Method for establishing a craniocerebral trauma prognosis model based on multi-modal data fusion and ai

Abstract

This invention provides a method and system for establishing a prognostic model for traumatic brain injury (TBI) based on multimodal data fusion and AI, belonging to the field of medical artificial intelligence technology. The method includes collecting patient clinical text and CT image data, performing preprocessing, extracting clinical text features and image data features, and constructing three major models: a TBI prognostic prediction model, a clinical model of the predictive efficacy of edema expansion on TBI prognosis, and a TBI cerebral edema expansion prediction model. The final output of the prognostic model prediction results includes patient risk stratification results, neurological function prognostic prediction results, and cerebral edema expansion prediction results. The system includes a data acquisition module, a data preprocessing module, a feature extraction module, a multimodal model construction module, and a result output module. This invention can accurately achieve risk stratification and prognostic prediction for TBI, effectively improving the accuracy and efficiency of diagnosis and treatment, reducing the risk of misdiagnosis and missed diagnosis, and providing strong support for improving patient prognosis.

Description

Technical Field

[0001] This invention relates to the field of medical artificial intelligence technology, and in particular to a method and system for establishing a prognostic model for craniocerebral trauma based on multimodal data fusion and AI. Background Technology

[0002] Traumatic brain injury (TBI) is a major global health and socioeconomic challenge. It has a high incidence rate, rapid disease progression, and high mortality and disability rates, causing enormous losses and a heavy burden on patients, their families, and society. In military scenarios, war traumatic brain injury (bTBI) has the highest disability and mortality rates among all types of combat trauma. The effective treatment window for TBI is extremely short, and early and accurate diagnosis and intervention are crucial for improving patient prognosis.

[0003] Currently, the diagnosis, severity assessment, and treatment planning of total brain injury (TBI) in clinical practice mainly rely on the clinical experience, physical examination, and imaging interpretation of senior neurologists. However, junior frontline physicians and primary care units often lack extensive experience and imaging interpretation skills, and traditional diagnostic methods suffer from high subjectivity and insufficient accuracy, making it difficult to achieve precise risk stratification and individualized treatment for TBI. Furthermore, existing research has not fully explored the imaging markers related to the prognosis of TBI, and there is a lack of auxiliary models that can rapidly and intelligently assess secondary damage caused by TBI, failing to meet the clinical need for precision diagnosis and treatment.

[0004] Artificial intelligence (AI), as an emerging science and technology, has seen significant applications in the field of medical imaging, particularly in areas such as tumor benign / malignant diagnosis, image segmentation, and prognostic prediction, based on machine learning or deep learning. Currently, AI can utilize multiple computer vision, machine learning, and deep learning technologies in the medical field. In the clinical diagnosis and treatment of tumor-related brain injury (TBI), computed tomography (CT) is one of the most commonly used examination methods. By acquiring high-resolution images, it determines the location, type, and severity of TBI, playing a crucial role in the diagnosis and treatment of TBI patients. However, in the field of TBI diagnosis and treatment, existing technologies often focus on single-modal data, such as using only CT images, without fully integrating multimodal information from clinical text and imaging data. This results in limited model predictive efficiency, making it difficult to comprehensively reflect the patient's condition and provide comprehensive auxiliary decision support for clinical practice. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for establishing a prognostic model for traumatic brain injury (TBI) based on multimodal data fusion and AI. It uses artificial intelligence technology (AI) to construct a TBI risk stratification model, mine imaging markers related to TBI prognosis, construct a clinical model of the predictive efficacy of edema expansion on TBI prognosis, and finally use machine learning methods to construct a predictive model of TBI cerebral edema expansion. This helps clinicians make personalized TBI treatment decisions, thereby improving patient prognosis.

[0006] To achieve the above objectives, this invention proposes a method for establishing a prognostic model for traumatic brain injury based on multimodal data fusion and AI, comprising the following steps: Step S1: Collect clinical text and imaging data of patients with traumatic brain injury; Step S2: Preprocess the clinical text and image data; Step S3: Automatically extract clinical text feature I and image data feature I using a deep learning model; Step S4: Construct a multimodal attention module, manually extract image data feature II and clinical text feature II required for constructing the TBI prognostic prediction model, based on DenseNet, combine gated multimodal units to fuse image data feature II and clinical text feature II, construct the TBI prognostic prediction model, use focal loss to alleviate class imbalance, verify the predictive effect of the TBI prognostic prediction model, and perform statistical analysis on the TBI prognostic prediction model; Step S5: Define the prognostic criteria for neurological function, divide the initial hematoma volume range, set the follow-up time window to draw the hematoma time-volume curve, define the indicators of hematoma expansion and edema expansion, construct a clinical model of the predictive efficacy of edema expansion on TBI prognosis, define delayed edema expansion (DPE), validate the clinical model of the predictive efficacy of edema expansion on TBI prognosis, and perform statistical analysis on the clinical model of the predictive efficacy of edema expansion on TBI prognosis. Step S6: Extract the clinical text feature III and image data feature III required for constructing the TBI brain edema expansion prediction model; extract and filter the image group data feature III from the image data feature III; fuse the clinical text feature III and the processed image data feature III; construct the TBI brain edema expansion prediction model based on machine learning methods; construct training sample sets for the early edema expansion research cohort and the delayed edema expansion research cohort; select the model with the highest area under the curve (AUC) value as the TBI brain edema expansion prediction model classifier; train and verify the performance of the TBI brain edema expansion prediction model; and perform statistical analysis on the TBI brain edema expansion prediction model. Step S7: Output the prognostic model prediction results, including patient risk stratification results, neurological function prognostic prediction results, and cerebral edema expansion prediction results. The results are evaluated using the Glasgow Outcome Scale and the modified Rankin Scale.

[0007] Preferably, in step S1, the clinical text is derived from the patient's electronic medical record, including medical documents, laboratory reports, demographic data and treatment records. The imaging data includes the patient's initial head imaging information within 24 hours of admission and subsequent follow-up imaging information. The data inclusion criteria are patients aged 10 to 80 years who meet the diagnostic criteria for traumatic brain injury, and cases with incomplete medical history or laboratory results or missing head examination imaging information are excluded.

[0008] Preferably, in step S2, clinical text preprocessing includes anonymization, missing data supplementation, and standardization, supplementing missing age, gender, and past medical history data by retrieving medical records; image data preprocessing includes image cropping, window width and level settings and adjustments, data augmentation, and region of interest segmentation.

[0009] Preferably, in step S3, clinical text feature I includes patient demographic information, Glasgow Coma Scale (GCS) score, symptoms of headache, vomiting, and coma, hematoma rupture into the ventricle, past medical history, and treatment methods. Past medical history includes hemorrhagic stroke, ischemic stroke, hypertension, diabetes, hyperlipidemia, coronary heart disease, heart failure, arrhythmia, anticoagulation therapy, and antiplatelet therapy. Treatment methods include conservative drug therapy (CMT), conventional craniotomy (CC), and minimally invasive treatment (MIA). MIA includes stereotactic hematoma aspiration and neuroendoscopic hematoma aspiration. Imaging data feature I includes traditional imaging feature I and radiomics feature I. Traditional imaging feature I includes the location of the hemorrhage and skull fracture status. Radiomics feature I includes histogram features and texture features.

