| To Register SMDM Homepage | |
|
|
|
||||
Purpose: To contrast various meta-analysis techniques for the assessment of the diagnostic test performance of sentinel lymph node biopsy (SNB) following preoperative chemotherapy in patients with breast cancer.
Methods: A systematic review was conducted of studies which examined the results of SNB following preoperative chemotherapy. Inclusion criteria required completion axillary lymph node dissection as a test “gold standard”. Robust resistant regression method was used to construct SROC curves and compared to test results using Bayesian hierarchical models. Two distinct Bayesian models were considered. A beta model using sensitivity derived from each study as random draws from a beta distribution was analyzed using an exponential prior for the beta parameters. Within a study, the observed number of SNB test positives out of the true positives was assumed to be a binomial random variable. The second model, logit model, assumed that the studies were derived from a population of studies in which the log odds of the sensitivity was a normal distribution. Both Bayesian models used prior parameters derived from published data. Sensitivity analyses were performed to examine the effects of prior selection on posterior estimates.
Results: Fourteen studies were identified. The range for reported sensitivity was 61-100%. The specificity was 100% in all studies. Pooling of data resulted in the sensitivity of SNB of 89%. The adjusted parameters using the SROC curve revealed a global sensitivity of 87% (95% confidence interval, 82-93). In the Bayesian analysis, the beta model resulted in a posterior estimate of sensitivity of 83% (95% credible interval, 72-91), while the logit model estimated the sensitivity of SNB at 90% (95% credible interval, 84-94). The logit model showed little sensitivity to prior parameters, while the beta model was more sensitive.
Conclusions: The estimate of sensitivity for SNB following preoperative chemotherapy derived from meta-analysis of published studies varies from 83% to 90% depending on the analytic approach. Model assumptions are important in deriving summary estimates. Both Bayesian hierarchical models generated a wider variation in the estimate because between-study variation was incorporated into these models. Bayesian approaches provide a flexible framework to incorporate trial heterogeneity, realistically assess uncertainty, and may result in better input for decision models.
See more of Poster Session - Public Health; Methodological Advances
See more of The 26th Annual Meeting of the Society for Medical Decision Making (October 17-20, 2004)