Bayesian modelling and ROC analysis to predict placebo responders using clinical score measured in the initial weeks of treatment in depression trials.

WHAT IS ALREADY KNOWN ABOUT THIS SUBJECT: * In major depressive disorder an appreciable percentage (40%) of patients in antidepressant trials will have a placebo response. * In these trials, early changes (i.e. within the first 4 weeks) of the clinical score scale (e.g. HAMD-17) are associated with response at end-point. * Unpredictable placebo response is one of the major reasons for clinical trial failure in the evaluation of antidepressant drugs. WHAT THIS STUDY ADDS: * Provides a model to describe the time course of individual and population placebo response. * Provides a methodology to forecast the individual probability to be placebo responder based on early HAMD-17 measurements with an assessment of the prognostic power. * Provides a methodological framework to implement a population enrichment strategy in the design of clinical trials for the assessment of novel antidepressant drugs. AIMS: To develop a probabilistic and longitudinal model describing the time course of Hamilton's Rating Scale for Depression (HAMD-17) total score in patients with major depressive disorders treated with placebo and to develop predictive models to estimate the response at end-point given HAMD-17 measurements at weeks 2 and 4.
METHODS: Patients (n = 691) from seven clinical trials were analysed in WinBUGS using a Bayesian approach. The whole dataset was randomly split in a learning (359 patients for model definition) and a test dataset (332 patients for assessment of model predictive performance). The analysis of the learning dataset assumed uninformative priors, whereas the analysis of the test dataset used the posterior parameter estimates of the learning dataset as priors. ROC curve analysis estimated the optimal sensitivity/specificity cut-off between false-negative and false-positive rates and determined the prognostic allocation rule for patients to responder and nonresponder groups.
RESULTS: A Weibull/linear model accurately described the population and individual HAMD-17 time course. The total area under the ROC curve, ranging from 0.76 (logistic model with data at week 2) to 0.86 (longitudinal model with data at week 4), provided a measure of the prognostic discriminatory power of early HAMD-17 measures using the two models. The best placebo-responder classification score (86.32% true and 13.68% false positive) was associated with the longitudinal model with HAMD-17 measures at week 4.
CONCLUSION: Results showed the relevance of the Bayesian approach to predict HAMD-17 score at study end and to classify a patient as a placebo responder given the uncertainty in parameters derived from historical data and early HAMD-17 measurements.