Room 6C/6E Model-Free Machine Learning Method for Survival Probability Prediction

Friday, October 12, 2012: 8:00 PM
6C/6E (WSCC)
Yuan Geng , Statistics, North Carolina State University, Raleigh, NC
Wenbin Lu , Statistics, North Carolina State University, Raleigh, NC
Hao Helen Zhang , Statistics, North Carolina State University, Raleigh, NC
It is of great interest to predict cancer patients’ survival probability, i.e. the chance of surviving certain years, after a treatment, based on their medical records and gene expression of the tumor. Doctors may choose the best treatment according to the predicted survival probability.

 The traditional statistical methods would suffer systematic bias if the model assumption is violated. Also they have difficulty in handling high dimensional gene data. We propose a flexible model-free machine learning method, using weighted support vector machine (SVM) with the inverse censoring probability weight (ICPW), which is robust of model specification and performs well for the high dimensional data.  We have proved via numerical simulations that the proposed method outperforms the existing model based methods when the model is misspecified. Also, our method can handle high dimensional data with ease due to the advantage of the machine learning method.

We have applied the method on two real data. It successfully predicted the survival for the recurrent gliomas data by Piantadosi (1997), which could not be directly predicted by traditional methods. Our method also showed advantage on the breast cancer data by Van Houwelingen et al. (2006) which included thousands of gene expression data as predictors.