LC risk assessment models

Statistical models relating multiple risk factors to
cancer risk can identify high-risk subsets of smokers.
There are three criteria to evaluate the performance
of risk assessment models: calibration (reliability),
discrimination, and accuracy [280]. Calibration as-
sesses the ability of a model to predict the num-
ber of endpoint events in subgroups of the popula-
tion and is evaluated by using the goodness-of-fit
statistic. Discrimination is a measure of a model’s
ability to distinguish between those who will and
will not develop disease, and is quantified by cal-
culating the concordance statistic, or area under a
receiver operating characteristic (ROC) curve. Ac-
curacy including positive and negative predictive
values refers to themodel’s ability to categorize spe-
cific individuals. The best-known cancer prediction
model is the Gail model for breast cancer [281]. It
has been validated in several populations [282–285]
and appears to give accurate predictions for women
undergoing routine mammographic screening but
probably overestimates the risk for young women
not undergoing routine mammography [286]. The
modest discrimination ability of the Gail model calls
for the incorporation of promising biological fac-
tors [287–290]. Prediction models for other cancers
(melanoma [291,292], colorectal cancer [293], and
LC [7,8]) have also emerged.
The few published LC risk assessment models
mainly focus on smoking behavior and demo-
graphic characteristics. Bach et al. [8] used data col-
lected from CARET, a large, randomized trial of LC
prevention, to derive a LC risk prediction model.
The model used the subject’s age, sex, asbestos ex-
posure history, and smoking history to predict LCrisk andwas derived by use of data fromfive CARET
study sites and then validated by assessing the ex-
tent it could predict events in the sixth study site.
The model was then applied to evaluate the risk of
LC among smokers enrolled in a study of LC screen-
ing with computed tomography (CT). The model
identified smoking variables (duration of smoking,
average number of cigarettes smoked per day, dura-
tion of abstinence), age, asbestos exposure and the
study drug, β-carotene and retinyl palmitate as sig-
nificant predictors of LC. Themodel provided strong
evidence that LC risk varies greatly among smok-
ers and was internally validated and well calibrated
with a cross-validated concordance index of 0.72.
Bach’s model is most applicable to heavy smokers
aged between 50 and 75 years. Recently, Spitz et al.
[294] developed lung cancer risk models for never,
former, and current smokers, respectively. In their
models, factorswith strong etiological roles, e.g., en-
vironmental tobacco smoke, family history of can-
cer, dust exposure, prior respiratory disease, and
smoking history variables were all identified as sig-
nificant predictors of lung cancer risk. The models
were internally validated with cross-validated con-
cordance statistics for the never, former, and current
smoker models of 0.57, 0.63, and 0.58, respectively.
The computed 1-year absolute risk of lung cancer
for a hypotheticalmale current smokerwith an esti-
mated relative risk close to 9was 8.68%. The ordinal
risk index performed well in that true-positive rates
in the designated high-risk categories were 69%
and 70% for current and former smokers, respec-
tively. When externally validated, this risk assess-
ment procedure could use easily obtained clinical
information to identify individualswhomay benefit
from increased screening surveillance for lung can-
cer. In summary, current LC risk prediction models
have been focused on smoking variables and there
is potential to developmore accuratemodels by col-
lecting more data and incorporating additional risk
factors. Moreover, external validation of existing
models to independent populations is important.Concluding remarks
The results of many reported associations of single
polymorphism analyses are incongruent and couldnot be replicated even with key study parameters
similar to the original ones. Beyond a possible effect
frompopulation heterogeneity, shortcomings in ex-
perimental design and statistical methodology such
as small sample size, lack of control for confound-
ing, selection bias, and multiple comparisons may
account for a large part of these discrepancies. Since
cancer is a multistep and multifactorial disease, the
influence of individual variants identified frommost
candidate gene approach studies on overall cancer
riskmight beminimal.Moreover,many cancer risk-
associated genetic variants lack functional valida-
tion. To circumvent these caveats, pathway-based
approaches have been exploited that simultane-
ously analyze the impact of multiple variants in
the same carcinogenesis-related signaling or func-
tion pathway on cancer predisposition. This strategy
might amplify the effect from single variants; how-
ever, the pathway-based approach also depends on
a priori knowledge from basic investigations sug-
gesting the involvement of the pathway in tumori-
genesis. A haplotype-based genome scan approach
has also been proposed to identify causal variants
in the whole-genome scale without any presump-
tion based on prior knowledge, as has been success-
fully applied to isolate causal polymorphisms in a
variety of common human diseases. This approach
mandates stringent study designs, adequate sample
size, and statistical power. In addition, high-power
computational methodologies of data analysis and
error shooting should be developed to probe the vast
amount of interactions amongst genetic and envi-
ronmental factors, and molecular function assays
should be carried out to determine the genotype–
phenotype correlations and validate the biological
significance of the identified high risk alleles.