Table 4.

Model evaluation for fracture risk prediction

	Models	AUC	95% CI of AUC	Delong’s test p-value	IDI	IDI p-value	NRI	NRI p-value
Discovery cohort	Model 1: conventional risk factors alone	0.70	0.66, 0.74	N.A.	N.A.	N.A.	N.A.	N.A.
	Model 2: conventional risk factors + 27 selected metabolites	0.74	0.70, 0.77	0.001	0.031	<0.001	0.07	0.03
	Model 1+ FN-BMD	0.72	0.68, 0.76	N.A.	N.A.	N.A.	N.A.	N.A.
	Model 2+ FN-BMD	0.75	0.72, 0.79	0.002	0.029	<0.001	0.08	<0.001
	Model 1+ LS-BMD	0.71	0.67, 0.75	N.A.	N.A.	N.A.	N.A.	N.A.
	Model 2+ LS-BMD	0.74	0.71, 0.78	0.003	0.032	<0.001	0.12	<0.001
Replication cohort	Model 1: conventional risk factors alone	0.81	0.73, 0.89	N.A.	N.A.	N.A.	N.A.	N.A.
	Model 2: conventional risk factors + 18 selected metabolites	0.87	0.81, 0.94	0.013	0.110	<0.001	0.23	0.03
	Model 1+ FN-BMD	0.87	0.80, 0.93	N.A.	N.A.	N.A.	N.A.	N.A.
	Model 2+ FN-BMD	0.91	0.86, 0.96	0.015	0.079	0.003	0.09	0.16
	Model 1+ LS-BMD	0.85	0.78, 0.92	N.A.	N.A.	N.A.	N.A.	N.A.
	Model 2+ LS-BMD	0.89	0.83, 0.95	0.022	0.089	0.001	0.26	0.01

Model 1: conventional risk factors includes sex, age, BMI, current smoking status, menopausal status. Model 2: conventional risk factors which includes sex, age, BMI, current smoking status, menopausal status plus 27 or 18 selected metabolites. Participants were grouped into two categories, [0, 20%) and [20%, 1], based on their fracture risk for NRI. Delong’s test, IDI, and NRI were used to compare model 2 and model 1, and additionally accounting for FN-BMD or LS-BMD for both model 2 and model 1. AUC=Area under curve. IDI=Integrated discrimination improvement. NRI=Net reclassification index. CI=confidence interval. BMD=bone mineral density. FN=femoral neck. LS=lumbar spine. N.A.=not applicable