Multi-center collaborative modeling and risk prediction method and system based on quality perception federated learning

Abstract

The invention discloses a multi-center collaborative modeling and risk prediction method and system based on quality perception federated learning. Firstly, data preprocessing and feature standardization are performed; then executing quality perception aggregation, integrating client updating by the server by adopting a quality perception self-adaptive federated aggregation algorithm, and generating a new global model; then federation training is carried out, the server issues the global model to the clients, the clients use local data to train, calculate parameter updating and carry out differential privacy protection and then upload, and iterative optimization is carried out until the model converges; and finally, implementing model deployment, deploying the final global model to each participating client, and providing real-time auxiliary diagnosis. According to the method, the problems of time sequence alignment and missing value intelligent coding of multi-center heterogeneous data can be solved, dynamic weight aggregation of data volume, data quality and data distribution similarity is comprehensively considered, deep interactive fusion of electrocardiograms, images, texts and structured data is realized, and resource-limited hospitals are supported to deploy lightweight models.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of medical artificial intelligence and computer technology, and in particular relates to a method and system that combines federated learning and deep learning technologies for collaborative data modeling, disease-assisted diagnosis, and prognostic risk prediction for multi-center patients. Background Technology

[0002] The hospital established its treatment center to provide rapid and accurate diagnostic and treatment support for patients—especially those suffering from acute high-risk diseases such as acute myocardial infarction. However, the system currently faces several key challenges: First, treatment-related data (such as electronic medical records, test results, electrocardiograms, and imaging data) from different medical institutions are fragmented, making effective cross-institutional sharing and integrated analysis difficult, resulting in significant "data silos." Second, due to the highly sensitive nature of the patient health information involved, direct data centralization and sharing are not only strictly restricted by laws and regulations such as the Personal Information Protection Law and the Data Security Law, but also bring significant privacy and compliance risks. Furthermore, AI-assisted diagnostic models built upon limited data from a single institution generally suffer from insufficient generalization ability, often resulting in decreased accuracy when applied to other medical scenarios. Simultaneously, the diagnostic and treatment procedures and risk assessments for high-risk diseases are not yet standardized across hospitals and even among different physicians, necessitating the introduction of intelligent tools capable of integrating multi-source data and possessing strong generalization capabilities to promote standardized and regulated treatment.

[0003] In the intelligent construction of hospitals, existing technological approaches all have significant limitations. Traditional centralized AI modeling (Shaikh AAS, et al. Weighted aggregation through probability basedranking: An optimized federated learning architecture to classify respiratory diseases.) relies on aggregating raw data from various centers, which is not only difficult to implement in practice but also faces serious compliance challenges due to its handling of sensitive medical data. If a training model based solely on data from a single center is adopted, the model is prone to overfitting due to a lack of data volume and diversity, severely limiting its performance and generalization ability (Cai YQ, et al. Pitfalls in developing machine learning models for predicting cardiovascular diseases: challenge and solutions.). As for traditional statistical methods such as logistic regression, although they have long been used in clinical risk assessment (such as the classic GRACE score), their model expressive power is limited, making it difficult to fully explore the characteristics of high-dimensional, nonlinear, multimodal medical data, leading to a bottleneck in prediction accuracy. In recent years, federated learning, as an emerging distributed machine learning paradigm, has provided a potential direction for solving the above-mentioned dilemmas. It has been systematically elaborated by Yang et al. in their foundational work (Yang, Qiang, et al. Federated machine learning: Concept and applications), and related patents such as "Heart Image Processing Method Based on Federated Deep Learning" (Patent No. CN114612408A) have explored it.However, most of these initial attempts remain at the level of a general framework, which leads to significant differences in the diagnosis and treatment pathways and risk assessments of emergency patients among different hospitals and doctors (Hicks, KA, et al. (2018). "2017 Cardiovascular and Stroke EndpointDefinitions for Clinical Trials." Circulation, 137(9), 961-972.). There is still a lack of in-depth integration for the specific clinical scenario of the treatment center, especially in how to systematically integrate multimodal data such as electronic medical records, electrocardiogram signals, and medical images, and determine their standards, and seamlessly embed them into the complete end-to-end business loop of "pre-hospital screening-risk prediction-assisted decision-making" for in-depth optimization, so as to form a truly usable, credible, and reliable standardized intelligent assisted diagnosis and treatment solution.

