Artificial Intelligence-Based Methods and Systems for Assessing Postoperative Risks of Jaw Cysts

Abstract

This invention provides an artificial intelligence-based method and system for assessing postoperative risk of jaw cysts. The method includes: acquiring target clinical data; extracting risk features from the target clinical data and using an SVM machine learning risk assessment model to predict the risk features to obtain a first postoperative risk assessment result; extracting patient basic information, abnormal indicator information, and surgical contraindication information from the target clinical data; using the patient basic information and abnormal indicator information as search terms to search a medical expert knowledge base to obtain risk retrieval results; inputting a prompt word template constructed based on the patient basic information, surgical contraindication information, and risk retrieval results into a large language model to generate a second postoperative risk assessment result and the basis for obtaining the current assessment result; when the first and second postoperative risk assessment results are inconsistent, inputting the first and second postoperative risk assessment results into a risk decision model to output a final postoperative risk assessment result.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an artificial intelligence-based method and system for assessing postoperative risks of jaw cysts. Background Technology

[0002] Jawbone cysts are fluid-filled cystic lesions within the jawbone. They can occur anywhere on the jawbone, initially without any symptoms. As they grow slowly, the cyst can compress and expand the surrounding bone. Jawbone cysts are one of the most common diseases in oral and maxillofacial surgery, characterized by high incidence, insidious nature, and high misdiagnosis rate. Many patients with jawbone cysts are diagnosed with significant bone destruction at the time of diagnosis. Continued disease progression or improper treatment can lead to gingival retraction, tooth displacement, malocclusion, and in severe cases, maxillofacial deformities, significantly impacting the patient's quality of life.

[0003] Surgery is the primary treatment for jaw cysts. However, a number of risks exist post-operatively, including but not limited to: Cyst recurrence: Especially for aggressive cysts (such as ameloblastomas and keratocysts), recurrence is a core challenge in long-term post-operative management. Pathological fracture: Large cysts can cause severe bone resorption and thinning of the jawbone, reducing post-operative weight-bearing capacity and increasing the risk of fracture. Nerve damage (inferior alveolar nerve): Cysts are closely related to nerves, and surgery may cause permanent or temporary sensory dysfunction. Infection and wound non-healing: This is related to the type of cyst, surgical trauma, and the patient's overall condition. Damage to adjacent teeth: Surgery may lead to pulp devitalization, loosening, or loss of adjacent teeth.

[0004] Currently, risk assessment for jaw cyst surgery relies heavily on the individual experience of clinicians, resulting in significant subjectivity, non-quantification, and inconsistency. Specifically: 1. Limited Assessment Dimensions: Doctors primarily rely on two-dimensional X-rays (such as curved tomography) or CT images for visual assessment, failing to adequately extract the three-dimensional morphology, internal texture features, and subtle relationships with surrounding tissues of the lesion. This also hinders effective utilization of preoperative risk assessment. 2. Difficulty in Information Integration: Risk assessment requires integrating multi-source information such as imaging features (size, borders, density), pathological type (puncture or postoperative pathology), patient age, overall health status, and clinical symptoms. Doctors must integrate this unstructured information mentally, lacking standardized and systematic fusion analysis models. 3. Strong Experience Dependence: The "intuitive" judgments developed by senior doctors over many years are difficult to quantify and impart, leading to significant differences in assessment standards between different hospitals and doctors, and a long learning curve for younger doctors. 4. Inability to Achieve Dynamic Prediction: It is difficult to dynamically and personally predict the recurrence risk at different postoperative time points (e.g., 6 months, 1 year, 3 years post-surgery).

[0005] In recent years, although AI has made progress in the field of medical image analysis, specialized technologies for postoperative risk assessment of jaw cysts still have significant shortcomings: Limited functionality: Most existing technologies focus on cyst detection and segmentation (identifying "what it is" and "where it is"), or benign / malignant classification, with few systems deeply researching postoperative biological behavior prediction (predicting "what will happen after surgery"). Limited data modality: Most studies are based solely on single imaging modalities such as CT or MRI, failing to effectively integrate key information such as clinical textual data (medical records, pathology reports) and blood test indicators. This results in risk assessment models "seeing the trees but not the forest," limiting predictive effectiveness. Poor model interpretability: Many are "black box" models, only providing risk probability values ​​but failing to clearly explain which specific characteristics of the cyst drive this risk conclusion. This makes them difficult for doctors to understand and trust, hindering their use in clinical decision support. Limited generalization ability: Models are often trained on single-center, small-sample data, resulting in insufficient generalization ability across different hospital equipment, scanning parameter differences, and various rare cyst types, limiting their clinical applicability. Summary of the Invention

[0006] In view of the above problems, the present invention is proposed to provide an artificial intelligence-based method and system for assessing postoperative risks of jaw cysts to overcome the above problems.

