Table 3.

Top performing Extreme Gradient Boosting classifiers (XGBoost) for HemoPI-1, HemoPI-2, HemoPI-3 datasets before and after optimizing classifier hyperparameters and number N of descriptors.

Hyperparameters	Feature reduction	N	Acc. (%)	Prec. (%)	MCC statistic	CK statistic	AUC ROC
HemoPI-1 model and validation datasets
Default	MC (0.75)	26	95.7	93.7	0.914	0.914	0.957
			92.3	88.5	0.846	0.846	0.923
HemoPI-2 model and validation datasets
'colsample_bytree': 0.8, 'eta': 0.1, 'max_depth': 14, 'min_child_weight': 1, 'subsample': 0.7, 'tree_method': 'hist', 'objective':'binary:logistic'	RFECV	34	79.1	75.3	0.577	0.577	0.787
			70.3	67.2	0.398	0.397	0.697
HemoPI-3 model and validation datasets
'colsample_bytree': 0.8, 'eta': 0.2, 'max_depth': 14, 'min_child_weight': 0.2, 'subsample': 0.8, 'tree_method': 'approx', 'objective':'binary:logistic'	MC (0.75)	28	78.7	74.9	0.569	0.568	0.783
			72.6	69.1	0.445	0.444	0.720

Optimal number N of descriptors were determined using MC: multicollinearity and RFECV: recursive feature extraction with tenfold cross-validation.