Loosening of hip stem implants is a major reason for the revision of total hip replacements. Current diagnostic methods cannot detect a loosening in an early stage with high sensitivity and specificity. Previous analysis showed the potential of classifying a loosening by the analysis of extracorporeally captured structure borne sound. For this purpose, an existing hip-stem implant was modified to house a mechanical oscillator which can be excited by an extracorporeal magnetic coil. The excitation system is directly located in the hip-stem implant thus allowing for in vivo monitoring of implant loosening. Experiments with four human femur bones from body donations have been conducted in a previous study. Results showed that the loosening can be detected almost certain using individual models and with high accuracy using a non-individual model trained with data from multiple individuals.This contribution aims to discuss the generalization of the previously developed non-individual model. Data augmentation is used to emulate inter-individual variability and cross-validate the non-individual model. Alternative outlier detection strategies besides the previously applied mean average error method are evaluated.