In total, 59 community-dwelling elders (19 males and 40 females; mean age = 81.86 years, SD 6.95 years) participated in our study from September 2019 to December 2019. Participants with behavioral problems (eg, violence), unstable mental status (eg, paranoia), communication problems (eg, dialect), or severe hearing impairment were excluded from the study. Written informed consent was obtained from all participants before performing data collection. The imbalanced gender distribution was a direct reflection of the outnumbered male participants enrolling in community services [50]. Among the participants, six had experienced falls within the past year. A detailed breakdown of the demographics is shown in Table 1. This study was approved by the Research Ethics Committee of City University of Hong Kong (reference no. 3-2-201803-02).

The score followed the Tinetti POMA grading guideline [8]. The Tinetti POMA grading guideline is the mobility and balance assessment outcome used in this study. Age affected the Tinetti POMA performance, but gender did not. Age showed a significant difference between the sit-to-stand score=4 and <4 groups (Student t test, P<.001). Age also showed a significant difference between the stand-to-sit score=2 and <2 groups (Student t test, P<.001). There was no gender difference in all of these tasks (Fisher exact test P=.99, 0.44, and 0.50 for sit-to-stand, turning 360°, and stand-to-sit motions, respectively.)

In this step (Step 1 in Figure 1), balance assessments were conducted with an inertial sensor. Each participant was asked to perform three actions: standing up from an armless chair, turning around, and sitting down in the armless chair. There were 59 recorded inertial signals for each task from 59 participants. The three motions are commonly used in many functional assessments. All the three motions are adopted in the BBS [9], the Physical Performance Test (PPT) [50] and Tinetti POMA [8]; standing up and turning are adopted in the Fullerton Advanced Balance (FAB) [51] and the clinical Gait and Balance Scale (GABS) [52]; standing up and sitting down are adopted in the Postural Assessment Scale for Stroke Patients (PASS) [53], the Sensory Orientated Mobility Assessment Instrument (SOMAI) [54], and the Activity-based Balance Level Evaluation (ABLE) scale [55]. These three motions were considered sufficient to illustrate the idea of motion detection and evaluation in the developed framework.

A commercial inertial measurement unit (Wit-motion JY901B; including an accelerometer and a gyroscope at 3 axes with 16-bit resolution, sampling frequency 40 Hz, and a built-in Kalman filter) was attached on an elastic belt and placed on the L4 vertebra of each participant prior to each task. The inertial signal was transmitted via Bluetooth to the dedicated mobile application. The participants would then listen to the instruction and perform it accordingly. As soon as a participant initiated the task, a research assistant would press the button (green buttons and STOP buttons in Figure 2) on the app to mark the starting time and ending time, as shown in the clock on the mobile device. Figure 2 displays a screenshot of the data collection app.

Before starting the assessment, each participant was required to perform a trial. Sit-to-stand and stand-to-sit motions were conducted using a straight-backed armless chair. Participants were asked to complete the turning task in a fast yet safe manner according to their judgment. These motions were conducted in line with Tinetti POMA guidelines. Participants were asked whether they needed a break for rest, but none needed it.

In this part (Step 2 in Figure 1), we aimed to determine the starting and ending times for a known task from inertial signals containing only a known task (sit-to-stand, turning 360°, and stand-to-sit) and nontask activities. Since some participants needed extra time to comprehend and respond to the instruction, some recordings included pretask records. In addition, it also took a few seconds for users to stop the recording after the participants completed the motion. Therefore, some sensor signals irrelevant to the motion might be included in the data. Since the motion to be analyzed was known in the prescribed mobility and balance assessment list, we only needed to detect the motion. The developed framework leverages these facts to reduce the complexity of motion detection to a simpler binary classification problem. We built separate classifiers for each task motion, allowing each piece of signal that contained a known task (ie, sit-to-stand, turning 360°, or stand-to-sit) to go through the data pipeline separately, as introduced below.

After the signal was automatically annotated, sensor-based mobility and balance assessment was performed to evaluate the corresponding motion (Step 3 in Figure 1). Sensor features that have been used in several previous studies were extracted from the detected signal and evaluated in relation to its Tinetti POMA-B grading items. This part consisted of feature extraction and prediction modeling to achieve the goal.

