Introduction
Certain diseases are closely associated with the types of activities that patients perform in their daily lives. For example, individuals with chronic obstructive pulmonary disease (COPD) often experience breathlessness during routine activities such as brushing their teeth, eating, or climbing stairs (Satoh et al. 2009; Yousaf et al. 2019). Similarly, patients with arrhythmias may develop palpitations or chest discomfort when engaged in specific movements. Therefore, understanding the relationship between activity and symptom onset is crucial for the accurate assessment of disease severity and diagnosis in clinical practice.
Despite the clinical relevance, the evaluation of activity-related symptoms remains challenging. COPD severity is commonly assessed using spirometry (Agustí et al. 2023), symptom questionnaires such as the COPD Assessment Test (Kim et al. 2024), or functional tests such as the 6 min walk test (6 MWT) which can be physically demanding for patients with limited respiratory capacity and may be influenced by verbal encouragement from clinicians, thus reducing objectivity (Holland et al. 2014; Weir et al. 2013). For arrythmia, the 24–48 h Holter electrocardiogram (ECG) monitoring is often used in combination with patient-maintained activity logs (Chung 2013). However, these logs are burdensome to complete and inherently subjective, which can result in inaccuracies or omissions that compromise diagnostic precision.
Recent advantages in wearable sensor technologies that use accelerometers and other inertial sensors have enabled the automatic recognition of physical activities (Ankita et al. 2021; Ueda, Tamai & Yasumoto 2015; Bozkurt 2022). These methods can provide objective and continuous data on patient activities and offer promising alternatives to traditional self-reporting and manual assessments. In clinical settings, this technology can help clarify the relationship between symptom occurrence and specific movements, potentially improving the objectivity and reliability of severity assessment and diagnosis. For instance, one study successfully distinguished between sedentary and ambulatory activities during ambulatory ECG monitoring using a dual-axis accelerometer and position sensor (Luckhurst, Hughes & Shelley 2024).
Currently, inertial measurement units (IMUs) are widely used to achieve precise activity recognition. Modern 9-axis IMUs combine accelerometers, gyroscopes, and magnetometers to capture acceleration, angular velocity, and geomagnetic data at high sampling rates. These three components offer complementary information: acceleration reflects movement, vibration, and tilt; angular velocity captures rotational motion; and geomagnetic data provide directional orientation. Although these devices offer high accuracy, their relatively large size and high energy demands can hinder long-term clinical use.
In clinical settings, minimizing data volume is critical for several interrelated reasons. First, lower data loads allow faster processing and more efficient wireless network congestion. Second, the reduced data requirements decrease power consumption, enabling longer operation with smaller batteries. Third, compact devices with limited storage capacity benefit from lower data loads by extending the monitoring duration. Fourth, minimizing storage and power requirements contributes to overall device miniaturization and improves patient comfort and compliance. Finally, data reduction can lower device production costs, making widespread clinical adoption more feasible. These interrelated benefits make data reduction strategies particularly valuable in clinical applications that require long-term continuous monitoring, optimizing both technical performance and patient experience.
Although prior studies have examined the accuracy of activity recognition using different combinations of sensor axes (Wieland & Nigg 2023; Huang, Yan & Onnela 2022), few studies have focused on clinical use cases involving breathlessness in COPD or documentation in Holter ECG logs. To address this gap, the present study aimed to determine the extent to which data from 9-axis IMUs can be reduced while maintaining high activity recognition accuracy in clinically relevant contexts. Building on our previous work which demonstrated that data from a single wearable device placed on the nondominant wrist or chest can yield high classification accuracy (Yamane, Kimura & Morita 2024), we evaluated 9-, 6- and 3-axis subsets of IMU data collected from these two positions. In addition, we analyzed how each sensor component—acceleration, angular velocity, and geomagnetism—contributed to the recognition of activity types. As this study was conducted as a pilot, data were collected from healthy participants. We hypothesized that the 9-axis data would provide the highest accuracy, and certain 6- or 3-axis configurations might achieve comparable performance depending on the activity and sensor placement. This has important implications for the development of smaller, lower-power wearable devices that could support long-term patient monitoring and enhance the objectivity of clinical evaluations.
Methods
Experimental setup
Participants
Thirty healthy individuals (17 females and 13 males) were recruited for this study. The participants were recruited through advertisements posted on the university campus. Exclusion criteria included the presence of cardiovascular or respiratory disorders that could present a risk during exercise stress testing and pregnancy. All participants provided written informed consent prior to participation. The cohort had a mean age of 21.0 years (SD = 0.87; range: 19–23 years). Hand dominance distribution showed one left-handed and 29 right-handed individuals. This study was conducted in accordance with the Declaration of Helsinki and was approved by the Ethics Committee of Okayama University (approval number: R2203-001, on April 14, 2022).
