With the widespread use of ultrasonography (US) and computerized tomography (CT), the usefulness of alpha-fetoprotein assay in the diagnosis of hepatocellular carcinoma (HCC) has decreased. The aim of our study was to evaluate the best cut-off value for serum alpha-fetoprotein to discriminate between liver cirrhosis (LC) and HCC and the factors influencing levels in a Sicilian population. Three hundred and seventy-two patients with LC and 197 with HCC-associated LC were studied. The etiology was: HCV in 288 cases (77.4%) of LC and 147 cases (75%) of HCC; HBV in 31 cases (8.3%) of LC and 15 cases (7.6%) of HCC; HCV/HBV in 21 cases (5.6%) of LC and 6 cases (3.0%) of HCC; non-viral in 32 cases (8.6%) of LC and 29 cases (15%) of HCC. Hepatic function was estimated by the Child-Pugh's score; the TNM classification was used in HCC. The area under the ROC curve was 0.81 +/- 0.02; the best discriminant cut-off value, calculated as the value of the maximized likelihood ratio, was 30 ng/ml. At this level sensitivity (SE) was 65%, specificity (SP) 89%, positive predictive value (PPV) 74% and negative predictive value (NPV) 79%. When the patients were divided at this cut-off point into two groups according to viral or non-viral etiology, PPV was 70% versus 94%, respectively (p < 0.05). In the non-viral diseases PPV reached 100% for AFP serum levels of 100 ng/ml, while in the viral diseases PPV was 100% when AFP was greater than 400 ng/ml. There were no significant differences in SE, SP or NPV between viral and non-viral liver diseases. Child's classes B and C were more frequent in HCC (chi 2 of MH 7.7, p < 0.0001). There was a correlation between AFP serum values and TNM classification (p < 0.02) and on multiple logistic regression AFP levels > 30 ng/ml correlated positively only with the TNM stage (p < 0.0001). In conclusion, the best cut-off value for serum AFP in our study population was 30 ng/ml, but at this level sensitivity was low. This cut-off value was more useful in detecting non-viral HCC, because PPV was significantly higher than in viral HCC; therefore, our data confirm that the usefulness of AFP in the diagnosis of HCC of viral etiology is limited, being more useful in HCC of non-viral etiology.