Training method, scoring method and scoring system for trauma medical scoring model

Abstract

The invention discloses a trauma medical scoring model training method, a scoring method and a scoring system. The trauma medical scoring model training method comprises the following steps: deconstructing scoring rules to construct a structured rule library; acquiring case data marked with wound information and scores; respectively finely adjusting the information extraction model and the medical scoring model based on the same base model; constructing a difference model, comparing model output with expert annotation, generating less-evaluation, multi-evaluation and false-evaluation items and difference tags, and reversely optimizing the main model; a series information extraction and medical scoring model and a parallel difference model form a combined model, and closed-loop training is realized by dynamically adjusting the weight of a joint loss function. According to the method, a complex scoring task is decomposed into three stages of information understanding, rule application and deviation correction, the scoring efficiency and accuracy are remarkably improved, and an efficient and interpretable intelligent solution is provided for optimal distribution of trauma medical resources.

Description

Technical Field

[0001] The present invention relates to a trauma medical scoring model training method, and also relates to a trauma medical scoring method and a scoring system based on the trauma medical scoring model, belonging to the technical field of data processing. Background Art

[0002] Medical scoring systems (such as APACHE II, TRISS, AIS-ISS, GCS, NHISS, etc.) are important tools for assessing the overall severity of a patient's condition. These systems score different manifestations from different angles and different parts of the body based on the patient's subjective and objective observations and calculate the total score to assess the severity of the condition.

[0003] There are many medical scoring systems in the prior art. The first is rule-based medical scoring systems, which attempt to automate the processing and scoring of medical record data through computer algorithms. These systems usually rely on fixed medical rules or templates, searching for specific diagnostic information in medical record data and automatically calculating the trauma score according to the matching rules.

[0004] The second type is a medical question-answering system based on a large language model, which trains a general medical question-answering model through medical-related corpus. Although such a system can handle medical-related issues, it can only be used for general medical consultation documents and is difficult to cope with highly specialized and detailed scenarios such as medical scoring.

[0005] The third type is an automated scoring system based on the open interface call of a commercial large model, which cannot solve the problem of medical data security. And because the commercial large model is a black box, it can only be adapted to the medical scoring scenario through external tuning, but cannot be further optimized from the perspective of internal structure and parameters, and cannot meet the requirements of medical interpretability, so it is difficult to promote. Summary of the invention

[0006] The primary technical problem to be solved by the present invention is to provide a trauma medical scoring model training method.

[0007] Another technical problem to be solved by the present invention is to provide a trauma medical scoring method based on the trauma medical scoring model.

[0008] Another technical problem to be solved by the present invention is to provide a trauma medical scoring system based on the trauma medical scoring model.

[0009] In order to achieve the above technical objectives, the present invention adopts the following technical solutions: According to a first aspect of an embodiment of the present invention, a trauma medical scoring model training method is provided, comprising the following steps: Step 1: Deconstruct and simplify the scoring rules, expand them into multiple items that do not overlap and have no omissions, and build a scoring rule library; Step 2: Obtain basic cases, each of which is labeled with trauma-related information and scores; Step 3: Using the basic cases, different fine-tuning methods are used to tune the same base model, and pre-train the information extraction model and the medical scoring model; Step 4: construct a differential model based on the information extraction model and the medical scoring model; the differential model generates differential information, including a list of under-rated items, a list of over-rated items, a list of mis-rated items, and difference labels, by comparing the output of the medical scoring model with the expert annotation results, and triggers back propagation to optimize the main model parameters; Step 5: Connect the information extraction model and the medical scoring model in series and in parallel with the differential model to build a combined model; the output of the differential model dynamically adjusts the model weights through the joint loss function to form a closed-loop training mechanism; Step 6: Divide the basic cases into a sample set and a validation set, use the sample set to jointly tune the combined model, and use the validation set to verify it, so as to obtain a trauma medical scoring model.

[0010] Preferably, step 4 includes the following sub-steps: (41) Based on the medical scoring model trained in step 3, perform inference generation model scoring on all cases; (42) The model scores were corrected by experts to construct a differential result dataset, in which the under-rated, over-rated, and mis-rated items of each case were annotated; (43) aligning the trauma-related information dataset output by the information extraction model with the scoring information dataset output by the medical scoring model to construct a structured differential input dataset; (44) selecting a second base model independent of the first base model for quantization processing, and selecting a fine-tuning method through comparative experiments based on the differential input data set and the differential result data set; (45) Using the fine-tuning method selected in step (44), with the differential input data set as input and the differential result data set as output, the hyperparameters are optimized to achieve optimal model performance, and finally the differential model is trained to obtain the differential model.

[0011] Preferably, the information extraction model and the medical scoring model are based on the same first base model and are constructed by different fine-tuning methods; the differential model is based on an independent second base model and is constructed by a fine-tuning method different from the aforementioned models.

[0012] Preferably, the output of the information extraction model serves as the input of the medical scoring model, and the output of the medical scoring model is integrated with the trauma-related information and differential information judged by experts into a differential input data set, which serves as the input of the differential model.

[0013] Preferably, in step 3, the pre-trained information extraction model and medical scoring model are quantized to reduce memory usage and adapt to the hospital's local server deployment.

[0014] Preferably, in step 4, the difference label is used to indicate the difference type between the difference input data set of the difference model and the difference result data set of the difference model.

[0015] Preferably, in the case of multiple reviews, the difference information includes at least one list of multiple review items, wherein each multiple review item includes: 1) a difference label; 2) a description of the injury corresponding to the difference; In the case of few reviews, the difference information includes at least one list of few review items, wherein each of the few review items includes: 1) a difference label; 2) a description of the injury corresponding to the difference; In the case of mis-evaluation, the difference information includes at least one list of mis-evaluation items, wherein each mis-evaluation item includes: 1) a difference label; 2) a correct rule corresponding to the difference; 3) an incorrect rule corresponding to the difference.

