ENSEMBLED ARTIFICIAL NEURAL NETWORKS TO PREDICT THE FITNESS SCORE FOR BODY COMPOSITION ANALYSIS

To predict the nutrition and health status of staff and students in Yuan Ze University and select the influential variables from the total body composition variables, which should have similar predictive ability with the whole factors. Design: Spontaneous and voluntary physical examination. Setting: Sanitary & Health Care Section of Yuan Ze University in Taiwan. Participants: 1227 staff and students. Measurements: With the help of Inbody720TM, 139 body composition variables were measured and 60 variables were retained after data pre-processing. An ensembled artificial neural networks (EAnn) prediction model was established and seven different methods for assessing variables importance were applied. Besides, classical linear and logistic regression models were developed for comparison with EAnn prediction results. Results: The prediction performance of EAnn model was satisfactory (RMSE (train) = 0.2686, RMSE (validation) = 0.2648, RMSE (test) = 0.3492). Since both the actual and simulation fitness score were at the range of 0 to 100, according to rounding off rule, the simulated value was almost the same with actual value. Besides, 12 important variables were obtained by seven methods for quantifying variable importance in EAnn, which had similar predictive capability with 60 variables (RMSE (train) = 0.3263, RMSE (validation) = 0.322, RMSE (test) = 0.3226). The linear and logistic regression models results were both evidently worse than EAnn results. Conclusion: The results confirm that EAnn is appropriate to approximate such a complicated, non-invasive and highly non-linear problem as body composition analysis. It can be helpful for nutritionists to manage and improve the nutrition and health condition of staff and students, by adjusting the 12 most important variables.