Cluster-Analysis-Based User-Adaptive Fall Detection Using Fusion of Heart Rate Sensor and Accelerometer in a Wearable Device

This paper proposes an automatic fall detector in a wearable device that can reduce risks by detecting falls and promptly alerting caregivers. For this purpose, we propose cluster-analysis-based user-adaptive fall detection using a fusion of heart rate sensor and accelerometer. The objectives of the proposed fall detector are to have high accuracy with a low-complexity model regardless of diverse conditions. To meet the objectives, we propose the best 13-dimensional feature subset by using feature selection. In addition, we verify the performance increment of combining a heart rate sensor with an accelerometer and the effectiveness of the cluster-analysis-based anomaly detection. We also show the effectiveness of the user-adaptive method when using both heart rate and acceleration signals that were hardly covered in other papers. Finally, we prove that the performance of the proposed fall detector achieves is better than that of recent user-adaptive and user-independent approaches. This study is the first attempt to demonstrate the merits of the user-adaptive approach using a combination of heart rate and acceleration signals to detect falls. Moreover, this paper also contributes to fall detection area by providing the data we collected.


I. INTRODUCTION
Falls are the leading causes of injuries and injury-related deaths in the elderly [1]. Approximately half of the elderly population are unable to get up without help, even if they are not injured [2], [3]. In addition, lying on the floor for a long time often leads to dehydration, muscle damage, and fear of potential falls [4], [5]. Automatic fall detection systems can reduce risks by detecting falls and promptly alerting caregivers [6]- [14].
There are three types of fall detection approaches: wearable-based, ambient-based, and vision-based [15]. Although ambient-and vision-based approaches have better accuracy than that of the wearable-based approach, wearablebased approaches are advantageous in terms of cost, setup, The associate editor coordinating the review of this manuscript and approving it for publication was Jeonghwan Gwak . computational cost, and space restriction. In addition, wearable devices are essential because many elderly people want to live autonomously at home without locational restrictions [16]. Wearable-based fall detection systems typically use accelerometers [6]- [13], [17]- [19], gyroscopes [20], heart rate sensors [21], or a combination of these [22]- [25]. Recently, several smartwatches have started to provide fall detection services. However, to the best of our knowledge, the performance and detailed algorithms have not yet been reported. The growth of smartwatches and smartbands motivated us to propose a wristband-type fall detection approach. The proposed approach can be applied immediately to commercial smartwatches. This paper proposes combining heart rate sensor and accelerometer to achieve a higher accuracy than that obtained when using a single accelerometer. Wristband-type fall detectors have low accuracy and provide false alarms [15]. VOLUME 8, 2020 This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see http://creativecommons.org/licenses/by/4.0/ As alternatives, accelerometers combined with surface electromyography sensors [26], gyroscopes [22], [23], or various bio-signal sensors, including those used to record electrical activity of the heart, blood pressure, pulse oximetry, respiration rate, and body surface temperature [27]. Among these various options, we have chosen a heart rate sensor because higher accuracy levels have been achieved to an extent using a multi-dimensional combination of physiological and kinematic parameters [15]. In addition, a heart rate sensor is advantageous in terms of cost and size over other physiological sensors, and it is generally used in hospitals [28] and smartwatches.
The contributions of this study are summarized as follows. i) We propose the best feature subset of heart rate and acceleration signals for the purpose of demonstrating reliable performance and designing a low-complexity model. ii) We verify that combining a heart rate sensor with an accelerometer increases the effectiveness of detecting falls. iii) We show the effectiveness of the user-adaptive method when using both heart rate and acceleration signals that has been hardly covered in other papers. iv) We prove that the performance of the proposed is better than that of recent user-adaptive and user-independent approaches by comparing with 12 conventional approaches. To the best of our knowledge, this study is the first attempt to demonstrate the merits of a user-adaptive approach combining a heart rate sensor and an accelerometer to detect falls.
The remainder of the paper is organized as follows: In Section II, related work on fall detection approaches are introduced. Section III introduces the proposed method and evaluation progress. In Section IV, the experimental setup and design are described. In Section V, experimental results are presented and discussed. Lastly, Section VI presents the conclusions of the study.