Multi-sensor technology for human activity recognition
The ActiGraph GT9X Link (ActiGraph LLC, Pensacola, FL, USA) is an activity monitoring device equipped with a primary 3-axis accelerometer and an IMU. The IMU includes a secondary 3-axis accelerometer, 3-axis gyroscope, and 3-axis magnetometer, which are used to measure acceleration, rotational velocity, and magnetic flux, respectively. The IMU functionality enables the device to provide rotation and direction data for advanced applications. This device has been widely used in previous studies for human activity recognition (Niazi et al. 2017; Hendry et al. 2020; Leotta, Fasciglione & Verri 2021). The compact size and wearable design of the device render it suitable for long-term monitoring in clinical settings.
Multi-position wear protocol and structured activity sequence
Five ActiGraph GT9X Link devices were attached to each participant as follows:
One device was attached to the dominant wrist using a link wristband. The wristband was tightly secured to ensure minimal movement of the device.
One device was attached to the nondominant wrist using a link wristband.
One device was attached to the chest using a pouch and belt. The pouch was positioned on the sternum and secured using an adjustable elastic belt.
One device was attached to the hip contralateral to the dominant wrist using a link-belt clip.
One device was attached to the thigh contralateral to the dominant wrist using a pouch and belt. The pouch was positioned on the anterior aspect of the thigh midway between the hip and knee and secured with an adjustable elastic belt.
The IMU recordings were performed at a sampling frequency of 100 Hz for all positions. The idle sleep mode was disabled to ensure continuous data collection. The participants completed nine activities, each lasting 2 min at a self-selected pace, in the following order: lying in the supine position, standing, sitting, eating, brushing teeth, using the restroom, walking, ascending or descending stairs, and running. Activities between eating and brushing teeth were classified as ‘other movements’ because the activities in this section could not be clearly categorized into any of the other nine activities. A stopwatch was used for each activity to ensure accurate timing. The recorded times were then synchronized with the sensor data by aligning the stopwatch timestamps with the corresponding data points. This process enabled precise labeling of activities within the collected dataset.
Activity types
This study focused on nine activities from two distinct categories. The first category comprised fundamental activities (lying in the supine position, standing, sitting, walking, ascending or descending stairs, and running) identified using a 3-axis accelerometer in previous studies (Sumikawa et al. 2018; Trost, Zheng & Wong 2014). The second category encompassed specific activities of daily living, during which patients with COPD reported experiencing dyspnea in a 24-item questionnaire (Eakin et al. 1998) or as listed in Holter ECG recording forms (eating, brushing teeth, and using the restroom), as detailed in our previous study (Yamane, Kimura & Morita 2024). Notably, using the restroom refers to the use of a seated toilet, assuming no behavioral differences between males and females.
Activity recognition and evaluation of the machine learning model
Activity classification using nondominant wrist and chest sensor data
The participants wore the GT9X Link device on their dominant and nondominant wrists, chest, hip, and thigh. Because accelerometer data from devices worn on the nondominant wrist or chest demonstrated high accuracy in activity recognition (Yamane, Kimura & Morita 2024; O’Brien & Min 2022; Panahandeh et al. 2012), two approaches were employed to evaluate activity classification effectiveness. The nondominant wrist classifier used data from the nondominant wrist sensor to predict activities. Conversely, the chest classifier used chest sensor data for the predictions. These methods assessed the ability of the system to recognize activities based on the data from these body parts.
IMU sensor fusion: accelerometer-, gyroscope-, and magnetometer-axis combinations
The following sensor data combinations were evaluated (Figure 1):
Nine axes (Acc_Gyr_Mag): Accelerometer data (x, y, and z axes), gyroscope data (roll, pitch, and yaw), and magnetometer data (x, y, and z axes).
Six axes (Acc_Gyr): Accelerometer data (x, y, and z axes) and gyroscope data (roll, pitch, and yaw).
Six axes (Acc_Mag): Accelerometer data (x, y, and z axes) and magnetometer data (x, y, and z axes).
Six axes (Gyr_Mag): Gyroscope data (roll, pitch, and yaw) and magnetometer data (x, y, and z axes).
Three axes (Acc): Accelerometer data (x, y, and z axes).
Three axes (Gyr): Gyroscope data (roll, pitch, and yaw).
Three axes (Mag): Magnetometer data (x, y, and z axes).
Figure 1
Axis combinations investigated in this study. The ActiGraph GT9X Link is equipped with an accelerometer, a gyroscope, and a magnetometer. The relationship between axis combination and activity classification accuracy was evaluated.
Data processing and feature extraction
The data were processed using the ActiGraph ActiLife software (version 6.13.4), as previously described (Yamane, Kimura & Morita 2024). The extracted data were segmented into 10 s non-overlapping time- and frequency-domain features (Liu, Gao & Freedson 2012), which included mean, standard deviation, variance, maximum, minimum, root mean square, signal magnitude area, correlation between axes, entropy, energy, kurtosis, skewness, median, interquartile range, and autoregressive coefficients, with 156 features utilized per window during model training and testing.
