A dynamic evaluation method for the efficacy of Bragg therapy driven by multi-cycle, multimodal peripheral blood data.

Abstract

This invention belongs to the field of tumor detection and relates to a dynamic evaluation method for the efficacy of Bragg therapy driven by multi-cycle, multimodal peripheral blood data. The method includes the following steps: acquiring peripheral blood multimodal data from patients participating in Bragg therapy, and dividing the peripheral blood multimodal data into a training dataset and a validation dataset; constructing a Bragg therapy efficacy evaluation model using the training dataset, and validating the Bragg therapy efficacy evaluation model using the validation dataset; receiving peripheral blood multimodal data from patients participating in Bragg therapy, and calculating the evaluation results using the Bragg therapy efficacy evaluation model. This invention, by combining long short-term memory networks and graph neural networks, can effectively track the temporal changes of immune cell characteristics, thereby providing a more accurate basis for developing personalized treatment plans for Bragg therapy patients, improving treatment outcomes and survival rates.

Description

Technical Field

[0001] This invention belongs to the field of tumor detection and relates to a dynamic evaluation method for the efficacy of Bragg therapy driven by multi-cycle, multi-modal peripheral blood data. Background Technology

[0002] Current methods for evaluating cancer treatment primarily rely on predictive models built from immune cell characteristics and clinical data. This approach typically involves the following steps: First, collecting basic clinical information (such as age, sex, and tumor type) and biomarkers (including routine blood test results (such as tumor markers and lymphocyte levels), cytokine levels (such as interleukin-2 and interleukin-6), and refined immune cell typing data (such as effector memory CD4 T cells, effector memory CD8 T cells, and exhausted T cells)) from cancer patients. Then, statistical and machine learning methods (including logistic regression and random forests) are used to analyze this data and construct a predictive model for immunotherapy response. This method aims to assess patient treatment outcomes by identifying key features related to the immune response.

[0003] However, existing technologies are limited in that they typically rely on single-modality data analysis (i.e., a single type of data source, such as radiomics features from positron emission tomography / computed tomography (PET / CT), genomics data, immune cell typing, or hematological test indicators), without comprehensively considering multimodal data. Therefore, existing methods still fall short in terms of accuracy and clinical applicability, making it difficult to meet the needs of personalized medicine. Summary of the Invention

[0004] This invention provides a dynamic evaluation method for the efficacy of Bragg therapy driven by multi-cycle, multi-modal peripheral blood data, which can achieve a comprehensive assessment of the patient's immune status.

[0005] The technical solution provided by this invention is as follows:

[0006] A multi-cycle, multi-modal peripheral blood data-driven method for dynamic evaluation of the efficacy of Bragg therapy includes the following steps:

[0007] Peripheral blood multimodal data of patients participating in the Bragg treatment were obtained, and the peripheral blood multimodal data were divided into training dataset and validation dataset;

[0008] The peripheral blood multimodal data includes peripheral blood test data;

[0009] The peripheral blood test data includes cytokine level test results and detailed immune cell typing test results;

[0010] A Bragg treatment efficacy evaluation model was constructed using the training dataset, and the Bragg treatment efficacy evaluation model was validated using the validation dataset.

[0011] The Bragg treatment efficacy evaluation model is constructed using graph neural networks and long short-term memory networks;

[0012] Peripheral blood multimodal data from patients participating in Bragg therapy are received and calculated using the Bragg therapy tumor efficacy assessment model to output assessment results.

[0013] Furthermore, the peripheral blood multimodal data includes peripheral blood test data before treatment, at each treatment cycle, and during efficacy assessment.

[0014] Furthermore, the peripheral blood test data also includes routine blood test results; these routine blood test results include white blood cell count, hemoglobin, platelets, neutrophils, lymphocytes, liver function, kidney function, electrolytes, and thyroid function test results; the cytokine level test results include interleukin-10, granulocyte-macrophage stimulating factor, interleukin family, interferon, and tumor necrosis factor test results; and the refined immune cell typing test results include CD8+. + T cell rate of change, magnitude of PD-1 expression decline, early changes in interleukin-10 levels, CD4 + Changes in the proportion of effector memory T cells and NK cell activity.

[0015] Furthermore, a multi-head attention mechanism is used to integrate the preprocessed fine immune cell typing detection results to obtain a more comprehensive feature representation. The preprocessed routine blood test results, preprocessed cytokine level detection results, and clinical features are used as node features and input into the graph neural network. Through message passing and neighborhood aggregation, a high-dimensional representation vector of the node is obtained.

