Traffic accident injury condition analysis and AIS scoring method based on large model

Abstract

The invention relates to the technical field of injury condition analysis and injury scoring, and relates to a traffic accident injury condition analysis and AIS scoring method based on a large model. Firstly, a traffic accident injury condition document is obtained. Then document processing is carried out after document type identification is carried out on the traffic accident injury condition document, and a digital document is determined. And then, based on the error correction process, performing error correction on the digital document. And then, inputting the digital document subjected to error correction into the large disassembly model to obtain each injury condition unit output by the large disassembly model. Then, for each injury unit, various AIS score information matching the injury unit is retrieved in the database. And then, according to various scoring information, AIS scoring knowledge is determined, AIS scoring is performed on the injury condition unit based on the AIS scoring knowledge, and a scoring result is determined. According to the scheme, the text report and the image can be processed in a unified mode, the problem that multi-source heterogeneous data is difficult to comprehensively utilize in a traditional method is solved, the independent injury condition unit is accurately extracted from a complex document, and accurate AIS scoring is carried out.

Description

A Method for Traffic Accident Injury Analysis and AIS Scoring Based on a Large Model Technical Field

[0001] This manual relates to the field of injury analysis and injury scoring technology, and in particular to a method for traffic accident injury analysis and AIS scoring based on a large model. Background Technology

[0002] With the deepening application of artificial intelligence technology in the medical field, AI-based trauma scoring systems have become important tools for assisting trauma diagnosis and assessment. Some existing systems employ prompt engineering techniques, using predefined prompt templates to guide large language models to automatically extract injury information from medical documents. However, this approach suffers from a series of fundamental drawbacks, limiting its accuracy, robustness, and practicality in clinical practice.

[0003] This technical approach heavily relies on manually designed prompt templates, requiring tedious debugging for various medical documents. It also struggles to effectively handle complex medical descriptions beyond the preset scope, leading to failures in extracting key information. General retrieval technologies fail to adapt to the complex structure of professional medical knowledge, resulting in insufficient medical accuracy in search results. Furthermore, it lacks the ability to process aggregated injury descriptions: it cannot automatically parse continuous injuries (e.g., "fractures of the 1st-3rd thoracic vertebrae"), quantify multiple injuries (e.g., "multiple rib fractures"), or effectively decompose and combine descriptions (e.g., "fractures of the cervical and lumbar vertebrae"), and it lacks effective algorithms and medical verification mechanisms for this purpose.

[0004] Therefore, this specification provides a method for traffic accident injury analysis and AIS scoring based on a large model. Summary of the Invention

[0005] This specification provides a method for traffic accident injury analysis and AIS scoring based on a large model, in order to partially solve the aforementioned problems existing in the prior art.

[0006] This specification adopts the following technical solution: This specification provides a method for traffic accident injury analysis and AIS scoring based on a large model, including: S1. Obtaining traffic accident injury documents, which include at least outpatient medical records, imaging examination reports, autopsy reports, and accident scene photos; S2. Identifying the document type of the traffic accident injury documents and processing them according to the identified document type to determine digital documents; S3. Correcting errors in the digital documents based on a preset progressive error correction process; S4. Inputting the corrected digital documents into a trained large-scale model to obtain each injury unit output by the large-scale model; S5. For each injury unit, retrieving various AIS scoring information matching the injury unit from preset databases; S6. Determining AIS scoring knowledge based on the various AIS scoring information and performing AIS scoring on the injury unit based on the AIS scoring knowledge matching the injury unit to determine the scoring result.

[0007] Based on the aforementioned technical means, this solution can uniformly process text reports (medical records, images, autopsies) and images (scene photos), solving the problem that traditional methods struggle to comprehensively utilize multi-source heterogeneous data, thus improving information integrity and the basis for evaluation. The pre-defined progressive error correction process effectively corrects typos and non-standard terminology in documents, improving the quality and consistency of input data and reducing subsequent analysis errors caused by data noise. Unlike general models, training specifically for injury decomposition tasks allows for more accurate extraction of independent "injury units" from complex documents, solving the difficulty of incomplete or ambiguous key information extraction, which is crucial for ensuring accurate subsequent scoring. By structurally storing the AIS scoring criteria in a database and determining scores through retrieval and matching rather than pure model generation, the standardization and traceability of the scoring process are ensured, and the risk of "illusion" is effectively controlled, resulting in more reliable results.

[0008] Furthermore, S2 specifically includes: identifying the document type of the traffic accident injury document, wherein the document type includes structured documents and unstructured documents; determining whether the document type of the traffic accident injury document is a structured document; if so, using document layout analysis technology to identify the standard field structure of the traffic accident injury document, extracting the structured traffic accident injury text content, and determining the digital document based on the extracted traffic accident injury text content; if not, using a document quality assessment algorithm to determine the document readability, and determining whether the traffic accident injury document needs to be manually reviewed based on the document readability; in the case of manual review, receiving the traffic accident injury text content returned after manual review, and determining the digital document based on the traffic accident injury text content returned after manual review.

[0009] Based on the aforementioned technical methods, the inefficiency of sending all documents to humans is avoided, as is the high error rate caused by forcing machines to process all documents. Optimal resource allocation is achieved: "machines handle the easy-to-process documents, humans handle the difficult ones," controlling labor costs while ensuring quality. Document type recognition and structured information extraction provide clean and well-organized text input for subsequent error correction (S3) and injury analysis (S4). For example, accurately extracting the "diagnosis" field from a structured medical record is far more reliable than having a large model guess from the entire medical record. The system can handle documents of varying quality in the real world. Whether it's clearly printed reports or illegible on-site shorthand notes, the system has corresponding processing channels and will not completely fail due to poor formatting of individual documents.

[0010] Furthermore, the digital document also includes an image quality assessment report; step S2 further includes S21: performing an image quality assessment on the accident scene photos in the traffic accident injury document, determining the image quality assessment result, wherein the image quality assessment includes assessments of sharpness, contrast, and noise level; performing image correction on the accident scene photos based on the image quality assessment result, and determining an image quality assessment report containing the corrected accident scene photos.

[0011] Based on the aforementioned technical methods, while on-site photographs do not directly show internal injuries, they can provide crucial clues (such as the direction of impact, signs of speed changes, and intrusion into the vehicle's interior). High-quality images can help large models more accurately infer the types of injuries that may occur (such as the probability and severity assessment of whiplash injuries to the neck in rear-end collisions).

[0012] Furthermore, S3 specifically includes: extracting character image features from the digital document and calculating the similarity between the character image features and a preset standard character template; determining whether the similarity reaches a preset threshold; if yes, determining that the character image features do not need to be corrected; if no, determining that the character image features need to be corrected, and replacing the character image features with the standard character template to achieve character shape correction.

