Machine learning-based method for evaluating curative effect of hypogranulocyte accompanying fever after hematologic tumor chemotherapy

Abstract

The invention relates to the technical field of disease curative effect evaluation, and discloses a method for evaluating the curative effect of granulocyte deficiency accompanying fever after hematologic tumor chemotherapy based on machine learning, and the method comprises the steps: S1, collecting and preprocessing multi-center medical data, the data comprising structured clinical data and multi-modal unstructured data; s2, a curative effect prediction model is constructed based on a graph neural network, and a central moment constraint-based graph federal learning training model is adopted, wherein a patient graph is constructed as an original graph and an enhanced graph is generated; respectively processing the original image and the enhanced image through a two-channel image convolutional neural network, and extracting a high-order representation vector; constructing an enhanced perception loss function based on the high-order representation vectors of the original image and the enhanced image; combining a federated learning framework with multi-center data to train a model; and S3, predicting a curative effect evaluation result through the trained curative effect prediction model. The accuracy, the stability and the clinical interpretability of curative effect evaluation are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of disease efficacy evaluation, and in particular to a method for evaluating the efficacy of neutropenia with fever after chemotherapy for hematological tumors based on machine learning. Background Art

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

[0003] Currently, the evaluation of FN efficacy relies primarily on clinical experience, changes in blood counts, inflammatory indicators such as CRP, and assessment of infection sites. A lack of systematic, generalizable modeling approaches makes it difficult to extract stable patterns from multi-center, multimodal, and multivariate data. Furthermore, the occurrence and progression of FN are complex, influenced by multiple factors such as individual patient differences, chemotherapy regimens, underlying diseases, and supportive care plans. This creates a high-dimensional medical data landscape characterized by complex structure, strong heterogeneity, and significant noise perturbations.

[0004] In recent years, machine learning methods such as graph neural networks (GNNs), federated learning, and distributed regression have been widely used in the medical field, providing new directions for modeling complex FN states and predicting efficacy. However, existing methods still have the following technical bottlenecks: (1) Graph representation instability: When faced with data perturbations (such as missing features, different hospital record formats), graph models are prone to structural sensitivity, which affects the discrimination effect; (2) Heterogeneous distribution alignment is difficult: Data distributions vary greatly between multiple centers, and direct joint modeling has generalization risks; (3) Lack of deep modeling capabilities for graph structural changes: Traditional graph distances are difficult to simultaneously characterize node features and structural information; (4) Prediction targets are difficult to align 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 an efficacy evaluation method that has stronger difference perception ability, unified structure-feature measurement ability, cross-center generalization ability, and both robustness and interpretability. Summary of the Invention

[0006] The present invention proposes a machine learning-based method for evaluating the efficacy of neutropenia with fever after chemotherapy for hematological malignancies, aiming to address the problems of strong graph modeling sensitivity, poor model generalization ability, and insensitivity to structural changes in the existing technology, thereby improving the accuracy, stability and clinical interpretability of efficacy evaluation.

[0007] To achieve the above objectives, the following technical solutions are adopted:

[0008] A method for evaluating the efficacy of neutropenia with fever after chemotherapy for hematological malignancies based on machine learning, comprising the following steps:

[0009] Step S1: Collecting and preprocessing multi-center medical data, wherein the data includes structured clinical data and multimodal unstructured data; the multimodal unstructured data includes medical text data, medical imaging data and derived feature data thereof;

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

[0011] Step S3: Construct the new patient data into a patient graph, and output the probability distribution of effective and ineffective efficacy through the trained efficacy prediction model, and take the category with the maximum probability as the efficacy evaluation result.

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

[0013] Furthermore, the graph enhancement mechanism 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 of 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 is used to receive patient graph structure data; the dual-channel graph convolutional neural network includes two weight-sharing graph convolution layers, which process the original image and the enhanced image respectively; the global pooling layer is used to generate a graph-level representation vector through attention-weighted averaging; the orthogonality constraint module is used to impose orthogonality constraints on the weights of the graph convolution layer; the fully connected classification layer is used to map the graph representation vector to the category space; the Softmax output layer is used to generate a probability distribution of therapeutic efficacy;

[0016] The dual-channel graph convolutional neural network specifically includes: the first graph convolution layer: reducing the input dimension to 64 dimensions and using the ReLU activation function; the second graph convolution layer: reducing the 64-dimensional features to a 32-dimensional embedding vector;

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

[0018] Furthermore, an enhanced perception loss function is constructed based on the high-order representation vectors of the original image and the enhanced image, including:

[0019] The difference between the representation vectors of the original image and the enhanced image is predicted through a multi-layer perception network to obtain the distance between the predicted images;

[0020] The true inter-image distance between the original image and the enhanced image is calculated using the difference measurement method that integrates node features and graph structure;

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

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

[0023] Calculate the Wasserstein distance of node feature distribution to measure the difference in node features, i.e. distribution;

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

[0025] The final graph distance is obtained by dynamically balancing the weights of the difference in feature distribution and the difference in graph topology through the adjustable parameter α;

[0026] An optimal transmission algorithm is used to calculate the coupling matrix of node matching, which aligns node features and graph structure by minimizing the comprehensive difference between the original image and the enhanced image.

