Artificial intelligence-based brain trauma patient condition change prediction method and system

Abstract

The application discloses a brain trauma patient condition change prediction method and system based on artificial intelligence, relates to the fields of artificial intelligence and medical treatment, and comprises the following steps: S1, constructing an initial change prediction model; S2, acquiring a training data set; S3, training the change prediction model; S4, acquiring to-be-predicted analysis data; S5, analyzing the to-be-predicted analysis data to output cross-modal prediction results and single-modal prediction results; S6, fusing the cross-modal prediction results and the single-modal prediction results to obtain change prediction results; a CT sequence feature extraction module extracts CT sequence single-modal features of CT sequence data through preprocessing, dense feature extraction, time sequence-space alignment and time sequence modeling, completes accurate feature extraction and frame alignment of the CT sequence, ensures the accuracy and continuity of CT image features, and provides high-quality imaging support for condition prediction; the method can provide reliable auxiliary support for clinical doctors to judge the condition of patients and develop optimized treatment schemes, and realizes rapid and accurate prediction of the condition change of brain trauma.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and medical technology, and in particular to a method and system for predicting changes in the condition of patients with traumatic brain injury based on artificial intelligence. Background Technology

[0002] Traumatic brain injury (TBI) is one of the most common acute and critical illnesses in clinical neurosurgery. Its causes mainly include traffic accidents, falls from heights, violent blows, and accidental collisions. Globally, it has extremely high morbidity, mortality, and disability rates, placing a heavy burden on patients, their families, and society. According to relevant clinical data, the global incidence of TBI exceeds 200 per 100,000 people annually. The mortality rate for severe TBI patients is as high as 30%-50%, and approximately 60%-70% of survivors suffer from varying degrees of neurological dysfunction, including motor impairment, cognitive impairment, and language impairment, severely affecting their quality of life and significantly increasing the burden on families and the consumption of social medical resources. In clinical diagnosis and treatment, the progression of traumatic brain injury (TBI) patients' conditions is highly complex and uncertain, especially lesions such as intracranial hematomas and cerebral edema caused by trauma. The dynamic changes in the size, shape, and location of these lesions directly determine the patient's treatment plan and prognosis. Therefore, achieving rapid and accurate diagnosis and prediction of changes in the condition of TBI patients has become a key link in improving patient prognosis and reducing mortality and disability rates. Currently, computed tomography (CT) imaging is the core tool for clinical diagnosis and assessment of TBI. It has advantages such as fast examination speed, high resolution, relatively low cost, and strong repeatability. It can clearly present the specific details of lesions such as intracranial hematomas, cerebral contusions, cerebral edema, and skull fractures, providing important imaging evidence for clinicians to determine the severity of the condition and formulate treatment plans. Meanwhile, clinical diagnosis and treatment also require the integration of structured clinical data from patients, including age, gender, type of injury, Glasgow Coma Scale (GCS), blood pressure, blood glucose, complete blood count, and coagulation function, to comprehensively assess the trend of changes in the patient's condition and more fully and accurately evaluate the patient's prognosis. However, current technologies still have many shortcomings in the field of predicting the condition of traumatic brain injury, making it difficult to meet the needs of precise clinical diagnosis and rapid intervention. Therefore, developing an efficient, accurate, and objective technology for predicting changes in the condition of traumatic brain injury has significant clinical practical significance and application value.

[0003] Currently, there are several technical shortcomings in the field of predicting the condition of traumatic brain injury. The most similar traditional technical solutions have many deficiencies, as follows: First, traditional diagnostic and treatment methods rely heavily on doctors' manual interpretation of images. Doctors analyze CT sequence images based on their own clinical experience and judge the dynamic trend of lesions by comparing CT images at different time points. However, this method is affected by factors such as the doctor's experience level, subjective judgment differences, and fatigue from reading images, making it difficult to accurately capture subtle changes in lesions. Moreover, the efficiency of image reading is low, which cannot meet the "race against time" diagnostic and treatment needs of critically injured traumatic brain injury patients. Especially in primary hospitals, due to the relatively insufficient experience of doctors, the misdiagnosis and missed diagnosis rate can reach 30%-40%, which seriously affects the patient's treatment timing and prognosis. Secondly, conventional machine learning methods have also been attempted to be applied to the diagnosis of traumatic brain injury. These methods extract manual and clinical features from CT images to construct classification and prediction models, which improves the objectivity and efficiency of diagnosis to some extent. However, they have significant technical bottlenecks: manual feature extraction relies on manual design, requires extensive expertise in radiology, is cumbersome, and struggles to capture deep semantic features and temporal changes in CT sequence images, resulting in limited feature representation capabilities. Furthermore, traditional machine learning methods are highly dependent on data; when data volume is insufficient, data contains noise, or feature dimensionality is high, the model's generalization performance and diagnostic accuracy significantly decrease, making it difficult to adapt to the complex and diverse clinical cases of traumatic brain injury. Thirdly, existing deep learning models are mostly designed for 2D / 3D single images and cannot adapt to the special structure of 5D sequence data (batch, sequence length, channels, height, width) of traumatic brain injury CT images. Moreover, feature extraction and inter-frame alignment are separate processes, failing to form a co-optimized integrated framework. This leads to the loss of fine-grained details during the alignment process, insufficient capture of temporal features, and an inability to accurately capture the dynamic evolution of lesions. Finally, most existing bimodal fusion models adopt a fixed weight fusion method, which cannot adaptively allocate weights according to the disease characteristics, data quality, and importance of different modal features in disease diagnosis of different samples. The cross-modal semantic gap is obvious, and the model has poor robustness, weak adaptability to noise, missing data, and small sample scenarios, slow inference speed, and difficulty in adapting to the needs of primary medical equipment and rapid and accurate diagnosis in emergency departments. Summary of the Invention

