Evaluation of survival extrapolation in immuno-oncology using multiple pre-planned data cuts: learnings to aid in model selection

Table 4 Prediction accuracy key findings

Data cut	Model	Criteria for model selection	Prediction accuracy (months)
			PE		RMST
			Fitted (earlier) data cut	Latest (36-mo) data cut	Fitted (earlier) data cut	Latest (36-mo) data cut
12-mo	KM estimates		51.8	32.1	8.7	17.5
12-mo	Log-normal	Best AIC, BIC	−0.3	−10.2	−0.2	−0.9
12-mo	Log-logistic	Lowest RMST, PE	−0.7	−11.5	−0.2	−1.2
12-mo	1-knot Odds	Highest RMST, PE	−0.8	−6.6	−0.3	−0.5
18-mo	KM estimates		39.9	32.1	11.3	17.5
18-mo	Log-normal	Best AIC, BIC	+ 1.1	−8.7	+ 0.0	−0.5
18-mo	Log-logistic	Lowest RMST, PE	+ 0.0	−9.8	− 0.0	− 0.8
18-mo	1-knot Odds	Highest RMST	+ 1.2	−5.1	− 0.2	− 0.4
18-mo	3-knot Odds	Highest PE	+ 1.1	−4.8	−0.2	− 0.4
24-mo	KM estimates		35.8	32.1	13.5	17.5
24-mo	1-knot Odds	Best AIC	+ 1.2	−2.0	+ 0.0	+ 0.0
24-mo	Log-normal	Best BIC	−0.1	−6.0	+ 0.4	+ 0.1
24-mo	Log-logistic	Lowest RMST, PE	−1.6	−7.4	+ 0.2	−0.3
24-mo	Gen Gamma	Highest RMST	+ 1.2	−2.7	+ 0.2	+ 0.1
24-mo	3-knot Hazard	Highest PE	+ 0.9	−0.4	−0.1	− 0.1
36-mo	KM estimates		32.1	32.1	17.5	17.5
36-mo	1-knot Hazard	Best AIC, BIC, highest PE	+ 1.4	+ 1.4	+ 0.2	+ 0.2
36-mo	3-knot Hazard	Lowest RMST	+ 0.9	+ 0.9	−0.1	−0.1
36-mo	Log-normal	Highest RMST	−2.3	− 2.3	+ 0.9	+ 0.9
36-mo	Log-logistic	Lowest PE	−4.1	−4.1	+ 0.4	+ 0.4

Key:AIC Akaike’s information criterion; BIC Bayesian information criterion; Gen Generalized; mo Month(s); PE Point estimate; RMST Restricted mean survival time
Note: Negative values indicate that the model underestimates survival, whereas positive values indicate that the model overestimates survival. ‘Fitted’ refers to the data cut from which the models were fitted (i.e. the data cut stated within the left-hand column), and so a comparison is made between a model fitted to a given data cut and the Kaplan-Meier curve for this same data cut. ‘Latest’ refers to the 36-month data cut, and so a comparison is made between a model fitted to the specified data cut (which may be earlier) and the Kaplan-Meier curve for the 36-month data cut. Models were included in this table if one or more of the following criteria were met: (1) the model provided the ‘best’ AIC or BIC score, (2) the model provided either the ‘highest’ or ‘lowest’ estimate of RMST at 36 months, or (3) the model provided either the ‘highest’ or ‘lowest’ PE of survival at 36 months. Where RMST estimates were tied (to the nearest 0.1 month), the model with the lowest AIC or BIC was included here. Full prediction accuracy results are provided within the supplementary material

ISSN: 1471-2288