[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, enhanced perception loss, consistency loss, orthogonal constraint loss, and central moment difference loss for the efficacy classification task, 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 enhanced image of the same patient.

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

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

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

[0032] The server receives the encrypted central moment parameters of each center and performs weighted aggregation to generate the global central moment;

[0033] Constrain the local feature distribution to align with the global distribution through the central moment difference distance;

[0034] The federated learning only transmits feature means and high-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-order moment of the local and global data, i.e., the mean distance, and the weighted distance of the higher-order central moment; the higher-order central moment is the second-order and higher-order central moment, including the weighted distance of variance, skewness, and kurtosis;

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

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

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

[0040] 1. This paper proposes an explicit modeling-enhanced difference mechanism and constructs an enhanced perception modeling strategy: By performing both structural and feature perturbations on the patient graph (e.g., edge rewiring, feature masking, and subgraph sampling), this paper constructs multiple enhanced graphs, guiding the model to identify differences between graphs. Furthermore, an enhanced perception loss function is introduced, enabling the model to not only learn classification but also acquire "difference sensitivity," enabling more effective identification of subtle but critical physiological changes after chemotherapy.

[0041] 2. The present invention constructs a graph difference measurement mechanism that fuses the Wasserstein distance of node features and the Gromov-Wasserstein distance of graph structure: The present invention designs the Fused-Gromov-Wasserstein graph distance, which comprehensively considers the changes in graph node attributes and graph topology structure, stably measures the global differences between graphs under perturbations, and provides a precise geometric basis for enhancing perceptual modeling and consistency regularization.

[0042] 3. The present invention proposes a graph federated learning method based on central moment constraints: To address the problem of distribution differences among multi-center medical data, the present invention introduces a central moment (such as mean, variance, skewness, and kurtosis) constraint mechanism to achieve alignment of the embedded distributions of patient graphs from different institutions, thereby improving the model migration and generalization capabilities under the federated learning framework.

[0043] 4. The present 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, the present invention constructs a local orthogonal graph structure and designs a feature orthogonality loss term, which enables the feature dimensions learned by the model to be independent of each other and complementary in information, thereby enhancing the ability to identify key variables in the diagnostic process.

[0044] 5. The present invention proposes a central moment difference distance indicator to describe the degree of multi-center distribution deviation: The present invention constructs a central moment difference distance to quantify the deviation between the embedded distribution of each center and the global distribution, thereby providing a distribution alignment adjustment signal for federated optimization and realizing "homogeneous modeling of heterogeneous data."

[0045] 6. The present invention proposes an overall loss function that integrates multiple factors such as enhanced perception, consistency regularization, and center alignment: the present invention combines and optimizes 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 stability and discrimination ability.

[0046] It should be understood that the contents described in the summary of the invention are not intended to limit the key or important features of the embodiments of the present invention, nor are they intended to limit the scope of the present invention. Other features of the present invention will become readily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The above and other features, advantages and aspects of the embodiments of the present invention will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. The accompanying drawings are provided for a better understanding of the present invention and do not constitute a limitation of the present invention. In the accompanying drawings, the same or similar reference numerals represent the same or similar elements, among which:

[0048] Figure 1This is a flow chart of a method for evaluating the efficacy of granulocytosis with fever after chemotherapy for hematological malignancies based on machine learning according to an embodiment of the present invention;

[0049] Figure 2 This is a system framework diagram of a method for evaluating the efficacy of granulocytopenia with fever after chemotherapy for hematological malignancies based on machine learning according to an embodiment of the present invention;

[0050] Figure 3 Schematic diagram of the structure of the therapeutic effect prediction model according to an embodiment of the present invention;

[0051] Figure 4 Schematic diagram of the federated learning framework structure of an embodiment of the present invention;

[0052] Figure 5 This is a module diagram of a machine learning-based system for evaluating the efficacy of granulocytosis with fever after chemotherapy for hematological malignancies according to an embodiment of the present invention. DETAILED DESCRIPTION

[0053] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0054] In this document, the term "and / or" simply describes a relationship between related objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A exists alone, A and B exist simultaneously, or B exists alone. Furthermore, the character " / " in this document generally indicates that the related objects are in an "or" relationship.

