Real-Time Fatigue Evaluation Using Ecological Momentary Assessment and Smartwatch Data: An Observational Field Study on Construction Workers

We conducted an observational field study to evaluate fatigue among construction workers. To measure both subjective and objective fatigue using a real-time approach, a smartwatch equipped with a researcher-developed EMA application was selected as the experimental equipment. The focus of the experimental design phase was to: (1) develop a measurement process for the integrated measurement of subjective and objective fatigue, (2) collect data from construction workers in the field and mine data for evaluation, and (3) evaluate the collected multidimensional (objective and subjective) data between fatigue groups. To verify the framework, low-fatigue and high-fatigue groups were compared (Fig. 1).

Objective fatigue was measured using physiological indicators, such as HR accelerometer and gyroscope data. HR has been commonly used as an indicator of people’s level of fatigue in their daily life (Bishop et al. 2018; Lee et al. 2020). HR is the most commonly used physiological measure to assess and monitor ongoing work fatigue (Aryal et al. 2017; Chang et al. 2009; Hsu et al. 2008; Mehta et al. 2017; Umer et al. 2020). Although HR is an adequate factor for fatigue assessment, physical demands (i.e., activity, energy) can increase it (Chan et al. 2012), and it can be influenced by changes in body posture (Anwer et al. 2021). Therefore, some factors related to physical demands should be considered when HR intends to be used for fatigue assessment (Anwer et al. 2021). This study uses activity value estimated from an accelerometer as a physical demand indicator for complementation HR. The relationship between fatigue and activity has been previously verified in various groups (Lamberts et al. 2010), and activity was found as a crucial component of fatigue routine assessment (Van der Werf et al. 2000). As shown in Table 1, physical demands criteria employed in earlier research were energy (Lee et al. 2017), calories (Guo et al. 2017), and kinetic data (McDonald et al. 2016). Objective factors that combined HR with activity could improve the validity of fatigue assessment. The overall subjective fatigue was self-reported twice, before and after the experiment, using the Korean version of SOFI. In addition, real-time subjective fatigue was reported by the participants using the EMA application.

The experiment period included one day for setting up and adapting to wearing the equipment, three days for data collection, and half a day (6 h) for collecting the equipment and ensuring data loss prevention. Data collection was conducted in small groups at different construction sites, with different collection times depending on the environment and circumstances of each site. We recruited a convenience sample of 100 construction workers at five construction sites in Korea between July and November 2020. The inclusion criteria were (1) $\mathrm{being}\ge 19\mathrm{years}$ old, (2) being Korean, and (3) having at least six months of work experience in construction. Participants received gifts worth US $100 for completing three days of data collection. The researchers described the experiment procedure in person to each participant individually using the smartwatch and provided illustrative materials, including a guide and video demonstration to fully understand the experimental method (Appendix S3). All participants provided written informed consent, and the institutional review board of the affiliated university approved the study. There was no difference in baseline information depending on recruiting seasons and sites.

As seven participants had missing data in the baseline SOFI, 93 participants were classified into two groups based on the degree of general fatigue. The evaluation framework was validated by classifying the participant into two groups based on the SOFI average score ($\mathrm{mean}=1.56\pm 1.28$) considering the overall fatigue level (Hernandez Arellano et al. 2015; Lee 2016). The high-fatigue group ($\mathrm{mean}=2.64\pm 0.85$, $\mathrm{n}=41$) was above the mean, and the low-fatigue group ($\mathrm{mean}=0.77\pm 0.42$, $\mathrm{n}=52$) was below the mean (Fig. 2).

Data preprocessing began with downloading raw data and its conversion into analyzable data. Triaxial data were calculated as a degree of activity count. Based on the characteristics of the signal data collected in the time series at every 1-s interval, null or abnormal values 60 s or more in a row due to the user or device error were checked and excluded from the analysis. If a participant removed the device, the triaxial accelerometer recorded values of 0 to indicate the duration of time for which it was not worn. Natural human behavior involves micromovements sensed by the accelerometer even during sleep; therefore, periods with a continuous absence of movement indicated device removal.

