Layerwise feature selection in Stacked Sparse Auto-Encoder for tumor type prediction

Transcriptome data has been proved to be very valuable for clinical applications, such as diagnosis and prognosis of various cancers. In this paper, we present layer-wise feature selection in conjunction with stacked sparse auto-encoders (SSAE), a deep learning strategy for tumor classification with gene expression data. While SSAE learns high-level features from data, performing feature selection in every layer is a heuristic to obtain relevant features at every stage and also to assist in reducing the computation during fine-tuning procedure. The data in the new feature representation is finally used by classifier(s) to perform Tumor detection. The algorithm was tested on 36 datasets from the GEMLeR repository and w.r.t. AUC (Area under ROC curve) performance, it was found to outperform the GEMLeR benchmark results on 35 datasets (tied on the other dataset).