Contribution
Improving the energy efficiency of the detection of hip-stem implant loosening using structure-borne sound for an on-implant implementation
* Presenting author
Abstract:
The number of total hip arthroplasties will rise due to the increasing age of the population with higher demand by more active patients. Previously, we proposed autoencoders to detect the implant loosening of total hip stems using structure-borne sound. In particular, the setup consists of an embedded exciter and accelerometers, which measure the radiated structure-borne sound outside of the body. Our current investigations focus on embedding multiple accelerometers directly onto the implant to get a better signal quality and remove the need for external measurement hardware. To achieve such a self-contained system, autoencoder inference must be realized on the implant. Hence, a strict energy budget must be considered. Therefore, we assessed the existing set of autoencoders obtained from the hyperparameter tuning, focusing only on the quality of the results. We observed that the number of parameters spans a broad range, from a few thousand to several million. However, an increased complexity of the models typically did not yield better results. To achieve energy-efficient models while maintaining high detection accuracy, we use two approaches. First, we use a hardware-aware neural architecture search to yield more compact networks. Second, we apply model optimization techniques like pruning and quantization to simplify the computation.