[0013] According to the above technical means, for the character shape distortion caused by printing blurring, paper stains, scanning skew, and low resolution (such as the difficulty in distinguishing "已", "己", and "巳"; the confusion between "5" and "S"), this method has remarkable effects. It can directly repair the "bad characters" output by the OCR engine, preventing misreading of injury due to a single character difference at the root (such as misreading "肱骨骨折" as "肱骨骨拆"). Compared with the error correction of large models based on context semantics, image template matching is an operation with lightweight calculation, definite logic, and extremely fast speed. As the first line of defense in the error correction process, it can quickly filter out most low-level errors and reduce the burden of subsequent semantic error correction. This method is purely based on visual morphology and will not generate new semantic errors due to "understanding" the text content. For strings without context logic such as drug dosages and numbers, this shape-based error correction is often more reliable than semantic error correction.

[0014] Further, S3 specifically includes: determining the pronunciation of the characters in the digital document; according to the pronunciation of the characters in the digital document, matching numbered terms with a preset pronunciation similarity in a preset medical term dictionary; according to the context of the identified numbered terms in the medical term dictionary, matching it with the context of the characters in the digital document, and correcting the characters and their context in the digital document according to the matching result.

[0015] According to the above technical means, using the medical term dictionary as an authoritative reference system ensures the correct direction of error correction. Combining context matching makes the error correction decision no longer a simple "phonetically similar replacement", but an intelligent judgment based on the domain language model, greatly improving the accuracy. In Chinese medical texts, homophone errors caused by oral records, dialect accents, voice input, and mental sets are extremely common. This method targets the problem and can solve the problem of "correct in form, phonetically similar but wrong in meaning" that character image error correction cannot handle at all.

[0016] Further, S3 specifically includes: identifying the injury type text in the digital document; determining the injury site corresponding to the injury type text; determining various injury type descriptions related to the injury site in a preset medical knowledge base; judging whether there is an injury type description that matches the injury type text among the various injury type descriptions; if so, determining that the injury type text in the digital document is correct; if not, calculating the similarity between each injury type description and the injury type text respectively, and correcting the injury type text according to the calculated similarity.

[0017] Based on the aforementioned technical means, even if the original description is vague or non-standard (such as writing "the bone is broken"), the system can correct it to the standard medical term "fracture" through similarity calculation, and associate it with a specific bone, thus achieving information standardization and refinement. This demonstrates a common-sense judgment ability similar to that of a senior medical reviewer. This greatly enhances the professional authority of the entire AIS scoring system output, making it more likely to be accepted by clinical experts and forensic experts. This is a qualitative leap in complementing the first two error correction methods (characters, speech). It can identify and correct expressions that the former two are completely powerless to address and that are medically impossible. For example, mistakenly writing "skull fracture" as "skull inflammation" may have correct characters and pronunciation, but it is a serious medical logical error. This step can effectively correct this.

[0018] Furthermore, the method also includes step S7: among the various scoring information, determining the information of historical scoring cases matching the injury unit, and determining the scoring result of the historical scoring cases from the information of the historical scoring cases; determining whether the difference between the scoring result of the historical scoring cases and the scoring result of the injury unit is within a preset range; if yes, determining that the scoring result of the injury unit is normal; if no, providing feedback on the injury unit through a preset expert review interface, and displaying the information of the historical scoring cases matching the injury unit and the AIS scoring knowledge on the expert review interface, so that experts can refer to the information of the historical scoring cases and determine a new scoring result based on the AIS scoring; receiving the new scoring result and using the new scoring result as the scoring result of the injury unit.

[0019] Based on the aforementioned technical means, this is the last and most crucial line of defense against serious, systematic scoring errors in automated systems. When the AI's score differs significantly from historical consensus, the system automatically "applies the brakes," returning the decision-making power to human experts, fundamentally eliminating the risk of AI errors being directly applied to clinical or judicial practice. All cases marked as "differences exceeding the acceptable range" and their final rulings (whether upholding the AI ​​score or adopting the new expert score) are recorded. This forms a complete and traceable decision-making audit trail, which is crucial for meeting the requirements of medical quality management and judicial evidence compliance. Clearly demonstrating to end users that the system has a built-in expert oversight mechanism will greatly reduce their resistance to "black box AI" and increase the acceptability of the system's output.

[0020] This specification provides a device for traffic accident injury analysis and AIS scoring based on a large model, comprising: an acquisition module for acquiring traffic accident injury documents, wherein the traffic accident injury documents include at least outpatient medical records, imaging examination reports, autopsy reports, and accident scene photos; an identification module for identifying the document type of the traffic accident injury documents and processing the documents according to the identified document type to determine digital documents; an error correction module for correcting errors in the digital documents based on a preset progressive error correction process; a decomposition module for inputting the corrected digital documents into a trained decomposition large model to obtain each injury unit output by the decomposition large model; a retrieval module for retrieving various AIS scoring information matching each injury unit from preset databases; and a scoring module for determining AIS scoring knowledge based on the various AIS scoring information and performing AIS scoring on the injury unit based on the AIS scoring knowledge matching the injury unit to determine the scoring result.

[0021] This specification provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method for traffic accident injury analysis and AIS scoring based on a large model.

[0022] This specification provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement a method for traffic accident injury analysis and AIS scoring based on a large model.

[0023] The above-mentioned technical solutions adopted in this specification achieve the following beneficial effects: This solution can uniformly process text reports (medical records, images, autopsies) and images (scene photos), solving the problem that traditional methods are difficult to comprehensively utilize multi-source heterogeneous data, and improving information integrity and assessment basis. The preset progressive error correction process can effectively correct typos, non-standard terminology, etc. in documents, improve the quality and consistency of input data, and reduce subsequent analysis errors caused by data noise. Unlike general models, training for injury decomposition tasks can more accurately extract independent "injury units" from complex documents, solving the difficulty of incomplete or ambiguous key information extraction, which is the core to ensure the accuracy of subsequent scoring. By storing the AIS scoring criteria in a structured database and determining the score through retrieval matching rather than pure model generation, the standardization and traceability of the scoring process are ensured, and the risk of "illusion" is effectively controlled, resulting in more reliable results. Attached Figure Description

[0024] The accompanying drawings, which are included to provide a further understanding of this specification and form part of this specification, illustrate exemplary embodiments and their descriptions, serving to explain this specification and not constituting an undue limitation thereof. In the drawings: Figure 1 is a flowchart illustrating an automated construction method for data analysis based on a large model, as provided in an embodiment of this specification; Figure 2 is a schematic diagram illustrating traffic accident injury analysis and AIS scoring based on a large model, as provided in this specification. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this specification clearer, the technical solutions of this specification will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in this specification without creative effort are within the scope of protection of this application.

