Sonographic prediction of malignancy in adnexal masses using multivariate logistic regression analysis

The aim of the study was to assign a probability of malignancy for any patient with an adnexal tumor by the application of multivariate logistic regression analysis to variables recorded at the time of pelvic sonography. Sixty-seven women with known adnexal masses were examined using transvaginal B-mode and color Doppler imaging. For each patient the variables included: (1) age, (2) maximum tumor diameter, (3) tumor volume, (4) unilocularity (presence (0) or absence(1)), (5) papillary projections (presence (1) or absence (0)), (6) random echogenicity (presence (1) or absence (0)), (7) highest peak systolic velocity (PSV), (8) time-averaged maximum velocity (TAMXV), (9) pulsatility index (PI) and (10) resistance index (RI). The TAMXV, PI and RI were those associated with the highest PSV. These ten independent variables and the final histological diagnosis for each patient (the dependent variable) were used for the regression analysis. Approximately 75% of the entire dataset was randomly selected for generating the regression model. The remaining 25% was used as the testing set for cross-validation of the model. In the entire dataset there were 52 women with benign, three with borderline and 12 with invasive ovarian tumors. Regression analysis on the ten variables resulted in the retention of only 'age', 'papillary projection score' and 'TAMXV' as significantly contributing to predicting the presence or absence of malignancy. The probability of malignancy for any patient was given by solving the equation: Probability = 1/(1 + e-z) where e is the base value for natural logarithms and z = (0.1273 x Age) + (0.2794 x TAMXV) + (4.4136 x Papillary projections score) - 14.2046. Cross-validation of the model on the test set of data gave a 100% sensitivity and specificity. However, for the entire dataset the best sensitivity and specificity were 93.3 and 90.4%, respectively, at a cut-off value of 25% probability of malignancy. In conclusion, multivariate logistic regression analysis enables the calculation of probability of malignancy for any patient with a known adnexal mass. The accuracy of this prediction appears to be better than that of morphological or Doppler criteria when the latter are used independently. The value of this model needs to be tested prospectively.