| ICU is characterized by high cost of treatment and high mortality rate.However,compared with developed countries,domestic ICU resources are relatively scarce.In this context,intelligent medical enterprises with "AI+ICU" have emerged in China in recent years.By providing AI-assisted diagnosis and treatment services for hospital ICU,it can realize early warning of patients’ conditions,timely diagnosis and treatment,prevention over treatment,reduce the length of patients’ stay in hospital,reduce treatment costs and accelerate bed turnover rate.As one of the auxiliary diagnosis and treatment techniques,the prediction of ICU mortality has been an active topic in the medical field for decades.The prediction of ICU mortality is of great significance to the rescue of patients and the allocation of ICU resources.The current research on ICU mortality prediction mainly focuses on the application of machine learning and deep learning models to improve the prediction accuracy.The end-to-end hierarchical feature extraction ability of deep learning models makes the learning ability of deep learning models significantly higher than that of machine learning methods.However,the deep learning models used for medical tasks are mostly from the fields of image classification and natural language processing,etc.,and there is no specially designed model corresponding to the characteristics of sparse and unbalanced time series data in medical learning tasks.At present,the research of neural architecture search algorithm is changing rapidly and has achieved good results in target detection,image classification and other tasks.In this paper,the research work of neural architecture search is divided into search space,search strategy and performance evaluation strategy and briefly introduced,and then we improved the existing ENAS model to adapt to the task of ICU mortality prediction,and reconstructed the optimal subnetwork of Cell structure obtained by the model to construct the ICU mortality prediction model.The main work is as follows: 1)To improve the training of sub-network sharing parameters,we introduced the knowledge of unbalanced learning,such as cost sensitive function and resampling method,to the imbalance of medical data;2)To improve the incentive function in the search strategy,the calculation of incentive value is changed from considering the loss value of the subnetwork on the verification set to the evaluation score of the subnetwork on the verification set;3)The ICU mortality prediction model was constructed by reconstructing the searching subnetwork.The optimal subnetwork obtained from the model was taken as the Cell unit of the bidirectional neural network to construct the ICU mortality prediction model;4)Fully experiments,we respectively in open MIMICIII,e ICU mortality prediction data sets,through the experiment we found that the excitation function in the process of neural network search task of customization and the network sharing parameters in training imbalance of learning knowledge fusion,optimizing search convergence and son of the network training model has important significance.In addition,the prediction effect of the ICU mortality prediction model based on bidirectional circulating neural network architecture was 16% higher than that of the comparison model under the AUROC value,reflecting the feasibility and superiority of our improved model. |
