Diagram of Consensus Feature Elimination. We first generated 100 alternative 5-fold random splits of samples, upon which it constructs 500 classifiers with their AUCs as well as weight vectors. Each feature is then ranked by its average square weight. The lowest ranking feature was removed backward until the maximum average AUC was achieved. The procedure is repeated for 100 times, and the most frequently occurring marker set was regarded to be the ultimate marker.