Accuracy and the area under the curve (AUC) were used to evaluate the performance of this work. Accuracy was defined as the percentage of observations classified into the correct class, as (TP + TN)/(TP + TN + FP + FN), where TP is the true-positive class being classified as positive, TN is the true-negative class being classified as negative, FP is the false-negative class being classified as positive, and FN is the false-positive class being classified as negative. The AUC was obtained from the area under the receiver operating characteristic (ROC) curve. The ROC curve plots the sensitivity [TP/(TP + FN)]) against the false-positive rate [1 – (TN/{TN + FP})] at different levels of thresholds.

The agreement between the mobility and balance assessment results using the human-annotated label and the motion detection method was evaluated using the McNemar agreement test. The null hypothesis states that the two results show agreement in classifying the mobility and balance assessment evaluation results. The test statistic was calculated by z² = (n₁₂ – n₂₁)²/(n₁₂ + n₂₁), where n₁₂ is the prediction result of when the manual label sensor feature shows positive but the motion detection label sensor feature shows negative. In contrast, n₂₁ is the prediction result where the manual label sensor feature shows negative but the motion detection label sensor feature shows positive. The test statistic followed a chi-square distribution with 1 degree of freedom.

Leave-one-subject-out (LOSO) cross-validation (CV) was used for training and testing. Each time, all the participants’ inertial measurements except the i-th participant were used for training the motion detection and sensor-based balance assessment, and the inertial measurements from the i-th individual were used for testing. The convolutional LSTM neural network was built and trained using Keras [63] with TensorFlow [64].

The computer used for training was an Intel E5-2670 CPU with Nvidia Tesla K20 and CUDA version 10.2 on a Linux system. The batch size was set at 200, and the stopping rule was set for no improvement after 20 epochs. The rest of the computation was conducted using SciPy [57] and NumPy [65].

Our motion detection methods yielded moderate accuracy in detecting sit-to-stand, turning 360°, and stand-to-sit tasks from 85% to 88% at different levels (see Table 4). The difference in classification accuracy did not significantly vary between different mean filter sliding window sizes. The developed motion detection method detected the same motion performed by the participant not observed in the training stage, because the LOSO CV withheld the participant’s motion signal as a testing set in each train and validation cycle. The training time and testing time showed that it took more than 1500 s to train the model, but detection was conducted within 0.3 s/participant.

The sit-to-stand task motion was well detected and classified using the developed method. The detector had 85%-87% accuracy in detecting sit-to-stand motion that was not previously observed in the training set. The extracted feature, the peak count along the AP axis, exhibited a strong ability to discriminate between normal and abnormal motion in the sit-to-stand task through the k-NN method at 90% accuracy. The high accuracy could be attributed to the fact that multiple attempts to rise generate multiple peaks in the signals [58], which aligns with the Tinetti POMA grading criterion sit-to-stand transition. LDA showed the least discriminative capability because of the Gaussian distribution assumption. The result shows that the developed motion detection method and balance evaluation model can predict abnormal sit-to-stand motion.

The turning 360° task motion was detected using the developed method at an accuracy ranging from 85% to 86%. The classification showed 92% agreement with the professional opinion using k-NN in terms of the Tinetti POMA turning motion outcome. These results indicate that Tinetti grading items may correlate with the turning speed of the participants. Previous studies [66,67] have also reported that the turning speed is correlated with certain clinical mobility and balance assessment tools.

The stand-to-sit task motion was detected at 88%-89% accuracy by the developed method. Compared with the other two tasks, however, sensor-based prediction showed the least accuracy in predicting the functional assessment result, at 86% accuracy using the one-class SVM. This may be ascribed to Tinetti POMA’s grading to deduct a mark for participants who completed this task with the assistance of their arm, which occurred in the majority of the cases according to the report. This kind of information may not be revealed from the sensor signal, as the location of the sensor is at the participant’s lower back. Consequently, it could result in weaker performance in classifying the category of older adults.