Machine learning model training and testing
Machine learning model training and testing were performed as described previously (Yamane, Kimura & Morita 2024). Leave-one-subject-out (LOSO) cross-validation was performed following the approach outlined in a previous study (Gholamiangonabadi, Kiselov & Grolinger 2020). The dataset comprised 30 participants divided into two exclusive subsets: a training subset comprising data from 29 participants, and a test subset with data from a single participant. The training data were used to train the random forest (RF) classifier model (Breiman 2001), an algorithm chosen because of its ease of implementation and frequent use in activity recognition research (Matsuyama et al. 2019; Bozkurt 2022). The RF hyperparameters included n_estimators = 2,500; criterion = “gini”; max_depth = 30; min_samples_split = 2; min_samples_leaf = 1; and random_state = 42. For the n_estimators, the value was gradually increased from the default of 100, and the value at which the recognition accuracy plateaued was selected. For max_depth, instead of using the default ‘None,’ we set it to a numerical value and gradually increased it, selecting the value at which recognition accuracy plateaued. We set random_state to 42, which is commonly used, instead of the default ‘None,’ to ensure reproducibility of results. Default values were used for all other parameters. After training, model performance was evaluated using the test data. This process was iterated 30 times by varying the training and testing subset combinations.
Machine learning model evaluation
The classifier performance was assessed using the following indices:
Precision is defined as the proportion of predicted positive cases that were correctly identified. Precision was determined using the following equation:
Precision = (True positive)/(True positive + False positive). (1)
Recall is defined as the proportion of actual positive cases that were correctly identified. Recall was determined using the following equation:
Recall = (True positive)/(True positive + False negative). (2)
F-value is defined as the harmonic mean of precision and recall. The F-value was determined using the following equation:
F-value = 2 × (Precision × Recall)/(Precision + Recall). (3)
The F-value ranges from 0 to 1 (Bull et al. 2018). Activities with F-values equal to or greater than 0.7 were considered recognizable because an area under the receiver operating characteristic curve (AUC) of 0.7 or higher was deemed highly or moderately accurate in previous studies (Fischer, Bachmann & Jaeschke 2003; Swets 1988; Carrington et al. 2023). Confusion matrices were generated for activity classification predictions.
Results
Figures 2, S1, and S2 illustrate representative samples of the unprocessed data along the x-, y-, and z-axes, respectively, each containing acceleration, angular velocity, and magnetic field strength measurements. Throughout the experiments, the activities of the participants were documented and aligned with the accelerometer readings (denoted by numbers 0–9). Variations in activity types and sensor placements led to distinct patterns of fluctuation in acceleration, angular velocity, and magnetic field strength.
Figure 2
Waveform data depicting sensor readings from two distinct body locations. The nondominant wrist data are represented by the red line; the chest data are denoted by the green line. Labels (0–9) correspond to specific physical activities: 0, lying in a supine position; 1, standing; 2, sitting; 3, eating; 4, brushing teeth; 5, using the restroom; 6, walking; 7, ascending or descending the stairs; 8, running; and 9, other movements. (a) x-axis acceleration measurements. (b) x-axis angular velocity recordings. (c) x-axis magnetic field intensity observations.
The recognition accuracy for each activity was evaluated using the acceleration data from a 9-axis accelerometer. In total, 156, 104, and 52 features were utilized per window during model training and testing for the 9-axis, 6-axis, and 3-axis configurations, respectively. A list of the features used in each configuration is presented in Table S1. Activities with F-values of 0.7 or higher are considered recognizable. Figure 3a and Table 1 present the activity recognition results for different axis combinations with the accelerometer attached to the nondominant wrist. Acc_Gyr_Mag achieved the highest F-values for each activity, except for brushing teeth. For the 6-axis configuration, Acc_Gyr and Acc_Mag achieved the highest F-values, except for brushing teeth, for which Gyr_Mag performed better. In the 3-axis configuration, Acc achieved the highest F-values for most activities except for brushing teeth and using the restroom, which were most accurately recognized by Gyr and Mag, respectively. For activities such as lying in the supine position, sitting, eating, and running, the Acc F-values were equivalent to those on the 9-axis.
