Abstract
PURPOSE: Early identification of lung cancer (LC) will considerably facilitate the intervention and prevention of LC. The human proteome micro-arrays approach can be used as a "liquid biopsy" to diagnose LC to complement conventional diagnosis, which needs advanced bioinformatics methods such as feature selection (FS) and refined machine learning models. METHODS: A two-stage FS methodology by infusing Pearson's Correlation (PC) with a univariate filter (SBF) or recursive feature elimination (RFE) was used to reduce the redundancy of the original dataset. The Stochastic Gradient Boosting (SGB), Random Forest (RF), and Support Vector Machine (SVM) techniques were applied to build ensemble classifiers based on four subsets. The synthetic minority oversampling technique (SMOTE) was used in the preprocessing of imbalanced data. RESULTS: FS approach with SBF and RFE extracted 25 and 55 features, respectively, with 14 overlapped ones. All three ensemble models demonstrate superior accuracy (ranging from 0.867 to 0.967) and sensitivity (0.917 to 1.00) in the test datasets with SGB of SBF subset outperforming others. The SMOTE technique has improved the model performance in the training process. Three of the top selected candidate biomarkers (LGR4, CDC34, and GHRHR) were highly suggested to play a role in lung tumorigenesis. CONCLUSION: A novel hybrid FS method with classical ensemble machine learning algorithms was first used in the classification of protein microarray data. The parsimony model constructed by the SGB algorithm with the appropriate FS and SMOTE approach performs well in the classification task with higher sensitivity and specificity. Standardization and innovation of bioinformatics approach for protein microarray analysis need further exploration and validation.