. 2021 Sep 23;8:637355. doi: 10.3389/fmolb.2021.637355

TABLE 1.

Summary for iterative progress on model precision scores.

	Best method	Tuning grid	Best tune hyperparameters	Validation	Test
miRNA	PCA and undersample	Max depth: [2:8]	Max depth: 6	Accuracy: 86%	Accuracy: 81%
	AdaBoost	# of trees: [1:16]	Number of trees: 12	F-score: 86%	F-score: 82%
	AdaBoost	# of trees: [1:16]	Co-efficiency of learning: Breiman	F-score: 86%	F-score: 82%
miRNA and mRNA	PCA SMOTE	Cost: 10⁽⁻⁴⁾ × (20:150))	Cost: 0.0025	Accuracy: 81%	Accuracy: 93%
miRNA and mRNA	SVM (linear kernel)	Cost: 10⁽⁻⁴⁾ × (20:150))	Cost: 0.0025	F-score: 0.82%	F-score: 92%
miRNA, mRNA, and methylation	PCA SMOTE	Cost: 10⁽⁻⁴⁾ × (20:150))	Cost: 0.0027	Accuracy: 82%	Accuracy: 93%
miRNA, mRNA, and methylation	SVM (linear kernel)	Cost: 10⁽⁻⁴⁾ × (20:150))	Cost: 0.0027	F-score: 83%	F-score: 92%

The miRNA model applied by feature selection through importance (the hybrid model) and the class imbalance solution through undersampling is the method to be applied for prediction. For both “miRNA–mRNA” and the “miRNA–mRNA–methylation” triple model, principal component analysis for dimensionality and SMOTE for the class imbalance solution was the best method to increase predictive power and stability of the model.