-risk rule and a multi-report association analysis rule, and the probabilistic model is a Bayesian inference model.

[0079] Specifically, the "immediate high-risk rule" targets the clearest and most direct clues about old injuries, enabling rapid identification of high-risk cases. The "multi-report correlation analysis rule" is a major innovation of this invention, defining inference logic for different combinations of injuries in scenarios involving temporal comparisons of multiple reports on the same injury site. Bayesian inference models excel at handling uncertainty, calculating the probability of a conclusion based on prior knowledge and observed evidence. This makes them suitable for interpreting ambiguous scenarios such as "unmentioned" injuries, providing quantitative confidence support for fuzzy judgments.

[0080] In one possible embodiment, the immediate high-risk rule is as follows: the earliest examination result of the corresponding examination site within 15 days after the incident is judged, and if "old injury" is explicitly mentioned, it is marked as a high-risk clue, and the initial confidence value is set to the highest.

[0081] Specifically, according to medical common sense, fresh injuries will have obvious imaging characteristics within 15 days after the injury, such as edema, bleeding, and clear fracture lines. Examinations conducted at this time can best reflect the true extent of the injury at the time of the accident. If the earliest examination report within this period clearly mentions "old injury," it means that the injury existed before the accident and is unrelated to the current insured event. Therefore, it has the highest credibility and should be directly marked as a high-risk clue with the highest priority.

[0082] In one possible embodiment, the multi-report association analysis rule includes the following scenarios:

[0083] Scenario A: When all the ordered injury records for the same examination site are "not mentioned", the semantic consistency of the judgment basis corresponding to each record is judged. If the semantics are consistent, the Bayesian inference model is activated to calculate the posterior probability that there is an old injury in the corresponding examination site.

[0084] Specifically, if the same area is not mentioned as abnormal in multiple examinations, there are two possibilities: either the area is indeed undamaged, or the area has a stable, old injury that the doctor deems unnecessary to describe. If the semantics of the "judgment basis" in multiple reports are highly consistent, it indicates that the doctor's observation angle and descriptive method are the same each time, making the probability of omission extremely low. In this case, "continuous non-mention" is more likely because the injury is old and stable. Therefore, a Bayesian inference model is activated to quantify the probability of the existence of an old injury based on historical data, thus enabling the interpretation of "silent evidence."

[0085] Scenario B: When the orderly injury records of the same examination site are mixed with "not mentioned" and "old injury", if the record with the earlier time is "not mentioned" and the record with the later time is "old injury", then the semantic consistency of the judgment basis corresponding to the two is judged. If the semantics are consistent, the conclusion of "old injury" is supported.

[0086] Specifically, if a report stating "not mentioned" appears first, followed by a report stating "old injury," there is a possibility of misjudgment or fabrication. In this case, it is necessary to compare the "judgment basis" of the two reports: if the semantics are consistent, it means that the objective imaging features observed in the two examinations are the same, and the doctor only determined it to be an old injury later, thus supporting the conclusion of "old injury"; if the descriptions are inconsistent, it means that the observation results of the two examinations are different, and the later conclusion of "old injury" is less credible.

[0087] If the "old injury" is recorded first, then subsequent "unmentioned" records will not affect the existing conclusion;

[0088] Specifically, chronic injuries are pre-existing, permanent, or long-term lesions. Once clearly diagnosed in an early report, their absence in subsequent reports does not mean the injury has disappeared; it simply means the doctor may deem it unnecessary to repeat the description. Therefore, the early conclusion of "chronic injury" has higher priority, and subsequent "not mentioned" records cannot overturn this conclusion.

[0089] Scenario C: When the ordered injury records for the same examination site contain "new injuries" or are all "old injuries", the corresponding judgment conclusion is output directly.

[0090] Specifically, if the report explicitly mentions "new injury," it indicates that the damage was caused by this accident; if all reports explicitly mention "old injury," it indicates that the damage existed before the accident. For these deterministic scenarios, there is no need for complex probabilistic reasoning; conclusions can be directly output through deterministic rules, which can significantly improve the system's operational efficiency.

[0091] In one possible embodiment, the calculation process of the Bayesian inference model is as follows: Let event H be "an old injury exists in the examined area", and event E be "N consecutive reports conclude 'not mentioned' and the judgment basis is highly consistent";

[0092] Calculating posterior probability based on Bayes' theorem :

[0093]

[0094] In the formula, For prior probability, Let be the likelihood probability. This represents the total probability.