[0004] By reviewing existing technologies, the following technological gaps were identified: First, there is a lack of standardized protocols for federated learning data specific to the field of diagnosis and treatment, which cannot effectively handle issues such as temporal alignment between different centers and encoding of missing values; second, existing federated aggregation algorithms do not consider the heterogeneity of medical data quality, and simple data weighting can lead to low-quality data polluting the global model; third, there are insufficient methods for deep fusion of multimodal medical data, and existing methods are mostly simple splicing, failing to fully explore the interaction relationships between modalities. Summary of the Invention

[0005] The purpose of this invention is to solve the problems mentioned in the background art. This invention constructs a federated learning architecture that allows a central server to work collaboratively with multiple different central clients. Each client trains a local model using local data and only uploads model updates (such as weights and gradients) to the server. The server integrates these updates through a secure aggregation algorithm to generate a better global model and distributes it to each client. Through multiple iterations, the global model learns the data features of all participants while protecting data privacy.

[0006] To achieve the objectives of this invention, a multi-center collaborative modeling and risk prediction method based on quality-aware federated learning is disclosed, comprising the following steps:

[0007] Step 1, Data Preprocessing and Feature Standardization: Each client preprocesses its local multimodal data according to the different center federated data standardization protocols, including time alignment, missing pattern embedding, and cross-center normalization;

[0008] Step 2, Quality-Aware Aggregation: The server uses a quality-aware adaptive federated aggregation algorithm to integrate client updates and generate a new global model;

[0009] Step 3, Federated Training: The server distributes the global model to the clients, each client trains using local data, calculates and updates parameters, performs differential privacy protection, and then uploads the model; Iterative Optimization: Repeat steps S2-S3 until the model converges;

[0010] Step 4, Model Deployment: Deploy the final global model to all participating clients to provide real-time auxiliary diagnosis.

[0011] Furthermore, in step 1, time alignment: define the symptom onset time T0 as the global time reference; construct a unified timeline. , The sampling interval is k, and the time index is k; the original time series data is resampled to the standard time point by cubic spline interpolation.

[0012] Further, in step 1, missing pattern embedding: missing pattern embedding: constructing the missing mask matrix M and the missing type vector r; through a learnable embedding function Map missing types to the embedding space. This represents the code for the missing reason of the j-th feature in the i-th sample, with values ​​of 0, 1, 2, and 3, corresponding to "observed", "systematic not detected", "random detection failed", and "patient refused", respectively. This represents a learnable parameter matrix used to map one-hot encoding missing types to... 3D embedding vector; Will Convert to a 4-dimensional one-hot vector. The length of the embedding vector is represented by the length of the observation; the observation is concatenated with the missing embedding as the final feature representation.

[0013] Furthermore, in step 3, the quality-aware adaptive federated aggregation algorithm calculates the comprehensive contribution of client i as follows:

[0014]

[0015] in These represent hyperparameters, which respectively represent the amount of data. Data quality Similarity to data distribution The relative importance of satisfies α + β + γ = 1; This represents the overall contribution of client i in round t.

[0016] Furthermore, in step 3, the global model is a cross-modal attention fusion Transformer network, including:

[0017] Modality-specific encoders: handle structured data, electrocardiograms, laboratory tests, text, and images respectively;

[0018] Cross-modal attention layer: Calculates attention weights between modal pairs and fuses features;

[0019] Multi-head self-attention layer: Globally fuses information from all modalities;

[0020] Multi-task output layer: Simultaneously completes disease classification, risk scoring, mortality prediction, and readmission prediction.

[0021] Furthermore, the cross-modal attention layer computes attention weights for modality pairs (i,j):

[0022]

[0023] in, Represents the normalized activation function. The query matrix represents mode i. The bond matrix representing mode j. Indicates the scaling factor. The dimension of the attention head is represented, and the features are updated through residual connections: ,in These are the features of modality i after attention enhancement. This represents the feature representation of mode i after uniform dimensional projection.

[0024] Furthermore, the multi-task output layer adopts a dynamic weight adjustment strategy.

[0025]

[0026] in This represents the weight of task k in the t-th round of training. Let j be the loss of task k in round t-1, and j represent the task index. The weights are adaptively assigned based on the loss of each task in the previous round.

[0027] A multi-center collaborative modeling and risk prediction system based on quality-aware federated learning includes a federated learning server responsible for global model management, client scheduling, security aggregation, and anomaly detection; multiple hospital clients, including a local database, a data preprocessing module, a local training module, a differential privacy protection module, and a clinical application interface; and a secure communication layer, which is a data transmission channel based on Transport Layer Security (TLS) and homomorphic encryption.

[0028] Furthermore, the federated learning server includes: a global model management module for storing and version-controlling global model parameters; a quality assessment module for calculating the data quality and distribution similarity of each client; a secure aggregation engine for performing aggregation calculations of encrypted parameters; a privacy budget management module for tracking and allocating differential privacy budgets; and an anomaly detection module for identifying and removing abnormal client updates.

[0029] Furthermore, the client includes: a data preprocessing module for implementing the CPCFed protocol, performing time alignment, missing pattern embedding, and standardization; a feature extraction module for extracting deep features from electrocardiograms, images, and text; a local training module for training models using local data; and a clinical application module for providing real-time diagnosis, risk warnings, and treatment recommendations.

