W E Lowell1, G E Davis. 1. Information Services, Augusta Mental Health Institute, ME 04330.
Abstract
OBJECTIVE: To test the effect of diagnosis on training an artificial neural network (ANN) to predict length of stay (LOS) for psychiatric patients involuntarily admitted to a state hospital. DESIGN: A series of ANNs were trained representing schizophrenia, affective disorders, and diagnosis-related group (DRG) 430. In addition to diagnosis, variables used in training included demographics, severity of illness, and others identified to be significant in predicting LOS. RESULTS: Depending on diagnosis, ANN-predictions compared with actual LOS indicated accuracy rates ranging from 35% to 70%. The validity of ANN predictions was determined by comparing LOS estimates with the treatment team's predictions at 72 hours following admission, with the ANN predicting as well as or better than did the treatment team in all cases. CONCLUSIONS: One problem in traditional approaches to predicting LOS is the inability of a derived predictive model to maintain accuracy in other independently derived samples. The ANN reported here was capable of maintaining the same predictive efficiency in an independently derived cross-validation sample. The results of ANNs in a cross-validation sample are discussed and the application of this tool in augmenting clinical decision is presented.
OBJECTIVE: To test the effect of diagnosis on training an artificial neural network (ANN) to predict length of stay (LOS) for psychiatricpatients involuntarily admitted to a state hospital. DESIGN: A series of ANNs were trained representing schizophrenia, affective disorders, and diagnosis-related group (DRG) 430. In addition to diagnosis, variables used in training included demographics, severity of illness, and others identified to be significant in predicting LOS. RESULTS: Depending on diagnosis, ANN-predictions compared with actual LOS indicated accuracy rates ranging from 35% to 70%. The validity of ANN predictions was determined by comparing LOS estimates with the treatment team's predictions at 72 hours following admission, with the ANN predicting as well as or better than did the treatment team in all cases. CONCLUSIONS: One problem in traditional approaches to predicting LOS is the inability of a derived predictive model to maintain accuracy in other independently derived samples. The ANN reported here was capable of maintaining the same predictive efficiency in an independently derived cross-validation sample. The results of ANNs in a cross-validation sample are discussed and the application of this tool in augmenting clinical decision is presented.