Table 1
Evaluation results of the nondominant wrist classifier.
| ACC | GYR | MAG | ACC_GYR | ACC_MAG | GYR_MAG | ACC_GYR_MAG | ||
|---|---|---|---|---|---|---|---|---|
| Lying in the supine position | Precision | 0.9462 | 0.7696 | 0.5148 | 0.9510 | 0.9511 | 0.8287 | 0.9446 |
| Recall | 0.9635 | 0.6889 | 0.5942 | 0.9686 | 0.9791 | 0.8250 | 0.9737 | |
| F-value | 0.9496 | 0.6947 | 0.5402 | 0.9556 | 0.9614 | 0.8183 | 0.9552 | |
| Standing | Precision | 0.9176 | 0.7698 | 0.5038 | 0.9298 | 0.9316 | 0.8843 | 0.9350 |
| Recall | 0.9353 | 0.8410 | 0.5641 | 0.9485 | 0.9404 | 0.8331 | 0.9459 | |
| F-value | 0.9186 | 0.7890 | 0.4975 | 0.9272 | 0.9261 | 0.8192 | 0.9306 | |
| Sitting | Precision | 0.8087 | 0.6978 | 0.4160 | 0.8578 | 0.7625 | 0.8150 | 0.9171 |
| Recall | 0.8026 | 0.7425 | 0.4949 | 0.8282 | 0.7618 | 0.8115 | 0.8690 | |
| F-value | 0.7812 | 0.6984 | 0.4269 | 0.8258 | 0.7509 | 0.7894 | 0.8769 | |
| Eating | Precision | 0.8594 | 0.6457 | 0.7353 | 0.8571 | 0.8506 | 0.7169 | 0.8650 |
| Recall | 0.8829 | 0.7930 | 0.7226 | 0.9093 | 0.8904 | 0.8225 | 0.9068 | |
| F-value | 0.8576 | 0.7053 | 0.6982 | 0.8684 | 0.8543 | 0.7525 | 0.8712 | |
| Brushing teeth | Precision | 0.7800 | 0.8805 | 0.4199 | 0.8439 | 0.7945 | 0.8287 | 0.8433 |
| Recall | 0.7442 | 0.8515 | 0.3921 | 0.7876 | 0.7393 | 0.8410 | 0.7979 | |
| F-value | 0.7524 | 0.8486 | 0.3640 | 0.8013 | 0.7545 | 0.8316 | 0.8104 | |
| Using the restroom | Precision | 0.6773 | 0.5862 | 0.7384 | 0.7043 | 0.7441 | 0.7764 | 0.7904 |
| Recall | 0.6484 | 0.4075 | 0.7173 | 0.6544 | 0.7782 | 0.6681 | 0.7634 | |
| F-value | 0.6444 | 0.4646 | 0.7127 | 0.6640 | 0.7376 | 0.6921 | 0.7588 | |
| Walking | Precision | 0.8781 | 0.7993 | 0.7427 | 0.8761 | 0.9170 | 0.8436 | 0.9295 |
| Recall | 0.8310 | 0.7620 | 0.8097 | 0.8632 | 0.9033 | 0.8465 | 0.9010 | |
| F-value | 0.8241 | 0.7421 | 0.7675 | 0.8413 | 0.9017 | 0.8314 | 0.9039 | |
| Ascending or descending the stairs | Precision | 0.8634 | 0.7314 | 0.8118 | 0.8721 | 0.9105 | 0.8326 | 0.8979 |
| Recall | 0.8418 | 0.7237 | 0.8527 | 0.8333 | 0.8947 | 0.8324 | 0.8920 | |
| F-value | 0.8374 | 0.7114 | 0.8287 | 0.8413 | 0.8976 | 0.8269 | 0.8890 | |
| Running | Precision | 0.9974 | 0.9005 | 0.8461 | 0.9974 | 0.9974 | 0.9247 | 0.9974 |
| Recall | 0.9897 | 0.9032 | 0.8729 | 0.9897 | 0.9897 | 0.9034 | 0.9897 | |
| F-value | 0.9926 | 0.8835 | 0.8549 | 0.9926 | 0.9926 | 0.8968 | 0.9926 | |
| Other movements | Precision | 0.6256 | 0.5614 | 0.5704 | 0.6604 | 0.7210 | 0.6962 | 0.7316 |
| Recall | 0.6574 | 0.5713 | 0.4783 | 0.6944 | 0.7322 | 0.6539 | 0.7423 | |
| F-value | 0.6256 | 0.5547 | 0.4903 | 0.6624 | 0.7038 | 0.6586 | 0.7192 |
Figure 3
F-value comparison of different accelerometer axes combinations with respect to each activity. (a) Accelerometer attached to the nondominant wrist. (b) Accelerometer attached to the chest.
Figure 3b and Table 2 present the activity-recognition results for different axis combinations with the accelerometer attached to the chest. Acc_Gyr_Mag achieved the highest F-value for each activity. For the 6-axis configuration, Acc_Mag achieved the highest F-values, except for eating and running, for which Gyr_Mag performed better. When using the 3-axis configuration, the axis with the highest F-value depended on the type of activity recognized. Acc had the highest recognition accuracy when lying in the supine position, brushing teeth, walking, and ascending or descending stairs. In particular, Acc had considerably better recognition accuracy for activities such as lying in the supine position and walking than those of the other two axes. Gyr exhibited the highest recognition accuracy for eating and running. In particular, for eating, the recognition accuracy of Gyr was considerably higher than that of the other two axes. Mag achieved the highest recognition accuracy for standing, sitting, and using the restroom. This accuracy was considerably higher than that of the other two axes. When lying in the supine position, sitting, standing, using the restroom, ascending or descending stairs, and running, the F-values of the 6-axis and 3-axis were comparable to those of the 9-axis.