[0010] Preferably, when extracting clinical text feature I, a pre-set case report form (CRF) is used for extraction; when extracting image data feature I, a deep learning model is used to automatically segment the hematoma and perihemorrhagic edema in standard DICOM format image data. Subsequently, medical image processing software is used to manually calibrate the region of interest in the image data automatically segmented by the deep learning model. Among these features, traditional image data feature I is extracted using the DenseNet46 method, and radiomics feature I is extracted from the segmented image data.

[0011] Preferably, step S4 includes the following steps: Step S41: Construct a multimodal attention module including Grid, SE, Non-local, and Proposed to identify image regions in the image data that contribute to the prediction of poor prognosis of TBI; Step S42: Using medical image processing software, manually identify and delineate hematomas on imaging slices containing lesions in image regions that contribute to poor prognosis prediction of TBI. Review the accuracy of hematoma delineation, integrate all delineated slices for each patient into a complete segmentation file, and automatically calculate and output the hematoma volume using medical image processing software as image data feature II. Researchers independently record clinical text, and use a pre-designed CRF table to collect clinical characteristics and demographic data for each patient as clinical text feature II. Step S43: The extracted image data feature II and clinical text feature II are combined with DenseNet, which is densely connected by a convolutional neural network (CNN) structure. By constructing a gated multimodal unit (GMU), the clinical text feature II and image data feature II are fused, and adaptive weights are set in the subspace to construct a prognostic prediction model for traumatic brain injury (TBI). Step S44: Train the fused clinical text feature II and image data feature II, set the batch size and initial learning rate, use focal loss to alleviate the imbalance of prediction categories, and use 5-fold cross-validation to verify the prediction effect of the traumatic brain injury (TBI) prognosis prediction model. Step S45: Perform statistical analysis on the prognostic prediction model for traumatic brain injury (TBI).

[0012] Preferably, step S5 includes the following steps: Step S51: Define the prognostic criteria for neurological function and use the modified Rankin Scale (mRS) to evaluate the neurological functional outcomes of TBI patients; Step S52: Divide the initial hematoma volume range, set the follow-up time window, and plot the hematoma time-volume curve; Step S53: Define hematoma expansion and edema expansion, and construct a clinical model of the predictive efficacy of edema expansion on TBI prognosis. The specific steps are as follows: Step S531: Define the early absolute hematoma expansion value as the difference between the hematoma volume examined within 48 to 72 hours after onset and the initial hematoma volume within 24 hours after onset. The calculation formula is as follows: ; in, This represents the early absolute hematoma enlargement value. The hematoma volume should be re-examined within 48 to 72 hours after the onset of illness. The initial hematoma volume within 24 hours of onset; Step S532: Define the early absolute edema expansion value as the difference between the edema volume examined within 48 to 72 hours after onset and the initial edema volume within 24 hours after onset. The calculation formula is as follows: ; in, This represents the value of early absolute edema expansion. The edema volume should be re-examined within 48 to 72 hours after the onset of illness. The initial edema volume within 24 hours of onset; Step S533: Calculate the early relative edema expansion value X, using the following formula: ; in, This represents the value of early relative edema expansion; Step S534: Calculate the early absolute hematoma combined with edema expansion value Z. The calculation formula is as follows: ; in, This represents the value of early-stage absolute hematoma combined with edema expansion. The total lesion volume should be re-examined within 48 to 72 hours after the onset of illness. The initial total lesion volume within 24 hours of onset; Step S54: Based on the definitions of early absolute hematoma expansion, absolute edema expansion, relative edema expansion, and combined hematoma and edema expansion, plot the receiver operating characteristic (ROC) curve and calculate the sensitivity and specificity for predicting poor patient prognosis. Step S55: Define delayed edema expansion (DPE) and validate the clinical model of the predictive efficacy of edema expansion on TBI prognosis. The specific steps are as follows: Step S551: Based on the plotted hematoma time-volume curve, determine the peak and stable periods of edema growth, and define the specific time window for delaying edema expansion. Step S552: Using the defined delay period as the time range, calculate the change in edema volume at different time points within this period, initially defining delayed edema expansion. The calculation formula is as follows: ; in, To delay the expansion of edema, This represents the edema volume at subsequent time points in the delayed period. This represents the edema volume at the initial time point of the delay period. Step S553: ​​Plot the operating characteristic curve (ROC) of delayed edema expansion (DPE), evaluate the predictive performance of ROC on poor prognosis of TBI patients, determine the absolute edema volume expansion threshold of delayed edema expansion using the Youden method, determine the quantitative definition of delayed edema expansion, and verify the clinical model performance of edema expansion on TBI prognosis. Step S56: Perform statistical analysis on the clinical model of the predictive efficacy of edema expansion on TBI prognosis.

[0013] Preferably, step S6 includes the following steps: Step S61: Extract clinical text feature III and image data feature III that may be related to the expansion of edema in the clinical text feature I and image data feature I. Use the 3DUnet segmentation network to segment the skull and edema or hematoma, calculate radiomics features and screen radiomics features. Then extract the histogram features and texture features of edema and hematoma as radiomics features III in image data feature III. Step S62: Screen the radiomics features III. The specific steps are as follows: Step S621: Calculate the Pearson coefficients of radiomics features and edema expansion label, and retain the 100 variables most relevant to the edema expansion label to form 100-dimensional radiomics features; Step S622: Select an algorithm to finally filter out the 5 radiomics features most relevant to the edema expansion label from the 100-dimensional radiomics features; Step S63: The extracted and processed clinical text features III and image data features III are fused together using a simple splicing method to form a 31-dimensional feature input vector, and a TBI brain edema expansion prediction model is constructed based on machine learning methods. Step S64: Match the 31-dimensional feature input vector with the corresponding edema expansion label to construct training sample sets for the early edema expansion research cohort and the delayed edema expansion research cohort, respectively; Step S65: Select four classic machine learning classifiers: Random Forest (RF), Logistic Regression (LR), TabNet, and Catboost. Use the area under the curve (AUC) as the classifier performance metric. Calculate the AUC values ​​of the four classifiers in the early edema expansion cohort and the delayed edema expansion cohort respectively. Select the classifier model with the highest AUC value as the TBI brain edema expansion prediction model classifier. Step S66: Train and validate the TBI cerebral edema expansion prediction model using a 4-fold cross-validation method; Step S67: Perform statistical analysis on the TBI cerebral edema expansion prediction model.

[0014] Preferably, SPSS 24.0 was used to perform statistical analysis on the constructed prognostic prediction model for traumatic brain injury (TBI), the clinical model of the predictive efficacy of edema expansion on TBI prognosis, and the TBI cerebral edema expansion prediction model. The normality of the data was verified by the Kolmogorov-Smirnov test. Normally distributed data were analyzed by t-test or one-way ANOVA, and non-normally distributed data were analyzed by Mann-Whitney U test or Kruskal-Wallis H test. Categorical variables were analyzed by chi-square test. Significant predictive factors in univariate analysis were included in multivariate logistic regression analysis. The effectiveness of the model was evaluated by plotting ROC curves and calculating the area under the ROC curve (AUC), sensitivity, specificity, and 95% confidence interval. The DeLong method was used to compare the ROC curves.