[0007] This invention provides an artificial intelligence-based method for assessing postoperative risk of jaw cysts, the method comprising: Acquire target clinical data, which includes patient basic information, past medical history, personal history, preoperative condition, laboratory indicators, and imaging examination data; Risk features of the target clinical data are extracted, and a pre-trained SVM machine learning risk assessment model is used to learn and predict the risk features to obtain the first postoperative risk assessment result. Extract basic patient information, abnormal indicator information, and surgical contraindication information of the patients to be evaluated from the target clinical data; The patient's basic information and abnormal indicator information are combined into search terms. Based on the search terms, a pre-built medical expert knowledge base is searched to obtain risk search results based on the expert knowledge base. The risk search results are the complications triggered by the abnormal indicator information. Based on the patient's basic information, surgical contraindications information, and risk retrieval results, a prompt word template is constructed. The prompt word template is then input into a preset large language model to generate a second postoperative risk assessment result and the basis for obtaining the current assessment result. When the results of the first postoperative risk assessment and the second postoperative risk assessment are inconsistent, the results of the first and second postoperative risk assessments are input into the pre-trained risk decision model, and the final postoperative risk assessment result is output.

[0008] Furthermore, before extracting risk characteristics from the target clinical data, the method includes determining risk characteristics for postoperative risk assessment; Identify risk characteristics for postoperative risk assessment, including: Obtain the first training dataset, which includes data samples from desensitized historical patients' clinical data and surgical complications. Extract the original data features of each data sample in the first training dataset; Dummy variable transformation and normalization are performed on each feature information in the original data. Each processed feature was compared with the surgical complication results using an independent samples t-test to remove features from the original data that were not related to the surgical complication results. Based on the Pearson correlation coefficient, feature filtering is performed on the remaining feature information after removing irrelevant feature information in order to remove highly redundant strongly correlated features; We construct a minimum absolute shrinkage and selection operator Lasso regression model. During the feature selection regression process, we penalize the feature coefficients to compress the coefficients of unimportant features to 0, while retaining the features with non-zero coefficients. The selected core features are then used as risk features for machine learning.

[0009] Furthermore, before extracting the patient's basic information, abnormal indicator information, and surgical contraindication information from the target clinical data, the method further includes: Detect and remove abnormal statistical data from the target clinical data.

[0010] Furthermore, the training steps of the machine learning risk assessment model include: Obtain the first training dataset, which includes data samples from desensitized historical patients' clinical data and surgical complications. Extract risk features from each data sample in the first training dataset, and train a machine learning risk assessment model based on the risk features of each data sample and the corresponding surgical complications. During the training of the machine learning risk assessment model, the GridSearchCV grid search method is used to optimize the key parameters of the machine learning risk assessment model to determine the optimal parameter combination and obtain the optimal machine learning risk assessment model.

[0011] Furthermore, the method also includes: Using the prediction function of the optimal machine learning risk assessment model as the input to the SHAP interpreter, a portion of the samples in the training dataset are randomly selected as the background dataset to calculate the baseline contribution value of each risk feature. Randomly select a portion of samples from the training dataset as the test dataset, call the shap_values() method of the SHAP interpreter, and calculate the SHAP value corresponding to each core feature in each test sample in the test dataset. The SHAP values ​​of each core feature in all test samples are displayed in a graphical visualization.

[0012] Furthermore, the steps for constructing the medical expert knowledge base include: Obtain a second training dataset, in which data samples include desensitized clinical data of historical patients. A medical expert knowledge base is constructed based on the second training dataset and the guidance of oral medicine experts. The knowledge entries in the medical expert knowledge base include abnormal indicator information, the normal value range of the abnormal indicator information, complications exceeding the normal value, and surgical contraindications.

[0013] Furthermore, the training steps of the risk decision-making model include: Obtain a third training dataset, whose data samples include desensitized historical patients' clinical data and surgical complications; Obtain the first postoperative risk assessment result and the second postoperative risk assessment result corresponding to each data sample in the third training dataset; Using the first and second postoperative risk assessment results as inputs and the surgical complications of the corresponding data samples as labels, a decision tree algorithm is used for modeling and training to obtain a risk decision model.