[0016] Furthermore, the preprocessing method for the routine blood test results and cytokine level test results is to standardize them using a normalization method; the preprocessing method for the refined immune cell typing test results is to standardize them using a Z-score method.

[0017] Furthermore, the data processed by the graph neural network is input into the long short-term memory network, and through temporal feature learning, the temporal evolution characteristics of the patient's treatment cycle are obtained.

[0018] Furthermore, based on the time evolution characteristics of the patient's treatment cycle, the probability prediction results of whether the patient responds effectively to immunotherapy are output.

[0019] Furthermore, the Bragg treatment efficacy evaluation model for tumors is trained using a binary cross-entropy loss function and an adaptive moment estimation optimizer.

[0020] This invention also provides a Bragg therapy efficacy evaluation system, which is used to implement the above-mentioned multi-cycle, multi-modal peripheral blood data-driven dynamic evaluation method for Bragg therapy efficacy, including:

[0021] The acquisition module is used to acquire peripheral blood multimodal data of patients participating in the Bragg treatment, and divide the peripheral blood multimodal data into a training dataset and a validation dataset;

[0022] The peripheral blood multimodal data includes peripheral blood test data;

[0023] The peripheral blood test data includes cytokine level test results and detailed immune cell typing test results;

[0024] The model building module is used to build a Prague tumor treatment efficacy evaluation model using pre-selected model parameters and the training dataset, and to validate the Prague tumor treatment efficacy evaluation model using the validation dataset.

[0025] The Bragg treatment efficacy evaluation model is constructed using graph neural networks and long short-term memory networks;

[0026] The evaluation module is used to receive peripheral blood multimodal data from patients participating in Bragg therapy and to calculate the evaluation results using the Bragg therapy tumor efficacy evaluation model.

[0027] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method as described above.

[0028] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method as described above.

[0029] Beneficial effects

[0030] Existing technologies for predicting the efficacy of immunotherapy primarily rely on analysis of single biomarkers and clinical characteristics. This approach often lacks integration of multimodal data, limiting predictive results. Specifically, the use of a single data source may fail to fully reflect the complex immune environment and its dynamic changes within the patient. This makes effective comparison and treatment matching between different patients difficult. This application addresses this by integrating peripheral blood multimodal data (routine blood test results, cytokine levels, and refined immune cell typing) and combining graph neural networks and long short-term memory networks to construct a more complex predictive model. This allows us to capture the interactions between various immune cells and time-series information. This comprehensive analysis method helps improve the accuracy and personalization of predictions. Furthermore, the multimodal data fusion method employed in this application effectively overcomes the shortcomings of existing technologies, enabling a comprehensive assessment of the patient's immune status.

[0031] Furthermore, existing technologies often fail to fully utilize the interaction information between immune cells during model construction. Many predictive models are primarily based on simple statistical methods, often lacking in-depth exploration of the potential relationships between different immune characteristics and ignoring the complex network relationships between immune cells. This results in weak model performance and limited predictive ability when capturing subtle differences in immune responses. Therefore, existing solutions still fall short in terms of accuracy and clinical applicability, making it difficult to meet the needs of personalized medicine. In contrast, this application proposes to introduce graph neural networks (GNNs), which can effectively capture the interrelationships between immune cells, thereby improving the accuracy and reliability of predictions.

[0032] Finally, existing technologies often lack in-depth analysis of time-series data, resulting in insufficient dynamic observation of immunotherapy responses. This leads to uncorrected residual issues regarding changes in the patient's immune status during multiple treatments, causing unnecessary treatment delays. The method of this invention, by incorporating Long Short-Term Memory (LSTM) networks, can effectively track the temporal changes in immune cell characteristics, thereby providing a more precise basis for developing personalized treatment plans for patients undergoing Bragg therapy, improving treatment outcomes and survival rates. Attached Figure Description

[0033] Figure 1 This is a flowchart of the prediction method of the present invention;

[0034] Figure 2 ROC curve for dynamic evaluation of the efficacy of Prague treatment;

[0035] Figure 3 This figure shows the changes in biomarkers and model predictions after treatment with Bragg. Detailed Implementation

[0036] Example 1

[0037] Peripheral blood multimodal data of patients participating in the Prague treatment were obtained, including relevant data before treatment, at each treatment cycle, and during efficacy evaluation. All necessary clinical data were collected through the hospital information system: the attending physician was responsible for recording clinical data, the laboratory physician was responsible for blood sample collection and testing, and the dedicated research nurse was responsible for data collation and entry. All data were stored in the clinical research database after standardized operating procedures and quality control.