[0026] In embodiments of this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0027] The approach of using predefined prompt templates to guide large language models in automatically extracting injury information from medical documents has several fundamental drawbacks: First, its adaptability heavily relies on manually designed prompt templates. These templates require extensive and tedious manual adjustments for different types of medical documents, such as outpatient medical records, imaging reports, and autopsy reports, resulting in a huge engineering workload. More importantly, when the injury description uses natural language expressions or complex medical sentences that exceed the scope of the preset templates, the system often fails to understand and process them effectively, leading to the failure to extract key information.

[0028] Secondly, this technical approach typically mandates the output of structured data in specific formats, such as JSON. This rigid constraint on structured output severely limits the system's ability to adapt to complex, ambiguous, or implicit medical expressions in natural language. When faced with injury descriptions containing numerous technical terms, nested logic, and non-standard grammatical structures, forced formatting parsing can easily lead to information loss, semantic distortion, or incorrect associations.

[0029] Third, existing methods often employ general vector retrieval or text matching techniques, which fail to deeply adapt to the complex medical knowledge structure inherent in professional trauma scoring standards such as AIS2015. The AIS standards have strict hierarchical anatomical relationships and injury type constraints, while general retrieval techniques cannot understand, for example, the distinction between "injuries of different severity levels in the same anatomical location," nor can they verify whether the injury description conforms to basic human anatomy, resulting in insufficient medical accuracy of the retrieval results.

[0030] Fourth, in the fundamental stage of injury information extraction, the accuracy of general optical character recognition (OCR) technology in recognizing medical terminology is relatively low. Tests show that it has a high error rate in recognizing key anatomical terms and injury description terms such as "thoracic vertebrae 1-3" and "compression fracture," with an overall accuracy rate typically below 85%. This fundamentally affects the reliability of all subsequent analyses and scoring.

[0031] Of particular concern is the limitations of existing technologies in handling common "aggregated descriptions" in injury descriptions. For example, when faced with a continuous injury description like "compression fracture of the 1st to 3rd thoracic vertebrae," the system cannot automatically identify that it represents three independent injury instances and accurately resolve the numbering boundaries (e.g., "1-3") and connection patterns. For general descriptions like "multiple rib fractures," it lacks the ability to determine the specific number and location of fractures and to complete the quantitative supplementation and differential assessment of injury information. For expressions like "cervical and lumbar vertebrae fractures," which are composed of conjunctions (e.g., "and"), the system lacks an effective algorithm for decomposing the boundaries and cannot accurately separate them into two independent injury concepts for processing. Furthermore, the decomposed results lack a medical rationality verification mechanism based on anatomical constraints and the Abbreviated Injury Scale (AIS) scoring rules.

[0032] Finally, existing technologies face significant challenges in the representation and retrieval of medical knowledge. The AIS2015 standard contains thousands of scoring rules with complex logical dependencies, exhibiting a deeply hierarchical knowledge structure and multidimensional constraints. General knowledge graph and graph retrieval technologies struggle to efficiently and accurately represent and retrieve this type of specialized medical knowledge, exhibiting structural defects in representation. In terms of efficiency, traditional retrieval algorithms are slow to respond to large-scale medical knowledge graphs containing tens of thousands of nodes, failing to meet the real-time requirements of scoring within 10 seconds in scenarios such as traffic accident handling. Furthermore, their excessive demands on memory and computing resources restrict the practical deployment and application of the system.

[0033] Therefore, this specification provides a method for traffic accident injury analysis and AIS scoring based on a large model to solve the above problems.

[0034] The technical solutions provided in the various embodiments of this specification are described in detail below with reference to the accompanying drawings.

[0035] Figure 1 is a flowchart illustrating a method for traffic accident injury analysis and AIS scoring based on a large model provided in an embodiment of this specification, including the following steps: S1: Obtain traffic accident injury documents, which include at least outpatient medical records, imaging examination reports, autopsy reports, and accident scene photos.

[0036] This specification describes the process of performing traffic accident injury analysis and AIS scoring based on a large model. In the embodiments described herein, this process can be executed by a server. However, this specification does not limit the type of device or platform used to perform this large-model-based traffic accident injury analysis and AIS scoring process; for example, a personal computer, mobile terminal, or other such device or platform can also be used. For ease of description, the following description uses a server as the executing entity.

[0037] In one or more embodiments of this specification, the server can obtain traffic accident injury documents, which include at least outpatient medical records, imaging examination reports, autopsy reports, accident scene photos, and other original traffic accident data in different formats.

[0038] S2: The traffic accident injury document is identified by document type, and document processing is performed on the traffic accident injury document according to the identified document type to determine the digital document.

[0039] In one or more embodiments of this specification, since the traffic accident injury documents obtained by the server contain original traffic accident data in different formats, the server can identify the document type of the traffic accident injury documents and perform document processing on the traffic accident injury documents according to the identified document type to determine the digital document.

[0040] The traffic accident injury documents are subjected to document type identification to determine their document type, which includes structured documents and unstructured documents. Specifically, the server can determine whether the traffic accident injury document is a structured document.

[0041] If the traffic accident injury documents contain structured documents such as outpatient medical records, imaging reports, and autopsy reports, the server can use document layout analysis technology to identify the standard field structure of the traffic accident injury documents and extract the structured text content. For example, for outpatient medical records containing "chief complaint: headache, nausea, vomiting," document layout analysis technology can automatically extract structured patient symptom information such as "headache, nausea, vomiting." Then, based on the extracted traffic accident injury text content, a pre-defined standardized digital document is determined.

[0042] If not, the server can use a document quality assessment algorithm to determine document readability and, based on readability, determine whether the traffic accident injury document needs manual review. If readability is low (e.g., below a preset readability threshold), manual review is required, whereby the human reviewer identifies the traffic accident injury text content in documents with low readability (e.g., messy handwritten medical records). In the case of manual review, the server receives the traffic accident injury text content returned after manual review and determines a preset standardized digital document based on this content.

[0043] Furthermore, in this specification, the server can also process the document types of outpatient medical records, imaging examination reports, and autopsy reports in the traffic accident injury documents, extract the text content of the traffic accident injury, and determine the standardized digital document by combining the text content of each traffic accident injury.

[0044] Specifically, the server can identify the document type of each document in the traffic accident injury documentation system, including structured and unstructured documents. The documents in the traffic accident injury documentation system include outpatient medical records, imaging reports, and autopsy reports. Then, for each document in the traffic accident injury documentation system, the server can determine whether the document type is a structured document. If so, document layout analysis technology is used to identify the standard field structure of the traffic accident injury documentation document, extract the structured traffic accident injury text content, and determine the digital document based on the extracted text content. If not, a document quality assessment algorithm is used to determine the document readability, and the readability determines whether the traffic accident injury documentation document needs to be manually reviewed. In the case of manual review, the server receives the manually reviewed traffic accident injury text content and determines the digital document based on the manually reviewed text content.