[0014] Another aspect of the present invention provides an artificial intelligence-based risk assessment system for jaw cyst surgery, the system comprising: The acquisition module is used to acquire target clinical data, which includes patient basic information, past medical history, personal history, preoperative conditions, laboratory indicators, and imaging examination data. The first risk assessment module is used to extract the risk features of the target clinical data, and to learn and predict the risk features using a pre-trained SVM machine learning risk assessment model to obtain the first postoperative risk assessment result. The prompt word information extraction module is used to extract basic patient information, abnormal indicator information, and surgical contraindication information of the patient to be evaluated from the target clinical data; The prompt word information retrieval module is used to merge the patient's basic information and abnormal indicator information into search words, search a pre-built medical expert knowledge base based on the search words, and obtain risk retrieval results based on the expert knowledge base. The risk retrieval results are the complications triggered by the abnormal indicator information. The second risk assessment module is used to construct a prompt word template based on the patient's basic information, surgical contraindications information and risk retrieval results. The prompt word template is then input into a preset large language model to generate a second postoperative risk assessment result and the basis for obtaining the current assessment result. The decision module is used to input the first and second postoperative risk assessment results into the pre-trained risk decision model and output the final postoperative risk assessment result when the first postoperative risk assessment result is inconsistent with the second postoperative risk assessment result.

[0015] Furthermore, the system also includes: The feature screening module is used to determine risk features for postoperative risk assessment before extracting risk features from the target clinical data; Identify risk characteristics for postoperative risk assessment, including: Obtain the first training dataset, which includes data samples from desensitized historical patients' clinical data and surgical complications. Extract the original data features of each data sample in the first training dataset; Dummy variable transformation and normalization are performed on each feature information in the original data. Each processed feature was compared with the surgical complication results using an independent samples t-test to remove features from the original data that were not related to the surgical complication results. Based on the Pearson correlation coefficient, feature filtering is performed on the remaining feature information after removing irrelevant feature information in order to remove highly redundant strongly correlated features; We construct a minimum absolute shrinkage and selection operator Lasso regression model. During the feature selection regression process, we penalize the feature coefficients to compress the coefficients of unimportant features to 0, while retaining the features with non-zero coefficients. The selected core features are then used as risk features for machine learning.

[0016] In another aspect, the present invention provides a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the artificial intelligence-based method for assessing postoperative risk of jaw cysts as described above.

[0017] This invention provides an artificial intelligence-based method and system for assessing postoperative risks of jaw cyst surgery. It integrates multi-source medical data, including patient basic information, personal history, preoperative conditions, examination indicators, and imaging data. Based on collected historical patient data, a machine learning risk assessment model is constructed. Furthermore, an innovative large-scale risk model for jaw cyst surgery is built based on a constructed medical expert knowledge base. These two risk assessment models provide postoperative risk assessment results from different perspectives: one based on real patient data, and the other on a risk assessment model built using physician professional knowledge. Finally, a dynamic decision tree model provides the final risk assessment result for jaw cyst surgery, achieving accurate and interpretable clinical risk assessment for maxillofacial surgery.

[0018] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0019] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. In the drawings: Figure 1 This is a flowchart of an artificial intelligence-based method for assessing postoperative risk of jaw cysts according to an embodiment of the present invention. Figure 2 This is a network flowchart of the artificial intelligence-based risk assessment method for jaw cyst surgery in an embodiment of the present invention. Figure 3 This is a schematic diagram of model training for the decision tree model in an embodiment of the present invention. Figure 4 This is a structural block diagram of an artificial intelligence-based risk assessment system for jaw cyst surgery, according to an embodiment of the present invention. Detailed Implementation

[0020] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0021] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art and should not be interpreted in an idealized or overly formal sense unless specifically defined.

[0022] Figure 1 The flowchart illustrates an artificial intelligence-based method for assessing postoperative risk of jaw cysts, as provided in an embodiment of the present invention. Figure 1 As shown, the artificial intelligence-based risk assessment method for jaw cyst surgery proposed in this invention includes the following steps: S11. Obtain target clinical data, which includes patient basic information, past medical history, personal history, preoperative condition, laboratory indicators, and imaging examination data.

[0023] S12. Extract the risk features of the target clinical data, and use a pre-trained SVM machine learning risk assessment model to learn and predict the risk features to obtain the first postoperative risk assessment result.

[0024] S13. Extract the patient's basic information, abnormal indicator information, and surgical contraindication information from the target clinical data.

[0025] Furthermore, before performing step S13, the method further includes: detecting and removing abnormal statistical data in the target clinical data.

[0026] S14. The patient's basic information and abnormal indicator information are combined into search terms. Based on the search terms, a pre-built medical expert knowledge base is searched to obtain risk search results based on the expert knowledge base. The risk search results are the complications triggered by the abnormal indicator information.

[0027] S15. Construct a prompt word template based on the patient's basic information, surgical contraindications information, and risk retrieval results. Input the prompt word template into a preset large language model to generate a second postoperative risk assessment result and the basis for obtaining the current assessment result.

