Introduction
Chronic obstructive pulmonary disease (COPD) is a chronic disease in which prolonged exposure to noxious substances such as tobacco smoke and airborne pollutants (Vos et al. 2012) causes persistent inflammation of the lungs, resulting in decreased respiratory function. It is the fourth leading cause of death worldwide (Lozano et al. 2012). Its main symptoms include cough, phlegm, and shortness of breath during exercise. The 6 min walk test is currently used to determine the severity of COPD. The 6 min walk test requires patients to walk as far as possible for 6 min in a straight line over a distance of 30 meters and return (Holland et al. 2014). However, it is burdensome and exhausting for the patients and medical staff. Furthermore, the language used by medical staff to instruct patients affects the distance the patient walks in 6 minutes (Weir et al. 2013), and the instruction language is not standardized, making the 6 min walk test unfair and less objective. Therefore, another way to measure the severity of COPD should be conceived.
Using the UCSD Shortness of Breath Questionnaire 3 developed at the University of California, San Diego (Eakin et al. 1998), a previous study reported that COPD patients exhibited dyspnea even during light daily activities such as eating, standing up from a chair, brushing teeth, shaving, brushing hair, bathing, changing clothes, tidying, and weeding a garden (Satoh et al. 2009). COPD patients exhibit lower levels of physical activity because of dyspnea (Kawagoshi et al. 2011), and such COPD patients have a higher risk of death (Waschki et al. 2011). Thus, the severity of COPD is typically evaluated based on the level of physical activities that make patients feel shortness of breath. A questionnaire (Esteban et al., 2010) is commonly used to assess the level of such activities. However, this test is based on the recall of participants and not the objective (Shephard, 2003).
Recently, several wearable devices capable of measuring physical activity levels such as the Apple Watch, ActiGraph, Garmin, and Fitbit have been made available in the market. These wearable devices estimate the level of physical activities more objectively than questionnaires. Therefore, measuring the level of daily activities resulting in shortness of breath in COPS patients using a wearable device and comparing the activity level with that of healthy people of the same age could provide an alternative to the 6 min walk test used in hospitals.
Based on the above discussion, the long-term objective of this study is to develop a simple test method that determines the severity of COPD. Our strategy is based on the automatic detection of the type of activities patients perform and the measurement of the corresponding activity amount for each activity. The activity amount is to be subsequently compared with that of the healthy people of the same age to determine the severity of COPD in the patients. If patients suffer from severe COPD, they would frequently feel shortness of breath while performing the daily activities, implying that their activity amount would decrease compared with that of healthy people of the same age. Along with the activity type, the difference in activity amount between COPD patients and healthy people of the same age could be used as indicators of the severity of COPD. The classification of activities is realized through a machine learning approach using acceleration data acquired using an accelerometer. For example, Trost et al. (2014) classified seven types of activities of children wearing ActiGraph GT3X+ on their right-hip and non-dominant wrist. Ueda et al. (2015) classified six daily activities using two different devices. However, to the best of our knowledge, no previous studies have applied machine learning algorithms to distinguish daily activities that cause shortness of breath in COPD patients to measure the severity of COPD. The accelerometer should be able to identify not only the basic activities such as sitting, walking, and running that have been considered in a previous study (Sumikawa et al. 2018) but also the daily activities that cause COPD patients to feel shortness of breath. Therefore, in this study, we aim at clarifying the accuracy of the classification of activities that cause shortness of breath in COPD patients.
Methods
Study population
Forty-six healthy participants (21 males and 25 females) provided written informed consent for participation. The mean age of the participants was 21.2 ± 0.95 years (range: 20–23 years) (the ages of five participants were not recorded, and they were excluded from the calculation of the average age). This study was performed according to the Declaration of Helsinki. It was approved by the Ethics Committee of Okayama University (approval number R2003-006). Selection criteria included healthy people aged between 18 and 64 years. As per the exclusion criteria, those suffering from cardiac or respiratory diseases that expose people to the risk of exercise load have been excluded.