II. RELATED WORK A. FEATURE SELECTION
One of the objectives of this study is to select the best feature subset that exhibits reliable performance to design a lowcomplexity model. A goal of feature selection is to build simpler models and to increase classification performance by choosing a best set of features. Feature selection is achieved using filters and wrappers [29]. Filters require independent objective functions to judge the separability of features from classifiers. For this reason, filters are computationally efficient because of using independent functions but has low estimation accuracy. Through repeated experiments performed using filters, our previous study proposed the best feature subset of a heart rate sensor for fall detection [24]. Because the combination of acceleration and heart rate signal features were hardly considered, the accuracy of the feature subset obtained using filters still remains to be improved [24].
As an alternative to filters, wrappers are used because wrappers directly take advantage of a classifier or a regression model to assess the quality of selected features.
Wrappers have high computational complexity because classifier model selection is concurrent with feature selection. For instance, Xiong et al. proposed an acceleration feature subset by using wrappers. Because a high accuracy of the subset is achieved using all the feature subsets [12], wrappers have a high computational cost. While many studies have independently introduced various features for acceleration and heart rate signals, their validation for feature vector construction is usually omitted, and there have been very few attempts to combine the features from both sensors. In this study, we select the best fall detection feature subset of a heart rate and an acceleration signals by using both filter and wrapper to design a low-complexity model.
Supervised learning can help in designing an elaborate fall detector by optimizing fall patterns; however, such an approach has serious drawbacks. First, it is challenging to obtain sufficient fall data from an actual user for training. Recent studies have collected factitious fall data. Second, it is difficult to learn various fall patterns because fall accidents can occur under several scenarios such as slipping, tripping or dropping in various directions.
An unsupervised anomaly detection method can be an alternative to supervised learning. An anomaly is defined as a point in a certain time step where the behavior of system is significantly different from the previous normal status [31]. Owing to their simplicity and ability to handle huge amounts of process data, unsupervised-learning-based anomaly detection methods have been widely used for fall detection [9], [24], [32], [33]. The models are trained using normal behavior data and anomaly data are used to detect falls.
One of the merits of using unsupervised anomaly detection is that the training model requires only normal behavior data that can be easily obtained and can be non-artificial data. Another merit is that all the various fall patterns need not be considered because fall data instances are not used as training samples. Therefore, the proposed unsupervised anomaly detection approach has high applicability. Among the different unsupervised anomaly detection methods, clusteranalysis-based models are widely used because of their low complexity [9], [24], [33]. Therefore, this study focused on detecting falls using cluster-analysis-based anomaly detection method. 40390 VOLUME 8, 2020

C. USER-ADAPTIVE FALL DETECTION
The last objective of the study is to achieve a highperformance fall detector, regardless of the conditions of a user. Most studies proposed fall detectors using userindependent models without considering factors such as the age of the users, their symptoms, or whether they use any assistive devices [10], [11], [12], [32]. Although these studies have an advantage of using the fall detector immediately because of using generic detector, the low accuracy at the actual user remains a challenging issue. In other words, the performance of the fall detector trained using the data of one user will degenerate when used by other users.
To resolve this issue, Medrano et al. showed that an increase in performance can be achieved by personalizing the fall detector [34]. The distribution of inertial sensor data can be influenced to a great extent by various users. To this end, Zhao et al. proposed a user-adaptive algorithm [33]. Their model can adapt to new users with good recognition performance. Lee et al. [9] and Nho et al. [24] also presented a useradaptive fall detector, but only a few users were evaluated. Instead of using user-independent approaches, we propose a user-adaptive fall detector.

