BRIDGING BEYOND REGRESSION--ARTIFICIAL NEURAL NETWORK IN MAPPING FROM SF-36 TO HEALTH UTILITY
Author(s)
Yu AP1, Yu YF2, Nichol MB31 University of Southern California, Los Angeles, CA, USA; 2 HealthCore, Inc, Wilmington, DE, USA; 3 University of Southern California, School of Pharmacy, Los Angeles, CA, USA
OBJECTIVES: Linear regression was often used to develop models bridging SF-36 to utility, with assumptions of linearity and independence among subscales. Artificial neural networks (ANN) are capable to detect complex patterns among data, thus may provide as a promising alternative algorithm to bridge from SF-36 subscales to utility. This study is to investigate the predictive power of an ANN algorithm, in contrast to that of linear regression. METHODS: Data included a cohort of 6923 Southern California Kaiser Permanente members who filled both SF-36 and HUI2 during 1994-1995. A random sample containing half of total members was used as a training sample to generate mapping algorithms based on linear models and ANN respectively, and the other half was used as testing sample to compare accuracy of prediction between the two methods. Both models used 8 domain scores plus age and gender as predictors. Model prediction accuracy was evaluated based on mean squared error, mean absolute error, percentage of correct prediction within 0.01, 0.03 and 0.1 error range. Observed and predicted utility scores by both methods in testing sample were summarized and compared. This random sample tests were conducted multiple times in order to examine the stability of such cross-validation procedure. RESULTS: The “trained” neural network model was a 10-3-1 network. After cross-validation, the mean squared error was 0.128 for linear model vs. 0.127 for ANN, and absolute error was 0.095 vs. 0.093. Within 0.01 error range, linear model predicted 8.58% correctly, and ANN 8.73%. The rates were 23.8% vs. 24.9% with 0.03 error range, and 63.4% vs. 65.4% with 0.10 error range. The testing process was repeated multiple times and results were stable. CONCLUSIONS: When testing to an independent sample, the bridging algorithm based on ANN showed improved predictive accuracy and can be used as an alternative to linear prediction algorithm.
Conference/Value in Health Info
2005-05, ISPOR 2005, Washington, DC, USA
Value in Health, Vol. 8, No. 3 (May/June 2005)
Code
PMC33
Topic
Methodological & Statistical Research, Patient-Centered Research
Topic Subcategory
Modeling and simulation, Patient-reported Outcomes & Quality of Life Outcomes
Disease
Multiple Diseases