[0015] This invention also proposes a system for establishing a prognostic model for craniocerebral trauma based on multimodal data fusion and AI, including a data acquisition module, a data preprocessing module, a feature extraction module, a multimodal model construction module, and a result output module; The data acquisition module includes a clinical text acquisition unit, an image data acquisition unit, and a data filtering unit; The data preprocessing module is connected to the data acquisition module and includes a clinical text preprocessing unit and an image data preprocessing unit. The feature extraction module is connected to the data preprocessing module and includes a clinical text feature extraction unit, a traditional image feature extraction unit, and a radiomics feature extraction unit. The multimodal model building module is connected to the feature extraction module, including a TBI prognostic prediction model building unit, a clinical model building unit for the predictive efficacy of edema expansion on TBI prognosis, and a TBI brain edema expansion prediction model building unit. The results output module is connected to the multimodal model construction module.

[0016] Therefore, this invention proposes a method and system for establishing a prognostic model for traumatic brain injury based on multimodal data fusion and AI, with the following beneficial effects: (1) This invention provides doctors with more objective and accurate auxiliary decision-making basis through artificial intelligence technology, which improves the efficiency and quality of doctors' diagnosis and treatment and reduces the risk of misdiagnosis and missed diagnosis.

[0017] (2) The novel risk stratification model and the comprehensive prognostic assessment system of this invention enable precise stratification and personalized treatment of patients with traumatic brain injury, avoid unnecessary examinations and treatments, and improve the efficiency of medical resource utilization.

[0018] (3) This invention uses artificial intelligence technology to assist doctors in conducting remote consultations and online consultations, breaking the geographical limitations of diagnosis and treatment and providing efficient medical services for patients in other places.

[0019] (4) The intelligent system of the present invention automatically preprocesses image data, extracts key features, trains models, realizes process automation, and improves the work efficiency of doctors. Attached Figure Description

[0020] Figure 1 A flowchart illustrating a method for establishing a prognostic model for traumatic brain injury based on multimodal data fusion and AI; Figure 2 This is a schematic diagram of a system structure for establishing a prognostic model for craniocerebral trauma based on multimodal data fusion and AI. Detailed Implementation

[0021] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0022] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0023] Example 1 like Figure 1 As shown, this invention provides a method for establishing a prognostic model for traumatic brain injury based on multimodal data fusion and AI. It combines multiple technologies such as medical image processing, machine learning, and deep learning to comprehensively and systematically analyze imaging data of traumatic brain injury, improving the accuracy and reliability of the prognostic model. The method includes the following steps: Step S1: Collect clinical text and imaging data of patients with traumatic brain injury. Clinical text is derived from the patient's electronic medical record, including medical documents, laboratory reports, demographic data and treatment records. Imaging data includes the patient's initial head imaging information within 24 hours of admission and subsequent follow-up imaging information. The data inclusion criteria are patients aged 10 to 80 years who meet the diagnostic criteria for traumatic brain injury. Cases with incomplete medical history or laboratory results, or missing head examination imaging information are excluded.

[0024] Step S2: Preprocess the clinical text and image data. Clinical text preprocessing includes anonymization, missing data supplementation, and standardization. Missing data such as age, gender, and past medical history are supplemented by retrieving medical records. Image data preprocessing includes image cropping, window width and level settings and adjustments, data augmentation, and region of interest segmentation.

[0025] Step S3: Utilize a deep learning model to automatically extract clinical text features I and imaging data features I. Clinical text features I include patient demographic information, Glasgow Coma Scale (GCS) score, symptoms of headache, vomiting, and coma, hematoma rupture into the ventricles, past medical history, and treatment methods. Past medical history includes hemorrhagic stroke, ischemic stroke, hypertension, diabetes, hyperlipidemia, coronary heart disease, heart failure, arrhythmia, anticoagulation therapy, and antiplatelet therapy. Treatment methods include conservative drug therapy (CMT), conventional craniotomy (CC), and minimally invasive treatment (MIA). MIA includes stereotactic hematoma aspiration and neuroendoscopic hematoma aspiration. Imaging data features I include traditional imaging features I and radiomics features I. Traditional imaging features I include the location of the hemorrhage and the extent of skull fracture. Radiomics features I include histogram features and texture features.

[0026] Clinical text feature I was extracted using a pre-defined case report form (CRF). Image data feature I was extracted using a deep learning model to automatically segment hematoma and perihemorrhagic edema in standard DICOM format image data. Subsequently, medical image processing software was used to manually calibrate the regions of interest in the image data automatically segmented by the deep learning model. Traditional image data feature I was extracted using the DenseNet46 method, while radiomics feature I was extracted from the segmented image data.

[0027] Step S4 involves constructing a multimodal attention module. Image data feature II and clinical text feature II, required for building a traumatic brain injury (TBI) prognostic prediction model, are manually extracted. Based on DenseNet, image data feature II and clinical text feature II are fused using gated multimodal units to construct the TBI prognostic prediction model. Focal loss is used to alleviate class imbalance. The predictive performance of the TBI prognostic prediction model is verified, and statistical analysis is performed. The specific steps are as follows: Step S41: Construct a multimodal attention module including Grid, SE, Non-local, and Proposed to identify image regions in the image data that contribute to the prediction of poor prognosis of TBI; Step S42: Using medical image processing software, manually identify and delineate hematomas on imaging slices containing lesions in image regions that contribute to poor prognosis prediction of TBI. Review the accuracy of hematoma delineation, integrate all delineated slices for each patient into a complete segmentation file, and automatically calculate and output the hematoma volume using medical image processing software as image data feature II. Researchers independently record clinical text, and use a pre-designed CRF table to collect clinical characteristics and demographic data for each patient as clinical text feature II. Step S43: The extracted image data feature II and clinical text feature II are combined with DenseNet, which is densely connected by a convolutional neural network (CNN) structure. By constructing a gated multimodal unit (GMU), the clinical text feature II and image data feature II are fused, and adaptive weights are set in the subspace to construct a prognostic prediction model for traumatic brain injury (TBI). Step S44: Train the fused clinical text feature II and image data feature II, set the batch size and initial learning rate, use focal loss to alleviate the imbalance of prediction categories, and use 5-fold cross-validation to verify the prediction effect of the traumatic brain injury (TBI) prognosis prediction model. Step S45: Perform statistical analysis on the prognostic prediction model for traumatic brain injury (TBI).