[0055] Figure 1 This is a flow chart of a method for evaluating the efficacy of granulocytosis with fever after chemotherapy for hematological malignancies based on machine learning according to an embodiment of the present invention; Figure 2 This is a system framework diagram of a method for evaluating the efficacy of neutropenia with fever after chemotherapy for hematological malignancies based on machine learning according to an embodiment of the present invention. Figure 1 and Figure 2 As shown, a method 100 for evaluating the efficacy of granulocytopenia with fever after chemotherapy for hematological malignancies based on machine learning comprises the following steps:

[0056] Step S1: Collect multi-center medical data and pre-process them, wherein the data includes structured clinical data and multimodal unstructured data, and the multimodal unstructured data includes medical text data, medical imaging 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: neutrophil count, total white blood cell count, CRP value and other time-varying sequence data;

[0062] Treatment process records: event logs such as chemotherapy drug dosage, timing of antibiotic use, and duration of fever;

[0063] Physiological status score: ECOG score, KPS score and other regular assessment data.

[0064] (2) Multimodal unstructured data:

[0065] Medical text data: medical records, nursing assessment texts, and imaging diagnostic reports in electronic medical records; for example, the imaging department report describes "chest CT shows multiple ground-glass shadows in both lungs, suggesting lung infection"; the medical record records "today's highest temperature is 39.0°C, and vancomycin is added to cover Gram-positive bacteria."

[0066] Medical imaging data: CT chest films, peripheral blood smear images, PET-CT scans; for example, a DICOM file of a patient's chest CT image.

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

[0068] Step S1.2: Medical data preprocessing.

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

[0070] Data cleaning: Eliminate data from patients with incomplete information, worsening of condition or severe complications during the study, or withdrawal midway, remove duplicate records, and correct obviously erroneous data.

[0071] Data standardization: Normalize continuous data (such as age and blood index values) and convert them into values ​​in the range [0,1]. Perform one-hot encoding or label encoding on categorical data (such as gender, pathological type, and medication type) to convert them into numerical data that can be recognized by machine learning models.

[0072] Missing value processing: For a small amount of missing data, methods such as mean filling, median filling or mode filling are used for processing; for data with a large proportion of missing data, whether to eliminate the variable or sample is decided based on the actual situation.

[0073] Multimodal data alignment and annotation: Perform named entity recognition and field extraction on unstructured data (such as medical record text and imaging examination reports); use a unified patient ID to align structured and unstructured data to build an integrated multimodal sample representation. Table 1 shows the extracted field mapping relationship during the unstructured text data alignment process, and Table 2 shows the feature extraction method for unstructured original images (such as CT chest films); manually annotate samples to determine whether they are infected, and construct labels 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 the imaging examination report to extract key diagnostic conclusions (such as infection site, lesion size, and imaging score) and map them into structured features: infection site: multi-label encoding is used (such as lung = 1, blood flow = 2, urinary tract = 3); lesion severity: quantified according to the 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 using a pre-trained NER model to process the text of the imaging report. For example, extracting the entity "Infection Site: Lung" maps it to the multi-label feature [1, 0, 0] (assuming the encoding is [Lung = 1, Blood Flow = 0, Urinary Tract = 0]). Extracting the entity "Maximum Body Temperature: 39.0" maps it to the continuous numerical feature 39.0 (which is then normalized). Image processing involves inputting patient A's chest CT image into a pre-trained ResNet50 model. The model outputs a 1024-dimensional feature vector, such as [0.12, -0.45, 0.88, ...], as a quantitative representation of the image.

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

[0081] Step S2.1 constructs the explicit modeling enhancement difference.

[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 more robust and discriminative graph representations.

[0083] Step S2.1.1: Modeling the patient graph and constructing the enhanced graph.

[0084] Each patient is regarded as a graph structure G=(V,E), where the node V represents various key variables (such as chemotherapy regimens, laboratory indicators, demographic characteristics, etc.); the edge E reflects the statistical correlation or clinical correlation between clinical variables (such as multivariate synergistic changes, combined drug use, etc.).

[0085] For example, define a node (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 maximum temperature (eigenvalue: 0.89); ... (other relevant variables).

[0086] The construction of edge E is based on the association between clinical variables and is determined by 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 determined by calculating the absolute value of the Pearson correlation coefficient between node variables in the training set and setting a threshold > 0.3 to establish a connection. An edge is established between the corresponding nodes, and the edge weight is the correlation coefficient value. The edge weight threshold can be determined by grid search. The thresholds [0.2, 0.4, 0.6] are tested on the validation set, and the threshold that gives the highest model AUC is selected (finally 0.3).