A. FEATURE CANDIDATES
The 11 heart rate feature candidates collected from the literature are introduced through a comparison of heart rate signals between falls and non-falls. The features of the heart rate signals depend on the time interval between R peaks (RR interval, R peaks is the same as N peaks) from the QRS complex [35] in the time domain. The 11 candidates we consider in the study are presented in Table 1. The mean and standard deviation of the RR intervals in the segmented temporal windows are represented by MRR and SDNN, respectively [36]. RMSSD and SDSD denote the root mean square and standard deviation of successive differences between adjacent NNs. NN50 and NN20 refer to the number of pairs of successive NNs that differ by more than 50 and 20 ms, respectively [36]. TI is defined by the total RR interval occurrences divided by the maximum height in a histogram based on the measurement of RR intervals with bins of 20 ms [37]. The deceleration and acceleration capacity are characterized by the capacity to slow down or speed up the heart rate [38]. The Feature ranges between before-falls and after-falls for each feature are compared in Table 1. It shows that heart rate sensor detects the difference right before and after falling. A large standard deviation is caused by individual difference, and it is supported by Figure 1 which shows RR differences of each subject. In Figure 1, the red plot is for before-fall and the other is for after fall in the same subject column. Because MRR indicates the average value of RR interval, it was chosen as an easy-to-understand feature. MRR values in most subjects decreased due to the falls in Figure 1.
For accelerometer signals, the 11 feature candidates were extracted from the magnitude of the accelerometer signal mentioned in the literature. Because motion accelerometer   signals tend to fluctuate aperiodically, the data in the segmented temporal windows were described using data central tendency and dispersion in statistics: mean, median (Med), standard deviation (SD), mean absolute deviation (MAD), skewness (Skew), and kurtosis (Kurt). In addition, the energy distribution of data in signal processing was used for describing data: signal magnitude area (SMA), data maximum (Max), spectral entropy (SE), the summation of time-domain energy (TDE), and activity count (AC), which increases each time the magnitude of the accelerometer signal exceeds a given threshold (3 g in our case).
For both the heart rate and acceleration, feature calculation is based on a size-fixed temporal window, and 4, 5, 6, 7, and 8 s are considered as window size candidates. Because the RR interval is a basic processing period in heart rate, the window size is longer than 3 s. If the window size exceeds 8 s, the fall detection performance degrades because heart rates return to normal after a fall. In addition, note that the size of the temporal window is closely related to the response time of the entire detection system [39].

B. FEATURE SELECTION
The objectives of feature selection are to build simpler and more comprehensible models and increase the data classification performance by choosing proper feature sets. The performance of the proposed fall detector is strongly influenced by the quality of feature selection. For feature selection, we selected the initial feature combinations in filters and then optimized them using a wrapper. We used two objective functions in filters.
Fisher score (FS) is one of the linear objective functions. It selects features such that the feature vectors within the same class are similar, while vectors from different classes are dissimilar in (1). Equations (2) and (3) indicate the ratio of the within-class scatter matrix S W and the between-class scatter matrix S B , respectively, where n i (i = 1, 2, · · · , C) are the C classes and N is the total number of instances. where where Mutual information (MI), MI(V k ;w c ), is a non-linear objective function that measures the amount of information shared by feature candidates V k and their C class labels w c in (6), where H is the entropy function. By using threefold cross-validation, the initial feature combinations were selected from the filter results. It is determined by voting three groups of filter results.
Wrappers were used to select the best feature subset from among the initial feature combinations roughly chosen by the filters. In addition, the temporal window size was determined by the MI scores from 4, 5, 6, 7, or 8 s candidates.

C. CLUSTER-ANALYSIS-BASED ANOMALY DETECTION
We propose a cluster-analysis-based anomaly detection in which the clusters are generated by normal instances, and any anomaly is considered as a fall. GMMs are used as anomaly detectors. GMMs are parametric probability density functions represented as a weighted sum of Gaussian component densities [40], which represent normal behavior in this case. We decide that the Youden's index (YI) is the most important criteria so that a threshold distance to separate falls and non-fall segmented windows is determined when YI is maximized. This overcomes the limitations of accuracy-based threshold determination to be introduced in the results. The details that we chose YI is described in the following subsection. The performance is evaluated using 10-fold cross-validation that includes only fall instances for validation. Bayesian information criterion (BIC) is used to determine the optimal number of clusters. It is a criterion for selecting a model from among a finite set of models; the model with the lowest BIC is preferred [41]. Increasing the likelihood by adding parameters causes overfitting; the BIC solves this problem by introducing a penalty term for the number of parameters in the model. GMMs are generated with individually tuned parameters for the cluster-analysisbased fall detection approach.

D. EVALUATION CRITERIA
To evaluate each classification result presented in this study, accuracy (Acc) (7), sensitivity (Sen) (8), specificity (Spe) (9), and YI (10) were computed. While assigning the true positives (TPs), true negatives (TNs), false positives (FPs), and false negatives (FNs), each TP was labeled as fall if the data includes a fall event; otherwise, it was labeled as nonfalls. The accuracy was determined by counting TNs when they were greater than falls (TP), given that the dataset was imbalanced. Even if the accuracy was high, the sensitivity was low while the specificity was high. For this, YI played a significant role as a main index for fall detection.