[0038] The specific data collected included: routine blood tests (complete blood count: white blood cell count (WBC), hemoglobin (Hb), platelets (PLT), neutrophils (Neut), lymphocytes (Lymph), etc.; biochemical indicators: liver function (LFT), kidney function (RFT), electrolytes (Electrolyte), etc.; thyroid function: triiodothyronine (T3), thyroid-stimulating hormone (TSH), etc.), cytokine levels (interleukin-10 (IL-10), interleukin-6 (IL-6), tumor necrosis factor-α (TNF-α), granulocyte-macrophage colony-stimulating factor (GM-CSF), interferon-γ (IFN-γ), etc.), and refined immune cell typing (CD4+). + Effector memory T cells (CD4) + effector memory T cells), CD8 + T cells (CD8) + T cells, natural killer (NK) cells, dendritic cells, and programmed death receptor 1 / programmed death ligand 1 expression levels (PD-1 / PD-L1 expression levels, etc.).

[0039] Data preprocessing specifically includes standardizing and normalizing blood test results, cytokine levels, and immune cell typing data to eliminate batch effects and ensure data consistency. For continuous variables (such as routine blood test data and cytokine level data), a minimum-maximum normalization method is used to standardize the data, mapping it to the [0,1] interval. The calculation formula is as follows:

[0040] x 标准化 = (x - x 最小值 ) / (x 最大值 - x 最小值 ),

[0041] Where x 标准化 Here, x represents the standardized data, and x is the data to be preprocessed. 最小值 x is the minimum value of this data.最小值 This represents the maximum value of the data.

[0042] Preprocessed routine blood test data and preprocessed cytokine level data are obtained to eliminate systematic differences between different test batches. Z-score normalization is used for refined immune cell typing data, i.e.:

[0043] x 标准值 = (x -μ) / σ,

[0044] Where x 标准化 For the standardized data, x is the data to be preprocessed, μ is the sample mean, and σ is the sample standard deviation, so that the mean of the transformed data is 0 and the standard deviation is 1, in order to eliminate the batch effect caused by different detection platforms and obtain the preprocessed fine immune cell typing data.

[0045] The first tumor progression event occurring in the predicted target was selected as a positive example, and the five treatment cycles with an interval of at least 14 days prior to the predicted target were selected as input features for the Prague cancer treatment efficacy assessment model.

[0046] Discrete features are represented using one-hot encoding to facilitate subsequent model training.

[0047] A multi-head attention mechanism was employed to integrate preprocessed routine blood test data, cytokine level data, and refined immune cell typing data, thereby constructing a more comprehensive feature representation and laying the foundation for subsequent model training. The preprocessed refined immune cell typing data included CD4+. + Effector memory T cells (CD4) + effectormemory T cells), CD8 + T cells (CD8) + Immune-related indicators include T cells, natural killer (NK) cells, dendritic cells, and PD-1 / PD-L1 expression levels. These features represent different dimensions of immune status information in the tumor microenvironment.

[0048] The integration process mainly includes four key steps: feature encoding, multi-head projection, attention calculation, and feature fusion. First, different types of immune features are converted into vector representations. Then, they are mapped to multiple attention heads through linear projection. Next, a single attention head score is calculated in each head. Finally, the outputs of multiple single attention head scores are concatenated to obtain the multi-head attention score, and then mapped through a linear layer to obtain the integrated feature representation.

[0049] Specifically: (1) Feature encoding steps: convert different types of immune features into vector representations.