[0016] Preferably, in step 6, the verification process of the verification set includes experts correcting the output results of the differential model to form a new differential result data set, and iteratively optimizing the parameters of the combined model.

[0017] According to a second aspect of an embodiment of the present invention, a trauma medical scoring method is provided, wherein a trauma medical scoring model obtained by the aforementioned trauma medical scoring model training method is used to score the trauma of the wounded to optimize the allocation of medical resources.

[0018] According to a third aspect of an embodiment of the present invention, there is provided a trauma medical scoring system, comprising a processor and a memory, wherein the processor and the memory are coupled; wherein the memory is used to store a computer program; and the processor runs the computer program stored in the memory to implement the aforementioned trauma medical scoring method.

[0019] Compared with the prior art, the present invention has the following technical effects: First, through model combination and step-by-step fine-tuning, the subjectivity and inefficiency of traditional manual scoring are solved, while avoiding the defects of traditional large language models in terms of security, cost and controllability, providing an efficient, accurate and feasible intelligent medical scoring solution. Secondly, although the information extraction model and the medical scoring model are based on the same base model, they are optimized through different fine-tuning methods, decomposing the complex medical scoring tasks into two orthogonal dimensions of "information understanding" and "rule application", realizing professional model division of labor, and effectively solving the knowledge loss and task conflict problems of general large language models in professional scenarios. Thirdly, after quantization, the obtained trauma medical scoring model can reduce the computing power requirements, adapt to the local server deployment of the hospital, and ensure data security and privacy protection. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is a flow chart of a trauma medical scoring model training method in the first embodiment of the present invention; Figure 2 This is a logical structure diagram of a trauma medical scoring model training method in the first embodiment of the present invention; Figure 3 This is a typical medical record example diagram in the first embodiment of the present invention; Figure 4 This is a schematic diagram of the logical structure of the medical scoring model in the first embodiment of the present invention; Figure 5 This is a schematic diagram of the logical structure of the differential model in the first embodiment of the present invention; Figure 6 This is an example diagram of a multi-evaluation situation in the first embodiment of the present invention; Figure 7 This is an example diagram of a case where there are few comments in the first embodiment of the present invention; Figure 8 This is an example diagram of a misjudgment situation in the first embodiment of the present invention; Fig. 9 FIG. 4 is a schematic diagram of the structure of a trauma medical scoring system in the third embodiment of the present invention. DETAILED DESCRIPTION

[0021] The technical content of the present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] The technical concept of the embodiment of the present invention is to decompose the complex medical scoring task into three orthogonal dimensions: information understanding, rule application, and deviation correction, which are processed by the information extraction model (MR), the medical scoring model (MS), and the differential model (MD) respectively. The information extraction model focuses on extracting structured relevant trauma information from unstructured medical records; the medical scoring model maps this information to the deconstructed scoring rule library to generate scores; the differential model generates differential labels to optimize the model by comparing the trauma description keywords with the scoring rules. In the data processing process, the scoring rules are deconstructed into structured items to reduce the computational complexity; by comparing the expert correction data with the model output, the model weights are dynamically adjusted, and the parameters of the information extraction model and the medical scoring model are reversely optimized; quantitative and localized deployment is adopted to ensure data security and reduce costs. The embodiment of the present invention improves accuracy and efficiency through model division of labor, uses differential models to output differential labels to enhance interpretability, and takes into account both economy and safety, providing an efficient, reliable and feasible intelligent solution for trauma medical assessment.

[0023] First embodiment like Figure 1 and Figure 2 As shown, the first embodiment of the present invention provides a training method for a trauma medical scoring model. The trauma medical scoring model can be used to quickly evaluate trauma, thereby optimizing the allocation of medical resources. For example, at a large emergency scene (such as a car accident scene), the trauma medical scoring model can automatically complete injury analysis and scoring based on trauma-related information provided by on-site rescue personnel, and recommend the most appropriate medical resources based on the scoring results.

[0024] In one embodiment of the present invention, the training method comprises at least the following steps: Step 1: Deconstruct and simplify the scoring rules, expand them into multiple items that do not overlap and have no omissions, and build a scoring rule library.

[0025] It should be noted that although the training method provided in the embodiment of the present invention takes trauma information as an example, it is not limited to such information. Those skilled in the art can understand that the training method is also applicable to other medical information. This embodiment can adopt the 2005 ISS (Injury Severity Score) rule. The ISS rule divides injuries into 6 major areas according to body regions, and each major area is further divided into 43 small parts. In order to simplify the rules, the original tens of thousands of injury conditions are highly condensed into more than a hundred scoring rules, for example, reduced to more than 500, so as to make scoring more efficient.

[0026] Taking bone injuries of the limbs, pelvis and buttocks as an example, the rewriting and integration methods of the scoring rules are described in detail. In this process, the scoring rules are expanded into multiple non-overlapping and non-omitted items, so as to construct a comprehensive and accurate scoring rule library. In order to achieve this goal, the Markdown language is used in this embodiment to structure the content, and various types of injuries in all parts of different body regions (a total of 6 major regions) are clearly displayed by setting titles, subtitles and lists. In the specific item settings, it is ensured that each item contains only one form of one (or similar) trauma in one part (or a specific type) of one body region, so as to avoid overlaps and omissions between items and make the rule base more rigorous and comprehensive.

[0027] Specifically, structured processing meets the following principles: 1) Use body region headings and place items in the same region under the same heading (e.g., Region 5 is “Limbs, Pelvis, Hips”).

[0028] 2) Use each part in the same area as a subheading under the corresponding title (for example, the part "Bone" is under the title of Area 5).

