Clinical frailty syndrome assessment using inertial sensors embedded in smartphones

A Galán-Mercant; A I Cuesta-Vargas

doi:10.1088/0967-3334/36/9/1929

1. Introduction

Clinical frailty syndrome is a common geriatric syndrome characterized by decreases in physiological reserves and increased vulnerability which may, in the event of unexpected intercurrent processes, result in falls, hospitalization, institutionalization or even death (Fried et al 2001). Specifically, it has been shown that falls are a significant source of morbidity and mortality in the syndrome in frail elderly people (Tinetti et al 1988).

The timed up and go (TUG) test is a clinical tool frequently used to assess mobility and the risk of falls (Berg et al 1992, Whitney et al 2004, Salarian et al 2009, Weiss et al 2011). The TUG test was designed to predict the risk of falls in the elderly (Whitney et al 2004) and identify balance deficits (Mathias et al 1986). The clinical potential of the TUG test lies in the sequencing of several basic functional abilities, such as standing and sitting transitions, transitions requiring balance (e.g. turning) and walking in a straight line (Benecke et al 1987, Rogers et al 1998). These are common day-to-day activities and are often associated with falls (Tinetti et al 1988, Aberg et al 2010).

Having reviewed the literature, there is no consensus that the TUG test time is optimal in discriminating the risk of falls among population groups (Mathias et al 1986, Trueblood et al 2001, Buatois et al 2006, Thrane et al 2007, O'Sullivan et al 2009, Shahar et al 2009), or on the use of the TUG test over other methods to classify patients as frail or non-frail (Syddall et al 2003, Savva et al 2013). The TUG test, despite being clinically widespread, has a number of limitations, the main ones being that: (1) it focuses only on the time variable and does not take into account other variables related to deficits in kinematics and kinetics which may affect balance or the risk of a fall; (2) it measures the total time to perform the test, without taking into account partial times in the different functional tasks which make up the TUG (Salarian et al 2009, 2010, Zampieri et al 2010, 2011).

Recently, several studies have examined the functional tasks which make up the TUG test and the instruments used, together with improvements to the instruments, specifically attaching inertial sensors to the body (Najafi et al 2002, Moe-Nilssen and Helbostad 2004, 2005, Bidargaddi et al 2007, Ganea et al 2007, 2011, Botolfsen et al 2008, Higashi et al 2008, Janssen et al 2008, Marschollek et al 2009, Salarian et al 2009, 2010, Weiss et al 2010, 2011, Zampieri et al 2010, 2011, Martínez-Ramírez et al 2011, Galán-Mercant and Cuesta-Vargas 2013a, 2013b, 2014, Greene et al 2014). A previous study concluded that inertial sensors can offer an accurate and reliable method for studying human motion, but the degree of accuracy and reliability is site- and task-specific (Cuesta-Vargas et al 2010).

Nowadays, the latest generation smartphones often include inertial sensors with subunits such as accelerometers and gyroscopes which can detect acceleration and inclination (Shaw et al 2011). The specific applications for these elements are already being developed for use in a wide range of situations. Such uses include assessing and quantifying kinematic variables related to functional tasks (Galán-Mercant and Cuesta-Vargas 2013a, 2013b, 2014), identifying trembling in people suffering from Parkinson's disease (Bidargaddi et al 2007) and measuring Cobb angles in x-rays (Shaw et al 2011). Another application is the development of an objective method to classify levels of physical activity and an indicator of the degree of functional capacity and quality of life (Xia et al 2011) and gait characteristic analysis and identification based on the smartphone's accelerometer and gyrometer (Sun et al 2014).

The aim of this study was to determine the series of kinematic variables derived from accelerometry which show the greatest level of precision in discriminating between two groups of elderly persons (frail and non-frail) in the 10 m expanded timed up and go test, using the inertial sensors embedded in a smartphone—the iPhone 4^®—compared to the variable traditionally used, namely time.

