A Multi-constraint Deep Semi-supervised Learning Method for Ovarian Cancer Prognosis Prediction

Hua Chai, Long-Yi Guo, Minfan He, Zhongyue Zhang, Yuedong Yang

Abstract: Evaluating ovarian cancer prognosis is important for patients’ follow-up treatment. However, the limited sample size tends to lead to overfitting of the supervised evaluation task. Considering to get more useful information from different perspectives, we proposed a semi-supervised deep neural network method called MCAP. MCAP introduced the heterogeneity information of the tumors through unsupervised clustering constraint, to help the model better distinguish the difference in the prognosis of ovarian cancer. Besides, the data recovering constraint is used to ensure learning a high-quality and low-dimensional representation of the genes in the network. For making a comprehensive analysis for ovarian cancer, we applied MCAP to seven gene expression datasets collected from TCGA and GEO databases. The results proved that the MCAP is superior to the other prognosis prediction methods in both 5-fold cross-validation and independent test.