[0045] It is worth noting that if the document quality assessment algorithm determines that the document has high readability (readability is lower than the preset readability threshold), then the document can be considered to have a clear structure and structured features. Document layout analysis technology can be used to identify standard field structures, extract structured traffic accident injury text content, and determine the digital document based on the extracted traffic accident injury text content.

[0046] In one or more embodiments of this specification, the digital document may also include an image quality assessment report.

[0047] Step S2 further includes S21, whereby the server performs image quality assessment on the accident scene photos in the traffic accident injury document, determines the image quality assessment result, and the image quality assessment includes evaluations of sharpness, contrast, and noise level. Then, based on the image quality assessment result, the server performs image correction on the accident scene photos, and determines an image quality assessment report containing the corrected accident scene photos.

[0048] For example, image quality assessment is performed on poorly captured accident scene photos, including metrics such as sharpness, contrast, and noise level. For blurry, tilted (i.e., low-sharpness) car accident scene photos, automatic image tilt correction and super-resolution reconstruction are performed. For photos with uneven lighting, an adaptive contrast enhancement algorithm is applied for correction.

[0049] S3: Based on a preset progressive error correction process, perform error correction on the digital document.

[0050] In one or more embodiments of this specification, since digital documents may contain medical text or other content that may be misidentified, the server can correct the identified digital documents based on a preset progressive error correction process. The preset progressive error correction process may include character shape correction, medical terminology correction, and contextual semantic correction.

[0051] Specifically, the server can extract character image features from digital documents, that is, the image features of each text character in the digital document, and calculate the similarity between the character image features and a preset standard character template. It then determines whether the similarity reaches a preset threshold. If yes, the character image features do not need correction. If no, the character image features need correction, and the standard character template replaces the character image features to achieve character shape correction. For example, if "thoracic vertebra" is misidentified as "chest who," the server extracts the character image features of the incorrect character "chest who," matches it with the standard character template, calculates the similarity, and corrects it to the correct term "thoracic vertebra."

[0052] In this specification, the server can also determine the pronunciation of characters in a digital document. Based on the pronunciation of the characters in the digital document, it matches numbered terms with a preset pronunciation similarity value from a pre-defined medical terminology dictionary. Based on the context of the identified numbered terms in the medical terminology dictionary, it matches the context of the characters in the digital document, and corrects the characters and their context based on the matching results. For example, in the case where "C6" is incorrectly identified as "6", based on the context of related numbered terms such as cervical spine numbered terms identified in the medical terminology dictionary, and combined with pronunciation similarity analysis, it automatically corrects it to "C6".

[0053] In this specification, the server can also identify injury type text in digital documents. It then determines the injury site corresponding to the injury type text. Next, it identifies various injury type descriptions related to the injury site in a pre-defined medical knowledge base. The server can then determine if any injury type description matches the injury type text. If so, the server determines the injury type text in the digital document is correct. If not, the server calculates the similarity between each injury type description and the injury type text, and corrects the injury type text based on the calculated similarity. Correction can be based on the injury type description with the highest similarity. For example, for a statement like "liver fracture," which does not conform to medical common sense, the server automatically corrects it to a description from the medical knowledge base based on the mismatch between the injury site and injury type.

[0054] Of course, any one of the three error correction methods mentioned above can be used to correct digital documents, or any two of them can be used, or even all three methods can be used sequentially in a progressive manner.

[0055] S4: Input the corrected digital document into the trained large-scale decomposition model to obtain the injury units output by the large-scale decomposition model.

[0056] In one or more embodiments of this specification, the server can input the corrected digital document into a trained large-scale injury model to obtain the individual injury units output by the large-scale injury model. Through this specially trained large-scale injury model, the digital document can be accurately decomposed into different independent injury units.

[0057] For example, for a text description of “compression fracture of the 1st to 3rd thoracic vertebrae” in a digital document, the decomposition model can identify the numerical range “1-3” and resolve it into three independent compression fractures of the 1st, 2nd, and 3rd thoracic vertebrae (equivalent to decomposing the compression fracture of the 1st to 3rd thoracic vertebrae into three different injury units), and verify its anatomical continuity on the human anatomy knowledge graph.

[0058] For example, for a digital document containing a vague text description of "multiple rib fractures", the large model can be broken down to extract the specific number and location information of the fractures, such as calculating a total of 5 different fractures from "fractures of the 3rd to 5th ribs on the left side and fractures of the 7th and 9th ribs on the right side".

[0059] Alternatively, for a digital document containing a relatively complex text description of “compression fracture of the 1st-2nd thoracic vertebrae and the 4th lumbar vertebra with spinal cord injury”, the decomposition model can process continuous descriptions, parallel relationships and complication descriptions in layers, breaking them down into multiple independent injury units.

[0060] In this specification, each injury unit can be listed as an independent injury unit list, and each injury unit contains data such as the specific anatomical location, injury type, severity assessment, and confidence score.

[0061] In this specification, the training method for the large-scale decomposition model can be as follows: Historical digital documents are acquired as training data, and the decomposition results of these historical documents are used as annotations. The historical training documents are input into the large-scale decomposition model to be trained, yielding the output decomposition results. The loss value is determined based on the difference between the annotations and the output decomposition results of the large-scale decomposition model. The large-scale decomposition model is trained with the minimum loss value as the optimization objective, where the loss value and the difference are positively correlated.

[0062] S5: For each injury unit, retrieve various AIS score information that matches the injury unit from the preset databases.

[0063] S6: Based on the various AIS scoring information, determine the AIS scoring knowledge, and perform AIS scoring on the injury unit based on the AIS scoring knowledge that matches the injury unit, and determine the scoring result.

[0064] In one or more embodiments of this specification, for each injury unit, the server can retrieve various AIS score information matching the injury unit from preset databases.

[0065] The server can employ a hierarchical retrieval strategy to search for relevant knowledge in the AIS-specific knowledge graph, vector database, and expert case library. This includes AIS standard knowledge, historical cases, expert experience, and original injury data, among other AIS scoring information matching the injury unit. Then, based on a multi-source information fusion algorithm using a large model, the server strictly adheres to the AIS 2015 standard to perform scoring calculations and generates detailed scoring criteria and confidence assessment methods to determine the AIS scoring knowledge matching the injury unit.

