Selecting Modeling Techniques for Outcome Prediction: Comparison of Artificial Neural Networks, Classification and Regression Trees, and Linear Regression Analysis for Predicting Medical Rehabilitation Outcomes

• Published in: Proceedings - AMIA Symposium 1999-01, p.1187-1187
• Main Authors: Walters, Deborah K. W., Linn, Richard T., Kulas, Margaret, Cuddihy, Elisabeth, Wu, Chonghua, Granger, Carl V.
• Format: Article
• Language: English
• Subjects: Poster Sessions
• ISSN: 1531-605X

A multitude of techniques exists for modeling medical outcomes. One problem for the researcher is how to select an appropriate modeling technique for a given task. This paper addresses the problem through: an analysis of the strengths and weaknesses of three techniques; and, a case study in which the three techniques are applied to the task of predicting medical rehabilitation outcomes. The three techniques selected where linear regression analysis (LRA), classification and regression trees (CART) and artificial neural networks (ANN). The analysis illustrates that when the relationship between the independent and dependent variables is a linear one, that LRA is adequate. However, when a nonlinear relationship exists, CART or ANN analysis will yield better models. When a nonlinear, nonorthogonal relationship exists, then ANN analysis will yield a better model. The results of the case study show that the ANN model is more accurate than both LRA and CART in predicting the discharge motor FIM from admission data for stroke patients admitted to medical rehabilitation facilities. However, the increased accuracy comes at an increase in the computational cost of the model, thus a decision about which technique to use must be made by weighing the increased accuracy against the increased cost.