[0029] 3) Under the sub-heading, put items with the same AIS score under the same sub-heading (for example, clavicle fracture, scapula fracture, humerus fracture, pelvic ring fracture, acetabulum fracture, pelvic fracture, etc. with an AIS score of 2 points are listed under the sub-heading "AIS score: 2 points". In this way, the severity of each injury is quantified by the AIS score, which facilitates the subsequent calculation of the total ISS score).

[0030] 4) Open injuries and closed injuries at the same site are listed as one item each (e.g., “open clavicle fracture” and “closed clavicle fracture” are two items).

[0031] 5) Any part of the injury with unclear details (NFS) is a separate item.

[0032] Here is an example: #Zone 5: Limbs, pelvis, buttocks ##Part: Bones ###AIS score: 2 points: - Closed clavicle fracture; Open clavicle fracture; Closed scapula fracture; Open scapula fracture; - Closed humeral fracture; closed radius fracture; closed ulna fracture (including Monteggia fracture, Smith fracture, Galeazzi fracture, Colles fracture, Barton fracture, etc.); - Open simple diaphyseal fracture of ulna; Open simple diaphyseal fracture of radius; Open simple diaphyseal fracture of humerus; - Radius fracture; ulna fracture (including Monteggia fracture, Smith fracture, Galeazzi fracture, Colles fracture, Barton fracture, etc.); - Open fracture of the ulnar styloid process; - Closed carpal fracture; Open carpal fracture; - Closed metacarpal fracture; Open metacarpal fracture; - Closed fracture of the tibia; - Closed fibula fracture; Open fibula fracture; - Closed patellar fracture; Open patellar fracture; - Closed talus fracture; Open talus fracture; - Closed calcaneal fracture; Open calcaneal fracture; - Closed cuneiform fracture; Open cuneiform fracture; - Closed cuboid fracture; Open cuboid fracture; - Closed metatarsal fracture; Open metatarsal fracture; - Closed scaphoid fracture; Open scaphoid fracture; - Ankle fracture; medial malleolus fracture; lateral malleolus fracture; - Fractures of the pelvic ring (including sacrum, coccyx, hip, ilium, pubis and ischium), NFS: Sacrum fracture, NFS: Coccyx fracture, NFS: Hip fracture, NFS: Ilium fracture, NFS: Pubis fracture, NFS: Ischium fracture, NFS: Sciatic fracture, NFS. - A stable fracture of the pelvic ring (comprising the sacrum, coccyx, hip bones, ilium, pubis, and ischium) with an intact posterior ring. -Acetabular fracture: closed acetabular fracture.

[0033] - Pelvic fracture, NFS: Stable fracture of the pelvis with intact posterior ring.

[0034] It should be noted that the present invention rewrites the ISS rules into structured entries, which is of great significance in many aspects. First of all, this structured processing makes the data easier to process and parse, and can significantly reduce the computational complexity compared to unstructured text. By adopting the structured form of "region-site-score-trauma information", complex conditional judgments and calculation steps can be simplified, thereby improving the accuracy of judgments and reducing the required computing power. Such a design is particularly conducive to the deployment of trauma medical scoring models in hospitals, and can take into account the hospital's needs for data security, privacy protection and low cost.

[0035] In addition, the optimization of the ISS scoring rules has significantly reduced the number of entries, for example, it can be reduced to 1 / 10 to 1 / 100 of the original. This optimization not only makes the system more efficient in processing trauma tasks, but also avoids processing a wide range of natural language understanding tasks like large language models. Therefore, the trauma medical scoring model provided by the embodiment of the present invention can be constructed based on a large language model with small-scale parameters (1 billion level). This design can not only ensure the accuracy of the scoring, but also effectively reduce the computing power requirements, further improving the practicality and economy.

[0036] Step 2: Obtain basic cases, each of which is labeled with trauma-related information and scores.

[0037] Extract trauma or orthopedic medical records from the hospital's management system, and obtain desensitized original text from the medical record data, including current medical history, physical examination, auxiliary laboratory examination, diagnosis, etc., to construct a case data set (DS_Input1). That is, the case data set includes current medical history, physical examination, auxiliary laboratory examination, diagnosis, etc.

[0038] In the trauma-related information annotation stage, after expert judgment, the trauma-related information corresponding to each case is extracted from the case data set to construct a trauma-related information data set (DS_Output1). That is, the trauma-related information data set includes: 1) trauma-related information extracted from the aforementioned case data set as "injury description"; 2) the expert's "main diagnosis" extracted from the medical record.

[0039] by Figure 3 Taking the medical record shown as an example, the main diagnosis is "1. Injury of the medial meniscus of the left knee; 2. Injury of the cruciate ligaments of both knees; 3. Effusion of both knee joints"; the content before the main diagnosis is the description of the injury, which is the trauma-related content extracted from the aforementioned current medical history, physical examination, auxiliary examinations, etc.

[0040] From the above examples we can see that: 1) The primary diagnosis is a distillation of trauma-related information: The first point of the main diagnosis, "injury of the medial meniscus of the left knee", corresponds to the following two trauma-related information: • Injury of the medial meniscus body and posterior horn of the left knee joint (Grade II) • Separately listed "Left knee medial meniscus injury" The second point of the main diagnosis, "injury of the cruciate ligaments of both knee joints", corresponds to the following two trauma-related information: • Left anterior cruciate ligament + posterior cruciate ligament injury • Partial tear of the right anterior cruciate ligament Point 3 of the main diagnosis, "bilateral knee joint effusion", corresponds to the following two trauma related information: • Small amount of effusion in the left knee capsule • Small amount of effusion in the right knee capsule 2) The main diagnosis is based on the "Guidelines for the Grading of Orthopedic Injuries", and needs to be listed first: • Injuries requiring surgical intervention (eg, meniscus grade II injury) • Injuries that affect joint stability (such as complete rupture of the cruciate ligament) • The presence of effusion suggests an acute inflammatory lesion 3) The main diagnosis does not need to be listed: • Injuries that do not require surgical intervention (e.g., mild bone bruises) • Lesions that naturally resolve as the primary injury heals (e.g., soft tissue edema, which is a secondary change after trauma) • Old injuries Trauma-related information that does not need to be listed in the above major diagnoses needs to be pre-trained to initially avoid multiple or mis-evaluations in the trauma medical scoring model.

