| Modality | N_sens | N_spec |
|---|---|---|
| Ultrasound | 22 | 7 |
| Mammography | 18 | 3 |
| MRI | 7 | 2 |
Meta analysis of breast cancer imaging in pregnant and lactating patients
Statistical analysis
As descriptive analysis, sensitivity and/or specificity of each primary study, along with 95% confidence intervals (CI), were calculated, provided the study has corresponding data. For studies with both types of data, the diagnostic odds ratio (DOR) was also calculated. Meta analysis was performed on studies with complete data using the recommended bivariate modeling approach (Reitsma et al. 2005), with the R-package mada (Sousa-Pinto 2022). Summary receiver operating characteristic (SROC) analysis was used to combine these studies to generate a summary area under the curve (AUC). All analyses were performed in R version 4.3.1 (R Foundation for Statistical Computing, Vienna, Austria).
Note: another approach is to use multiple imputations to fill in the missing data (mainly for specificity).
Results on available data
The numbers of primary studies with sensitivity or specificity data are tabulated below. For ultrasound, mammography, and MRI, meta analysis will be based on the 7, 3, and 2 studies with complete data, respectively.
Overview
Ultrasound
Characteristics of all primary studies on ultrasound are summarized below.
| Study | Sensitivity (95% CI) | Specificity (95% CI) | DOR (95% CI) |
|---|---|---|---|
| Ahn | 0.975 (0.8, 0.997) | ||
| Bock | 0.5 (0.246, 0.754) | 0.875 (0.669, 0.96) | 7 (1.2, 41.3) |
| Chung | 0.917 (0.517, 0.991) | 0.702 (0.628, 0.767) | 26 (1.4, 478.8) |
| Cordoba | 0.935 (0.76, 0.985) | ||
| Espinosa | 0.917 (0.517, 0.991) | ||
| Haliloglu | 0.875 (0.396, 0.987) | 0.98 (0.918, 0.995) | 343 (11.7, 10028.5) |
| Jafari | 0.984 (0.867, 0.998) | ||
| Langer | 0.767 (0.659, 0.848) | ||
| Liberman | 0.929 (0.561, 0.992) | ||
| Myers | 0.985 (0.874, 0.998) | ||
| Nishanova | 0.855 (0.692, 0.939) | ||
| Obenauer | 0.5 (0.17, 0.83) | 0.979 (0.828, 0.998) | 47 (1.7, 1280) |
| Oh | 0.95 (0.655, 0.995) | ||
| Qian | 0.971 (0.884, 0.993) | 0.593 (0.515, 0.667) | 49 (9.4, 256.6) |
| Reyes | 0.956 (0.829, 0.99) | ||
| Robbins | 0.9 (0.463, 0.989) | 0.853 (0.778, 0.905) | 52.2 (2.7, 1012.9) |
| Son | 0.929 (0.561, 0.992) | 0.783 (0.609, 0.894) | 47 (2.3, 948.5) |
| Taron | 0.975 (0.8, 0.997) | ||
| Taskin | 0.99 (0.908, 0.999) | ||
| Taylor | 0.977 (0.815, 0.998) | ||
| Wang | 0.857 (0.786, 0.907) | ||
| Yang | 0.977 (0.815, 0.998) |
Summary ROC analysis combining the 7 primary studies with complete data is presented below.
Mammography
Characteristics of all primary studies on mammography are summarized below.
| Study | Sensitivity | Specificity | DOR |
|---|---|---|---|
| Ahn | 0.844 (0.604, 0.95) | ||
| Bock | 0.773 (0.478, 0.927) | 0.962 (0.717, 0.996) | 85 (3.6, 2001.3) |
| Cordoba | 0.66 (0.464, 0.813) | ||
| Espinosa | 0.917 (0.517, 0.991) | ||
| Langer | 0.806 (0.712, 0.874) | ||
| Liberman | 0.771 (0.573, 0.894) | ||
| Myers | 0.897 (0.752, 0.962) | ||
| Nishanova | 0.79 (0.619, 0.897) | ||
| Obenauer | 0.5 (0.17, 0.83) | 0.967 (0.747, 0.997) | 29 (1, 802) |
| Oh | 0.9 (0.463, 0.989) | ||
| Reyes | 0.761 (0.618, 0.863) | ||
| Robbins | 0.9 (0.463, 0.989) | 0.921 (0.842, 0.962) | 104.5 (5.1, 2162.8) |
| Samuels | 0.611 (0.309, 0.847) | ||
| Taron | 0.625 (0.409, 0.8) | ||
| Taskin | 0.865 (0.741, 0.935) | ||
| Taylor | 0.725 (0.506, 0.872) | ||
| Wang | 0.827 (0.698, 0.908) | ||
| Yang | 0.881 (0.682, 0.962) |
Summary ROC analysis combining the 3 primary studies with complete data is presented below.