Procedure
Participants wore ActiGraph wGT3X-BT tri-axial accelerometers (ActiGraph Inc., Pensacola, FL, USA) on the dominant wrist and attached to an elastic belt on the hip (vertical side). ActiGraph is ISO-13485:2016-certified and FDA 510(k)-cleared Class II medical devices in the U.S. It is one of the most frequently used wearable devices in the world, not only for health management but also for research activities. Its size is 4.6 × 3.3 × 1.5 cm and weighs 19 g. It provides raw acceleration data probably containing valuable movement information, such as direction and orientation. Data are sampled at high frequencies from 30 to 100 Hz. In this study, the sampling rate was 100 Hz, and the idle sleep mode was disabled. The participants performed nine activities in the following order: changing clothes, sitting, standing, lying in a supine position, brushing teeth, moving luggage, going up/down the stairs, running, and walking. Activities performed between lying in a supine position and brushing teeth were categorized into “other movements.” The activities from changing clothes to moving luggage were performed in one room, those after going up/down the stairs were performed in the same building, and the walking and running activities were performed on the rooftop of the same building in sunny weather and in a gymnasium in rainy weather. All the activities were performed for 2 min and were self-paced wherever applicable. After all the activities were completed, the participants returned to the first room, removed the accelerometers, and changed back into their original clothes. Subsequently, they were provided a JPY 1,000 prepaid card as reward. One experimenter was always present during the experiments. One experimenter occasionally conducted experiments on two subjects simultaneously to shorten the experiment time. In such a case, cloth changing measurements were obtained sequentially, whereas measurements corresponding to brushing teeth and moving luggage were acquired simultaneously; while one subject was brushing teeth, the other subject was moving luggage. Thereafter, the subject who had brushed teeth moved luggage, and the other subject who had moved luggage brushed teeth. All the remaining movements performed by the two subjects were measured simultaneously in a different way from the above; the two subjects performed the same activities simultaneously.
Activity types
The nine activities selected for this study can be categorized into two types of activities: basic activities (sitting, standing, lying in a supine position, going up/down the stairs, walking, and running) that have been recognized by a tri-axial accelerometer in previous studies (Sumikawa et al. 2018; Trost, Zheng & Wong 2014) and daily living activities during which COPD patients reported feeling shortness of breath in the 24 item questionnaire (Eakin et al. 1998). Although we have not specifically tested, there are other daily activities due to which COPD patients tend to experience dyspnea, for example, eating a meal, bathing, and shaving.
Changing clothes
Participants took off their upper and lower clothes and thereafter changed back into other or the same clothes. They were instructed to remain still for approximately 10 s before and after changing clothes to enable improved understanding of the activity from the data during analysis.
Sitting
Participants remained seated in the chair for 2 min.
Standing
After sitting, participants stood up, and subsequently, the standing measurement was performed for another 2 min.
Lying in a supine position
After the participants were instructed to lie down on a sofa, a 30 s rest was ensured. Thereafter, the measurements corresponding to the activity of lying in a supine position for 2 min were acquired. Subsequently, they rose from the sofa and stood up. The movements from the standing position to lying down on the sofa and the movements of rising from the sofa after lying down in a supine position were not used for data analysis.
Brushing teeth
Participants held a toothbrush in their hands and brushed their teeth for 2 min. Thereafter, they were asked to rinse their mouth. Data during rinsing the mouth were not considered part of the overall brushing activity.
Moving luggage
Participants carried two textbooks weighing approximately 1.5 kg from a shelf approximately 150 cm above the floor to a sofa approximately 1 m away at a height of approximately 40 cm above the floor. They repeated this back-and-forth movement for 2 min.