[0030] Compared with existing technologies, the significant advancements of this invention are: 1) It is the first to deeply integrate federated learning with multimodal data from medical centers, designing a heterogeneous data feature extraction and fusion scheme for different medical scenarios; 2) It proposes a federated learning optimization strategy for imbalanced medical data. Addressing the issues of large differences in data volume among different centers and a small number of positive samples (such as AMI), a weighted average algorithm (e.g., assigning different weights based on data volume or sample quality of each center) is introduced during federated aggregation, or a federated focus loss function is adopted to improve the model's ability to identify minority diseases; 3) It constructs a multi-task federated learning model integrating diagnosis and risk. A single model simultaneously completes disease classification and risk prediction, better meeting actual clinical needs and improving efficiency; 4) It designs an incentive mechanism that includes evaluation of model performance and data contribution. The server can evaluate the contribution of each client to the global model, providing a basis for future value allocation or academic collaboration; 5) Under the premise of absolute data privacy protection, it significantly improves the generalization ability and robustness of the AI ​​model, providing a feasible solution for the standardization and intelligent construction of medical centers.

[0031] To more clearly illustrate the functional characteristics and structural parameters of the present invention, further explanation is provided below in conjunction with the accompanying drawings and specific embodiments. Attached Figure Description

[0032] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0033] Figure 1 This is a schematic diagram of the overall architecture of the system of the present invention;

[0034] Figure 2 This is an architecture diagram of the MSECG Encoder in this invention;

[0035] Figure 3 This is a structural diagram of the deep learning model in this invention;

[0036] Figure 4 This is a flowchart of the federal learning training process. Detailed Implementation

[0037] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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.

[0038] This invention aims to construct a collaborative intelligent platform for multiple treatment centers using federated learning technology. Its core value lies in breaking down "data silos" between institutions without sharing raw data, aggregating the data value of various centers (hospitals, institutions) to train a more powerful and generalizable AI model. This method strictly adheres to data privacy regulations, ensuring that all participants' raw data remains locally, and collaborative modeling is achieved only through the exchange of encrypted intermediate parameters (such as model gradient updates). Based on this, we expect the joint model to achieve rapid and accurate risk identification for the causes of acute high-risk diseases (such as acute myocardial infarction, pulmonary embolism, aortic dissection, stroke, and COPD), and to effectively stratify and predict the short-term and long-term major adverse cardiovascular events in patients, thereby comprehensively improving the accuracy and efficiency of clinical diagnosis and treatment. Ultimately, the "global intelligent model" produced through federated learning will provide unified, high-level AI-assisted decision support for all participating treatment centers, which helps to reduce inter-hospital differences in treatment and powerfully promotes the standardization and homogenization of treatment processes across different centers.

[0039] like Figure 1 As shown, Figure 1 This is a schematic diagram of the system architecture of this invention. The federated learning server is responsible for global model initialization, client scheduling, secure aggregation of model updates, global model distribution, and version management; the multi-center clients are subsystems within each hospital, including local databases and local training modules; the secure communication channel ensures the encryption and security of data transmission between the server and clients. The overall workflow is summarized as follows:

[0040] 1. The server sends global model parameters to the selected clients;

[0041] 2. Each client uses local private data to train the model and calculate parameter updates;

[0042] 3. The client performs differential privacy noise addition and encryption on the update before uploading;

[0043] 4. The server uses a quality-aware aggregation algorithm to integrate and update, generating a new global model;

[0044] 5. Repeat the iteration until the model converges.

[0045] 6. The final model is deployed to various hospitals to provide real-time assisted diagnosis.

[0046] Specifically, a multi-center collaborative modeling and risk prediction method based on quality-aware federated learning includes the following steps:

[0047] Step S1: Multimodal data preprocessing and feature engineering standardization

[0048] 1. Data source definition

[0049] 1.1 Each client performs local multimodal data acquisition and preprocessing according to a unified standard.

[0050] The data and features include:

[0051] Modality 1: For structured clinical data , Number of features:

[0052] Demographic characteristics: age, sex, height, weight, BMI;

[0053] Vital signs: systolic blood pressure, diastolic blood pressure, heart rate, respiratory rate, body temperature, and blood oxygen saturation;

[0054] Past medical history: hypertension, diabetes, coronary heart disease, smoking history, family history;

[0055] Mode 2: For time-series ECG data T is the duration of time.