[0041] Medical score annotation: Annotate the case data set to obtain the standard score for each case that complies with ISS rules, including the following sub-steps: Step (21): For a small number of cases, after being annotated by experts, a standard score DS_Output2-1 for a small number of cases is constructed; Step (22): For most cases, score each case using the tuned Internet big language model (see prior patent application CN119337327A), and after expert correction, construct a standard score DS_Output2-2 for most cases; Step (23): Combine the data sets obtained in the above two steps to construct a score-related information data set (DS_Output2) for all cases. The score-related information data set includes region-site-trauma-related information-main diagnosis-score.

[0042] Step 3: Using the basic cases, different fine-tuning methods are used to tune the same base model, and pre-train the information extraction model and the medical scoring model.

[0043] In this embodiment, the selected base model (Base Model) is based on the internationally accepted Benchmark standard, and a large language model with excellent performance and open source small parameter scale is selected from many models. Specifically selected is the open source large language model GLM4-9B developed by Zhipu AI, which has 9 billion parameters. The GLM4-9B model is based on the Transformer architecture design and is optimized for Chinese semantic understanding and generation tasks. Its 9 billion parameter scale is between traditional deep learning models (millions) and ultra-large-scale models (hundreds of billions). It has strong semantic understanding capabilities and avoids the high computing power requirements of commercial large models. It is suitable for deployment in scenarios with limited resources (such as local servers in hospitals). In addition, as an open source model, GLM4-9B supports users to deeply tune the model structure and parameters, which is convenient for field adaptation in combination with medical scoring scenarios, such as adding classification layers, adjusting attention mechanisms, etc.

[0044] In order to reduce the accuracy of model parameters, the base model can be quantized. For example, FP16 precision compression can be used to significantly reduce memory usage and inference latency at the expense of a small amount of performance, thereby meeting clinical real-time requirements. At the same time, the base model can be run on a single machine with low hardware investment.

[0045] In the process of selecting the fine-tuning method, this embodiment is optimized through comparative experiments. The specific operation is as follows: use the case data set constructed in step 2 as the input of the first base model, and the trauma-related information data set as the expected output of the model. A variety of fine-tuning methods are used in the experiment, including but not limited to LoRA, AdaLoRA, LoKr, LoHA, IA3, P-Tuning, Prefix-Tuning, etc. During the experiment, the external hyperparameters of all methods and the conventional parameters corresponding to each fine-tuning method are kept consistent. Finally, based on the expert's preliminary judgment on the fine-tuning results, the fine-tuning method with the highest accuracy is selected as the preferred first fine-tuning method (FTM1), such as Prefix-Tuning, which is particularly suitable for enhancing the model's ability to capture context.

[0046] In order to further optimize the first base model, this embodiment uses the first fine-tuning method selected above to tune it. The specific operation is: using the case data set constructed in step 2 as the model input, and the trauma-related information data set as the model output, by adjusting the hyperparameters, using the first fine-tuning method to make the first base model achieve the best effect.

[0047] After fine-tuning, the fine-tuned model is further enhanced through retrieval enhancement generation (RAG) and prompt engineering to obtain a trauma-related information extraction model. Using this information extraction model, all cases can be processed to generate a data set containing trauma-related information, providing accurate input data for subsequent medical scoring. The following are the specific sub-steps: Set the model role: Set the role of the first base model to an emergency doctor, who is good at analyzing medical records and identifying trauma-related injury information in the medical records.

[0048] Clarify the processing goal: Given the first base model processing goal, that is, to extract trauma-related diagnostic information from medical records.

[0049] Prompt processing flow: The processing flow of the first base model is clarified through the prompt project, which specifically includes the following steps: Extract diagnostic information: Extract diagnostic content from the input text.

[0050] Determine relevance: Identify whether each diagnosis is relevant to this acute trauma.

[0051] Extract detailed descriptions: For the diagnoses related to this score, extract detailed descriptions related to the diagnosis from the medical records, including medical history, physical examination, auxiliary examinations, etc.

[0052] Specify the output format: Set the output format of the model to JSON format, which includes three fields: "Diagnosis", "Relevant", and "Disease Description".

[0053] In actual operation, the input medical record data comes from the case data set (DS_Input1), including unstructured text such as current medical history, physical examination, auxiliary test examination, diagnosis, etc. Structured tags (tokens) are inserted into the original medical record text to clearly identify the region and part, and provide the model with explicit region-part boundary information to facilitate subsequent attention range control. Finally, the output of the first base model (through the prompt engineering setting) is structured trauma-related information (JSON format), including: Diagnosis: Trauma-related diagnostic entries extracted from medical records.

[0054] IsRelevant: Determines whether this diagnosis is relevant to the current acute trauma (Boolean value).

[0055] Injury description: A detailed description related to the diagnosis extracted from the medical record (such as medical history, physical examination, auxiliary examinations, etc.).

[0056] The information extraction model in the embodiment of the present invention is constructed based on the GLM4-9B model, which can accurately extract trauma-related information and effectively solve the problem that the existing technology is difficult to process unstructured text. The model supports efficient parameter fine-tuning methods such as LoRA and Prefix-Tuning, and only requires a small amount of labeled data to adapt the medical scoring rules, avoiding the high cost of training from scratch. In addition, for injuries that do not need to be listed (such as injuries that do not require surgical intervention, old injuries, etc.), the model can effectively filter and process them.

