Machine Learning-Based Efficacy Assessment Method for Chemotherapy in Hematologic Malignancies with Pleuronecrosis and Fever

Abstract

This invention relates to the field of disease efficacy assessment technology, and discloses a machine learning-based method for assessing the efficacy of chemotherapy for hematologic malignancies with fever, including: S1: collecting and preprocessing multi-center medical data, the data including structured clinical data and multimodal unstructured data; S2: constructing an efficacy prediction model based on a graph neural network and training the model using graph federated learning based on central moment constraints, including: constructing a patient graph as the original graph and generating an enhanced graph; processing the original graph and enhanced graph separately using a dual-channel graph convolutional neural network to extract higher-order representation vectors; constructing an enhancement-perception loss function based on the higher-order representation vectors of the original graph and enhanced graph; training the model using a federated learning framework with multi-center data; and S3: predicting the efficacy assessment results using the trained efficacy prediction model. This invention improves the accuracy, stability, and clinical interpretability of efficacy assessment.

Description

Technical Field

[0001] This invention relates to the field of disease efficacy evaluation technology, and in particular to a machine learning-based method for evaluating the efficacy of chemotherapy for hematologic malignancies with fever. Background Technology

[0002] Febrile neutropenia (FN) is a common and serious complication of chemotherapy in patients with hematologic malignancies, characterized by rapid onset, rapid progression, high risk of infection, and high mortality. Accurate and timely assessment of treatment efficacy in FN patients helps guide individualized treatment decisions, reduce unnecessary antibiotic use, minimize waste of medical resources, and improve patient prognosis.

[0003] Currently, the assessment of FN efficacy mainly relies on clinical experience, changes in blood counts, inflammatory markers such as CRP, and judgment of infection sites. It lacks systematic and generalizable modeling methods, making it difficult to extract stable patterns from multi-center, multimodal, and multivariate data. Furthermore, the occurrence and development mechanisms of FN are complex, influenced by multiple factors such as individual patient differences, chemotherapy regimens, underlying diseases, and supportive treatment plans, forming a high-dimensional medical data scenario characterized by structural complexity, high heterogeneity, and significant noise disturbance.

[0004] In recent years, machine learning methods such as graph neural networks (GNN), federated learning, and distributed regression have been widely used in the medical field, providing new directions for modeling complex states of FN and predicting efficacy. However, existing methods still have the following technical bottlenecks: (1) unstable graph representation: graph models are prone to structural sensitivity when faced with data perturbations (such as missing features or different hospital record formats), affecting the discrimination effect; (2) difficulty in aligning heterogeneous distributions: the data distribution differences among multiple centers are large, and direct joint modeling carries the risk of generalization; (3) lack of deep modeling ability for graph structure changes: traditional graph distance is difficult to simultaneously characterize node features and structural information; (4) difficulty in aligning prediction targets with actual risk perception: small clinical changes may be crucial in the early stages, and existing loss functions are insufficient to characterize them.

[0005] Therefore, there is an urgent need for a efficacy assessment method that is more capable of perceiving differences, has a unified structural-feature measurement capability, can generalize across centers, and is both robust and interpretable. Summary of the Invention

[0006] This invention proposes a machine learning-based method for evaluating the efficacy of chemotherapy for hematologic malignancies with fever. It aims to address the problems of high sensitivity of graph modeling, poor model generalization ability, and insensitivity to structural changes in existing technologies, thereby improving the accuracy, stability, and clinical interpretability of efficacy evaluation.

[0007] To achieve the above objectives, the following technical solution is adopted:

[0008] A machine learning-based method for evaluating the efficacy of chemotherapy for hematologic malignancies with fever after chemotherapy includes the following steps:

[0009] Step S1: Collect and preprocess multi-center medical data, including structured clinical data and multimodal unstructured data; the multimodal unstructured data includes medical text data, medical image data and their derived feature data;

[0010] Step S2: Construct an efficacy prediction model based on a graph neural network and train the model using graph federated learning based on central moment constraints. This includes: using clinical variables in the preprocessed data as nodes V, constructing edges E based on the statistical or clinical correlation between clinical variables, and constructing a patient graph as the original graph G=(V,E); applying structural and feature perturbations to the original graph based on a graph augmentation mechanism to generate an augmented graph; processing the original graph and the augmented graph separately using a dual-channel graph convolutional neural network with shared weights to extract higher-order representation vectors; constructing an augmentation-aware loss function based on the higher-order representation vectors of the original graph and the augmented graph; and training the model using a federated learning framework with multi-center data and introducing central moment constraints to align feature distributions.

[0011] Step S3: Construct a patient graph from the new patient data, and output the probability distribution of effective and ineffective efficacy using the trained efficacy prediction model. Take the category with the highest probability as the efficacy evaluation result.

[0012] Furthermore, the preprocessing described in step S1 includes: data cleaning and standardization, missing value handling, and multimodal data alignment and labeling.

[0013] Furthermore, the graph enhancement mechanism described in step S2 includes at least one of the following perturbation methods: randomly deleting nodes or edges; reconnecting edge relationships; masking node feature values; linearly mixing features from different patients; and sampling feature subgraphs.

[0014] Furthermore, the structure of the efficacy prediction model includes an input layer, a dual-channel graph convolutional neural network, an orthogonal constraint module, a global pooling layer, a fully connected classification layer, and a Softmax output layer;

[0015] The input layer receives patient graph structure data; the dual-channel graph convolutional neural network includes two graph convolutional layers with shared weights, processing the original graph and the enhanced graph respectively; the global pooling layer generates graph-level representation vectors through attention-weighted averaging; the orthogonal constraint module applies orthogonal constraints to the weights of the graph convolutional layers; the fully connected classification layer maps the graph representation vectors to the category space; and the Softmax output layer generates the efficacy probability distribution.

[0016] Specifically, the dual-channel graph convolutional neural network includes: a first graph convolutional layer that reduces the input dimension to 64 dimensions and uses the ReLU activation function; and a second graph convolutional layer that reduces the 64-dimensional features to a 32-dimensional embedding vector.

