Xiaolin, LiLiXiaolin2025-12-012025-12-012024 the A2024http://hdl.handle.net/10197/30555The imperative to strike a harmonious balance between precision and efficiency has led to the exploration of novel approaches for decentralized inferencing. This thesis addresses the critical challenge of minimizing power consumption on edge devices while transmitting minimal data to the cloud, thereby achieving low latency, low complexity, and high accuracy in the detection of abnormal biomedical signals. This research introduces three innovative solutions to tackle this multifaceted problem. The first solution combines large and small models, strategically leveraging their unique strengths to enhance the efficiency and accuracy of biomedical signal inferencing. By combining the comprehensiveness of large models with the agility of small models, this approach optimizes performance while judiciously managing computational resources. This fusion not only enables the recognition of subtle patterns in biomedical signals but also ensures that the inferencing process remains feasible for edge devices with limited processing capabilities. The second solution focuses on addressing the challenges of resource-intensive deep learning models by investigating techniques for network compression and pruning. These methods aim to streamline model architectures without compromising on accuracy. By reducing the redundancy within neural networks and eliminating less critical connections, this approach not only minimizes computational complexity but also enables faster inferencing, making it well-suited for low-latency requirements in biomedical signal analysis. By allowing models to make informed decisions with varying degrees of confidence, the third solution not only optimizes computational resources but also ensures that the system remains adaptable to different signal scenarios using early exit points. This adaptability is particularly valuable in scenarios where maintaining high accuracy is paramount while managing the computational load efficiently. These three approaches present a comprehensive framework for advancing decentralized abnormal biomedical signal inferencing. Through this work, we aim to contribute to the realization of highly efficient, low-power, and low-latency biomedical signal analysis systems, fostering advancements in healthcare and remote monitoring technologies.enDecentralized inferingEarly exitsBiomedical signal classificationECG classificationDoctoral Thesishttps://creativecommons.org/licenses/by-nc-nd/3.0/ie/