[0028] Based on steps S13 to S15, this embodiment of the invention innovatively constructs a postoperative risk assessment model based on a large language model, which is referred to as Struct-RAG. This embodiment employs a structured database to store the knowledge base and constructs a retrieval algorithm for risk assessment of jaw cyst surgery based on this structured knowledge base. The RAG based on this retrieval algorithm is defined as Struct-RAG. The implementation principle and working mechanism of Struct-RAG are described in detail below.

[0029] To reduce retrieval accuracy and model illusion, this invention implements RAG retrieval preprocessing through a custom Tools algorithm, ensuring the accuracy of knowledge base retrieval while eliminating some of the illusions associated with large models. Specifically, it can be divided into four Tools algorithms, whose implementation functions are detailed below: Tool 1: Abnormal Indicator Detection: The core function of this algorithm is to remove abnormal statistical data such as missing values ​​and outliers from patient clinical data, while retaining normal statistical data.

[0030] Tool2: Abnormal Indicator Extraction: The core function of this algorithm is to compare the value range of normal indicators with the normal clinical statistical data retained in Tool1, and finally obtain abnormal indicator information from the patient's normal clinical statistical data. The abnormal indicator information includes the abnormal examination items and the specific values ​​of the indicators.

[0031] Tool3: Surgical Contraindication Extraction: The core function of this algorithm is to obtain the abnormal indicators and indicator values ​​retained in Tool2, and compare them to see if they trigger surgical contraindications in order to obtain surgical contraindication information. Surgical contraindication information includes the indicator and indicator value that triggers the surgical contraindication, as well as what the specific contraindication is.

[0032] Tool 4: Extraction of basic patient information. This algorithm mainly retains basic patient information such as age, gender, body fat percentage, etc.

[0033] The following details how Struct-RAG performs risk assessment for jaw cyst surgery: 1. Query: This contains basic information about patients undergoing jaw cyst surgery, including examination results, medical records, etc., and is generally semi-structured data. 2. Tools: RAG retrieval preprocessing toolset, including anomaly detection (detecting and removing abnormal indicators in the submitted query, such as missing indicators, incorrect indicator data types, etc.), anomaly extraction (extracting abnormal indicators from the query, such as high blood pressure, the specific value of the blood pressure, smoking history, etc.), surgical contraindication extraction (obtaining contraindications and their indicator values ​​from the query), and patient basic information extraction (extracting basic patient information, such as age, gender, body fat percentage, etc.).

[0034] 3. Retrieval Tool: The obtained patient basic information and abnormal indicator information are retrieved from the knowledge base through a semi-structured query to obtain a description of the complications caused by the abnormal indicators.

[0035] 4. Prompt: The Prompt engine constructs a Prompt from the data obtained by the Tools and the search engine.

[0036] 5. LLM: Generates surgical risk assessment results and the basis for those results.

[0037] S16. When the results of the first postoperative risk assessment and the second postoperative risk assessment are inconsistent, the results of the first and second postoperative risk assessments are input into the pre-trained risk decision model, and the final postoperative risk assessment result is output.

[0038] In this embodiment, before a new patient undergoes preoperative examinations and is prepared for surgery, the patient's clinical data can be input into SVM and Struct-RAG respectively to obtain the patient's interpretable first postoperative risk assessment result A and second postoperative risk assessment result B. If the decision results of A and B are consistent, the decision result will not be input into the risk decision model trained on the C4.5 decision tree for risk decision-making; the patient's decision result directly adopts the surgical risk assessment result and assessment basis of Struct-RAG. If the decision results of A and B are inconsistent, the risk assessment results A and B are input into the risk decision model. The final maxillofacial surgery risk assessment result refers to the C4.5 classification result, and the assessment basis corresponds to SVM (if consistent with SVM) or Struct-RAG (if consistent with Struct-RAG). The network flow implementation is as follows: Figure 2 As shown.

[0039] This invention provides an artificial intelligence-based method for assessing the postoperative risk of jaw cysts. It integrates multi-source medical data, including patient basic information, personal history, preoperative conditions, examination indicators, and imaging data. Based on collected historical patient data, a machine learning risk assessment model is constructed. Furthermore, based on a built medical expert knowledge base, a Struct-RAG jaw cyst surgery risk model is innovatively constructed. These two risk assessment models provide postoperative risk assessment results from different perspectives: one is based on real patient data, and the other is based on a risk assessment model built using the doctor's professional knowledge. Finally, a dynamic decision tree model is used to provide the final risk assessment result for jaw cyst surgery, achieving accurate and interpretable clinical risk assessment for maxillofacial surgery.

