44 / 220
intervals (CIs) were calculated using the DerSimonian and
Laird random effects model because of anticipated heteroge-
neity. Random effects modeling takes into account both
within-study and between-study variation. To correct for any
continuity errors, 0.5 was added to all cells with a frequency
of 0 in order to calculate the pooled estimates.
Summary receiver operating characteristics (SROC)
curves were fitted using the Moses-Shapiro-Littenberg
method, and the area under the curve (AUC), Q
*
index, and
their respective standard errors were estimated.
The Spearman’s correlation coefficient was calculated to
assess for threshold effect, and variability between individual
studies was evaluated by plotting the diagnostic accuracy
estimates on a forest plot. Heterogeneity was quantified using
the
I
2
index. Potential heterogeneity between individual stud-
ies was explored using single-factor meta-regression with the
following covariates: sample size, QUADAS score, site of
initial tumor, imaging type, timing of posttreatment scan,
method of image interpretation (visual vs semiquantitative or
quantitative), and clinical presentation of recurrence (sympto-
matic vs asymptomatic or not reported). Covariates were con-
sidered to be explanatory for the heterogeneity if the
regression coefficients were statistically significant (
P
\
.05).
Publication bias was quantified using the Egger’s regres-
sion model, with the effect of bias assessed using the fail-
safe number and trim-and-fill method. The fail-safe number
was the number of studies that we would need to have
missed for our observed result to be nullified to statistical
nonsignificance at the
P
\
.05 level. Publication bias is
generally regarded as a concern if the fail-safe number is
less than 5
n
1
10, with
n
being the number of studies
included in the meta-analysis.
The impact of imaging modality, method of image inter-
pretation, and timing of scan on sensitivity and specificity
separately was also assessed using subgroup analysis, and a
Z test was performed to determine the statistical differences
between subgroups.
Statistical analyses were performed using Meta-Disc
(version 1.4, Unit of Clinical Biostatics, Ramon y Cajal
Hospital, Madrid, Spain), GraphPad Prism (version 6.0,
GraphPad Software, San Diego, CA), and Microsoft Excel
(version 14.2.0, Microsoft, 2011).
Results
Study Selection
The search strategy identified 3411 citations, of which 312
abstracts were considered relevant. Based on the predeter-
mined selection criteria, 150 full-text articles were evalu-
ated, and 27 studies met our inclusion criteria and provided
test accuracy data (
Table 1
;
Figure 1
).
Study Characteristics
There were a total of 1195 patients in the 27 selected stud-
ies, with the number of patients in each study varying from
12 to 98. The time from treatment to imaging ranged from 2
to 260 weeks. The timing or duration of follow-up was
noted in 23 studies and ranged from 6 to 86 months.
Twenty-two studies reported on the diagnostic accuracy of
FDG-PET, while 5 studies reported on the use of FDG-PET/
CT. Scans were assessed qualitatively in 13 studies and
semiquantitatively in 10 studies, with a specific cutoff value
reported in 3 studies; 4 studies did not specify whether
scans were interpreted visually or semiquantitatively.
The vast majority of studies included SCCs from a variety
of locations on the head and neck; 1 study
23
reported specifi-
cally on oral cancer and 1 study
37
on nasopharyngeal cancers.
Fourteen studies reported on the stage of the initial tumor,
with 8 of these studies
17,20,26,27,30,32,34,35
specifically enrolling
patients with stage III or IV head and neck cancers. Six stud-
ies
16,19,21,29,30,35
included only patients in whom there was no
evidence of distant metastases at initial diagnosis, another 4
studies
13,15,20,37
did not have such an inclusion criterion, but
the study population consisted only of patients in whom dis-
tant metastases were not present initially, and 4 stud-
ies
18,22,23,36
included at least 1 patient in whom distant
metastases were detected at the initial diagnosis.
Treatment involved radiotherapy without chemotherapy in
3 studies,
13,19,34
radiotherapy with chemotherapy in 9 stud-
ies,
*
radiotherapy with and without chemotherapy in 8 stud-
ies,
11,12,18,21,27,32,33,35
and intra-arterial chemotherapy in 3
studies.
20,23,24
The remaining 4 studies
14-16,25
included at
least 1 patient either who underwent radiotherapy postopera-
tively or in whom neck dissection was performed in addition
to radiotherapy. We could not meaningfully compare the
diagnostic accuracy of using PET to detect residual/recurrent
disease after radiotherapy alone versus radiotherapy with che-
motherapy, as there were insufficient studies once we consid-
ered primary site and neck recurrences separately.
Only 1 study
19
specified that the study population was
clinically asymptomatic for disease. Four studies
24,25,31,32
recruited clinically symptomatic patients or patients with
suspected recurrence, 1 study
13
noted that at least some of
the patients in the study population were symptomatic,
while 21 studies did not report on the patient’s clinical pre-
sentation at recurrence.
Publication Bias
The primary and nodal groups were assessed for publication
bias using an Egger’s regression model; no publication was
observed for primary sites (
P
= .48). However, publication
bias was detected for nodal sites (
P
= .006), with the fail-
safe number being 1445 studies. Given the comprehensive
literature search strategy used, we feel it is extremely
unlikely that this large number of studies was missed.
Quality Assessment of Studies
The QUADAS score ranged from 10 to 13 out of a maximum
of 14, with a median of 11.5. Most papers scored well on the
items relating to variability and reporting. However, the scores
for presence of bias were more variable. Only 4 studies
23,26,34,35
reported that all patients received the same reference test
*
References 17, 19, 22, 26, 28, 30, 31, 36, 37.
Cheung et al
24