The raw HR and accelerometer data were aggregated into 1-min epochs optimized for aggregating this data using a script written in Python. Data were excluded from the analysis when (1) there were outlier measures in 60 s despite the indication of wearing status or (2) more than 50% of a participant’s data were missing in one day. In the data processing phase, 17 subjects, from the high-fatigue group ($\mathrm{n}=8$) and the low-fatigue group ($\mathrm{n}=9$), were excluded due to missing EMA reports, devices, and networking programs. Finally, data from 76 participants (high-fatigue group: $\mathrm{n}=32$, low-fatigue group: $\mathrm{n}=44$) were used for analysis (Fig. 2). The average age of the participants was $46\pm 11.29\mathrm{years}$. Most participants ($\mathrm{n}=74$, 97%) were men, and approximately half of the participants had up to high school education ($\mathrm{n}=35$, 46%). On the basis of the occupational fatigue assessment, 42% ($\mathrm{n}=32$) of the participants were classified into the high-fatigue group. Raw sensor data were processed using the aggregation, data selection, and reaggregation steps to standardize the assessment outcome.

Activity was calculated using an accelerometer in the smartwatch. Among the data collected by the smartwatch sensors, an accelerometer detects the degree of movement, whereas a gyroscope has been mainly used for motion recognition through movement patterns. We focused on measuring the degree of activity rather than patterns among construction workers, as work types were diverse. The activity calculation process (Fig. 3) comprised three steps: (1) aggregating raw accelerometer data at an interval of 60 s, (2) adjusting the weight of the three-axis of difference value (Kim et al. 2021)), and (3) converting difference value to energy value using signal vector magnitude.

The fatigue evaluation process was designed based on the collected physiological and self-reported data. Objective factors were evaluated as physical fatigue, and the subjective factor was evaluated as mental and cognitive fatigue. The evaluation was performed with three factors, the EMA score as the subjective factor and HR and activity (physiological indicators) as objective factors collected from the smartwatch. Physical status evaluation using HR and physical routine awareness using activity value are two interpretations of the assessment of physical fatigue. The process contained three parts of (Fig. 4): (1) interaction between each factor, (2) congruence between subjective and objective fatigue in each group, and (3) overall fatigue patterns between high- and low-fatigue groups.

The Mann–Whitney U test was performed to identify daily fluctuations in EMA, activity, and HR and compare the mean differences between the two groups at a specific time (Table 3). Although EMA reports were collected hourly, the initiation of individual report times slightly differed. Therefore, the results are provided as time frames to ensure a sufficient number of samples. Time frames were determined based on the general break and lunch times of construction workers. The high-fatigue group reported higher EMA scores throughout the day. However, the high-fatigue group had lower levels of activity and HR in most frames compared with the low-fatigue group.

Furthermore, both groups showed the highest EMA scores at the end of the workday, between 3 p.m. and 5 p.m. This indicates that workers felt greater fatigue when finishing work. The highest levels of activity and HR in both groups were observed between 12 p.m. and 3 p.m., followed by 7 a.m. to 10 a.m. The high-fatigue group had higher EMA scores and lower HR and activity than the low-fatigue group in all time frames.

EMA scores were processed as follows: (1) exploring fatigue awareness patterns based on each participant’s daily EMA scores, (2) confirming congruence between subjective and objective fatigue, and (3) examining differences in the overall patterns between the groups. Fig. 5 shows changes in EMA scores, activity, and HR during work for all participants over three days. Changes in activity and HR were similar. As activity increased or decreased, HR followed the same pattern, indicating a correlation between the two factors.

