To improve the QoL of the elderly, it is essential to predict their stress states. In general, the stress state varies from day to day or time to time depending on what activities are performed and how long/strong. However, most existing studies predict the stress state using biometric data and specific activities (e.g., sleep time, exercise time and amount) as explanatory variables, but do not consider all daily living activities. Therefore, it is necessary to predict the stress state by linking various daily living activities and biometric information. In this paper, we propose a method to improve the prediction accuracy of stress estimation by linking daily living activities data and biometric data. Specifically, we construct a machine learning model in which the objective variable is the result of a stress status questionnaire obtained every morning and evening, and the explanatory variables are the types of daily living activities performed in the 24 h prior to the questionnaire and the feature values calculated from the biometric data during each of the performed activities. The results of the evaluation experiments using the one month data collected from five elderly households, show that the proposed method (using per-activity biometric features) improves the prediction accuracy by more than 10% from the baseline methods (with biometric features without considering activities).
1 Introduction
With the aging of the population in developed countries, there is an urgent need for effective measures to promote the care of the elderly and the extension of their healthy life expectancy. However, there are many elderly people who wish to continue to live independently at home. To grant this wish, it is important to build an environment in which elderly people can understand their daily activities and manage their health conditions by themselves. There have been many studies on activities sensing/recognition technologies in the home and health condition monitoring using these technologies to build an environment for self health management and improve lifestyle habits of the elderly [1].
In order to provide monitoring services and improve the lifestyle of the elderly, it is necessary to develop indicators of health status and investigate the factors that cause these indicators to change. Currently, the World Health Organization (WHO) has developed a questionnaire index called WHOQOL-100 [2] as a representative indicator of health status. However, the method of measurement using such a questionnaire index places a heavy burden on respondents due to a large number of questions. For this reason Amenomori et al. [3] conducted a study to measure HRQOL (Health Related Quality of Life) with less burden by using devices such as smartphones and smartwatches. They aimed to improve HRQOL which is strongly related to physical and mental stress states, by continuously measuring HRQOL and detecting signs of stress early to prevent it.
Advertisement
Stress estimation has also been studied using devices rather than questionnaires to reduce the burden of estimation. Stress estimation using devices has been applied to many situations such as daily life [4‐6]. Fukuda et al. [4] collected wake-up times and sleep data obtained from wearable devices and answers of occupational health and safety questionnaire called the Depression and Anxiety Mood Scale (DAMS) in office workers for two to three weeks at a general company, and constructed a machine learning model to estimate the questionnaire answer. In addition, Natasha et al. [5] proposed a method to quantitatively measure and predict the health, stress, and happiness of the next day using smart devices, respectively, in order to understand the stress state at work.
In this paper, as a method to improve the prediction accuracy of stress estimation, we propose a method to predict stress by linking daily living activities data and biometric data. Specifically, we construct a machine learning model with the results of a stress status questionnaire obtained every morning and evening as the objective variable, and the types of daily activities performed in the 24 h before the questionnaire and the features calculated from biometric data during each daily activity as the explanatory variables. As biometric data, we use the Lorenz plot area calculated from the heart rate data collected by a smartwatch (Fitbit). The Lorenz plot area is known to be useful for stress estimation because it can visualize the activity level of the parasympathetic nervous system [7].
To evaluate the effectiveness of the proposed method, we applied the method to the dataset [1] consisting of daily living data, biometric data, and stress status questionnaires (two questions in the morning and two questions in the evening) collected from five households of elderly people over 60 years old for one month. To confirm the validity of the per-activity biometric features, we compared the baseline method 1 (using 24-h RRI variance and Lorenz plot area [7, 8] as features), the baseline method 2 (adding sleep time, which has been validated in the previous study [4], as a feature to the baseline method 1), and the proposed method (using RRI variance and Lorenz plot area for each activity type as features). For three class classification, the accuracy of the four questionnaires on average was 34.5% for the baseline method 1 and 47.25% for the baseline method 2, while the average accuracy of the proposed method was 58.25%, confirming the effectiveness of considering the per-activity biological features.
2 Related Research
This section surveys existing studies on QoL estimation, stress estimation and health management using smart home technologies.
Advertisement
2.1 QOL Estimation
Quality of life (QoL) is a measure of satisfaction and quality in the daily life. The QoL [9] research originally started as a concept to discuss the quality of life after treatment in the medical field, but it is now used not only in the medical field but also as a concept related to the quality of life in general, such as work-life balance and happiness.