[0056] Standard 12-lead electrocardiogram, sampling frequency 500Hz;

[0057] The recording duration is denoted as T;

[0058] Raw waveform + Heart Rate Variability (HRV) indicator;

[0059] Modal 3: For laboratory test data , Number of features:

[0060] Cardiac markers: troponin I / T (cTnI / T), creatine kinase isoenzyme (CK-MB), myoglobin;

[0061] Inflammatory markers: white blood cell count, C-reactive protein, procalcitonin;

[0062] Coagulation function: D-dimer, fibrinogen;

[0063] Kidney function: creatinine, blood urea nitrogen, estimated glomerular filtration rate (eGFR);

[0064] Blood lipids: total cholesterol, low-density lipoprotein cholesterol (LDL-C), high-density lipoprotein cholesterol (HDL-C), and triglycerides;

[0065] Others: N-terminal pro-Brain natriuretic peptide (NT-proBNP), blood glucose;

[0066] Modality 4: Targeting textual chief complaint and medical history data ,in Representative string:

[0067] Chief complaint description (e.g., "squeezing pain behind the sternum for 3 hours");

[0068] Present medical history text;

[0069] Maximum length 512 characters;

[0070] Modal 5: For medical imaging

[0071] Coronary CTA: DICOM format, resolution 512×512×number of slices N slices;

[0072] Echocardiography: Video format, 30fps;

[0073] Chest X-ray: 2048*2048 resolution;

[0074] 2. Time alignment mechanism to address the issue of inconsistent data collection timestamps across different hospitals.

[0075] 2.1 Define the global time base

[0076] Let the patient's self-reported onset time be... Construct a unified timeline:

[0077]

[0078] in For standard sampling intervals, electrocardiogram vital signs Laboratory tests , , ;

[0079] 2.2 Timestamp Mapping

[0080] For the raw data of client i Calculate relative time: ;

[0081] Will Mapped to the most recent standard time point ;

[0082] 2.3 Time Series Data Resampling

[0083] For any standard time point If the observation for that moment is not present in the original data, use cubic spline interpolation:

[0084]

[0085] in For cubic B-spline basis functions:

[0086]

[0087] 2.4 Cross-center normalization

[0088] Federal statistical calculations use global mean and standard values ​​to reduce the impact of reference ranges across different hospitals and equipment.

[0089] Each client calculates its local statistical mean.

[0090] ,

[0091] in This represents the local difference of client i. This represents the local mean of client i.

[0092] Server aggregation:

[0093]

[0094] This represents the number of samples for client i.

[0095] After each client downloads the global statistics, they are standardized:

[0096]

[0097] in For standardized feature values, This represents the global mean. Represents global standard deviation

[0098] 3. Missing Pattern Embedding: Missing mechanisms for medical data are divided into three categories: Completely Random Missing (MCAR): Missing data is unrelated to any variable (e.g., equipment failure); Random Missing (MAR): Missing data is related to the observed variable (e.g., mild illness where advanced examinations were not performed); Non-Random Missing (MNAR): Missing data is related to unobserved variables (e.g., patient refuses examinations). Missing patterns are then constructed.

[0099] 3.1 Constructing the missing mask matrix

[0100] For the characteristic matrix Define a binary mask matrix M.

[0101]

[0102] in," "represents being observed," "Represents something missing."

[0103] 3.2 Missing type annotation

[0104] Introducing missing type vectors

[0105] coding meaning Example 0 Observed Normal troponin levels were detected. 1 Systemic undetected Primary hospitals lack NT-prOBNP testing capabilities 2 Random detection failed Hemolysis of blood sample renders test invalid 3 The patient refused The patient refused coronary angiography.

[0106] Annotation method:

[0107] Automatic identification via Hospital Information System (HIS) logs;

[0108] If the test request form exists but there is no result → Type 2;

[0109] If this test has never been requested and is routinely performed at this hospital → Type 1;

[0110] If the test has never been applied for and the hospital does not meet the requirements → Type 1;

[0111] 3.3 Learnable Missing Embeddings

[0112] Define embedded functions Map missing types to a dedimensional embedding space:

[0113]

[0114] in ;

[0115] 3.4 Feature Fusion

[0116] The final input model features are the concatenation of observations and missing embeddings:

[0117]

[0118] Where [.;.] denotes vector concatenation, and 0 represents vector concatenation. de It is a zero-dimensional vector.

[0119] 4. Electrocardiogram Deep Feature Extraction

[0120] 4.1 Preprocessing of the original waveform

[0121] 12-lead electrocardiogram

[0122] First, median filtering is used for baseline drift correction (window 200ms). Second, a 40Hz low-pass Butterworth filter is used to denoise the high-frequency noise. Finally, the Pan-Tompkins algorithm is used to detect the integrity of the R-wave.

[0123] 4.2 Hierarchical Feature Extraction Network

[0124] like Figure 2 As shown, the architecture of the "Federated ECG Feature Extractor" (MSECG Encoder) is as follows.