[0017] The orthogonal constraint module constrains the weight matrix to satisfy the orthogonality condition through the Frobenius norm, and generates an orthogonal constraint loss function based on the orthogonality condition, so that the learned feature dimensions are independent of each other.

[0018] Furthermore, an enhancement-aware loss function is constructed based on the higher-order representation vectors of the original and enhanced graphs, including:

[0019] The distance between the predicted images is obtained by predicting the difference between the representation vectors of the original image and the augmented image using a multilayer perceptron.

[0020] The true inter-graph distance between the original graph and the augmented graph is calculated using a difference metric method that combines node features and graph structure.

[0021] The error between the predicted difference and the distance to the true graph is used as the loss constraint, and the L2 norm is used to constrain the consistency between the predicted distance and the distance between the true graph.

[0022] Furthermore, the method for calculating the true inter-graph distance between the original graph and the enhanced graph using the difference metric of fused node features and graph structure is a Fused-Gromov-Wasserstein distance calculation using the difference in fused node feature distribution and graph topology, including:

[0023] Calculate the Wasserstein distance between the node feature distributions to measure the differences in node features, i.e., their distributions;

[0024] Calculate the Gromov-Wasserstein distance of the graph topology to measure the differences in graph topology;

[0025] The final graph distance is obtained by dynamically balancing the weights of the differences in feature distribution and graph topology using an adjustable parameter α.

[0026] The optimal transmission algorithm is used to calculate the coupling matrix for node matching. This coupling matrix aligns node features with the graph structure by minimizing the comprehensive difference between the original graph and the enhanced graph.

[0027] Furthermore, the efficacy prediction model also includes a multi-objective optimizer for multi-objective joint optimization training of the efficacy prediction model; the loss function of the multi-objective joint optimization includes: cross-entropy loss of efficacy classification task, enhancement perception loss, consistency loss, orthogonal constraint loss, and central moment difference loss, and each loss term is dynamically weighted and balanced through hyperparameters;

[0028] The consistency loss function forces the model to output similar classification probabilities for the original image and the enhanced image of the same patient.

[0029] Furthermore, the central moment difference loss is calculated using the federated learning framework through the following steps:

[0030] Each medical center trains a local graph neural network model to extract patient graph embedding vectors.

[0031] Calculate the central moments of the patient embedding vector, including the mean, variance, and higher-order moments;

[0032] The server receives encrypted central moment parameters from each center and generates a global central moment through weighted aggregation.

[0033] Align local feature distributions with global distributions by constraining the difference in central moments;

[0034] The federated learning method only transmits the feature mean and higher-order central moment statistics between each center and the server.

[0035] Furthermore, the calculation of the central moment difference distance includes:

[0036] Calculate the first moment (mean distance) and the weighted distance of higher-order central moments between the local and global centers; where higher-order central moments are second-order and above central moments, including the weighted distance of variance, skewness and kurtosis.

[0037] The mean distance and weighted distance are normalized and scaled according to the range of eigenvalues.

[0038] Furthermore, the method also includes: employing differential privacy technology to add Laplace noise to the central moment parameters uploaded to each center to achieve privacy protection.

[0039] Compared with the prior art, the present invention achieves the following beneficial effects:

[0040] 1. This invention proposes an explicit modeling enhancement mechanism to construct an enhanced perception modeling strategy: By subjecting the patient graph structure to dual perturbations of structure and features (such as edge reconnection, feature masking, subgraph sampling, etc.), this invention constructs multiple enhanced graphs, guiding the model to identify differences between graphs. Simultaneously, an enhanced perception loss function is introduced, enabling the model not only to learn classification but also to possess "difference sensitivity" capabilities, thereby more effectively identifying subtle but crucial physiological changes after chemotherapy.

[0041] 2. This invention constructs a graph difference measurement mechanism that integrates Wasserstein distance based on node features and Gromov-Wasserstein distance based on graph structure: This invention designs the Fused-Gromov-Wasserstein graph distance, which comprehensively considers changes in graph node attributes and graph topology, and stably measures the global differences between graphs under perturbation, providing an accurate geometric basis for enhanced perception modeling and consistency regularization.

[0042] 3. This invention proposes a graph federated learning method based on central moment constraints: To address the problem of distribution differences among multi-center medical data, this invention introduces a central moment (such as mean, variance, skewness, and kurtosis) constraint mechanism to align the graph embedding distribution of patients from different institutions, thereby improving the model transfer and generalization capabilities under the federated learning framework.

[0043] 4. This invention proposes a local orthogonal graph representation and a feature orthogonality loss function: In order to improve the interpretability and redundancy removal of graph embedding, this invention constructs a local orthogonal graph structure and designs a feature orthogonality loss term, which makes the feature dimensions learned by the model independent of each other and complementary in information, thereby enhancing the ability to identify key variables in the diagnostic process.

[0044] 5. This invention proposes a central moment difference distance index to describe the degree of offset of multi-center distribution: This invention quantifies the offset between the embedded distribution of each center and the global distribution by constructing the central moment difference distance, thereby providing an adjustment signal for distribution alignment for federated optimization and realizing "homogeneous modeling of heterogeneous data".

[0045] 6. This invention proposes an overall loss function that integrates multiple factors such as enhanced perception, consistency regularization, and center alignment: This invention combines and optimizes the enhanced perception loss, structural consistency loss, feature orthogonality loss, and central moment difference loss to form an integrated, multi-objective guided joint loss function, which comprehensively improves the model's stability and discriminative power.

[0046] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of the present invention, nor is it intended to restrict the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0047] The above and other features, advantages, and aspects of the various embodiments of the present invention will become more apparent from the accompanying drawings and the following detailed description. The drawings are provided for a better understanding of the invention and are not intended to limit the invention. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0048] Figure 1This is a flowchart illustrating a machine learning-based method for evaluating the efficacy of chemotherapy for hematologic malignancies with fever after chemotherapy, according to an embodiment of the present invention.

[0049] Figure 2 This is a schematic diagram of the system framework of a machine learning-based method for evaluating the efficacy of chemotherapy for hematologic malignancies with fever after chemotherapy, according to an embodiment of the present invention.