[0095] Specifically, Bayes' theorem updates beliefs about the probability of event H based on observed evidence E. The prior probability P(H) is the base probability of "an old injury exists in a certain area" derived from historical claims data, reflecting the prevalence of this type of injury; the likelihood probability P(E|H) is the probability that "given the existence of an old injury in that area, N consecutive reports do not mention it and the judgment criteria are consistent," reflecting the industry practice of doctors not describing stable old injuries; the total probability... It represents the total probability of the observed evidence E under all possible circumstances. The posterior probability P(H|E), calculated using Bayes' theorem, is the exact probability that an old injury exists at that location after the current evidence is observed.

[0096] The prior probability is obtained based on the statistical analysis of historical claims data, and the likelihood probability is obtained based on the statistical analysis of the probability of observing N consecutive "not mentioned" reports when there is an old injury.

[0097] Specifically, the prior probability and likelihood probability are not arbitrarily set, but are derived through statistical analysis of massive amounts of historical claims data and medical reports, thus possessing a solid data foundation. As system data accumulates, these two probabilities can be continuously iterated and optimized, constantly improving the accuracy of the Bayesian inference model.

[0098] When the posterior probability exceeds a preset risk threshold, it is determined to be a highly suspected old injury, and an alert is triggered.

[0099] Specifically, the preset risk threshold is calibrated based on the insurance institution's risk preference and historical misjudgment rate. When the posterior probability exceeds this threshold, it indicates that the possibility of old injuries has reached a level requiring close human attention. The system triggers an alert, marking the case as high-risk, which is then reviewed by auditors, thus balancing the system's automated processing capabilities with risk control requirements.

[0100] In one possible embodiment, the method for determining semantic consistency is as follows: A Sentence-BERT pre-trained model is used to convert the judgment criteria text into a fixed-length semantic vector; the cosine similarity between the two semantic vectors is calculated and mapped to the [0,1] interval; when the semantic similarity value... If the semantic consistency threshold τ is greater than or equal to the preset semantic consistency threshold, it is determined to be semantically consistent.

[0101] Specifically, traditional text similarity calculation methods can only capture the similarity of surface words and cannot understand deep semantics. Sentence-BERT is a pre-trained model specifically optimized for sentence-level semantic representation, capable of mapping sentences with the same meaning but different expressions to similar positions in a high-dimensional space. Cosine similarity is used to measure the angle between two vectors in a high-dimensional space; the smaller the angle, the higher the similarity, which can accurately quantify the semantic consistency of two texts. Mapping similarity to the [0,1] interval and setting a threshold τ achieves standardization and automation of semantic consistency judgment.

[0102] The semantic similarity The calculation formula is:

[0103]

[0104] In the formula, , These are the semantic vectors corresponding to the text used for the judgment.

[0105] Specifically, this formula is the standard mathematical expression for cosine similarity. Wherein, It is the dot product of two vectors. Let |b|| and |b|| be the magnitudes of the two vectors, respectively. The cosine similarity value ranges from [-1, 1]. The closer the value is to 1, the more consistent the directions of the two vectors are, and the more similar the corresponding text semantics are; the closer the value is to -1, the more opposite the semantics are; and the closer the value is to 0, the more unrelated the semantics are.

[0106] In one possible embodiment, in S5, the comprehensive old injury risk score... The calculation is performed using a linear weighted aggregation model, and the formula is as follows:

[0107]

[0108] In the formula, This is an indicator value indicating whether a high-risk rule has been triggered immediately. The posterior probability output by the Bayesian inference model; The semantic similarity value of the key comparison points; , , These are weighting coefficients calibrated using historical data;

[0109] Specifically, This reflects the most direct high-risk clues, and the weighting Usually the largest; This reflects the probabilistic risk of latent, old injuries; This reflects the risks arising from semantic inconsistency. Weighting coefficients. , , The comprehensive risk score is derived by calibrating historical case data using machine learning, ensuring that it accurately reflects the true risk level of the case.

[0110] Based on the comprehensive old injury risk score, cases are divided into low-risk, medium-risk, and high-risk categories. Low-risk cases are automatically approved, medium-risk cases are recommended for manual review, and high-risk cases are marked as cases requiring priority review.