[0088] (2) Construction based on clinical prior knowledge: Based on the key variable relationship pairs predefined in medical guidelines, clinical pathways or expert experience, connections are forced to be established between their corresponding nodes. These relationship pairs include but are not limited to: chemotherapy drug nodes and neutrophil count nodes; fever duration nodes and CRP (C-reactive protein) value nodes; antibiotic type nodes and infection site nodes; imaging feature nodes and specific infection sign nodes. The weight of the edge 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, defining edges (E): Nodes are connected based on rules determined during training (e.g., Pearson correlation coefficient > 0.3 or clinical prior knowledge). For example, the neutrophil count node and the CRP value node are highly negatively correlated in the training set, so a high-weight edge is established between them. Based on clinical knowledge, an edge is mandatory between the antibiotic type node and the infection site node. A CT image feature node might be connected to a lung infection node and a CRP value node. Ultimately, a graph structure G_patientA = (V, E) is constructed that uniquely represents the current status of patient A.

[0090] Based on the original graph G, an enhanced graph G is constructed based on the enhancement mechanism + The graph enhancement mechanism includes at least one of the following perturbations: random deletion of nodes or edges; rewiring of edge relationships; masking of node feature values; linear mixing of features from different patients; sampling of 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 (simulating missing information or differences in clinical records).

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

[0094] 2. Feature perturbation:

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

[0096] Feature blending: Linearly fuses the features of two patient images to construct a new sample representation. This is a technical regularization method. The generated augmented samples do not directly correspond to real patients and are only used to improve the model's robustness to feature perturbations and improve model generalization.

[0097] 3. Mixed disturbance:

[0098] Sub-image sampling: A subset of patient features is selected to form an enhanced sample to simulate diagnostic analysis from different perspectives.

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

[0100] This step S2.1.1 constructs multiple “similar but not identical” enhanced graphs G + , guiding the model to learn a robust patient graph representation, reducing the risk of overfitting to local noise or feature shift.

[0101] like Figure 3 As shown in the figure, 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 is used to receive patient graph structure data; the dual-channel graph convolutional neural network contains two weight-shared graph convolution layers, which process the original image and the enhanced image respectively; the global pooling layer is used to generate graph-level representation vectors through attention-weighted averaging; the orthogonality constraint module is used to impose orthogonality constraints on the weights of the graph convolution layer; the fully connected classification layer is used to map the graph representation vector to the category space; and the Softmax output layer is used to generate the probability distribution of therapeutic efficacy.

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

[0104] Step S2.1.2 enhances perceptual modeling.

[0105] The core goal of enhanced perception modeling is to quantify the difference in representation between the original and enhanced images and explicitly model this difference as a training target. In this way, the model not only learns how to classify but also learns how to identify potential risk changes caused by data perturbations, thereby improving its ability to perceive subtle changes in patient status.

[0106] The construction of the enhanced perception loss function includes: predicting the difference between the representation vectors of the original image and the enhanced image through a multi-layer perception network to obtain the predicted inter-image distance; using the difference measurement method that integrates node features and graph structure to calculate the true inter-image distance between the original image and the enhanced image; using the error between the predicted difference and the true image distance as the loss constraint and using the L2 norm to constrain the consistency between the predicted inter-image distance and the true image distance. Specifically, define the enhanced perception loss function , used to measure the original image G and the enhanced image G + The error between the predicted difference between the characterization vectors and the actual graph distance:

[0107]

[0108] in, and : The high-order representation vector of the original image and the enhanced image (output by the graph neural network encoder). : Multi-layer Perceptron Network (MLP) for Estimating Graph Difference. : Represents the square of the L2 norm, which is used to calculate the size of the prediction error. : A comprehensive structural-feature difference measure between two images. : Enhanced perceptual loss function.

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

[0110] Step S2.1.3 Graph difference measurement.

[0111] Furthermore, in order to calculate the true distance between the original image and the enhanced image The present invention uses a method to measure the difference between node features and graph structure to calculate the true inter-graph distance between the original graph and the enhanced graph. Specifically, the Fused-Gromov-Wasserstein distance calculation is based on the difference in node feature distribution and graph topology structure. The method includes: calculating the Wasserstein distance of node feature distribution to measure the difference in node features, i.e., distribution (reflecting the distribution shift of clinical variable values); calculating the Gromov-Wasserstein distance of graph topology to measure the difference in graph topology (reflecting the change in the correlation between variables); dynamically balancing the weights of the difference in feature distribution and the difference in graph topology through an adjustable parameter α∈[0,1] to obtain the final graph distance; and using the optimal transmission algorithm to calculate the node matching coupling matrix. The coupling matrix aligns node features and graph structure by minimizing the comprehensive difference between the original graph and the enhanced graph. The above process is specifically expressed by the following formula:

