Self-Aware Machine Learning for Multimodal Workload Monitoring during Manual Labor on Edge Wearable Sensors

Abstract

The design of reliable wearable technologies for real-time and long-term monitoring presents a major challenge. Self-awareness is a promising solution that enables the system to monitor itself in interaction with the environment and to manage its resources more efficiently. In this work, we aim to utilize the notion of self-awareness to improve the battery life of edge wearable sensors for multimodal health and workload monitoring. Specifically, we consider cognitive workload detection during manual labor as a case study to illustrate the impact of our proposed technique in wearable technologies. Our multimodal machine-learning algorithm is able to detect cognitive workload during manual labor with a performance of 81.75%. By adopting the notion of self-awareness, we achieve an improvement of 27.6% in energy consumption, with less than 6% of performance loss.

SPARE: A Spectral Peak Recovery Algorithm for PPG Signals Pulsewave Reconstruction in Multimodal Wearable Devices

“…Table 2 reports the error rates on estimating the HR for the data set from [ 9 ] using as reference the HR from the ECG signal after applying SPARE and fastSPARE and the baseline methodology without MA removal. It can be seen that SPARE is able to reduce the mean absolute error ( ) in HR estimation from PPG by 45.05%.…”

“…The second data set includes data collected during an experimental session where subjects simulate manual labor (with asynchronous and sudden movements) [ 9 ]. Similarly, the third database (from [ 23 ]) includes data from subjects walking or running on a treadmill.…”

“…The validation relied upon two already existing data sets. The first one encompassed data from twelve subjects that were performing various activities, including tightening some screws in blocks of wood at different heights, walking, and solving arithmetic tasks (about 18 min per subject, ethical approval HREC:037-2019; see [ 9 ] for more details about the experimental protocol). The PPG sensor used was an Empatica E4 (with a sampling frequency of 64 Hz) and, for the ECG, a Shimmer3 ECG (with a sampling frequency of 512 Hz).…”

“…For this reason, it allows researchers and clinicians to evaluate various cardiovascular-related diseases, such as atherosclerosis and arterial stiffness [ 2 ], and can be even used for biometric identification [ 3 , 4 ]. Moreover, several biomarkers can be extracted from each pulsewave and correlated with cognitive workload, stress, and emotional state of subjects [ 5 , 6 , 7 , 8 , 9 ]. Indeed, these biomarkers allow assessing the physiological changes induced by both physical activity and cognitive tasks (e.g., cardiac response, blood volume, and peripheral blood vessel resistance).…”

AMSER: Adaptive Multi-modal Sensing for Energy Efficient and Resilient eHealth Systems

“…Other techniques target model optimization to reduce the compute intensity of sense-making algorithms [10]. Existing optimizations for multi-modal machine learning (MMML) based eHealth applications address multi-modal sensing and sense-making independently [11]- [13]. Improving end-to-end system metrics of performance, energy consumption, and prediction accuracy necessitates joint optimization of sensing and sense-making phases.…”