Do Advanced Statistical Techniques Really Help in the Diagnosis of the Metabolic Syndrome in Patients Treated With Second-Generation Antipsychotics?

Abstract

OBJECTIVE: Metabolic and cardiovascular diseases in patients with schizophrenia have gained a lot of interest in recent years. Developing an algorithm to detect the metabolic syndrome based on readily available variables would eliminate the need for blood sampling, which is considered expensive and inconvenient in this population. METHOD: All patients fulfilled DSM-IV diagnosis of schizophrenia or schizoaffective disorder. We used the International Diabetes Federation criteria (European population) to diagnose the metabolic syndrome. We used logistic regression and optimized artificial neural networks and support vector machines to detect the metabolic syndrome in a cohort of schizophrenic patients of the University Psychiatric Center Kortenberg, KU Leuven, Belgium. Testing was done on one-third of the included cohort (202 patients); training was performed using a 10-fold stratified cross-validation scheme. The data were collected between 2000 and 2008. RESULTS: All 3 methods yielded similar results, with satisfying accuracies of about 80%. However, none of the advanced statistical methods could improve on the results obtained using a very simple and naive model including only central obesity and information on blood pressure. CONCLUSIONS: Although so-called pattern recognition techniques bear high promise in improving clinical decision making, the results should be presented with caution and preferably in comparison with a less complicated technique.