[0125] 1) Given input C is the number of channels, N is the number of sequences, and T is the sequence length. First, a set of convolutional kernels with a size of... The 1D causal convolutional layer extracts multi-scale temporal dependencies while mapping channels to 2C:

[0126]

[0127]

[0128]

[0129] 2) Using the first scale For example, we first divide it into two equal parts along the channel direction:

[0130]

[0131] E represents the segmentation, which is used to capture the numerical fluctuation information of the features; F represents the gating branch, which generates a 0-1 weight map and controls the "retention / suppression" of the features in the activation branch.

[0132] Then, the gated Tanh unit (GTU) is defined as follows:

[0133]

[0134] Here, tanh is the activation function of the output part, i.e., the hyperbolic tangent activation function, and sigmoid is the gate unit used to generate weights, i.e., the sigmoid activation function. The two parts are multiplied point by point. Obtain the output of GTU

[0135] 3) By stacking multiple GTUs, the receptive field in the time dimension can be expanded, thereby improving the model's ability to extract long-term temporal correlations in the data. The M-GTU is defined as follows:

[0136] in, Indicates pooling, , , These are the outputs of GTUs at three different scales. The Concat(.) operation concatenates the output features of the three GTUs at three different scales, producing a time-dimension... The characteristics of this are then transformed by a pooling layer with a window of size W, changing the time dimension to... ,Right now Here, by setting the window size, you can... It can also be directly mapped to the T-dimensional dimension through a linear layer. Therefore, Further transformed into .

[0137] 4) Finally, we... Adding residual connections and the ReLU activation function yields the output of the M-GTU. :

[0138]

[0139] Step S2: Cross-modal attention fusion of Transformer network

[0140] 1. Modal encoder implementation, such as... Figure 3 As shown

[0141] Encoder1: Structured data encoder: through two fully connected (FC) layers;

[0142] Encoder2: Electrocardiogram Encoder: The aforementioned "Multi-scale Electrocardiogram Feature Extractor" (MSECG Encoder);

[0143] Encoder3: Lab test encoder: via two fully connected layers;

[0144] Encoder4: Text encoder: using pre-trained BioBERT (BERT for the medical field);

[0145] Encoder5: Medical Image Encoder: Uses ResNet50 pre-trained on the ImageNet visual database.

[0146] Formula: For modal :

[0147]

[0148] This represents the original input data for the m-th mode; This represents the trainable parameters of the m-th modal encoder; This represents the feature representation output by the encoder, where d_m is the modality-specific feature dimension.

[0149] Then, a unified dimension projection layer is used to project all modalities onto a unified dimension d. h =256

[0150]

[0151] Represented as the projected weight matrix of mode m; This is the projection bias vector; To unify the feature dimensions after projection.

[0152] 2. Cross-modal attention mechanism

[0153] Pairwise Cross-Modal Attention:

[0154] For mode pairs Calculate mode m i For mode m j attention.

[0155] 1) Query, Key, Value Transformation:

[0156]

[0157]

[0158]

[0159] in , are the weight matrices for Q, K, and V, respectively. These are the bias vectors for Q, K, and V, respectively.

[0160] 2) Attention weight calculation:

[0161]

[0162] in, Represents the normalized activation function. =16 is the scaling factor to prevent gradient vanishing.

[0163] 3) Enhanced features:

[0164]

[0165] For the enhanced features, This represents the feature representation of mode i after uniform dimensional projection, and the residual connection preserves the original mode information.

[0166] 4) Traverse all modal pairs:

[0167] For M=5 modes, there are a total of 10 mode pairs, where each mode... Calculate its attention to all other modalities:

[0168]

[0169] 5) Multi-head self-attention global fusion:

[0170] Concatenate all enhanced features:

[0171]

[0172] This represents the concatenation of all modality enhancement features, with the following dimensions:

[0173] Multi-head attention (8 heads):

[0174]

[0175] in This represents the multi-head self-attention computation function. This represents the hidden features after splicing. This indicates the 1st to 8th points of attention. The linear transformation matrix of the multi-head attention output. , :

[0176]

[0177] in , (Dimensions of each head) This represents the attention calculation function, where Q is the query matrix, K is the key matrix, and V is the value matrix. For attention dimension.

[0178] 6) Feedforward Network (FFN):

[0179]

[0180] in, This indicates the hidden features after fusion. Indicates multi-head attention output

[0181]

[0182] ReLU is a rectified linear unit, and FFN is a feedforward network. This represents the first-layer weight matrix. This is the first layer bias vector. This is represented as the second-layer weight matrix. This represents the second-layer bias vector, with the following dimensions:

[0183] 7) Feature fusion and multi-task output:

[0184] Final fusion representation:

[0185]

[0186] in, Representation layer normalization, This indicates that the feature dimension after fusion is 256.

