Research On Composite Quantile Regression Neural Network Method For Right-censored Data

With the development of the field of survival analysis,statistical inference of rightcensored data is of great importance for the study of medical diagnosis.In recent years,machine learning methods have developed rapidly and offer greater advantages over traditional statistical methods in some respects.As a result,there is a growing trend in statistical research to integrate the two.Composite quantile regression can effectively handle complex relationships between multiple independent and dependent variables,and the model is robust and has broader application conditions.In this paper,we take right-censored data as an entry point to directly predict the survival time of events by fusing composite quantile regression and different neural network methods based on in-depth study and research of statistical and machine learning methods.The main research components are as follows:First,an inverse probability weighting method is introduced to predict the survival of right-censored data by combining composite quantile regression with the loss function of a multi-hidden layer feedforward neural network.Also,the hyperparameters involved in the neural network were adjusted using a whale optimization algorithm,integer encoding and one-hot encoding were used to encode the classification features,and a binary whale optimization algorithm was used for variable selection for data with high variance.The performance of the model was evaluated through simulations and example data.The results show that the composite quantile regression neural network model with double hidden layers performs optimally across data sets,and that the C-index results are more stable using soloheat coded data,making the proposed prediction model sufficiently flexible to assist in medical decision making.Secondly,convolutional neural networks and penalty functions are used to implement feature filtering,and a composite quantile regression gated recurrent neural network method is proposed for survival time prediction of high-dimensional right-censored data.The model first passes the data pre-processed features through a one-dimensional convolutional and pooling layer,and the resulting output is passed into a composite quantile regression gated recurrent neural network with a loss function containing a penalty term,and then combined with an inverse probability weighting method to obtain the survival time prediction results.Finally,we validated the model on two simulated and two example data,respectively.Comparing the composite quantile regression neural network,Deep Quantreg and Cox proportional model,the results show that when K=5,it is the method proposed in this paper that has the highest survival prediction accuracy for high-dimensional censored data among all evaluation metrics.And as the value of K increases the prediction accuracy also decreases significantly.In summary,the method proposed in this paper not only effectively prevents overtreatment and wastage of medical resources,but also provides a scientific basis for medical staff and patients’ families when making medical decisions,and by achieving cross-domain data set validation of the proposed method,it also finds an effective way to address the problem of telecommunication customer churn,which is of great practical importance.