A systematic review of tests for lymph node status in primary endometrial cancer

Our search attempted to capture all the studies that reported the diagnostic accuracy of sentinel node biopsy, positron emission tomography (PET), magnetic resonance imaging (MRI), computer tomography (CT) and ultrasound scanning for the detection of lymphatic spread in primary endometrial cancer. Bibliographic databases MEDLINE (1966–2006), EMBASE, Cochrane Library (issue II, 2006) and MEDION (1980–2006) were searched without language restrictions. The search strategy used relevant medical subheadings (MeSH), text words and word variants for endometrial cancer and combined these with the terms for the index tests and lymphadenopathy (see Additional file 1). Hand searches of reference lists from primary articles and other reviews were carried out to identify manuscripts missed by electronic searching. Experts in the field were contacted for unpublished studies and conference abstracts were reviewed.

The selection of studies involved a two-stage process and two reviewers (TJS, CHM). The electronic searches were examined and complete manuscripts of potentially relevant citations retrieved for a final decision on inclusion based on pre-defined selection criteria. Studies were selected if they reported accuracy of the index tests, compared to histological examination of the lymph nodes (reference standard) in women with a primary presentation of endometrial cancer of any histological type or stage and allowed data extraction to create two by two tables. No language restrictions were applied. In cases of duplicate publications the most recent manuscript was selected. Final inclusion or exclusion was decided after examining the complete manuscripts. All were examined in duplicate by the two reviewers with any discrepancies resolved by a third reviewer (KSK).

A piloted data extraction form was used to collect information on study characteristics, quality and accuracy results from each of the selected manuscripts. The study characteristics extracted were the stage of disease, the index test and reference standard methodology and the setting and date of the study. Accuracy data from the studies were reordered in two by two tables. For the purpose of analysis when a manuscript reported the accuracy of more than one index test, the tests were reported on separately. Non diagnostic test results and a failure to perform the test, such as an inability to detect the sentinel node or inadequate histology were excluded from the two by two tables, but their occurrence was recorded, along with the results from the reference standard in each case, if provided.

All of the manuscripts meeting the selection criteria were assessed for their methodological quality, defined as the confidence that the study design, conduct and analysis minimised biases in the estimation of test accuracy. Existing, well developed tools were used to generate items for our assessment of methodological quality [7‐9], this process was again carried out in duplicate. For the population, consecutive or random recruitment of eligible women in to the study was considered ideal. Convenience sampling, such as arbitrary recruitment or non-consecutive recruitment was deemed inadequate. Prospective recruitment of patients was considered to be associated with potentially a lesser degree of bias than retrospective recruitment. The description of the population was considered ideal if the study clarified the stage of disease and the body mass index of a patient, which can affect the accuracy of techniques. We recorded the stage of disease in accordance with FIGO classification. The reporting of the index test was considered ideal if the study documented the test in sufficient detail to allow replication by other researchers. It was considered important for the time interval between the index test and the reference standard to be described and an interval of four or less weeks was considered suitable [10]. For the reference standard itself, a description of method of histological verification was important and it was considered preferable for the readers of the reference standard to be blind to the index test results. Information on the number of women recruited into the study and those on whom outcome data were known was sought from the manuscripts to examine partial and differential verification. Verification was considered ideal if all women originally enrolled into the study, without legitimate exclusions were included in the data analysis. We examined if withdraws from the study were explained and if uninterpretable results were reported.

The main strengths and weaknesses in respect of each of the above items for all studies included in the systematic review were tabulated. We did not attempt to collapse our assessment of quality into a score, as suggested methods have little validity and may have a tendency to obscure the strengths and weaknesses of a study rather than clarify them.

From the two by two tables, sensitivity (true positive rate) and specificity (true negative rate), along with their exact confidence intervals were computed. These estimates were plotted in a ROC space to evaluate the degree of correlation between these indices. When two by two tables contained zero cells we applied a standard correction of adding 0.5 to all four cells of that table [11].

We anticipated that in common with other diagnostic reviews [10, 12, 13] there would be heterogeneity of results amongst involved studies. We examined heterogeneity visually using forest plots of sensitivity, specificity and LRs and statistically using Cochran Q [14]. The small number of studies did not allow for detailed exploration of reasons for heterogeneity using meta regression techniques. However studies were instead divided into index test type, which in previous reviews has represented a major source of heterogeneity [16] and difference in accuracy were tested for statistical significance.

We used bivariate random-effect meta-analysis [16] to obtain summary estimates of sensitivity and specificity and other derived measures such as positive and negative likelihood ratios (LRs). LRs allow estimation of the probability of lymphatic spread with a specific test result [17‐19]. The bivariate model assumes that logit transformations of sensitivity and specificity are negatively correlated and follow a bivariate normal distribution. This analysis also incorporates the different precision by which sensitivity and specificity have been measured in each study. The model produces random effect estimations for the mean logit sensitivity and specificity with corresponding 95% confidence intervals, it produces also an estimation for the amount of between-study variation for sensitivity and specificity separately, and finally an estimation of the covariance between sensitivity and specificity. Confidence regions in logit-ROC space can be constructed using these estimates. The ellipse in logit-ROC space can be back-transformed to conventional scale, and plotted in ROC space giving a confidence region for the summary operating point.

Meta-DiSc version 1.4 [14] was used for initial analyses and forest plots and the PROC MIXED procedure in SAS version 8.2 for Windows (SAS Institute) was used to fit bivariate models.