aꢀꢀꢁ ꢂt ꢃꢄ.
in the literature8,12 to be associated with survival, including
NLR. Secondary aims were to compare our model with the
MSKCC and Heng models and to analyze survival in the
subset of patients on first-line sunitinib therapy.
were drawn with replacement from the observed data. The
model building strategy described above was conducted on
all 500 bootstrapped data sets. Finally, the frequency with
which each predictor variable appeared in the final model
from all 500 samples was counted. Variables appearing in
>
50% of the models were retained. In the next step, an
Methods
additional 500 bootstrap samples were obtained. With each
sample, a Cox proportional hazards model was fit using the
retained variables from the first step. Using the results from
the 500 estimated models, mean parameter estimates, haz-
ard ratios, and confidence intervals (CIs) were calculated.
A risk group variable was created by summing the num-
ber of risk factors each patient had from the final model.
The area under the receiver operating characteristic curve
(ROC) was used to determine the predictive accuracy of the
risk group variable. Additional risk group variables were
Patients
Retrospective data were obtained from the medical records
of all patients diagnosed with mRCC at two Canadian cen-
tres, from July 2007 until December 2011. The inclusion
9
criteria of Heng and colleagues were used.
We recorded basic demographic, survival, and clinical
data, including all variables that have been shown in the
literature to be important prognostic factors in mRCC. We
8
,12
calculated using prognostic models from the literature.
1
9
used an NLR categorization from the literature (≤3 vs. >3).
The predictive accuracy of the risk group variables from dif-
ferent prognostic models was assessed with the area under
the ROC. Kaplan Meier curves for each of the different risk
group variable were calculated.
We retrospectively reviewed data from 120 patients, and
included the 89 patients who received at least one cycle of
active treatment; 31 patients did not receive treatment and
were excluded from the analysis. Approval for this study
was obtained from the Horizon Health Network Research
Ethics Board.
All analyses were conducted using R version 3.0.1.
Results
Statistical analysis
Patient characteristics, treatment, and survival
All patients who started first-line treatment, mostly with
sunitinib, were included in the analysis. The main outcome,
overall survival, was defined as time from treatment initia-
tion until death, otherwise censored at last their last followup
or contact. The survival distribution and median survival
were assessed via Kaplan Meier estimates. Univariate asso-
ciations between overall survival and baseline demographic
and clinical factors were examined. Significance was taken
at p<0.05. Log-rank tests were used to test the presence of
a significant difference between survival among categor-
ical variables and overall survival. For prognostic purposes,
during the model-building phase all continuous variables
were dichotomized at the upper or lower levels of normal
except for age, which was dichotomized at >65 years and
Baseline characteristics are provided in Table 1. Eighty-nine
patients were treated for mRCC during the study period.
First-line treatments included sunitinib, being the most com-
mon (n=71, 79.8%), followed by interferon alpha (n=14,
15.7%), pazopanib (n=2, 2.2%), sorafenib (n=1, 1.1%), and
temsirolimus (n=1, 1.1%). Mean patient age was 63.1 years
(standard deviation [SD] 9.9 years, range 38‒88). Thirty-
eight out of the 89 patients (42.7%) were metastatic at diag-
nosis. There were 17 patients who did not undergo nephrec-
tomy. Of these 17 patients, seven were due to comorbidities;
the remaining 10 patients did not have nephrectomy for
unknown reasons.
At a median followup of 24.6 months (95% CI 19.2, 39.7)
for the entire cohort, the median overall survival was 20.9
months (95% CI 14.2–50.6) (Fig. 1). By the end of the study
period, 44 patients (49.4%) had died. One-year survival was
63.7% (95% CI 0.54–0.76).
≤
65 years.
A multivariate Cox proportional hazards regression model
was constructed using a stepwise procedure to identify the
most significant variables affecting the disease-free sur -
vival. The stepwise algorithm used the Aikake Information
Criterion (AIC), a penalized likelihood measure, to choose a
final model. When comparing two models, the model with a
lower AIC value more closely resembles reality. The propor-
tional hazards assumption of the final model was examined
with a global test of proportionality. The test examines cor-
relations between model residuals and log (time).
Univariate analysis
Factors that were significantly associated with poor overall
survival were age >65 years, absence of prior nephrectomy,
non-clear-cell histology, presence of two or more metastatic
sites, presence of brain metastases, time interval from diag-
nosis to treatment of <1 year, and hemoglobin below the
lower limit of normal (Table 2).
The internal validity of the final model was examined with
a two-step bootstrap procedure. In the first step, 500 samples
1
14
CUAJ • March-April 2016 • Volume 10, Issues 3-4