[0050] Figure 3 This is a schematic diagram of the therapeutic effect prediction model according to an embodiment of the present invention;

[0051] Figure 4 This is a schematic diagram of the federated learning framework structure according to an embodiment of the present invention;

[0052] Figure 5 This is a schematic diagram of a module of a machine learning-based chemotherapy efficacy evaluation system for hematologic malignancies with fever after chemotherapy, according to an embodiment of the present invention. Detailed Implementation

[0053] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0055] Figure 1 This is a flowchart illustrating a machine learning-based method for evaluating the efficacy of chemotherapy for hematologic malignancies with fever after chemotherapy, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the system framework of a machine learning-based method for evaluating the efficacy of chemotherapy for hematologic malignancies with fever after chemotherapy, according to an embodiment of the present invention. Figure 1 and Figure 2 As shown, a machine learning-based method 100 for evaluating the efficacy of chemotherapy for hematologic malignancies with fever includes the following steps:

[0056] Step S1: Collect and preprocess multi-center medical data. The data includes structured clinical data and multimodal unstructured data. The multimodal unstructured data includes medical text data, medical image data, and their derived feature data.

[0057] Step S1.1 Medical data collection.

[0058] The types of data collected include:

[0059] (1) Structured time-series clinical data:

[0060] Structured time-series clinical data:

[0061] Laboratory time-series indicators: sequence data of changes over time such as neutrophil count, total white blood cell count, and CRP value;

[0062] Treatment process record: event log including chemotherapy drug dosage, timing of antibiotic use, duration of fever, etc.;

[0063] Physiological status scores: Regularly assessed data such as ECOG score and KPS score.

[0064] (2) Multimodal unstructured data:

[0065] Medical text data: progress notes, nursing assessment texts, and imaging diagnostic reports in electronic medical records; for example, a radiology report describes "multiple ground-glass opacities in both lungs visible on chest CT, suggesting lung infection"; a progress note states "today's highest body temperature was 39.0℃, and vancomycin was added to cover Gram-positive bacteria."

[0066] Medical imaging data: chest CT scans, peripheral blood smear images, PET-CT scans; for example, DICOM files of chest CT images of patients.

[0067] Derived feature data: semantic features extracted from text / images (such as descriptions of signs of lung infection and quantitative features of cell morphology).

[0068] Step S1.2 Medical data preprocessing.

[0069] Preprocessing of the collected multi-center medical data lays the foundation for the subsequent construction of a multimodal learning model. Preprocessing specifically includes:

[0070] Data cleaning: Remove data from patients with incomplete information, whose condition worsened or developed serious complications during the study, or who withdrew midway; remove duplicate records; and correct obvious errors.

[0071] Data standardization: Continuous data (such as age, blood index values, etc.) are normalized and transformed into values ​​in the range [0,1]; categorical data (such as gender, pathology type, drug type, etc.) are one-hot encoded or labeled and transformed into numerical data that can be recognized by machine learning models.

[0072] Missing value handling: For a small number of missing data, methods such as mean imputation, median imputation, or mode imputation are used; for data with a large proportion of missing values, it is decided whether to remove the variable or sample based on the actual situation.

[0073] Multimodal data alignment and annotation: Named entity recognition and field extraction are performed on unstructured data (such as medical record text and imaging examination reports); structured and unstructured data are aligned using a unified patient ID to construct an integrated multimodal sample representation. Table 1 shows the field mapping relationship extracted during the alignment of unstructured text data, and Table 2 shows the feature extraction method of unstructured original images (CT chest X-rays, etc.); samples are manually labeled to indicate whether infection has occurred, and labels are constructed for supervised model training.

[0074] Table 1

[0075]

[0076] Table 2

[0077]

[0078] Among them, named entity recognition (NER) is performed on the text of imaging examination reports to extract key diagnostic conclusions (such as infection site, lesion size, and imaging score) and map them into structured features: infection site: multi-label coding is used (e.g., lung = 1, blood flow = 2, urinary tract = 3); lesion severity: quantified according to description (no abnormality = 0, mild = 1, moderate = 2, severe = 3); specific indicators: such as pleural effusion volume (mm) and nodule diameter (cm) are directly used as continuous node features.

[0079] Optionally, in one embodiment, text processing involves processing the image report text using a pre-trained NER model. For example, entity extraction: Infection site: Lung -> mapped to multi-label feature [1, 0, 0] (assuming encoding as [Lung=1, Blood flow=0, Urinary tract=0]). Entity extraction: Highest body temperature: 39.0 -> mapped to continuous numerical feature 39.0 (which will then be normalized). Image processing involves inputting the chest CT image of patient A into a pre-trained ResNet50 model, which outputs a 1024-dimensional feature vector, such as [0.12, -0.45, 0.88, ...], as a quantized representation of the image.

[0080] Step S2: Construct an efficacy prediction model based on a graph neural network and train the model using graph federated learning based on central moment constraints. This includes: constructing a patient graph as the original graph G=(V,E) using clinical variables in the preprocessed data as nodes V and edges E based on the statistical or clinical correlation between clinical variables; applying structural and feature perturbations to the original graph based on a graph augmentation mechanism to generate an augmented graph; processing the original graph and the augmented graph separately using a dual-channel graph convolutional neural network with shared weights to extract higher-order representation vectors; constructing an augmentation-aware loss function based on the higher-order representation vectors of the original graph and the augmented graph; and training the model using a federated learning framework with multi-center data and introducing central moment constraints to align feature distributions.

[0081] Step S2.1 Construct explicit modeling to enhance differences.

[0082] To improve the accuracy and generalization ability of the efficacy prediction model for patients with neutropenia and fever after chemotherapy, this paper introduces a graph augmentation modeling framework. By explicitly perturbing the medical graph structure, the model is forced to recognize differential semantic features, thereby learning a more robust and discriminative graph representation.

[0083] Step S2.1.1 Patient image modeling and construction of enhanced image.

[0084] Each patient is considered as a graph structure G=(V,E), where nodes V represent various key variables (such as chemotherapy regimens, laboratory indicators, demographic characteristics, etc.); edges E reflect the statistical or clinical correlations between clinical variables (such as multivariate synergistic changes, combined drug use, etc.).

