A new survival analysis model in adjuvant Tamoxifen-treated breast cancer patients using manifold-based semi-supervised learning.

Breast cancer researchers are very interested in studying and analyzing the survival time of patients. However, choosing the appropriate model for survival time analysis is the main challenge in the survival analysis of these patients. Using a dataset of patients with breast cancer, a new survival analysis model is proposed in this study. First, using the Cox-PH model, the whole selected data was categorized into high-risk and low-risk groups. The classified data was given to the AFT model in the next step to estimate the survival time. In some instances, the value of the metabolic status of the Tamoxifen feature was missing. Next, using the information obtained from the first and second steps, in the proposed Manifold-based semi-supervised learning module, the value of missing data was estimated. Using all labeled data, the final survival analysis was performed, and the relative hazard obtained. The comparison of the proposed model via other previous models in survival analysis demonstrates that the proposed model is 26% more accurate than when only the Cox-PH model was used and 28% more accurate than when only the AFT model was used in the testing set. Furthermore, the findings demonstrate that when the Manifold-based SSL model was used in the learning module, the accuracy was 19% higher than the Co-training SSL method and 12% higher than the SVMSSL method. The advantages of the proposed survival analysis method are high capability for identifying the survival risk classes of the patient and high predictive accuracy for the relative hazard. • We present a novel survival analysis method by combining the Cox-PH model, the AFT model and the Manifold-based semi-supervised learning (SSL) model to effectively predict survival analysis. • Studies and analysis of the survival time of breast cancer patients, using both clinical and Pharmacogenomics information. • The proposed model is high capability for identifying the survival risk classes of patient. • The proposed model is high accuracy in estimating the label of unlabeled data, due to the use of knowledge contained in this data and the model's stability against noise. [ABSTRACT FROM AUTHOR]