Table 2
Evaluation results of the chest classifier.
| ACC | GYR | MAG | ACC_GYR | ACC_MAG | GYR_MAG | ACC_GYR_MAG | ||
|---|---|---|---|---|---|---|---|---|
| Lying in the supine position | Precision | 0.9922 | 0.6341 | 0.8043 | 0.9922 | 0.9922 | 0.7997 | 0.9956 |
| Recall | 1.0000 | 0.6380 | 0.7923 | 1.0000 | 1.0000 | 0.8154 | 1.0000 | |
| F-value | 0.9956 | 0.6282 | 0.7700 | 0.9956 | 0.9956 | 0.7855 | 0.9976 | |
| Standing | Precision | 0.5640 | 0.5938 | 0.8219 | 0.6455 | 0.8345 | 0.7981 | 0.8528 |
| Recall | 0.5951 | 0.6212 | 0.8756 | 0.6874 | 0.8641 | 0.8459 | 0.8675 | |
| F-value | 0.5434 | 0.5647 | 0.8139 | 0.6333 | 0.8118 | 0.7905 | 0.8094 | |
| Sitting | Precision | 0.5480 | 0.5538 | 0.7834 | 0.6019 | 0.7496 | 0.7120 | 0.8119 |
| Recall | 0.5410 | 0.5660 | 0.7846 | 0.6103 | 0.7821 | 0.6949 | 0.7667 | |
| F-value | 0.5032 | 0.5070 | 0.7295 | 0.5624 | 0.7555 | 0.6691 | 0.7591 | |
| Eating | Precision | 0.6444 | 0.6970 | 0.6323 | 0.7315 | 0.8037 | 0.8095 | 0.8442 |
| Recall | 0.6884 | 0.7562 | 0.5881 | 0.7408 | 0.7516 | 0.7933 | 0.8161 | |
| F-value | 0.6513 | 0.7069 | 0.5875 | 0.7230 | 0.7558 | 0.7774 | 0.8187 | |
| Brushing teeth | Precision | 0.9271 | 0.8518 | 0.7378 | 0.9112 | 0.9282 | 0.8897 | 0.9386 |
| Recall | 0.8731 | 0.8466 | 0.7342 | 0.8825 | 0.8968 | 0.8904 | 0.9171 | |
| F-value | 0.8861 | 0.8353 | 0.6917 | 0.8820 | 0.8998 | 0.8801 | 0.9224 | |
| Using the restroom | Precision | 0.6606 | 0.5777 | 0.8377 | 0.7006 | 0.8543 | 0.8405 | 0.8567 |
| Recall | 0.5841 | 0.4117 | 0.8074 | 0.6178 | 0.8768 | 0.8068 | 0.8569 | |
| F-value | 0.5876 | 0.4627 | 0.8028 | 0.6337 | 0.8491 | 0.8096 | 0.8420 | |
| Walking | Precision | 0.9371 | 0.8527 | 0.8053 | 0.9711 | 0.9606 | 0.8876 | 0.9761 |
| Recall | 0.9454 | 0.7020 | 0.9082 | 0.9354 | 0.9547 | 0.9099 | 0.9593 | |
| F-value | 0.9358 | 0.7219 | 0.8486 | 0.9362 | 0.9541 | 0.8912 | 0.9661 | |
| Ascending or descending the stairs | Precision | 0.9518 | 0.7427 | 0.9302 | 0.9383 | 0.9642 | 0.8965 | 0.9532 |
| Recall | 0.9331 | 0.8711 | 0.9078 | 0.9601 | 0.9679 | 0.9117 | 0.9785 | |
| F-value | 0.9360 | 0.7785 | 0.9144 | 0.9422 | 0.9643 | 0.8993 | 0.9641 | |
| Running | Precision | 0.9956 | 0.9906 | 0.9263 | 0.9952 | 1.0000 | 0.9906 | 1.0000 |
| Recall | 0.9750 | 0.9947 | 0.9374 | 0.9861 | 0.9750 | 0.9974 | 0.9889 | |
| F-value | 0.9776 | 0.9923 | 0.9283 | 0.9888 | 0.9800 | 0.9937 | 0.9933 | |
| Other movements | Precision | 0.6476 | 0.6179 | 0.6795 | 0.7025 | 0.7815 | 0.7213 | 0.7917 |
| Recall | 0.6389 | 0.5684 | 0.5762 | 0.6980 | 0.7577 | 0.6548 | 0.7777 | |
| F-value | 0.6186 | 0.5741 | 0.6081 | 0.6886 | 0.7581 | 0.6725 | 0.7739 |
Figures 4 and 5 present excerpts from the confusion matrix for activity recognition using the nondominant wrist and chest classifiers. Despite misclassifications, most activities were correctly classified along the diagonal. These patterns varied based on the attachment position and combination of the axes. The confusion matrices of all axis combinations are shown in Figures S3 and S4.