[0085] For example: Define nodes (V): Each node represents a clinical variable. The graph for patient A might contain the following nodes: Node 1: Normalized neutrophil count (eigenvalue: 0.15); Node 2: Normalized CRP value (eigenvalue: 0.72); Node 3: Encoded antibiotic type (eigenvector: [0,1,0,...]); Node 4: “Lung infection” label extracted from text (eigenvalue: 1); Node 5: 1024-dimensional feature vector extracted from CT images; Node 6: Normalized highest body temperature (eigenvalue: 0.89); ... (other relevant variables).

[0086] The construction of edge E is based on the association between clinical variables and is determined through one or a combination of the following two methods:

[0087] (1) Construction based on statistical correlation: Calculate the statistical correlation coefficients (such as Pearson correlation coefficient and Spearman rank correlation coefficient) between all node variables in the training set. For example, the weight of edge E can be calculated by calculating the absolute value of the Pearson correlation coefficient between node variables on the training set, and setting a threshold > 0.3 to establish a connection. An edge is established between the corresponding nodes, and the weight of the edge is the correlation coefficient value. The edge weight threshold can be determined by grid search. The threshold [0.2, 0.4, 0.6] is tested on the validation set, and the threshold that makes the model AUC the highest is selected (0.3 is finally selected).

[0088] (2) Construction based on prior clinical knowledge: Based on predefined key variable relationships according to medical guidelines, clinical pathways, or expert experience, connections are forcibly established between their corresponding nodes. These relationship pairs include, but are not limited to: chemotherapy drug node and neutrophil count node; fever duration node and CRP (C-reactive protein) value node; antibiotic type node and infection site node; imaging feature node and specific infection sign node. The weight of the edges constructed based on prior knowledge can be set to a fixed value of 1, or different weight values ​​can be assigned according to clinical importance.

[0089] For example, define edges (E): establish connections between nodes based on rules determined during the training phase (e.g., Pearson correlation coefficient > 0.3 or clinical prior knowledge). For instance, the neutrophil count node and the CRP value node are highly negatively correlated in the training set, so a high-weighted edge will be established between them. Based on clinical knowledge, an edge will be forcibly established between the antibiotic type node and the infection site node. CT image feature nodes may be connected to the lung infection node and the CRP value node. Ultimately, a graph structure G_patientA = (V, E) uniquely representing the current state of patient A is constructed.

[0090] Based on the original graph G, an enhanced graph G is constructed using the enhancement mechanism. + The graph augmentation mechanism includes at least one of the following perturbation methods: randomly deleting nodes or edges; reconnecting edge relationships; masking node feature values; linearly mixing features from different patients; and sampling feature subgraphs. Specifically, perturbation differences are introduced in the following ways:

[0091] 1. Structural disturbance:

[0092] Node / edge deletion: Randomly remove some feature nodes or relationships between variables (due to missing simulation information or discrepancies in clinical records).

[0093] Edge reconnection: Replace some edges to simulate potential causal structure misjudgments or data measurement errors.

[0094] 2. Characteristic perturbations:

[0095] Feature masking: Randomly masking certain feature values ​​(such as neutrophil count, duration of fever, etc.) to improve the model's robustness to missing information.

[0096] Feature fusion: Linearly fuses the graph features of two patients to construct a new sample representation. This is a technical regularization method; the generated enhanced samples do not directly correspond to real patients, but are only used to improve the model's robustness to feature perturbations and its generalization ability.

[0097] 3. Mixed disturbances:

[0098] Subgraph sampling: Select a subset of patient features to form an enhanced sample, simulating diagnostic analysis from different perspectives.

[0099] Global topology adjustment: Simulates changes in the coupling structure between variables caused by differences between different hospitals or clinical pathways.

[0100] This step S2.1.1 involves constructing multiple "similar but not identical" augmented graphs G. + This guides the model to learn robust patient graph representations, reducing the risk of overfitting to local noise or feature bias.

[0101] like Figure 3 As shown, the structure of the efficacy prediction model includes an input layer, a dual-channel graph convolutional neural network, an orthogonal constraint module, a global pooling layer, a fully connected classification layer, and a Softmax output layer.

[0102] The input layer receives patient graph structure data; the dual-channel graph convolutional neural network contains two graph convolutional layers with shared weights, processing the original graph and the enhanced graph respectively; the global pooling layer generates graph-level representation vectors through attention-weighted averaging; the orthogonal constraint module applies orthogonal constraints to the weights of the graph convolutional layers; the fully connected classification layer maps the graph representation vectors to the class space; and the Softmax output layer generates the efficacy probability distribution.

[0103] Specifically, the dual-channel graph convolutional neural network includes: a first graph convolutional layer that reduces the input dimension to 64 dimensions and uses the ReLU activation function; and a second graph convolutional layer that reduces the 64-dimensional features to a 32-dimensional embedding vector.

[0104] Step S2.1.2 Enhanced perception modeling.

[0105] The core objective of augmented perception modeling is to quantify the representational differences between the original and augmented images and to explicitly model these differences as the training objective. In this way, the model not only learns how to classify data but also how to identify potential risk changes caused by data perturbations, thereby improving its ability to perceive subtle changes in patient condition.

[0106] The construction of the augmented perception loss function includes: predicting the difference between the representation vectors of the original image and the augmented image using a multilayer perceptron to obtain the predicted inter-graph distance; calculating the true inter-graph distance between the original image and the augmented image using a difference metric that fuses node features and graph structure; using the error between the predicted difference and the true graph distance as a loss constraint and constraining the consistency between the predicted inter-graph distance and the true graph distance using the L2 norm. Specifically, the augmented perception loss function is defined. Used to measure the difference between the original image G and the enhanced image G. + The error between the predicted difference between the representation vectors and the actual graph distance:

[0107]

[0108] in, and : Higher-order representation vectors of the original and enhanced graphs (output by the graph neural network encoder). : Multilayer perceptual network (MLP) for estimating graph differences. : Represents the square of the L2 norm, used to calculate the magnitude of the prediction error. : A measure of the combined structural and feature differences between two graphs. : Enhanced perception loss function.

