Statistical Analysis
Survival analysis was performed by transplant type (single or tandem) at
each mIBG scan time point (diagnosis or EOI). The primary endpoint was
an EFS event; death was a secondary endpoint.
An optimum cut point for total
Curie score was determined for each combination (transplant type and
time of mIBG scan) by maximizing the Youden index with respect to how
well the Curie score differentiated patients who did and did not have an
EFS event. The Youden index represents the maximum value (sensitivity +
specificity – 1) over all CS threshold values.19Event-free survival (EFS) time was calculated with two different
starting points, from enrollment or from the EOI mIBG scan, until the
occurrence of an event; or, if no event, until the date of last
follow-up. Events included death from any cause, disease relapse or
progression, or the development of a secondary malignancy. Overall
survival (OS) time was calculated with two different starting points,
from enrollment, or from the EOI scan to death from any cause or, if the
patient was still alive, the date of last follow-up. Three-year EFS and
OS estimates,20 with standard errors per Peto et
al.,21 were computed for all patients for each
variable investigated at the corresponding time point and transplant
group. Log-rank tests were performed to compare EFS and OS, with a
p-value <0.05 considered statistically significant. The data
used for this analysis were current as of June 30, 2020.
Several variables were investigated for their effect on outcome,
including an optimal CS cut point from the initial diagnostic and EOI
mIBG scans. Outcomes based on a relative reduction in Curie scores from
diagnosis to EOI were determined, using reductions of 50% and 75%
as in previous analyses.
To determine the independent prognostic strength of the Curie score for
survival in the presence of various prognostic factors, including age
(<18 vs. ≥18 months), MYCN status (non-amplified vs.
amplified), end-induction response (CR/VGPR vs. PR vs. MR/NR), and
immunotherapy (yes vs. no), Cox proportional hazards (PH) models with
the Efron method of handling tied event times were fit. Tests for
violations of the PH assumption were performed. Backward selection was
used to determine the most parsimonious
model, with a threshold
p-value<0.05 to remain in the model.