E. PERFORMANCE EVALUATION
A 10-fold cross-validation is conducted during a performance evaluation to prevent overfitting. We gathered 8280 non-fall and 2458 fall segmented windows from all subjects. In the case of the user-independent approach for the comparison group [10], [11], [17], [25], [42], the 8280 non-fall and 2458 fall windows are divided into 10-equal sized groups. For example, one group consisting of 828 non-fall and 246 fall segmented windows is taken as a test dataset. Then the remaining groups are taken as a training dataset. A model is fitted on the training set and evaluated on the test set. This procedure is repeated 10 times, with each of the 10 groups used exactly once as the validation data. The 10 results can then be averaged to produce a single estimation.
In the case of user-adaptive approach including the proposed method and the comparison group [9], [24], [33], a model is trained individually for 21 subjects. For example, for subject number one, there are 345 non-fall and 122 fall segmented windows which are divided into 10 equal-sized groups. Each group consists of 34 non-fall and 12 fall segmented window. One group is taken as a test dataset and the remaining groups are taken as a training dataset. The procedure of the evaluation is the same as in the user-independent. The performance of the user-independent approach was evaluated using 10-fold cross-validation to all of the segmented windows. On the other hand, 10-fold cross-validation is used to each of subject in the case of the user-adaptive approach.

IV. EXPERIMENTAL SETUP A. DATA COLLECTION AND PREPROCESSING
A total of 21 participants (mean=25.8 yr, SD=3.6 yr) wore a tri-axial accelerometer (EBIMU24GV4, E2BOX) and a heart rate sensor (HRM-2511B, Kyto Electronics) on the wrist for the experiments as shown in Figure 2. Their detailed specifications are given by Table 2. The use of accelerometers is tricky for their calibration that detailed processing is given as follows. Acceleration by human motion can be given by (11), where a motion (t) is pure dynamic acceleration, a bias (t) is a bias such as thermal noise in the accelerometer, and g sin(θ) is the influence of gravity containing the initial angle φ, and the time-dependent angle variation θ(t), during the motion [39]. a total (t) = a motion (t) + gsin(φ + θ(t)) + a bias (t) (11) Of the total acceleration acquired by sensors, the component of interest is mainly pure dynamic motion for motion analysis, and fall detection relies on the detection of falling impact represented by the pure acceleration. The pure motion is obtained by removing, from entire acceleration, g sin(θ) and a bias (t), which cause a sort of drift error which is the actual change in measurement value under the same conditions at different temporal points. For g sin(θ) removal, we use gravity effect removal function of EBIMU24GV4, which detects orientation change using embedded gyroscope and cancels gravity component. For a bias (t) removal, we set up bandpass filters with 5-20hz bandwidth referring to [43]. In addition, it is why we chose the accelerometer for our research that EBIMU24GV4 minimizes thermal errors to 0.026%/ • C by using thermometers inside. Finally, the 11 feature candidates mentioned in the literature were extracted from the sum magnitude vector (SMV). SMV is defined as the intensity of triaxial acceleration value, which is shown as (12) where x[n], y[n], and z[n] are the acceleration of X-, Y-, and Z-axis, respectively, at the sampling time n. For the analysis, 13 activities of daily livings (ADLs) and 6 fall signals were collected for each subject. Data collection was carried out because there are no public datasets that include accelerometer and heart rate signals for falls and nonfalls. Basically, data on four near-falls (brushing teeth, rapidly zipping up and down, hitting the sensor, and lying down and up on the bed) and nine ADLs (wearing a cloth, washing, brushing hair, eating, writing, tying a shoelace, walking, sitting down and up, and climbing up and down on stairs) were collected. The list of falls, near-falls, and ADLs simulated in the laboratory experiment is shown in Figure 3. Because the heart rate signals might be affected by the simulated falls, an experimenter drops subjects suddenly while they lie on the edge of the bed wearing an eye patch and earplugs to avoid the influence of external stimuli. In addition, the experimenter suddenly pushed when the heart rate of the subject returned within the normal range because the subjects who know to fall may have an effect on the heart rate signal.
The overall performances are reported based on the results of 10-fold cross-validation on the collected 21 subjects data. The collected dataset can be accessed via https://github. com/nhoyh/HR_IMU_falldetection_dataset. The database samples are constructed through a sliding window that provides framing of the full-length data (falls: 2458 segmented windows, non-falls: 8280 segmented windows). For database construction, fall data are manually segmented based on the magnitude of the accelerometer signals to segment the data containing fall signals from the acquired data. This process is unnecessary for non-fall data because it is segmented during data acquisition; however, fall data is manually segmented to remove redundancy for analysis. In other words, the windows of falls are set based on the accelerometer signals from the first to the last inclusion. These temporal windows overlap the first window to the end. Details about the fall data segmentation to remove redundancy are described in Algorithm 1. The segmented windows are affected by the window size and the overlap period which are covered in Section V-A. Note that, all the fall data are only used for validation and testing and not for training. Therefore, manual segmentation of fall data is unnecessary in actual applications. In actual application, the proposed algorithm detects whether falls or non-falls within the automatic temporal window.

