Survival Prediction of Breast Cancer Patients Using Deep Learning Analysis
Mohammed Rahman1, Geoffrey Yoo2, Konstantinos Krampis3
1Department of Statistics, Hunter College, City University of New York
2Computer Science Program, The Graduate Center, City University of New York
3Department of Biological Sciences, Hunter College, City University of New York
Survival models have been explored in clinical settings to investigate the relationships between cancer patients’ prognostic covariates, such as genetic and clinical features, and their survival probability. The most common model used to assess prognostic variables in cancer patient survival is Cox proportional hazards, which may not accurately predict cancer patients’ outcomes because of the assumption that the outcome is a linear combination of covariates. We present a study of 873 breast cancer patient data collected from The Cancer Proteome Atlas (TCPA) Portal. We applied a Deep Learning method for survival prediction in Breast Cancer (BRCA) patients and validated its predictive accuracy. Survival predictions by means of DeepSurv, a deep feed-forward neural network, was compared with the Cox proportional hazard model and the random survival forest model. We validate that the DeepSurv method performs better than the more common survival models, and its stronger predictive ability with multi-dimensional data. DeepSurv displayed the highest model performance among the three models, with the c-index of the testing sets reaching a perfect score of 1.00, followed by the Cox Proportional Hazard model with 0.94 and Random Survival Forest with 0.91. The predictive and prognostic abilities of DeepSurv, enables clinical researchers to use deep neural networks as a means in predicting the effects of a patient’s clinical and genetic features on their risk of death.
Email questions and comments about this abstract to mohammed.rahman622@myhunter.cuny.edu.