[0028] Step S5: Define the prognostic criteria for neurological function, divide the initial hematoma volume range, set the follow-up time window to plot the hematoma time-volume curve, define the indices for hematoma expansion and edema expansion, construct a clinical model for the predictive efficacy of edema expansion on TBI prognosis, define delayed edema expansion (DPE), validate the clinical model for the predictive efficacy of edema expansion on TBI prognosis, and perform statistical analysis on the clinical model for the predictive efficacy of edema expansion on TBI prognosis. The specific steps are as follows: Step S51: Define the prognostic criteria for neurological function and use the modified Rankin Scale (mRS) to evaluate the neurological functional outcomes of TBI patients; Step S52: Divide the initial hematoma volume range, set the follow-up time window, and plot the hematoma time-volume curve; Step S53: Define hematoma expansion and edema expansion, and construct a clinical model of the predictive efficacy of edema expansion on TBI prognosis. The specific steps are as follows: Step S531: Define the early absolute hematoma expansion value as the difference between the hematoma volume examined within 48 to 72 hours after onset and the initial hematoma volume within 24 hours after onset. The calculation formula is as follows: ; in, This represents the early absolute hematoma enlargement value. The hematoma volume should be re-examined within 48 to 72 hours after the onset of illness. The initial hematoma volume within 24 hours of onset; Step S532: Define the early absolute edema expansion value as the difference between the edema volume examined within 48 to 72 hours after onset and the initial edema volume within 24 hours after onset. The calculation formula is as follows: ; in, This represents the value of early absolute edema expansion. The edema volume should be re-examined within 48 to 72 hours after the onset of illness. The initial edema volume within 24 hours of onset; Step S533: Calculate the early relative edema expansion value X, using the following formula: ; in, This represents the value of early relative edema expansion; Step S534: Calculate the early absolute hematoma combined with edema expansion value Z. The calculation formula is as follows: ; in, This represents the value of early-stage absolute hematoma combined with edema expansion. The total lesion volume should be re-examined within 48 to 72 hours after the onset of illness. The initial total lesion volume within 24 hours of onset; Step S54: Based on the definitions of early absolute hematoma expansion, absolute edema expansion, relative edema expansion, and combined hematoma and edema expansion, plot the receiver operating characteristic (ROC) curve and calculate the sensitivity and specificity for predicting poor patient prognosis. Step S55: Define delayed edema expansion (DPE) and validate the clinical model of the predictive efficacy of edema expansion on TBI prognosis. The specific steps are as follows: Step S551: Based on the plotted hematoma time-volume curve, determine the peak and stable periods of edema growth, and define the specific time window for delaying edema expansion. Step S552: Using the defined delay period as the time range, calculate the change in edema volume at different time points within this period, initially defining delayed edema expansion. The calculation formula is as follows: ; in, To delay the expansion of edema, This represents the edema volume at subsequent time points in the delayed period. This represents the edema volume at the initial time point of the delay period. Step S553: ​​Plot the operating characteristic curve (ROC) of delayed edema expansion (DPE), evaluate the predictive performance of ROC on poor prognosis of TBI patients, determine the absolute edema volume expansion threshold of delayed edema expansion using the Youden method, determine the quantitative definition of delayed edema expansion, and verify the clinical model performance of edema expansion on TBI prognosis. Step S56: Perform statistical analysis on the clinical model of the predictive efficacy of edema expansion on TBI prognosis.

[0029] Step S6: Extract clinical text features III and image data features III required for constructing the TBI cerebral edema expansion prediction model. Extract and filter image group data features III from image data features III. Fuse clinical text features III with processed image data features III. Construct the TBI cerebral edema expansion prediction model based on machine learning methods. Construct training sample sets for early edema expansion research cohorts and delayed edema expansion research cohorts. Select the model with the highest area under the curve (AUC) as the TBI cerebral edema expansion prediction model classifier. Train and validate the performance of the TBI cerebral edema expansion prediction model. Perform statistical analysis on the TBI cerebral edema expansion prediction model. The specific steps are as follows: Step S61: Extract clinical text feature III and image data feature III that may be related to the expansion of edema in the clinical text feature I and image data feature I. Use the 3DUnet segmentation network to segment the skull and edema or hematoma, calculate radiomics features and screen radiomics features. Then extract the histogram features and texture features of edema and hematoma as radiomics features III in image data feature III. Step S62: Screen the radiomics features III. The specific steps are as follows: Step S621: Calculate the Pearson coefficients of radiomics features and edema expansion label, and retain the 100 variables most relevant to the edema expansion label to form 100-dimensional radiomics features; Step S622: Select an algorithm to finally filter out the 5 radiomics features most relevant to the edema expansion label from the 100-dimensional radiomics features; Step S63: The extracted and processed clinical text features III and image data features III are fused together using a simple splicing method to form a 31-dimensional feature input vector, and a TBI brain edema expansion prediction model is constructed based on machine learning methods. Step S64: Match the 31-dimensional feature input vector with the corresponding edema expansion label to construct training sample sets for the early edema expansion research cohort and the delayed edema expansion research cohort, respectively; Step S65: Select four classic machine learning classifiers: Random Forest (RF), Logistic Regression (LR), TabNet, and Catboost. Use the area under the curve (AUC) as the classifier performance metric. Calculate the AUC values ​​of the four classifiers in the early edema expansion cohort and the delayed edema expansion cohort respectively. Select the classifier model with the highest AUC value as the TBI brain edema expansion prediction model classifier. Step S66: Train and validate the TBI cerebral edema expansion prediction model using a 4-fold cross-validation method; Step S67: Perform statistical analysis on the TBI cerebral edema expansion prediction model.

[0030] Step S7: Output the prognostic model prediction results, including patient risk stratification results, neurological function prognostic prediction results, and cerebral edema expansion prediction results. The results are evaluated using the Glasgow Outcome Scale and the modified Rankin Scale.

[0031] SPSS 24.0 was used to perform statistical analysis on the constructed prognostic prediction model for traumatic brain injury (TBI), the clinical model of the predictive efficacy of edema expansion on TBI prognosis, and the TBI cerebral edema expansion prediction model. The normality of the data was verified by the Kolmogorov-Smirnov test. Normally distributed data were analyzed by t-test or one-way ANOVA, and non-normally distributed data were analyzed by Mann-Whitney U test or Kruskal-Wallis H test. Categorical variables were analyzed by chi-square test. Significant predictive factors in univariate analysis were included in multivariate logistic regression analysis. The effectiveness of the models was evaluated by plotting ROC curves. The area under the ROC curve (AUC), sensitivity, specificity, and 95% confidence interval were calculated. The DeLong method was used to compare the ROC curves.

[0032] Example 2 like Figure 2 As shown, the present invention also provides a system for establishing a prognostic model for craniocerebral trauma based on multimodal data fusion and AI, including a data acquisition module, a data preprocessing module, a feature extraction module, a multimodal model construction module, and a result output module; The data acquisition module is used to collect multimodal data from TBI patients, including a clinical text acquisition unit, an image data acquisition unit, and a data filtering unit. The clinical text acquisition unit interfaces with the hospital's electronic medical record system to obtain medical records, laboratory reports, demographic data, and treatment records. The image data acquisition unit receives image data in standard DICOM format, such as head CT images. The data filtering unit filters data according to inclusion or exclusion criteria to ensure data validity.

[0033] The data preprocessing module connects to the data acquisition module, receiving raw multimodal data and outputting standardized clinical text and medical image data, providing high-quality input for feature extraction. The data preprocessing module includes a clinical text preprocessing unit and an image data preprocessing unit. The clinical text preprocessing unit performs anonymization, missing data completion, and standardization, supplementing missing age, gender, and past medical history data by retrieving medical records, and recording and validating clinical features using a pre-set case report form (CRF). The image data preprocessing unit performs image cropping, window width and level settings and adjustments, data augmentation, and region of interest segmentation.

[0034] The feature extraction module is connected to the data preprocessing module, receives standardized data, and outputs clinical features, traditional imaging features, and screened radiomics features to provide feature vectors for model construction. The feature extraction module includes a clinical text feature extraction unit, a traditional image feature extraction unit, and a radiomics feature extraction unit. The clinical text feature extraction unit extracts 22 clinical features, including patient demographic information, Glasgow Coma Scale (GCS) score, symptoms of headache, vomiting, and coma, hematoma rupture into the ventricles, past medical history, and treatment methods. Past medical history includes hemorrhagic stroke, ischemic stroke, hypertension, diabetes, hyperlipidemia, coronary heart disease, heart failure, arrhythmia, anticoagulation therapy, and antiplatelet therapy. Treatment methods include conservative drug therapy (CMT), conventional craniotomy (CC), and minimally invasive treatment (MIA). MIA includes stereotactic hematoma aspiration and neuroendoscopic hematoma aspiration. The traditional image feature extraction unit uses the DenseNet46 method to extract the location of the hemorrhage and skull fracture. The radiomics feature extraction unit extracts histogram and texture features from the image information and removes redundancy through a two-step screening process.