Figure 4
Confusion matrices comparing anticipated and actual activities, based on data from the nondominant wrist-mounted sensor. The vertical entries represent the actual activities performed, and horizontal entries depict the activities inferred by the activity recognition classifier. (a) Gyroscope. The orange highlights denote errors in distinguishing between using the restroom, sitting, eating, and other movements. (b) Magnetometer. The purple highlights denote errors in distinguishing between brushing teeth, lying in a supine position, standing, sitting, and eating. (c) Gyroscope and magnetometer. The orange and purple highlights signify enhanced recognition precision when contrasted with (a) and (b), respectively.
Figure 5
Confusion matrices comparing anticipated and actual activities based on data from the chest-mounted sensor. (a) Accelerometer. The green highlight denotes errors in distinguishing between standing and sitting. (b) Gyroscope. The blue highlight denotes errors in distinguishing between lying in the supine position, standing, and sitting. The orange highlight denotes errors in distinguishing between using the restroom, sitting, eating, and other movements. The red highlight denotes errors in distinguishing between walking and ascending or descending the stairs. (c) Magnetometer. The light blue highlight denotes errors in distinguishing between eating, brushing teeth, and using the restroom. (d) Accelerometer and gyroscope. The blue and red highlights signify enhanced recognition precision when contrasted with (b). (e) Accelerometer and magnetometer. The green highlight signifies enhanced recognition precision when contrasted with (a). (f) Gyroscope and magnetometer. The orange and light blue highlights signify enhanced recognition precision when contrasted with (b) and (c), respectively.
Discussion
Study Overview and Key Findings
This study aimed to determine the minimum number of sensor axes required for accurate activity recognition in clinical settings, particularly for activities relevant to COPD and arrhythmia. Our key findings indicate that for monitoring activities using a wearable sensor on the nondominant wrist, 3-axis accelerometer data provide comparable accuracy to 9-axis data for recognizing activities such as lying in the supine position, standing, eating, and running. Furthermore, when monitoring activities with a sensor placed on the chest, 6-axis data combining accelerometer and magnetometer readings achieve similar accuracy to 9-axis data for recognizing lying in the supine position, standing, sitting, using the restroom, and ascending or descending stairs. These results suggest that in specific clinical contexts, simpler sensor configurations may be sufficient for accurate activity recognition, thereby reducing device complexity and power consumption.
Healthy individuals were recruited to conduct activity recognition in this pilot study. The target activities included those known to cause breathlessness in patients with COPD and those listed on the recording cards of the Holter ECG monitors. Based on these results, we will consider collecting data from these patients in the future.
Sensor Performance on the Nondominant Wrist
First, activity recognition performance was examined with the wearable device worn on the nondominant wrist. We evaluated various combinations of sensor axes to identify the most efficient configuration for reliable activity classification.
Key Role of Accelerometer Data
Except for brushing teeth, the 9-axis sensor recognized all activities with high accuracy. This was likely due to the comprehensive information provided by all nine axes. Previous studies demonstrated that using data from multiple sensors can improve activity recognition accuracy (Huang, Yan & Onnela 2022; Kavuncuoğlu, Özdemir & Uzunhisarcıklı 2024). Among the 6-axis combinations, those including acceleration data were more accurate, possibly because acceleration can capture movement intensity and contribute to posture estimation when combined with gyroscope data, critical for distinguishing between sedentary and dynamic activities. Similarly, in the 3-axis combinations, acceleration data often yielded more accurate results than those of the other axes. Therefore, acceleration data are crucial for activity recognition when the device is worn on the nondominant wrist. Using only 3-axis acceleration data, activities such as lying in the supine position, sitting, eating, and running were recognized with an accuracy comparable to that of the 9-axis sensor. This finding suggests that 3-axis acceleration data may be sufficient for accurately recognizing certain activities. In a previous study, six basic human activities (standing, sitting, jogging, walking, going downstairs, and going upstairs) were classified with a high recognition rate of over 95.06% using a 3-axis accelerometer, thereby demonstrating the effectiveness of acceleration data (Thuong et al. 2024).