[0109] In clinical settings, this mechanism can effectively characterize subtle but potentially critical changes such as a slight decrease in neutrophil count or a slight increase in infection markers, providing the model with a stronger risk perception capability.

[0110] Step S2.1.3 Graph difference measurement.

[0111] Furthermore, in order to calculate the true distance between the original image and the augmented image... This invention employs a method that fuses node features and graph structure differences to calculate the true inter-graph distance between the original and enhanced graphs. Specifically, it uses the Fused-Gromov-Wasserstein distance, which fuses differences in node feature distribution and graph topology, to calculate the distance. This includes: calculating the Wasserstein distance of node feature distribution to measure the difference in node features (reflecting the distribution shift of clinical variable values); calculating the Gromov-Wasserstein distance of graph topology to measure the difference in graph topology (reflecting changes in the correlation between variables); dynamically balancing the weights of the differences in feature distribution and graph topology using an adjustable parameter α∈[0,1] to obtain the final graph distance; and using an optimal transport algorithm to calculate the coupling matrix for node matching. This coupling matrix aligns node features and graph structure by minimizing the overall difference between the original and enhanced graphs. The specific process described above is expressed by the following formula:

[0112]

[0113] : Gromov-Wasserstein distance, which integrates features and structure, to measure the combined difference between graph G and augmented graph G⁺ at the structural and nodal feature levels. α: Hyperparameter balancing feature and structural differences. A larger α indicates a greater focus on differences in node features, enhancing sensitivity to feature changes (such as fluctuations in neutrophil counts); a smaller α indicates a greater focus on structural differences, enhancing sensitivity to structural changes (such as reconstruction of causal relationships between variables). π: The coupling matrix for node matching, belonging to the set... , where E and Q are the node distributions on graphs G and G⁺. Wasserstein distance, calculated based on π to compare the feature matrix X of the original graph nodes with the feature matrix X of the enhanced graph nodes. ⁺ The distribution differences. Gromov-Wasserstein distance, calculated based on π, for the original graph's structural connectivity matrix ε and the enhanced graph's structural connectivity matrix. The differences reflect changes in the graph's topology. X: The node feature matrix of the original graph G. + Enhanced graph G + The node feature matrix of the original graph G. ε: The structural connectivity matrix of the original graph G (such as the adjacency matrix or edge weight matrix). + : The structural connectivity matrix of the enhanced graph G⁺.

[0114] FGWD is approximated using entropy regularization:

[0115]

[0116] Sinkhorn and Gromov-Sinkhorn are implemented using the Python library POT (Python OptimalTransport). The hyperparameter α ranges from [0.3, 0.7], and its optimal value is determined by grid search on the validation set to balance the importance of feature differences and structural differences for specific clinical tasks.

[0117] This difference measure integrates the Wasserstein distance of node features with the Gromov-Wasserstein distance of graph structure, enabling the measurement of global similarity between graphs while simultaneously considering changes in node attributes and graph topology. The Fused-Gromov-Wasserstein distance dynamically aligns the structure and features of the original and enhanced graphs. It is a differentiable, low-complexity graph difference measure with clear geometric meaning, outperforming traditional graph edit distances and particularly suitable for modeling complex clinical graph data.

[0118] This invention improves the robustness and discriminative power of the model against noise, missing data, and subtle clinical changes through graph augmentation and augmentation perception loss; it effectively aligns multi-center feature distributions through federated learning based on central moment constraints, improving the generalization performance of the joint model while protecting privacy; and graph structure modeling captures the complex relationships between clinical variables.

[0119] Step S2.1.4 Consistency Regularization.

[0120] Consistency regularization aims to constrain the model's predictions for the same patient sample under different enhancement perspectives, preventing the model from producing drastically different judgments due to weak perturbations, thereby enhancing prediction robustness. Let the model... The classification prediction is For enhanced maps The prediction is The consistency loss function is defined as follows:

[0121]

[0122] Original image Category The predicted probability. Enhanced graph Category The predicted probability. Total number of categories (i.e., "effective treatment" and "ineffective treatment"). Consistency loss function.

[0123] The consistency loss function forces the model to output similar classification probabilities for the original and enhanced images of the same patient. This mechanism helps improve the model's stability under conditions of "different assessment stages of the same patient" or "slight noise in the data," ensuring the consistency and clinical interpretability of its predictions.

[0124] Step S2.2 Graph federation learning based on central moment constraints.

[0125] To address the clinical management needs of hematologic malignancy patients experiencing granulocytopenia and fever after chemotherapy, multiple medical institutions have accumulated heterogeneous electronic medical records, laboratory test data, biochemical indicators, and records of fever symptom development. Significant differences exist among centers in examination procedures, patient populations, and recording methods, resulting in a non-independent and identically distributed characteristic distribution of these features, which affects the generalization performance of multi-center collaborative modeling.

[0126] To address the aforementioned issues, step S2.2 employs a graph federated learning method based on central moment constraints. This method utilizes graph neural networks and incorporates multi-center medical sample structure information (such as patient-examination-symptom relationship graphs). It leverages central moment statistics to constrain the feature distributions of each participant to converge, effectively improving the stability and diagnostic accuracy of joint learning from multi-source medical data. Figure 4 The diagram shown is a schematic diagram of the federated learning framework structure according to an embodiment of the present invention.

[0127] Step S2.2.1 Local orthogonal graph model.

[0128] Within each medical center, patient clinical data (such as temperature upon admission, granulocyte count, C-reactive protein, etc.) are represented as node features in a graph structure. Assuming that these features locally follow a multivariate Gaussian distribution, orthogonal constraint modules are constructed, and the hidden layer feature matrix is ​​constrained using the Frobenius norm. Satisfying the orthogonality condition: Orthogonal constraint relaxation factor =1e-4, forced to satisfy by gradient descent. Based on this constraint, an orthogonal constraint loss function is generated, making the learned feature dimensions independent of each other.