[0112]

[0113] : The Gromov-Wasserstein distance that integrates features and structures is used to measure the comprehensive difference between the graph G and the enhanced graph G⁺ in terms of structure and node features. α: A hyperparameter that balances feature differences and structural differences The larger α is, the more attention is paid to the differences in node features, which can enhance the sensitivity to feature changes (such as fluctuations in neutrophil counts); the smaller α is, the more attention is paid to structural differences, which can enhance the sensitivity to structural changes (such as reconstruction of causal relationships between variables). π: The coupling matrix of node matching, which belongs to the set , where E and Q are the node distributions on graphs G and G⁺. : Wasserstein distance, based on π, calculates the original graph node feature matrix X and the enhanced graph node feature matrix X ⁺ distribution differences. : Gromov-Wasserstein distance, based on π, calculates the original graph structure connection matrix ε and the enhanced graph structure connection matrix The difference reflects the change in the topological structure of the graph. X: The node feature matrix of the original graph G. + : Enhanced graph G + Node feature matrix of the original graph G. ε: Structural connection matrix of the original graph G (such as adjacency matrix or edge weight matrix). + : The structural connectivity matrix of the enhanced graph G⁺.

[0114] The entropy regularization is used to approximate the FGWD:

[0115]

[0116] Sinkhorn and Gromov-Sinkhorn were implemented using the Python library POT (Python Optimal Transport). The hyperparameter α ranged from 0.3 to 0.7, and its optimal value was determined through grid search on the validation set to balance the importance of feature and structural differences for specific clinical tasks.

[0117] This difference metric combines the Wasserstein distance of node features with the Gromov-Wasserstein distance of graph structure. It can measure the 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 augmented graphs. It is a differentiable, low-complexity, and geometrically meaningful graph difference metric that outperforms traditional graph edit distances and is particularly suitable for modeling complex clinical graph data.

[0118] This invention improves the model's robustness and discriminability to noise, missingness, and subtle clinical changes through graph augmentation and enhanced perceptual loss; 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 to maintain stable prediction results for the same patient sample under different enhanced perspectives, avoiding the model from producing completely different judgment results due to weak perturbations, thereby enhancing prediction robustness. The classification prediction is , for the enhanced graph The prediction is , then the consistency loss function is defined as follows:

[0121]

[0122] : Original image Belong to category The predicted probability of . : Enhanced graph Belong to category The predicted probability of . : Total number of categories (i.e., “treatment effective” and “treatment ineffective”). : 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 across different evaluation stages of the same patient or under the influence of mild data noise, ensuring that its predictions are consistent and clinically interpretable.

[0124] Step S2.2: Graph federated 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 tests, biochemical parameters, and records of fever symptom progression. Significant differences among centers in examination processes, patient populations, and recording methods result in non-independent and identically distributed feature distributions, impacting the generalization performance of multi-center joint modeling.

[0126] To solve the above problem, this step S2.2 adopts a graph federated learning method based on central moment constraint, which is based on graph neural network, combined with multi-center medical sample structure information (such as patient-examination-symptom relationship graph), and uses central moment statistics to constrain the feature distribution of each participant to be consistent, effectively improving the stability and diagnostic accuracy of multi-source medical data joint learning. Figure 4 The figure shows a schematic diagram of the federated learning framework structure of an embodiment of the present invention.

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

[0128] In each medical center, the clinical data of patients (such as body temperature, granulocyte count, C-reactive protein, etc.) are represented as node features of the graph structure. Assuming that the features are locally subject to multivariate Gaussian distribution, an orthogonal constraint module is constructed to constrain the hidden layer feature matrix by the Frobenius norm. Satisfy orthogonality conditions: Orthogonality constraint relaxation factor =1e-4, forced to satisfy by gradient descent , and based on this constraint, an orthogonal constraint loss function is generated to make the learned feature dimensions independent of each other.

[0129] The orthogonality constraint loss function formula is defined as:

[0130]

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

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

[0133]

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

[0135] Step S2.2.2 constructs the central moment difference constraint.

[0136] Furthermore, the central moment difference loss is calculated through the following steps of the federated learning framework: each medical center locally trains the graph neural network model to extract the patient graph embedding vector; calculates the central moment of the patient embedding vector, including the mean, variance and high-order moment; the server receives the encrypted central moment parameters of each center and performs weighted aggregation to generate the global central moment; and uses the central moment difference distance to constrain the local feature distribution to align with the global distribution.

