LAUSR

Although cancer diagnosis research has continuously made breakthroughs in a single indicator, it is a challenging task to improve its multiple joint indicators. This study proposes a multi-category multi-state information ensemble-based classification method. We fuse protein-coding and non-coding genes to construct co-expression profiles, which ensemble the field information of classical genetics and epigenetics. A hierarchical feature selection algorithm based on control groups is put forward to quickly remove irrelevant and redundant features without the bias caused by unbalanced dataset. Multiple heterogeneous diagnosis models, which ensemble multiple diagnosis model structures and model states, are constructed and a competition mechanism is then introduced to automatically select the best model from multiple heterogeneous models without deeply grasping the positive and negative fusion effects between different algorithms and features. We apply the proposed method to classify three high-incidence cancers, in which the classification accuracy and sensitivity are over 99.23% and the classification specificity is over 97.37%. This illustrates that the proposed method has upgraded the three joint indicators of cancer diagnosis at the same time. Compared with the state-of-the-art classification methods, the classification accuracy has been improved by 2.23–9.23%, the sensitivity by 6.25–37.40%, and the specificity by 0–12.02%. In addition, feature analysis reveals three biological findings.