[0035] The multimodal model building module is connected to the feature extraction module, receives the extracted data features, builds a prediction model, and outputs a prediction model that has been trained and validated. The multimodal model building module includes a TBI prognostic prediction model building unit, a clinical model building unit for the predictive efficacy of edema expansion on TBI prognosis, and a TBI cerebral edema expansion prediction model building unit. The TBI prognostic prediction model building unit uses a 3D DenseNet network to extract image information features, combines a gated multimodal unit (GMU) with clinical text features to build a prognostic prediction model, and uses a Glasgow Outcome Scale (GOS) score of 1-3 as the criterion for judging adverse outcomes. The clinical model building unit for the predictive efficacy of edema expansion on TBI prognosis plots hematoma time-volume curves based on different initial hematoma volume ranges, defines early and delayed edema expansion indicators, plots the operating characteristic curve (ROC) of delayed edema expansion (DPE), evaluates the predictive performance of ROC on poor prognosis of TBI patients, determines the absolute edema volume expansion threshold for delayed edema expansion using the Youden method, and defines the quantitative definition of delayed edema expansion. The TBI cerebral edema expansion prediction model building unit uses the Catboost algorithm to integrate clinical text features, traditional imaging features, and screened radiomics features to build a binary classification prediction model.

[0036] In the multimodal model construction module, SPSS 24.0 was used to perform statistical analysis on the constructed prognostic prediction model for traumatic brain injury (TBI), the clinical model of the predictive efficacy of edema expansion on TBI prognosis, and the TBI cerebral edema expansion prediction model to validate the models. Specifically, the normality of the data was verified by the Kolmogorov-Smirnov test. Normally distributed data were tested using t-tests or one-way ANOVA, and non-normally distributed data were tested using Mann-Whitney U tests or Kruskal-Wallis H tests. Categorical variables were tested using chi-square tests. Significant predictive factors in univariate analysis were included in multivariate logistic regression analysis. The effectiveness of the models was evaluated by plotting ROC curves and calculating the area under the ROC curve (AUC), sensitivity, specificity, and 95% confidence interval. The DeLong method was used to compare the ROC curves.

[0037] The results output module is connected to the multimodal model construction module. It receives the prediction model, inputs the preprocessed multimodal data into the model, and outputs the prognostic model prediction results. The results of risk stratification, prognostic prediction, and edema expansion prediction are displayed through a visual interface, such as the integrated interface of a hospital HIS system, to assist clinical decision-making.

[0038] The invention will be further illustrated below through specific implementation examples.

[0039] The data used in this embodiment comes from the electronic medical record database of a hospital's neurosurgery department from 2016 to 2024, and includes medical records, test reports, imaging data, etc.

[0040] Step S1: Collect clinical text and imaging data of patients with traumatic brain injury. The clinical text comes from the patient's electronic medical record, including medical documents, test reports, demographic data and treatment records. The imaging data includes the patient's initial head imaging information within 24 hours of admission and subsequent follow-up imaging information. The inclusion criteria are patients aged 10 to 80 years who meet the diagnostic criteria for traumatic brain injury. The exclusion criteria are cases with incomplete medical history or test results, or missing CT results. Step S2: Preprocess the clinical text and image data. Clinical text preprocessing includes anonymization, missing data supplementation, and standardization. Missing data such as age, gender, and past medical history are supplemented by retrieving medical records. Image data preprocessing includes image cropping, window width and level settings and adjustments, data augmentation, and region of interest segmentation. Step S3: Utilize a deep learning model to automatically extract clinical text features I and imaging data features I. Clinical text features I include patient demographic information, Glasgow Coma Scale (GCS) score, symptoms of headache, vomiting, and coma, hematoma rupture into the ventricles, past medical history, and treatment methods. Past medical history includes hemorrhagic stroke, ischemic stroke, hypertension, diabetes, hyperlipidemia, coronary heart disease, heart failure, arrhythmia, anticoagulation therapy, and antiplatelet therapy. Treatment methods include conservative drug therapy (CMT), conventional craniotomy (CC), and minimally invasive treatment (MIA). MIA includes stereotactic hematoma aspiration and neuroendoscopic hematoma aspiration. Imaging data features I include traditional imaging features I and radiomics features I. Traditional imaging features I include the location of the hemorrhage and the extent of skull fracture. Radiomics features I include histogram features and texture features.

[0041] Step S4: Construct a deep learning-based prognostic model for TBI using convolutional neural networks to predict the neurological functional outcomes of TBI patients at discharge. The Glasgow Outcome Scale (GOS) is used to evaluate patient prognosis, with a GOS score of 1-3 defined as an adverse outcome. The specific steps are as follows: Step S41: Automatically crop the image below the foramen magnum, and reduce the interference of invalid information on model performance for images with an average slice thickness of 5.0 mm. Set the window width and window level to 90 HU and 45 HU respectively, followed by data augmentation, random cropping, and rotation. Considering the large amount of original image data and memory and complexity, the image is finally scaled to 192. 192 For each patient, a baseline hematoma volume was calculated using a 80-size scale. The ICH and ICH-GS scales were scored on each patient. This process used Insight Toolkit SNAP (ITK-SNAP) software. On the initial CT plain scan data, a trained researcher identified and delineated all imaging slices containing the lesion, followed by a review of the accuracy of the hematoma delineation by another researcher. All delineated slices for each patient were integrated into a complete segmentation file, and the ITK-SNAP software automatically calculated and output the hematoma volume as image data feature II. Clinical text was independently recorded by two researchers, and clinical text feature II and demographic data for each patient were collected using a pre-designed CRF table.

[0042] Step S42: Construct a convolutional neural network. Considering the need for robustness and effectiveness of neural network structure in predicting prognosis, the DenseNet46 method is used to extract image data feature II. A 3DDenseNet structure with a network depth of 121 is used. Based on the DenseNet structure which is densely connected by a convolutional neural network (CNN), DenseNet is easier to train with the backpropagation algorithm. A gated multimodal unit (GMU) is constructed to fuse image data feature II extracted from CT images and clinical text feature II extracted from electronic medical record information. The GMU can effectively merge visual and text information and set adaptive weights in their subspaces to construct a prognostic prediction model for traumatic brain injury (TBI).

[0043] Step S43: Train the fused clinical text feature II and image data feature II on four NVIDIA Tesla V100 GPUs (32GB each) using PyTorch 1.5.1, with a batch size of 32 and an initial learning rate of 1e. -4 During training, focal loss was used to alleviate the imbalance of predicted classes, and 5-fold cross-validation was used to verify the predictive effect of the prognostic prediction model for traumatic brain injury (TBI).

[0044] Step S44: Statistical analysis was performed on the prognostic prediction model for traumatic brain injury (TBI) using SPSS 24.0. The Kolmogorov-Smirnov test was used to determine whether the data were normally distributed. Data that were normally distributed were expressed as mean (± standard deviation). Intergroup comparisons were performed using... Tests or one-way ANOVA were performed; non-normally distributed data were represented as medians, and comparisons between groups were performed using the Mann-Whitney U test or the Kruskal-Wallis H test. Categorical variables were represented as number of cases, incidence rate, or proportion, and comparisons between groups were performed using the chi-square test. Significant predictive factors in univariate analysis were included in multivariate logistic regression analysis, and a two-sided P < 0.05 was considered statistically significant. The efficacy of prognostic prediction models (two CNN models and two traditional rating scales) was evaluated by plotting receiver operating characteristic (ROC) curves, and the area under the curve (AUC), sensitivity, specificity, and corresponding 95% confidence intervals (CI) were calculated. ROC curves were plotted using GraphPad Prism software (version 9.0), and comparisons between groups were performed using the DeLong method.

