Energy-Aware Adaptive Ensemble Controller for Biomedical Signal Classification Using Dual-Model Inference

In recent years, edge computing has made realtime biomedical signal processing in the healthcare domain possible, especially for monitoring vital signs using portable devices. However, the majority of methods currently available are based on Deep Learning (DL) models, which are static and ignore differences in the constraints of the device, resulting in a performance that is not optimally efficient under constrained energy or processing environments. Hence, this research proposes an Energy-Aware Adaptive Ensemble Controller (EAEC) to intelligently adapt the inference process based on current system resources. Initially, Electrocardiogram (ECG) signals were attained from the Massachusetts Institute of Technology Beth Israel Hospital Arrhythmia Dataset. Then, the data are pre-processed with low-pass filtering and contrastlimited adaptive histogram equalization (CLAHE) to eliminate high-frequency noise and increase local contrast. that, a 1Dimensional Convolutional Neural Network (1D-CNN) and two bidirectional long short-term memories (Bi-LSTM) are initialized to obtain the best accuracy for more complex patterns. Finally, the selected models were executed for realtime prediction disease in order to maintain the best trade-off between energy prediction efficiency and diagnostic accuracy. The proposed EAEC achieved better results in terms of accuracy (94.872%) than the existing DL-healthcare disease diagnosis model.