guest ::
%0 Thesis %A Loh, Johnson Luo %T Neuro-inspired hardware acceleration for efficient biomedical signal processing in mobile sensing devices %I Rheinisch-Westfälische Technische Hochschule Aachen %V Dissertation %C Aachen %M RWTH-2025-05740 %P 1 Online-Ressource : Illustrationen %D 2025 %Z Veröffentlicht auf dem Publikationsserver der RWTH Aachen University %Z Dissertation, Rheinisch-Westfälische Technische Hochschule Aachen, 2025 %X The processing of biomedical signals in mobile sensing devices enables the continuous monitoring of health parameters for early detection of threatening arrhythmia in the population through convenient wearable devices, such as smartwatches. The design of processing modules, which are feasible in this resource-constrained environment, is subject to multiple constraints for their deployment in-field. High quality classification is desired for accurate detection to trigger treatment by trained personnel. Robust classification beyond available training data is necessary to generalize system feasibility across the general population. Low-power operation is necessary for long-termscreening for sparse features indicating abnormal health conditions. The co-optimization of neuro-inspired algorithms on dedicated hardware shows the promise to address all desired specifications in an application-specific device. This work explores neuro-inspired concepts for low-power digital processing of biomedical signals. Artificial neural networks has shown superior classification capabilities in the machine learning domain. Using an artifical neural network as a baseline, a systematic design space exploration methodology is applied to design an ECG classifier and co-optimize the system from the algorithm level down to a hardware design for ultra-low power consumption and high classification quality. Then, the system is extended with a domain generalization method for robust classification across multiple datasets. The method is designed for direct integration into a pre-trained neural network with low overhead regarding inference and training. At last, the temporal coding of information in spikes is adopted from the human brain as a data processing mechanism for low power processing. The investigated temporal coding method shows equivalent numerical values after processing with reduced operations compared to conventional fixed-point arithmetic. In the end, the neuro-inspired concepts show promising directions to improve specialized ANN hardware accelerators for biomedical signal processing both for low-power processing and robust high-quality classification. %F PUB:(DE-HGF)11 %9 Dissertation / PhD Thesis %R 10.18154/RWTH-2025-05740 %U https://publications.rwth-aachen.de/record/1013853