[0057] In the process of building the medical scoring model, the first base model (BM-A) is also used as the basis. On this basis, a multi-layer classification network (such as Figure 4 As shown in Figure 1, a new classification model is constructed. The number of output units of the classification model strictly corresponds to the number of medical scoring items. For example, in the AIS-ISS scoring system, the scoring items include 6 regions and 43 parts. After rewriting in step 1, the number of medical scoring items is more than 500, so the corresponding output units are also more than 500, and each scoring item corresponds to an output unit.

[0058] The main function of the base model is to receive input layer data (such as medical record text) to complete feature extraction and semantic understanding. The superposition of multiple layers of classification networks is used to refine the extracted features into specific scoring items. The classification network usually consists of a fully connected layer and Dropout. The output layer is adapted according to the scoring requirements. Since the number of output units strictly corresponds to the number of items in the AIS-ISS scoring system, each output unit outputs the score of a part. The activation function of the output layer uses Softmax.

[0059] In the selection of fine-tuning methods, similar to the information extraction model, the case data set constructed in step 2 is used as the model input, and the score-related information data set is used as the model output. LoRA, AdaLoRA, LoKr, LoHA, IA3, P-Tuning, Prefix-Tuning and other fine-tuning methods are used for comparative experiments. In the experiment, the same external hyperparameters and conventional parameters corresponding to the fine-tuning method are maintained. Based on the preliminary judgment of the experts, the appropriate second fine-tuning method (FTM2) is selected. This method is different from the first fine-tuning method (FTM1). For example, LoRA can be used to efficiently learn the scoring rules. By adjusting the hyperparameters, the model achieves the best effect, and finally the optimized trauma medical scoring model is obtained.

[0060] It should be noted that although the information extraction model and the medical scoring model are built on the same base model, different fine-tuning methods are used during the optimization process. This strategy enables the two models to maximize performance at their respective stages. Specifically, the information extraction model focuses on accurately extracting trauma-related information from medical record texts, while the medical scoring model focuses on accurately mapping this extracted information to specific scoring rules.

[0061] This division of labor actually decomposes the complex medical scoring task into two orthogonal dimensions: "information understanding" and "rule application". The information extraction model is responsible for understanding the key information in the medical record text, while the trauma medical scoring model is responsible for applying this information according to the established scoring rules. Through this specialized division of labor, the common knowledge loss and task conflict problems of large language models in professional scenarios are effectively solved.

[0062] Ultimately, this design achieves a highly accurate trauma medical scoring model that can be deployed in hospitals. It not only improves the efficiency and accuracy of scoring, but also reduces the demand for computing power, allowing the model to run efficiently in a hospital environment with limited resources.

[0063] Step 4: Based on the information extraction model and the medical scoring model, a differential model is constructed; the differential model generates differential information, including a list of under-rated items, a list of over-rated items, a list of mis-rated items and difference labels, by comparing the output of the medical scoring model with the expert annotation results, and triggers back propagation to optimize the main model parameters.

[0064] This step combines expert rules and data-driven automatic adjustment, and uses the expert annotation results as the differential result data set to iteratively optimize the model. In one embodiment of the present invention, step 4 may include the following sub-steps: (41) Based on the medical scoring model trained in step 3, perform inference generation model scoring on all cases; (42) The model scores were corrected by experts to construct a differential result dataset, in which the under-rated, over-rated, and mis-rated items of each case were annotated; (43) aligning the trauma-related information dataset output by the information extraction model with the scoring information dataset output by the medical scoring model to construct a structured differential input dataset; For example, the process of constructing a differential input dataset can be to align the trauma-related information (DS_Input2) and the score-related information (DS_Output2) of the same case by field based on the case ID number to form a complete record containing detailed information of "trauma description + model score".

[0065] The differential outcome data set contains "trauma-related information and differential information", and its specific content includes at least the following three categories: Under-rated information: refers to the injury items that the information extraction model failed to identify, but the experts clearly pointed out during the annotation process. For example, the injury item "fracture of the fifth rib on the right side" was not scored by the model, but the experts clearly pointed out its existence during the annotation.

[0066] Multi-criteria information: refers to the information extraction model incorrectly identifying non-trauma related items and misclassifying them as acute injuries. For example, misclassifying "old lumbar disc herniation" as an injury related to current trauma.

[0067] Mis-rating information: refers to the difference between the score given by the medical scoring model and the score annotated by the expert exceeding the set threshold. For example, the AIS score given by the medical scoring model is 3 points, while the score annotated by the expert is 4 points. This score difference exceeds the allowable error range.

[0068] (44) selecting a second base model independent of the first base model for quantization processing, and selecting a fine-tuning method through comparative experiments based on the differential input data set and the differential result data set; Here, a large language model that is different from the first base model but has a comparable number of parameters is selected as the second base model (BM-B), and a model quantization and fine-tuning method is selected.

[0069] (45) Using the fine-tuning method selected in step (44), with the differential input data set as input and the differential result data set as output, the hyperparameters are optimized to achieve optimal model performance, and finally the differential model is trained to obtain the differential model.

[0070] It should be noted that the information extraction model is first trained using the complete data set. After the training is completed, the medical scoring model will adjust the data set weight according to the deviations in the prediction results of the information extraction model to complete its own training. At the same time, the information extraction model will also calculate its own weight in the final decision based on these deviations. The final decision result is based on the weighted sum of each classifier. In this way, the deviation can be effectively controlled, thereby improving the overall prediction accuracy.

