Multi-center Clinical Data-driven Heart Failure Combined With Acute Kidney Injury Prediction Method Study

The prediction of heart failure combined with acute kidney injury provides a reasonable diagnosis and treatment plan,results in a risk reduction and prognosis improvement of patients,so as to lower the medical expenditure.With the development of medical informatization,a large number of Electronic Health Records(EHRs)are able to be accumulated and applied to provide conditions for data-driven prediction methods for heart failure and acute kidney injury.Datadriven prediction method for heart failure combined with acute kidney injury makes full use of the patient information in the EHRs.It reflects the real clinical environment,and facilitates the inclusion of new risk factors.The naive data-driven prediction method for heart failure combined with acute kidney injury is only applicable to patients with similar characteristic information.However,data-driven prediction method for heart failure combined with acute kidney injury in a single clinical center is only applicable to patients with similar distribution of characteristics.It suffers from the distribution bias of multi-center clinical data sets,lacking the generalizability of clinical practice.Therefore,it is not able to be applied to multiple patients from other countries or regions.The multi-center clinical data-driven prediction method for heart failure combined with acute kidney injury can extract common characteristics of patients in different environments to improve prediction performance and to support clinical decision-making.Therefore,the prediction method of heart failure combined with acute kidney injury driven by multi-center clinical data is more authoritative and generalizable.However,there are still several challenges in multi-center clinical data-driven research methods,such as: 1)It is difficult to extract potential common features because the characteristics of patients in distinct environments are quite different;and 2)Current prediction models have not fully utilized the medical knowledge related to heart failure complicated with acute kidney injury,and the semantics of patient features are not rich enough.To address the above challenges,this thesis proposes a knowledge-aware multi-center clinical dataset adaption method.First,in order to extract the common features of patients from different clinical centers,this thesis proposes a feature extraction method based on adversarial learning.Specifically,the generative adversarial network is used to extract the potential common features of multi-center clinical datasets to predict whether a patient has heart failure with acute kidney injury.After constructing a knowledge graph related to heart failure combined with acute kidney injury,this thesis merges existing medical knowledge and the common features extracted from adversarial learning to improve prediction performance of the multi-center clinical datadriven heart failure combined with acute kidney disease.A case study was conducted on two clinical datasets which consist of 5,075 heart failure patient samples collected from a China upper first-class hospital and 1,006 patient samples collected from MIMIC-III respectively to evaluate the performance of the proposed method.The experimental results show that: 1)The prediction method based on adversarial learning can better extract the potential common features of multi-center datasets,so it is more suitable for predicting heart failure combined with acute kidney injury than traditional machine learning algorithms;2)Knowledge-aware method incorporating the knowledge graph into the model makes full use of existing medical knowledge and effectively improve the predictive performance of multi-center patients with heart failure and acute kidney injury.The multi-center clinical data-driven prediction method for heart failure combined with acute kidney injury proposed in this thesis has significantly better performance than state-of-the-art methods.It effectively extracts the potential common features of multi-center clinical patients and makes full use of medical information of the knowledge graph.This proposed methodology is able to be smoothly shifted to tackle with other diseases by adaptable utilization of multi-centers clinical datasets,as a crucial advantage over other techniques for clinical risk prediction and prevention.