B. EXPERIMENTAL DESIGN
The best feature combination of heart rate and acceleration signals is selected through a two-step feature selection. The filters select several feature combinations by increasing feature dimension and comparing the scores computed within the same dimension group. The initial feature combinations are reported based on the results of 4-fold cross-validation on the overall features. Then, the wrapper determines the best subset among the initial combinations. As a result, To validate the effectiveness of the proposed algorithm, two comparisons of the proposed algorithm were carried out. One comparison was carried out to verify the effectiveness of combining the heart rate sensor and accelerometer in Section V-B. In detail, the fall detection performance using a combination of heart rate sensor and accelerometer was compared with that obtained when using a single accelerometer. In addition, we verified how potential the cluster-analysisbased anomaly detection is when using acceleration and heart rate signals (Section V-C). Moreover, we show the effectiveness of the user-adaptive method when using both heart rate and acceleration signals that were hardly covered before in other papers.
Finally, Section V-D presents a comparison of the proposed algorithm with previous studies. We prove that the proposed fall detector exhibits better performance than recent user-adaptive and user-independent approaches. We chose 12 approaches from the studies listed in Table 3 and investigated their performance on our database. More details about 12 approaches are described following paragraph. In addition, the improvement of the user-adaptive approach over the userindependent approach is proved.
We verify the performance of the proposed approach by comparing it with three other user-adaptive approaches [9], [24], [33]. Lee et al. proposed a cluster-analysis-based abnormal activity detection algorithm using accelerometer signals [9]. In this case, Gaussian mixture models (GMMs) were utilized for the clustering model. The mean and standard deviation for each of the three accelerometer axis were used. Our previous work proposed a user-adaptive fall detection algorithm combining accelerometer and heart rate sensors [24]. As a result, a 5-D feature subset (root mean square of successive differences, standard deviation of successive differences, normal to normal 50, normal to normal 20 for heart rate signal feature, and mean absolute deviation for acceleration feature) was proposed. Our previous work is the first attempt of combining a heart rate sensor and an accelerometer, but their user-adaptiveness was not verified. In addition, using wrappers is essential for selecting the best feature subset to improve performance. Zhao et al. proposed a user-adaptive fall detection method combining a gyroscope and accelerometer [33]. From each window, a vector of 13-D feature subset was obtained by calculating variables in the time domain. The mean, standard deviation, energy, mean crossing rate, maximum value, and minimum value were extracted from the magnitude of the synthesized accelerometer signal and angular velocity. In addition, the tilt angle was used as an extra feature. Their user-adaptive algorithm based on k-means clustering, local outlier factor (LOF), and multivariate Gaussian distribution outperformed the other user-adaptive methods.
In the case of the user-independent studies [10], [11], [17], [25], [42], Abbate et al. presented a smartphone-based fall detection system that includes novel techniques for recognizing ADLs that could be misdetected as falls [10]. Khojasteh et al. used sensor placed on the wrist for fall detection [11]. The algorithm proposed by Khojasteh was, in fact, an improvement over that proposed by Abbate with threshold tuning of features. Both these studies used feed forward neural networks (NN) with 8-D acceleration feature subset. Aziz et al.. compared the accuracy of two approaches (threshold-based and machine learning-based), and proved machine learning algorithms get greater performance [42]. In this case, we reproduce four approaches which were used for the machine learning algorithms: support vector machine (SVM), decision tree (DT), k-nearest neighbor (kNN), and naive bayes (NB). Saleh et al. proposed two-segment feature extraction and SVM-based fall detection algorithm [17]. The 12-D feature vector was introduced from the sliding window into two equal segments noted left and right. Then, linear and quadratic kernel SVMs were used as a fall detector. Hussain et al. extracted 24-D feature from an accelerometer and a gyroscope [25]. The highest accuracy was achieved with kNN and the second-highest accuracy with SVM for fall detection.