Contribution of Angular Velocity: Brushing Teeth
The angular velocity was helpful in effectively recognizing the brushing teeth activity, providing key rotational data that enhanced the distinction between brushing teeth and other activities, making it particularly valuable when the device was worn on the nondominant wrist; the triaxial angular velocity provided more accurate results than the 9-axis sensor. Among the 6-axis combinations, those excluding angular velocity exhibited the lowest accuracy, indicating that angular velocity is crucial for classifying the activity of brushing teeth when the device is worn on the nondominant wrist. In a previous study, brushing teeth was recognized with high accuracy using a 6-axis stretchable electronic sensor that measured acceleration and angular velocity attached to the dominant hand (Garlant, Ammann & Slepian 2018). Comparing the Mag and Gyr_Mag confusion matrices, incorporating angular velocity reduced misclassification of brushing teeth with respect to stationary movements (such as lying in the supine position, standing, or sitting) or eating. The wrist of the non-brushing hand remained suspended while brushing the teeth. The angular velocity data likely aided in recognizing this state, particularly in helping differentiate between brushing teeth and standing, where both postures involve freely hanging wrists; however, standing typically involves less movement. Angular velocity is instrumental in differentiating between these two states based on the degree of movement.
Contribution of the Magnetometer: Using the Restroom
The magnetometer was useful for recognizing using the restroom. The triaxial magnetometer demonstrated greater accuracy in recognizing using the restroom than the other three axes. In 6-axis comparisons, combinations excluding the magnetometer exhibited the lowest accuracy, indicating that the magnetometer is crucial for classifying using the restroom when the device is worn on the nondominant wrist. Comparing the Gyr and Gyr_Mag confusion matrices, after incorporating the magnetometer, misclassification of using the restroom as eating, sitting, or other movements was reduced significantly. During the act of using the restroom, considerable changes in the direction of the wrist likely facilitated improved recognition, as the magnetometer can be used to detect the direction of a person (Lowe & Ólaighin 2014).
Sensor Performance on the Chest
Next, we evaluated recognition accuracy when the device was attached to the chest. This location provides a central view of the body movements and may offer different strengths across sensor-axis combinations.
Recognition Accuracy Across Axis Combinations
Considering the comprehensive information from all nine axes, the 9-axis sensor recognized all the activities with high accuracy. A previous study demonstrated that activity classification using acceleration and ECG data from a patch-type sensor achieved higher accuracy than activity classification using acceleration data alone (Ren et al. 2024). Among the 6-axis combinations, the acceleration and magnetometer data recognized most activities with high accuracy, except for eating and running. Although the subject was an animal, a previous study demonstrated that magnetometry, in addition to acceleration, enhanced activity recognition (Sakai et al. 2019). Activities, such as lying in the supine position, standing, sitting, using the restroom, and ascending or descending stairs, were classified with an accuracy comparable to that of the 9-axis sensor.
Accuracy varied substantially across activities when using the 3-axis configuration. Acceleration was particularly effective in recognizing lying in the supine position and walking, whereas angular velocity was most effective in recognizing eating. The magnetometer demonstrated high accuracy for standing, sitting, and using the restroom. Some activities, such as lying in the supine position, standing, and running, were recognized with comparable accuracy compared with the 9-axis configuration. When utilizing the 3-axis configuration, the axis type should be adjusted according to the specific activity being recognized. However, the overall accuracy was lower for the 3-axis configuration, making the 6- or 9-axis configuration preferable, except when targeting specific activities.
Effectiveness of Acceleration: Lying in the Supine Position and Walking
Acceleration was helpful in successfully recognizing the activities of lying in the supine position and walking. Even with the 6-axis configuration, the combinations with acceleration (Acc_Gyr and Acc_Mag) were more accurate than those without acceleration (Gyr_Mag) in recognizing the activities of lying in the supine position and walking. This demonstrates that acceleration data are particularly effective in recognizing these activities when the device is attached to the chest. Adding acceleration data eliminated the confusion between lying in the supine position and standing or sitting, as was evident in the Gyr and Acc_Gyr confusion matrices. This improvement likely reflects the ability to capture the tilt of the body, which differs by approximately 90° across postures, using acceleration. Similarly, when comparing the confusion matrices of Gyr and Acc_Gyr, introducing acceleration data reduced the confusion between walking and ascending or descending stairs, possibly because acceleration can capture bodily movements. While walking involves minimal movement in the vertical direction (z-axis), ascending or descending stairs involves considerable vertical movement. In a previous study, five types of basic activities (running, walking on level ground, walking on an incline, standing, and squatting) were classified using a 3-axial accelerometer attached to the chest (Vakacherla et al. 2023).
Effectiveness of Angular Velocity: Eating
Angular velocity was helpful in recognizing the activity of eating. In the case of 6-axis configurations, the Gyr_Mag combination demonstrated higher accuracy in recognizing the eating activity than the other 6-axis combinations. This result, combined with the high accuracy of the angular velocity among 3-axis options, suggests that angular velocity effectively recognizes eating when the device is attached to the chest. A comparison of the confusion matrices of Mag and Gyr_Mag revealed that adding angular velocity data reduced the misclassification of eating as either brushing teeth or using the restroom. This improvement may be attributed to the ability of the angular velocity to capture the rotation of the body associated with eating.