[0187] 3. Multi-task output layer

[0188] Task 1: Disease Classification (Examples of 5 categories: AMI, unstable angina, non-ACS chest pain, pulmonary embolism, aortic dissection):

[0189]

[0190] in This represents the final fusion representation. This represents the weight matrix for the classification task. This represents the bias vector for the classification task. This represents the classification prediction probability vector.

[0191] Task 2: Risk Score (0-100 points, continuous values):

[0192]

[0193] in This represents the risk scoring task weight vector. This indicates a bias in the risk scoring task. This is the sigmoid function, mapped to [0,1], then multiplied by 100.

[0194] Task 3: Risk of In-Hospital Mortality

[0195] in This represents the weight vector for the mortality prediction task. Death prediction task bias This represents the predicted probability of in-hospital mortality.

[0196] Task 4: Risk of readmission after 30 days:

[0197]

[0198] in This represents the weight vector for the readmission prediction task. This indicates a bias in the readmission prediction task. This represents the predicted probability of readmission after 30 days.

[0199] 4. Multi-task learning loss function

[0200] Adaptive task weight adjustment:

[0201]

[0202] in, For the total loss, L represents the dynamic weights of task k in the t-th round of training. k This represents the single-task loss for task k, and the weights of task k during the t-th training round:

[0203]

[0204] When a task has a high learning difficulty, its weight is automatically increased, making the model pay more attention to that task.

[0205] Loss functions for each task:

[0206] 1) Classification loss (cross-entropy + focal loss, to resolve class imbalance)

[0207] in, Represents the classification loss function. For the balancing weights of category c, To focus on parameters and reduce the severity of easily classified samples, Let i be the true label of the i-th sample. Let be the predicted probability that the i-th sample belongs to category c.

[0208] 2) Risk score loss:

[0209]

[0210] in This represents the smoothing L1 loss function. This represents the risk score predicted by the model for the i-th sample; Let represent the true risk score of the i-th sample.

[0211]

[0212] 3) Binary classification loss (binary cross-entropy):

[0213]

[0214]

[0215] in, Indicates predicted loss due to death. This indicates a 30-day readmission risk prediction. Indicates the true label, This represents the model's predicted probability for the i-th sample.

[0216] 4) Total Losses

[0217]

[0218] in, This indicates a 30-day readmission risk prediction.

[0219] Step S3: Federated Training Process

[0220] 1. Quality-Aware Adaptive Federated Aggregation Algorithm (QA-FedAvg)

[0221] Client-side local model quality assessment:

[0222] AUC-ROC calculation (Hanley-McNeil method):

[0223]

[0224] in, This represents the AUC value of client i. This represents the number of positive samples (AMI patients) in the local validation set of client i. This represents the number of negative samples (non-AMI patients) in the local validation set of client i; This represents the model's predicted probability (confidence level) for the m-th positive sample. This represents the model's predicted probability (confidence level) for the nth negative sample; 1[.] is an indicator function that returns 1 if the condition is true, and 0 otherwise.

[0225] Normalized quality score:

[0226] in This represents the AUC value calculated by client i on the local validation set in round t, reflecting the model's performance on this central dataset; This represents the normalized quality score.

[0227] 2. Calculation of data distribution similarity

[0228] Maximum Mean Discrepancy (MMD) represents the square of the maximum mean difference, measuring the distance between two distributions:

[0229]

[0230] This represents the data probability distribution of client i, describing the overall characteristics of patients in that center. To represent the global data distribution, it is approximated by a mixed distribution of data from all clients; : ℝ^p → ℋ, a feature mapping function that maps samples in the original feature space to the reproducing kernel Hilbert space (RKHS); ℋ: a high-dimensional or even infinite-dimensional reproducing kernel Hilbert space, which is convenient for calculating the distance between distributions.

[0231] Unbiased estimation:

[0232]

[0233] in: : The number of samples for client i; The number of samples in the global dataset (usually composed of samples from various centers); : The feature vector of the k-th sample of client i, where p is the feature dimension; The p-th sample in the global dataset.

[0234] Gaussian kernel function:

[0235]

[0236] in, It represents the Euclidean distance between samples x and y, and measures the similarity of the feature spaces; This represents the bandwidth parameter of the kernel function, which controls the rate of similarity decay.

[0237] Median heuristic determination :

[0238]

[0239] in: The set of distances between all sample pairs; : Take the median function so that σ adapts to the data scale.

[0240] Similarity weight:

[0241]

[0242] in: The temperature coefficient controls the sensitivity of the similarity weight to distribution differences. This represents the MMD squared estimate between client i and the global distribution; a larger value indicates a greater distributional difference.