Going up/down the stairs
Participants performed one round trip of going up/down the stairs from the first floor to the fourth floor of the building at their usual pace. The time from the first instruction to their return to the same place was measured as the activity of going up/down the stairs. Similar to the case of changing clothes, they were instructed to remain still for approximately 10 s before ascending the stairs and after returning to the same place.
Walking
Participants walked around the 30-m perimeter of the roof of the building or the gymnasium at a normal walking pace for 2 min. They were instructed to remain still to the extent possible for 10 s before and after the 2-min walk, similar to the case of changing clothes.
Running
Participants ran along the perimeter of the roof of the building or the gymnasium for approximately 30 m at the same pace as jogging for 2 min. They were instructed to remain still to the extent possible for 10 s before and after the movement, similar to the case of walking.
Data processing and feature extraction
Raw tri-axial acceleration data (100 Hz) (Figure 1) were extracted from wGT3X-BT using ActiGraph propriety software (ActiLife Version 6.13.4). The data were saved in raw format as .gt3x files and converted to .csv format for data processing. They were segmented into sequences of the 10 s non-overlap-time or frequency-domain features, as shown below (Liu, Gao & Freedson 2012): mean, standard deviation, variance, maximum, minimum, root mean square, signal magnitude area, correlation between axes, entropy, energy, kurtosis, skewness, median, interquartile range, and autoregressive.
Figure 1
Example of acceleration data. The red and blue lines correspond to the wrist and hip data, respectively. The alphabets indicate the activities; a: changing clothes. b: sitting. c: standing. d: lying in a supine position. e: other movements. f: moving luggage. g: brushing teeth. h: going up/down the stairs. i: walking. j: running. (a) acceleration along the horizontal axis. (b) acceleration along the vertical axis. (c) acceleration along the lateral axis. Green squares indicate activities that are suitably recognized (F > 0.9). Silver squares indicate activities that are not recognized (F < 0.7). Gold squares indicate activities that are not recognized by the hip classifier but are recognized by the wrist classifier.
Model training and testing
Owing to data acquisition failure caused by the wrong settings of wGT3X-BT, fifteen datasets were discarded for analysis; in other words, 31 out of the 46 datasets were used for analysis. First, the idle sleep mode was enabled, which made the device enter a low power state after 10 s of inactivity. Some data were missing because of the sleep mode of ActiGraph. After we realized this, we disabled the idle sleep mode. Moreover, we occasionally forgot to program ActiLife. Without setting ActiLife appropriately, ActiGraph does not collect data. Therefore, we discarded the data for these two cases. The dataset of 31 subjects was split into two non-overlapping subsets: a training subset comprising the data of 30 subjects and a test subset comprising the data of 1 subject. Label-encoding, an encoding technique for handling categorical variables, was used. Random forest classifier model (Breiman 2001) was trained on the training data. Random forest is a supervised learning algorithm that combines multiple decision trees. Decision trees categorize or make predictions based on how a previous set of questions was answered. This process is realized through a tree diagram, thus called decision trees. Random forest is used to solve regression and multi-class classification problems and applied to several fields, such as finance, healthcare, and e-commerce (IBM Cloud Education 2020). After training, the trained model was evaluated on the test data. The sequence of training and evaluation was repeated 31 times with different combinations of training and test subsets. Scikit-learn (Pedregosa et al. 2011), a Python module for machine learning, was used for model training and testing. The version of Python is 3.6.8, and that of scikit-learn is 0.23.2. The following functions of scikit-learn were used: train_test_split, accuracy_score, RandomForestClassifier, classification_report, and confusion_matrix. The hyperparameters of random forest include the following: n_estimators = 2,500, criterion = “gini,” max_depth = 30, min_samples_split = 2, min_samples_leaf = 1, and random_state = 42. Here, “n_estimators” is the number of trees in the forest, and its default value is 100. “criterion” is a function for measuring the quality of a split. The supported criteria are “gini” for Gini impurity and “log_loss” and “entropy” both for the Shannon information gain. Its default value is “gini”. “max_depth” is the maximum depth of the tree. If None, then nodes are expanded until all leaves are pure or until all leaves contain less than min_samples_split samples. Its default value is None. “min_samples_split” is the minimum number of samples required to split an internal node, and its default value is 2. “min_samples_leaf” is the minimum number of samples required to be at a leaf node, and its default value is 1. “random_state” controls both the randomness of the bootstrapping of the samples used when building trees and the sampling of the features to consider when looking for the best split at each node, and its default value is None (Scikit.ensemble.RandomForestClassifier, 2022).
