Instrumented assessment of test battery for physical capability using an accelerometer: a feasibility study

Older healthy participants (OHP, 50–70 years) were recruited in the North East of England as part of a pilot intervention study⁷ within LiveWell which evaluated internet-based lifestyle interventions in people in the retirement transition (approximately 2 years before/after retirement). Ethical consent for the project was granted by the Newcastle University Faculty of Medical Sciences ethics committee (00745/2014) and all participants gave informed written consent. Participant recruitment was arranged at baseline through large employers on Teesside and on Tyneside. Testing took place at Newcastle University facilities or at a leisure centre in Redcar.

Each participant wore a low cost (<£100) tri-axial accelerometer-based device⁸ (figure 1, dimensions: 2.3  ×  3.3  ×  0.8 cm, weight 9 g) on the fifth lumbar vertebrae (L5). This location was chosen to minimise device attachment during instrumented testing while also optimising algorithm usage, i.e. numerous algorithms developed for use on L5. The device was held in place by double sided tape and Hypafix⁹. The device recorded at a sampling frequency of 100 Hz (16 bit resolution) and at a range of ±8 g. A trained researcher used a stop watch and measurement tape (as appropriate) to record outcomes for each physical capability task.

Zoom In Zoom Out Reset image size

The HAP assessment consisted of the following:

• (a)  
  Postural control, standing balance: five tests were performed each lasting 50 s without shoes, arms folded across their chest while focusing on a wall-mounted fixed point (target) at a horizontal distance of 1 m. Variations included: (i) flat surface, feet together, eyes open (FLFTEO), (ii) flat surface, feet together, eyes closed (FLFTEC), (iii) foam surface¹⁰ (50.0  ×  41.0  ×  6.0 cm), feet together, eyes open (FOFTEO), (iv) foam surface, feet together, eyes closed (FOFTEC) and (v) flat surface, tandem stance, eyes open (FLTMEO). Due to the nature of the task and derived outcomes (section 2.5) this was quantified by the BWM only.
• (b)  
  Locomotion, 4 m walk gait speed (×2): after a practice walk, participants walked at their usual speed between two markers. Manual and BWM timing began upon the first footfall, i.e. participant's first step over the starting point. Recording ended after the participant completed the walk (manual) or last 'purposeful' footfall (BWM), i.e. vertical acceleration (a_V) exceeded a predetermined threshold. The threshold excluded any non-purposeful steps after the 4 m marker where the participant had slowed. Time to complete the 4 m walk was converted into a meters-per-second metric.
• (c)  
  Endurance, 2 min walk: participants walked continuously and as fast as they could without running. The route consisted of walking back and forth around cones placed 25 feet (7.62 m) apart. Once completed, the total distance walked was calculated manually.
• (d)  
  Lower limb strength, repeated sit-to-stand-to-sit (×2): after a practice, participants performed five sit-to-stand-to-sit posture transitions (PT), with arms folded across their chest, as quickly as possible. Participants were instructed to stand fully and not to touch the back of the chair during each repetition.
• (e)  
  Lower limb strength and locomotion, TUG (×3): after a practice, participants stood up from a chair (height: 40–50 cm), walked 2 m at a normal pace, around a cone, back to the chair, turned and sat down. The TUG time was recorded manually as the time from initiation of chair rise to the time when the participant's back touched the backrest of the chair at the end of the manoeuvre.

The following is a brief overview of the algorithms used to instrument each of the tasks outlined in section 2.3:

• (a)  
  Algorithm #1 (postural control): jerk (rate of change of a_M, equation(1)), root mean square (RMS, a_M magnitude equation (2)) and frequency components (95% percentile (F95%), ellipsis) were implemented (Moe-Nilssen and Helbostad 2002, Mancini et al 2012). Figure 2(a) shows an example of the data with the segmentation markers to extract the exact period of standing data. Due to its sensitivity, we present acceleration data within the mediolateral (a_M) direction only (Moe-Nilssen and Helbostad 2002).

  $$\begin{eqnarray}&&\text{Jerk}=\frac{1}{2}\underset{0}{\overset{t} \int} {{\left(\frac{\text{d}{{a}_{M}}}{\text{d}t}\right)}^{2}}\end{eqnarray} \tag{ 1 }$$

  $$\begin{eqnarray}&&\text{RMS}=\sqrt{\frac{1}{n}\left(a_{M1}^{2}+\cdots +a_{Mn}^{2}\right)}\end{eqnarray} \tag{ 2 }$$
• (b)  
  Algorithm #2 (locomotion, endurance): we utilised a Gaussian continuous wavelet transform to estimate the initial contact (IC) and final contact (FC) gait time events from a_V (McCamley et al 2012). From the calculation of IC/FC times, we recorded total time to complete the 4 m (time between the first IC and last FC). In addition, during the endurance task, we calculated numerous temporal gait characteristics based on the estimated IC/FC events and their sequence within the gait cycle, i.e. step, stride, stance and swing times.
• (c)  
  Algorithm #3 (locomotion, endurance): we applied the inverted pendulum model to estimate step length (Zijlstra and Hof 2003) and, from this, calculated the total distance walked during the endurance task (summation of step lengths). The model is based on the vertical movement of the centre of mass (h) due to the double integration of a_V and length of the pendulum (l, height of sensor from ground), equation (3).