In particular, QoL, which is directly related to human health, is called HRQOL (Health Related Quality of Life), and is evaluated by categorizing it into various domains such as physical, psychological, social interaction, economic and occupational, and religious and spiritual states. The World Health Organization (WHO) has developed various indicators to quantitatively assess HRQOL, such as WHOQOL [2] and Short Form [10]. These indicators are assessed using paper questionnaires. However, the WHOQOL-100 [2] requires 100 items in 6 domains, while the SF-36 [10] requires 36 items in 8 domains. The labor to answer these questions make it difficult to assess the quality of life on a daily basis.
Amenomori [3] et al. have proposed a method to continuously measure HRQOL using mobility and biometric information obtained from smartphones and smartwatches, and have shown that HRQOL can be predicted using a small number of questionnaires and information from smart devices. They aimed to improve HRQOL by continuously measuring HRQOL, which is strongly related to physical and mental stress states, to detect early signs of stress and to prevent it.
Unlike Amenomori’s work, the goal of our work in this paper is to detect and prevent causes of daily living stress in the elderly at an early stage. For this goal, we try to use per-activity biometric information to predict the stress state so as to understand which activity causes the stress.
2.2 Stress Estimation
There have been many studies on stress estimation using devices in many situations. There are two groups of stress estimation. The first group is the studies conducted in a controlled laboratory environment [11]. In these studies, the researchers intentionally generated stress using some kind of stress test, where the researcher has complete control over the stress level, and high-stress detection accuracy (80–97%) is usually reported.
The second group is studies that analyze stress in real life [4, 5, 12, 13] which have reported low accuracy [14]. Therefore, various systems are being researched to improve the prediction accuracy. To analyze stress in workers, not only smartwatches but also chest-mounted heart rate sensors (e.g., WHS-3 [15]) have been used [16]. However, systems that use multiple wearable sensors place a high burden on the subject and are not suitable for real-time stress detection in daily living. Therefore, it is desirable to use wearable devices as minimum and unobtrusive as possible.
2.3 Health Management in Smart Home
Since the elderly spend more time at home than out of the house, there has been development of systems to manage health conditions using activities data in the home, such as smart homes. Some smart home research has attempted to correlate activities and predict the well-being of occupants within a living space. The goal of Intel Research’s Computer-Supported Coordinated Care (CSCC) project [17] is to identify the care network characteristics and needs of older adults who wish to live at home. In addition, Jakkula et al. [6] aimed to identify trends in health status over time, with the goal of creating a smart home system focused on next-generation health care capable of home health care. However, most existing studies related to stress estimation in the smart home have been conducted in a controlled smart home, not in the subject’s home or other ordinary residence. Therefore, there are few studies that analyze stress in the home in real life.
In addition, in most of the related studies described above, the sensor data and feature values used do not take into account all aspects of the subject’s home life. Therefore, if there is no information related to activities of daily living, it is not possible to understand the activities that causes stress, and it may not be possible to improve the stress state. Therefore, it is necessary to incorporate data related to activities and biometric information in the home.
3 Proposed Method
In this study, we propose a method for estimating the stress state at the end of a day (or the beginning of the next day) from the values of the stress index for each of daily living activities which are performed in a day. In the following sub-sections, we describe the method of obtaining daily living activities and stress indicators used in the proposed method, the definition of the stress state to be predicted, and the method of constructing the model.
3.1 Acquisition of Daily Living Activities
In this study, we assume an environment in which the activities of residents can be automatically obtained by using an in-home daily living activities recognition system (such as SALON [18]). We target primary activities such as cook, meal, rest, work, clean, wash, go out, bath, and sleep. We assume that the in-home activity estimation system records the time of day when each activity is performed. Table 1 shows an example of activities log. The SALON system does the following to stabilize the annotation accuracy. The SALON system places annotation buttons near the activity location with an aim to improving annotation accuracy. When a user forgets to press the annotation button, annotation data is corrected by the user.
3.2 Acquisition of Stress Index
In this study, we assume that residents wear a wearable biometric sensor (ECG) that can measure heart rate and heart rate variability. In the proposed method, we focus on the RRI variance and Lorenz plot area as stress indicators that can be calculated from heart rate variability data [7, 8].
In a Lorenz plot (LP), denoting RR interval at time t by RRI(t), a dot is plotted at the coordinate \((RRI(t), RRI(t+1))\) in a two dimensional plane. The area of the ellipse that covers the plotted dots is calculated as the Lorenz plot area as shown in Fig. 3.
