Types of studies

We included:

1. either prospective or retrospective diagnostic cohort studies that enrolled patients with blunt trauma who:

   1. underwent any type of POCS as primary imaging modality to screen for thoracoabdominal injuries; and

   2. also underwent predefined imaging or invasive reference tests to verify POCS results;

2. studies that provided 2 x 2 tables (or sufficient information to tabulate results) to allow for calculating sensitivity, specificity, and other indices of diagnostic test accuracy.

We excluded:

1. diagnostic case‐control studies comparing patients with known case status to healthy controls, as this creates artificial populations and tends to overestimate sensitivity of the index test;

2. case series and case reports;

3. studies with unclear index or reference tests; and

4. studies that did not allow for creating 2 x 2 tables.

Participants

The target population of this review comprised people of any age or gender who sustained any type of blunt trauma in a civilian scenario and were transferred to a hospital of any care level. Also, in order to be eligible participants had to have undergone POCS as the primary imaging tool and to have been followed up either as inpatients or outpatients with different diagnostic modalities to verify whether the condition of interest was present or absent.

Because of clear differences in clinical management, we deliberately excluded people with penetrating injuries, as well as members of armed forces wounded in the battlefield.

Index tests

Any type of POCS performed in a trauma setting (e.g. FAST ultrasonography of the abdomen or thorax, or both, or any advanced ultrasound protocol) intended to detect:

1. free fluid (as a surrogate of bleeding) in the abdomen, retroperitoneal space, or chest;

2. injuries to solid organs such as the liver or spleen (including attempts to grade their severity);

3. lesions of major vessels; and

4. other injuries (e.g. pneumothorax, as indicated by air in the pleural space).

Variation in POCS technology and application (e.g. specification of ultrasound machines and probes and how up‐to‐date they were, and handling of inconclusive test results) is addressed in the Assessment of methodological quality section of the review. We planned to examine its potential influence on diagnostic accuracy estimates in the Investigations of heterogeneity section of the review.

Target conditions

This review focused on blunt thoracoabdominal and multiple trauma, meaning any blunt, non‐penetrating force to the abdomen and chest and both solid and hollow viscera, as well as both major vessels. Target conditions considered by this review included:

1. free fluid in the:

   1. thoracic cavity (uni‐ or bilateral, where specified);

   2. abdominal cavity (by abdominal quadrant, where specified);

   3. retroperitoneal space;

   4. pericardium; or

   5. mediastinum;

2. organ injuries, defined as:

   1. liver injuries (e.g. capsular tears, haematoma, tissue lacerations);

   2. splenic injuries (e.g. capsular tears, haematoma, tissue lacerations);

   3. injuries to other solid organs (e.g. pancreas, kidneys);

   4. injuries to hollow viscera; or

   5. any other organ laceration detected by ultrasonography;

3. vascular lesions, defined as:

   1. dissection or rupture of the thoracic or abdominal aorta, or both;

   2. rupture of other vessels such as the iliac arteries;

4. other injuries (e.g. pneumothorax, as indicated by air in the pleural space in the thoracic cavity).

We analysed the effect of different types of target conditions as part of our Investigations of heterogeneity. We categorised target conditions into surrogates of blunt trauma (i.e. free fluid and free air, named limited assessment), and both surrogates and direct signs of organ damage (i.e. organ injuries and vascular lesions, named complete assessment).

Reference standards

In order to be accepted as a diagnostic reference standard, the deliberate use (and the reasoning for its use) of the particular method needed to be specified. To avoid verification bias, all participants were required to undergo an independent imaging or invasive test, regardless of the initial POCS scan.

We classified the following tests as reference standards to confirm the presence or absence of the target condition:

1. any type of CT scan of the major body cavities (i.e. chest, abdomen, pelvis), either selective or performed as a whole‐body scan. We planned to stratify results for the use of intravenous or oral contrast agents, or both, and the time interval between POCS and CT;

2. any type of MRI of the major body cavities;

3. laparotomy (by a median or transverse approach), or laparoscopy, either diagnostic or therapeutic;

4. thoracotomy (by median sternotomy or a clamshell approach), or thoracoscopy, either diagnostic or therapeutic;

5. autopsy, either done by pathologists or forensic examiners.