MRI
Characteristics of all primary studies on MRI are summarized below.
| Study | Sensitivity | Specificity | DOR |
|---|---|---|---|
| Espinosa | 0.917 (0.517, 0.991) | 0.833 (0.31, 0.982) | 55 (0.8, 3650.7) |
| Myers | 0.972 (0.888, 0.994) | ||
| Obenauer | 0.9 (0.463, 0.989) | 0.917 (0.517, 0.991) | 99 (1.6, 6052.7) |
| Oh | 0.95 (0.655, 0.995) | ||
| Taron | 0.975 (0.8, 0.997) | ||
| Taskin | 0.974 (0.791, 0.997) | ||
| Taylor | 0.929 (0.561, 0.992) |
Summary ROC analysis combining the 2 primary studies with complete data is presented below.
Summary and publication bias
Overall sensitivity and specificity
| Modality | TP | FN | FP | TN | Sensitivity | Specificity |
|---|---|---|---|---|---|---|
| Ultrasound | 532 | 49 | 137 | 443 | 0.916 | 0.764 |
| Mammography | 359 | 86 | 6 | 101 | 0.807 | 0.944 |
| MRI | 113 | 1 | 0 | 7 | 0.991 | 1 |
Funnel plot and publication bias
For each modality, the following shows the funnel plot for study-specific log sensitivity/specificity odds centered by the pooled estimates in Table 4 (hence log-odds ratio) vs their standard errors. The Egger et al. (1997) test for asymmetry suggests a potential bias for larger sensitivities/specificities in ultrasound (skewed to the right; P = 0.001). There is no visible bias in mammography (P = 0.376) or MR (P = 0.152).
Other summaries combining sensitivity and specificity
- Positive likelihood ratio (LR+): probability of a case testing positive over that of a non-case testing positive, i.e., sensitivity/(1-specificity);
- Negative likelihood ratio (LR-): probability of a case testing negative over that of a non-case testing negative, i.e., (1-sensitivity)/specificity;
- Diagnostic odds ratio (DOR = LR+/LR-).
These are tabulated below.
| Modality | N | LR+ | LR- | 1/LR- | DOR |
|---|---|---|---|---|---|
| Ultrasound | 7 | 5.14 (3.09, 8.87) | 0.22 (0.08, 0.5) | 4.49 (2.02, 12.4) | 23.8 (9.48, 59.9) |
| Mammography | 3 | 9.83 (4.62, 20.9) | 0.31 (0.13, 0.57) | 3.25 (1.75, 7.44) | 32.8 (9.27, 117) |
| MRI | 2 | 7.41 (1.61, 54.7) | 0.11 (0.01, 0.56) | 9.12 (1.79, 67.2) | 74.9 (4, 1410) |
Results of imputation analysis
As an alternative to the above complete-case (CC) analysis, we use multiple imputations (MI) to fill in the missing FP and TN data for ultrasound and mammography (MR has too few complete data points to generate reliable imputations). Specifically, a two-level model is used to first generate the prevalence, which leads to imputation of the total (FP + TN), and then specificity TN / (FP + TN), with study-specific random effects to account for the between-study heterogeneity (Mao 2024). A total of \(J=500\) imputed samples are generated to compute the average of summary AUC, sensitivity, and specificity (Reitsma et al. 2005), along with their uncertainties arising from the missing data. The results are plotted below (gray dots represent one instance of imputed data points out of \(J=500\)).
The MI summary AUC for ultrasound (0.929) is higher than the CC result (0.813). Recall the latter is based on only 7 studies which happen to include two with rather low sensitivity (0.5). This likely makes it conservative. For mammography the results are more similar between CC and MI. In general, summary sensitivity and specificity under MI are better aligned with the overall sensitivity and specificity, as MI allows all studies to be included.
References
Egger, M., G. D. Smith, M. Schneider, and C. Minder. 1997. “Bias in Meta-Analysis Detected by a Simple, Graphical Test.” BMJ 315 (7109): 629–34. https://doi.org/10.1136/bmj.315.7109.629.
Mao, Lu. 2024. “Mtroc: Meta Analysis of Diagnostic Accuracy with Missing Data.” https://github.com/lmaowisc/mtroc.
Reitsma, Johannes B., Afina S. Glas, Anne W. S. Rutjes, Rob J. P. M. Scholten, Patrick M. Bossuyt, and Aeilko H. Zwinderman. 2005. “Bivariate Analysis of Sensitivity and Specificity Produces Informative Summary Measures in Diagnostic Reviews.” Journal of Clinical Epidemiology 58 (10): 982–90. https://doi.org/10.1016/j.jclinepi.2005.02.022.
Sousa-Pinto, Philipp Doebler with contributions from Bernardo. 2022. “Mada: Meta-Analysis of Diagnostic Accuracy.” https://CRAN.R-project.org/package=mada.