[0045] Step S5: Construct a clinical model of the predictive efficacy of edema expansion on TBI prognosis. The specific steps are as follows: Step S51: Define neurological prognosis. Use the modified Rankin score (mRS) to evaluate the neurological outcome of TBI patients. An mRS score of 4-6 at 90 days after TBI (early edema expansion cohort) or a discharge mRS score of 3-6 (delayed edema expansion cohort) is defined as poor neurological prognosis. Step S52: Before formal analysis, anonymize the patient's imaging and clinical baseline information, use a deep learning model to automatically segment the hematoma and peri-hematoma edema in standard DICOM format CT images, and then use the lnsight ToolkitSNAP software to manually calibrate the regions of interest in the CT images automatically segmented by the deep learning model. Step S53: Divide the initial hematoma volume range and the follow-up imaging time window, and plot the hematoma time-volume curve. The initial hematoma volume (B) range is set into the following three categories: ; Considering that the timing of CT follow-up examinations varies among patients, we set a series of follow-up imaging time windows as follows: day 1, day 2 to 4, day 5 to 8, day 13 to 15, and day 19 to 21. Step S54: Define hematoma expansion and edema expansion, and construct a clinical model of the predictive efficacy of edema expansion on TBI prognosis. The specific steps are as follows: Step S541: Define the early absolute hematoma expansion value as the difference between the hematoma volume examined within 48 to 72 hours after onset and the initial hematoma volume within 24 hours after onset. The calculation formula is as follows: ; in, This represents the early absolute hematoma enlargement value. The hematoma volume should be re-examined within 48 to 72 hours after the onset of illness. The initial hematoma volume within 24 hours of onset; Step S542: Define the early absolute edema expansion value as the difference between the edema volume examined within 48 to 72 hours after onset and the initial edema volume within 24 hours after onset. The calculation formula is as follows: ; in, This represents the value of early absolute edema expansion. The edema volume should be re-examined within 48 to 72 hours after the onset of illness. The initial edema volume within 24 hours of onset; Step S543: Calculate the early relative edema expansion value X, using the following formula: ; in, This represents the value of early relative edema expansion; Step S544: Calculate the early absolute hematoma combined with edema expansion value Z. The calculation formula is as follows: ; in, This represents the value of early-stage absolute hematoma combined with edema expansion. The total lesion volume should be re-examined within 48 to 72 hours after the onset of illness. The initial total lesion volume within 24 hours of onset; Step S55: Based on the definitions of early absolute hematoma expansion, absolute edema expansion, relative edema expansion, and combined hematoma and edema expansion, plot the receiver operating characteristic (ROC) curve and calculate its sensitivity and specificity in predicting poor patient prognosis. Step S56: Define delayed edema expansion (DPE) and validate the clinical model of the predictive efficacy of edema expansion on TBI prognosis. The specific steps are as follows: Step S561: Based on the plotted hematoma time-volume curve, determine the peak and stable periods of edema growth, and define the specific time window for delaying edema expansion. Step S562: Using the defined delay period as the time range, calculate the change in edema volume at different time points within this period, initially defining delayed edema expansion. The calculation formula is as follows: ; in, To delay the expansion of edema, This represents the edema volume at subsequent time points in the delayed period. This represents the edema volume at the initial time point of the delay period. Step S563: Plot the operating characteristic curve (ROC) of delayed edema expansion (DPE), evaluate the predictive performance of ROC on poor prognosis of TBI patients, determine the absolute edema volume expansion threshold using the Youden method, determine the quantitative definition of delayed edema expansion, and verify the clinical model performance of edema expansion on TBI prognosis prediction efficacy. Step S57: Statistical analysis was performed on the clinical model of the predictive efficacy of edema expansion on TBI prognosis using SPSS 24.0. The Kolmogorov-Smirnov test was used to determine whether the data were normally distributed. Data that were normally distributed were expressed as mean (± standard deviation). Intergroup comparisons were performed using... Tests or one-way ANOVA were performed; non-normally distributed data were represented as medians; comparisons between groups were performed using the Mann-Whitney U test or the Kruskal-Wallis H test; categorical variables were represented as number of cases, incidence rate, or proportion; comparisons between groups were performed using the chi-square test; significant predictors in univariate analysis were included in multivariate logistic regression analysis; a two-sided P < 0.05 was considered statistically significant.

[0046] Step S6: Construct a TBI cerebral edema expansion prediction model based on machine learning methods. The specific steps are as follows: Step S61: Before formal analysis, anonymize the patient's imaging and clinical baseline information. Using a deep learning model, automatically segment hematoma and edema in standard DICOM format CT images. Subsequently, the neurosurgeon uses Imsight Toolkit SNAP software to manually calibrate the regions of interest in the automatically segmented CT images. Clinical and demographic data are collected using a pre-defined medical record report form (CRF). The neurosurgeon reviews the completeness and accuracy of the data, and retrieves and supplements any missing data (such as age, gender, medical history, etc.) from the medical records.

[0047] Step S62: Construct machine learning models and train them for both the early perihematoma edema PHE expansion study cohort and the delayed perihematoma edema PHE expansion study cohort. The specific steps are as follows: Step S621: Extract clinical text features III and image data features III that may be related to the expansion of the patient's edema. Use the 3DUnet segmentation network to segment the skull and edema or hematoma, calculate the radiomic features, and screen the radiomic features in the image data features III. Then use the open-source radiomics toolkit pyradiomics to extract the histogram features and texture features of edema and hematoma as radiomic features III in the image data features III. Step S622: Perform feature data filtering, the specific steps are as follows: Step S6221: Calculate the Pearson coefficients of radiomics feature III and the edema expansion label, and retain the 100 variables most relevant to the edema expansion label; Step S6222: Using the Logistic Regression model selection algorithm in the open-source machine learning library Scikit-Leam, the 5 radiomic features most relevant to the edema expansion label are finally selected from 100 radiomic features. Step S63: Combine the clinical text feature III with the processed image data feature III to form an input vector, and combine it with the corresponding edema expansion label to form training samples. The fusion method uses a simple concatenation method to form 31-dimensional features. Train binary classifiers for early edema expansion and delayed edema expansion respectively. Use the popular CatBoost model, set the number of iterations to 2000, the depth of the decision tree to 10, the initial learning rate of the early edema expansion model to 0.5, and the learning rate of the delayed edema expansion model to 0.25. Use the default settings for other parameters. The objective function of the CatBoost model is logloss. Step S64: Select three classic machine learning classifiers, Random Forest (RF), Logistic Regression (LR), and TabNet, and compare their performance with the Catboost model. Use the area under the curve (AUC) as the performance metric for the model. Calculate the AUC values ​​of the four models in the early edema expansion cohort and the delayed edema expansion cohort respectively. Select the model with the highest AUC value as the classifier for the TBI brain edema expansion prediction model. Step S65: Train and validate the TBI cerebral edema expansion prediction model using a 4-fold cross-validation method; Step S66: Statistical analysis was performed on the TBI cerebral edema expansion prediction model using SPSS 24.0. The Kolmogorov-Smirnov test was used to determine whether the data were normally distributed. Data that conformed to a normal distribution were expressed as mean (± standard deviation). Intergroup comparisons were performed using... Tests or one-way ANOVA were performed; non-normally distributed data were represented as medians, and comparisons between groups were performed using the Mann-Whitney U test or the Kruskal-Wallis H test. Categorical variables were represented as number of cases, incidence rate, or proportion, and comparisons between groups were performed using the chi-square test. Significant predictive factors in univariate analysis were included in multivariate logistic regression analysis, and a two-sided P < 0.05 was considered statistically significant. The efficacy of prognostic prediction models (two CNN models and two traditional rating scales) was evaluated by plotting receiver operating characteristic (ROC) curves, and the area under the curve (AUC), sensitivity, specificity, and corresponding 95% confidence intervals (CI) were calculated. ROC curves were plotted using GraphPad Prism software (version 9.0), and comparisons between groups were performed using the DeLong method.

