Energy Estimation of Sleep Monitoring Algorithms for Application Specific Integrated Circuit Implementation

Sleep monitoring is crucial for a variety of health issues. The emergence of wearable technologies allows access to valuable sleep data to a wide range of users. For such wearable technologies to be unobtrusive and practical for daily living, their size and weight need to be minimized. Consequently, the work presented in this paper is with the ultimate goal of the implementation of a wireless wearable sleep monitoring device using Application Specific Integrated Circuits (ASICs). Sophisticated machine learning algorithms are commonly used for sleep stage classification, however, the high computational cost of some of these algorithms leads to high electrical power consumption and the need for large and heavy batteries. This paper investigated the energy consumption of four of the commonly used classification algorithms used for sleep monitoring, i.e. k-nearest neighbor, neural network, logistic regression, and random forest, based on the low-level counts of arithmetic and relational operations and simulated energy consumed by required operations for an ASIC hardware implementation. The method used in this paper ensures software-implementation independence of the estimated results by investigating low-level operations. Based on the assessments performed on a publicly available sleep dataset, results showed that the logistic regression and neural network classification algorithms were able to perform with considerably lower energy consumption when ASIC implementation is considered.