2006 Scientific Session Abstracts
Scientific Session Presentations & Schedule
E-Poster Presentations
Submission Compliance and Guidelines
Copyright Transfer Form (PDF)
Abstract Archive
Bayesian Inference to Assist IHC Interpretation
Rodney Schmidt MD, PhD (schmidtr@u.washington.edu)1; Robin Vollmer MD2; Lawrence True MD1; and Erin Grimm, MD1. 1Department of Pathology, University of Washington, Seattle, WA; 2Dept of Pathology, VA Medical Center and Duke University Medical Center, Durham, NC.
Context: Pathologists currently interpret the results of immunohistochemistry (IHC) studies using a subjective inferential process that produces no quantitative estimate of diagnostic certainty. The consistency and rigor of this process might be improved using Bayes’ theorem. Bayes’ theorem posits that the likelihood of each possible diagnosis can be calculated from a pre-test estimate of the likelihood of each diagnosis, antibody sensitivity data, and test results for each antibody. We sought to implement Bayes’ theorem in a diagnostically accessible form, determine whether the output probabilities were diagnostically plausible, and generate interpretation guides for common clinical scenarios and antibodies.
Technology: A generalized form of Bayes’ theorem was implemented in SQL using a simplifying assumption of statistical independence of results for different antibodies.
Design: For proof of concept, incidence rates for 17 common neoplasms were estimated from published SEER incidence data. Sample antibody sensitivity data for cytokeratins 7 and 20, S100, PSA, and TTF1 were taken from published data. Output diagnostic probability tables were generated for various combinations of test results for each antibody and reviewed for diagnostic plausibility by 3 experienced surgical pathologists. The underlying mathematical assumptions and the starting incidence and sensitivity data were then reviewed to identify issues that could significantly influence the results.
Results: The algorithm to implement Bayes’ theorem produces CK7/CK20 interpretation guides and probability estimates for output diagnoses that are clinically plausible. The greatest difficulty with applying Bayes’ theorem lies with the antibody sensitivity data – published results vary by author, may not reflect the local lab experience, and different criteria may be used for cancers when reporting incidence and sensitivity data. Pre-test estimates of tumor likelihood also influence the output probabilities. The assumption of statistical independence for each antibody is also known to be violated in some situations.
Conclusions: Our proof-of-concept studies show that Bayes’ theorem can be implemented in a form that can be slipstreamed into the diagnostic/reporting process. This tool can: 1) assist in the interpretation of IHC results by estimating the likelihoods of possible diagnoses; 2) facilitate teaching by providing an objective basis for interpretation; 3) provide a framework for further investigations to refine and improve the output probability estimates.