A. BEST HEART RATE AND ACCELERATION SIGNALS FEATURE COMBINATION
Using the filter and wrapper methods for heart rate and accelerometer signals at an overlap of 90% and window size of 6 s, a 13-D feature subset consisting of 5-D heart rate and 8-D acceleration feature was selected.
We present the feature combinations with the highest scores obtained by the filters for candidates with the same dimensions in Table 4. As the results of FS, AC for acceleration and SDSD for heart rate were chosen as initial combinations because of recording the highest score. Candidates with the highest score among each group are similar with respect to accelerometer signals or identical with respect to heart rate signals, regardless of the filter objective functions. However, the quantitative tendency of their scores is contradictory to each other depending on the objective function. As feature dimension increases, FS scores decrease and MI scores increase. Based on this conflicting result, we conclude that AC for acceleration and SDSD for heart rate chosen by FS are only valid and accepted as initial feature combinations because we interpret the result as a constraint of linear filters that estimate the data with non-linear distribution. In addition, it should be noted that the priority of MAD-based combinations in acceleration has been repeatedly supported by our previous studies [24]. Then, all the features chosen by MI were investigated again with the wrapper.
Because MI-filtered estimates show the correlation with the dimensions, the wrapper is additionally applied on the candidates chosen by MI. Table 5 lists the results obtained by the wrapper method recorded by the initial combinations of heart rate and acceleration features chosen by the filters in Table 4. This result prevents interpretive misunderstandings that higher-dimension features have higher performance than the proposed approach. If the dimension is greater than 8-D, filters were not evaluated because there were a small number of remain subsets. The remaining subsets were estimated using the wrapper. VOLUME 8, 2020 TABLE 4. Initial feature combinations of acceleration and heart rate signals using filters.

TABLE 5.
Results of the wrapper recorded by the initial subsets of heart rate and acceleration chosen by MI.
As a result, an elbow point was observed when combining SDSD, SHR, FHR, NN50, and RMSSD for heart rate and Med, MAD, SD, Skew, SE, AC, TDE, Kurt for acceleration, which are listed in the initial combination. Finally, we propose the best feature combination of the 13-D feature subset that combines 5-D heart rate features and 8-D Acceleration Features for fall detection and denote it as HAF in Table 5. Using the combination of features obtained by MI for both heart rate and acceleration results in a higher performance than that of FS (AC, SDSD recorded 74.76%). Because MI and FS are a nonlinear and a linear filter respectively, acceleration shows bigger differences according to the filter types implying more serious nonlinearity of the feature distribution than heart rate, which reaffirms previous research [44].
We investigated the influence of the temporal window and overlapping size together and found that the optimal window size and the overlap period are 6 s and 90%, respectively. This investigation was conducted on heart rate with an MI filter because the MI filter is more complicated and sensitive to extracting features based on the signal period compared to an accelerometer signal. In addition, selecting a window size based on acceleration is insufficient to include the effectiveness of the heartbeat because the small size is chosen [42], [45]. The overlapping period was selected as 90% of a window size to include as many segments as possible to prevent data loss. Figure 4 shows the MI estimates affected by window sizes. We confirm that the prominence of 6 s is constant, regardless of the overlap size. This result implies that a narrow window size decreases the number of RR intervals. Thus, the influence of the heart rate is reduced. Moreover, a wide window size cannot represent changes in heart rate features when it returns to the normal state after a fall.

