Research On Feature Processing And Discount Factor Selection Methods In Computer-aided Diagnos

Computer-Aided Diagnosis means that the machine can better collect the symptoms of patients and make more reasonable inferences about patients’ diseases in communication with patients.However,in the process of disease identification,with the increase in the number of diseases,the data set of diseases and symptoms will become very large,resulting in the problem of dimension disaster.However,the existing Computer-Aided Diagnosis system cannot completely solve the problem of dimension disaster caused by the increase of data sets and does not extract the collected information more deeply.Therefore,starting from these two directions,this paper proposes to use symptom filtering and weight calculation methods to improve the judgment accuracy of the model.At the same time,the fixed discount factor scheme adopted by most tasks based on reinforcement learning has been proven to have great defects,so the hyperbolic function scheme is proposed to replace it.The main research content of this paper is divided into the following three aspects:Based on hierarchical reinforcement learning,we build a Computer-Aided Diagnosis model.When simple reinforcement learning models solve practical problems,they often encounter the problem that the state space and action space are too large,which leads to the unsatisfactory training effect of the model.The hierarchical reinforcement learning model uses the idea of "divide and conquer" based on the original,decomposes the space,and abstracts different levels of control layers.Each lower level is responsible for a disease group and keeps collecting the symptoms of the disease group,thus achieving the goal of reducing the action space.The upper level’s task is to manage the lower level and decides which lower-level subject to activate.Aiming at the low accuracy of disease recognition,we propose two auxiliary methods.First,set up a corresponding symptom filter for each disease group based on the frequency of symptoms,and effectively eliminate the symptoms with low frequency or high overlap,to achieve the purpose of manual dimensionality reduction.The second is the mining of the correlation between symptoms,mainly referring to SENet(Squeeze and Exception Nets).Through the modeling of channels,the correlation and weight values between different symptoms can be obtained,so that the new network model can achieve the purpose of amplifying effective symptoms and suppressing invalid symptoms.Compared with the original model,the model with these two auxiliary methods has a significant improvement in the accuracy of disease diagnosis.Secondly,we want to discuss the parameter discount factor of reinforcement learning,and a hyperbolic function is proposed to replace the fixed discount factor scheme adopted by most reinforcement learning tasks.The discount factor is used to convert and estimate the future rewards,to ensure the convergence of the Bellman Equation.Setting a reasonable value for the discount factor is a problem that needs to be solved for all tasks based on reinforcement learning.According to the research of psychology and neurology,hyperbolic function is more in line with the preferences of humans and animals in learning and selection.Finally,different comparative experiments have shown that the disease diagnosis model based on hierarchical reinforcement learning effectively improves the accuracy of disease diagnosis by adding auxiliary diagnostic methods and hyperbolic function discount factor schemes.Among them,two auxiliary methods,symptom filtering and weight calculation,have improved the diagnostic accuracy of the model by more than10%;The hyperbolic function discount factor scheme has improved the diagnostic accuracy of the model by about 4%.