Dr. George Davis of the Augusta Mental Health Center (Augusta, ME) has trained a BrainMaker neural network which predicts the length of stay (LOS) for psychiatric patients. His system (available through Psybernetics, Inc. Augusta, ME) allows state hospitals and private facilities to determine which patients would benefit most from short stays and which require long-term (thirty days or more) hospitalization.
The system has the potential of providing an annual savings of $100,000 to a 300 bed private facility, and up to $600,000 for a tertiary (state) facility. Separating short term from long term stay patients upon admission rather than after some period of observation saves time and money. Fewer inappropriate hospitalizations occur, which not only saves the state money, but allows the short term patient to benefit from community settings and support systems, and reduces the psycho-social stigma of hospitalization. In addition, there is a lessened burden on the legal system and law enforcement agencies, and less paperwork. Short term patients who require hospitalization are more likely to be admitted to a general hospital because they may still retain insurance benefits.
The neural network performs better than traditional approaches in predicting the length of stay (LOS). Only 8-30% of the variance in LOS could be correlated to a combination of demographic, diagnostic and clinical variables. By comparison neural networks were able to explain 39-86% of the variance.
There are 48 or 49 kinds of input data used to train two different neural networks. The inputs include basic demographics, admission history, family support systems, ability to care for self, diagnosis information, etc.
There are four output neurons. The outputs define four classes of LOS: 1) less than 1 week, 2) greater than 1 week but less than thirty days, 3) greater than thirty days but less than six months, 4) greater than six months but less than one year. Four networks were trained using 500-600 cases, two each on two different years of annual data. In this way, the predicted LOS for particular patients could be compared between two years at the institution which had undergone major organizational changes.
Psychiatric diseases are the most difficult to predict. In addition, varying standards and funding policies make care more susceptible to chaos and difficult to compare between locations. Neural networks provide a means to predict the effectiveness of care for a specific location. Several neural networks can be trained with patient data from particular time periods, which will provide a method of judging the effectiveness of changing policy, procedure, or available resources. For example, if several "typical" patient cases are created or particular troublesome cases are selected, these can be run through the different networks to determine which changes in treatment would be most beneficial.
INPUTS
age
sex
residence prior to admission
date of admission
admission history
diagnosis information
employment
family support
severity of illness
dangerousness
self care ability
assaults
substance abuse
(etc.)
OUTPUT
length of hospital stay