B. EFFECTIVENESS OF COMBINING HEART RATE SENSOR WITH ACCELEROMETER
In this subsection, we verify fusing heart rate sensor to a single accelerometer is significantly effective to detect falls. In order to validate the effect of fusing heart rate sensor, comparison with a single accelerometer is demonstrated.
Compared to the use of a single accelerometer, the effectiveness of the proposed is analyzed in that the each state representing falling process is different between when it is described by acceleration only and by the fused one. Figure 5a explains the acceleration profile changes in every phase during falling, which consists of five steps: before falling, during the fall, during impact with ground, after falling, and returning to initial status. The difference of each state between when represented by 10-D acceleration feature vectors and when represented by the proposed features is shown by T-stochastic neighbor embedding visualization [46] in Figure 5b and 5c where each number represents instance sequences describing the state. First of all, while each state is separated clearly using the fused one, acceleration makes it intractable. It is confirmed there is occlusion in between 'during impact with ground' and 'during the fall' and between 'before falling' and 'returning to initial status' in Figure 5c. These occlusions is dissolved by using heart rate signal because it is enabled to differentiate subjects' physiological shifts by pains or sentimental shock that physical kinematics hardly carries. It is reaffirmed that this situation is not coincidental by analyzing the difference between the phase of 'before falling' and 'after falling' from multiple subjects' response in Figure 6. While two states are simplified extremely compared to the above five ones, the occlusions between the both states are found again.  According to Figure 7, combining heart rate sensor with a single accelerometer is significantly effective in detecting falls. YI is calculated for all 21 subjects. The average and standard deviation of the 21 subjects are depicted in Table 6. Moreover, the best YI were obtained for single accelerometer results in the same dimension. For single accelerometers, YI values exhibited large standard deviation for subjects, even with the user-adaptive approach. This is because the single accelerometers recorded high variances for YI for different Sen. For the 11-D acceleration feature, the Spe variance was 2.67%, but Sen was 7.57%. This phenomenon occurs when all the case of using the single accelerometer. It is difficult to show reliable fall detection performance for all users if there is standard deviation in Sen is large for users. In contrast, there is a small standard deviation for subjects in the proposed combination. Moreover, the variances of Spe and Sen were obtained as for the proposed combination 2.51% and 3.77%, respectively. Therefore, we verified that the combination of heart rate sensor with a single accelerometer provides better average performance and reliable fall detection performance regardless of the user compared with a single accelerometer.

C. CLUSTER-ANALYSIS-BASED ANOMALY DETECTION
One of the contributions of this study is to verify how potential the cluster-analysis-based anomaly detection is when using acceleration and heart rate signals. Aziz et al. [42] VOLUME 8, 2020 compared the performance of two approaches (thresholdbased and machine learning-based) in detecting falls, and they found the machine learning algorithms performed better than the threshold-based algorithms. In this case, SVM, DT, kNN, NB, and logistic regression were chosen for the machine learning algorithms. In this subsection, we compared fall detection performance using supervised learning (SVM, DT, kNN, and NB) and using GMM under the same subset of features (HAF). Table 7. shows the performances of the five methods. Firstly, in an aspect of Sen, using GMMs show better performance with small standard deviation than the other methods. In other words, the proposed cluster-analysis-based anomaly detection with GMMs fairly detects falls in average under individual subjects' diverse conditions. On the other hand, the others achieve better Spe than GMMs, which means less false alarms. However the relatively big gaps between Sen and Spe in supervised algorithms may be interpreted as affected by data imbalance, and they seem to be trained to maximize Acc by making more FNs in the condition that positive samples are less than negative ones. As we mentioned at Section III-D, the imbalanced data distribution causes the gap between spe and sen to be mistaken for high acc. For this, YI played a significant role as a main index, and it is clearly shown the cluster-analysis-based anomaly detection is more promising with its highest YI and lowest standard deviation.

D. USER-ADAPTIVE FALL DETECTION
Finally, we show the proposed fall detector records better performance than recent user-adaptive and user-independent studies. Our approach is compared with aforementioned three user-adaptive approaches and six user-independent approaches listed in Table 3. In Table 8, we show the proposed approach is effective to detect falls in terms of YI (91.34%). In addition, we can expect reliable performance for every user because the proposed approach achieves a low standard deviation in every evaluation criteria. This indicates that it is rare to record particularly low performance when the approach applied to a new user.
Another contribution is to show the effectiveness of the user-adaptive method when using both heart rate and acceleration signals that were hardly covered in other papers. A detailed comparison of a user-adaptive detector and userindependent detector in terms of YI is shown in Figure 8; the difference is statistically significant. The performance is improved from 6.60% to 26.71% without exception when the user-adaptive approach is used. Few studies have reported personalized fall detectors using only a single accelerometer; however, these detectors are not effective in some cases [34]. Even though we used different dataset, the effectiveness of combining a heart rate sensor with an accelerometer seems to eliminate the exceptions. The global average improvement for YI was obtained as 12.77% for all subjects. These results lead show that the user-adaptive detector exhibits improved performance compared with a user-independent detector when a heart rate sensor combined with an accelerometer is used. Since we verified the proposed approach has a promising improvement over the user-independent fall detector, reliable performance can be expected when the proposed adaptive approach is used than the independent approach, regardless of diverse conditions (e.g., ages, symptoms, or the use of assistive devices).