Electronic searches

We searched the following electronic sources.

1. Ovid MEDLINE (1946 to 15 July 2017).

2. PubMed (not MEDLINE) (1947 to 15 July 2017).

3. Ovid Embase (1974 to 15 July 2017).

Search strategies are shown in Appendix 1.

We performed a further search on 6 December 2018; details of the eight potentially relevant studies identified have been added to the Characteristics of studies awaiting classification and Studies awaiting classification sections, and may be incorporated into the review at the next update.

Searching other resources

A systematic review by Scherer and colleagues showed that results from studies that have not been published in a full‐text format are systematically different from fully published results (Scherer 2007). We therefore searched the BIOSIS database for conference abstracts to identify potentially relevant studies that had not yet been published in a journal format (see Appendix 1).

We planned to contact authors of individual studies by email, letter, or phone, if we considered their results to be important but needed further explanation or raw data. We guaranteed that any data exchange complied with the International Conference on Harmonisation Good Clinical Practice (ICH‐GCP) principles and rules and regulations of data safety and security.

Selection of studies

Two review authors (AH, JL) independently screened the titles and abstracts of the identified reports, documenting details of selected studies in a predefined electronic spreadsheet and assessing studies for eligibility in terms of the predefined inclusion and exclusion criteria. If it was not possible to make a decision based on title and abstract alone, the full texts of potentially relevant studies were assessed. Any disagreements between authors regarding the selection of studies were resolved by a third expert (DS). The study selection process is documented in a detailed flow chart (Figure 1).

Figure 1

---

---

Study flow diagram for the search conducted on 15 July 2017.

Data extraction and management

As stated above, we established an SOP for data extraction for systematic reviews, meta‐analyses, and health technology assessment (HTA) reports. We adhered to ICH‐GCP, Good Epidemiological Practice (GEP), and other relevant rules and recommendations. We have trained personnel on site to record, manage, and audit data, and our data storage modes comply with federal legislation on data safety for research purposes. Two review authors (AH, JL) independently extracted data from original papers in duplicate, and resolved discrepancies by discussion, moderated by a third review author (DS). They extracted the following information from published papers:

1. study characteristics (author, year of study, year of publication, journal reference, study design, inclusion/exclusion criteria, operator characteristics, hardware specifications, index test used, reference test used, general setting (urban/rural), mass casualty (yes/no));

2. patient characteristics (age, gender, type of trauma, type of injury, injury severity, haemodynamic stability, probability of survival);

3. outcome of the index test as assessed in the individual studies by diagnosing the target condition and, if available, the number of participants with inconclusive results or who had no test result;

4. diagnostic 2 x 2 tables, cross‐classifying the disease status on the basis of the reference test (i.e. number of true‐positive, false‐positive, false‐negative, and true‐negative results).

Diagnostic accuracy was expressed by individual and pooled indicators such as sensitivity and specificity with 95% confidence intervals (CI), positive and negative likelihood ratios (LR), positive predictive value (PPV), negative predictive value (NPV), and the summary receiver operating characteristic curve (SROC).

Assessment of methodological quality

Two review authors (AH, JL) independently used the QUADAS‐2 tool (the revised and updated version of the original Quality Assessment of Diagnostic Accuracy Studies list of items) to assess methodological quality of individual studies (Whiting 2011). Any discrepancies were resolved by discussion, moderated by a third review author (DS). QUADAS‐2 includes four main domains, namely: patient selection, index test, reference standard, and flow and timing. We assessed each domain with regard to risk of bias, rating them as 'low', 'high', or 'unclear'. We assessed concerns regarding applicability only for the first three domains, categorising them as 'low', 'high', or 'unclear'. Signalling questions were answered as 'yes', 'no', or 'unclear' (see also Appendix 2). By using tailored review‐specific signalling questions we were able to perform a custom‐made assessment of the methodological quality of all included studies. We omitted the signalling question "Was any case‐control design avoided?" in Domain 1: Patient selection since we did not include any case control study, case series, or case reports. We added three signalling questions to Domain 2: Index test and two signalling questions to Domain 3: Reference test referring to operator's expertise and background, technical features of the hardware, and appropriateness of the ultrasound protocol and reference imaging standard. In Domain 4: Flow and timing, we included the signalling question "Did all participants receive a reference standard?" in order to explore the risk of partial verification bias.