[0129] The orthogonal constraint loss function is defined as follows:

[0130]

[0131] in, The Frobenius norm ensures that features are orthogonally distributed in the latent space, thus mitigating feature redundancy and information bias caused by differences in medical institutions. k: Index of neural network layers, ranging from 2 to l−1. l: Total number of network layers. : The characteristic orthogonality loss function of patients in the i-th hospital at the k-th layer. : The weight matrix of patients in the i-th hospital at the k-th layer. : The transpose of . : The identity matrix, used as an orthogonal reference.

[0132] From a graph structure perspective, normalized graph convolution operations are used to learn patient feature representations:

[0133]

[0134] : Represents the normalized adjacency matrix, adjacency matrix After adding self-connection and normalization of the longitude matrix H, the clinical graph relationship between different patients or medical examination items is represented. : The feature matrix of the i-th hospital patient. : The feature map representation of the i-th hospital patient at the k-th layer. From the input layer to the first The sequence of weight matrices for each layer. : The k-th power of the normalized adjacency matrix, representing the information propagation in a multi-order graph structure.

[0135] Step S2.2.2 Construct central moment difference constraints.

[0136] Furthermore, the central moment difference loss is calculated using the federated learning framework through the following steps: each medical center trains a local graph neural network model to extract patient graph embedding vectors; the central moments of the patient embedding vectors are calculated, including the mean, variance, and higher-order moments; the server receives the encrypted central moment parameters from each center, and generates a global central moment through weighted aggregation; the local feature distribution is aligned with the global distribution by constraining the central moment difference distance.

[0137] The server aggregates the patient feature means and higher-order statistics uploaded by participating hospitals to construct a global distribution. Since only feature means and higher-order central moment statistics (such as mean, variance, etc.) are transmitted instead of the raw data, patient privacy can be effectively protected.

[0138] Further calculation of the central moment difference distance includes:

[0139] Calculate the first-order moments (mean distance) and weighted distance (higher-order central moments) between the local and global centers. The higher-order central moments are second-order and above, and the distance is a weighted distance including variance, skewness, and kurtosis. Specifically, the global mean is calculated as follows:

[0140]

[0141] R: Mean of global patient characteristics; R: Total number of hospitals participating in federated learning; : Number of patient samples in the i-th hospital; : The mean of patient characteristics in the i-th hospital (i.e., the expected value of the characteristics);

[0142] Calculation of the weighted distance between local and global higher-order central moments: For the j-th order central moment, its local and global expressions are as follows:

[0143]

[0144]

[0145] : The j-th order central moment of the i-th hospital; : The j-th order central moment of the global system; : The feature vector of the m-th patient in the i-th hospital.

[0146] Central moments are used to characterize the higher-order structure of the distribution of medical features. For example, j=2 indicates that the variance reflects the variability of vital signs among patients, j=3 indicates that the skewness reflects the distribution tilt of abnormal samples, and j=4 can capture peak abnormal trends (such as the rate of deterioration of granulomatosis). The unified alignment of higher-order moments is the key to improving the generalization ability of medical data.

[0147] The mean distance and weighted distance are normalized and scaled according to the range of eigenvalues. Specifically, the central moment difference distance is defined as:

[0148]

[0149] Where [ɑ,b] represents the upper and lower bounds of the feature value range, which are determined based on the minimum value ɑ=min(X) and the maximum value b=max(X) of the global features in the training set; : Loss due to difference in central moments.

[0150] The central moment difference distance jointly considers first-order (mean) and higher-order statistical differences, serving as a metric for measuring the consistency between a hospital's characteristics and global features. Minimizing this distance guides the local model to learn to approximate the globally independent and identically distributed distribution, improving the consistency of prognostic analyses for granulocytopenia and fever. It also addresses the problems of feature distribution bias (such as equipment differences) and label distribution imbalance.

[0151] Optionally, in some embodiments, to further ensure the security of patient privacy data, encrypted privacy protection technologies (such as differential privacy DP, secure multi-party computation SMPC, and homomorphic encryption HE) are used to encrypt the data. After calculating the central moment parameters locally, Laplace noise is added using the ε-differential privacy (DP) mechanism: the privacy budget ε = 0.1 is set, and the sensitivity Δf is calculated as (ba) based on the feature range [a, b]; for the first moment... Add noise: For higher order moments Add noise: ; Encrypted parameters Uploaded to the server for aggregation.

[0152] Step S2.2.3 Overall optimization goal.

[0153] Furthermore, the efficacy prediction model also includes a multi-objective optimizer for multi-objective joint optimization training. The loss function for multi-objective joint optimization includes: cross-entropy loss for the efficacy classification task, enhancement perception loss, consistency loss, orthogonality constraint loss, and central moment difference loss, and each loss term is dynamically weighted and balanced through hyperparameters. Finally, the local optimization objective function for each hospital i is defined as follows:

[0154]

[0155] : The total loss function for the i-th hospital patient, used to jointly optimize task performance and distribution alignment. Cross-entropy loss is used for downstream tasks such as predicting patient fever status and classifying granulation degree. : The corresponding real label (such as fever status, grain deficiency classification); Orthogonalization loss maintains feature orthogonality, mitigates distribution shift, and reduces interference between local features. Central moment difference loss improves the alignment effect of multi-center features. The enhanced perception loss function in step S2.1.2 improves the model's ability to perceive subtle changes in the patient's state. The consistency loss function in step S2.1.4 enhances the ability to distinguish key clinical states. Hyperparameters ρ and φ are used to balance the weights of each loss term. Specifically, the loss weights are determined through Bayesian optimization: on the validation set, search for ρ∈ [1e-4,1e-3], φ∈[1,100], and select the combination that maximizes the weighted F1 score (ρ=0.0005, φ=10).

[0156] The training of the efficacy prediction model of this invention is a multi-objective collaborative optimization process. At the start of training, each medical center first calculates the forward propagation on its local data, obtaining graph embeddings and various losses. Subsequently, the central moments are calculated locally and encrypted before being uploaded to the server for federated aggregation. The server returns the global central moments to calculate the central moment difference loss. Finally, all loss terms are weighted and summed to obtain the total loss, which is used to perform backpropagation and local model parameter updates. Through multiple iterations, the model gradually gains robustness to data perturbations, sensitivity to small changes, and consistency across center distributions while optimizing the main classification task.

