A Center for Sustainable Cloud Computing

Analytics & Applications

Multimodal and Personalized Methods for Health and Wellness Monitoring

Self-awareness for dynamic energy management on mobile and wearable platforms.

Wearable sensors are widely used for monitoring health and wellness. Among them, edge wearable devices based on multimodal sensors use the concept of self-awareness, which refers to the capability of a system to evaluate its self-performance by interacting with the environment and take corrective measures. One of the rapidly developing applications for wearable technologies is daily monitoring of work-related stress including cognitive workload. This is especially critical during manual labor, where workers need to follow a specific procedure with time and quality constraints. Efficient monitoring of the process requires a fully wearable and minimally obtrusive system.

In this project, we adopt the notion of self-awareness for dynamic energy management on mobile and wearable platforms and propose a self-aware machine-learning technique to improve the lifetime of such systems without any major performance loss. A multimodal monitoring system for cognitive workload assessment comprises three mains steps: signal acquisition, preprocessing, and workload detection. In our technique, each signal is filtered and different biomarkers are extracted. From those biomarkers, several features in time and frequency domain are extracted and a reduced set of the most important features is used for workload detection.

We adopt the Random Forest algorithm for both feature selection and classification. In the first stage of feature selection, statistically irrelevant features are removed. Thereafter, we select the most important feature that shows a Pearson’s correlations coefficient higher than 0.99. In the final step, we apply the Recursive Feature Elimination technique (RFE) to select the most informative features for cognitive workload detection during manual labor.

Our evaluation of the proposed self-aware machine-learning algorithm demonstrates that it is able to distinguish among baseline, cognitive workload, and physical work with an 81.75% of gmean. The model performs similarly to the state of art accuracy. Moreover, the proposed two-mode system improves the energy consumption by 27.6% without any major loss in classification performance.

Suggested Reading

https://infoscience.epfl.ch/record/275468/files/StressWorkloadDetection_D_T__Final_.pdf