2. Methods

2.1. Design and participants

A cross-sectional study was conducted involving 30 participants over 65 years old, 14 of whom were frail participants (mean Fried frailty score 3.95 out of 5) and 16 of whom were non-frail participants (mean Fried frailty score 0.05 out of 5). Non-frail participants were recruited through advertisements in the Sport and Health Centre in Torremolinos, Spain. Frail participants were recruited through advertisements in the Geriatric Centres in Torremolinos and Benalmadena, Spain. Written informed consent was obtained from each individual. The study was approved by the ethics committee of the Faculty of Medicine at the University of Malaga, Spain. The participants were assessed for frailty syndrome using the Fried criteria for classification (unintentional weight loss, self-reported exhaustion, weakness, slow walking speed and low physical activity). The participants who met at least three of the criteria were defined as 'frail' and the participants who did not meet three of the criteria were defined as 'non-frail' (Fried et al 2001). Combining the frail and pre-frail classes into a single class (based on the relatively small sample) reduced the frailty classification to a binary problem. As can be seen from the mean scores, there was a marked difference between the groups, the frail participants scoring highly for frailty and the non-frail group being very strong. The exclusion criteria were a history of acute pain in the previous 12 months, previous surgery, the presence of a tumour and ongoing treatment for musculoskeletal disorders in the upper or lower extremities. Patients with impaired cognition, musculoskeletal back co-morbidities and problems associated with exercise intolerance were also excluded. All participants were clinically examined by a physiotherapist and all examined participants were included in the study.

2.2. Instruments and procedure

The participants carried the iPhone 4 smartphone in a small neoprene sleeve fixed with non-elastic tape around the trunk, positioned on the middle third of the sternum (see figure 1). Previous studies on the variability of inertial sensor readings in different body segments (Zijlstra et al 2008, Dijkstra et al 2010) have shown that the data obtained from sensors located at sternum level provide reliable data compared to other more frequently used locations, such as close to the centre of gravity. Dijkstra et al (2010) and Zijlstra et al (2008) recommend positioning the sensor on the sternum due to overestimation of the acceleration values obtained from the participant with positioning elsewhere.

Figure 1. Refer to the following caption and surrounding text. — **Figure 1.** The accelerometer measures velocity along the x, y and z axes. The gyroscope measures rotation around the x (pitch), y (roll), and z axes (yaw).
Download figure:
Standard image High-resolution image

Apple^® uses a tri-axial gyroscope, an accelerometer and a magnetometer in the iPhone 4. The application used to obtain kinematic data was the xSensor^® Pro (Crossbow Technology, Inc., available from the Apple AppStore). In this study, the application had a capacity of 20GB (iPhone 4 storage ranges from 16 GB to 64 GB depending on the model). The data sampling rate was set to 32 Hz. The iPhone 4 is required to obtain data from the accelerometer, gyroscope and magnetometer simultaneously as earlier versions do not support integrated data collection. A previous study (Galán-Mercant et al 2014) has shown that the iPhone 4 accelerometer is accurate and precise compared to the gold standard, with an intra-class correlation coefficient of between 0.819 and 0.987.

2.3. Expanded timed up and go (ETUG) test

All participants performed the ETUG test three times. The time taken to complete the task was recorded by the clinician using a stopwatch and the 'best' trial was selected for inclusion in the study, based on the time taken, i.e. the 'best' trial was the one that took the least amount of time. Participants had a 5 min rest between trials during which the iPhone was not removed. Although in the traditional ETUG test an armchair is used (Podsiadlo and Richardson 1991), previous studies have explored using armless chairs (Benecke et al 1987, Wall et al 2000, Janssen et al 2008) and it has been argued that using armless chairs could reduce the variability between participants by eliminating the choice of whether or not to use the armrests to stand (Zampieri et al 2010). We therefore followed others in using an armless chair and participants were instructed not to use their arms to stand up.

