Table 3.

Performance of CoxPH models trained with 1000 bootstrap samples. Inputs were varied based on number of lesion-ROI and different feature combinations. The best model was compared to previously established predictors. C-index (± standard deviation) was obtained. P-values of the overall model were obtained using Score (logrank) test. Models included all information unless specified in the table

Comparison	Input model	DLBCL		NSCLC
		C-index	P-value	C-index	P-value
Lesion-ROI subgroup	All features, all lesion-ROI	0.84 ± 0.04	< 0.001	0.71 ± 0.03	< 0.001
	5 biggest lesion-ROI	0.78 ± 0.06	< 0.001	0.67 ± 0.03	0.006
	1 hottest lesion-ROI	0.77 ± 0.05	< 0.001	0.68 ± 0.03	< 0.001
Feature subgroup	All features, all lesion-ROI	0.84 ± 0.04	< 0.001	0.71 ± 0.03	< 0.001
	BL + FU + Patient-level Response	0.67 ± 0.06	0.03	0.70 ± 0.03	< 0.001
	FU	0.64 ± 0.08	0.2	0.65 ± 0.03	< 0.001
	BL	0.67 ± 0.06	0.03	0.64 ± 0.03	< 0.001
Previously Established Predictors	All features, all lesion-ROI	0.84 ± 0.04	< 0.001	0.71 ± 0.03	< 0.001
	RECIST	0.59 ± 0.05	0.14	0.50 ± 0.01	0.47
	PERCIST	0.58 ± 0.05	0.2	0.51 ± 0.01	0.51
	Deauville	0.62 ± 0.07	0.25	-	-
	Global SUVpeak2	-	-	0.57 ± 0.03	0.01
	Global Volume 1	-	-	0.62 ± 0.03	< 0.001
	Percent Change Global SUVmax	0.58 ± 0.06	0.3	-	-