[0004] The purpose of this invention is to design an artificial intelligence-based method and system for predicting changes in the condition of patients with traumatic brain injury in order to solve the above problems.

[0005] The present invention achieves the above objectives through the following technical solutions:

[0006] Artificial intelligence-based methods for predicting changes in the condition of patients with traumatic brain injury include:

[0007] S1. Construct an initial change prediction model, which includes a CT sequence feature extraction module, a clinical feature encoding module, a cross-modal feature distillation module, a cross-modal dynamic fusion module, a cross-modal prediction head, and two single-modal prediction heads. The CT sequence feature extraction module preprocesses and extracts features from the input CT sequence data, outputting CT sequence single-modal features. The clinical feature encoding module standardizes and encodes structured clinical data, outputting global clinical features with dimensions consistent with the CT sequence single-modal features. The cross-modal feature distillation module uses a teacher-student network structure to refine the global clinical features and CT sequence single-modal features. The cross-modal dynamic fusion module adaptively and dynamically fuses the refined CT sequence single-modal features and global clinical features, outputting bimodal features. The cross-modal prediction head outputs the cross-modal prediction result based on the bimodal features. The two single-modal prediction heads... The modal prediction head is used to predict and output CT prediction results and clinical prediction results based on single-modal features of CT sequences and global clinical features, respectively. The CT sequence feature extraction module includes a preprocessing module, a dense feature extraction module, a temporal-spatial alignment module, and a temporal modeling module from input to output. The preprocessing module is used to preprocess the CT sequence data. The dense feature extraction module is used to extract fine-grained features from the processed CT sequence data, including spatial and temporal features. The temporal-spatial alignment module includes a bidirectional difference elimination sub-block (DF-Block) and a stepwise alignment sub-block (BiA-Block). The difference elimination sub-block (DF-Block) unifies the grayscale distribution through fine-grained feature calibration and adaptive normalization of the lesion region. The BiA-Block achieves inter-frame alignment by performing multi-step cumulative deformation field prediction on the fine-grained features after grayscale distribution unification. The deformation field of the nth frame in the CT sequence data is... Represented as: , This represents the deformation field of the (n-1)th frame in the CT sequence data. This represents the deformation field increment of the nth frame relative to the (n-1)th frame; the temporal modeling module fuses the fine-grained features after inter-frame alignment and outputs the single-modal features of the CT sequence;

[0008] S2. Obtain the training dataset, which includes multiple training samples, each of which includes CT sequence data and clinical structured data.

[0009] S3. Use the training dataset to train and optimize the initial change prediction model to obtain the optimized change prediction model.

[0010] S4. Obtain the data to be predicted and analyzed, which includes CT sequence data and structured clinical data;

[0011] S5. Analyze the data to be predicted using the optimized change prediction model, and output cross-modal prediction results, CT prediction results and clinical prediction results.

[0012] S6. Perform fusion analysis on cross-modal prediction results, CT prediction results and clinical prediction results to obtain the change prediction results of the data to be analyzed.

[0013] An AI-based system for predicting changes in the condition of patients with traumatic brain injury includes:

[0014] Storage; storage is used to store computer programs;

[0015] An actuator; the actuator is used to execute a computer program stored in a memory, which, when executed, implements the artificial intelligence-based method for predicting changes in the condition of patients with traumatic brain injury as described above.