Statistical analysis and data synthesis

If at least one of the target conditions was detected (i.e. pneumothorax, free fluid, organ or vessel injury), we considered the patient (participant) to be traumatised or test‐positive. Otherwise, we considered the participant to be uninjured or test‐negative. Our observational unit of interest was thus the individual participant, not a particular injury, and we did not evaluate single target conditions separately in the primary analysis. We used inconclusive test results as reported in the primary studies. For individual studies, we calculated sensitivity and specificity with 95% CI, tabulated pairs of sensitivity and specificity with CI, and depicted estimates by coupled forest plots using Review Manager 5 (RevMan 5) (Review Manager 2014).

Due to the subjective nature of the interpretation of POCS findings, we expected implicit thresholds in test positivity. We assessed a possible threshold effect visually by plotting true‐positive rates (sensitivity) from each study against false‐positive rates (1 − specificity) in a receiver operating characteristic (ROC) space and coupled forest plots of sensitivity and specificity. Since the dichotomous operationalisation of the test result does not enable explicit thresholds, we used the bivariate model according to Reitsma 2005, which is a robust statistical model taking the underlying relationship between sensitivity and specificity into account. The random‐effects approach allows for calculating sensitivity and specificity estimates while controlling for heterogeneity across studies. We fitted the models in Stata (Stata 2017) using the 'metandi' command and produced SROC plots using RevMan 5. We estimated average sensitivities and specificities using the bivariate model. We obtained likelihood ratios post estimation using the parameters of the bivariate model (see summary of findings Table 1).

Investigations of heterogeneity

We assessed heterogeneity visually by inspecting the coupled forest plots and plots of study results in the SROC space. We also investigated possible sources of heterogeneity by adding single covariates to the basic bivariate random‐effects model. We conducted fitting of the bivariate model via the 'xtmelogit' command in Stata (Takwoingi 2016). We investigated the effect of adding covariates by conducting a likelihood ratio test that compared the ‐2 log likelihoods of the basic bivariate model to a model including a single covariate. If a significant reduction in the ‐2 log likelihood was detected (indicated by a P value of < 0.05), test performance was considered to be associated with the particular covariate. For statistically significant test results, we determined whether the covariate was associated with the estimated sensitivity, specificity, or both (Macaskill 2011), by removing the covariate terms for either sensitivity or specificity, and comparing the fit of each alternative model using likelihood ratio tests.

For tests of heterogeneity, we required a minimum of 10 studies in total and at least four studies per subgroup. We had to dichotomise the covariates we investigated, and differentiate between paediatric and non‐paediatric (i.e. adult/mixed populations), surrogates of injury (e.g. air, free fluid) and organ lacerations, abdominal injuries (i.e. injuries exclusively located in the abdomen) or chest injuries (i.e. injuries exclusively located in the chest), and single CT versus CT plus laparotomy used as reference standard (see summary of findings Table 2).

Sensitivity analyses

We performed sensitivity analyses to investigate how individual QUADAS‐2 key domains (i.e. patient selection, index test, reference standard, and flow and timing) affected accuracy estimates, and to explore whether the different evaluations of the two independent review authors within two original studies influenced pooled sensitivities or specificities, or both, of the index test. Moreover, we examined the impact of participants' age on accuracy estimates by only including paediatric studies.

Assessment of reporting bias

We did not assess reporting bias because there are no accepted ways of doing this for diagnostic test accuracy studies (Deeks 2005).