[0050] Vector representation of routine blood test data: X 血常规 ∈ R^(d 血常规 ) ;

[0051] Among them, X 血常规 Let R be a vector representation of routine blood test data, where R is the set of real numbers and d is the vector representation of the data. 血常规 These are the results of routine blood tests after pretreatment;

[0052] Cytokine level data vector representation: X 细胞因子 ∈ R^(d 细胞因子 ) ;

[0053] Fine-grained immune cell typing data vector representation: X 精细免疫细胞分型 ∈ R^(d 精细免疫细胞分型 ) ;

[0054] The initial feature vector is obtained by concatenating the above three vectors: X = [X 血常规 ; X 细胞因子 ; X 精细免疫细胞分型 ] ∈ R^d, where d = d 血常规 + d 细胞因子 + d 精细免疫细胞分型 ;

[0055] Where X represents the initial feature vector obtained by concatenating the three vectors, which are: X representing conventional blood test data. 血常规 X represents cytokine level data 细胞因子 And X, representing detailed immune cell typing data 精细免疫细胞分型 In this formula, R represents the set of real numbers, and d represents the total dimension of the initial feature vector X after concatenation. Its value is obtained by adding the dimensions of each part, i.e., d 血常规 (Dimensions of a routine blood test data vector), d 细胞因子 (Dimensions of the cytokine level data vector) and d 精细免疫细胞分型 The sum of the dimensions of the fine immune cell typing data vector.

[0056] (2) Multi-head projection step: Mapping to multiple attention heads via linear projection. For the i-th attention head, the initial feature vector X is transformed into query (Q), key (K), and value (V) matrices through three linear transformations:

[0057] Q i =X×W i Q W iQ ∈ R^(d×d q ) is a learnable query projection matrix;

[0058] K i =X×W i K W i K ∈ R^(d×d k ) is a learnable bond projection matrix;

[0059] V i =X×W i V W i V ∈ R^(d×d v ) is a learnable value projection matrix;

[0060] Here, i represents the i-th attention head in the multi-head attention mechanism. Q i K i 、 and V i These are the query matrix, key matrix, and sum-value matrix of the i-th attention head, respectively. i Q W i K and W i V These are the learnable query projection matrix, key projection matrix, and value projection matrix specifically for the i-th head, respectively. The parameters of these matrices are continuously optimized during model training. Furthermore, R represents the set of real numbers; d is the dimension of the initial feature vector X; and d... q d k and d v These are the dimensions of the query, key, and value vectors after projection transformation, respectively. This step multiplies the initial feature vector X, formed by concatenating various data types, with three independent learnable weight matrices to obtain the three core matrices—query (Q), key (K), and value (V)—which are linearly transformed (or projected) into the attention mechanism.

[0061] (3) Attention calculation steps: Calculate the individual attention head score in each head.

[0062] The formula for calculating a single attention head is:

[0063]

[0064] in For the output of the i-th attention head, This is a scaling factor used to prevent gradient vanishing. In the attention calculation formula, i represents the i-th attention head in the multi-head mechanism, and its calculation involves three core matrices: the query matrix Q. i Key matrix K i and the value matrix V i S() is the softmax function. During calculation, the key matrix K is first... i Perform the transpose (denoted by the superscript T), then combine it with the query matrix Q. i A dot product operation is performed. The resulting score is divided by a scaling factor to prevent gradient vanishing during training, thus making the training more stable. Then, a softmax function converts these scaled scores into a set of probability weights that sum to 1. Finally, these probability weights are used to adjust the value matrix V. i By performing a weighted summation, we obtain the final output of the attention head.

[0065] (4) Feature fusion step: concatenate and map the score outputs of multiple individual attention heads.

[0066] Concatenate the outputs of h attention heads:

[0067] Con(attention head 1, ..., attention head h) ∈ R^(h×d) v ) ;

[0068] The final multi-head attention output is obtained through linear layer mapping:

[0069] MHA(X) = Con(attention head 1, ..., Attention head h)W^O;

[0070] W^O∈R^(h×d v ×d m );

[0071] Where Con() is the concatenation function (Concat function), R is the set of real numbers, h is the h-th attention head, and d v This represents the dimension of the value vector after projection transformation, MHA(X) is the final multi-head attention output obtained through linear layer mapping, W^O is the learnable output parameter matrix, and d mThe final feature representation is obtained by integrating the dimensions of the final feature representation, realizing the learning of correlations between different immune features. This step takes multiple independent individual attention head scores as direct input and applies them through two core operations: First, all h independent attention head scores (attention head 1, ..., attention head h) are concatenated together to form a larger combination matrix; then, this concatenated matrix is ​​multiplied by a learnable output parameter matrix W^O to perform a final linear mapping, thereby obtaining the integrated multi-head attention output MHA(X).

[0072] The core formula used is the standard formula for multi-head attention mechanisms, and the formula for calculating a single attention head is as follows:

[0073]

[0074] in For a single attention head score, S() is the softmax function, Q is the query matrix, K is the key matrix, and T denotes the transpose. Let d be the scaling factor, V be the value matrix, and d be the scaling factor. k is the dimension of the key vector.