The ETUG test uses a 10 m walkway to include gait cycles in the assessment (Wall et al 2000). The beginning and the end of the walkway were marked with 2.5 cm green tape on the floor and were shown to the participants before the trials. Participants were instructed to sit straight with their back touching the back of the chair. After they were given the 'go' signal by the clinician, they rose from the chair, walked at their fastest walking speed (but not running) to the end of the walkway, turned around to the right or left after passing the green tape, returned to the chair, turned around and sat down. The time taken to complete the task was recorded by the clinician using a stopwatch.

2.4. Phases of the expanded timed up and go (ETUG) test

Off-line processing was used to identify and analyse the different phases of the ETUG test, of which there were five (see figure 2): sit-to-stand (Si–St), gait go (GG), turning (T), gait come (GC) and turn-to-stand-to-sit (T–St–Si). Each phase of the ETUG test was detected based on previous studies using signal data on acceleration and yaw degrees from the iPhone 4 accelerometer; Si–St and T–St–Si transitions were detected and analysed based on a method published by Weiss et al (2010) and T transitions were detected and analysed based on a method published by Salarian et al (2009).

Figure 2. Refer to the following caption and surrounding text. — **Figure 2.** Illustration of the analysis of different intervals based on the ETUG acceleration and yaw degrees signal in the anterior–posterior direction.
Download figure:
Standard image High-resolution image

2.5. Data processing

Off-line analysis was developed to process the inertial sensor data. This analysis was designed to obtain kinematic data systematically for further statistical analysis and was performed using the basic R^® software package. The analysis was conducted to obtain kinematic information from the accelerometer for each participant in each of the five phases of the ETUG test. From accelerometer we obtained peak positive and peak negative acceleration values, means and standard deviations (SD) of accelerations for the three movement axes x, y and z (see figure 1).

2.6. Statistical analysis

The analysis was performed using the SPSS software (version 19.0 for Windows, IBM^®). Student's t-tests (two-tailed) were used to compare the frail and non-frail groups. Receiver operating characteristic (ROC) curves were drawn to evaluate the level of precision in the discriminatory capability of the kinematic variables compared to the variable traditionally used, namely time. Based on a previous classification (Swets 1988), the levels of precision in discriminatory capability were classified in the area under the curve (AUC) and were as follows: low precision (0.5–0.7), moderate to high precision (0.7–0.9) and high precision (0.9).

3. Results

Table 1 shows the characteristics of the sample and timed values using the stopwatch in the ETUG test. Figures 3 and 4 represent graphical displays of typical recordings of acceleration in the three orthogonal axes for subjects performing the ETUG procedure. Figure 3 represents a typical normal volunteer and figure 4 represents a similar recording for a frail subject. Note that compared with the normal subject in the case of the frail subject the amplitudes of accelerations are reduced, the total number of steps taken is raised and the total time taken for the ETUG investigation is increased.

Table 1. Descriptive statistics for the clinical data of study groups and the timing score from the ETUG test (n = 30).

	Frail (n = 14)	Non-frail (n = 16)	t-test
	Mean ± SD	Mean ± SD	t-test
Gender	10 female, 4 male	10 female, 6 male	–
Age (years)	83.71 ± 6.37	70.25 ± 3.32	p < 0.001
Weight (kg)	56.21 ± 9.64	71.03 ± 13.11	p < 0.005
Height (cm)	155.79 ± 7.81	159.44 ± 10.61	p < 0.298
BMI (kg m⁻²)	23.36 ± 3.48	27.87 ± 3.79	p < 0.005

Figure 3. Refer to the following caption and surrounding text. — **Figure 3.** A graphical representation of the three axis acceleration recorded during a typical performance of the ETUG procedure for a non-frail subject.
Download figure:
Standard image High-resolution image

Figure 4. Refer to the following caption and surrounding text. — **Figure 4.** A graphical representation of the three axis acceleration recorded during a typical performance of the ETUG procedure for a frail subject.
Download figure:
Standard image High-resolution image

Table 2 shows the results of the AUC analysis in each of the five phases and as a whole on the ETUG test. Variables related to the accelerometer which obtained an AUC value of less than 0.9 (high precision) are not shown in the results. Values related to the time have been derived from the analysis of the kinematic data (time in each phase).