[0048] Step S7: Output the prognostic model prediction results, including patient risk stratification results, neurological function prognostic prediction results, and cerebral edema expansion prediction results. The results are evaluated using the Glasgow Outcome Scale and the modified Rankin Scale.

[0049] It is worth noting that all contents not described in detail in this invention are existing technologies and are well known to those skilled in the art.

[0050] Therefore, this invention provides a method and system for establishing a prognostic model for traumatic brain injury (TBI) based on multimodal data fusion and AI. Relying on artificial intelligence technology and deep learning models, it utilizes deep neural networks to automatically extract and classify TBI prognostic-related image data features from medical imaging data, constructing a TBI prognostic prediction model. Machine learning algorithms are used to build a clinical model of the predictive efficacy of edema expansion on TBI prognosis and a TBI cerebral edema expansion prediction model. The model is optimized through cross-validation and parameter tuning to improve generalization ability and reduce overfitting. Secondly, medical image processing such as denoising, enhancement, and segmentation is performed on CT image data of TBI patients. AI algorithms are used to extract a large number of image features, and the features most relevant to TBI prognosis are selected and retained to provide reliable data support for the model. Finally, actual clinical data is collected to verify the effectiveness and reliability of the model. The research results are then used in conjunction with clinicians for clinical decision-making, developing precise and personalized treatment plans for patients, helping to clarify the pathogenesis and prognostic patterns of TBI, and providing a clinical auxiliary decision-making tool.

[0051] 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for establishing a prognostic model for traumatic brain injury based on multimodal data fusion and AI, characterized in that, Includes the following steps: Step S1: Collect clinical text and imaging data of patients with traumatic brain injury; Step S2: Preprocess the clinical text and image data; Step S3: Automatically extract clinical text feature I and image data feature I using a deep learning model; Step S4: Construct a multimodal attention module, manually extract image data feature II and clinical text feature II required for constructing the TBI prognostic prediction model, based on DenseNet, combine gated multimodal units to fuse image data feature II and clinical text feature II, construct the TBI prognostic prediction model, use focal loss to alleviate class imbalance, verify the predictive effect of the TBI prognostic prediction model, and perform statistical analysis on the TBI prognostic prediction model; Step S5: Define the prognostic criteria for neurological function, divide the initial hematoma volume range, set the follow-up time window to draw the hematoma time-volume curve, define the indicators of hematoma expansion and edema expansion, construct a clinical model of the predictive efficacy of edema expansion on TBI prognosis, define delayed edema expansion (DPE), validate the clinical model of the predictive efficacy of edema expansion on TBI prognosis, and perform statistical analysis on the clinical model of the predictive efficacy of edema expansion on TBI prognosis. Step S6: Extract the clinical text feature III and image data feature III required to construct the TBI brain edema expansion prediction model, extract and screen the radiomics feature III from the image data feature III, fuse the clinical text feature III and the processed image data feature III, construct the TBI brain edema expansion prediction model based on machine learning methods, construct training sample sets for the early edema expansion research cohort and the delayed edema expansion research cohort, select the model with the highest area under the curve (AUC) value as the TBI brain edema expansion prediction model classifier, train and verify the performance of the TBI brain edema expansion prediction model, and perform statistical analysis on the TBI brain edema expansion prediction model; Step S7: Output the prognostic model prediction results, including patient risk stratification results, neurological function prognostic prediction results, and cerebral edema expansion prediction results. The results are evaluated using the Glasgow Outcome Scale and the modified Rankin Scale.

2. The method for establishing a prognostic model for traumatic brain injury based on multimodal data fusion and AI according to claim 1, characterized in that: In step S1, the clinical text is derived from the patient's electronic medical record, including medical documents, laboratory reports, demographic data and treatment records. The imaging data includes the patient's initial head imaging information within 24 hours of admission and subsequent follow-up imaging information. The data inclusion criteria are patients aged 10 to 80 years who meet the diagnostic criteria for traumatic brain injury, and cases with incomplete medical history or laboratory results or missing head examination imaging information are excluded.

3. The method for establishing a prognostic model for traumatic brain injury based on multimodal data fusion and AI according to claim 1, characterized in that: In step S2, clinical text preprocessing includes anonymization, missing data supplementation, and standardization, supplementing missing age, gender, and past medical history data by retrieving medical records; image data preprocessing includes image cropping, window width and level settings and adjustments, data augmentation, and region of interest segmentation.

4. The method for establishing a prognostic model for traumatic brain injury based on multimodal data fusion and AI according to claim 1, characterized in that: In step S3, clinical text feature I includes patient demographic information, Glasgow Coma Scale (GCS) score, symptoms of headache, vomiting, and coma, hematoma rupture into the ventricles, past medical history, and treatment methods. Past medical history includes hemorrhagic stroke, ischemic stroke, hypertension, diabetes, hyperlipidemia, coronary heart disease, heart failure, arrhythmia, anticoagulation therapy, and antiplatelet therapy. Treatment methods include conservative drug therapy (CMT), conventional craniotomy (CC), and minimally invasive treatment (MIA). MIA includes stereotactic hematoma aspiration and neuroendoscopic hematoma aspiration. Imaging data feature I includes traditional imaging feature I and radiomics feature I. Traditional imaging feature I includes the location of the hemorrhage and skull fracture status, while radiomics feature I includes histogram features and texture features.

5. The method for establishing a prognostic model for traumatic brain injury based on multimodal data fusion and AI according to claim 4, characterized in that: Clinical text feature I was extracted using a pre-defined case report form (CRF). Image data feature I was extracted using a deep learning model to automatically segment hematoma and perihemorrhagic edema in standard DICOM format image data. Subsequently, medical image processing software was used to manually calibrate the regions of interest in the image data automatically segmented by the deep learning model. Traditional image data feature I was extracted using the DenseNet46 method, while radiomics feature I was extracted from the segmented image data.