[0016] The beneficial effects of this invention are as follows: The CT sequence feature extraction module extracts CT sequence single-modal features from CT sequence data through preprocessing, dense feature extraction, temporal-spatial alignment, and temporal modeling, completing accurate feature extraction and frame alignment of the CT sequence, ensuring the accuracy and continuity of CT image features, and providing high-quality imaging support for disease prediction; the clinical feature encoding module extracts and outputs global clinical features consistent with the dimensions of CT sequence single-modal features from structured clinical data, strengthening the role of key clinical indicators and laying the foundation for subsequent fusion; the cross-modal dynamic fusion module automatically allocates fusion weights for the two modalities through a dynamic weight allocation network, achieving sample-level adaptive fusion and eliminating cross-modal semantic gaps. The system leverages the complementary advantages of two types of data to generate fused bimodal features. The cross-modal feature distillation module, through a "teacher-student" model architecture and a bidirectional semantic alignment mechanism, achieves knowledge transfer and feature refinement, suppresses the influence of noise and redundant features, improves the robustness and lightweight nature of the model, and ensures that the model can run efficiently on primary healthcare equipment. The classification prediction module captures the long-term temporal dependencies of lesions through the Transformer enhanced temporal classification head, completes the binary classification prediction of disease changes, and outputs accurate prediction results and probabilities. This provides reliable auxiliary support for clinicians to judge the patient's condition and formulate optimized treatment plans, enabling rapid and accurate prediction of changes in the condition of traumatic brain injury. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the change prediction model in the artificial intelligence-based method for predicting changes in the condition of patients with traumatic brain injury in this invention;

[0018] Figure 2 This is a schematic diagram of the CT sequence feature extraction module in the artificial intelligence-based method for predicting changes in the condition of patients with traumatic brain injury in this invention;

[0019] Figure 3 This is a schematic diagram of the dense feature extraction module in the artificial intelligence-based method for predicting changes in the condition of patients with traumatic brain injury in this invention;

[0020] Figure 4 This is a schematic diagram of the clinical feature encoding module in the artificial intelligence-based method for predicting changes in the condition of patients with traumatic brain injury in this invention;

[0021] Figure 5 This is a schematic diagram of the cross-modal feature distillation module in the artificial intelligence-based method for predicting changes in the condition of patients with traumatic brain injury in this invention;

[0022] Figure 6 This is a schematic diagram of the cross-modal dynamic fusion module in the artificial intelligence-based method for predicting changes in the condition of patients with traumatic brain injury in this invention;

[0023] Figure 7 This is a schematic diagram of the cross-modal prediction head in the artificial intelligence-based method for predicting changes in the condition of patients with traumatic brain injury in this invention;

[0024] Figure 8 This is the ROC curve comparing the change prediction model of this method with other existing technical models. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0026] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0027] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0028] In the description of this invention, it should be understood that the terms "upper," "lower," "inner," "outer," "left," "right," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product of this invention is in use, or the orientation or positional relationship commonly understood by those skilled in the art. They are only used to facilitate the description of this invention and to simplify the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0029] Furthermore, the terms "first," "second," etc., are used only to distinguish descriptions and should not be interpreted as indicating or implying relative importance.

[0030] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, terms such as "set" and "connection" should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be a connection within two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0031] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0032] like Figures 1-7 As shown, an artificial intelligence-based method for predicting changes in the condition of patients with traumatic brain injury includes:

[0033] S1. Construct an initial change prediction model. This model includes a CT sequence feature extraction module, a clinical feature encoding module, a cross-modal feature distillation module, a cross-modal dynamic fusion module, a cross-modal prediction head, and two single-modal prediction heads. The CT sequence feature extraction module preprocesses and extracts features from the input CT sequence data, outputting CT sequence single-modal features. The clinical feature encoding module standardizes and encodes structured clinical data, outputting global clinical features consistent with the dimensions of the CT sequence single-modal features. The cross-modal feature distillation module uses a teacher-student network structure to refine the global clinical features and the CT sequence single-modal features. The cross-modal dynamic fusion module adaptively and dynamically fuses the refined CT sequence single-modal features and the global clinical features, outputting bimodal features. The cross-modal prediction head outputs cross-modal prediction results based on the bimodal features. The two single-modal prediction heads respectively predict and output CT prediction results and clinical prediction results based on the CT sequence single-modal features and the global clinical features.

[0034] The CT sequence feature extraction module, from input to output, includes a preprocessing module, a dense feature extraction module, a temporal-spatial alignment module, and a temporal modeling module. The preprocessing module preprocesses the CT sequence data. The dense feature extraction module extracts fine-grained features from the processed CT sequence data, including spatial and temporal features. The temporal-spatial alignment module includes a bidirectional difference elimination sub-block (DF-Block) and a progressive alignment sub-block (BiA-Block). The DF-Block unifies the grayscale distribution through fine-grained feature calibration and adaptive normalization of the lesion region. The BiA-Block achieves inter-frame alignment by performing multi-step cumulative deformation field prediction on the fine-grained features after grayscale distribution unification. The deformation field of the nth frame in the CT sequence data... Represented as: , This represents the deformation field of the (n-1)th frame in the CT sequence data. This represents the deformation field increment of the nth frame relative to the (n-1)th frame; the temporal modeling module fuses the fine-grained features after inter-frame alignment and outputs the single-modal features of the CT sequence; the dense feature extraction module is an improved DenseNet network.

