CN-122000084-A - Multi-center collaborative modeling and risk prediction method and system based on quality perception federal learning

CN122000084ACN 122000084 ACN122000084 ACN 122000084ACN-122000084-A

Abstract

The invention discloses a multi-center collaborative modeling and risk prediction method and system based on quality perception federal learning. The method comprises the steps of firstly, preprocessing data and standardizing features, then, executing quality perception aggregation, integrating client updating by a server through a quality perception self-adaptive federation aggregation algorithm to generate a new global model, then, federating training, transmitting the global model to the clients through the server, using local data training, calculating parameter updating, uploading after differential privacy protection, iterating and optimizing until the model converges, and finally, deploying the model, deploying the final global model to each participating client to provide real-time auxiliary diagnosis. The method can solve the problems of time sequence alignment and missing value intelligent coding of multi-center heterogeneous data, comprehensively consider dynamic weight aggregation of data quantity, data quality and data distribution similarity, realize deep interaction fusion of electrocardiogram, images, texts and structured data, and support resource-limited hospital deployment lightweight model.

Inventors

GUO LIQUAN
ZHANG BOCHAO
WANG JIPING
XIONG DAXI
ZHOU WEINAN
ZHOU WEI

Assignees

中国科学院苏州生物医学工程技术研究所

Dates

Publication Date: 20260508
Application Date: 20260407

Claims (10)

1. The multi-center collaborative modeling and risk prediction method based on quality perception federal learning is characterized by comprising the following steps of: Step 1, preprocessing and feature standardization, wherein each client performs preprocessing on local multi-mode data according to different central federal data standardization protocols, including time alignment, deletion mode embedding and cross-center normalization; Step 2, quality perception aggregation, wherein a server integrates client update by adopting a quality perception self-adaptive federation aggregation algorithm to generate a new global model; Step 3, federal training, in which a server transmits a global model to clients, and each client uses local data training, calculates parameter updating and carries out differential privacy protection and then uploads the parameter updating; And 4, model deployment, namely deploying a final global model to each participating client to provide real-time auxiliary diagnosis.
2. The multi-center collaborative modeling and risk prediction method based on quality-aware federal learning of claim 1, wherein in step 1, the time alignment is defined by defining symptom onset time T 0 as a global time reference, constructing a unified time axis , And carrying out cubic spline interpolation resampling on the original time sequence data to a standard time point.
3. The multi-center collaborative modeling and risk prediction method based on quality-aware federal learning according to claim 1, wherein in step 1, the missing pattern is embedded by constructing a missing mask matrix M and a missing type vector r by a learnable embedding function The miss type is mapped to the embedding space, The deletion reason code of the jth characteristic of the ith sample is represented, and the values of 0, 1, 2 and 3 are respectively corresponding to 'observed', 'systematic undetected', 'random detection failure', 'patient rejection', Representing a matrix of learnable parameters for mapping onehot encoded deletion types to Maintaining an embedding vector; Will be Converted into a onehot vector in 4 dimensions, Representing the length of the embedded vector, and splicing the observed value and the missing embedded vector as a final characteristic representation.
4. The multi-center collaborative modeling and risk prediction method based on quality-aware federation learning according to claim 1, wherein in step 3, a quality-aware adaptive federation aggregation algorithm calculates the comprehensive contribution of a client i as Wherein the method comprises the steps of Representing super parameters, respectively representing data volume Quality of data Similarity to data distribution Is used to determine the relative importance of the (c) in terms of the (c), satisfy alpha +: beta+gamma=1; And the comprehensive contribution degree of the client i at the t-th round is represented.
5. The multi-center collaborative modeling and risk prediction method based on quality-aware federal learning according to claim 1, wherein in step 3, the global model is a cross-modal attention fusion transducer network, comprising: A modality specific encoder for processing structured data, electrocardiogram, laboratory examination, text, and images, respectively; a cross-mode attention layer, namely calculating attention weights between mode pairs and fusing features; A multi-head self-attention layer for globally fusing all mode information; And the multi-task output layer is used for simultaneously completing disease classification, risk scoring, death prediction and readmission prediction.
6. The multi-center collaborative modeling and risk prediction method based on quality aware federal learning according to claim 5, wherein the cross-modal attention layer calculates attention weights for modal pairs (i, j): Wherein, the A query matrix representing a modality i is presented, The key matrix representing the modality j, The scale factor is represented as such, Represents the dimension of the attention header and updates the feature by residual connection: Wherein Is the feature of the mode i after attention enhancement, Representing a feature representation of modality i after a unified dimensional projection.
7. The multi-center collaborative modeling and risk prediction method based on quality-aware federal learning according to claim 5, wherein the multi-tasking output layer employs a dynamic weight adjustment strategy Wherein the method comprises the steps of Representing the weight of task k in the t-th round of training, And (3) the loss of the task k in the t-1 th round, j represents a task index, and the weight is adaptively allocated according to the loss of the previous round of each task.
8. The multi-center collaborative modeling and risk prediction system based on quality-aware federal learning is based on the multi-center collaborative modeling and risk prediction method based on quality-aware federal learning, and is characterized by comprising a federal learning server, a plurality of hospital clients, a safety communication layer and a safety communication layer, wherein the federal learning server is responsible for global model management, client scheduling, safety aggregation and anomaly detection, the hospital clients comprise a local database, a data preprocessing module, a local training module, a differential privacy protection module and a clinical application interface, and the safety communication layer is based on TLS and homomorphic encryption data transmission channels.
9. The multi-center collaborative modeling and risk prediction system based on quality-aware federal learning of claim 8, wherein the federal learning server comprises a global model management module for storing and versioning global model parameters, a quality assessment module for computing data quality and distribution similarity of each client, a security aggregation engine for performing an aggregate computation of encryption parameters, a privacy budget management module for tracking and allocating differential privacy budgets, and an anomaly detection module for identifying and rejecting anomalous client updates.
10. The multi-center collaborative modeling and risk prediction system based on quality-aware federal learning according to claim 8, wherein the client comprises a data preprocessing module for implementing CPCFed protocols for time alignment, deletion pattern embedding and standardization, a feature extraction module for extracting depth features from electrocardiography, images and texts, a local training module for training a model by using local data, and a clinical application module for providing real-time diagnosis, risk early warning and treatment recommendation.