Table 2. Results of AUC and gyroscope and acceleration-based values in the ETUG test and the five phases.

			Frail (n = 14)		Non-frail (n = 16)
	AUC	p-value	Mean	SD	Mean	SD
ETUG test
ETUG time (s)	0.888	0.025	53.64	24.12	15.52	2.91
Min^a y axis acc^b (m s⁻²)	0.955	0.016	− 2.054	0.741	− 6.027	4.224
Min z axis acc (m s⁻²)	0.908	0.023	− 2.435	1.553	− 6.776	2.434
Si–St phase
ETUG time (s)	0.942	0.057	4.989	1.743	2.883	2.251
Min y axis acc (m s⁻²)	1.000	< 0.001	− 1.471	0.788	− 6.182	2.415
GG phase
ETUG time (s)	1.000	< 0.001	20.148	10.36	4.511	1.243
Min x axis acc (m s⁻²)	0.960	0.040	− 2.312	0.641	− 6.293	2.877
Min t axis acc (m s⁻²)	0.942	0.057	− 2.691	0.783	− 10.680	7.113
Min z axis acc (m s⁻²)	1.000	< 0.001	− 2.396	1.613	− 8.104	1.836
Mean z axis acc (m s⁻²)	0.920	0.057	− 0.587	1.493	− 3.377	1.387
T phase
ETUG time (s)	0.938	0.061	5.329	1.344	2.815	2.069
Min x axis acc (m s⁻²)	0.938	0.050	− 2.053	0.962	− 5.779	2.432
Min y axis acc (m s⁻²)	0.978	0.024	− 2.004	− 9.448	0.945	6.937
Min z axis acc (m s⁻²)	0.987	0.015	− 1.815	1.619	− 7.204	2.438
GC phase
ETUG time (s)	0.964	0.037	17.103	9.238	4.462	0.998
Min x axis acc (m s⁻²)	0.942	0.050	− 2.460	0.937	− 6.029	2.548
Min y axis acc (m s⁻²)	0.942	0.047	− 2.824	1.002	− 11.387	8.656
Min z axis acc (m s⁻²)	1.000	< 0.001	− 2.287	1.540	− 8.360	1.720
Mean z axis acc (m s⁻²)	0.924	0.050	− 0.584	1.503	− 3.447	1.320
T–St–Si phase
ETUG time (s)	0.991	0.012	5.769	2,727	1.352	0.717
Min y axis acc (m s⁻²)	0.938	0.045	− 2.950	2.441	− 9.003	4.324

^aMinimum (peak negative). ^bAcceleration.

4. Discussion

This study has systematically evaluated and analysed the discriminating properties of the values of a series of readings based on the accelerometer function of the inertial sensor found in the iPhone 4 during the ETUG test carried out amongst frail and non-frail persons aged over 65 years. The search for discriminating variables focused on the five phases into which the test is divided (Si–St, GG, T, GC, T–St–Si) together with the test as a whole.

In phases Si–St and T–St–Si, the results indicate that the variables for acceleration show greater sensitivity in terms of discriminating between the population groups in the study. Specifically, these were the peak negative acceleration variables for motion axes x, z and y. In phases GG and GC, accelerations on the x axis showed a greater level of discrimination between the frail and non-frail groups. For the T phase, the peak negative acceleration variables which obtained area values below the curve greater than those for time were the accelerations on the z and y axes.

This study is, to the best of our knowledge, the first to have used readings taken directly from inertial sensors embedded in a smartphone, in this case the iPhone 4, in each of the five phases of the ETUG test. A series of variables derived from the acceleration readings have been shown to be sensitive in discriminating between the two groups in the study, whereas the time measured with a conventional stopwatch, a discriminating variable in the non-instrumented test, showed equal or lower values for discriminatory capability. This latter finding is supported by previous studies which have also shown that the time taken by a participant to carry out the ETUG test is not always a sensitive variable in discerning, for example, the risk of falls among the elderly (Boulgarides et al 2003, Thrane et al 2007, Shahar et al 2009, Salarian et al 2010, Greene et al 2014).