E. DISCUSSION
We had to collect activity data on the subjects and evaluate the proposed model and the other models. In this process, we thoroughly implemented the experimental procedures of the other approaches. The differences between recorded performance in their papers and the implemented method's performance were caused by a few factors.
First factor is the distinct non-fall scenarios. In the case of [9], [24], [33], there is no near-fall case such as lying on the bed or hitting the sensor. Overall performance can be degraded because we conducted various non-fall scenarios including near-falls. This difference is also caused by nearfall data acquisition. When the non-fall scenarios are simple, a model to distinguish between falls and non-falls becomes a simple problem. Therefore, the subjects were instructed by the experimenters to be more active in the near-fall scenarios. For example, in the case of brushing teeth, the subjects were requested to brush teeth rapidly and violently as possible to gather data that could be indistinguishable from falls. One of the reasons that the user-independent approaches recorded lower performance than user-adaptive approaches was caused by indistinguishable non-fall data. In that respect, one of the best ways to get a high-performance fall detector is by implementing a user-adaptive detector.
To verify the differences were caused by the inclusion of the near-fall data, comparison without near-fall activities was carried out in Table 9. When we reproduced two approaches [10], [11] with near-fall in Table 8, the loss functions of the models did not converge to zero even though we use very low learning rate and high iteration. However, in the case of no near-fall activities, the loss functions did converge with 0.001 of learning rate. Not only this case, overall performances were improved.
Lastly, the sensor attachment location creates discrepancies between their record and the implemented record. It was reported that the fall detection sensitivity varies by location such as 95% for waist, and 36-91% for the wrist [47]. Because the approaches of [9] and [33] were waist-mounted detectors and [10] was a smartphone-based detector, it can be expected that the performance degraded due to the change of sensor attachment location to the wrist.
Because of the aforementioned factors, it is necessary to validate the algorithm in various experimental settings. We welcome further researchers to make use of our published database in their performance evaluations.

VI. CONCLUSION
This study proposed a cluster-analysis-based user-adaptive fall detection approach that combines a heart rate sensor and an accelerometer to reduce the risk of falls. While many studies have omitted the feature vector construction of heart rate sensor and acceleration, we proposed the best 13-D feature subset (SDSD, SHR, FHR, NN50, RMSSD, Med, MAD, SD, Skew, SE, AC, TDE, Kurt) using both filter and wrapper to design a low-complexity model. Then, we verified that the combination of heart rate sensor and a single accelerometer is significantly effective in detecting falls. One of the merits is that the proposed fall detector has high applicability in an actual environment. Because we propose unsupervised anomaly detection, the training model requires only normal behavior data that can be easily obtained and can be non-artificial data. In this process, we verified how potential the cluster-analysis-based anomaly detection is when using acceleration and heart rate signals. In addition, reliable performance can be expected when applying the proposed adaptive approach, regardless of diverse conditions (e.g., ages, symptoms, usage of assistive devices) because we verified that the proposed user-adaptive approach demonstrates promising improvement over the user-independent approach (an improvement of 12.77%). The last contribution is to achieve better performance than those of recent useradaptive and user-independent methods. The effectiveness of the proposed user-adaptive approach is verified through a comparison with 12 conventional approaches. The results reveal that the proposed approach shows an improvement of at least 2.83% over other recent fall detectors.
Because we implemented the comparison group approaches by using acquired data, there are differences between recorded performance in their paper and the implemented performance. First reason is the distinct non-fall scenarios including near-fall cases. In addition, near-fall data was acquired which was as close as possible to being VOLUME 8, 2020 indistinguishable from fall data. Through additional experiment, we verify the differences were caused by the inclusion of near-fall data. The last reason for the differences is the sensor attachment location. We hope that the collected data will not only be useful for fall detection studies but will also be used to evaluate the performance of other researchers.