Table 3.

Comparative assessment of candidate biomarker multivariate models at a fixed specificity and sensitivity for the diagnosis of liver injury

Metric	Model	Threshold	Confirmatory cohort						Replication cohort
Specificity ≥ 0.90			Specificity	Sensitivity	TN	TP	FN	FP	Specificity	Sensitivity	TN	TP	FN	FP
Logistic regression	FBP1 + GSTA1	0.50	0.92	0.31	70	10	22	6	0.90	0.13	37	3	21	4
	FBP1 + GSTA1 + LECT2	0.45	0.91	0.56	69	18	14	7	0.83	0.33	34	8	16	7
	FBP1 + CES1 + LECT2	0.52	0.91	0.47	69	15	17	7	0.85	0.21	35	5	19	6
Random forest	FBP1 + LECT2	0.46	1.00	1.00	76	32	0	0	0.83	0.42	34	10	14	7
	FBP1 + LECT2 + CPS1	0.46	1.00	1.00	76	32	0	0	0.76	0.46	31	11	13	10
Sensitivity ≥ 0.90
Logistic regression	FBP1 + GSTA1	0.14	0.36	0.91	27	29	3	49	0.29	0.96	12	23	1	29
	FBP1 + GSTA1 + LECT2	0.14	0.39	0.91	30	29	3	46	0.46	0.88	19	21	3	22
	FBP1 + CES1 + LECT2	0.16	0.41	0.91	31	29	3	45	0.46	0.79	19	19	5	22
Random forest	FBP1 + LECT2	0.46	1.00	1.00	76	32	0	0	0.83	0.42	34	10	14	7
	FBP1 + LECT2 + CPS1	0.46	1.00	1.00	76	32	0	0	0.76	0.46	31	11	13	10

Each model compared onset non-DILI (NDO) cases versus DILI cases (DO) and was trained using the confirmatory cohort and validated using the replication cohort.

TP true positive, TN true negative, FP false positive, FN false negative.