[0137] The server aggregates the mean and higher-order statistics of patient features uploaded by participating hospitals to construct a global distribution. Because only the mean and higher-order central moment statistics (such as mean and variance) are transmitted, rather than the raw data, patient privacy is effectively protected.

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

[0139] Calculate the first-order moment between the local and global, i.e., the mean distance and the weighted distance of the higher-order central moment. The higher-order central moment is the second-order and higher-order central moment, including the weighted distance of variance, skewness, and kurtosis. Specifically, the calculation of the global mean is:

[0140]

[0141] : global patient feature mean; R: total number of hospitals participating in federated learning; : The number of patient samples in the i-th hospital; : The mean of the patient characteristics of the i-th hospital (i.e., the expectation of the characteristics);

[0142] Calculation of local and global high-order central moment weighted distance: For the j-th order central moment, its local and global expressions are:

[0143]

[0144]

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

[0146] Central moments are used to characterize the higher-order structure of the distribution of medical features. For example, when j=2, the variance reflects the inter-patient fluctuation of physical features; when j=3, the skewness reflects the distribution of abnormal samples; and when j=4, it can capture abnormal spikes (such as the rate of worsening granulocytosis). The unified alignment of higher-order moments is key to improving the generalization of medical data.

[0147] The mean distance and weighted distance are normalized and scaled according to the eigenvalue range. 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 by the minimum value ɑ=min(X) and the maximum value b=max(X) of the global feature of the training set; : Central moment difference loss.

[0150] The central moment difference distance combines first-order (mean) and higher-order statistical differences as a metric for measuring the consistency of a hospital's features with those of the global model. Minimizing this distance guides local models toward a global independent and identically distributed distribution, improving the consistency of prognostic analysis for neutropenia and fever. This addresses issues such as feature distribution shift (e.g., device differences) and label imbalance.

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

[0152] Step S2.2.3 Overall optimization objective.

[0153] 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 for the efficacy classification task, enhanced perception loss, consistency loss, orthogonal constraint loss, and central moment difference loss, and each loss term is dynamically weighted and balanced through hyperparameters. Ultimately, the local optimization objective function for each hospital i is defined as follows:

[0154]

[0155] : The total loss function for patients in the i-th hospital, which is used to jointly optimize task performance and distribution alignment. : Cross entropy loss, used for downstream tasks such as patient fever status prediction and granulocytosis degree classification. : The corresponding true label (such as fever status, granulocytosis classification); : Orthogonalization loss, maintaining feature orthogonality, alleviating distribution shift, and reducing interference between local features. : Center moment difference loss improves the multi-center feature alignment effect. : The enhanced perception loss function in step S2.1.2 improves the model's ability to perceive subtle changes in the patient's condition. The consistency loss function in step S2.1.4 enhances the ability to discriminate key clinical states. The hyperparameters ρ and φ are used to balance the weights of each loss term. Specifically, the loss weights are determined using Bayesian optimization: ρ ∈ [1e-4, 1e-3], φ ∈ [1, 100] are searched on the validation set, and the combination that yields the highest weighted F1 score is selected (ρ = 0.0005, φ = 10).

[0156] The training of the efficacy prediction model of the present invention is a multi-objective collaborative optimization process. At the beginning of training, each medical center first calculates the forward propagation on the local data to obtain the graph embedding and various losses. Subsequently, the central moment is calculated locally and encrypted and uploaded to the server for federated aggregation. The server returns the global central moment to calculate the central moment difference loss. Finally, the weighted sum of all loss terms is used to obtain the total loss, which is used to perform backpropagation and local model parameter updates. Through multiple iterations, the model gradually obtains robustness to data perturbations, sensitivity to small changes, and consistency in cross-center distribution while optimizing the main classification task.

[0157] Step S2.3: Output of efficacy prediction results.

[0158] The efficacy prediction output is:

[0159]

[0160] : Feature representation of the patient in the i-th hospital in the l-th layer (the last layer) of the graph neural network; : The weight matrix of the fully connected output layer, which is used to map high-dimensional features to category space (such as treatment effectiveness / ineffectiveness); : bias term of the output layer; : Normalization function, converting the output into the probability distribution form of two categories; : The predicted probability distribution of the therapeutic effect of patients in the i-th hospital (the output is a vector of length 2, indicating the probability of "effective treatment" and "ineffective treatment").

[0161] The final classification decision is:

[0162]

[0163] : The final efficacy classification label of the patients in the i-th hospital, with a value of 0 or 1; argmax: the category label corresponding to the maximum probability (such as 0 for ineffective treatment and 1 for effective treatment).