[0035] The clinical feature encoding module processes both discrete and continuous features. It includes a first fully connected layer, a first normalization layer for continuous feature processing, a second fully connected layer, a third fully connected layer, a clinical attention mechanism submodule, a first fusion layer, and a first multilayer perceptron (MLP). The first fully connected layer uses one-hot encoding to embed discrete clinical data. The first normalization layer uses Min-Max normalization to process continuous clinical data. The second fully connected layer enhances the dimensionality of the normalized continuous clinical data to increase its nonlinear representation capability. The third fully connected layer reduces dimensions to eliminate dimensional differences and improves the discriminative ability of continuous features. The clinical attention mechanism submodule calculates the correlation weights between various clinical indicators and disease changes and weights the features output by the third fully connected layer. The first fusion layer fuses the outputs of the first fully connected layer and the clinical attention mechanism submodule to obtain fused features. The first MLP performs nonlinear transformation and deep feature extraction on the fused features and outputs global clinical features with dimensions consistent with the single-modal features of the CT sequence.

[0036] The cross-modal feature distillation module includes a teacher network, a student network, a first linear layer, a second fusion layer, and a back alignment layer. The outputs of the CT sequence feature extraction module and the clinical feature encoding module serve as the inputs to the student network and the teacher network, respectively. The output of the student network serves as the input to the first linear layer. Both the teacher network and the student network include a second linear layer and a ReLU activation layer from input to output. The outputs of the first linear layer and the CT sequence feature extraction module serve as the inputs to the second fusion layer. The back alignment layer performs back distillation on the CT global features using clinical global features, outputting the MSE loss.

[0037] The cross-modal dynamic fusion module, from input to output, consists of a third fusion layer, a Dynamic Weight Assignment Network (DWAN), and a weighted fusion layer. The outputs of the CT sequence feature extraction module and the clinical feature encoding module serve as inputs to the third fusion layer and the weighted fusion layer. The DWAN, from input to output, consists of a fourth fully connected layer, a dropout layer, a fifth fully connected layer, and an activation layer. The DWAN adaptively assigns fusion weights between CT sequence single-modal features and clinical global features, expressed as: ,in, , These represent the fusion weights of CT sequence unimodal features and clinical global features, respectively. This is a single-modal feature of CT sequences. As a clinical feature, , These are the fourth and fifth fully connected layers, respectively. The activation function for the activation layer. The regularization layer; the weighted fusion layer multiplies the assigned weights element-wise with the unimodal features of the CT sequence and the global clinical features, and then fuses them to obtain bimodal features. , is represented as: ,in, , For cross-modal semantic mapping functions, , For learnable parameters, for Activation function.

[0038] The cross-modal prediction head consists of a global average pooling layer, a multi-head attention mechanism, a second normalization layer, a feedforward network layer, a third normalization layer, a flattening layer, and a second multilayer perceptron (MLP) from input to output. The multi-head attention mechanism is used to capture the long-term temporal dependencies of lesions in CT sequence data based on the optimized bimodal features, and to deeply explore the dynamic evolution of lesions.

[0039] S2. Obtain the training dataset, which includes multiple training samples, each of which includes CT sequence data and clinical structured data.

[0040] S3. Train and optimize the initial change prediction model using the training dataset to obtain the optimized change prediction model; the specific steps of training and optimizing the change prediction model include:

[0041] (1) The t-th training sample is input into the current CT sequence feature extraction module;

[0042] (2) The CT sequence feature extraction module extracts the CT sequence single-modal features of the CT sequence data in the t-th training sample;

[0043] (3) Calculate the first loss , is represented as: ,in, This represents the total number of frames in the CT sequence data. The features of the nth frame in the CT sequence data extracted by the CT sequence feature extraction module. Indicates bilinear interpolation transformation;

[0044] (4) Utilize the first loss Optimize the current CT sequence feature extraction module;

[0045] (5) Input the t-th training sample into the current change prediction model for prediction analysis;

[0046] (6) The reverse alignment layer performs reverse distillation of CT global features using clinical global features and outputs MSE loss;

[0047] (7) Calculate the fusion loss Cross-modal prediction head loss Single-mode prediction head loss and cross-modal characteristic distillation loss ; indicates as: ; ; , ; For cross-entropy loss, Weighted-importance similarity loss The weighting coefficients, For positive contrast loss, For inverse MSE alignment loss, , , , ;in, This refers to the batch sample size. For the true labels of the samples, For cross-modal prediction results, , These are the weighting coefficients for positive and negative categories in cross-modal predictions, respectively; and These are the fusion weights of the single-modal features and the global clinical features of the CT sequence of the i-th sample, respectively; and These represent the instantaneous accuracy of the current batch of CT prediction results and clinical prediction results, respectively. and These are the CT prediction results and the clinical prediction results, respectively. For cosine similarity, For temperature coefficient, and These are the refined CT sequence monomodal features and clinical global features, respectively.

