Neural network processed 41 variables of 289 psychiatric inpatients and successfully converged (97% correlation between output and target signals) on a good 8-week outcome. The sensitivity of each parameter was calculated by ditherring the input signals one by one at 1% signal increase and computing the corresponding output. Fifteen parameters were predictors of good outcome and 17 parameters were predictors of bad outcome. The most sensitive good outcomes were ECT, group, behavioral and dynamic psychotherpeutic approaches, clozapine administration, noradrenergic antidepressant treatment, schizopositive symptoms, and being of high socioeconomic level. The most sensitive bad outcomes were long duration of the last in-hospital stay or last remission, suffering from OCD or resistant depression, being treated with low potency neuroleptic drugs, and having community support. Sensitivity detection by neural network is a new approach in psychiatry. A 70% concordance level was found with outcome measures reported in literature and that have been traditionally calculated. The fact that the system converged in spite of the wide spectrum of input signals means that underlying interconnections were detected by the neural network.