On the \(y=x\) axis, let m denote the mean of the distance from the origin (0, 0) (on the \(y = x\) axis), \(\delta _x\) the standard deviation from the origin. On the \(y = -x\) axis, let \(\delta _{(-x)}\) denote the standard deviation from the origin (0, 0). In this case, the area of the ellipse with major axis \(\delta _x\) and minor axis \(\delta _{(-x)}\) is calculated as follows:
3.3 Association of Activities with Stress Indicators
In the proposed method, for each activity obtained in Sect. 3.1, the RRI variance and Lorenzplot area are calculated from the heart rate data while the activity is performed by the method described in Sect. 3.2. If the same activity is performed repeatedly in different time intervals (e.g., Cook, Meal and Rest in Table 1), the biometric data measured during those intervals are integrated and stress indicators are calculated for the integrated data. Finally, per-activity stress indicators values are calculated as shown in Table 2.
Fig. 2.
Tabulation procedure for morning and evening questionnaires.
×
3.4 Prediction Model Construction
In this study, we construct a machine learning model to predict the stress state where we used as groundtruth the answers for the questions that residents give at the end of the day, from per-activity stress indicators values explained in Sect. 3.3.
4 Evaluation Experiment
For the evaluation experiment, we collected a dataset consisting of activities data, biometric data and questionnaire answers. In each home of the elderly participants, we set up sensors (up to 10 motion and ambient sensors and a few door sensors), five annotation buttons (corresponding to activities) and a server of a SALON shown in Fig. 1. In the experiment, first we extract features from the dataset, and then construct a random forest classifier to predict the stress level.
Fig. 3.
How to calculate the Lorenz plot area.
×
4.1 Dataset
We collected the dataset from five general elderly households (one single, four married couples, all in 60’s), which consists of daily living activities data, biometric data, sensor data, and stress state data (obtained through questionnaire). The data collection period is one month for each household. The details of each data are described below.
Daily Living Activities Data. We collected data on five typical daily living activities (bathing, cooking, eating, going out, sleeping, and other). Residents record the start and end of each activity by pressing the annotation buttons Shown in Fig. 1 installed at the locations where the activities are performed.
Biometric Data. We use the heart rate data collected by a smartwatch (Fitbit Alta HR) as biometric data. Fitbit collects the heart rate [bpm] once every 15 s for one minute.
We convert the heart rate to RRI [ms] and generate features for stress estimation, i.e., the variance of RRI and Lorenz plot area. The conversion equation for RRI is shown in Eq. (2).
where RRI is defined as the heartbeat interval in ms, and Heartrate is the number of heartbeats per minute.
Stress State Data. Stress state data is collected by asking residents to fill out questionnaires immediately after waking up and before going to bed each day. We adopted four questionnaire items related to mental stress and physical stress. In the questionnaire, a five-point Likert scale was used. The specific questionnaire items are shown below.
Table 1.
Activities log
Time
Activity
−7:00
Sleep
7:30–8:00
Cook
8:00–8:30
Meal
8:30–9:00
Rest
9:00–18:00
Go Out
18:00–18:30
Cook
18:30–19:30
Meal
19:30–21:00
Rest
21:00–21:30
Bath
21:30–23:00
Rest
23:00-
Sleep
Table 2.
Per-activity stress indicators
Activity
RRI variance
LP area
Sleep
0.8
1.0
Cook
0.3
0.5
Meal
0.2
0.4
Rest
0.3
0.3
Go out
0.1
0.3
Bath
0.5
0.7
Sleep
0.9
0.9
Questionnaire Items Immediately After Waking Up
M1. Did you feel physically refreshed this morning?
M2. Did you feel mentally refreshed this morning?
Questionnaire Items Before Bedtime
N1. Do you experience any physical stress due to physical pain or discomfort?
N2. Do you experience mental stress?
4.2 Selection of Features and Baseline Methods
The proposed method uses per-activity biometric features of all activities performed in 24 h before the questionnaire. In this evaluation experiment, two other methods (baseline 1 and baseline 2) are set as baselines.
The common basic features used for all three methods are as follows:
Average RRI value in the last 24 h
Standard deviation of RRI in the last 24 h
Average RRI for 3 h after waking up
Standard deviation of RRI for 3 h after waking up
The specific features used for each method are as follows:
Baseline 1 Lorenz plot area in the last 24 h
Baseline 2 Lorenz plot area and sleep time in the last 24 h
Proposed method Lorenz plot area for each activity (up to 6 types), Lorenz plot area for each activity in 3 h after waking up (up to 6 types)
4.3 Model Construction and Validation
We use random forests to construct stress level prediction. This time, we divided the data into three subsets and used a three-fold cross-validation method for model training and validation. In constructing models, the answers (in 5 level likert scale) of the questionnaire were reorganized into three levels: good, bad, and neither good nor bad.