[0066] For example, for a single injury, "compression fracture of the second thoracic vertebra," a base score of 3 is calculated based on the AIS scoring knowledge summarized from the relevant AIS scoring rules, and the final score is adjusted according to the degree of injury. When a patient has multiple injuries, the score for each injury is calculated separately, and then a comprehensive score is calculated according to the AIS standards (i.e., AIS scoring knowledge) to ensure the accuracy and consistency of the scoring results.

[0067] Of course, the server can also calculate a confidence score based on the consistency of multi-source information for the scoring results. When the information is sufficient and consistent, a high confidence score is given, and when the information is insufficient or there are conflicts, it is marked as requiring manual review.

[0068] Finally, detailed AIS scoring results can be generated, which may include specific scores for each injury, overall scores, detailed reasoning processes, analysis of key influencing factors, and confidence interval estimates.

[0069] Furthermore, in one or more embodiments of this specification, step S7 is included: the server determines information on historical scoring cases matching the injury unit from various scoring information, and determines the scoring results of the historical scoring cases from the information on the historical scoring cases. It is then determined whether the difference between the scoring results of the historical scoring cases and the scoring results of the injury unit is within a preset range.

[0070] If so, then the scoring result of that injury unit is considered normal.

[0071] If not, the injury unit will be fed back through a preset expert review interface. This interface will display information about historical scoring cases and AIS scoring knowledge matched with that injury unit, allowing experts to refer to this information and determine a new scoring result based on the AIS scores. The new scoring result will then be received and used as the final score for that injury unit.

[0072] Therefore, the system can automatically identify scoring results that significantly differ from historical cases, such as scores that are marked significantly higher or lower than similar cases, and mark them as requiring expert review. It provides experts with an intelligent review interface, displaying the scoring basis, similar cases, and key factor analysis, supporting experts in quickly confirming or modifying the scoring results. Furthermore, it can automatically integrate expert feedback and new experience into the knowledge base, continuously optimizing and improving the accuracy of the scoring algorithm. Finally, it generates a final confirmed AIS scoring report, including the final score reviewed by experts, a complete interpretability analysis, system performance evaluation, and a knowledge base update log.

[0073] In this manual, the large-scale model can be trained based on specialized large-scale language models (such as GPT-4, ChatGLM-Med, etc.). This large-scale language model enables context-aware analysis, allowing the model to understand the true meaning of medical terms within specific contexts. Multimodal information fusion combines text, images, and structured data for comprehensive analysis. Medical knowledge reasoning is based on a built-in medical knowledge graph for logical reasoning. Adaptive learning capabilities continuously optimize recognition accuracy based on new medical cases.

[0074] Large models can be precisely decomposed using the attention mechanism of the Transformer architecture to achieve self-attention computation: recognizing the semantic relationships between elements (such as text segmentation) in injury description text of digital documents. The expression is:

[0075] It can be used to calculate the attention weights between words in injury description text, identifying which words are most important for the current analysis. For example, it can be used to identify the association between the numerical range and anatomical location in "compression fracture of the 1st-3rd thoracic vertebrae," providing a semantic basis for subsequent decomposition. Q is the query matrix, representing the target word vector representation that needs to be focused on. K is the key matrix, representing the context word vector representation used for matching. V is the value matrix, representing the information vector representation that needs to be weighted and aggregated. is the dimension of the key vector, used to scale the dot product to prevent gradient vanishing.

[0076] It also designed a cross-attention mechanism: linking the medical knowledge base and the AIS standard.

[0077]

[0078] The system correlates and matches injury descriptions with a medical knowledge base to extract relevant AIS scoring criteria and medical knowledge. For example, it can match "thoracic vertebral compression fracture" with the AIS scoring rules for thoracic vertebral fractures, providing a standardized basis for scoring. : The query matrix of the input text, representing the vector representation of the injury description. The key matrix of the knowledge base stores vector indexes of medical knowledge and AIS standards. The knowledge base's value matrix stores specific medical knowledge and scoring rules. : Key vector dimension, a scaling factor used for cross-modal attention computation.

[0079] The expression to convert injury description text into a list of independent injury units, i.e., individual injury units, is:

[0080] The input aggregated injury description text (such as "compression fracture of the 1st to 3rd thoracic vertebrae") is progressively converted into multiple independent injury units. : The injury unit token that is output at any time. : All injury unit sequences generated before the specified time. : The aggregated description text input. : The hidden state vector of the time-decoder. The weight matrix of the output layer maps the hidden states to the vocabulary dimension. : The bias vector of the output layer.

[0081] The expression for calculating the confidence level is:

[0082] This tool is used to assess the reliability and accuracy of the breakdown results of aggregated injury descriptions. It allows for the evaluation of the accuracy of the breakdown results after "multiple rib fractures" is broken down into specific fracture locations, determining whether manual review is required. : The original confidence score of the current disassembly result. : No. The confidence score of each candidate decomposition result. : The total number of candidate decomposition results. Denominator: The sum of the fractional exponents of all candidate results, used for normalization.

[0083] In one or more embodiments of this specification, during the error correction and standardization process of digital documents, medical terms can first be converted into high-dimensional vector representations using the BERT model, and context enhancement can be performed using a self-attention mechanism. Then, the semantic similarity between OCR erroneous terms and standard terms is calculated based on cosine similarity, and associative reasoning is performed by combining knowledge graph node information propagation. Finally, confidence is assessed by comprehensively evaluating multi-dimensional similarity scores, and character-level accurate error correction is achieved through an n-gram language model, medical rationality verification, and Bayesian posterior probability calculation, selecting the optimal error correction result. This technology achieves a 96.7% recognition rate for medical professional terms and an overall OCR error correction accuracy of 94.2% even under adverse conditions such as handheld photography, inability to select angles, distances, and lighting.

[0084] For example, this technology can be applied to situations where traffic police officers take handheld photos of traffic accident injury diagnosis reports in their offices. Due to limitations in shooting angle, distance, and lighting conditions, errors such as misidentifying "thoracic vertebral compression fracture" as "thoracic vertebral compression osteophyte" can occur. This technology can automatically correct these errors to standard medical terminology.

[0085] The expression for converting medical terms into high-dimensional vector representations using the BERT model is as follows:

[0086] in, : BERT embedding vector representation of medical terminology. : A BERT model specifically fine-tuned for the medical field. : The string of medical terms entered. The total number of tokens (lexical units) contained in a term. : No. An embedding vector for each token.

[0087] The expression for context enhancement using self-attention is:

[0088] in, : Term vector representation after context enhancement. : The term vector from the original BERT output.

[0089] The expression for calculating the semantic similarity between OCR error terms and standard terms is:

[0090] in, : Vector representation of query terms (error terms identified by OCR). : Vector representation of candidate standard terms. : Semantic similarity score, ranging from [-1, 1].

