Statistical models for predicting a beneficial response to interferon-alpha in patients with chronic hepatitis B

Therapy with interferon-alpha has been reported to induce remissions in 35% of patients with chronic hepatitis B. The ability to identify patients likely to respond would be helpful in making recommendations for treatment. In this statistical analysis we included 82 patients with chronic hepatitis B who received interferon-alpha in clinical trials at the National Institutes of Health between 1984 and 1991. A response was defined as the loss of hepatitis B virus (HBV) DNA and hepatitis B e antigen (HBeAg) within 1 year of therapy. Multiple clinical parameters measured at pretreatment (month 0) and after the first month (month 1) of therapy were selected by stepwise regression to support the development of the prognostic models: the two-stage logistic regression model and a neural network that utilized higher-order non-linear interactions between variables. Among the 82 patients, 24 (29%) were responders. The two-stage logistic model using pretreatment variables: sex, hepatic fibrosis and alanine aminotransferase (ALT) levels correctly identified 61% of responders and 76% of non-responders. When HBV DNA at month 1 along with sex, initial ALT and fibrosis was included, the resultant model correctly identified 69% of responders and 77% of non-responders. The neural network, by incorporating interactions between variables, correctly identified 77% and 86% of responders, and 87% and 92% of non-responders, using pretreatment factors alone and the combination of pretreatment and month 1 factors respectively. Hence, the neural network was more accurate than the simple logistic regression model in predicting a response to interferon-alpha in chronic hepatitis B. The universality of these models needs to be further verified.