A review of the existing literature reveals a previous study by Savva et al (2013), which concluded that TUG (the 3 m version) can be used as a sensitive and specific proxy for frailty and a specific proxy for pre-frailty that can be employed when the application of the Fried criteria is not practicable. The main limitation is the lack of gold standard for frailty. Our study calculates the ability of the expanded version of TUG to identify frail members of the population using the phenotypic definition of frailty as the gold standard and assumes that this was measured without error. Other recent studies (Ganea et al 2007, Martínez-Ramírez et al 2011, Greene et al 2014), as is the case in our study, have suggested that the readings derived from accelerometry are more sensitive in discriminating the characteristics associated with frailty. In particular, Greene et al (2014) investigated a fast method for the automatic, quantitative assessment of the frailty state of a patient based on a simple protocol employing a body-worn inertial sensor. Their study found that a model classifying participants according to two frailty categories (frail and non-frail) yielded a mean cross-validated classification accuracy of 72.88% and a TUG time of 72.09%, thus showing that the acceleration variables had a higher level of accuracy in discrimination than the TUG time variable. The aim of Martínez-Ramírez et al's (2011) study was to examine orientation and acceleration signals from a tri-axial inertial magnetic sensor during quiet standing balance tests using the wavelet transform in three populations: frail, pre-frail and healthy. Unlike our study, their research was focused on a quiet standing test. Ganea et al's (2007) study aimed to propose a methodology allowing detailed characterization of sit-to-stand/stand-to-sit postural transitions in the elderly through systematic evaluation of accelerometry values, whereas we examined a wider range of functions. None of the above studies used smartphone inertial sensor technology to capture kinematic variables.

In terms of existing literature on accelerometry studies in populations of fallers, a research group has recently worked on finding discriminating variables in the instrumented TUG test using accelerometry (Weiss et al 2011). This study systematically evaluated accelerometry values in the elderly with a risk of falls on the traditional 3 m TUG test, focusing only on Si–St and T–St–Si transitions. Weiss et al (2011) used multivariate analysis to assess the discriminatory capability of accelerometry, employing different regression models to find the best combination of multiple variables related to acceleration in order to differentiate between groups. As in our study, they found a series of acceleration variables that appeared to be highly discriminating between groups with and without the risk of falls. However, we used ROC curves to show the level of precision of discrimination of the accelerometer variables one by one.

Other recent studies, like ours, have instrumented the TUG test, analysing kinematic data in search of variables with a high level of discrimination between elderly population groups with diverse functional capabilities (Bidargaddi et al 2007, Zampieri et al 2010, 2011). However, these studies focused on elderly persons with Parkinson's disease and carried out the test over a distance of 7 m. Also, unlike this study, neither group used smartphone technology to capture kinematic variables. In another recent study, Cuesta-Vargas et al (2010) quantified accelerometry during the TUG test in persons suffering from Parkinson's disease and found a subseries of acceleration variables which seemed to be particularly sensitive to inter-group differences. However, they did not use the expanded TUG test, which presents more gait cycles for analysis during the GG and GC phases (Wall et al 2000).

Analysing the data obtained here, it is noteworthy that from the values derived from the accelerometer readings, several variables were found to discriminate between the participating groups in the study; specifically, these were the peak negative acceleration variables for motion axes x, z and y. In terms of sensitivity, the values were equal to or greater than those for the variable traditionally used to discriminate in the ETUG test, namely time. Examining the discriminatory capability of these three variables compared to time, the peak negative acceleration variables for axes z and y in particular displayed greater sensitivity than the time variable when analysing the five phases of the test as a whole (see table 2). Looking at the AUCs of these variables compared to time, we find 0.955 for peak negative acceleration on the y axis and 0.908 for peak negative acceleration on the z axis. The AUC for time was 0.888. During the Si–St phase, the AUCs obtained for accelerometry showed high levels of discrimination for peak negative acceleration on the y axis (see table 2). The traditional time variable obtained an AUC of 0.942.