[0243] Normalized distribution similarity:

[0244]

[0245] in satisfy =1

[0246] Data volume weights (logarithmic):

[0247]

[0248] in This indicates the number of data samples for client i; Logarithmic transformation prevents large data centers from monopolizing weights due to excessive data volume, and +1 avoids the logarithm being 0.

[0249] Overall contribution:

[0250]

[0251] in These represent hyperparameters, which respectively represent the amount of data. Data quality Similarity to data distribution The relative importance of satisfies α + β + γ = 1. This represents the overall contribution of client i in round t.

[0252] 3. Adaptive Aggregation Formula

[0253] Momentum aggregation:

[0254]

[0255] in This indicates that the parameter update vector of client i reflects the changes to the model caused by local training. This represents the momentum coefficient, which controls the proportion of historical information retained; the larger the value, the more conservative the data retention. This represents the global model parameters in round t.

[0256] Momentum scheduling strategy:

[0257]

[0258] in, Let represent the momentum coefficient in the t-th training round. The initial momentum coefficient, is the final momentum coefficient, t is the current training round, and T is the total training rounds.

[0259] 4. Anomaly detection and removal mechanism

[0260] Weighted average update:

[0261]

[0262] in To represent the weighted average direction of updates from all clients

[0263] Cosine similarity calculation:

[0264]

[0265] in: This refers to the dot product operation of vectors. Let L2 be the L2 norm (Euclidean length) of the vector. This represents the cosine similarity; a larger value indicates a more consistent update direction.

[0266] Anomaly detection criteria:

[0267] Client i has been removed ⟺

[0268] Where: τ is the similarity threshold, set based on experience (normal client similarity is usually >0.5); if a client is marked as abnormal for 3 consecutive rounds, its participation will be suspended and manual review is required.

[0269] Step S4: Federated Knowledge Distillation and Lightweight Deployment

[0270] Federal knowledge distillation framework, such as Figure 4 As shown:

[0271] 1. Teacher Model Training

[0272] choose A large, well-resourced hospital (such as a top-tier tertiary hospital) was used to train a teacher model. .in This represents the trainable parameters.

[0273] Teacher model architecture:

[0274] The above describes the complete CPMAFNet Cross-modal Attention Fusion Transformer Network (CPMAFNet).

[0275] Loss function:

[0276]

[0277] Federated aggregation (using QA-FedAvg):

[0278]

[0279] Indicates the number of clients. This represents the parameters of the teacher model in the (t+1)th round, where the number of training rounds is T1=100.

[0280] 2. Knowledge Distillation

[0281] Simplified version of CPMAFNet, main compression strategies:

[0282] Reduce the number of Transformer layers: 6 layers → 2 layers

[0283] Reduce the number of heads to focus on: 8 heads → 4 heads

[0284] Reduce the hidden layer dimension: 256 → 128

[0285] Replace ResNet50 (image encoder) with MobileNetV3.

[0286] Number of parameters: 10 8 →10 7

[0287] Soft tag generation:

[0288] The probability of the teacher model generating soft labels at temperature τ:

[0289]

[0290] in: The output represents the log-odds ratio of the teacher model for category k; temperature parameter. (τ>1 softened distribution). The effect of temperature: the smaller τ is, the closer it is to: standard softmax (one-hot encoding); the larger τ is, the closer it is to: a smoother probability distribution, containing "dark knowledge" (e.g., a sample is not AMI, but has a similarity of 0.2 to AMI).

[0291] 1) Student model training

[0292] Soft prediction probability of the student model:

[0293]

[0294] in This represents the logit output of the student model for class k.

[0295] Distillation loss (KL divergence):

[0296]

[0297]

[0298] Denotes KL divergence, the scale of soft-label gradients, and Proportional; multiplied by Compensation is used to bring the soft label loss to the same order of magnitude as the hard label loss.

[0299] 2) Hard-label loss (standard cross-entropy):

[0300]

[0301] in For real labels (one-hot encoded) The prediction is for τ=1.

[0302] 3) Total loss of the student model:

[0303]

[0304] in This represents the balance parameter in distillation loss, controlling the ratio of hard label loss to distillation loss.

[0305] 3. Federal distillation polymerization:

[0306] After each hospital (including smaller hospitals) trains its local student model, it uploads updated parameters, and the server aggregates the data.

[0307]

[0308] in Represents the updated parameters of the student model.

[0309] The global student model was broadcast to all hospitals.

[0310] Feedback to update the teacher model:

[0311]

[0312] The Expand(.) function maps the student model to the teacher model dimension through weight interpolation. The main focus is on retaining the teacher's knowledge while appropriately incorporating the student's generalization abilities.

[0313] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, 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 process, method, article, or apparatus.