To confirm the effectiveness of the per-activity biometric features, we compared the baseline method 1 (using 24-h RRI variance and Lorenz plot area as features), the baseline method 2 (adding sleep time as a feature to the baseline method 1), and the proposed method (using RRI variance and Lorenz plot area for each type of activities performed in 24 h as features).
Fig. 4.
Confusion matrix for the N1 (physical stress) questionnaire at night.
×
5 Results
5.1 Effects of Different Features
The results of predicting the questionnaire results of three methods are shown in Table 3. The proposed method achieved prediction accuracy of 0.57, 0.56, 0.63, 0.57 for questions M1, M2, N1, and N2, respectively while baseline 1/2 did 0.32/0.47, 0.35/0.49, 0.33/0.54, and 0.38/0.39, respectively. This suggests that adding sleep time features (baseline 2) improves accuracy so some extent but adding per-activity biometric features (the proposed method) greatly improves the accuracy compared to baseline 1.
5.2 Stress Estimation Model Construction
The stress level prediction model is constructed using random forest using features described in Sect. 4.2. Each question’s answer obtained as stress state is used as the ground truth data. The stress state data is rated on a 5-point scale, but since it is only necessary to judge whether the stress state is good or bad, the scale is reorganized to 3 levels: good, neutral, and bad. A three-fold cross-validation method is used to validate the model.
Table 3.
Prediction accuracy of three methods for each question
Baseline Method 1
Baseline Method 2
Proposed Method
M1
0.32
0.47
0.57
M2
0.35
0.49
0.56
N1
0.33
0.54
0.63
N2
0.38
0.39
0.57
We show more detailed results of N1 prediction in Table 4 and the confusion matrix in Fig. 4.
Table 4.
Evaluation of recall and f-measure for questionnaire N1
baseline method1
baseline method2
proposed method
accuracy
33%
accuracy
54%
accuracy
63%
recall
f-measure
recall
f-measure
recall
f-measure
Bad
0.34
0.37
Bad
0.53
0.59
Bad
0.63
0.69
Good
0.35
0.40
Good
0.60
0.64
Good
0.76
0.70
Table 5.
Importance of nighttime questionnaire N1 (physical stress).
Feature
Importance
RRIMean
0.119
Lorenz plot _Bathing
0.114
RRIMean_3hour
0.112
RRIStd
0.093
Lorenz plot_Goingout
0.076
Lorenz plot_Sleeping
0.073
Lorenz plot_Eating
0.073
Lorenz plot_Other_3hour
0.065
Lorenz plot_Other
0.060
Lorenz plot_Cooking_3hour
0.052
Lorenz plot_Cooking
0.048
RRIStd_3hour
0.046
Lorenz plot_Eating_3hour
0.044
Lorenz plot_Goingout_3hour
0.026
5.3 Evaluation of Feature Importance
In order to identify the features that contribute to the estimation, we investigated their importance. Table 5 shows the contribution of features in the model to predict N1’s answer. In Table 5, the mean value and standard deviation of RRI had the high contribution (1st and 4th rank). Lorenz plot areas for some activities also show high contribution. Especially, Lorenz plot area for bathing activity is ranked in the second. We also confirmed that the Lorenz plot area of bathing often ranked high in other questionnaires. Therefore, it is suggested that the stress prediction accuracy could be improved with per-activity stress indicators more finely than just using biometric features. Also, it is suggested that the Lorenz plot during bathing is somehow related to stress.
6 Conclusion
In this paper, to improve the prediction accuracy of stress estimation, we proposed a method to predict stress level by linking daily living activities data and biometric data. Through the experiment using the dataset collected from elderly households, it is shown that the proposed method with per-activity stress indicators features in the last 24 h improves the prediction accuracy to a great extent compared to the baseline methods using only stress indicators without distinguishing activities.
As future work, we plan to apply the results of stress estimation obtained in this study to create a system that can predict QoL with high frequency and provide feedback on daily living activities based on the results of QOL estimation, aiming to build a system that encourages changes of activities to improve health state.
Open Access This chapter is licensed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license and indicate if changes were made.
The images or other third party material in this chapter are included in the chapter's Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the chapter's Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.
