Survival Analysis For Maintenance Hemodialysis Patients

With the development of our society,end-stage renal disease has become one of the public health problems that seriously endanger human health and increase the economic burden both in China and around the world.The most important treatment is hemodialysis.Survival analysis is an important branch of statistical science and has been widely used in medicine and other related studies.However,the traditional principal analytical model——Cox model,assumes that event risk is a linear combination of patient attributes and often requires manual manipulation of attributes.This hypothesis is likely to be too ideal and too simple in real-world data to lead to predictable event risk rates because of 1)there may be a non-linear relationship between attributes,2)manual additions and deletions of attributes require medical experience,3)attributes are not always independent of each other.This paper addresses this important issue by using neural networks to automatically identify the complex relationships among patient attributes,thereby fitting out a model that can meet the actual demand more,and improve the prediction accuracy.This paper analyzed the theoretical deficiencies of the traditional model in detail and used the advantages of automatic learning of the complex relationships between data attributes to try to build an improved deep model by combining the advantages of deep learning technique.On the one hand,it eliminated some tedious steps,such as the traditional manual selection of important attributes,adding interaction property items according to prior experience.On the other hand,it expanded the linear assumption,which makes the model more applicable and more practical.In the process of building multi-layer network,the number of layers and the number of neurons are tuned to the optimum by grid search.The pre-training normalization,regularization and dropout are used to improve the model and prevent over-fitting.In order to validate the model's effect,this paper used the maintenance hemodialysis patient data set from the real hospital blood purification room.After preprocess step,the training and prediction of the traditional model and the deep model are carried out respectively on the dataset.The experiment results show that the traditional model is more suitable than deep model when the data is less.And when the data amount is large enough,the accuracy of traditional model is low and the model is easy to fail,but deep model can always Maintain good results.The average accuracy rate of increase of nearly 20 percent,which fully reflects the superiority of the deep model.High accuracy prediction is critical to improving the quality of patients' life on maintenance hemodialysis.The deep model can compare the event risk level between patients with different attributes,so medical staff can more easily identify the key factors and provide patients with better treatment options.On the other hand,the improvement of the method not only can only be used to solve the problem of hot medicine,but also can be applied to survival analysis and reliability analysis in many fields.