[0164] In step S2, the graph enhancement technology is applied to the patient variable relationship graph and the enhanced perceptual loss ( ) explicitly uses FGWD distance to guide the model to learn perturbation invariance and sensitivity. In the medical federated learning scenario, the central moment difference constraint is introduced ( ) Align high-order feature distributions to solve the problem of non-IID (Non-Independent and Identically Distributed) medical data, going beyond simple mean alignment (such as first-order moment). Use dual-channel GNN to process original and enhanced images, orthogonal constraints ), consistency loss ( ), and the cross entropy loss for specific efficacy prediction tasks, through a carefully designed joint optimization objective ( The synergistic effect of the above technologies solves the technical challenge of accurately and robustly evaluating the efficacy of granulocytopenia with fever using multi-center, multimodal, and heterogeneous medical data while protecting privacy.

[0165] Step S3: Construct the new patient data into a patient graph, and output the probability distribution of effective and ineffective efficacy through the trained efficacy prediction model, and take the category with the maximum probability as the efficacy evaluation result.

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

[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 level 75%)."

[0168] Optionally, in some embodiments, to assist clinical decision-making, the system can provide an explanation: "The model's judgment of 'effective' is primarily based on the following: the CT imaging features are highly consistent with 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 rate is within expectations." This enhances doctors' trust in AI recommendations, transforming them from "black box" decisions to understandable auxiliary references. The model can output a quantitative, objective, and explainable efficacy evaluation result, which can provide clinicians with strong support for accurate decision-making.

[0169] According to the above-mentioned embodiment of the present invention, a machine learning-based method for evaluating the efficacy of post-chemotherapy granulocytopenia with fever in hematologic malignancies achieves cross-center, highly robust efficacy prediction by fusing multimodal medical data and utilizing graph neural networks and a federated learning framework. Through key technical approaches such as multimodal data fusion and alignment, representation learning based on graph enhancement and dual-channel GNNs, enhanced perceptual loss (FGWD) fused with Wasserstein distance, and central moment constraints (CMD) within a federated learning framework for distribution alignment and privacy protection, the system effectively captures the complex relationships between clinical variables, enhances robustness to noise and missing data, and improves the model's generalization and prediction accuracy in cross-center scenarios, while ensuring patient privacy.

[0170] Figure 5 is a module diagram of a system for evaluating the efficacy of granulocytosis with fever after chemotherapy for hematological malignancies based on machine learning according to an embodiment of the present invention; Figure 5 As shown, a machine learning-based system 200 for evaluating the efficacy of neutropenia with fever after chemotherapy for hematological malignancies includes:

[0171] Data acquisition and preprocessing module 210: collects 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 construction and training module 220: constructing an efficacy prediction model based on a graph neural network and adopting a graph federated learning training model based on central moment constraints, including: constructing a patient graph as the original graph G=(V,E) using clinical variables in the preprocessed data as nodes V and statistical or clinical correlations between variables as edges E; applying structural perturbations and feature perturbations to the original graph based on a graph enhancement mechanism to generate an enhanced graph; processing the original graph and the enhanced graph separately through a dual-channel graph convolutional neural network with shared weights to extract high-order representation vectors; constructing an enhanced perception loss function based on the high-order representation vectors of the original graph and the enhanced graph; training the model with multi-center data through a federated learning framework, and introducing central moment constraints to align feature distributions;

[0173] Therapeutic effect prediction output module 230: constructs new patient data into a patient graph, outputs the probability distribution of effective and ineffective therapeutic effects through the trained therapeutic effect prediction model, and takes the category with the maximum probability as the therapeutic effect evaluation result.

[0174] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the described module can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0175] In the technical solution of this application, the acquisition, storage and application of user personal information involved are in compliance 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. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the methods.

[0177] It should also be noted that, in the embodiments of the present application, relational terms such as first and second, etc. are merely used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprise", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the statement "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device comprising the elements.

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

Claims

1. A method for evaluating the efficacy of granulocytosis with fever after chemotherapy for hematological malignancies based on machine learning, characterized in that: The following steps are involved: Step S1: Collecting and preprocessing multi-center medical data, wherein the data includes structured clinical data and multimodal unstructured data; the multimodal unstructured data includes medical text data, medical imaging data and derived feature data thereof; Step S2: constructing an efficacy prediction model based on a graph neural network and adopting a graph federated learning training model based on central moment constraints, including: using clinical variables in the preprocessed data as nodes V, constructing edges E based on statistical correlation or clinical correlation between clinical variables, and constructing a patient graph as the original graph G=(V,E); applying structural perturbations and feature perturbations to the original graph based on a graph enhancement mechanism to generate an enhanced graph; processing the original graph and the enhanced graph separately through a two-channel graph convolutional neural network with shared weights to extract high-order representation vectors; constructing an enhanced perception loss function based on the high-order representation vectors of the original graph and the enhanced graph; training the model with multi-center data through a federated learning framework, and introducing central moment constraints to align feature distributions; Step S3: Construct the new patient data into a patient graph, and output the probability distribution of effective and ineffective efficacy through the trained efficacy prediction model, and take the category with the maximum probability as the efficacy evaluation result.