[0040] The artificial intelligence-based risk assessment method for jaw cyst surgery provided in this invention first requires the collection of patient data and the construction of an expert knowledge base.

[0041] This invention utilizes a pre-developed patient data collection system for doctors, which encrypts and de-identifies patient information data and collects and stores it in real-time within the hospital's intranet. Based on this system, this invention can accurately and quickly collect information including: basic patient information, past medical history, personal history, preoperative conditions, laboratory indicators, imaging examinations, and the patient's surgical complications. Preoperative conditions include respiratory function and cardiac function classification; laboratory indicators include complete blood count, coagulation function, blood lipids, blood glucose, myocardial infarction markers, and electrolytes. Complications include 16 types as shown in Table 1. Table 1. Postoperative complications of jaw cyst surgery Furthermore, the steps for constructing the medical expert knowledge base include: obtaining a second training dataset, in which data samples include desensitized clinical data of historical patients; constructing a medical expert knowledge base based on the second training dataset and guidance from oral medicine experts, wherein the knowledge entries in the medical expert knowledge base include abnormal indicator information and the normal range of the abnormal indicator information, complications exceeding the normal value, and surgical contraindications.

[0042] The second training dataset can be selected from the data collected and stored by the aforementioned patient data collection system.

[0043] Specifically, this invention incorporates a structured medical expert knowledge base, which includes the patient's basic information, past medical history, personal history, preoperative condition, laboratory indicators, and imaging characteristics. Here, we will use hypertension in the personal history as an example for illustration.

[0044] Medical experts have provided a normal range for blood pressure, a description of complications arising from exceeding normal values, and contraindications for surgery. For example, normal blood pressure is: systolic pressure 90-139 mmHg, diastolic pressure 60-89 mmHg, pulse pressure 30-40 mmHg. Complications are described as follows: systolic pressure greater than or equal to 140 mmHg and / or diastolic pressure greater than or equal to 90 mmHg, which may cause complications such as cerebral hemorrhage and postoperative hemorrhage; systolic pressure of 140-159 mmHg leads to mild complications; 160-179 mmHg leads to moderate complications; and above 180 mmHg leads to severe complications. Contraindications for surgery are described as: the surgical risk is too high, requiring the surgery to be postponed; the blood pressure range triggering contraindications is: diastolic pressure greater than 115 mmHg, systolic pressure greater than 200 mmHg.

[0045] Other features of expert knowledge are constructed similarly to hypertension in a person's personal history. In this embodiment of the invention, the expert knowledge base is stored in a lightweight SQLite database, which allows for structured querying of relevant expert knowledge.

[0046] The collection of patient data will be used for training and building machine learning risk assessment models; the construction of a medical expert knowledge base will be used to build the Struct-RAG surgical risk assessment system.

[0047] In this embodiment of the invention, before extracting risk characteristics from the target clinical data, the method includes determining risk characteristics for postoperative risk assessment. Determining risk characteristics for postoperative risk assessment specifically includes the following steps: Obtain the first training dataset, which includes desensitized historical patients' clinical data and surgical complications; the second training dataset can be selected from the data collected and stored by the aforementioned patient data collection system.

[0048] Extract the original data features of each data sample in the first training dataset.

[0049] Dummy variable transformation and normalization are performed on each feature information in the original data.

[0050] Each processed feature was compared with the surgical complication results using an independent samples t-test to remove features from the original data that were not related to the surgical complication results.

[0051] Based on the Pearson correlation coefficient, feature filtering is performed on the remaining feature information after removing irrelevant feature information to remove highly redundant strongly correlated features.

[0052] We construct a minimum absolute shrinkage and selection operator Lasso regression model. During the feature selection regression process, we penalize the feature coefficients to compress the coefficients of unimportant features to 0, while retaining the features with non-zero coefficients. The selected core features are then used as risk features for machine learning.

[0053] This invention first constructs 66 machine learning features based on collected patient data, including seven major categories: basic patient information, past medical history, personal history, preoperative conditions, surgery-related information, laboratory indicators, and imaging examinations, such as complete blood count (CBC) among the laboratory indicators. Then, feature engineering is performed on these 66 original features, retaining the most important features as risk features for machine learning.

[0054] The aforementioned risk features are also used to train the machine learning risk assessment model.

[0055] In this embodiment of the invention, the training steps of the machine learning risk assessment model include: Obtain the first training dataset, which includes data samples from desensitized historical patients' clinical data and surgical complications.

[0056] Risk features are extracted from each data sample in the first training dataset, and a machine learning risk assessment model is trained based on the risk features of each data sample and the corresponding surgical complications.