[0075] The calculation process for multi-head attention scores is as follows: (1) Obtain the score of a single attention head, and calculate the i-th attention head using the following formula:

[0076] Attention head i = Attn(Q×W) i Q , K×W i K V×W i V ),

[0077] Where attention head i is the score of the i-th attention head, Attn() is the score of a single attention head, and Q, K, and V are the query matrix, key matrix, and value matrix, respectively; W i Q W i K and W i V These are the learnable query projection matrix, key projection matrix, and value projection matrix specifically for the i-th head;

[0078] (2) Using the results from each attention head, the final multi-head attention output is obtained through concatenation and linear transformation:

[0079] MHA(Q,K,V) = Con(attention head 1,..., Attention head h)W^O

[0080] Where MHA(Q,K,V) is the multi-head attention output, Con() is the function (Concat function), and W^O is the learnable output parameter matrix;

[0081] (3) Multi-head attention output is adopted to obtain integrated feature representations through this mechanism and realize the correlation learning between different immune features.

[0082] First, different types of immune features are converted into vector representations, then mapped to multiple attention heads via linear projection. Next, a single attention head score is calculated in each head. Finally, the outputs of multiple single attention head scores are concatenated to obtain a multi-head attention score, which is then mapped through a linear layer to obtain the integrated feature representation.

[0083] More comprehensive feature representation is mainly reflected in four aspects: multi-angle information fusion, interrelationships between features, hierarchical representation, and context-aware representation. Each attention head can focus on different aspects of the feature, effectively learn the interaction relationships between different immune cell types, capture multi-level immune information from cells to molecules, and consider the role and status of each immune feature in the overall immune microenvironment, thereby providing a more comprehensive and accurate feature foundation for subsequent model training.

[0084] This study combines graph neural networks (GNNs) to analyze interactions between immune cells. Specifically, it constructs a heterogeneous relational graph by integrating the feature representations of routine blood test results, cytokine level test results, and refined immune cell typing results. Utilizing the node characteristics and edge relationships of the GNN, it captures complex interaction patterns between cells. Clinical features from each treatment cycle are used as nodes, and edges are established based on temporal relationships, feature correlations, and hierarchical relationships. Edge weights are defined based on a multidimensional edge weight calculation mechanism and weight aggregation formula. Patient clinical information and peripheral blood features are input into the GNN as node features. Through message passing and neighborhood aggregation, high-dimensional node representation vectors are obtained. The GNN model employs a heterogeneous relational graph convolutional network (HeteroRGCN) to learn node representations. The learned graph embeddings are concatenated with the original feature matrix and used as input to a Long Short-Term Memory (LSTM) model.

[0085] The high-dimensional representation vectors of nodes are input into a Long Short-Term Memory (LSTM) network, and temporal feature learning is used to obtain the temporal evolution features of the patient's treatment cycle. By training an integrated model of the LSTM network and a graph neural network, an integrated network structure of a two-layer heterogeneous relational graph convolutional network (HeteroRGCN) and a two-layer LSTM network is realized. The model is trained using a binary cross-entropy loss function and an adaptive moment estimation optimizer (Adam optimizer), thereby constructing a powerful predictive model to improve the accuracy of predicting the patient's response to immunotherapy.

[0086] Clinical characteristics include basic patient characteristics (age, sex, performance status score, body mass index, smoking status), tumor characteristics (tumor type and subtype, stage, tumor size and number, metastasis), treatment-related characteristics (previous treatment history, type of treatment drugs, number of lines of treatment, combination therapy, treatment-related adverse reactions), routine blood test results (lactate dehydrogenase level, neutrophil-to-lymphocyte ratio, platelet-to-lymphocyte ratio, albumin level, inflammatory markers, liver and kidney function indicators), and clinical outcome indicators (previous treatment response, progression-free survival, overall survival, objective response rate, efficacy classification), etc. These characteristics are usually obtained from electronic medical record systems, pathology reports, imaging examinations, and laboratory test results.