[0071] In one embodiment of the present invention, the structure of the differential model is as follows: Figure 5As shown in the figure, it mainly includes input layer, base model (feature extraction), parallel dual classification network (main / auxiliary task processing), output layer (result fusion) and fine-tuning network (dynamic parameter optimization). The fine-tuning network forms a closed-loop training mechanism with the base model and classification network through gradient backpropagation or attention weight adjustment. Among them, the function of the input layer is to receive the differential input data set (DS_Input3) from the information extraction model or the medical scoring model, and pass the data to the base model after standardization. The differential input data set is structured, and each record contains aligned and spliced ​​fields, including trauma-related information (DS_Input2 from the information extraction model, such as injury description, location, injury description) and scoring information (DS_Output2 from the medical scoring model, such as AIS score). This information is uniformly encoded into an input format that can be processed by the differential model, such as "trauma description: [injury item]; model score: [scoring item]".

[0072] The base model is a pre-trained model, usually based on the Transformer architecture, using a large language model with the same parameter size as the information extraction model and the medical scoring model as the third base model (e.g., GLM4-9B). This model is independent of the first base model or the second base model of the information extraction model and the medical scoring model to better adapt to the characteristics of the differential task. The base model semantically encodes the concatenated text passed by the input layer and extracts high-level features, such as the correlation between injuries and scores and potential conflict patterns. At the same time, the self-attention mechanism is used to capture the long-range dependency between the trauma description and the score, and identify differential clues, such as the contradiction between the keywords "oldness" and "acuteness".

[0073] Based on the feature vector output by the base model, the differential model processes three types of differential tasks through parallel multi-task classification heads: under-review detection, over-review detection, and misreview detection. Each task head contains an independent classification network (consisting of fully connected layers and activation functions), and the parameters of each task head are not shared. The output layer contains at least a fully connected layer or a Softmax activation function to integrate the results of the classification network. Finally, the output layer generates the final output, such as a difference tolerance label, by weighting or splicing.

[0074] For example, the output layer of the difference model produces the following structured results: 1) List of items with few comments (e.g., “fracture of the fifth rib on the right”).

[0075] 2) A list of multiple evaluation items (such as "chronic lumbar disc herniation").

[0076] 3) List of mis-evaluated items (e.g. “AIS score 3 points → 4 points”).

[0077] 4) Difference labels, which represent the difference type between the differential input data set and the differential result data set. Different difference labels are provided for under-review, over-review, and mis-review. The difference signals of the corresponding output units will trigger the gradient back propagation to the medical scoring model through the feedback path of the differential model to optimize its parameters, forming a closed-loop training mechanism.

[0078] The output format can be JSON or a specific tag sequence (such as "less comments on the right 5th rib bone").

[0079] To facilitate understanding, the following provides examples of the input and output (differential information) of the differential model in the cases of over-rating, under-rating, and erroneous rating.

[0080] Example 1: Multiple reviews, see Figure 6 .

[0081] As can be seen from Example 1, in the case of multiple reviews, the differential model input information (the output information of the medical scoring model) includes one or more records, each of which includes the following information: 1) Matching content (i.e. rules) 2) Part: bones 3) Area: limbs, pelvis, buttocks 4) Injury description: Sequelae of fracture of the 7th and 8th anterior ribs on the left side 5) Injury score: 2 In the case of multiple reviews, the differential information output by the differential model includes at least one list of multiple review items, each of which includes: 1) Differentiate labels: Don’t focus on old and non-traumatic information; 2) Injury description corresponding to the difference: sequelae of fracture of the 7th and 8th anterior ribs on the left side.

[0082] Example 2: Few reviews, see Figure 7 .

[0083] As can be seen from Example 2, in the case of few reviews, the differential model input information (the output information of the medical scoring model) includes one or more records, each of which includes the following information: 1) Matching content (i.e., the rules that apply to the injury description) 2) Part: joints 3) Area: limbs, pelvis, buttocks 4) Injury description: right shoulder dislocation 5) Injury score: 2 In the case of few reviews, the differential information output by the differential model includes at least one list of few-reviewed items, each of which includes: 1) Differential label: Pay special attention to information related to this trauma; 2) Injury description corresponding to the difference: fracture of the greater tuberosity of the right humerus.

[0084] Example 3: In the case of misjudgment, see Figure 8 .

[0085] As can be seen from Example 3, in the case of misjudgment, the differential model input information (the output information of the medical scoring model) includes one or more records, each of which includes the following information: 1) Matching content (i.e., the rules that apply to the injury description) 2) Part: bones / joints 3) Area: Face 4) Injury description: Nasal bone fracture (bilateral) 5) Injury score: 1 In the case of misjudgment, the differential information output by the differential model includes at least one list of misjudged items, and each multi-judgment item includes: 1) Difference labels: Pay attention to the AIS-ISS related rules; 2) The correct rules for differences; 3) Wrong rules corresponding to differences.

[0086] It should be noted that when the scoring result of the medical scoring model is completely correct, the difference label will be empty. This means that when the model's score is exactly the same as the expert's annotated score, the difference model will not generate any difference labels. In addition, the specific content of the difference label can be changed flexibly, as long as it is different in the three cases of under-rating, over-rating and mis-rating, so as to clearly distinguish different difference scenarios.

[0087] The fine-tuning network works in parallel with the base model. Through the feedback mechanism, the fine-tuning network can dynamically adjust the parameters of the base model or classification network. This dynamic adjustment is based on the difference between the model's output and the expert annotation. The fine-tuning network also introduces an attention mechanism and an adaptive loss function, which are applied to optimize the model's adaptability to specific differential scenarios (such as trauma-related differences). In this way, the model can more accurately identify and handle various complex differential situations, thereby improving the overall scoring accuracy and reliability.

[0088] Step 5: Connect the information extraction model in series with the medical scoring model and in parallel with the differential model to build a combined model; the output of the differential model dynamically adjusts the model weights through the joint loss function to form a closed-loop training mechanism.