[0057] During the training of the machine learning risk assessment model, the GridSearchCV method is used to optimize the key parameters of the model to determine the optimal parameter combination and obtain the optimal machine learning risk assessment model. GridSearch is a parameter tuning technique; exhaustive search: among all candidate parameter selections, every possibility is tried through a loop, and the parameter that performs best is the final result.

[0058] Furthermore, the method also includes: using the prediction function of the optimal machine learning risk assessment model as the input of the SHAP interpreter, randomly selecting a portion of the samples in the training dataset as the background dataset, and calculating the baseline contribution value of each risk feature; randomly selecting a portion of the samples in the training dataset as the test dataset, calling the shap_values() method of the SHAP interpreter, and calculating the SHAP value corresponding to each core feature in each test sample in the test dataset. The SHAP values ​​of each core feature in all test samples are displayed in a graphical visualization.

[0059] To demonstrate the importance of each feature in interpretability, this embodiment of the invention performs SHAP analysis. The core idea of ​​SHAP originates from the Shapley value in cooperative game theory. The Shapley value is used to fairly distribute the benefits brought by multiple participants in cooperation. SHAP introduces this concept into the interpretation of machine learning models to calculate the contribution of each feature to the model's prediction results. Through SHAP, the positive and negative impacts and the degree of impact of each feature on risk assessment can be clearly obtained.

[0060] The training steps of the risk decision-making model in this embodiment of the invention specifically include the following: Obtain a third training dataset, whose data samples include desensitized historical patients' clinical data and surgical complications; Obtain the first postoperative risk assessment result and the second postoperative risk assessment result corresponding to each data sample in the third training dataset; Using the first and second postoperative risk assessment results as inputs and the surgical complications of the corresponding data samples as labels, a decision tree algorithm is used for modeling and training to obtain a risk decision model.

[0061] In this embodiment of the invention, the SVM machine learning risk assessment model, trained on real clinical data, utilizes authentic clinical data, representing a value extraction from clinical data. Meanwhile, the Struct-RAG large-scale risk assessment model aims to simulate an expert's decision-making system based on a large language model, with the knowledge base behind Struct-RAG derived from expert experience. To address the need for both in-depth risk assessment of clinical data and integration of oral medicine expert experience in jaw cyst surgery risk assessment decisions, this embodiment proposes a dynamic risk assessment method based on decision trees.

[0062] Specifically, based on historical training data, this embodiment of the invention trained a risk assessment algorithm SVM based on machine learning and a risk assessment model based on Struct-RAG. This embodiment of the invention also constructed a second batch of data in the same manner. This data will be used directly to perform risk assessment inference using SVM and Struct-RAG. The risk assessment results and the true labels of this data will be used to model and train the C4.5 decision tree algorithm. The training process of the C4.5 decision tree model is as follows: Figure 3 As shown. Based on the trained decision tree model C4.5, this embodiment of the invention constructs a dynamic risk assessment decision-making strategy. This strategy is simple yet effective, embodying the process of "dynamic" risk assessment decision-making.

[0063] For the sake of simplicity, the method embodiments are described as a series of actions. However, those skilled in the art should understand that the embodiments of the present invention are not limited to the described order of actions, because according to the embodiments of the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions involved are not necessarily essential to the embodiments of the present invention.

[0064] This invention provides an artificial intelligence-based risk assessment system for jaw cyst surgery, the system comprising a functional module for implementing the artificial intelligence-based risk assessment method for jaw cyst surgery as described in any of the preceding claims. Figure 4 This schematic diagram illustrates the structure of an artificial intelligence-based risk assessment system for jaw cyst surgery, as provided in an embodiment of the present invention. (Refer to...) Figure 4 The system described in this embodiment of the invention includes: The acquisition module 401 is used to acquire target clinical data, which includes patient basic information, past medical history, personal history, preoperative conditions, laboratory indicators and imaging examination data. The first risk assessment module 402 is used to extract the risk features of the target clinical data, and to learn and predict the risk features using a pre-trained SVM machine learning risk assessment model to obtain the first postoperative risk assessment result. The prompt word information extraction module 403 is used to extract the patient's basic information, abnormal indicator information, and surgical contraindication information from the target clinical data; The prompt word information retrieval module 404 is used to merge the patient's basic information and abnormal indicator information into a search word, search a pre-built medical expert knowledge base based on the search word, and obtain risk retrieval results based on the expert knowledge base. The risk retrieval results are the complications triggered by the abnormal indicator information. The second risk assessment module 405 is used to construct a prompt word template based on the patient's basic information, surgical contraindication information and risk retrieval results, input the prompt word template into a preset large language model, generate a second postoperative risk assessment result and the basis for obtaining the current assessment result; The decision module 406 is used to input the first and second postoperative risk assessment results into the pre-trained risk decision model and output the final postoperative risk assessment result when the first postoperative risk assessment result is inconsistent with the second postoperative risk assessment result.