[0048] Cross-entropy loss This is used to constrain the Dynamically Weighted Network (DWAN) to avoid excessive weight for a single modality and to balance the contributions of CT and clinical modalities. Higher entropy indicates a more balanced weight distribution; weight-importance similarity loss. To match the modal weights to the single-modal prediction accuracy, higher weights are assigned to modes with better discrimination performance; Distance alignment weights and modality discrimination capability; Single-mode prediction head loss KL divergence is used to constrain the CT and clinical single-modal prediction distributions to approximate and fuse the cross-modal prediction distributions, while preserving the unique discriminative information of the single modality;

[0049] Cross-modal characteristic distillation loss Including positive clinical guidance imaging (InfoNCE contrast loss), reverse image verification clinical (MSE alignment loss); positive contrast loss (Clinical features guide CT feature refinement); Projecting CT and clinical features onto a unified 128-dimensional contrast space, bringing similar samples closer together and dissimilar samples further apart. Temperature coefficient; reverse MSE alignment loss (CT feature verification of clinical feature semantic consistency) constrains the distribution of projection features of the two modalities of the same sample to be close, eliminating the cross-modal semantic gap; 0.1 is the inverse loss weight to avoid the excessive coverage of image visual features by clinical indicators.

[0050] (8) Calculate the joint loss using CMDF loss, cross-modal loss, single-modal loss, and cross-modal characteristic distillation loss. and through joint losses The current change prediction model is optimized to obtain an updated change prediction model; joint loss Represented as: , , and Cross-modal prediction head loss Single-mode prediction head loss and cross-modal characteristic distillation loss Weighting coefficients;

[0051] (9) Determine whether the training termination condition has been met. If yes, remove the teacher network and back alignment layer of the cross-modal feature distillation module in the updated change prediction model to obtain the optimized change prediction model and proceed to S4; otherwise, proceed to (10).

[0052] (10) Let t = t + 1 and return (1).

[0053] S4. Obtain the data to be predicted and analyzed, which includes CT sequence data and structured clinical data;

[0054] S5. Analyze the data to be predicted using the optimized change prediction model, and output cross-modal prediction results and single-modal prediction results;

[0055] S6. Perform fusion analysis on the cross-modal prediction results and the single-modal prediction results to obtain the change prediction results of the data to be analyzed. Specifically, this includes:

[0056] S61. Calculate the classification accuracy of CT prediction results and clinical prediction results based on the current change prediction model. and Calculate adaptive weights for CT sequence data and structured clinical data. and , is represented as , , ;

[0057] S62, Utilizing Adaptive Weights and The CT prediction results and clinical prediction results are weighted by confidence level to obtain a single-modal prediction value. , is represented as: ;

[0058] S63, Using global balance coefficients For cross-modal prediction results and single-modal prediction values Weighted fusion is performed to obtain the final predicted probability of disease progression. , is represented as: ;

[0059] S64. Predicting probability based on the final change in the condition. The results are analyzed against a preset judgment threshold to obtain the predicted changes; the judgment threshold is 0.5. To determine if the patient's condition has changed; if The patient's condition was determined to be stable.

[0060] First, the 5D CT sequence data and structured clinical data acquired clinically are input into their respective processing modules. The CT sequence data enters the CT sequence feature extraction module, where it undergoes a series of operations including preprocessing, dense feature extraction, temporal-spatial alignment, and temporal modeling. This completes the extraction and frame alignment of CT sequence monomodal features, ensuring the accuracy and continuity of CT image features and providing high-quality imaging support for disease prediction. Simultaneously, the structured clinical data enters the clinical feature encoding module, where it undergoes discrete feature embedding, continuous feature normalization, and clinical attention mechanism processing to convert it into clinical features with dimensions consistent with the monomodal features of the CT sequences. Global features are used to enhance the role of key clinical indicators and lay the foundation for subsequent fusion. Then, the extracted CT sequence monomodal features are input into the cross-modal feature distillation module for refinement to obtain refined CT sequence monomodal features. The refined CT sequence monomodal features and the encoded clinical global features are then input into the cross-modal dynamic fusion module. The dynamic weight allocation network automatically assigns fusion weights to the two modalities. Combined with the cross-modal semantic alignment mechanism, sample-level adaptive fusion is achieved, eliminating the cross-modal semantic gap and giving full play to the complementary advantages of the two data to generate fused bimodal features. Next, the bimodal features are input into the cross-modal prediction head and the unimodal prediction head, respectively. The Transformer-enhanced cross-modal prediction head captures the long-term temporal dependencies of the lesions and outputs the cross-modal prediction results. The unimodal prediction head outputs the unimodal prediction results based on the bimodal features. Finally, the cross-modal prediction results and the unimodal prediction results are fused and analyzed to obtain the change prediction results of the data to be analyzed. The change prediction results include binary classification prediction results and prediction probabilities, providing reliable auxiliary support for clinicians to judge the patient's condition and formulate optimized treatment plans, and realizing rapid and accurate prediction of changes in the condition of traumatic brain injury.