As shown in Fig. 5, the fatigue pattern of construction workers differed in the morning and afternoon. In the morning, EMA scores increased slightly as working time passed (0.71 to 1.67) and activity constantly stayed high, whereas in the afternoon, EMA scores increased and activity decreased, indicating concurrence of subjective and objective fatigue. In particular, the time frame between 3 p.m. and 5 p.m. demonstrated a sharp increase in subjective and objective fatigue. Furthermore, unique fatigue patterns and significant changes in activity were reported for 7 a.m., which is the start of the work day, and 1 p.m., which is the usual time of returning to work after lunch. This pattern implies that resting or sleeping helps relieve physical fatigue and indicates the need for rest before 3 p.m. to 5 p.m. when there is a rapid increase in objective fatigue.

The contrast between the two fatigue groups is shown in Fig. 6 to examine (1) the congruence between subjective and objective fatigue, and (2) the difference in overall pattern between the high- and low-fatigue groups.

This study successfully constructed a framework for measuring fatigue among construction workers in real-time using subjective and objective observations. Our evaluation framework was constructed to (1) perform in actual construction fields and not in laboratory settings (Anwer et al. 2020; Umer et al. 2020; Zhang et al. 2019); and (2) propose a multidimensional system for measuring and evaluating subjective and objective fatigue, which has only been performed separately in previous studies (Lee et al. 2017; Mehta et al. 2017; Techera et al. 2019). Particularly, this study provided a method to evaluate subjective fatigue in real-time, which was previously assessed using survey methods. Further, the result indicates that defining fatigue groups is valid in this study similar to previous studies (Chen et al. 2003; Pelders and Nelson 2019; Zhang et al. 2015).

The framework has weaknesses when compared to existing studies. First, the framework could not reflect diverse physical data as existing studies (Aryal et al. 2017; Guo et al. 2017; Ueno et al. 2018) have done, as we used only smartwatches to collect data. Especially, we used raw heart rates measured on the PPG sensor rather than HR variability. Second, the framework could not consider a predicting phase to assess fatigue automatically (Techera et al. 2018; Umer et al. 2020; Yu et al. 2019). Defining a formula for increased subjective fatigue (EMA score) based on increased physical fatigue (HR, activity) is a challenge, as the construction sector has too many variables, such as various work types, career deviation, and weather, to generalize. Despite these limitations, this study has made significant progress in that measurement and evaluation were attempted in actual work conditions with 100 on-site construction workers, and multiple aspects of fatigue were reflected using EMA and smartwatches. This study performed field data collection. Therefore, field conditions may have caused some data collection issues. First, field data collection may be affected by seasonal effects. Data collection was planned for the summer season when fatigue is the highest over a short period (Ueno et al. 2018). However, there is significant variability of human resources at construction sites, and data collection was delayed. To improve the generalizability of the results of the proposed framework, seasonal effects were examined. Experiments were divided into two groups: the summer and the autumn groups. We checked the seasonal effects by comparing the baseline SOFI result between the two groups (Table 5). The seasonal effect on the baseline SOFI result was not significant; to be specific, baseline fatigue was not affected by seasonal differences.

Moreover, the framework could not be performed using time-stamp matching of all EMA responses. The time when the miss-match occurred was mainly in the morning, which was caused by a time-stamp matching error. This occurred due to time lag caused by several conditions at each construction site (i.e., work start time difference, construction progress difference). Therefore, EMA response time was not controlled inside the experiments, and a time discordance occurred in work start times. Thus, further studies may consider time-stamp matching because unintended external variables could be controlled.

Currently, the construction site’s fatigue management entirely depends on the subjective fatigue reports of the workers. Many researchers have attempted to improve the current fatigue management by investigating objective fatigue assessment factors (Table 1). This study also attempted to consider physical, mental, and cognitive aspects of fatigue and suggested a multidimensional approach using EMA and smartwatch. However, there is still a limitation in that all fatigue elements have not been covered. Therefore, the discrepancy between subjective and objective fatigue may still occur. This is because of the inevitable constraints, such as high variability of sensor data, external variables (i.e., harsh environment, work type), and restrictive information. The causes of discrepancy should be improved for a more accurate assessment. The multidimensional approach for fatigue assessment is still in the preliminary stage, and further studies should be conducted based on the findings of this study.