Effectiveness of Magnetometer: Standing, Sitting, and Using the Restroom
The magnetometer was helpful in effectively recognizing standing, sitting, and using the restroom. Even with 6-axis configurations, the magnetometer combinations (Acc_Mag and Gyr_Mag) demonstrated considerably higher accuracy in recognizing the activities of standing, sitting, and using the restroom than without a magnetometer (Acc_Gyr). This demonstrates that the magnetometer is particularly effective in recognizing these activities when the device is attached to the chest. A comparison of the confusion matrices of Acc and Acc_Mag revealed that adding a magnetometer to the acceleration data reduced the confusion between standing and sitting by capturing differences in chest height. Adding a magnetometer considerably reduced the misclassification of using the restroom as eating, sitting, or other movements, as observed in the Gyr and Gyr_Mag confusion matrices. When using the restroom, the frequent directional changes in the chest may have made the magnetometer particularly helpful in recognizing this activity.
Summary of Optimal Sensor Configurations
In conclusion, when attached to the nondominant wrist, the 9-axis configuration provides optimal performance. However, 6- or even 3-axis configurations can recognize specific activities with comparable accuracy. For 3- and 6-axis configurations, the inclusion of acceleration data is recommended. When attached to the chest, the 9-axis configuration provided optimal performance. Similar to wrist attachment, 6-axis or 3-axis configurations can recognize specific activities with comparable accuracy. However, because the 3-axis configuration can only recognize specific activities with high accuracy, using a 6-axis setup is recommended unless the recognition of specific activities is targeted. Combining an accelerometer and magnetometer is recommended for a 6-axis configuration on the chest.
Adequacy of Sampling Frequency
According to the Nyquist–Shannon sampling theorem, to accurately digitize a periodic signal, the sampling frequency must be at least twice the highest-frequency component of that signal (Jerri 1977). In our study, brushing teeth was the highest frequency activity. Because this involves manual brushing (not including the use of electric toothbrushes), the dominant frequency components fall within the range of a few hertz. Consequently, the sampling rate of 10 Hz exceeded the Nyquist requirement, making it adequate for activity recognition accuracy. Therefore, the 100 Hz sampling frequency used in this study appears adequate for capturing most movements of patients with COPD and those with chronic diseases.
Study Limitations, Clinical Implications, and Future Applications
This study had several limitations. First, all participants were young and healthy, which may not reflect the slower and irregular movement patterns of older adults or patients with COPD, potentially limiting the generalizability of the results. Second, although a LOSO cross-validation approach was used to improve generalizability, the robustness of the model under the conditions of noise or missing data, common in real-world settings, was not evaluated. Third, all activities were performed in controlled environments, which may not capture the variability and unpredictability of daily behaviors.
These limitations suggest the need for further research involving patient populations, real-world data collection, and more robust modeling approaches. Future studies should consider techniques such as longer or adaptive time windows, noise injection, and the use of stability metrics (e.g., cross-fold variance or feature consistency) to assess and improve model performance in clinically realistic contexts.
Despite these limitations, the findings have practical implications for the development of wearable systems tailored to clinical use. The fact that simpler sensor configurations provide comparable accuracy for specific activities suggests the feasibility of designing low-power, compact, and user-friendly devices. This may be particularly beneficial for long-term home monitoring of patients with COPD or arrhythmias, as continuous activity tracking can support symptom assessment, detect early signs of deterioration, or evaluate the impact of interventions.
Future work should evaluate whether wearable systems optimized using this axis-reduction approach can enhance clinical workflows, improve patient adherence to monitoring protocols, and contribute to better health outcomes in real-world settings. Furthermore, integrating these systems with clinical decision support tools or remote healthcare platforms can be a promising direction for expanding their impact in both outpatient and telemedicine contexts.
Conclusions
This study demonstrated that activity recognition relevant to COPD severity and arrhythmia diagnosis can be achieved with reduced IMU configurations. While the 9-axis sensors provided the highest accuracy, 3-axis acceleration for the nondominant wrist and 6-axis combined acceleration and magnetometer for the chest achieved comparable performances. These findings support the development of simpler, smaller, and more energy-efficient wearable devices that can enhance long-term patient monitoring and facilitate objective clinical assessments.
Data Accessibility Statement
The data presented in this study are available upon request from the corresponding author. The data are not publicly available due to confidentiality issues.
Additional File
The additional file for this article can be found as follows:
Acknowledgements
We thank all participants for being part of our study.
Competing Interests
The authors have no competing interests to declare.
Author contributions
All authors conducted literature search. M. M. and M. K. conceived the study design. M.K. recruited participants and collected acceleration data. T.Y. processed the acceleration data, extracted features, and performed machine learning. T.Y. wrote the manuscript. M.M. reviewed the manuscript.