[0061] To verify the effectiveness, superiority, and practicality of this method, a systematic and comprehensive experimental verification was conducted. The relevant experimental data are as follows: The dataset used in the experiment came from 700 patients with traumatic brain injury from three tertiary hospitals, including 560 cases in the training set, 70 cases in the validation set, and 70 cases in the test set. All patients were clinically diagnosed. The dataset includes the patients' CT sequence data and corresponding structured clinical data. The CT sequence data are all in DICOM format, with a sequence length of 24 frames and a uniform image size of 512×512 pixels per frame. The clinical data includes key indicators such as the patients' age, gender, type of injury, Glasgow Coma Scale (GCS) score, blood pressure, blood glucose, complete blood count, and coagulation function. The data quality meets the clinical diagnosis and treatment standards and can effectively support the experimental verification.

[0062] The evaluation metrics used in the experiment included accuracy, sensitivity, specificity, and AUC-ROC (area under the receiver operating characteristic curve). Accuracy measures the overall accuracy of the model's predictions; sensitivity measures the model's ability to identify patients with changing conditions; specificity measures the model's ability to identify patients with stable conditions; and AUC-ROC comprehensively measures the model's classification performance. Higher values ​​indicate better model performance. The experimental comparisons included traditional manual image reading methods, prediction models based on traditional machine learning (Support Vector Machine, Random Forest), prediction models based on single-modal deep learning such as ResNet50 and ViT, and existing dual-modal fusion models.

[0063] like Figure 8 As shown in the experimental results, the performance indicators of the proposed method are significantly better than those of the comparison model: the accuracy of the proposed method is 84.56%, which is 16.33 percentage points higher than the traditional manual image reading method (68.23%), 12.41 percentage points higher than the support vector machine model (72.15%), 6.22 percentage points higher than the ResNet50 model (78.34%), and 4.44 percentage points higher than the existing bimodal fusion model (80.12%); the sensitivity of the proposed method is 92.35%, which can effectively identify most patients with changes in their condition and avoid missed diagnoses, an improvement of 6.64 percentage points compared to the existing bimodal fusion model (85.71%); the AUC-ROC of the proposed method is 91.74%, significantly higher than the AUC-ROC values ​​of the comparison models (traditional manual image reading 69.50%, support vector machine 75.30%, ResNet50...). (82.40%, compared to 87.60% for existing dual-modal fusion models), indicating that the model has superior overall classification performance.

[0064] Ablation experiments were conducted to compare the proposed model with existing models, and the implementation structures are shown in Tables 1, 2, and 3. Robustness test results show that the proposed model has strong anti-interference capabilities: under noisy data interference (adding Gaussian noise of different intensities to CT images), the model performance decreased by ≤2.41%, while the existing bimodal fusion model's performance decreased by ≥5.32%; in feature loss scenarios (randomly missing 10%-30% of clinical features), the model performance decreased by ≤2.58%, while the existing bimodal fusion model's performance decreased by ≥6.15%; in small sample scenarios (training set reduced to 140 cases), the model performance decreased by ≤4.44%, while the existing bimodal fusion model's performance decreased by ≥8.76%. All test results are significantly better than traditional models, fully demonstrating the superiority, robustness, and practicality of the proposed method, and its ability to adapt to complex and diverse clinical diagnostic and treatment scenarios.

[0065] Table 1. Comparison of performance indicators of the ablation experimental test set (mean ± standard deviation)

[0066]

[0067] Table 2 Comparison of performance metrics for the three types of model test sets

[0068]

[0069] Table 3. Performance comparison of each model in the CT classification task of traumatic brain injury sequence.

[0070]

[0071] An AI-based system for predicting changes in the condition of patients with traumatic brain injury includes:

[0072] Storage; storage is used to store computer programs;

[0073] An actuator; the actuator is used to execute a computer program stored in a memory, which, when executed, implements the artificial intelligence-based method for predicting changes in the condition of patients with traumatic brain injury as described above.

[0074] The technical solutions of the present invention are not limited to the specific embodiments described above. Any technical modifications made in accordance with the technical solutions of the present invention fall within the protection scope of the present invention.