2. The method for evaluating the efficacy of neutropenia with fever after chemotherapy for hematological malignancies based on machine learning according to claim 1, characterized in that: The preprocessing in step S1 includes: data cleaning and standardization, missing value processing, and multimodal data alignment and labeling.

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

4. The method for evaluating the efficacy of neutropenia with fever after chemotherapy for hematological malignancies based on machine learning according to claim 3, 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 is used to receive patient graph structure data; the dual-channel graph convolutional neural network includes two weight-sharing graph convolution layers, which process the original image and the enhanced image respectively; the global pooling layer is used to generate a graph-level representation vector through attention-weighted averaging; the orthogonality constraint module is used to impose orthogonality constraints on the weights of the graph convolution layer; the fully connected classification layer is used to map the graph representation vector to the category space; the Softmax output layer is used to generate a probability distribution of therapeutic efficacy; The dual-channel graph convolutional neural network specifically includes: the first graph convolution layer: reducing the input dimension to 64 dimensions and using the ReLU activation function; the second graph convolution layer: reducing the 64-dimensional features to a 32-dimensional embedding vector; The orthogonality constraint module constrains the weight matrix to meet the orthogonality condition through the Frobenius norm, and generates an orthogonality 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 neutropenia with fever after chemotherapy for hematological malignancies based on machine learning according to claim 2, characterized in that: An enhanced perception loss function is constructed based on the high-order representation vectors of the original image and the enhanced image, including: The difference between the representation vectors of the original image and the enhanced image is predicted through a multi-layer perception network to obtain the distance between the predicted images; The true inter-image distance between the original image and the enhanced image is calculated using the difference measurement method that integrates node features and graph structure; The error between the predicted difference and the true image distance is used as the loss constraint and the L2 norm is used to constrain the consistency between the predicted image distance and the true image distance.

6. The method for evaluating the efficacy of neutropenia with fever after chemotherapy for hematological malignancies based on machine learning according to claim 5, characterized in that: The method for calculating the true inter-graph distance between the original graph and the enhanced graph by using the difference measurement method of fusing node features and graph structure is a Fused-Gromov-Wasserstein distance calculation using the difference in node feature distribution and graph topology structure, including: Calculate the Wasserstein distance of node feature distribution to measure the difference in node features, i.e. distribution; Calculate the Gromov-Wasserstein distance of the graph topology to measure the difference in graph topology; The final graph distance is obtained by dynamically balancing the weights of the difference in feature distribution and the difference in graph topology through the adjustable parameter α; An optimal transmission algorithm is used to calculate the coupling matrix of node matching, which aligns node features and graph structure by minimizing the comprehensive difference between the original image and the enhanced image.

7. The method for evaluating the efficacy of neutropenia with fever after chemotherapy for hematological malignancies based on machine learning according to claim 6, 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, enhanced perception loss, consistency loss, orthogonal constraint loss, and central moment difference loss for the efficacy classification task, 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 enhanced image of the same patient.

8. The method for evaluating the efficacy of neutropenia with fever after chemotherapy for hematological malignancies based on machine learning according to claim 7, characterized in that: The central moment difference loss is calculated through the following steps in the federated learning framework: Each medical center locally trains a graph neural network model to extract patient graph embedding vectors; Calculate the central moment of the patient embedding vector, including mean, variance and high-order moments; The server receives the encrypted central moment parameters of each center and performs weighted aggregation to generate the global central moment; Constrain the local feature distribution to align with the global distribution through the central moment difference distance; The federated learning only transmits feature means and high-order central moment statistics between each center and the server.

9. The method for evaluating the efficacy of neutropenia with fever after chemotherapy for hematological malignancies based on machine learning according to claim 8, characterized in that: The calculation of the central moment difference distance includes: Calculate the first-order moment of the local and global data, i.e., the mean distance, and the weighted distance of the higher-order central moment; the higher-order central moment is the second-order and higher-order central moment, including the weighted distance of variance, skewness, and kurtosis; The mean distance and the weighted distance are normalized and scaled according to the feature value range.

10. The method for evaluating the efficacy of granulocytosis with fever after chemotherapy for hematological malignancies based on machine learning according to claim 8, characterized in that: The method further comprises: Differential privacy technology is used to achieve privacy protection by adding Laplace noise to the central moment parameters uploaded by each center.

Images