Existing criteria for separating clinically benign but architecturally complex or cytologically atypical endometrial proliferations (hyperplasia or metaplasia) from well-differentiated endometrial carcinoma are underspecified and poorly reproducible, in part due to the absence of a uniformly agreed on methodologically independent outcome against which to judge the efficacy of competing sets of criteria. Because myoinvasion is the first unambiguous indicator of clinically aggressive behavior for proliferations in this spectrum, we have employed the presence or absence of myoinvasion as a tool to develop clinically meaningful diagnostic criteria for the separation of complex atypical hyperplasia/metaplasia from well-differentiated carcinoma (CAHM/WDCA). We obtained the paired endometrial samplings and hysterectomy specimens of 520 patients; these were split into a training set of 306 cases and a test set of 214. The presence or absence of myoinvasion was assessed from an examination of the hysterectomy specimen. For the purposes of this study, myoinvasion was defined as the presence of irregular intramyometrial glands surrounded by a granulation tissue-like response. To determine the morphologic features that were most predictive of myoinvasion, a series of endometrial architectural, cytological, and stromal features was initially evaluated on the training set (149 myoinvasive and 157 nonmyoinvasive). Using a variety of exploratory data techniques including the classification algorithm CART, we developed a diagnostic rule for predicting myoinvasion that employed one architectural feature (glandular complexity captured by a pictorial architectural index) and two cytological features (nuclear pleomorphism and prominence of nucleoli). Extensive squamous differentiation, fibroblastic stroma, necrosis, stromal foam cells, and other cytologic features did not provide additional predictive value when cross-validated. The true misclassification rate of the CART-generated prediction rule was further assessed by applying the rule to the test set drawn largely from community hospitals. The sensitivity and specificity of this rule for detecting myoinvasion was 99.5 and 57%. The likelihood ratio was 2:1, (i.e., using prior odds of myoinvasion in the CAHM/WDCA spectrum of 1:10, the posterior odds on myoinvasion using the CART-generated rule would be 1:5). Comparison of the CART-generated myoinvasion prediction rule with the Kurman and Norris endometrial stromal invasion criteria for well differentiated endometrial carcinoma (25), using receiver operator characteristic curve (ROC) techniques, demonstrated a significant improvement in the ability to separate myoinvasive from non-myoinvasive endometrial proliferations with the CART-generated rule; the average area under the curve for the CART-generated rule was 0.78 (SE = 0.02) versus 0.67 (SE = 0.03) for the endometrial stromal invasion criteria.(ABSTRACT TRUNCATED AT 400 WORDS)