Application Of The Deep Gradient Boosting Model In Stroke Prediction

Ensemble learning has been a pervasive machine learning technique for its high accuracy.However,ensemble learning has limited expression ability for complex problems,resulting in poor generalization,while deep learning achieved better performance with its strong expression ability.To improve the performance of machine learning systems for complex problems.we propose the Deep Gradient Boosting(DGB)model,which is a novel machine learning model that combines the principles from both deep learning and ensemble learning.Stroke is one of the most dangerous diseases in China,which has a death rate per 100,000 of 127.2.Currently,the diagnosis of stroke is mainly manual,which is hard and costly.Therefore,it is significant for the diagnosis and preventing of stroke to build a risk prediction model with the DGB from real medical history and further benefits the allocation of medical resources.We made the following contributions:The DGB model.The DGB consists of the input layer,the hidden layers,and the output layer.The input layer extract features,and can be subdivided into multiple random subspaces to leverage the parallel processing during the training stage.The outputs of hidden layers are stacked together with the original features to retain information,and the number of hidden layers is self-adjusted according to the rate of change of neighboring layers.The output layer generates final results through learning strategies.We evaluate the performance and compare the DGB model with the traditional individual model and the simple ensemble model on various data set,the performance superiority of the DGB model is proved.Furthermore,we perform the sensitivity analysis on the DGB model to obtain the optimal range of parameters.Finally,we parallelize the input layer nodes to accelerate the training process.The prediction of stroke.Stroke has various causes and medical grades,corresponding to different diagnostic and assessment methods.We reviewed the pathogenesis,the clinical manifestations,the diagnostic and assessment methods of stroke from different classes and grades.By training the DGB model with real cerebral apoplexy data sets,we can classify and predict the indication for people at risk above medium.Stroke monitor mobile applet.We build a mobile applet based on the model,which can display the predicted risk of cerebral apoplexy in graphics based on the user-input information,i.e.,basic examination,the living,the medical history,the diet,and the physique.