[0157] Step S2.3 Output the efficacy prediction results.

[0158] The output of the efficacy prediction is:

[0159]

[0160] : The feature representation of the i-th hospital patient in the l-th (last) layer of the graph neural network; The weight matrix of the fully connected output layer is used to map high-dimensional features to the category space (e.g., treatment effective / ineffective). : The bias term of the output layer; : Normalization function, which converts the output into a two-class probability distribution form; : The probability distribution of the predicted treatment effect for the i-th hospital patient (output is a vector of length 2, representing the probability of belonging to "treatment effective" and "treatment ineffective").

[0161] The final classification decision is:

[0162]

[0163] : The final efficacy classification label for the i-th hospital patient, with a value of 0 or 1; argmax: Take the category label corresponding to the highest probability (e.g., 0 indicates ineffective treatment, 1 indicates effective treatment).

[0164] In step S2, graph augmentation techniques are applied to the patient variable relationship graph, and an enhanced perceptual loss is designed ( Explicitly utilizing the FGWD distance to guide the model in learning perturbation invariance and sensitivity. In the medical federated learning scenario, a central moment difference constraint is introduced. Aligning high-order feature distributions addresses the challenge of non-IID (Non-Independent and Identically Distributed) medical data, surpassing simple mean alignment (such as first-order moment). It utilizes a dual-channel GNN to process the original and augmented images, along with orthogonal constraints. Consistency loss ), and cross-entropy loss for specific efficacy prediction tasks, through a carefully designed joint optimization objective ( The above-mentioned technologies work together to solve the technical challenge of accurately and robustly assessing the efficacy of granulocyte-depleted disease with fever using multi-center, multimodal, and heterogeneous medical data while protecting privacy.

[0165] Step S3: Construct a patient graph from the new patient data, and output the probability distribution of effective and ineffective efficacy using the trained efficacy prediction model. Take the category with the highest probability as the efficacy evaluation result.

[0166] In step S3, the previously constructed graph structure G_patientA representing the current state of patient A is input into the trained efficacy prediction model. The model's dual-channel GNN processes the graph representation, ultimately generating a probability distribution in the Softmax output layer. Specifically, the model returns a two-dimensional probability vector, for example: =[0.25,0.75], where, =0.25 indicates that the model predicts a 25% probability of "treatment ineffectiveness"; =0.75 means that the model predicts a 75% probability that the treatment will be effective.

[0167] According to the argmax rule, the category with the highest probability is taken as the final evaluation result, and the prediction result = argmax([0.25, 0.75]) = 1. The system outputs the clinical conclusion to the doctor: "Efficacy evaluation result: Treatment is effective (confidence 75%)".

[0168] Optionally, in some embodiments, to assist clinical decision-making, the system can provide an explanation: "The main basis for the model's 'effective' judgment is that the CT imaging features highly match the typical pattern of effective treatment, and the CRP value shows a downward trend. Although the neutrophil count is still low, the model believes that its recovery speed is within expectations." This enhances doctors' trust in the AI's suggestions, transforming it from a "black box" decision into an understandable auxiliary reference. The model can output a quantitative, objective, and interpretable efficacy assessment result, providing strong support for clinicians to make accurate decisions.

[0169] According to the above embodiments of the present invention, a machine learning-based method for evaluating the efficacy of chemotherapy for hematologic malignancies with neutropenia and fever after chemotherapy is proposed. This method integrates multimodal medical data and utilizes graph neural networks and a federated learning framework to achieve cross-center, highly robust efficacy prediction. Through key technologies such as multimodal data fusion and alignment, representation learning based on graph augmentation and dual-channel GNNs, enhanced perceptual loss (FGWD) incorporating Wasserstein distance, and central moment constraints (CMD) under the federated learning framework to achieve distribution alignment and privacy protection, the system can effectively capture the complex relationships between clinical variables, enhance robustness to noise and missing data, improve the model's generalization ability and prediction accuracy in cross-center scenarios, and simultaneously ensure patient privacy and security.

[0170] Figure 5 This is a schematic diagram of a module of a machine learning-based chemotherapy efficacy evaluation system for hematologic malignancies with fever following chemotherapy, according to an embodiment of the present invention; Figure 5 As shown, a machine learning-based system 200 for evaluating the efficacy of chemotherapy for hematologic malignancies with fever includes:

[0171] Data acquisition and preprocessing module 210: Acquires and preprocesses multi-center medical data, including structured clinical data and unstructured text and image data; the multimodal unstructured data includes medical text data, medical image data and their derived feature data;

[0172] Model building and training module 220: Constructs an efficacy prediction model based on a graph neural network and trains the model using graph federated learning based on central moment constraints. This includes: constructing a patient graph G=(V,E) as the original graph using clinical variables in the preprocessed data as nodes V and statistical or clinical correlations between variables as edges E; applying structural and feature perturbations to the original graph based on a graph augmentation mechanism to generate an augmented graph; processing the original graph and augmented graph separately using a dual-channel graph convolutional neural network with shared weights to extract higher-order representation vectors; constructing an augmentation-aware loss function based on the higher-order representation vectors of the original and augmented graphs; and training the model using a federated learning framework with multi-center data and introducing central moment constraints to align feature distributions.

[0173] The efficacy prediction output module 230 constructs a patient graph from the new patient data, and outputs the efficacy as a probability distribution of effective and ineffective efficacy through the trained efficacy prediction model, taking the category with the highest probability as the efficacy evaluation result.

[0174] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the described module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0175] The acquisition, storage, and application of user personal information involved in the technical solution of this application all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0176] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0177] It should also be noted that, in the embodiments of this application, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0178] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined in the embodiments of this application may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown in this application, but is to be accorded the widest scope consistent with the principles and novel features disclosed in the embodiments of this application.

