HISTOLOGIC FEATURE RELIABILITY IN CHILDHOOD NEURAL TUMORS

We studied intraobserver reproducibility in recognizing the presence o r absence of 57 histologic features or patterns in a random subset of tumors (822) from the Childhood Brain Tumor Consortium database. The s tudy protocol maximized consistency of the observer. We found that onl y six histologic features had high (greater than or equal to 0.75) rel iability estimates while a large number had intermediate estimates of 0.50-0.74. Supratentorial or infratentorial tumor location sometimes a ltered reliability. Reliability estimates were unacceptable for certai n histologic features often used as diagnostic criteria, descriptors o f tumor characteristics, or markers of anaplasia. We hypothesize that low reliability reflects, in part, the need for more specific operatio nal definitions, particularly those with subjective boundaries (e.g. t he cellular density constituting the histologic boundary between high and intermediate cell density). Rarity of a feature in the microscopic section (e.g. granular bodies) may also contribute to low reliability . We also show that the kappa statistic, a commonly used measure of re liability, is inappropriate for very common or uncommon histologic fea tures (e.g. features at the extremes of prevalence in the study cases) and we offer a simple empiric method for determining when an alternat ive measure, the Jaccard statistic, is appropriate.