[0091] The expression for information propagation among knowledge graph nodes is:

[0092] This formula is used to propagate and aggregate node information in a medical knowledge graph using graph neural networks, and to update the vector representation of medical terms through multi-hop reasoning, thereby enhancing the ability to reason about the associations between terms. : The hidden state of node i in the k-th iteration. : The weight matrix and bias vector of the k-th layer. : Neighbor node aggregation functions (such as mean, max, sum). : The set of neighboring nodes of node i. : Non-linear activation function.

[0093] The expression for confidence assessment based on multi-dimensional similarity scores is as follows:

[0094] This formula is used to calculate the overall confidence score of medical term mapping by integrating three dimensions: semantic similarity, structural similarity, and contextual similarity, providing a quantitative basis for error correction decisions. : Final mapping confidence score, in the range [0, 1]. Semantic similarity score. : Structural similarity score. Context similarity score. Weight parameters, satisfying .

[0095] The expression for the n-gram language model is:

[0096] This formula is used to calculate the conditional probability of a target word under given context based on an n-gram language model, providing linguistic constraints and rationality verification for character-level error correction. : given before Under the condition of the word, the first The word is The probability of. From the first To the The context sequence of each word. Sequences in the training corpus Number of times it appears. : All possible words.

[0097] The expression for calculating the Bayesian posterior probability is:

[0098] This formula is used to calculate the posterior probability of candidate characters based on Bayes' theorem. Combining image feature likelihood and contextual prior probability, it provides a probabilistic theoretical basis for character-level error correction. :character The posterior probability. :character Generate image features The likelihood probability. :character The prior probability. Character image features recognized by OCR. : Text context information. : The set of all candidate characters.

[0099] The expression for medical rationality verification is:

[0100] This formula is used to score the medical rationality of candidate characters using a set of multidimensional constraint rules, ensuring that the error correction results meet the norms and requirements of the medical field. :character The medical rationality score. : A set of medical constraint rules. : No. The weight of each constraint rule. : No. The set of valid characters under each constraint rule. : Characters to be verified.

[0101] The expression for selecting the optimal error correction result is:

[0102] This formula is used to combine Bayesian posterior probability and medical rationality score to select the optimal character as the error correction result, thereby achieving a unified decision-making process that combines probabilistic reasoning with medical constraints. : The optimal character to select. : Find the parameter that maximizes the function value.

[0103] In one or more embodiments of this specification, a large-model-driven intelligent architecture can be employed to achieve intelligent and efficient AIS knowledge representation through deep semantic understanding and knowledge reasoning capabilities. The algorithm first calculates the probability of medical entity recognition using the entity classification layer of the BERT model, and then evaluates the confidence level of relationships between entities using neural network transformations and the sigmoid activation function. Next, it calculates the comprehensive confidence level of triples by integrating entity recognition confidence, relationship extraction confidence, and contextual consistency, and dynamically allocates expert knowledge weights based on conflicts between standard knowledge and expert opinions. Finally, it integrates multiple expert opinions and standard knowledge through a weighted voting mechanism to achieve balanced multi-expert judgment and construct a 6-layer hierarchical knowledge graph specifically for the AIS2015 standard. This technology can handle the multi-layered and complex structure of human anatomy knowledge, supports medical entity relationship reasoning and expert experience fusion, and provides a high-quality knowledge foundation for AIS scoring.

[0104] The expression for calculating the probability of medical entity recognition is:

[0105] This formula is used to calculate the probability distribution of a text token belonging to a specific medical entity type through the entity classification layer of the BERT model, thereby achieving accurate identification and classification of medical entities. :time The token belongs to the entity. The probability of. : No. Types of physical entities (such as anatomical locations, injury types, etc.). : The current position of the token being processed. BERT in The hidden state vector output at each time step. : Weight matrix of entity classification layer. : Bias vector of entity classification layer.

[0106] The expression for assessing the confidence level of relationships between entities is:

[0107] This formula is used to calculate the confidence score of a specific relationship between two medical entities. It achieves a quantitative assessment of the relationship between entities through neural network transformation and sigmoid activation function. :entity With entity There is a relationship The confidence level. Two medical entities (e.g., "thoracic vertebra" and "compression fracture"). : Relationship type (such as "location-damage", "symptom-disease", etc.). : Vector representation of an entity. :relation The transformation matrix. :relation The bias term. : sigmoid activation function.

[0108] The expression for calculating the overall confidence score of a triple is:

[0109] This formula is used to calculate the comprehensive confidence score of knowledge triples by integrating three dimensions: entity recognition confidence, relation extraction confidence, and contextual consistency, thus providing quality assurance for knowledge graph construction. : Overall confidence level of knowledge triples. Confidence score for entity recognition. : Confidence score of relation extraction. Context consistency score. Weight parameters, satisfying .

[0110] The expression for assigning expert knowledge weights is:

[0111] This formula is used to dynamically allocate the weight of expert knowledge based on the conflict between standard knowledge and expert opinions and the expert confidence level, so as to achieve the optimal integration of standardized knowledge and expert experience. : Weighting of expert knowledge. : Conflict detection function between standard knowledge and expert opinions. Knowledge of AIS standards. Expert opinion. Confidence level of expert judgment.

[0112] The expression for multi-expert balanced decision-making is:

[0113] This formula is used to integrate the opinions of multiple experts and standard knowledge through a weighted voting mechanism, so as to achieve a balanced decision by multiple experts and ensure the scientific nature and reliability of the decision results. Final verdict : Weighted voting function. The total number of experts involved in the decision-making process. : No. The weight of each expert. : No. The opinion of an expert. Weight of standard knowledge. AIS standard knowledge content. Establish a multi-level index structure: node name index, relationship type index, and attribute value index. Implement a B+ tree-based path index, supporting fast hierarchical traversal and path lookup. Set up a distributed storage architecture to support horizontal scaling of large-scale knowledge graphs.

[0114] In one or more embodiments of this specification, a hierarchical medical terminology dictionary system can be constructed. The system includes three sub-databases: an anatomical terminology database, an injury terminology database, and an examination terminology database. Each sub-database uses a hash table data structure to store terminology mapping relationships, with the key being the original terminology string and the value being a standardized terminology string. The system employs a prefix tree index structure to achieve fast terminology lookup. Each node in the prefix tree includes a character mapping table, a set of child node pointers, and a Boolean flag indicating whether it is a terminal node. The dictionary construction process includes loading 50,000 anatomical terms, 30,000 injury terms, and 20,000 examination terms, and establishing synonym conversion relationships through terminology standardization mapping. The technical performance indicators of the dictionary system include a total of 100,000 terms, a lookup response time of less than 1 millisecond, a memory usage of 200MB, and an index construction time of 30 minutes.