Description

Multi-center collaborative modeling and risk prediction method and system based on quality perception federal learning Technical Field The invention belongs to the crossing field of medical artificial intelligence and computer technology, and particularly relates to a method and a system for collaborative data modeling, disease auxiliary diagnosis and prognosis risk prediction of a multi-center patient by combining federal learning and deep learning technologies. Background The hospital establishes a diagnosis and treatment center to provide rapid and accurate diagnosis and treatment support for patients, particularly for the group suffering from acute high-risk diseases such as acute myocardial infarction. However, the system currently faces a plurality of key challenges in the operation process, on one hand, diagnosis and treatment related data (such as electronic medical records, test results, electrocardiograms, image data and the like) among medical institutions are mutually split, effective cross-institution sharing and integrated analysis are difficult to realize, a remarkable data island phenomenon is formed, and on the other hand, due to extremely sensitive related patient health information, direct data concentration and sharing are strictly limited by laws and regulations such as personal information protection law and data security law, and outstanding privacy protection and compliance risks are brought. In addition, the artificial intelligent auxiliary diagnosis model constructed by the limited data of a single institution has the problem of insufficient generalization capability, and the phenomenon of reduced accuracy rate is often generated when the model is applied to other medical scenes, meanwhile, the diagnosis and treatment flow and risk assessment of high-risk diseases among hospitals and even different doctors at the present stage are not unified, and an intelligent tool which can fuse multiparty data and has strong generalization capability is required to be introduced so as to promote the realization of standardized and standardized diagnosis and treatment. In the intelligent construction of hospitals, the existing technical paths have obvious limitations. Traditional centralized AI modeling （Shaikh A A S ,et al. Weighted aggregation through probability based ranking: An optimized federated learning architecture to classify respiratory diseases.） relies on pooling raw data from various centers together, which is not only difficult and heavy to practice, but also faces serious compliance challenges due to touch sensitive medical data. If a training mode based on single central data is adopted, the model is very easy to be subjected to fitting due to lack of data quantity and diversity, the performance and generalization capability of the model are severely restricted （Cai Y Q,et al. Pitfalls in developing machine learning models for predicting cardiovascular diseases: challenge and solutions.）., and although the model is applied to traditional statistical methods such as logistic regression and the like in clinical risk assessment (such as classical GRACE scoring), the model expression capability of the model is limited, and high-dimensional and nonlinear multi-mode medical data characteristics are difficult to fully mine, so that the prediction precision encounters a bottleneck. In recent years, federal learning has been an emerging paradigm of distributed machine learning, providing a potential direction for solving the above-mentioned dilemma, and has been systematically elucidated by Yang et al in their founding work (Yang, qiang, et al FEDERATED MACHINE LEARNING: concept and applications), and has been explored by related patents such as "federal deep learning-based cardiac image processing methods" (patent number CN114612408 a). However, most of these preliminary attempts stay at the general framework level, which makes significant differences （Hicks, K. A., et al. (2018). "2017 Cardiovascular and Stroke Endpoint Definitions for Clinical Trials." Circulation, 137(9), 961-972.）, among diagnosis and treatment paths and risk evaluations of emergency patients of different hospitals, and the depth fusion of a specific clinical scene of a diagnosis and treatment center is not yet available, especially how to systematically integrate multi-mode data such as electronic medical records, electrocardiosignals, medical images and the like, determine the standard of the multi-mode data, and seamlessly embed the multi-mode data into a complete end-to-end business closed loop of "pre-hospital screening-risk prediction-auxiliary decision" for depth optimization, so that a truly available, reliable and standardized intelligent auxiliary diagnosis and treatment scheme is formed. The technical blank is found by combing the prior art, namely, the problems of time sequence alignment, missing value coding and the like of different centers cannot be effectively processed due to the lack of a federal l