Diagnostic accuracy and prediction increment of markers of epithelial-mesenchymal transition to assess cancer cell detachment from primary tumors

Evan L. Busch, Prabhani Kuruppumullage Don, Haitao Chu, David B. Richardson, Temitope O. Keku, David A. Eberhard, Christy L. Avery, Robert S. Sandler

Metastases play a role in about 90% of cancer deaths. Markers of epithelial-mesenchymal transition (EMT) measured in primary tumor cancer cells might provide diagnostic information about the likelihood that cancer cells have detached from the primary tumor. Used together with established diagnostic tests of detachment-lymph node evaluation and radiologic imaging-EMT marker measurements might improve the ability of clinicians to assess the patient's risk of metastatic disease. Translation of EMT markers to clinical use has been hampered by a lack of valid analyses of clinically-informative parameters. Here, we demonstrate a rigorous approach to estimating the sensitivity, specificity, and prediction increment of an EMT marker to assess cancer cell detachment from primary tumors. We illustrate the approach using immunohistochemical measurements of the EMT marker E-cadherin in a set of colorectal primary tumors from a population-based prospective cohort in North Carolina. Bayesian latent class analysis was used to estimate sensitivity and specificity in a setting of multiple imperfect diagnostic tests and no gold standard. Risk reclassification analysis was used to assess the extent to which addition of the marker to the panel of established diagnostic tests would improve mortality prediction. We explored how changing the latent class conditional dependence assumptions and definition of marker positivity would impact the results. All diagnostic accuracy and prediction increment statistics varied with the choice of cut point to define marker positivity. When comparing different definitions of marker positivity to each other, numerous trade-offs were observed in terms of sensitivity, specificity, predictive discrimination, and prediction model calibration. We then discussed several implementation considerations and the plausibility of analytic assumptions. The approaches presented here can be extended to any EMT marker, to most forms of cancer, and to different kinds of EMT marker measurements, such as RNA or gene methylation data. These methods provide valid, clinically-informative assessment of whether and how to use a given EMT marker to refine tumor staging and consequent treatment decisions.