Data analysis
We evaluated the diagnostic accuracy for each index test for each individual food. The data was synthetized by tabulating the index test’s sensitivity, specificity, true positives, true negatives, false positives, and false negatives. For allergens with variable allergenic profiles resulting from extensive heating or cooking, separate analysis was conducted for each allergenic configuration. For hen’s egg (HE) protein, the analysis was divided into baked HE, cooked (extensively heated) HE and raw HE. For cow’s milk (CM), it was separated into baked milk and fresh milk.
Where three or more studies for a given combination of index test and food were available, a meta-analyses was performed with a generalized linear mixed model of the binomial family with a logit link. This approach was chosen to perform a random effect estimate of both sensitivity and specificity, accounting for their correlation, computing the pooled sensitivity and specificity and performing the summary receiver operating curves (ROC)[14]. Briefly, every study contributed with its own contingency table for its specific cut-off value (i.e. true positive, true negative, false positive and false negative) were included in the model as a count. These analyses resulted in a bivariate random effect estimation of sensitivity and specificity along with heterogeneity assessed by I-squares defined according to Zhou and Dendukuri, 2014 [15]. We defined tests with high accuracy as those which had a sensitivity or specificity of ≥90% with I-squares under 50%. Low sensitivity and specificity were considered for test performing under 75%.
We performed sensitivity and specificity analysis using the optimal cut-off reported by the individual studies using the optimal cut-off reported by the individual studies, e.g. Youden’s Index or other methods. To obtain the estimated cut-offs used for each meta-analysis, we reported the median and interquartile range of all cut-offs considered optimal by the different authors. Further analyses were performed and focused on the maximum values for sensitivity and specificity as reported by the authors of included studies.
Further analyses were undertaken with the pre-established 95% positive predictive values (PPV) available in literature [16]. For skin prick tests (SPT) we used values of 8 mm for peanut [17] and CM and 7 mm for HE [18]. For sIgE, we used the following values: 15 kUA/L for peanut [17], CM and tree nuts, 7 kUA/L for HE and 20 kUA/L for fish [19, 20]. We included only values which have been previously validated thus this are not available for all foods. [18, 21-23].
As the PPV is dependent on the prevalence of allergic disease in a specific population, we looked at the sensitivity and specificity of pooled data for these cut-offs and defined them as highly accurate if they reached a value ≥ 90%.
In supplementary analyses, studies were stratified by test-specific threshold values, age of the participants (below 24 months, 24 months to 16 years and above 16 years) and by the country of origin. Where data on at least three different tests on the same food were available, a comparison was performed. To this end, the relative ratio of sensitivity and specificity was computed using an intercept only model [24]. Data for differences in subgroups were considered significant if there was a change in sensitivity or specificity over 7% (CI 95%) or they reached high diagnostic accuracy (over 90% of sensitivity or specificity for any given test).
To reduce heterogeneity in the meta-analyses, only index tests using the same characteristics where combined. For SPT, results are shown for studies using commercial extracts separate from those using skin prick to prick tests (SPP) with fresh foods. For sIgE testing, results from different platforms were used individually for meta-analyses (ImmunoCAP Specific IgE, ImmunoCAP™ ISAC, etc). Throughout the manuscript when talking about sIgE this refers to ImmunoCAP, if other methods are used for analysis, it will be specified accordingly. The random effect bivariate meta-analysis was performed using the metadata function of the STATA software version 15.