CN110348241A - A kind of multicenter under data sharing strategy cooperates with prognosis prediction system - Google Patents

Abstract

The invention discloses a multi-center collaborative prognosis prediction system under a data sharing strategy. The system can realize privacy-protected data sharing under multiple medical institution centers, thus providing enough data for model construction. The present invention uses an integrated learning algorithm that can obtain better prediction results than weak classifiers to construct the system. The system processes sensitive patient-level data at each center and simultaneously constructs sub-classifiers of the ensemble learning model, exchanging only less sensitive intermediate results to build a complete ensemble learning model, thus ensuring that the proposed multi-center model is compatible with The centralized model has the same or even better results. The multi-center collaborative prognosis prediction system of the present invention protects the personal privacy of patients and does not need to run algorithm models on large centralized data sources. In actual clinical applications, it provides a reliable solution for the lack of samples for constructing prediction models in a single medical institution. solution.

Description

A multi-center collaborative prognosis prediction system under a data sharing strategy

technical field

The invention belongs to the fields of medical treatment and machine learning, and in particular relates to a multi-center collaborative prognosis prediction system under a data sharing strategy.

Background technique

Prognosis prediction plays an important role in clinical research and practice. Predictive models constructed based on electronic health record (EHR) data from a single medical institution may lack sufficient statistical power and good generalization ability. Therefore, the construction of a prognosis prediction model based on the collaborative analysis of electronic health record data in multiple medical institution centers can be used to increase the number and coverage of patients used for model training, enrich the prognostic characteristics of patients, and ultimately improve the accuracy and accuracy of the model's prognosis prediction. Generalization. Integrated learning is an algorithm that is widely used in clinical prognosis. Unlike linear models such as logistic regression and cox models, integrated learning algorithms usually have better accuracy and have the ability to capture nonlinear relationships between variables. Avoid overfitting problems common in machine learning. Therefore, the use of ensemble learning algorithms for model construction provides an ideal solution for the construction of a multi-center collaborative prognosis prediction system. In addition, while performing multi-center prognosis prediction, the privacy of patients must be protected. Most of the existing privacy-preserving ensemble learning training models under multiple centers are based on encryption methods, such as using additive homomorphic encryption and other methods. Aslett et al. proposed an ensemble learning model based on fully homomorphic encryption. Magkos et al. used a protocol framework based on homomorphic encryption to build an encryption module to train an ensemble learning classifier. Although these encryption methods can prevent information leakage and data exchange, they will significantly affect computing and storage efficiency, have poor scalability, and are not suitable for processing large clinical data under multi-center.

Contents of the invention

The purpose of the present invention is to provide a multi-center collaborative prognosis prediction system under a novel data sharing strategy to address the deficiencies in the prior art.

The object of the present invention is achieved through the following technical solutions: a multi-center collaborative prognosis prediction system under a data sharing strategy, the system includes the following four modules:

(1) Data acquisition module: collect the data of each variable required for patient prognosis prediction in each medical institution center, and use it as the source data set of the medical institution center.

(2) Data anonymization module: randomly sample the source data set of each medical institution center with a percentage p, use an anonymization algorithm to generate anonymized data for the sampled data, and use the remaining data as the local training set of the medical institution center; The anonymized data of each medical institution center is collected by the central server to synthesize the enhanced data set; the enhanced data set is divided into two parts, namely the additional training set and the verification set; the additional training set is used for return and distributed to each medical institution center; The validation set is used to select the hyperparameters of the ensemble learning model.

(3) Model training module: each medical institution center trains the sub-classifier of the integrated learning model locally, and the training data in the training process includes the local training set of the medical institution center and the data returned by the central server to the medical institution center Additional training set; this indicates that the training set used to train the sub-classifiers for each medical institution center comes not only from the center itself but also from the datasets of other centers, thereby increasing the randomness of the dataset to improve the overall performance of the ensemble learning model. During training, the hyperparameters of the ensemble learning model are selected using a validation set created from the augmented dataset.

(4) Prognosis model application module: the central server collects sub-classifiers trained locally in each medical institution center to form a complete integrated learning model; new patient data is input into the integrated learning model to perform prognosis prediction.

Further, in the data anonymization module, the random sampling percentage p of the central source data set of each medical institution is selected as 50%. Fixing the proportion of anonymized data p at 50% can improve the prediction effect of the ensemble learning model, and the direct integration of sub-classifiers or the complete anonymization of data and centralized training cannot achieve the best results; the size of p can be adjusted to adapt to complex Decision support scenarios for prognosis prediction of patients in clinical practice under different scenarios.

Further, the anonymization algorithm may choose k-anonymity algorithm (k-anonymity), l-diversity (l-diversity), t-closeness (t-closeness) and differential privacy and other anonymous algorithms. Among them, the specific method for realizing k-anonymity can choose suppression, which means completely hiding certain information and not releasing certain data items.

Further, the system considers horizontal-partitioned data, that is, the source data set of each medical institution center has the same kind of variables.

The beneficial effects of the present invention are: the present invention innovatively proposes a multi-center data sharing strategy, which can realize privacy-protected data sharing under multiple medical institution centers, thereby providing sufficient data for model construction. The present invention uses an integrated learning algorithm (such as a random forest algorithm) that can obtain better prediction results than weak classifiers to construct the system. The system processes sensitive patient-level data at each center and simultaneously constructs sub-classifiers of the ensemble learning model, exchanging only less sensitive intermediate results to build a complete ensemble learning model, thus ensuring that the proposed multi-center model is compatible with The centralized model has the same or even better results. The multi-center collaborative prognosis prediction system of the present invention protects the personal privacy of patients and does not need to run algorithm models on large centralized data sources. solution.

