A neural network that predicts psychiatric length of stay.

We developed a neural network to predict length of stay after admission to a state psychiatric hospital. The network was trained with data from 1064 cases randomly selected over a 3 1/2-year period, and its accuracy was tested against actual length-of-stay patterns and predictions made by a team of clinicians 72 hours after admission. The network performed at least as well as the team of clinicians. Successful prediction of length of stay could result in more appropriate use of services, more timely initiation of treatment, better resource planning, and cost control--highly sought dividends in an era of diminishing fiscal resources.