[0089] In one embodiment of the present invention, step 5 specifically includes the following sub-steps: Model initialization: Initialize the parameters of the information extraction model and medical scoring model.

[0090] Forward propagation: The case data set is extracted through the features of the information extraction model and output to the input layer of the medical scoring model.

[0091] Loss calculation: Calculate the value of the joint loss function.

[0092] Joint loss function total_loss = α score_loss+β under_loss+γ error_loss+δ over_loss Among them, α is the weight of the trauma scoring task; β is the weight of the few-review detection task; γ is the weight of the false-review detection task; δ is the weight of the multiple-review detection task; score_loss is the loss of the trauma scoring task, under_loss is the loss of the few-review detection task; error_loss is the loss of the false-review detection task; over_loss is the loss of the multiple-review detection task.

[0093] Back propagation: The differential information corresponding to the under-rated, mis-rated, and over-rated items output by the differential model is input into the medical scoring model, and the model parameters are updated by back propagation according to the value of the joint loss function.

[0094] Weight adjustment: Dynamically adjust the weight coefficients of each task according to the loss during training. If the loss of any one or more of under-rating / over-rating / wrong rating is large, its weight β, γ, δ can be dynamically increased to make the combined model pay more attention to these three situations.

[0095] Repeat training: Repeat the above steps until the combined model converges.

[0096] The trauma medical scoring model provided by the embodiment of the present invention realizes an efficient closed-loop optimization training mechanism through the collaborative work of an information extraction model (MR), a medical scoring model (MS) and a differential model (MD). The output of the information extraction model serves as the input of the medical scoring model, and the output of the medical scoring model serves as the input of the differential model. The construction of the differential input data set integrates the original trauma information, model scoring results and rule verification labels, and the differential result data set not only contains the error type, but also implies the correct score after expert correction. This mechanism forms a closed-loop optimization training mechanism: when the differential information is intolerable (that is, the "no" branch, such as Figure 2 When the difference information is tolerable, the model iterative optimization is triggered; when the difference information is tolerable, the scoring result is directly output. Iterative optimization through the expert-annotated difference result dataset can effectively reduce the scoring bias.

[0097] The scoring rule base plays a key role in this process. Figure 2As shown in RAG, the scoring rule base is used to verify whether the output of the information extraction model, medical scoring model and differential model meets the ISS standard. This ensures that the output results of the entire system are highly accurate and reliable.

[0098] In the closed-loop training mechanism of the trauma medical scoring model, the differential model establishes a bidirectional gradient propagation path by structured processing of three types of differential information: under-evaluation, mis-evaluation, and over-evaluation. This path realizes the coordinated optimization of the medical scoring model (MS) and the information extraction model (MR). For example, when the information extraction model begins to extract the feature of "fracture displacement degree", the medical scoring model will simultaneously strengthen the mapping relationship between this feature and the AIS score. In addition, through the fine feedback of differential information, it is possible to achieve a scoring accuracy that exceeds the hundreds of billions of large language models while maintaining a small parameter scale. This design not only improves the efficiency and accuracy of the system, but also reduces the demand for computing power, making it more suitable for deployment in resource-limited environments such as hospitals.

[0099] Step 6: Divide the basic cases into a sample set and a validation set, use the sample set to jointly tune the combined model, and use the validation set to verify it, so as to obtain a trauma medical scoring model.

[0100] In one embodiment of the present invention, the basic cases are divided into a sample set and a validation set. The sample set is used to jointly tune the combined model, while the validation set is used to verify the performance of the model, thereby obtaining the final trauma medical scoring model. It is worth noting that the medical scoring model and the information extraction model in the combined model both use the same basic case set, but these case sets are extracted independently, so the input and output content of each sub-model are different.

[0101] In addition, the information extraction model and the medical scoring model are built on the same base model, and then the two are connected in series and in parallel with the differential model. The base model of the differential model is different from the information extraction model and the medical scoring model. This design enables the differential model to improve the accuracy of the main model's scoring results while maintaining a low computational overhead, thereby improving the accuracy of the entire trauma medical scoring model.

[0102] In addition, by tuning the information extraction model and the medical scoring model separately, the information extraction model focuses on semantic understanding (converting unstructured data into structured data), while the medical scoring model focuses on rule mapping (converting structured data into scores). This division of labor avoids the problem of difficulty in finding a unified optimization path due to conflicts in the objective functions of different tasks, thereby improving the accuracy of the combined model.

[0103] The embodiment of the present invention also achieves independent modularization, interpretability and continuous improvement capabilities of the trauma medical scoring model by separating the process of "information extraction → scoring → error correction". When the scoring results are biased, the independent model architecture can quickly locate the problem link (whether it is an information extraction error or a misuse of the scoring rules). Combined with the parallel design of the differential model, errors in specific stages can be corrected in a targeted manner.

[0104] The differential model itself uses the self-attention mechanism to locate logical contradictions between trauma descriptions and scores (such as the mismatch between "oldness" and high scores) by calculating the weights of inter-word associations. It also uses long-range dependencies to allow the model to directly associate remote fields (such as keywords in trauma descriptions and scoring results) without relying on local context. In addition, the differential model converts abstract features into specific difference types through a multi-task classification head and generates a final report in combination with expert rules. These features significantly improve the prediction effect of the differential model by dynamically fine-tuning the network and continuously optimizing the attention pattern using expert annotation data.

[0105] In summary, the trauma medical scoring model provided by the embodiment of the present invention is very suitable for deployment in hospitals and can effectively ensure data security and privacy protection.

[0106] Second embodiment The second embodiment of the present invention provides a trauma medical scoring method based on the first embodiment. The method uses the trauma medical scoring model obtained by the aforementioned trauma medical scoring model training method to score the trauma of the injured to optimize the allocation of medical resources.