Model evaluation
The mean F-value (Bull et al. 2018) over 31 training–testing iterations was used to evaluate the relative performance of the classifier. The followings are elements that we use to define the F-value:
We define the F-value as follows:
The F-value varies from 0 to 1. An activity with an F-value of 0.7 or higher was considered recognizable because an area under the curve (AUC) of 0.7 or higher was considered high or moderately accurate in previous studies (Fischer, Bachmann & Jaeschke 2003; Swets 1988). Confusion matrices were developed to summarize the wrong classification of each activity.
The subjects wore the wGT3X-BT on their wrist and hip. The following three approaches were adopted to evaluate the performance of activity classification: (1) using data from both the hip and wrist positions for prediction (wrist+hip classifier), (2) using data from only the wrist position for prediction (wrist classifier), and (3) using data from only the hip position for prediction (hip classifier).
Results
We investigated the accuracy with which each activity would be recognized using the acceleration data measured by the tri-axial accelerometer. We assessed that the activity was sufficiently recognizable when the F-value was 0.7 or higher according to the criterion proposed in the literature (Fischer, Bachmann & Jaeschke 2003; Swets 1988).
In the case of classification using the data from both the wrist and hip positions (wrist + hip), the F-values were the highest for all activities among the three combinations (wrist + hip, wrist, and hip) of the wearing position of the accelerometer. The average F-value was 0.7 or higher for eight out of nine types of activities: changing clothes, standing, lying in a supine position, brushing teeth, moving luggage, going up/down the stairs, walking, and running (Table 1). Particularly, the F-values for lying in a supine position and running were higher than 0.9, indicating that these activities were appropriately recognized. The F-values of sitting and the other movements were lower than 0.7, implying that these activities were not suitably recognized. In the case of using data from only the wrist position, the F-value was 0.7 or higher for the same eight activities of changing clothes, standing, lying in a supine position, brushing teeth, moving luggage, going up/down the stairs, walking, and running. In the case of using data from only the hip position, the F-value was 0.7 or higher for the five activities of lying in a supine position, moving luggage, going up/down the stairs, walking, and running. These results suggest that the recognition of most of the activities examined in this study, except sitting and the other movements, is feasible when the data from both the wrist and hip positions are used or when only the wrist-position data are used.
Table 1
Evaluation results. Emboldened values represent F-values of 0.7 or higher.