[0111] Specifically, for cases with extremely low risk, the system automatically approves them, significantly improving review efficiency; for medium-risk cases with some uncertainty, they are reviewed manually to ensure accuracy; and for high-risk cases, they are marked for priority review and conducted in-depth by senior reviewers to effectively prevent claims fraud. This tiered processing model maximizes the efficiency advantages of the automated system while ensuring review quality.

[0112] The application scenarios of this invention are as follows:

[0113] Intelligent review of old injuries in auto insurance personal injury claims: Applicable to the entire review process of personal injury claims for private cars and commercial vehicles, it is compatible with multimodal materials such as X-ray, CT, MRI medical imaging reports, disability assessment reports, and images and scans. After the materials are uploaded to the claims end, the system extracts text through OCR, fine-tunes the large model to parse semantics and completes the structured information output. It organizes injury records according to the time sequence of the examination site, uses a hybrid inference engine of deterministic rules and Bayesian to determine old and new injuries, and combines Sentence-BERT to complete the semantic comparison of the judgment basis. It outputs risk classification conclusions through a linear weighted model, meeting the needs of auto insurance claims to quickly identify old injuries, distinguish new injuries from accidents, prevent fraudulent claims, and improve the efficiency of batch case review.

[0114] Accurate Verification of Old Injuries in Personal Accident Insurance Claims: Applicable to personal accident insurance, group accident insurance, and other personal insurance claims review processes, adaptable to multimodal documents such as outpatient / inpatient medical examination reports and accidental injury disability assessment reports; after receiving the report images uploaded by the claimant, the system automatically extracts structured data such as examination date, accident time, and injury category, constructs a timeline of single-site injury, calls a dedicated rule engine to parse implicit injury clues such as "not mentioned," quantifies semantic consistency, calculates the confidence level of old injuries, and outputs interpretable graded review results, meeting the needs of excluding previous old injury payments in accident insurance claims, avoiding excessive claims, and ensuring the compliant use of insurance funds.

[0115] Assistance in the review of forensic injury assessment scenarios: Applicable to professional services such as injury assessment, past injury verification, and injury tracing in forensic appraisal institutions. It is compatible with scanned copies and electronic images of forensic appraisal reports, medical imaging diagnostic reports, and multiple follow-up examination reports. After the system uploads materials, it completes the structured extraction of multimodal information, correlates the results of multiple examinations of the same body part in a time sequence, analyzes the logic of injury changes through the main rule system, and combines a Bayesian model to quantify the probability of old injuries in fuzzy scenarios, outputting high-confidence qualitative reference conclusions. This meets the needs of forensic appraisal to reduce human subjective bias, improve the objectivity of distinguishing between new and old injuries, and strengthen the accuracy and credibility of appraisal conclusions.

[0116] Screening and review of old injuries for medical insurance reimbursement: Applicable to medical insurance center services such as expense reimbursement for insured persons, review of outpatient and inpatient expenses, and investigation of abnormal reimbursements. It is compatible with multimodal materials such as insured persons' previous physical examination reports, current medical imaging reports, and scanned copies of inpatient medical records. The system automatically extracts key information such as location, injury, and examination time, and sequentially links previous and current medical examination records. It uses a rule engine to screen for injuries related to old injuries, verifies the semantic consistency of report descriptions, and outputs risk classification results. This meets the medical insurance needs to prevent duplicate reimbursement of expenses for old injuries, standardize the use of medical insurance funds, and reduce unreasonable expenditures.

[0117] Intelligent screening of pre-existing injuries for health insurance underwriting: Applicable to health insurance and critical illness insurance underwriting, verification of past health status, and risk assessment of insurance applications. It is compatible with multimodal documents such as the insured's physical examination report, past medical imaging report, and outpatient diagnosis report. After receiving the materials uploaded by the underwriting end, the system extracts the injury and examination time information in a structured manner, sorts out the insured's past health records in chronological order, identifies old injuries and past medical history based on the rule engine, quantifies semantic similarity, and outputs a pre-existing injury risk classification report. This meets the needs of health insurance underwriting to quickly screen for past risks, provide objective basis for underwriting pricing, and reduce the risk of insurance claims disputes.

[0118] It should be noted that the method of this embodiment can be executed by a single device, such as a computer or server. The method of this embodiment can also be applied to a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method of this embodiment, and these multiple devices will interact with each other to complete the described multimodal review method for old injuries for insurance claims.