6. The method for establishing a prognostic model for traumatic brain injury based on multimodal data fusion and AI according to claim 1, characterized in that: Step S4 includes the following steps: Step S41: Construct a multimodal attention module including Grid, SE, Non-local, and Proposed to identify image regions in the image data that contribute to the prediction of poor prognosis of TBI; Step S42: Using medical image processing software, manually identify and delineate hematomas on imaging slices containing lesions in image regions that contribute to poor prognosis prediction of TBI. Review the accuracy of hematoma delineation, integrate all delineated slices for each patient into a complete segmentation file, and automatically calculate and output the hematoma volume using medical image processing software as image data feature II. Researchers independently record clinical text, and use a pre-designed CRF table to collect clinical characteristics and demographic data for each patient as clinical text feature II. Step S43: The extracted image data feature II and clinical text feature II are combined with DenseNet, which is densely connected by a convolutional neural network (CNN) structure. By constructing a gated multimodal unit (GMU), the clinical text feature II and image data feature II are fused, and adaptive weights are set in the subspace to construct a prognostic prediction model for traumatic brain injury (TBI). Step S44: Train the fused clinical text feature II and image data feature II, set the batch size and initial learning rate, use focal loss to alleviate the imbalance of prediction categories, and use 5-fold cross-validation to verify the prediction effect of the traumatic brain injury (TBI) prognosis prediction model. Step S45: Perform statistical analysis on the prognostic prediction model for traumatic brain injury (TBI).

7. The method for establishing a prognostic model for traumatic brain injury based on multimodal data fusion and AI according to claim 1, characterized in that: Step S5 includes the following steps: Step S51: Define the prognostic criteria for neurological function and use the modified Rankin Scale (mRS) to evaluate the neurological functional outcomes of TBI patients; Step S52: Divide the initial hematoma volume range, set the follow-up time window, and plot the hematoma time-volume curve; Step S53: Define hematoma expansion and edema expansion, and construct a clinical model of the predictive efficacy of edema expansion on TBI prognosis. The specific steps are as follows: Step S531: Define the early absolute hematoma expansion value as the difference between the hematoma volume examined within 48 to 72 hours after onset and the initial hematoma volume within 24 hours after onset. The calculation formula is as follows: ； in, This represents the early absolute hematoma enlargement value. The hematoma volume should be re-examined within 48 to 72 hours after the onset of illness. The initial hematoma volume within 24 hours of onset; Step S532: Define the early absolute edema expansion value as the difference between the edema volume examined within 48 to 72 hours after onset and the initial edema volume within 24 hours after onset. The calculation formula is as follows: ； in, This represents the value of early absolute edema expansion. The edema volume should be re-examined within 48 to 72 hours after the onset of illness. The initial edema volume within 24 hours of onset; Step S533: Calculate the early relative edema expansion value X, using the following formula: ； in, This represents the value of early relative edema expansion; Step S534: Calculate the early absolute hematoma combined with edema expansion value Z. The calculation formula is as follows: ； in, This represents the value of early-stage absolute hematoma combined with edema expansion. The total lesion volume should be re-examined within 48 to 72 hours after the onset of illness. The initial total lesion volume within 24 hours of onset; Step S54: Based on the definitions of early absolute hematoma expansion, absolute edema expansion, relative edema expansion, and combined hematoma and edema expansion, plot the receiver operating characteristic (ROC) curve and calculate the sensitivity and specificity for predicting poor patient prognosis. Step S55: Define delayed edema expansion (DPE) and validate the clinical model of the predictive efficacy of edema expansion on TBI prognosis. The specific steps are as follows: Step S551: Based on the plotted hematoma time-volume curve, determine the peak and stable periods of edema growth, and define the specific time window for delaying edema expansion. Step S552: Using the defined delay period as the time range, calculate the change in edema volume at different time points within this period, initially defining delayed edema expansion. The calculation formula is as follows: ； in, To delay the expansion of edema, This represents the edema volume at subsequent time points in the delayed period. This represents the edema volume at the initial time point of the delay period. Step S553: ​​Plot the operating characteristic curve (ROC) of delayed edema expansion (DPE), evaluate the predictive performance of ROC on poor prognosis of TBI patients, determine the absolute edema volume expansion threshold of delayed edema expansion using the Youden method, determine the quantitative definition of delayed edema expansion, and verify the clinical model performance of edema expansion on TBI prognosis. Step S56: Perform statistical analysis on the clinical model of the predictive efficacy of edema expansion on TBI prognosis.

8. The method for establishing a prognostic model for traumatic brain injury based on multimodal data fusion and AI according to claim 1, characterized in that: Step S6 includes the following steps: Step S61: Extract clinical text feature III and image data feature III that may be related to the expansion of edema in the clinical text feature I and image data feature I. Use the 3DUnet segmentation network to segment the skull and edema or hematoma, calculate radiomics features and screen radiomics features. Then extract the histogram features and texture features of edema and hematoma as radiomics features III in image data feature III. Step S62: Screen the radiomics features III. The specific steps are as follows: Step S621: Calculate the Pearson coefficients of radiomics features and edema expansion label, and retain the 100 variables most relevant to the edema expansion label to form 100-dimensional radiomics features; Step S622: Select an algorithm to finally filter out the 5 radiomics features most relevant to the edema expansion label from the 100-dimensional radiomics features; Step S63: The extracted and processed clinical text features III and image data features III are fused together using a simple splicing method to form a 31-dimensional feature input vector, and a TBI brain edema expansion prediction model is constructed based on machine learning methods. Step S64: Match the 31-dimensional feature input vector with the corresponding edema expansion label to construct training sample sets for the early edema expansion research cohort and the delayed edema expansion research cohort, respectively; Step S65: Select four classic machine learning classifiers: Random Forest (RF), Logistic Regression (LR), TabNet, and Catboost. Use the area under the curve (AUC) as the classifier performance metric. Calculate the AUC values ​​of the four classifiers in the early edema expansion cohort and the delayed edema expansion cohort respectively. Select the classifier model with the highest AUC value as the TBI brain edema expansion prediction model classifier. Step S66: Train and validate the TBI cerebral edema expansion prediction model using a 4-fold cross-validation method; Step S67: Perform statistical analysis on the TBI cerebral edema expansion prediction model.

9. The method for establishing a prognostic model for traumatic brain injury based on multimodal data fusion and AI according to claim 1, characterized in that: SPSS 24.0 was used to perform statistical analysis on the constructed prognostic prediction model for traumatic brain injury (TBI), the clinical model of the predictive efficacy of edema expansion on TBI prognosis, and the TBI cerebral edema expansion prediction model. The normality of the data was verified by the Kolmogorov-Smirnov test. Normally distributed data were analyzed by t-test or one-way ANOVA, and non-normally distributed data were analyzed by Mann-Whitney U test or Kruskal-Wallis H test. Categorical variables were analyzed by chi-square test. Significant predictive factors in univariate analysis were included in multivariate logistic regression analysis. The effectiveness of the models was evaluated by plotting ROC curves. The area under the ROC curve (AUC), sensitivity, specificity, and 95% confidence interval were calculated. The DeLong method was used to compare the ROC curves.

10. A system for establishing a prognostic model for traumatic brain injury based on multimodal data fusion and AI, characterized in that: It includes a data acquisition module, a data preprocessing module, a feature extraction module, a multimodal model construction module, and a result output module; The data acquisition module includes a clinical text acquisition unit, an image data acquisition unit, and a data filtering unit; The data preprocessing module is connected to the data acquisition module and includes a clinical text preprocessing unit and an image data preprocessing unit. The feature extraction module is connected to the data preprocessing module and includes a clinical text feature extraction unit, a traditional image feature extraction unit, and a radiomics feature extraction unit. The multimodal model building module is connected to the feature extraction module, including a TBI prognostic prediction model building unit, a clinical model building unit for the predictive efficacy of edema expansion on TBI prognosis, and a TBI brain edema expansion prediction model building unit. The results output module is connected to the multimodal model construction module.

Images