| PRECISION | RECALL | F-VALUE | ||
|---|---|---|---|---|
| Changing clothes | wrist+hip | 0.8588 | 0.7858 | 0.7897 |
| wrist | 0.7698 | 0.7165 | 0.7124 | |
| hip | 0.7202 | 0.4874 | 0.5509 | |
| Sitting | wrist+hip | 0.5923 | 0.5504 | 0.5192 |
| wrist | 0.5141 | 0.4157 | 0.4092 | |
| hip | 0.4606 | 0.4407 | 0.4171 | |
| Standing | wrist+hip | 0.7667 | 0.8088 | 0.7663 |
| wrist | 0.7311 | 0.7608 | 0.7229 | |
| hip | 0.5150 | 0.5030 | 0.4601 | |
| Lying in supine position | wrist+hip | 0.9977 | 0.9978 | 0.9977 |
| wrist | 0.8256 | 0.7374 | 0.7301 | |
| hip | 0.9957 | 0.9978 | 0.9966 | |
| Brushing teeth | wrist+hip | 0.9278 | 0.8415 | 0.8635 |
| wrist | 0.8324 | 0.8124 | 0.8090 | |
| hip | 0.6780 | 0.5645 | 0.5823 | |
| Moving luggage | wrist+hip | 0.8769 | 0.8972 | 0.8799 |
| wrist | 0.8335 | 0.8571 | 0.8397 | |
| hip | 0.7288 | 0.7933 | 0.7448 | |
| Going up/down stairs | wrist+hip | 0.9104 | 0.8405 | 0.8486 |
| wrist | 0.8023 | 0.7363 | 0.7281 | |
| hip | 0.8663 | 0.8216 | 0.8169 | |
| Walking | wrist+hip | 0.8857 | 0.9132 | 0.8922 |
| wrist | 0.7981 | 0.7756 | 0.7621 | |
| hip | 0.8948 | 0.8914 | 0.8828 | |
| Running | wrist+hip | 1.0000 | 0.9851 | 0.9923 |
| wrist | 0.9904 | 0.9777 | 0.9834 | |
| hip | 0.9786 | 0.9628 | 0.9682 | |
| Other movements | wrist+hip | 0.6322 | 0.6985 | 0.6245 |
| wrist | 0.5436 | 0.6506 | 0.5630 | |
| hip | 0.5127 | 0.6093 | 0.5289 |
Figure 1 presents an example of the raw acceleration data. An experimenter observed the participant activities and recorded them during the experiments that were synchronized with the accelerometer output (marked as a to j in the figure). Depending on the type of activity and on the position of the accelerometer, acceleration values fluctuate in a different manner (Figure 1).
Figure 2 presents the confusion matrix for the activity-type predictions provided by the wrist + hip, wrist, and hip classifiers. Most proportions of each activity window have been correctly classified, as shown in the diagonal of the confusion matrix. However, there are some parts that have been misclassified.
Figure 2
Confusion matrix: wrist+hip, wrist, and hip classifiers. The rows show the actual activities, while the columns show the activities predicted by the activity classifiers. The green underline indicates misclassification between sitting and other movements. The orange underline indicates misclassification between lying in a supine position and sitting, standing, or other movements. The red underline indicates misclassification between going up/down the stairs and walking. The purple underline indicates misclassification between changing clothes and moving luggage. Light blue underline indicates misclassification between sitting and standing.
Discussion
Aiming at the development of a simple test of COPD, this is a pilot study conducted to develop and examine a machine learning-based activity classifier for the wrist- and hip-worn accelerometers. When the wrist and hip classifiers were combined, most activities, except for sitting and other movements, were recognized: changing clothes, standing, lying in a supine position, brushing teeth, moving luggage, going up/down the stairs, walking, and running. These daily activities would cause shortness of breath in COPD patients. Measurements of the amounts of activity for each activity and comparing the measurements with that of healthy people of the same age may lead to the determination of the severity of the disease.
We demonstrated that the attachment position of the accelerometer is flexible and depends on the purpose and preference of users. It was proved that activities were recognized most accurately when the wrist and hip classifiers were combined. This result is reasonable because the data amount of the wrist and hip classifier was twice as large as that of the wrist or hip classifier. More acceleration data provided more information on activities that resulted in improved recognition performance. This result is in accordance with that of a previous study that examined the relation between the number and attachment position of sensors and activity recognition accuracy. It is demonstrated that attaching many sensors to different parts of a body can recognize both dynamic and transitional activities (Khan et al. 2018). Some studies also tended to attach multiple sensors for better characterization of activities (Baños et al. 2013; Bao & Intille 2004). Therefore, if we use the accelerometer for research or on the condition that patients do not hesitate to wear multiple devices, it would be more effective to wear the devices on both the wrist and hip. Nevertheless, even the wrist classifier recognized the same activities, although the F-values were lower than those of the wrist and hip classifier. While wearing the accelerometer at home, for convenience, it is recommended to not wear any devices. Moreover, the wrist-worn accelerometer could be incorporated into a popular smartwatch, such as Apple Watch, to spread the device. Therefore, a wrist-worn accelerometer is useful when measuring physical activities outside the hospital. In addition, even the hip classifier recognized five activities, indicating that we could incorporate the accelerometer into a pedometer to recognize activities. Some people wear the pedometer to track their steps per day. For these people, measuring physical activities using the hip-worn accelerometer would be effective. In conclusion, our data show that one could choose where to wear the accelerometer depending on the purpose and preference.