[0087] A key outcome of this study using graph neural networks is the learned node representations (graph embeddings) that capture the complex interaction patterns and relationships between immune cells. Specifically, the Hetero-Relational Graph Convolutional Network (HeteroRGCN) processes heterogeneous graphs containing three node types—routine blood test results, cytokine levels, and clinical features—and learns low-dimensional vector representations that characterize the intrinsic relationships between these nodes. These graph embeddings preserve the topological information and node characteristics, reflecting the signal transduction and mutual regulation relationships between immune system components. These embeddings are then concatenated with the original feature matrix as input to a Long Short-Term Memory (LSTM) model, forming an integrated model of LSTM and graph neural networks, thereby enhancing the predictive power and accuracy of patient immunotherapy responses.

[0088] The training of the Long Short-Term Memory (LSTM) model ultimately yielded a temporal prediction model capable of accurately predicting patient responses to immunotherapy. By concatenating the graph embeddings generated by the graph neural network with the original feature matrix as input, the LSTM model learned the dynamic patterns of immunotherapy response over time. Guided by a binary cross-entropy loss function and an adaptive moment estimation optimizer (Adam optimizer), the model effectively captured long-term dependencies and temporal patterns during treatment through a two-layer LSTM structure, ultimately outputting a probabilistic prediction of whether a patient responds effectively to immunotherapy.

[0089] The binary cross-entropy loss function and the adaptive moment estimation optimizer (Adam optimizer) train the output of an ensemble model integrating a long short-term memory network (LSTM) and a graph neural network (HeteroRGCN). During training, the input data is first processed by the HeteroRGCN to obtain graph embeddings. These embeddings are then concatenated with the original feature matrix and input into the LSTM layer. Finally, a fully connected layer generates a predicted probability representing the patient's response to immunotherapy. The model uses binary cross-entropy to calculate the difference between the predicted value and the true label, then calculates the gradient of the loss function with respect to the model parameters through backpropagation, and updates the parameters using the adaptive moment estimation optimizer. This optimizer combines momentum and adaptive learning rate methods, automatically adjusting the learning rate for different parameters and accelerating convergence. The entire training process typically employs mini-batch gradient descent, dividing the data into multiple batches for iterative training. Early stopping and learning rate decay techniques may also be used to improve the model's generalization ability, ultimately resulting in a time-series prediction model that accurately predicts patient responses to immunotherapy.

[0090] Finally, cross-validation was used to evaluate the model's performance, and the prediction results were compared with traditional efficacy assessment strategies to ensure the model's high accuracy and reliability. Based on the model's predictions, personalized treatment plans were developed for patients to improve treatment outcomes and patient survival rates.

[0091] The comparison results (Table 1) show that the prediction method based on the integrated model of long short-term memory network and graph neural network is superior to the traditional efficacy assessment strategy in terms of accuracy, sensitivity and specificity.

[0092] The specific steps for developing personalized treatment plans based on the model's predictions include: First, inputting the new patient's routine blood test results, cytokine levels, and clinical characteristics into the trained model to obtain the predicted probability of the patient's possible response to Bragg therapy; then, classifying patients into three categories—high response, intermediate response, and low response—based on the predicted probability. For patients predicted to have a high response, the current Bragg treatment plan can continue; for patients with an intermediate response, adjustments to the immunotherapy drug dosage or the addition of adjuvant therapies can be considered; and for patients predicted to have a low response, adjustments to other combination therapy strategies or conversion to other treatment plans can be made as early as possible; finally, patient data is collected periodically during treatment and re-predicted, dynamically adjusting the treatment plan to achieve real-time optimization of treatment, thereby maximizing treatment effectiveness, reducing unnecessary side effects, and improving the patient's quality of life and overall survival rate.

[0093] This study constructed several prediction models ranging from simple to complex.

[0094] First, basic machine learning models were established, including logistic regression for binary classification prediction using an activation function (sigmoid function) and random forest based on multi-decision tree ensemble.

[0095] Secondly, a series of sequence models were developed, including long short-term memory networks that use gating mechanisms to handle long-term dependencies, deep neural networks that use non-linear activation functions, bidirectional long short-term memory networks that can process sequence information bidirectionally, and transformers that use self-attention mechanisms.

[0096] Three hybrid architecture models were finally constructed: 1) Long Short-Term Memory Network-Graph Neural Network (LSTM-GNN) that combines temporal features; 2) Long Short-Term Memory Network-Graph Convolutional Network (LSTM-GCN) that integrates spatial relationships; and 3) Long Short-Term Memory Network-HeteroGraph Neural Network (LSTM-HeteroGNN) that models through heterogeneous graph structures.