Claims

1. A method for predicting changes in the condition of patients with traumatic brain injury based on artificial intelligence, characterized in that, include: S1. Construct an initial change prediction model. The change prediction model includes a CT sequence feature extraction module, a clinical feature encoding module, a cross-modal feature distillation module, a cross-modal dynamic fusion module, a cross-modal prediction head, and two single-modal prediction heads. The CT sequence feature extraction module is used to preprocess and extract features from the input CT sequence data, outputting CT sequence single-modal features. The clinical feature encoding module is used to standardize and encode structured clinical data, outputting global clinical features with dimensions consistent with the CT sequence single-modal features. The cross-modal feature distillation module has a teacher-student network structure and is used to refine the global clinical features and CT sequence single-modal features. The cross-modal dynamic fusion module is used to adaptively and dynamically fuse the refined CT sequence single-modal features and global clinical features, outputting bimodal features. The cross-modal prediction head outputs cross-modal prediction results based on bimodal features; two single-modal prediction heads are used to predict and output CT prediction results and clinical prediction results based on CT sequence single-modal features and clinical global features, respectively; the CT sequence feature extraction module includes a preprocessing module, a dense feature extraction module, a temporal-spatial alignment module, and a temporal modeling module from input to output. The preprocessing module is used to preprocess CT sequence data; the dense feature extraction module is used to extract fine-grained features from the processed CT sequence data, including spatial features and temporal features. The temporal-spatial alignment module includes a bidirectional difference elimination sub-block (DF-Block) and a progressive alignment sub-block (BiA-Block). The DF-Block unifies the grayscale distribution through fine-grained feature calibration and adaptive normalization of the lesion region. The BiA-Block achieves inter-frame alignment by performing multi-step cumulative deformation field prediction on the fine-grained features after grayscale distribution unification. The deformation field of the nth frame in the CT sequence data... Represented as: , This represents the deformation field of the (n-1)th frame in the CT sequence data. This represents the deformation field increment of the nth frame relative to the (n-1)th frame; the temporal modeling module fuses the fine-grained features after inter-frame alignment and outputs the single-modal features of the CT sequence; S2. Obtain the training dataset, which includes multiple training samples, each of which includes CT sequence data and clinical structured data. S3. Use the training dataset to train and optimize the initial change prediction model to obtain the optimized change prediction model. S4. Obtain the data to be predicted and analyzed, which includes CT sequence data and structured clinical data; S5. Analyze the data to be predicted using the optimized change prediction model, and output cross-modal prediction results, CT prediction results and clinical prediction results. S6. Perform fusion analysis on cross-modal prediction results, CT prediction results and clinical prediction results to obtain the change prediction results of the data to be analyzed.

2. The method for predicting changes in the condition of patients with traumatic brain injury based on artificial intelligence according to claim 1, characterized in that, The clinical feature encoding module includes a first fully connected layer, a first normalization layer, a second fully connected layer, a third fully connected layer, a clinical attention mechanism submodule, a first fusion layer, and a first multilayer perceptron (MLP). The first fully connected layer uses one-hot encoding to embed discrete clinical data; the first normalization layer uses Min-Max normalization to process continuous clinical data; the second fully connected layer is used to increase the dimensionality of the normalized continuous clinical data to enhance its nonlinear representation capability; the third fully connected layer eliminates dimensionality differences through dimensionality reduction; the clinical attention mechanism submodule calculates the correlation weights between each clinical indicator and changes in the patient's condition, and weights the features output by the third fully connected layer. The first fusion layer is used to fuse the output of the first fully connected layer and the output of the clinical attention mechanism submodule to obtain fusion features; The first multilayer perceptron (MLP) is used to perform nonlinear transformation and depth feature extraction on the fused features, and output clinical global features consistent with the dimensionality of the single-modal features of the CT sequence.

3. The method for predicting changes in the condition of patients with traumatic brain injury based on artificial intelligence according to claim 1, characterized in that, The cross-modal feature distillation module includes a teacher network, a student network, a first linear layer, a second fusion layer, and a back alignment layer. The outputs of the CT sequence feature extraction module and the clinical feature encoding module serve as the inputs to the student network and the teacher network, respectively. The output of the student network serves as the input to the first linear layer. Both the teacher network and the student network include a second linear layer and a ReLU activation layer from input to output. The outputs of the first linear layer and the CT sequence feature extraction module serve as the inputs to the second fusion layer. The back alignment layer performs back distillation on the CT global features using clinical global features, outputting the MSE loss.