[0119] It should be noted that the above description describes some embodiments of this disclosure. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in a different order than that shown in the above embodiments and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0120] Example 2

[0121] See Figure 3 Embodiment 2 of the present invention also provides a multimodal review system for old injuries for insurance claims, which adopts the above-mentioned multimodal review method for old injuries for insurance claims, including:

[0122] Multimodal data acquisition unit 100: used to receive multimodal unstructured data uploaded by users;

[0123] Information extraction and structuring unit 200: used to extract text information from the multimodal unstructured data using OCR technology, perform semantic parsing on the text information using a large language model fine-tuned by medical report data, and output structured key information;

[0124] Data association and organization unit 300: Used to sort the structured key information in chronological order according to the examination date, with the examination site as the key dimension, and generate a corresponding ordered list of injury records for each examination site;

[0125] Intelligent rule engine reasoning unit 400: Used to analyze the ordered injury record list using a hybrid reasoning architecture of deterministic rules and probabilistic models, and output the judgment result of suspected old injury with confidence score;

[0126] Comprehensive Risk Decision and Output Unit 500: This unit aggregates the output of the intelligent rule engine, calculates the comprehensive old injury risk score of the case, makes a graded decision based on the comprehensive old injury risk score, and generates an interpretable audit conclusion containing a key chain of evidence.

[0127] It should be noted that the information interaction and execution process between the modules of the above-mentioned device are based on the same concept as the method embodiment in Embodiment 1 of this application, and the resulting technical effects are the same as those in the method embodiment of this application. For details, please refer to the description in the method embodiment shown above in this application, and it will not be repeated here.

[0128] Example 3

[0129] Embodiment 3 of the present invention provides a non-transitory computer-readable storage medium storing program code for a multimodal review method for old injuries in insurance claims. The program code includes instructions for executing the multimodal review method for old injuries in insurance claims according to Embodiment 1 or any possible implementation thereof.

[0130] Computer-readable storage media can be any available medium that a computer can access, or a data storage device such as a server or data center that integrates one or more available media. The available medium can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

[0131] Example 4

[0132] Embodiment 4 of the present invention provides an electronic device, including: a memory and a processor;

[0133] The processor and the memory communicate with each other via a bus; the memory stores program instructions that can be executed by the processor, and the processor can call the program instructions to execute a multimodal review method for old injuries for insurance claims, as described in Embodiment 1 or any possible implementation thereof.

[0134] Specifically, a processor can be implemented in hardware or software. When implemented in hardware, the processor can be a logic circuit, an integrated circuit, etc. When implemented in software, the processor can be a general-purpose processor that reads software code stored in memory. This memory can be integrated into the processor or located outside the processor and exist independently.

[0135] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means.

[0136] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0137] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A multimodal review method for old injuries in insurance claims, characterized in that, include: S1. Receive multimodal unstructured data uploaded by users; S2. Extract text information from the multimodal unstructured data using OCR technology, and perform semantic parsing on the text information using a large language model fine-tuned with medical report data to output structured key information; S3. Using the examination site as the key dimension, sort the structured key information according to the examination date in chronological order, and generate a corresponding ordered injury record list for each examination site. S4. Using a hybrid reasoning framework of deterministic rules and probabilistic models, the ordered injury record list is analyzed, and the result of the suspected old injury determination with confidence score is output. S5. Aggregate the output results of the intelligent rule engine, calculate the comprehensive old injury risk score of the case, make a graded decision based on the comprehensive old injury risk score, and generate an interpretable review conclusion containing key evidence chains.

2. The multimodal review method for old injuries for insurance claims as described in claim 1, characterized in that, In S1, the multimodal unstructured data includes images from medical examination reports and images from disability assessment reports.

3. The multimodal review method for old injuries for insurance claims as described in claim 1, characterized in that, In S2, the training process of the large language model includes: collecting and labeling medical report text fragments, constructing a training set containing the medical report text fragments and corresponding standard labels, and using the training set to perform targeted training and fine-tuning of the large language model; The standard label includes the examination site, injury type, and judgment criteria; The structured key information includes the inspection date, time of the incident, inspection site, injury type, and judgment basis; The injury categories include old injuries, new injuries, no injuries, and injuries not mentioned.

