New method for evaluating artificial neural network algorithm with signal detection theory and full factorial design for detecting falls.

Fall is one of the most critical accidents resulting in serious injuries and significant financial losses among people in all ages. This paper presents the application of full factorial design (FFD) to investigate fall detection algorithms that have multiple hyperparameters which are very difficult to identify the best values for the dataset. In this study, the algorithm factors are investigated from two motion sensors and six artifact neural network (ANN) parameters on seven possible outcomes of signal detection theory (SDT). It is found that only one accelerometer and one gyroscope and small size ANN with scaled conjugate gradient (SCG) and radial basis function (RBF) provide a higher performance classification with lower computational complexity. Experimental outcomes show the new method using statistical theory for the selection of the most effective performance of fall detection algorithm parameters. Findings from the current study could be applied to various types of classification model problems in engineering applications, such as the design of products and systems. [ABSTRACT FROM AUTHOR]