4. The method for predicting changes in the condition of patients with traumatic brain injury based on artificial intelligence according to claim 1, characterized in that, The cross-modal dynamic fusion module, from input to output, consists of a third fusion layer, a Dynamic Weight Assignment Network (DWAN), and a weighted fusion layer. The outputs of the CT sequence feature extraction module and the clinical feature encoding module serve as inputs to the third fusion layer and the weighted fusion layer. The DWAN, from input to output, consists of a fourth fully connected layer, a dropout layer, a fifth fully connected layer, and an activation layer. The DWAN adaptively assigns fusion weights between CT sequence single-modal features and clinical global features, as shown below: ,in, , These represent the fusion weights of CT sequence unimodal features and clinical global features, respectively. This is a single-modal feature of CT sequences. As a clinical feature, , These are the fourth and fifth fully connected layers, respectively. The activation function for the activation layer. The regularization layer; the weighted fusion layer multiplies the assigned weights element-wise with the unimodal features of the CT sequence and the global clinical features, and then fuses them to obtain bimodal features. , is represented as: ,in, , For cross-modal semantic mapping functions, , For learnable parameters, for Activation function.

5. The method for predicting changes in the condition of patients with traumatic brain injury based on artificial intelligence according to claim 1, characterized in that, The cross-modal prediction head consists of a global average pooling layer, a multi-head attention mechanism, a second normalization layer, a feedforward network layer, a third normalization layer, a flattening layer, and a second multilayer perceptron (MLP) from input to output. The multi-head attention mechanism is used to capture the long-term temporal dependencies of lesions in CT sequence data based on the optimized bimodal features, and to deeply explore the dynamic evolution of lesions.

6. The method for predicting changes in the condition of patients with traumatic brain injury based on artificial intelligence according to claim 3, characterized in that, Training and optimizing the change prediction model specifically includes: (1) The t-th training sample is input into the current CT sequence feature extraction module; (2) The CT sequence feature extraction module extracts the CT sequence single-modal features of the CT sequence data in the t-th training sample; (3) Calculate the first loss , is represented as: ,in, This represents the total number of frames in the CT sequence data. The features of the nth frame in the CT sequence data extracted by the CT sequence feature extraction module. Indicates bilinear interpolation transformation; (4) Utilize the first loss Optimize the current CT sequence feature extraction module; (5) Input the t-th training sample into the current change prediction model for prediction analysis; (6) The reverse alignment layer performs reverse distillation of CT global features using clinical global features and outputs MSE loss; (7) Calculate the fusion loss Cross-modal prediction head loss Single-mode prediction head loss and cross-modal characteristic distillation loss ; indicates as: ; ; , ; For cross-entropy loss, Weighted-importance similarity loss The weighting coefficients, For positive contrast loss, For inverse MSE alignment loss, , , , ;in, This refers to the batch sample size. For the true labels of the samples, For cross-modal prediction results, , These are the weighting coefficients for positive and negative categories in cross-modal predictions, respectively; and These are the fusion weights of the single-modal features and the global clinical features of the CT sequence of the i-th sample, respectively; and These represent the instantaneous accuracy of the current batch of CT prediction results and clinical prediction results, respectively. and These are the CT prediction results and the clinical prediction results, respectively. For cosine similarity, For temperature coefficient, and These are the refined CT sequence monomodal features and clinical global features, respectively. (8) Calculate the joint loss using CMDF loss, cross-modal loss, single-modal loss, and cross-modal characteristic distillation loss. and through joint losses The current change prediction model is optimized to obtain an updated change prediction model; joint loss Represented as: , , and Cross-modal prediction head loss Single-mode prediction head loss and cross-modal characteristic distillation loss Weighting coefficients; (9) Determine whether the training termination condition has been met. If yes, remove the teacher network and back alignment layer of the cross-modal feature distillation module in the updated change prediction model to obtain the optimized change prediction model and proceed to S4; otherwise, proceed to (10). (10) Let t = t + 1 and return (1).

7. The artificial intelligence-based system for predicting changes in the condition of patients with traumatic brain injury according to claim 1, characterized in that, S6 includes: S61. Calculate the classification accuracy of CT prediction results and clinical prediction results based on the current change prediction model. and Calculate adaptive weights for CT sequence data and structured clinical data. and , is represented as , , ; S62, Utilizing Adaptive Weights and The CT prediction results and clinical prediction results are weighted by confidence level to obtain a single-modal prediction value. , is represented as: ; S63, Using global balance coefficients For cross-modal prediction results and single-modal prediction values Weighted fusion is performed to obtain the final predicted probability of disease progression. , is represented as: ; S64. Predicting probability based on the final change in the condition. The results are analyzed against preset judgment thresholds to obtain change prediction results.

8. An artificial intelligence-based system for predicting changes in the condition of patients with traumatic brain injury, characterized in that: include: Storage; Storage is used to store computer programs; Actuator; The actuator is used to execute a computer program stored in the memory, which, when executed, implements the artificial intelligence-based method for predicting changes in the condition of patients with traumatic brain injury as described in any one of claims 1-7.

Images