4. The multimodal review method for old injuries for insurance claims as described in claim 1, characterized in that, In S4, the deterministic rules include immediate high-risk rules and multi-report association analysis rules, and the probabilistic model is a Bayesian inference model.

5. The multimodal review method for old injuries for insurance claims as described in claim 4, characterized in that, The immediate high-risk rule is as follows: the earliest examination results of the corresponding examination site within 15 days after the incident are judged. If "old injury" is explicitly mentioned, it is marked as a high-risk clue, and the initial confidence value is set to the highest.

6. The multimodal review method for old injuries for insurance claims as described in claim 4, characterized in that, The multi-report association analysis rules include the following scenarios: Scenario A: When all the ordered injury records for the same examination site are "not mentioned", the semantic consistency of the judgment basis corresponding to each record is judged. If the semantics are consistent, the Bayesian inference model is activated to calculate the posterior probability that there is an old injury in the corresponding examination site. Scenario B: When the orderly injury records of the same examination site are mixed with "not mentioned" and "old injury", if the record with the earlier time is "not mentioned" and the record with the later time is "old injury", then the semantic consistency of the judgment basis corresponding to the two is judged. If the semantics are consistent, the conclusion of "old injury" is supported. If the "old injury" is recorded first, then subsequent "not mentioned" records will not affect the existing conclusion; Scenario C: When the ordered injury records for the same examination site contain "new injuries" or are all "old injuries", the corresponding judgment conclusion is output directly.

7. The multimodal review method for old injuries for insurance claims as described in claim 6, characterized in that, The calculation process of the Bayesian inference model is as follows: Let event H be "an old injury exists in the examined area", and event E be "N consecutive reports conclude 'not mentioned' and the judgment basis is highly consistent"; Calculating posterior probability based on Bayes' theorem : In the formula, For prior probability, Let be the likelihood probability. This represents the total probability. The prior probability is obtained based on the statistical analysis of historical claims data, and the likelihood probability is obtained based on the statistical analysis of the probability of observing N consecutive "not mentioned" reports when there is an old injury. When the posterior probability exceeds a preset risk threshold, it is determined to be a highly suspected old injury, and an alert is triggered.

8. The multimodal review method for old injuries for insurance claims as described in claim 6, characterized in that, The method for determining semantic consistency is as follows: A Sentence-BERT pre-trained model is used to convert the judgment criteria text into fixed-length semantic vectors. The cosine similarity between the two semantic vectors is calculated and mapped to the [0,1] interval. When the semantic similarity value... If the semantic consistency threshold τ is greater than or equal to the preset semantic consistency threshold, it is determined to be semantically consistent. The semantic similarity The calculation formula is: In the formula, , These are the semantic vectors corresponding to the text used for the judgment.

9. The multimodal review method for old injuries for insurance claims as described in claim 1, characterized in that, In S5, the comprehensive old injury risk score The calculation is performed using a linear weighted aggregation model, and the formula is as follows: In the formula, This is an indicator value indicating whether a high-risk rule has been triggered immediately. The posterior probability output by the Bayesian inference model; The semantic similarity value of the key comparison points; , , These are weighting coefficients calibrated using historical data; Based on the comprehensive old injury risk score, cases are divided into low-risk, medium-risk, and high-risk categories. Low-risk cases are automatically approved, medium-risk cases are recommended for manual review, and high-risk cases are marked as cases requiring priority review.

10. A multimodal review system for old injuries in insurance claims, employing the multimodal review method for old injuries in insurance claims as described in any one of claims 1-9, characterized in that, include: Multimodal data acquisition unit: used to receive multimodal unstructured data uploaded by users; Information extraction and structuring unit: used to extract text information from the multimodal unstructured data using OCR technology, perform semantic parsing on the text information using a large language model fine-tuned with medical report data, and output structured key information; Data association and organization unit: used to sort the structured key information in chronological order according to the examination date, with the examination site as the key dimension, and generate a corresponding ordered list of injury records for each examination site; Intelligent rule engine reasoning unit: Used to analyze the ordered injury record list using a hybrid reasoning architecture of deterministic rules and probabilistic models, and output the judgment result of suspected old injury with confidence score; Comprehensive Risk Decision and Output Unit: This unit aggregates the output of the intelligent rule engine, calculates the comprehensive risk score of old injuries in a case, makes a graded decision based on the comprehensive risk score of old injuries, and generates an interpretable audit conclusion containing a key chain of evidence.

Images