Fall-KAN: Fall impact time estimation Kolmogorov-Arnold Network
Nicholas Cartocci, Antonios E. Gkikakis, Fabio Pera, Maria Teresa Settino, Darwin Gordon Caldwell, Jesús Ortiz · 2024
A new application based on the Kolmogorov-Arnold Network (KAN) for estimating the time of impact of the fall is presented (Fall-KAN). The SisFall data set that contains fall scenarios with real subjects is used, and the approach is employed to estimate the impact time with the ground during a fall. The performance of the proposed approach is evaluated by comparing it with Linear Regression, Regression Tree, SVM, and MLP using the SisFall dataset in terms of RMSE and$R^{2}$. Fall-KAN achieved a RMSE of approximately 150$ms$after a single training epoch, which is more than 3% lower than other ML methods. It also demonstrated superior robustness in estimating the time of impact during a fall, with a low initial prediction error and a small settling value only close to the impact.