The sensor features from the manual label signal and the motion detection method yielded little difference, which barely affects the sensor-based balance assessment results. Both sit-to-stand AP peak counts and stand-to-sit peak counts showed no statistical difference (Wilcoxon signed-rank test P=.42 and .45, respectively). The average turn speed showed discrepancy (P=.03) with a mean difference of 5.68°/s. In summary, the motion detection label features showed statistically no difference in the sit-to-stand and stand-to-sit motions but a slight difference in turning.

Using the manual label motion signal features or the detected motion signal features showed no statistical difference in classifying the Tinetti POMA outcomes in all circumstances (Table 5). The McNemar test was conducted to determine whether there is a difference in the classification outcome between the manual label and the presented motion detection. The results revealed that there is no statistical difference. The motion detection accuracy was satisfactory enough to ensure that the sensor features were not affected by the detection method. Consequently, practitioners could be comfortable with adopting the motion detection method instead of traditional laborious works. None of the models rejected the null hypothesis that both labels will yield the same result. This indicates that even though the motion detection method did not fully agree with the manual label, it still captures vital information in the sensor signal to predict mobility and balance assessments.

Previous research [36] has reported clinician concerns about the real-time application of sensor-based balance assessments. The testing results in motion detection indicated only a little delay in obtaining the results of the developed framework. Therefore, the presented implementation showed acceptance by the clinician in terms of time efficiency. In contrast, a survey [68] reported that most of the Hong Kong community-dwelling older adults perceive motion-analyzing systems as useful. Accordingly, we believe that the framework would receive acceptance from both clinicians and older adults.

There are four limitations of our study. First, we only analyzed sit-to-stand, turning 306°, and stand-to-sit motions rather than the complete mobility and balance assessment that covers comprehensive mobility and balance aspects. Nonetheless, motions used in this study have been frequently analyzed in several mobility and balance assessments to assess lower-body strength [69] and dynamic balance [70]. Other popular mobility and balance assessments, such as the Timed Up and Go (TUG) test [71], would be included in our future work. Second, instead of solving a HAR problem, this work leveraged the application scenario and broke it down into a detection problem, which only required an indication of the start and end times for the known task. This assumption helped the model to finesse the large variability in different motions in older adults. Third, we would like to expand the population and observations to enhance motion detection performance and balance assessments. In the framework discussed in this study, motion detection still relied on a certain amount of postprocessing, such as postprocessing after motion detection and feature engineering. This additional processing may become unnecessary with more observations for training. In addition, a greater sample size can also facilitate sensor-based balance evaluation with more complex modeling techniques and a one-off multiclass classifier to detect all the motions like in other works [28,30]. We can also observe that there are more healthy adults in the community than those with deviation from the healthy populations. A large sample size provides more in-depth insights into functional assessment with further complex feature engineering and feature selection methods. The last limitation is that the technology acceptance of this approach was not validated with the users. Though some related work with similar populations has shown good acceptance, as mentioned in the Discussion section, it is worth investigating how this approach can provide better user experience in the future plan.

We presented human motion detection in a large-scale health care assessment. Existing works [72,73] on HAR have achieved satisfactory results in detecting activities from nonmovement validation on public data sets. Nonetheless, most of the public data sets are limited in the subject count and focus on healthy subjects than normal life. The network structure of this study was also inspired by several previous works [56,74] on HAR. This research aimed to provide a sensor-based balance assessment approach to clinical decision support with less human effort. We illustrated a framework using an inertial sensor for balance assessment. The major contributions are two-fold. First, few existing studies have integrated motion detection and sensor-based balance assessment to form the data analysis pipeline, to the best of our knowledge. Second, our method requires less human effort to identify and preprocess the inertial signal and enable a more efficient sensor-based balance assessment tool for medical professionals. This research examined the applicability of a deep-learning approach to detecting motion in the health care assessment context instead of general daily living.

This research illustrated the idea that functional assessment motions can be detected through HAR models. Therefore, the sensor data collection process can be conducted without additional labor if there is a sufficiently pretrained model in the future. Without the additional labor, the cost of sensor-based functional assessment can be reduced, providing more incentive to conduct large-scale implementation in identifying potential fallers in the community.