[0097] All models used the same input features and their performance was evaluated using cross-validation. Multiple metrics were used for comprehensive evaluation, including the area under the receiver operating characteristic curve (AUC-ROC), F1 score, accuracy, recall, and area under the precision-recall curve (AUC-PR).

[0098] Table 1. Performance evaluation of multiple models using cross-validation.

[0099]

[0100] As can be seen from the above results, the method of the present invention, Long Short-Term Memory Network-Heterogeneous Graph Neural Network, can achieve accurate prediction of the treatment effect of tumor Bragg therapy, which helps to optimize treatment decisions, improve patients' survival rate and quality of life, and has broad clinical application potential.

[0101] Example 2

[0102] In this invention, we used data from a total of 135 patients participating in the Bragg treatment clinical trial to construct the evaluation model, with 80% (108 patients) serving as the training dataset and 20% (27 patients) as the independent validation dataset. For missing values ​​inevitably present in the clinical data, temporal interpolation, multiple imputation, and K-nearest neighbor imputation were used to supplement the data, ensuring data integrity and analytical quality. The training process employed 5-fold cross-validation, using an optimizer (initial learning rate 0.001) and an early stopping mechanism to prevent overfitting. Key model parameters included: a multi-head attention mechanism (8 attention heads, 256 hidden layer dimensions), a graph neural network (3 layers of graph convolution, edge weight calculation covering four weight dimensions: temporal relationship, feature correlation, hierarchical relationship, and lymphocyte pairs), and a long short-term memory neural network (2 layers, 64 hidden units). On the independent validation set, the model ultimately achieved 83.5% accuracy, 82.3% precision, 84.2% recall, and an area under the receiver operating characteristic (AUC) of 0.87, with a cross-validation standard deviation of less than 4%, indicating good model reliability (e.g., Figure 2 (As shown).

[0103] This invention assesses patient efficacy by acquiring basic patient information (age, gender, tumor type, stage, etc.) and peripheral blood multimodal data (routine blood tests, biochemical indicators, thyroid function, cytokine levels, and refined immune cell typing, etc.). The data processing workflow includes standardization, logarithmic transformation of immune cell data, calculation of key immune cell ratios, and construction of temporal features. The assessment process mainly consists of data input, multi-head attention integration of different types of immune features, construction and analysis of patient-cycle-feature heterogeneity graphs, long short-term memory network capture of time-series patterns, and finally outputting the patient's expected response category and probability to Bragg treatment. Efficacy is quantified by determining whether tumor progression has occurred and calculating an efficacy index of 0-100.

[0104] In practical application, treatment data from a 58-year-old female patient with non-small cell lung cancer showed that after three treatment cycles, the proportion of CD8-positive T cells (CD8+) decreased. + T cell counts increased from 12% to 18%, programmed death receptor 1 (PD-1) expression decreased from 45% to 28%, and interleukin-10 (IL-10) levels significantly decreased. The model predicted tumor stability (probability 0.76) and a efficacy index of 70 (e.g., Figure 3 (As shown). Subsequent clinical evaluation confirmed that the tumor shrank by 35%, validating the accuracy of the model's predictions.

[0105] In a prospective validation study of 29 newly recruited patients receiving Bragg therapy, this method achieved a predictive accuracy of 82% after the second treatment cycle, predicting treatment outcomes approximately 25 days earlier on average than traditional imaging assessments. Compared with existing efficacy prediction methods, this method demonstrates significant advantages in accuracy, sensitivity, and specificity. Through model analysis, we identified the most valuable immune features for predicting response to Bragg therapy, including: CD8+ T cell change rate, PD-1 expression decline, early IL-10 level change, and the proportion of CD4+ effector memory T cells. + The effects include changes in the effector-memory T cell ratio and natural killer cell activity (NK cell activity change).

[0106] Clinical applications show that this method can provide doctors with early treatment decision support. For example, when a hepatocellular carcinoma patient has mild liver function abnormalities in the first cycle, the model predicts that the treatment will be effective. Based on this, the doctor continues the treatment, and the patient eventually achieves good results.

[0107] The main advantages of this invention are that it enables early prediction, provides high accuracy, supports individualized assessment, enables dynamic monitoring, and for the first time successfully integrates peripheral blood multimodal data for efficacy evaluation, providing innovative technical support and decision-making basis for precision tumor immunotherapy.