As mentioned earlier, evaluating objective and subjective fatigue is complicated, and fatigue patterns during work have not been considered in-depth. To address these limitations, this study developed a framework to evaluate the fatigue of construction workers using EMA and smartwatches. Fig. 8 exemplifies the utilization of the proposed framework by presenting the overall fatigue pattern and fatigue level at each time section. This facilitates a relative comparison of the overall fatigue pattern and fatigue level in each time frame when compared to the average. This can be used to investigate how severe the current fatigue is when compared generally. It can also be used to investigate fatigue patterns in a specific work environment. The result enables the establishment of a fatigue management plan according to the multidimensional fatigue assessment for construction workers. Thus, the proposed framework provides a more systematic and realistic process and can facilitate fatigue management in the construction field.

From an academic point of view, the proposed method verified that subjective fatigue can be assessed using EMA. Further research can be actively conducted, as it can be used as a real-time self-reporting method for construction workers’ physical, mental, and physiological conditions. Especially, it can be extended to research on developing integrated fatigue indicators combined with various objective fatigue factors. As work schedule generally does not consider fatigue among construction workers in the field, fatigue management planning can be established by referring to fatigue patterns, time-checking implications derived in this study, and objective factors, such as HR, skin temperature, activity, energy, and calories, which have been verified in previous studies (Anwer et al. 2020; Guo et al. 2017; Umer et al. 2020; Zhang et al. 2019). Moreover, it can contribute to lowering the possibility of accidents caused by human factors (He et al. 2020; Techera et al. 2018). This study contributes to the literature and knowledge on work-related fatigue management by proposing a framework for evaluating workers’ fatigue due to ongoing work.

Findings derived from this study also make some practical implications for managing construction workers’ health and safety. The significance lies in the following three aspects. First, to manage workers’ fatigue systematically, time sections need to be considered in the assessment interpretation (Fig. 8). The fatigue pattern is different for each time section, and fatigue management according to the characteristics of each section is required as follows. (1) Early morning: owing to personal health status affecting fatigue, it is necessary to check whether they had sufficient sleep and drinking the previous day. (2) 3 p.m.: is the time when physical fatigue increases rapidly. Managers should observe workers’ subjective fatigue and objective pattern and then continuously check their fatigue severity. Second, restrictions on work time should be adopted and implemented in the law aspect, like the n-hour workweek rule. Overwork (i.e., early work, night work) cause prolonged work hours and increases workers’ fatigue ultimately. Therefore, the manager should carefully monitor how many hours their workers are working per week. The construction industry has long work hours and physically demanding tasks, and overwork is a crucial issue for worker safety. Based on the findings, this study discovered that prolonged work without rest severely affects the two fatigue types and that timely checking of workers’ fatigue is needed. Finally, a health manager who can perform appropriate medical aid at construction sites should be employed regardless of the project size. To apply the multidimensional approach in real construction sites, the health manager can help make medical decisions or daily health checking, and manage workers’ fatigue factors.

Because of these results, one can conclude that existing assessment methods (e.g., only sensing, survey, and self-report) for construction workers’ ongoing work fatigue were insufficient for screening workers at risk for unsafe practices and accidents at sites. The current system determines the unsafe status of workers on the basis of age, blood pressure, and experience of site supervisors (Kamardeen and Hasan 2022; Pereira et al. 2020). Establishing advanced management is required because the current practices have limitations to prevent accidents caused by human errors and work-related fatigue is increasing (Techera et al. 2019; Wong et al. 2019a). Our multidimensional approach using EMA apps and wearable devices provides a possible solution by providing real-time data integrated with the objective and subjective information of construction workers in the field. By combining subjective and objective data, the integrated information also enhances the construction site’s current fatigue management strategy. The multidimensional approach can also be utilized for innumerable purposes, ensuring safe working conditions by reflecting fatigue patterns like monitoring high-risk workers, making decisions about work schedules, and innovating new methods for construction duration calculation.