Fatigue Monitoring Through Wearables: A State-of-the-Art Review

Abstract

The objective measurement of fatigue is of critical relevance in areas such as occupational health and safety as fatigue impairs cognitive and motor performance, thus reducing productivity and increasing the risk of injury. Wearable systems represent highly promising solutions for fatigue monitoring as they enable continuous, long-term monitoring of biomedical signals in unattended settings, with the required comfort and non-intrusiveness. This is a p rerequisite for the development of accurate models for fatigue monitoring in real-time. However, monitoring fatigue through wearable devices imposes unique challenges. To provide an overview of the current state-of-the-art in monitoring variables associated with fatigue via wearables and to detect potential gaps and pitfalls in current knowledge, a systematic review was performed. The Scopus and PubMed databases were searched for articles published in English since 2015, having the terms "fatigue," "drowsiness," "vigilance," or "alertness" in the title, and proposing wearable device-based systems for non-invasive fatigue quantification. Of the 612 retrieved articles, 60 satisfied the inclusion criteria. Included studies were mainly of short duration and conducted in laboratory settings. In general, researchers developed fatigue models based on motion (MOT), electroencephalogram (EEG), photoplethysmogram (PPG), electrocardiogram (ECG), galvanic skin response (GSR), electromyogram (EMG), skin temperature (T_sk), eye movement (EYE), and respiratory (RES) data acquired by wearable devices available in the market. Supervised machine learning models, and more specifically, binary classification models, are predominant among the proposed fatigue quantification approaches. These models were considered to perform very well in detecting fatigue, however, little effort was made to ensure the use of high-quality data during model development. Together, the findings of this review reveal that methodological limitations have hindered the generalizability and real-world applicability of most of the proposed fatigue models. Considerably more work is needed to fully explore the potential of wearables for fatigue quantification as well as to better understand the relationship between fatigue and changes in physiological variables.

Keywords: fatigue monitoring; imbalanced datasets; machine learning; occupational health and safety; physiological signal; signal quality assessment; validation; wearable.

Figure 1

PRISMA flowchart. Flow of information through the different phases of articles selection process.

Figure 2

Type of fatigue and technology application domain.

Figure 3

Wearables devised for fatigue monitoring. ECG, electrocardiogram; EEG, electroencephalogram; EOG, electrooculogram; GSR, galvanic skin response; IMU, inertial motion unit; NIRS, near-infra-red spectroscopy; PPG, photoplethysmography; RES, respiration; T_Sk, skin temperature. ¹Reprinted from Zhang et al. (2017). ²© (2015) IEEE. Reprinted, with permission, from Ko et al. (2015). ³Reprinted from Dhole et al., (2019). © (2019) with permission from Elsevier. ⁴© (2015) IEEE. Reprinted, with permission, from Li et al. (2015). ⁵Reprinted from Li and Chung (2015). ⁶Reprinted from Aryal et al. (2017). © (2017) with permission from Elsevier. ⁷Reprinted (adapted) with permission from Zeng et al. (2020). © (2020) American Chemical Society. ⁸Reprinted from Huang et al. (2018). © (2018), with permission from Elsevier. ⁹© (2018) IEEE. Reprinted, with permission, from Choi et al. (2018). ¹⁰© (2016) IEEE. Reprinted, with permission, from Ha and Yoo, (2016). ¹¹© (2018) IEEE. Reprinted, with permission, from Nakamura et al. (2018). ¹²Republished with permission of SAE International, from Niwa et al. (2016); permission conveyed through Copyright Clearance Center, Inc. ¹³Reprinted with permission of Fuji Technology Press Ltd., from (Wang et al., 2018). ¹⁴© (2016) IEEE. Reprinted, with permission, from Mokaya et al. (2016).

Figure 4

Physiological and motion signals for (A) mental fatigue, (B) vigilance, (C) drowsiness, (D) muscle fatigue, (E) physical fatigue monitoring and respective measurement locations. The fraction number in the boxes represents the number of studies on a specific signal based on the total literature reviewed on that type of fatigue. ECG, electroencephalogram; EEG, electroencephalogram; EMG, electromyogram; EOG, electrooculogram; EYE, eye movement; GSR, galvanic skin response; MOT, motion; NIRS, near-infra-red spectroscopy; PPG, photoplethysmogram; RES, respiration; T_Sk, skin temperature.

Figure 5

Risk of bias assessment. Studies with low risk of bias in all components were deemed to be of low risk of bias, studies with low or unclear risk of bias for all components were deemed to be of unclear risk and those with high risk of bias for one or more components were deemed to be of high risk of bias.