Artificial neural networks: a study in clinical psychopharmacology

Controlled trials in clinical psychopharmacology may fail to provide reliable information about the benefit of treatment when the patient is viewed in a real-life setting rather than as part of a well-defined sampling procedure. A viewpoint, rooted in systems theory, is proposed based on the identification of complex relationships among such dimensions as clinician's reasoning, drug properties, and patient's condition. Artificial Neural Network (ANN) technology provides efficient tools for data analysis within a systems-oriented approach. This study proposes a way to predict the outcome of psychopharmacological treatment. Analysis was conducted on retrospective data from clinical records of psychiatric patients treated with moclobemide. Twelve pharmacological, diagnostic, and topological variables were identified as the decisional items considered by six clinicians: age at onset, sex, previous treatment, duration and dose of moclobemide treatment, other drugs, psychiatric diagnosis and other clinical features. Data were binarily coded and transformed into observed frequencies in the sampling space; treatment outcome was binarily scored as the model's target. A Back-Propagation ANN based on the Delta rule with logistic transfer function was used. ANN correctly classified all cases of successful treatment (n = 51, 100%) but only half of the unsuccessful cases (n = 14, 52%). Patterns of response and areas of uncertainty were analyzed in a topological approach.