In the GG and GC phases, our results are supported by the findings of other authors (Najafi et al 2002). The results indicate that the accelerations obtained for the z axis (medial–lateral vector) present greater variability than for the other two axes, x (anterior–posterior vector) and y (vertical vector). This suggests that variability on the z axis could represent a differentiating aspect compared to the other two axes, x and y. Further research is necessary to determine sensitivity in the discrimination between these three motion axes. The results of this study show that the AUC for the z axis (medial–lateral vector) is greater than for axes y (vertical vector) and x (anterior–posterior vector) in both phases (GG and GC). The time variable obtained the best AUC in the GG phase (1.000) and 0.964 in the GC phase, similar to that obtained for the z axis (medial–lateral vector) (table 2).

Finally, with regard to the T phase, the peak negative acceleration variables which obtained AUC values greater than those for time were the accelerations on the z and y axes, 0.987 and 0.978 respectively (see table 2). The AUC value for acceleration on the x axis was equal to that obtained for time (AUC = 0.938), although acceleration variability was lower (see table 2). Having reviewed the literature, this study is, to the best of our knowledge, the first to show that there are variables with greater sensitivity in differentiating frail elderly people in the turning phase. Two previous studies have examined this phase through accelerometry and angular velocity in the elderly with Parkinson's disease (Salarian et al 2009, Zampieri et al 2011), concluding, unlike our study, that the variable corresponding to the duration of the phase showed greater values than those for acceleration and angular velocity.

5. Conclusions, limitations and directions for future research

Our study has some limitations. First, men and women have different characteristics and it would be very interesting to undertake a between-group investigation for gender in the future. Furthermore, although we employed ROC curves for the accelerometer variables individually, we did not do so for the resulting vector from three axes. Another significant limitation is the small number of participants, in particular those categorized as frail (N = 14). Given a larger cohort, it might be possible to create a robust multi-class statistical model that could reliably classify participants into each of the three frailty classes (frail, pre-frail and non-frail), whereas our participants were either categorized as frail or non-frail, the latter group being very healthy. Moreover, prospective studies are needed to determine if acceleration-derived measures, perhaps in combination with other metrics and previously described measures for the identification of frailty (Syddall et al 2003), would have predictive value. This study was specifically focused only on the timing of ETUG and a subset of the properties of the signal; analysis of the complete waveform and other time points may provide additional utility. Indeed, it appears that the proposed approach may not only offer a more refined scale for assessing older adults, but it may also help to pinpoint specific problems that give rise to abnormal performance on functional tasks.

Notwithstanding the above, the results obtained allow us to conclude that kinematic parameters obtained from inertial sensors embedded in smartphone technology are promising additions to the TUG analysis. There are encouraging signs that analyses of these parameters in the separate phases of the TUG procedure could offer the potential prospect of improved discrimination between frail and non-frail individuals, although further studies are needed to provide further detail and verify this. However, analysis of the five different phases of the ETUG test shows that the level of discrimination of the acceleration variables are at least as good as the headline ETUG time for completion and may well be better and that more work is required to establish the relative merits of the various parameters. Future studies will be required to research and analyse the level of discrimination in relation to the more traditional gait parameters, e.g. step time, contact time and cadence. Moreover, research is needed to examine the performance of the variables with high levels of discrimination in greater depth and also to investigate the potential and scope of the instrumented ETUG test with a view to determining which applications might detect changes in the progression of frailty amongst the elderly and their sensitivity to treatment, as well as the performance of these variables in relation to other (groups of) pathologies and/or population groups. Such contributions could be used to develop classification algorithms with the help of software designed for smartphones, thus allowing their application in clinical practice.

Author e-mails

Author affiliations

Dates