[0115] This specification also allows for the construction of a data structure for an AIS-specific knowledge graph. This graph includes five node types: body region node set, anatomical location node set, injury type node set, severity node set, and AIS-encoded node set. Each node contains data fields such as a node identifier, node type label, and attribute value dictionary. The edge set of the knowledge graph represents the relationships between nodes, including hierarchical relationship edges, association relationship edges, constraint relationship edges, and dependency relationship edges. The hierarchical relationship construction process defines a six-layer anatomical hierarchy, including layer 1 (body region), layer 2 (organ system), layer 3 (organ), layer 4 (anatomical location), layer 5 (sub-location), and layer 6 (specific location). The technical performance indicators of the graph database include a total of 50,000 nodes, a total of 200,000 edges, a query response time of less than 2 seconds, a memory usage of 4GB, and an index hit rate of 95%.

[0116] Furthermore, this specification describes an innovative expert-verified knowledge fusion mechanism for the knowledge base construction, comprising two levels of data sources: AIS2015 standard foundational data, which collects the original AIS2015 standard text, official interpretation documents, professional diagnostic standards, autopsy report specifications, historical scoring cases, and other professional materials. The foundational data undergoes standardized preprocessing, including format unification, content structuring, and terminology standardization, to ensure the accuracy and completeness of the basic knowledge.

[0117] Expert Verification Knowledge Base: This base collects practical application experience from experts through an expert verification mechanism, forming a criterion system with higher priority than the AIS2015 standard text. Expert verification knowledge includes: criteria for judging specific injury expressions, methods for distinguishing similar rule entries, strategies for handling special injury descriptions, and criteria for judging boundary situations. This expert knowledge is transformed into computable criterion rules through structured processing.

[0118] Furthermore, it enables knowledge fusion and priority management, organically integrating basic standard knowledge with expert-verified knowledge to establish a dynamic priority management mechanism. Expert-verified knowledge has higher application priority; when there is a conflict between basic standards and expert criteria, the expert criteria take precedence. Simultaneously, a knowledge version management and conflict resolution mechanism is established to ensure the consistency and accuracy of the knowledge base.

[0119] In one or more embodiments of this specification, a multi-head attention mechanism and Transformer architecture are employed to achieve intelligent fusion of multi-source information such as AIS standard knowledge, expert experience, and historical cases. The algorithm first aggregates information from multiple perspectives by concatenating the outputs of multiple attention heads and performing linear transformations. Then, it calculates the output of a single attention head by applying the attention mechanism after linear transformations of the input query, key, and value. Next, it dynamically calculates the weights of each information source based on reliability and timeliness, and fuses the feature representations of multiple information sources through weighted summation. Finally, it quantifies the degree of conflict between different information sources by calculating the product of similarity and consistency. Dynamic weight allocation and conflict detection ensure the logical consistency of the fusion result, providing a comprehensive information foundation for subsequent scoring.

[0120] The expression for concatenating multiple attention heads is:

[0121] This formula is used to aggregate and fuse information from multiple perspectives by splicing the outputs of multiple attention heads and performing a linear transformation. : The number of attention heads (usually 8 or 12). : No. The output of each attention head. : Output transformation matrix. : The input query, key, and value matrix. : splicing operation.

[0122] The expression for calculating the output of a single attention head is:

[0123] This formula is used to calculate the output of the i-th attention head, which applies the attention mechanism after linearly transforming the input query, key, and value. : No. The query, key, and value transformation matrix for each head. Attention calculation function.

[0124] The expression for calculating the weights of each information source is:

[0125] This formula is used to dynamically calculate the weights of each information source based on reliability and timeliness, and softmax normalization is used to ensure that the sum of the weights is 1. : No. Normalized weights of each information source. : No. Reliability score (0-1) for each information source. : No. Timeliness weight of each information source. : Adjustment coefficient for reliability and timeliness. Total number of information sources.

[0126] The expression for the feature representations that integrate multiple information sources is:

[0127] This formula is used to fuse feature representations from multiple information sources through a weighted summation method, with the weights determined by a dynamic allocation mechanism. : The fused feature vector representation. : Query vector (current injury description). : No. The feature vectors of each information source.

[0128] The expression for quantifying the degree of conflict between different information sources is:

[0129] This formula is used to detect the degree of conflict between different information sources, quantifying the level of conflict by calculating the product of similarity and consistency. Information conflict score, range [0, 1]. Similarity calculation function. Information consistency score.

[0130] Furthermore, a standardized scoring rule engine is constructed, strictly adhering to the AIS2015 standard to execute the scoring logic. First, based on three dimensions—damage severity, anatomical location importance, and dam type complexity—the final AIS score is determined through maximum computation. Then, the overall confidence level of the scoring rule matching is calculated by multiplying the probabilities of all conditions being met. Finally, the consistency of multiple scoring results is verified by calculating the coefficient of variation. By comprehensively considering damage severity, anatomical location importance, and dam type complexity, an accurate AIS score is generated, and confidence assessment and consistency verification are provided.

[0131] The expression for the AIS score is:

[0132] This formula is used to determine the final AIS score by maximizing the calculation based on three dimensions: the severity of the injury, the importance of the anatomical location, and the complexity of the injury type. Final AIS score. AIS severity levels (1-6). Severity weighting function. Importance weight of anatomical sites.

[0133] The expression for the overall confidence level of the damage type complexity factor is:

[0134] This formula is used to calculate the overall confidence level of the scoring rule matching, and evaluates the reliability of the rule matching by multiplying the probabilities of all conditions being met. : Rule matching confidence, range [0, 1]. The total number of conditions in the scoring rules. : No. Conditions (such as "fracture severity > 50%"). : Input evidence of injury. Given the evidence, the first The probability that a condition is met.

[0135] The expression for verifying consistency is:

[0136] This formula is used to verify the consistency of multiple scoring results, and the stability of the scoring is assessed by calculating the coefficient of variation. : Scoring consistency score, range [0, 1]. : Variance function. Mean function. : No. The results of the second rating. Number of times the rating was repeated.

[0137] The traffic accident injury analysis and AIS scoring method based on a large model, as shown in Figure 1, can uniformly process text reports (medical records, images, autopsies) and images (scene photos), solving the problem that traditional methods struggle to comprehensively utilize multi-source heterogeneous data, thus improving information integrity and the basis for assessment. A pre-defined progressive error correction process effectively corrects typos and non-standard terminology in documents, improving the quality and consistency of input data and reducing subsequent analysis errors caused by data noise. Unlike general models, this method is trained specifically for injury decomposition tasks, enabling more accurate extraction of independent "injury units" from complex documents, solving the difficulty of incomplete or ambiguous key information extraction, and is crucial for ensuring accurate subsequent scoring. By storing the AIS scoring criteria in a structured database and determining scores through retrieval and matching rather than pure model generation, the standardization and traceability of the scoring process are ensured, and the risk of "illusion" is effectively controlled, resulting in more reliable results.

