Survival Analysis on Relapse Predictions of Breast Cancer using Optimized ANN and Deepsurv Model
Poornima Sankar, Ramani Bai V · 2022 International Conference on Computing, Communication, Security and Intelligent Systems (IC3SIS) · 2022
Breast cancer is one of the widespread cancer types existing among women worldwide. The improvement in medical treatment and widespread awareness has reduced the mortality rate caused due to breast cancer. However, the patients have to face the next challenge, the relapse of breast cancer which may cause more severe effects leading to death. For some patients breast cancer may come back after treatment, sometimes even years later. This recurrence can be local, regional or in a distant area. The main aim of this research work is to predict the relapse of breast cancer among the patients using a machine learning approach. The proposed model also predicts the probability of relapse over time using a survival analysis approach. The data is obtained from Molecular Taxonomy of Breast Cancer International Consortium. It is a combination of clinical and molecular data of 1980 breast cancer patients. The clinical data comprises of basic clinical variables that are tumor grade, tumor size, patient age and number of lymph nodes positives. Molecular data includes segmented copy number alterations along 4794 consistent copy number regions. Artificial Neural Network is used to classify breast cancer patients as having a high or low risk of experiencing a recurrence. Survival Analysis is conducted on predicted high risk and low risk patient labels to predict the probability of relapse over time. The survival analysis is performed primarily using Linear Cox Proportional Hazards model. Further, the Deepsurv model is used to perform the survival analysis which outperforms the Linear Cox Proportional Hazards model. Deepsurv model generates individual and average survival analysis curves for predicted high risk patients. Artificial Neural Network predicts risk of breast cancer relapse with an accuracy of 0.95. The survival analysis models were evaluated by computing concordance index for the models. The Linear Cox Proportional Hazards model has a concordance index of 0.41. Further, the concordance index is improved to 0.71 using Deepsurv model.