Table 3.

Key differences of the HUNT Lung Cancer Model over externally validated risk prediction models developed in prospective population-based cohorts. AUC refers to prediction of 1-, 5- (EPIC), or 6-year cancer risk (PLCO, HUNT2).

Key studies Reference	LLPi Marcus et al., 2015, Raji et al., 2012	EPIC Hoggart et al., 2012	PLCOM2012 Weber et al., 2017, Tammemagi et al., 2013	HUNT2 Discovery cohort	CONOR Validation Cohort
Study group characteristics
Cohort type	Random selection (n=8760)	Multi-country health study (n=399 393)	Multicentre randomized screening (n=80 375)	One county 70% of total adult population (n=65 240)	One country, 11 health studies ever-smokers (n=45 341)
Age limit	45–79	35+	55–74	= 20	= 20
Median Pack-years	18·9	≈30a	≈30	10·3	11·5
Never-smokers analysed	Yes	Yes	No	Yes (n=24 725)	Not applicable
Follow-up, years	8·7 mean	5 max	6 max	13·2 mean	16 max
Feature selection	Yes backward	Yes, based on AUC and tdNRI	No, pre-specified	Yes backward	Not applicable
Number of variables	14 (6 selected)	12 (4 selected)	11	36 (7 selected)	7
Coding of non-linearities of continuous variables	No	Yes, including stratification	Yes	Yes	Yes
Report on missing data	Yes	Yes	No	Yes	Yes
MIc	No	No	No	Yes	Not applicable
MI with feature selectionc	No	No	No	Yes	Not applicable
Internal validation	Yes	No	Yes	Yesb	Not appliccable
External validation	Yes	EPIC test set	Yes	Yes
Discriminatory power (AUCd and/or C-indexe)
	C-index	AUC	AUC	C-index / AUC
Total Population	0·849	NR	NR	0·903
Ever-smokers	NR AUC 5y)	NR	0·803 (6 y)	0·869
External validation	0·67, 0·76, 0·82	0·787 (5 y) (Vlaanderen et al., 2014)	0·797 (6 y)	0·879 / 0·87 (6 y)

NR = not reported.

^aYears of smoking more than >15 cigarettes per day.

^bBootstrap in each of 30 multiply imputed datasets.

^cMI = multiple imputation.

^dArea under the receiver operating curve.

^eConcordance index.