Acceleration fluctuations played a vital role in distinguishing an activity from others. Two activities, lying in a supine position and running, were particularly recognized better than the other seven activities. This is attributed to the fact that they include dynamic and unique movements that differ from the other activities, as shown in the acceleration plots (“d” indicates lying in a supine position and “j” indicates running in Figure 1). The acceleration values of running changed and fluctuated more significantly than the others. The center of oscillation of the acceleration values for lying in a supine position was higher than the others. The two activities of sitting and other movements were not recognized; this could be explained by their acceleration plots that are similar to each other (“b” indicates sitting and “e” indicates other movements in Figure 1). In addition, the three activities of changing clothes, standing, and brushing teeth were not recognized by the hip classifier but recognized by the wrist+hip and wrist classifier. This is because these activities involve large wrist movements, but not many hip movements, as confirmed in the acceleration plots (“a” indicates changing clothes, “c” indicates standing, and “g” indicates brushing teeth in Figure 1). On the contrary, the five activities of lying in a supine position, moving luggage, going up/down the stairs, walking, and running were recognized by both the wrist and hip classifier. This is because these activities involve both wrist and hip movements (“f” indicates moving luggage, “h” indicates going up/down the stairs, and “i” indicates walking in Figure 1). These results suggest that acceleration fluctuations are the key to the recognition of activities.
Misclassification errors would occur when the target activity includes similar body-movements to other target activities. Some misclassifications do exist although most activities have been recognized based on F-values. Inspection of the confusion matrix for the wrist+hip classifier revealed that a substantial proportion of the sitting windows were misclassified as other movements (underlined in green in Figure 2). Investigation of the confusion matrix for the wrist classifier demonstrated that some windows pertaining to lying in a supine position were misclassified as sitting, standing, or other movements (underlined in orange in Figure 2). Furthermore, some windows corresponding to going up/down the stairs were misclassified as walking (underlined in red in Figure 2). Exploration of the confusion matrix for the hip classifier showed that some windows corresponding to changing clothes and sitting were misclassified as moving luggage (underlined in purple in Figure 2) and standing (underlined in light blue in Figure 2), respectively. Regarding the misclassification of the other movements, several movements (the windows corresponding to changing clothes, sitting, standing, lying in a supine position, brushing teeth, and moving luggage) were misclassified as other movements. In this study, we defined other movements as activities between lying in a supine position and brushing teeth or moving luggage. We chose this region for other movements because we supposed that it did not include similar body movements with other target activities. However, assessing the confusion matrix, several similar movements may be included in this region. Therefore, the other movements should be carefully defined. In conclusion, the findings suggest that both wrist and hip classifiers might be vulnerable to misclassification errors when the target activity includes similar body movements to other target activities.
Our results show that it is possible to develop a simple test that determines the severity of COPD using accelerometer(s). Assessing the F-values, we were able to identify changing clothes, standing, lying in a supine position, brushing teeth, moving luggage, going up/down the stairs, walking, and running by combining the wrist and hip data or by using only the wrist data. It is important to evaluate these daily activities because they help us understand what patients do and feel during these activities. These daily activities are likely to cause shortness of breath in COPD patients; measurement of activity amount for each activity and comparing the measurement with that of healthy people of the same age may lead to the determination of the severity of the disease.
