This systematic review adheres to the guidelines established in the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) statement [14]. The literature search, initiated on December 5, 2023, spanned four major databases: PubMed, Scopus, Web of Science, and Embase. Distinct search terms tailored for each database included (“nasal” OR “midfacial”) AND (“fracture*”) AND (“sonograph*” OR “ultrasonograph*” OR “ultrasound” OR “POCUS”) AND (“pediatric*” OR “paediatric*” OR “child*” OR “neonat*” OR “infant*” OR “toddler*” OR “preschool” OR “pre-school” OR “juvenile” OR “young adult*”). Additionally, a meticulous manual examination of references within the selected studies ensured comprehensive coverage. The review process involved a thorough evaluation of each article’s title, abstract, and/or full text, conducted independently by two co-authors. Uncertainties or ambiguities were addressed through consultation with a senior co-author. Deduplication, screening, and data extraction were facilitated by the AutoLit platform, developed by Nested Knowledge in St. Paul, Minnesota, USA.

All studies relevant to the diagnostic accuracy of ultrasound in pediatric patients (under 21 years old) were considered for inclusion if they presented at least one of the following diagnostic measures: sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), likelihood ratio (LR), diagnostic odds ratio (DOR), and area under the receiver operating characteristic curve (AUC). No restrictions were imposed on publication date, country of origin, patient characteristics, reference standard type, or study design. Non-English literature, case reports, case series with fewer than five eligible patients, conference abstracts, editorial comments, and review articles were excluded from the study.

Details extracted from each qualifying paper included the first author’s name, publication year, study design, sample size, participant and fracture characteristics, reference standard modality, ultrasound operator, type of ultrasound device, type of ultrasound probe, image acquisition methods, ultrasound features indicating nasal bone fractures, and diagnostic accuracy measures of ultrasound.

The quality assessment utilized the Diagnostic Accuracy Studies-2 (QUADAS-2) tool to evaluate included studies’ quality [15]. Independent assessments for potential bias and concerns regarding applicability were conducted for the four primary domains of the QUADAS-2 tool: patient selection, index test, reference standard, and flow and timing. Specific criteria outlined in the tool, such as the representativeness of the study population, blinding of test results, and completeness of outcome data, informed evaluations for each domain. Ratings of “low,” “high,” or “unclear” were assigned to determine the overall rigor and reliability of the evidence synthesis.

The analysis employed a random effects diagnostic test accuracy (DTA) model, specifically utilizing the bivariate model proposed by Reitsma et al. [16]. Summary Receiver Operating Characteristic (SROC) curves were generated based on the bivariate meta-analysis data, with study-specific estimates weighted in the random effects univariate DOR model. The AUC and its confidence interval (CI) for each subgroup were determined through 2000 sample bootstrapping, utilizing the bivariate model [17].

Heterogeneity assessment relied on the I2 metric, following the method outlined by Holling et al. [18], where an I2 CI exceeding 50% signified significant heterogeneity. Sensitivity analyses using the DOR univariate meta-analysis were conducted to identify potential outliers in the presence of significant heterogeneity. If outliers were identified, a re-analysis was performed to validate the results. The clinical relevance of the findings was explored using Fagan plots and likelihood ratio scattergrams, where positive likelihood ratios above 10 indicated confirmation suitability, and negative ratios below 0.1 suggested suitability for exclusion. Fagan nomograms were constructed for assumed pre-test prevalences of 25%, 50%, and 75%, based on the bivariate Reitsma model, as detailed by Zwinderman et al. [19].

All analyses were executed in R (version 4.3.2, R Foundation for Statistical Computing, Vienna, Austria), utilizing packages such as “mada,” “dmetatools” [20], “Metafor” [21], and “meta” [22].