Description of drawings

Figure 1 is a frame diagram of the multi-center collaborative prognosis prediction system under the data sharing strategy;

Figure 2 is a schematic diagram of a data sharing strategy;

Figure 3 is a schematic diagram of data transmission in each center;

Fig. 4 is a comparison chart of the prediction ability of the multi-center collaborative prognosis prediction system under the data sharing strategy of the present invention and the prognosis prediction system under centralized training.

Detailed ways

The present invention will be described in further detail below in conjunction with the accompanying drawings and specific embodiments.

A multi-center collaborative prognosis prediction system under a novel data sharing strategy provided by the present invention, as shown in Figure 1, includes the following four modules:

(1) Data acquisition module: collect the data of each variable required for patient prognosis prediction in each medical institution center, and use it as the source data set of the medical institution center. This embodiment adopts the data of colorectal cancer for experimental verification, wherein the number of medical institution centers is 5, and the electronic medical record data samples collected by each medical institution center through the data acquisition module are shown in Table 1, including age, gender , tumor size, T stage, N stage, and carcinoembryonic antigen index and other data information of six variables.

Table 1: Example of electronic medical record data collection at a single center for colorectal cancer patients

age gender Tumor size (mm) T stage N stage carcinoembryonic antigen index 1 65 male 4.8 I III positive 2 74 Female 1.5 II IV feminine … … … … … … …

(2) Data anonymization module: As shown in Figure 2, the source data set of each medical institution center is randomly sampled with a percentage p, and the anonymized data is generated using an anonymization algorithm for the sampled data, and the remaining data is used as the medical institution center local training set. Anonymized data from each medical institution center is collected by a central server to synthesize an augmented dataset; the augmented dataset is split into two parts, an additional training set and a validation set; the additional training set is used for backhaul and distributed to each medical institution center ; The validation set is used to select the hyperparameters of the ensemble learning model. In the experiment, the proportion of anonymized data p is set to 50%. The specific anonymization algorithm uses the suppression algorithm in k-anonymity. There are two hyperparameters that need to be selected through the verification set: the maximum number of features used by a single decision tree, sub The number of classifiers.

(3) Model training module: as shown in Figure 2, each medical institution center locally trains the sub-classifier of the integrated learning model, and the training data during the training process includes the local training set of the medical institution center and the central server return An additional training set for the facility center; this indicates that the training set used to train each facility center sub-classifier comes not only from the center itself but also from datasets from other centers, thereby increasing the randomness of the dataset to improve ensemble learning The overall performance of the model. During the training process, the hyperparameters of the ensemble learning model are selected using the validation set created from the augmented dataset, thereby solving the problem that the out-of-bag error (OOB) in the multi-center mode is not exactly the same as that of the standard random forest, which causes the unbiased estimation to be invalid. .

(4) Prognosis model application module: the central server collects sub-classifiers trained locally in each medical institution center to form a complete integrated learning model; new patient data is input into the integrated learning model to perform prognosis prediction. The experimental results are shown in Figure 4, and the prediction ability of the prognosis prediction system is measured by AUC. It can be seen that the multi-center collaborative prognosis prediction system under the data sharing strategy proposed by the present invention can achieve better prediction results than the prognosis prediction system under centralized training.

The above-mentioned embodiments are used to illustrate the present invention, rather than to limit the present invention. Within the spirit of the present invention and the protection scope of the claims, any modification and change made to the present invention will fall into the protection scope of the present invention.

Claims (4)

1. the multicenter under a kind of data sharing strategy cooperates with prognosis prediction system characterized by comprising

(1) data acquisition module: the number of each variable required for patient's prognosis prediction is collected respectively at each medical institutions center According to set of source data as the medical institutions center.

(2) data anonymous module: carrying out stochastical sampling to the set of source data at each medical institutions center with percentage p, to adopting Sample data generate anonymization data, local training set of the remaining data as the medical institutions center using anonymization algorithm；Come Enhancing data set is synthesized by central server collection from the anonymization data at each medical institutions center；Enhancing data set is divided into Two parts, i.e., additional training set and verifying collection；Additional training set is for returning and distributing to each medical institutions center；Verifying collection For selecting the hyper parameter (hyper parameter) of integrated learning model.

(3) it model training module: was being trained in the sub-classifier of the integrated learning model of locally training at each medical institutions center Training data in journey includes that the local training set at the medical institutions center and central server return to the medical institutions center Additional training set；This shows for training the training set of each medical institutions center sub-classifier to go back not only from center itself Data set from other centers, to increase the randomness of data set, to improve the overall performance of integrated study model.It is instructing During white silk, the hyper parameter of integrated learning model is selected using the verifying collection created from enhancing data set.

(4) prognostic model application module: the sub-classifier that each medical institutions center is trained is collected by central server and is constituted Complete integrated study model；New patient data is inputted into the integrated study model and executes prognosis prediction.

2. the multicenter under a kind of data sharing strategy according to claim 1 cooperates with prognosis prediction system, feature exists In, in the data anonymous module, the stochastical sampling percentage p selection 50% of each medical institutions center set of source data.

3. the multicenter under a kind of data sharing strategy according to claim 1 cooperates with prognosis prediction system, feature exists In k- anonymity algorithm (k-anonymity), l- diversity (l-diversity), t- proximity may be selected in the anonymization algorithm (t-closeness) and the anonymity algorithms such as difference privacy.The method for being wherein specifically used for realizing k- anonymity can choose inhibition (suppression), inhibit thoroughly to hide certain information, do not issue certain data item.

4. the multicenter under a kind of data sharing strategy according to claim 1 cooperates with prognosis prediction system, feature exists In the system considers horizontal segmentation data (horizontal-partitioned data), i.e., the source at each medical institutions center Data set has the variable of identical type.