[0138] Based on one or more embodiments of this specification, a method for traffic accident injury analysis and AIS scoring based on a large model is provided. Following the same idea, this specification also provides a corresponding device for traffic accident injury analysis and AIS scoring based on a large model, as shown in Figure 2.

[0139] Figure 2 is a schematic diagram of a traffic accident injury analysis and AIS scoring based on a large model provided in this specification. Specifically, it includes: an acquisition module 200 for acquiring traffic accident injury documents, which at least include outpatient medical records, imaging reports, autopsy reports, and accident scene photos; an identification module 202 for identifying the document type of the traffic accident injury documents and processing them according to the identified document type to determine the digital document; an error correction module 204 for correcting the digital document based on a preset progressive error correction process; a decomposition module 206 for inputting the corrected digital document into a trained decomposition large model to obtain each injury unit output by the decomposition large model; a retrieval module 208 for retrieving various AIS scoring information matching each injury unit from preset databases; and a scoring module 210 for determining AIS scoring knowledge based on the various AIS scoring information and performing AIS scoring on the injury unit based on the AIS scoring knowledge matching it, thus determining the scoring result.

[0140] This specification also includes, as in Embodiment Two, a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method for traffic accident injury analysis and AIS scoring based on a large model, as described in Embodiment One.

[0141] This specification also includes, in Embodiment 3, an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described method for traffic accident injury analysis and AIS scoring based on a large model.

[0142] The above description is merely an embodiment of this specification and is not intended to limit this specification. Various modifications and variations can be made to this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this specification should be included within the scope of the claims of this specification.

Claims

1. A method for traffic accident injury analysis and AIS scoring based on a large model, characterized in that, include: S1. Obtain traffic accident injury documents, which include at least outpatient medical records, imaging examination reports, autopsy reports, and accident scene photos; S2. The traffic accident injury document is identified by document type, and the document is processed according to the identified document type to determine the digital document; S3. Based on a preset progressive error correction process, perform error correction on the digital document; S4. Input the corrected digital document into the trained large-scale decomposition model to obtain each injury unit output by the large-scale decomposition model; S5. For each injury unit, retrieve various AIS scoring information that matches the injury unit from the preset databases; S6. Determine AIS scoring knowledge based on the various AIS scoring information, and perform AIS scoring on the injury unit based on the AIS scoring knowledge that matches the injury unit to determine the scoring result.

2. The method for traffic accident injury analysis and AIS scoring based on a large model as described in claim 1, characterized in that, S2 specifically includes: identifying the document type of the traffic accident injury document, wherein the document type includes structured documents and unstructured documents; determining whether the document type of the traffic accident injury document is a structured document; if so, using document layout analysis technology to identify the standard field structure of the traffic accident injury document, extracting the structured traffic accident injury text content, and determining the digital document based on the extracted traffic accident injury text content; if not, using a document quality assessment algorithm to determine the document readability, and determining whether the traffic accident injury document needs to be manually reviewed based on the document readability; in the case of manual review, receiving the traffic accident injury text content returned after manual review, and determining the digital document based on the traffic accident injury text content returned after manual review.

3. The method for traffic accident injury analysis and AIS scoring based on a large model as described in claim 2, characterized in that, The digital document also includes an image quality assessment report; step S2 further includes S21: performing an image quality assessment on the accident scene photos in the traffic accident injury document, determining the image quality assessment result, the image quality assessment including assessments of sharpness, contrast, and noise level; performing image correction on the accident scene photos based on the image quality assessment result, and determining an image quality assessment report containing the corrected accident scene photos.

4. The method for traffic accident injury analysis and AIS scoring based on a large model as described in claim 1, characterized in that, S3 specifically includes: extracting character image features from the digital document and calculating the similarity between the character image features and a preset standard character template; determining whether the similarity reaches a preset threshold; if yes, determining that the character image features do not need to be corrected; if no, determining that the character image features need to be corrected and replacing the character image features with the standard character template to achieve character shape correction.

5. The method for traffic accident injury analysis and AIS scoring based on a large model as described in claim 1, characterized in that, S3 specifically includes: determining the pronunciation of the characters in the digital document; matching numbered terms with a preset pronunciation similarity value from a preset medical terminology dictionary based on the pronunciation of the characters in the digital document; matching the context of the identified numbered terms in the medical terminology dictionary with the context of the characters in the digital document; and correcting the characters and their context in the digital document based on the matching result.

6. The method for traffic accident injury analysis and AIS scoring based on a large model as described in claim 1, characterized in that, S3 specifically includes: identifying injury type text in the digital document; determining the injury site corresponding to the injury type text; determining various injury type descriptions related to the injury site in a preset medical knowledge base; determining whether there is an injury type description in each injury type description that matches the injury type text; if so, determining that the injury type text in the digital document is correct; if not, calculating the similarity between each injury type description and the injury type text respectively, and correcting the injury type text according to the calculated similarity.

7. The method for traffic accident injury analysis and AIS scoring based on a large model as described in claim 1, characterized in that, The method further includes step S7: among the various scoring information, determining the information of historical scoring cases matching the injury unit, and determining the scoring result of the historical scoring cases from the information of the historical scoring cases; determining whether the difference between the scoring result of the historical scoring cases and the scoring result of the injury unit is within a preset range; if yes, determining that the scoring result of the injury unit is normal; if no, providing feedback on the injury unit through a preset expert review interface, and displaying the information of the historical scoring cases matching the injury unit and the AIS scoring knowledge on the expert review interface, so that experts can refer to the information of the historical scoring cases and determine a new scoring result based on the AIS scoring; receiving the new scoring result and using the new scoring result as the scoring result of the injury unit.

8. The device for traffic accident injury analysis and AIS scoring based on a large model as described in claim 1, characterized in that, include: The acquisition module is used to acquire traffic accident injury documents, which include at least outpatient medical records, imaging examination reports, autopsy reports, and accident scene photos; the identification module is used to identify the document type of the traffic accident injury documents, and to process the documents according to the identified document type to determine the digital documents. The error correction module is used to correct errors in the digital document based on a preset progressive error correction process; The decomposition module is used to input the corrected digital document into the trained decomposition model to obtain the injury units output by the decomposition model; the retrieval module is used to retrieve various AIS scoring information matching the injury unit from preset databases for each injury unit; the scoring module is used to determine AIS scoring knowledge based on the various AIS scoring information, and to perform AIS scoring on the injury unit based on the AIS scoring knowledge matching the injury unit to determine the scoring result.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the method described in any one of claims 1 to 7.

10. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in any one of claims 1 to 7.

Images