[0107] Third embodiment Based on the above trauma medical scoring method, this embodiment further provides a trauma medical scoring system. Fig. 9 As shown, the trauma medical scoring system includes one or more processors and a memory, wherein the memory is coupled to the processor and is used to store one or more programs, and when the programs are executed by the processor, the processor implements the trauma medical scoring method in the above embodiment.

[0108] The processor is used to control the overall operation of the trauma medical scoring system to complete all or part of the steps of the above-mentioned trauma medical scoring method. The processor can be a central processing unit (CPU), a graphics processing unit (GPU), a field programmable gate array (FPGA), an application-specific integrated circuit (ASIC), a digital signal processing (DSP) chip, etc. The memory is used to store various types of data to support the operation of the trauma medical scoring system, and these data may include, for example, instructions for any application or method used to operate on the trauma medical scoring system, and application-related data. The memory can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, etc.

[0109] In another exemplary embodiment, the present invention further provides a computer-readable storage medium including program instructions, which, when executed by a processor, implements the steps of the trauma medical scoring method in any of the above embodiments. For example, the computer-readable storage medium may be the above-mentioned memory including program instructions, and the above-mentioned program instructions may be executed by the above-mentioned processor to implement the above-mentioned method and achieve the same technical effect as the above-mentioned method.

[0110] It should be noted that the above embodiments are only examples, and the technical solutions of the various embodiments can be combined, and the order of the steps can be changed, all within the protection scope of the present invention.

[0111] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0112] The above is a detailed description of the trauma medical scoring model training method, scoring method and scoring system provided by the present invention. For those skilled in the art, any obvious changes made to it without departing from the essence of the present invention will constitute an infringement of the patent right of the present invention and will bear corresponding legal responsibilities.

Claims

1. A trauma medical scoring model training method, characterized in that The following steps are involved: Step 1: Deconstruct and simplify the scoring rules, expand them into multiple items that do not overlap and have no omissions, and build a scoring rule library; Step 2: Obtain basic cases, each of which is labeled with trauma-related information and scores; Step 3: Using the basic cases, different fine-tuning methods are used to tune the same base model, and pre-train the information extraction model and the medical scoring model; Step 4: constructing a differential model based on the information extraction model and the medical scoring model; The differential model generates differential information, including a list of under-rated items, a list of over-rated items, a list of mis-rated items, and differential labels, by comparing the output of the medical scoring model with the expert annotation results, and triggers back propagation to optimize the main model parameters; Step 5: Connect the information extraction model and the medical scoring model in series and in parallel with the differential model to build a combined model; the output of the differential model dynamically adjusts the model weights through the joint loss function to form a closed-loop training mechanism; Step 6: Divide the basic cases into a sample set and a validation set, use the sample set to jointly tune the combined model, and use the validation set to verify it, so as to obtain a trauma medical scoring model.

2. The trauma medical scoring model training method according to claim 1, characterized in that The step 4 includes the following sub-steps: (41) Based on the medical scoring model trained in step 3, perform inference generation model scoring on all cases; (42) The model scores were corrected by experts to construct a differential result dataset, in which the under-rated, over-rated, and mis-rated items of each case were annotated; (43) aligning the trauma-related information dataset output by the information extraction model with the scoring information dataset output by the medical scoring model to construct a structured differential input dataset; (44) selecting a second base model independent of the first base model for quantization processing, and selecting a fine-tuning method through comparative experiments based on the differential input data set and the differential result data set; (45) Using the fine-tuning method selected in step (44), with the differential input data set as input and the differential result data set as output, the hyperparameters are optimized to achieve optimal model performance, and finally the differential model is trained to obtain the differential model.

3. The trauma medical scoring model training method according to claim 2, characterized in that: The information extraction model and the medical scoring model are based on the same first base model and are constructed using different fine-tuning methods; the differential model is based on an independent second base model and is constructed using a fine-tuning method different from the aforementioned models.

4. The trauma medical scoring model training method according to claim 2, wherein: The output of the information extraction model serves as the input of the medical scoring model. The output of the medical scoring model is integrated with the trauma-related information and differential information judged by experts into a differential input data set, which serves as the input of the differential model.

5. The trauma medical scoring model training method according to claim 1, wherein: In step 3, the pre-trained information extraction model and medical scoring model are quantized to reduce memory usage and adapt to the hospital's local server deployment.

6. The trauma medical scoring model training method according to claim 1, characterized in that: In step 4, the difference label is used to indicate the difference type between the difference input data set of the difference model and the difference result data set of the difference model.

7. The trauma medical scoring model training method according to claim 6, wherein: In the case of multiple reviews, the difference information includes at least one list of multiple review items, where each multiple review item includes: 1) a difference label; 2) a description of the injury corresponding to the difference; In the case of few reviews, the difference information includes at least one list of few review items, wherein each of the few review items includes: 1) a difference label; 2) a description of the injury corresponding to the difference; In the case of mis-evaluation, the difference information includes at least one list of mis-evaluation items, wherein each mis-evaluation item includes: 1) a difference label; 2) a correct rule corresponding to the difference; 3) an incorrect rule corresponding to the difference.

8. The trauma medical scoring model training method according to claim 1, wherein: In step 6, the verification process of the verification set includes experts correcting the output results of the differential model to form a new differential result data set, and iteratively optimizing the parameters of the combined model.

9. A trauma medical scoring method, characterized in that The trauma medical scoring model obtained by using the trauma medical scoring model training method described in any one of claims 1 to 8 is used to score the trauma of the wounded to optimize the allocation of medical resources.

10. A trauma medical scoring system, characterized in that It comprises a processor and a memory, wherein the processor and the memory are coupled; wherein the memory is used to store a computer program; and the processor runs the computer program stored in the memory to implement the trauma medical scoring model training method described in any one of claims 1 to 8.

Images