Design of a Clinical Decision Support System for Fracture Prediction Using Imbalanced Dataset

Table 7

Comparisons of predictive performance among different sampling methods.

Group	Training phase			Testing phase
	AC (%)	SE (%)	SP (%)	AC (%)	SE (%)	SP (%)	G-mean	AUC
Random undersampling with OB1	68.57	74.82	62.32	63.16	77.00	62.36	0.6929	0.7590
Random undersampling with OB2	68.92	68.92	68.92	68.30	70.03	68.20	0.6909	0.7495
Clustering-based kNN undersampling with OB1	71.25	75.66	66.84	63.40	76.19	62.20	0.6884	0.7526
Clustering-based kNN undersampling with OB2	71.25	71.25	71.25	67.28	69.84	67.00	0.6840	0.7515
One-sided selection + OB1	77.80	82.59	35.62	94.54	0	100	0	0.7007
One-sided selection + OB2	89.75	1.41	99.79	71.43	41.98	73.13	0.5541	0.6626