[0065] This invention provides an artificial intelligence-based risk assessment system for jaw cyst surgery, further comprising a feature selection module (not shown in the accompanying drawings). This feature selection module is used to determine risk features for postoperative risk assessment before extracting risk features from the target clinical data. Determining risk features for postoperative risk assessment specifically includes: acquiring a first training dataset, the data samples in which desensitized historical patient clinical data and surgical complications are included; extracting the original data features of each data sample in the first training dataset; performing dummy variable transformation and normalization on each feature information in the original data features; performing independent samples t-tests on each processed feature information with the surgical complication results to remove feature information in the original data features that is irrelevant to the surgical complication results; performing feature filtering on the remaining feature information after removing irrelevant feature information based on the Pearson correlation coefficient to remove highly redundant strongly correlated features; constructing a minimum absolute contraction and selection operator Lasso regression model, and penalizing the feature coefficients during the feature selection regression process to compress the coefficients of unimportant feature information to 0, retaining feature information with non-zero coefficients, and using the selected core features as risk features for machine learning.

[0066] This invention provides an artificial intelligence-based risk assessment system for jaw cyst surgery, which also includes an information filtering module (not shown in the accompanying drawings). The information filtering module is used to detect and remove abnormal statistical data in the target clinical data before the prompt word information extraction module 403 extracts the patient's basic information, abnormal indicator information, and surgical contraindication information from the target clinical data.

[0067] The artificial intelligence-based risk assessment system for jaw cyst surgery provided in this invention achieves interpretable and accurate risk assessment for jaw cyst surgery. The system includes a first risk assessment module, which is a machine learning-based maxillofacial surgery risk assessment model constructed based on real clinical data; a second risk assessment module, which is a novel Struct-RAG maxillofacial surgery risk assessment model based on an expert knowledge base and enhanced retrieval algorithm; and a dynamic risk assessment decision-making process based on decision trees. This enables the system to accurately and interpretably assess the risk of maxillofacial surgery in clinical patients, and the results have been validated in clinical practice.

[0068] As the system implementation is basically similar to the method implementation, the description is relatively simple, and relevant parts can be found in the description of the method implementation.

[0069] Furthermore, another embodiment of the present invention provides a computer program product storing a computer program that, when executed by a processor, implements the steps described in the above embodiment of the artificial intelligence-based method for assessing postoperative risk of jaw cysts, for example... Figure 1 Steps S11-S16 are shown.

[0070] Furthermore, another embodiment of the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When executed by the processor, the computer program performs the steps described in the above embodiment of the artificial intelligence-based method for assessing postoperative risk of jaw cysts. Figure 1 Steps S11-S16 are shown.

[0071] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are intended to be within the scope of the invention and form different embodiments. For example, any of the claimed embodiments can be used in any combination.

[0072] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for assessing postoperative risk of jaw cysts based on artificial intelligence, characterized in that, The method includes: Acquire target clinical data, which includes patient basic information, past medical history, personal history, preoperative condition, laboratory indicators, and imaging examination data; Risk features of the target clinical data are extracted, and a pre-trained SVM machine learning risk assessment model is used to learn and predict the risk features to obtain the first postoperative risk assessment result. Extract basic patient information, abnormal indicator information, and surgical contraindication information of the patients to be evaluated from the target clinical data; The patient's basic information and abnormal indicator information are combined into search terms. Based on the search terms, a pre-built medical expert knowledge base is searched to obtain risk search results based on the expert knowledge base. The risk search results are the complications triggered by the abnormal indicator information. Based on the patient's basic information, surgical contraindications information, and risk retrieval results, a prompt word template is constructed. The prompt word template is then input into a preset large language model to generate a second postoperative risk assessment result and the basis for obtaining the current assessment result. When the results of the first postoperative risk assessment and the second postoperative risk assessment are inconsistent, the results of the first and second postoperative risk assessments are input into the pre-trained risk decision model, and the final postoperative risk assessment result is output.

2. The method according to claim 1, characterized in that, Before extracting risk characteristics from the target clinical data, the method includes determining risk characteristics for postoperative risk assessment; Identify risk characteristics for postoperative risk assessment, including: Obtain the first training dataset, which includes data samples from desensitized historical patients' clinical data and surgical complications. Extract the original data features of each data sample in the first training dataset; Dummy variable transformation and normalization are performed on each feature information in the original data. Each processed feature was compared with the surgical complication results using an independent samples t-test to remove features from the original data that were not related to the surgical complication results. Based on the Pearson correlation coefficient, feature filtering is performed on the remaining feature information after removing irrelevant feature information in order to remove highly redundant strongly correlated features; We construct a minimum absolute shrinkage and selection operator Lasso regression model. During the feature selection regression process, we penalize the feature coefficients to compress the coefficients of unimportant features to 0, while retaining the features with non-zero coefficients. The selected core features are then used as risk features for machine learning.