Claims

1. A multi-cycle multi-modal peripheral blood data-driven Bragg treatment efficacy dynamic evaluation method, characterized in that, The method comprises the following steps: acquiring peripheral blood multi-modal data of patients participating in the Prague therapy and dividing the peripheral blood multi-modal data into a training data set and a validation data set; the peripheral blood multi-modal data comprises peripheral blood detection data before treatment, at each treatment cycle, and at efficacy evaluation; the peripheral blood detection data comprises cytokine level detection results and fine immune cell typing detection results; a Prague therapy tumor efficacy evaluation model is constructed using the training data set, and the Prague therapy tumor efficacy evaluation model is verified using the validation data set; the Prague therapy tumor efficacy evaluation model is constructed using a graph neural network and a long short-term memory network; a multi-head attention mechanism is used to integrate the preprocessed peripheral blood detection data to obtain integrated feature representations; the integrated feature representations and clinical features are input into the graph neural network as node features, and high-dimensional representation vectors of the nodes are obtained through message passing and neighborhood aggregation processing; the data processed by the graph neural network is input into the long short-term memory network, and time evolution features of the treatment cycles of the patients are obtained through time series feature learning processing; receiving peripheral blood multi-modal data of patients participating in the Prague therapy and calculating the peripheral blood multi-modal data using the Prague therapy tumor efficacy evaluation model to output evaluation results.

2. The multi-cycle multi-modal peripheral blood data-driven Bragg treatment efficacy dynamic assessment method of claim 1, wherein, The peripheral blood detection data further comprises routine blood test results; the routine blood test results comprise white blood cell count, hemoglobin, platelets, neutrophils, lymphocytes, liver function, kidney function, electrolyte, thyroid function test results, the cytokine level detection results comprise granulocyte-macrophage stimulating factor, interleukin family, interferon, tumor necrosis factor detection results; the fine immune cell typing detection results comprise CD8 + T cell change rate, PD-1 expression decrease amplitude, early interleukin 10 level change, CD4 + Effector memory T cell proportion and NK cell activity change.

3. The multi-cycle multi-modal peripheral blood data-driven Bragg treatment efficacy dynamic evaluation method of claim 2, wherein, The preprocessing method for the conventional blood detection results and the cytokine level detection results is standardization processing using a normalization method; the preprocessing method for the fine immune cell typing detection results is Z-score standardization processing.

4. The multi-cycle multi-modal peripheral blood data-driven Bragg treatment efficacy dynamic assessment method of claim 1, wherein, According to the time evolution features of the treatment cycles of the patients, probability prediction results of whether the patients have effective responses to the immunotherapy are output.

5. The multi-cycle multi-modal peripheral blood data-driven Bragg treatment efficacy dynamic assessment method of claim 1, wherein, The Prague therapy tumor efficacy evaluation model is trained using a binary cross-entropy loss function and an adaptive moment estimation optimizer.

6. A Bragg tumor treatment efficacy evaluation system, the evaluation system is used to implement the multi-cycle multi-modal peripheral blood data-driven Bragg treatment efficacy dynamic evaluation method in claim 1, characterized in that, The method comprises: an acquisition module configured to acquire peripheral blood multi-modal data of patients participating in the Prague therapy and divide the peripheral blood multi-modal data into a training data set and a validation data set; the peripheral blood multi-modal data comprises peripheral blood detection data before treatment, at each treatment cycle, and at efficacy evaluation; the peripheral blood detection data comprises cytokine level detection results and fine immune cell typing detection results; a model construction module configured to construct a Prague therapy tumor efficacy evaluation model using preselected model parameters and the training data set, and to verify the Prague therapy tumor efficacy evaluation model using the validation data set; the Prague therapy tumor efficacy evaluation model is constructed using a graph neural network and a long short-term memory network; a multi-head attention mechanism is used to integrate the preprocessed peripheral blood detection data to obtain integrated feature representations; the integrated feature representations and clinical features are input into the graph neural network as node features, and high-dimensional representation vectors of the nodes are obtained through message passing and neighborhood aggregation processing; the data processed by the graph neural network is input into the long short-term memory network, and time evolution features of the treatment cycles of the patients are obtained through time series feature learning processing; An evaluation module is configured to receive peripheral blood multi-modal data of a patient undergoing a Bragg treatment and to perform a computation thereon using the Bragg treatment tumor response evaluation model to output an evaluation result.

7. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor implements the steps of the method of any of claims 1-5 when executing the program.

Images