Comparison of fuzzy inference, logistic regression, and classification trees (CART). Prediction of cervical lymph node metastasis in carcinoma of the tongue.

OBJECTIVES In this paper three statistical methods [logistic regression, classification and regression tree (CART), and fuzzy inference] for the prediction of lymph node metastasis in carcinoma of the tongue are compared. METHODS A retrospective collection of data in 75 patients treated for tongue cancer was carried out at the Clinic and Policlinic for Oral and Maxillo-facial Surgery at the University Hospital of Freiburg in Germany between January 1990 and December 1999; biopsy material was used for laboratory evaluations. Statistical methods for the prediction of lymph node metastasis were compared using ROC curves and accuracy rates. RESULTS All three methods show similar results for the prediction of lymph node metastasis with slightly superior results for fuzzy inference and CART. A great overlap is apparent in the ROC curves. The best result observed for fuzzy inference and CART was a sensitivity of 79.2% [95% confidence interval: (57.8%; 92.9%)] and a specificity of 86.3% (73.7%; 94.3%); the best result for predictions based on the logistic regression was a sensitivity of 66.7% (44.7%; 84.4%) and a specificity of 80.4% (66.9%; 90.2%). Accuracy rates of fuzzy method and CART were higher [accuracy rate for fuzzy method and CART: 84% (73.7%; 91.4%), for logistic regression method: 73.3%, 95%-CI: (61.9%; 82.9%)]. CONCLUSIONS From a clinical point of view, the predictive ability of the three methods is not sufficiently large to justify use of these methods in daily practice. Other factors probably on the molecular level are needed for the prediction of lymph node metastasis.