3. The method according to claim 1, characterized in that, Before extracting the patient's basic information, abnormal indicators, and surgical contraindications from the target clinical data, the method further includes: Detect and remove abnormal statistical data from the target clinical data.

4. The method according to claim 1, characterized in that, The training steps for the machine learning risk assessment model include: Obtain the first training dataset, which includes data samples from desensitized historical patients' clinical data and surgical complications. Extract risk features from each data sample in the first training dataset, and train a machine learning risk assessment model based on the risk features of each data sample and the corresponding surgical complications. During the training process of the machine learning risk assessment model, the key parameters of the machine learning risk assessment model are optimized to determine the optimal parameter combination of the machine learning risk assessment model, thereby obtaining the optimal machine learning risk assessment model.

5. The method according to claim 4, characterized in that, The method further includes: Using the prediction function of the optimal machine learning risk assessment model as the input to the SHAP interpreter, a portion of the samples in the training dataset are randomly selected as the background dataset to calculate the baseline contribution value of each risk feature. Randomly select a portion of the samples from the training dataset as the test dataset, and calculate the SHAP value corresponding to each core feature in each test sample of the test dataset; The SHAP values ​​of each core feature in all test samples are displayed in a graphical visualization.

6. The method according to claim 1, characterized in that, The steps for constructing the medical expert knowledge base include: Obtain a second training dataset, in which data samples include desensitized clinical data of historical patients. A medical expert knowledge base is constructed based on the second training dataset and the guidance of oral medicine experts. The knowledge entries in the medical expert knowledge base include abnormal indicator information, the normal value range of the abnormal indicator information, complications exceeding the normal value, and surgical contraindications.

7. The method according to claim 1, characterized in that, The training steps for the risk decision-making model include: Obtain a third training dataset, whose data samples include desensitized historical patients' clinical data and surgical complications; Obtain the first postoperative risk assessment result and the second postoperative risk assessment result corresponding to each data sample in the third training dataset; Using the first and second postoperative risk assessment results as inputs and the surgical complications of the corresponding data samples as labels, a decision tree algorithm is used for modeling and training to obtain a risk decision model.

8. An artificial intelligence-based risk assessment system for jaw cyst surgery, characterized in that, The system includes: The acquisition module is used to acquire target clinical data, which includes patient basic information, past medical history, personal history, preoperative conditions, laboratory indicators, and imaging examination data. The first risk assessment module is used to extract the risk features of the target clinical data, and to learn and predict the risk features using a pre-trained SVM machine learning risk assessment model to obtain the first postoperative risk assessment result. The prompt word information extraction module is used to extract basic patient information, abnormal indicator information, and surgical contraindication information of the patient to be evaluated from the target clinical data; The prompt word information retrieval module is used to merge the patient's basic information and abnormal indicator information into search words, search a pre-built medical expert knowledge base based on the search words, and obtain risk retrieval results based on the expert knowledge base. The risk retrieval results are the complications triggered by the abnormal indicator information. The second risk assessment module is used to construct a prompt word template based on the patient's basic information, surgical contraindications information and risk retrieval results. The prompt word template is then input into a preset large language model to generate a second postoperative risk assessment result and the basis for obtaining the current assessment result. The decision module is used to input the first and second postoperative risk assessment results into the pre-trained risk decision model and output the final postoperative risk assessment result when the first postoperative risk assessment result is inconsistent with the second postoperative risk assessment result.

9. The system according to claim 8, characterized in that, The system also includes: The feature screening module is used to determine risk features for postoperative risk assessment before extracting risk features from the target clinical data; Identify risk characteristics for postoperative risk assessment, including: Obtain the first training dataset, which includes data samples from desensitized historical patients' clinical data and surgical complications. Extract the original data features of each data sample in the first training dataset; Dummy variable transformation and normalization are performed on each feature information in the original data. Each processed feature was compared with the surgical complication results using an independent samples t-test to remove features from the original data that were not related to the surgical complication results. Based on the Pearson correlation coefficient, feature filtering is performed on the remaining feature information after removing irrelevant feature information in order to remove highly redundant strongly correlated features; We construct a minimum absolute shrinkage and selection operator Lasso regression model. During the feature selection regression process, we penalize the feature coefficients to compress the coefficients of unimportant features to 0, while retaining the features with non-zero coefficients. The selected core features are then used as risk features for machine learning.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 7.

Images