Claims

1. A machine learning-based method for evaluating the efficacy of chemotherapy for hematologic malignancies accompanied by fever, characterized in that, Includes the following steps: Step S1: Collect and preprocess multi-center medical data, including structured clinical data and multimodal unstructured data; the multimodal unstructured data includes medical text data, medical image data and their derived feature data; Step S2: Construct an efficacy prediction model based on a graph neural network and train the model using graph federated learning based on central moment constraints. This includes: constructing a patient graph as the original graph G=(V,E) using clinical variables in the preprocessed data as nodes V and edges E based on the statistical or clinical correlation between clinical variables; applying structural and feature perturbations to the original graph based on a graph augmentation mechanism to generate an augmented graph; processing the original graph and the augmented graph separately using a dual-channel graph convolutional neural network with shared weights to extract higher-order representation vectors; constructing an augmentation-aware loss function based on the higher-order representation vectors of the original graph and the augmented graph; and training the model using a federated learning framework with multi-center data and introducing central moment constraints to align feature distributions. The construction of the enhancement-aware loss function based on the higher-order representation vectors of the original graph and the enhancement graph includes: The distance between the predicted images is obtained by predicting the difference between the representation vectors of the original image and the augmented image using a multilayer perceptron. The difference metric method that combines node features and graph structure is used to calculate the true inter-graph distance between the original graph and the augmented graph. The error between the predicted difference and the distance to the true graph is used as the loss constraint, and the L2 norm is used to constrain the consistency between the predicted distance and the distance between the true graph. The method for calculating the true inter-graph distance between the original graph and the enhanced graph using the difference metric of fused node features and graph structure is the Fused-Gromov-Wasserstein distance calculation, which uses the difference in fused node feature distribution and graph topology. This includes: Calculate the Wasserstein distance between the node feature distributions to measure the differences in node features, i.e., their distributions; Calculate the Gromov-Wasserstein distance of the graph topology to measure the differences in graph topology; The final graph distance is obtained by dynamically balancing the weights of the differences in feature distribution and graph topology using an adjustable parameter α. The optimal transmission algorithm is used to calculate the coupling matrix for node matching. This coupling matrix aligns node features and graph structure by minimizing the comprehensive difference between the original graph and the enhanced graph. Step S3: Construct a patient graph from the new patient data, and output the probability distribution of effective and ineffective efficacy using the trained efficacy prediction model. Take the category with the highest probability as the efficacy evaluation result.

2. The method for evaluating the efficacy of chemotherapy for hematologic malignancies with fever after chemotherapy based on machine learning, as described in claim 1, is characterized in that... The preprocessing described in step S1 includes: data cleaning and standardization, missing value handling, and multimodal data alignment and labeling.

3. The method for evaluating the efficacy of chemotherapy for hematologic malignancies with fever after chemotherapy based on machine learning, as described in claim 2, is characterized in that... The graph enhancement mechanism described in step S2 includes at least one of the following perturbation methods: randomly deleting nodes or edges; reconnecting edge relationships; masking node feature values; linearly mixing features from different patients; and sampling feature subgraphs.

4. The method for evaluating the efficacy of chemotherapy for hematologic malignancies with fever after chemotherapy based on machine learning, as described in claim 3, is characterized in that... The structure of the efficacy prediction model includes an input layer, a dual-channel graph convolutional neural network, an orthogonal constraint module, a global pooling layer, a fully connected classification layer, and a Softmax output layer. The input layer receives patient graph structure data; the dual-channel graph convolutional neural network includes two graph convolutional layers with shared weights, processing the original graph and the enhanced graph respectively; the global pooling layer generates graph-level representation vectors through attention-weighted averaging; the orthogonal constraint module applies orthogonal constraints to the weights of the graph convolutional layers; the fully connected classification layer maps the graph representation vectors to the category space; and the Softmax output layer generates the efficacy probability distribution. Specifically, the dual-channel graph convolutional neural network includes: a first graph convolutional layer that reduces the input dimension to 64 dimensions and uses the ReLU activation function; and a second graph convolutional layer that reduces the 64-dimensional features to a 32-dimensional embedding vector. The orthogonal constraint module constrains the weight matrix to satisfy the orthogonality condition through the Frobenius norm, and generates an orthogonal constraint loss function based on the orthogonality condition, so that the learned feature dimensions are independent of each other.

5. The method for evaluating the efficacy of chemotherapy for hematologic malignancies with fever after chemotherapy based on machine learning, as described in claim 1, is characterized in that... The efficacy prediction model also includes a multi-objective optimizer for multi-objective joint optimization training of the efficacy prediction model; the loss function of the multi-objective joint optimization includes: cross-entropy loss function for efficacy classification task, enhancement perception loss function, consistency loss function, orthogonal constraint loss function, and central moment difference loss function, and each loss term is dynamically weighted and balanced through hyperparameters; The consistency loss function forces the model to output similar classification probabilities for the original image and the enhanced image of the same patient.

6. The method for evaluating the efficacy of chemotherapy for hematologic malignancies with fever after chemotherapy based on machine learning, as described in claim 5, is characterized in that... The central moment difference loss is calculated using the federated learning framework through the following steps: Each medical center trains a local graph neural network model to extract patient graph embedding vectors. Calculate the central moments of the patient embedding vector, including the mean, variance, and higher-order moments; The server receives encrypted central moment parameters from each center and generates a global central moment through weighted aggregation. Align local feature distributions with global distributions by constraining the difference in central moments; The federated learning method only transmits the feature mean and higher-order central moment statistics between each center and the server.

7. The method for evaluating the efficacy of chemotherapy for hematologic malignancies with fever after chemotherapy based on machine learning, as described in claim 6, is characterized in that... The calculation of the central moment difference distance includes: Calculate the first moment (mean distance) and the weighted distance of higher-order central moments between the local and global centers; where higher-order central moments are second-order and above central moments, including the weighted distance of variance, skewness and kurtosis. The mean distance and weighted distance are normalized and scaled according to the range of eigenvalues.

8. The method for evaluating the efficacy of chemotherapy for hematologic malignancies with fever after chemotherapy based on machine learning, as described in claim 6, is characterized in that... The method further includes: Differential privacy technology is used to achieve privacy protection by adding Laplace noise to the center moment parameters uploaded to each center.

Images