[0314] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-center collaborative modeling and risk prediction method based on quality-aware federated learning, characterized in that, Includes the following steps: Step 1, Data Preprocessing and Feature Standardization: Each client preprocesses its local multimodal data according to the different center federated data standardization protocols, including time alignment, missing pattern embedding, and cross-center normalization; Step 2, Quality-Aware Aggregation: The server uses a quality-aware adaptive federated aggregation algorithm to integrate client updates and generate a new global model; Step 3, Federated Training: The server distributes the global model to the clients, each client trains using local data, calculates and updates parameters, performs differential privacy protection, and then uploads the model; Iterative Optimization: Repeat steps S2-S3 until the model converges; Step 4, Model Deployment: Deploy the final global model to all participating clients to provide real-time auxiliary diagnosis.

2. The method for multi-center collaborative modeling and risk prediction based on quality-aware federated learning according to claim 1, characterized in that, In step 1, the time alignment involves defining the symptom onset time T0 as the global time reference and constructing a unified timeline. , The sampling interval is k, and the time index is k; the original time series data is resampled to the standard time point by cubic spline interpolation.

3. The method for multi-center collaborative modeling and risk prediction based on quality-aware federated learning according to claim 1, characterized in that, In step 1, the missing pattern embedding involves: constructing a missing mask matrix M and a missing type vector r; and using a learnable embedding function. Map missing types to the embedding space. This represents the code for the missing reason of the j-th feature in the i-th sample, with values ​​of 0, 1, 2, and 3, corresponding to "observed", "systematic not detected", "random detection failed", and "patient refused" respectively. This represents a learnable parameter matrix used to map the missing type of one-hot encoding to... 3D embedding vector; Will Convert to a 4-dimensional one-hot vector. The length of the embedding vector is represented by the length of the observation; the observation is concatenated with the missing embedding as the final feature representation.

4. The method for multi-center collaborative modeling and risk prediction based on quality-aware federated learning according to claim 1, characterized in that, In step 3, the quality-aware adaptive federated aggregation algorithm calculates the overall contribution of client i as follows: in These represent hyperparameters, which respectively represent the amount of data. Data quality Similarity to data distribution The relative importance of satisfies α + β + γ = 1; This represents the overall contribution of client i in round t.

5. The method for multi-center collaborative modeling and risk prediction based on quality-aware federated learning according to claim 1, characterized in that, In step 3, the global model is a cross-modal attention fusion Transformer network, including: Modality-specific encoders: handle structured data, electrocardiograms, laboratory tests, text, and images respectively; Cross-modal attention layer: Calculates attention weights between modal pairs and fuses features; Multi-head self-attention layer: Globally fuses information from all modalities; Multi-task output layer: Simultaneously completes disease classification, risk scoring, mortality prediction, and readmission prediction.

6. The method for multi-center collaborative modeling and risk prediction based on quality-aware federated learning according to claim 5, characterized in that, The cross-modal attention layer calculates attention weights for modality pairs (i,j): in, The query matrix represents mode i. The bond matrix representing mode j. Indicates the scaling factor. The dimension of the attention head is represented, and the features are updated through residual connections: ,in These are the features of modality i after attention enhancement. This represents the feature representation of mode i after uniform dimensional projection.

7. The method for multi-center collaborative modeling and risk prediction based on quality-aware federated learning according to claim 5, characterized in that, The multi-task output layer adopts a dynamic weight adjustment strategy. in This represents the weight of task k in the t-th round of training. Let j be the loss of task k in round t-1, and j represent the task index. The weights are adaptively assigned based on the loss of each task in the previous round.

8. A multi-center collaborative modeling and risk prediction system based on quality-aware federated learning, wherein the system is based on the multi-center collaborative modeling and risk prediction method based on quality-aware federated learning as described in any one of claims 1-7, characterized in that, It includes a federated learning server responsible for global model management, client scheduling, security aggregation, and anomaly detection; multiple hospital clients, including local databases, data preprocessing modules, local training modules, differential privacy protection modules, and clinical application interfaces; and a secure communication layer, with data transmission channels based on TLS and homomorphic encryption.

9. A multi-center collaborative modeling and risk prediction system based on quality-aware federated learning according to claim 8, characterized in that, The federated learning server includes: a global model management module for storing and version-controlling global model parameters; a quality assessment module for calculating the data quality and distribution similarity of each client; a secure aggregation engine for performing aggregation calculations of encrypted parameters; a privacy budget management module for tracking and allocating differential privacy budgets; and an anomaly detection module for identifying and removing abnormal client updates.

10. A multi-center collaborative modeling and risk prediction system based on quality-aware federated learning according to claim 8, characterized in that, The client includes: a data preprocessing module for implementing the CPCFed protocol and performing time alignment, missing pattern embedding, and standardization; a feature extraction module for extracting deep features from electrocardiograms, images, and text; a local training module for training models using local data; and a clinical application module for providing real-time diagnosis, risk warning, and treatment recommendations.

Images