Backpropagation and adaptive resonance theory in predicting suicidal risk

The ability of backpropagation and adaptive resonance theory (ART) neural networks to predict the probability of complete suicide, within a two year span, in major psychiatric patients was investigated. Variables associated with suicide risk were collected from the files of 161 hospitalized psychiatric patients with a 10 year or greater history of illness. 84 patients were hospitalized due to suicide attempts and 77 had no previous suicide attempts or ideations. Suicide attempts were rated as medically serious suicide attempts (MSSA) or non-MSSA and used for training the systems. The ability of the neural networks was evaluated by screening the extremes of the suicidal spectrum (1) 54 records of patients who committed suicide and (2) 150 records of patients who never had suicidal thoughts. The records were taken from 3 hospitals, in various geographic regions in Israel. Neither neural network system is reliable in predicting suicide, however, records from one hospital, Gehah Hospital, were better identified than those from the two other hospitals (p < 0.05 for PPV; p < 0.01 for specificity). At present, neural networks are not reliable instruments for evaluating suicidal risk due to the significant number of false positive results. When low risk was indicated reliability was greater (NPV = 75.28%, specificity = 97.10% with ART; NPV = 91.76%, specificity = 95.12% with backpropagation). However, PPV, NPV and specificity rates of both systems achieved with Gehah Hospital records suggest that using a direct-subjective questionnaire may produce better results in the future. ART and backpropagation performed similarly in all measurements.