In this study, sitting and other movements were not well recognized. The reason of the misclassification of sitting is probably that there were no hand movements or hip movements during sitting. It was difficult for the classifier to classify movements of sitting, and they were incorrectly categorized into other movements. However, considering the results of an earlier study (Trost, Zheng & Wong 2014), it may be possible to identify sitting if there is a movement of the wrists as in writing letters or working on a computer. Moreover, the recognition accuracy of sitting could be improved by incorporating activities before and after sitting as new target activities. Nevertheless, owing to the extreme difficulty of the recognition of sitting, it need not be recognized because it is considered to be less physically demanding and less likely to cause shortness of breath than the other daily activities. If we exclude sitting from the target activities, deterioration of recognition accuracy becomes more probable. Since sitting is a basic activity commonly performed in daily life, excluding it from the target activities would often lead to the misclassification of sitting as other activities. However, further experiments are required to examine the effects of eliminating the activity of sitting. In addition to sitting, we could not recognize a few other movements as well because such movements comprised a variety of different movements.
Using sensors with more data types, such as a 9-axis accelerometer, is expected to improve the recognition accuracy. GT9X Link is an alternative product to wGT3X-BT used in this study, and this version houses not only a tri-axial accelerometer but also a tri-axial gyroscope and tri-axial magnetometer that measure the rotational velocity and magnetic flux, respectively. Gyroscopes have been applied to gait analysis and recognition of postures and locomotion (Zhang et al. 2011). Using a gyroscope and accelerometer at the hip and ankles improve the individual-level prediction of energy expenditure (Hibbing et al. 2018). It is reported that the body angle when people sit or stand is related to the accuracy of recognition of sitting and standing (Van Lummel et al. 2013). Therefore, using GT9X Link instead of wGT3X-BT provides more information on activities and might enable the recognition of activities more accurately.
This study has a few limitations. First, participants wore the accelerometer on their dominant wrist. People usually wear their watch on their non-dominant wrist; therefore, if accelerometers installed on a smartwatch are used, it is atypical to wear the accelerometer on the dominant hand. Because the non-dominant wrist moves less dynamically than the dominant wrist, the results deteriorate when wearing the accelerometer on the non-dominant hand. Second, all participants were healthy and young people. Generally, COPD patients move more slowly than healthy people, and they are middle or old age; thus, the results might worsen if actual COPD patients are included in the study as participants. Finally, the measurements of the movements were obtained under specific experimental settings. In daily life, people perform a wider range of movements than those considered in this study. Therefore, the actual accuracy of recognition is expected to deteriorate further.
Conclusions
To develop a simple test for the measurement of the severity of COPD, this study tested the accuracy of a tri-axial accelerometer attached to the wrist and/or hip in recognizing daily activities. When the acceleration data obtained by accelerometers positioned on the wrist and hip were combined, the activities of changing clothes, standing, lying in a supine position, brushing teeth, moving luggage, going up/down the stairs, walking, and running were recognized. These recognized activities tend to cause shortness of breath in COPD patients. By measuring the amounts of activity for each of these activities and by comparing the measurements with that of the healthy people of the same age, we could determine the severity of COPD. These findings show the possibility of developing a simple test that can be performed at home to measure the severity of COPD, which will promote home medical care. This study could also provide new insights for research fields on activity recognition and COPD.
Data Accessibility statement
Data used in the research has not been made available. This is because we did not obtain permission from the participants.
Acknowledgements
We thank K. Rai for proposing the idea of this research and all the participants for participating in our study.
Funding Information
This work was supported by JSPS KAKENHI Grant Number JP21K12787.
Competing interests
The authors have no competing interests to declare.
Author contributions
All authors conducted a literature search. M.M. and Y.Y. conceived the study design. Y.Y. recruited participants and collected acceleration data. T.Y. and W.N. processed the acceleration data, extracted features, and carried out machine learning. T.Y. wrote the manuscript. M.M. reviewed the manuscript.