  $$\begin{eqnarray}&&\text{step}\text{ length}=2\sqrt{2lh-{{h}^{2}}}\end{eqnarray} \tag{ 3 }$$

  Algorithm #2 + #3 (locomotion, endurance): the estimates of step time and step length were combined to generate values for mean step velocity, equation (4).

  $$\begin{eqnarray}&&\text{step}\text{ velocity =}\text{}\frac{\text{step}\text{ length}}{\text{step}\text{ time}}\end{eqnarray} \tag{ 4 }$$
• (d)  
  Algorithm #4 (lower extremity strength, TUG): timed PT were estimated from a discrete wavelet transform (DWT, using a fifth-order approximation and Meyer wavelet) of the signal vector magnitude (SVM) from all three axes of the accelerometer (Bidargaddi et al 2007). The total time for five PT repetitions was derived from the initial trough (sit-to-stand) to the final peak (stand-to-sit), figure 2(b).

Zoom In Zoom Out Reset image size

This method also estimated total time to complete the TUG. Similar to the lower strength task the peak and trough of the transitions were used to calculate the time from start to finish, figure 2(c).

Paired sample t-tests were used to test for differences between manual and BWM quantified tasks. A repeated measures analysis of variance (ANOVA) with a Greenhouse–Geisser correction was used to examine differences in balance tests. For all analysis, statistical significance was set at p < 0.05.

Table 2 shows values for a_M for each task. The varying complexity of task from flat to foam surface is highlighted by the significantly increased RMS (p < 0.000) and Elliptical (p = 0.006) values. Jerk marginally increased between tests (p = 0.072) where greatest differences are observed between eye closed on flat surface compared to foam. There was no significant difference between tests for F95% (p = 0.122).

In trial 1, the BWM had a longer duration (greater) time compared with manual recording (−0.04 s) but this was not significant (p = 0.682). Mean difference (0.00 s) improved for trial 2 with no significant differences between times (p = 0.981).

The BWM significantly overestimated (p = 0.037) the total distance walked (approx. 11 m or 6.6%) compared with manually recording (table 3). It was also possible to determine additional gait characteristics from the BWM (table 4).

The BWM had significantly shorter duration (faster) time estimates for repeated PT (by approx. 1 s for both trials) compared with manual recordings (p < 0.000).

There were no significant differences between the BWM and manual recording for the total time to complete the TUG (p ≥ 0.319).

Quantifying balance with a BWM on the lower back has previously been shown to be reliable and suitable for vestibular impairment screening (Rine et al 2013). Moreover, similar work has shown accelerometry related outcomes (similar to those derived here) to be better or consistent with centre of pressure outcomes quantified using traditional methods, e.g. force plates (Whitney et al 2011). Exact comparison to other accelerometer-based postural control data from the literature is difficult due to the novelty of the work (use of a BWM) and variation in protocols. The outcomes quantified in this study were similar to other work (Mancini et al 2012) and also showed sensitivity to the difficulty of the protocol as seen previously (Rine et al 2013). We observed that F95% (ML direction) was the least discriminatory between trials. Therefore its use in HAP testing warrants further investigation.

This is the first study to utilise the timing sequence of IC/FC within the gait cycle to estimate total time to complete the 4 m walk and hence determine speed. Compared with manual recordings, the BWM had shorter durations (faster). Slight differences in trial 1 may be due to the subjective nature of manual recording, i.e. researcher error in timing the beginning/end of the walk. Conversely, the dependence of the BWM algorithm on identifying the final (purposeful) FC to mark the end of the trial may also impact upon timing error. However, in trial 2, mean difference between methods improved suggesting a learning effect for the observer and, perhaps, better participant compliance (complete stop after 4 m). One possible method for improving the FC event estimations is the alternative use of wavelet where it has been suggested that a bi-orthogonal spine wavelet is superior to that used here, i.e. Gaussian (Shao and Ma 2003). This will be evaluated in future studies.

The pendulum technique overestimated step length and subsequently total distance walked due to the spatial parameter being sensitive to linear deviations of gait (rounding/turning the cones) (Zijlstra and Hof 2003). Estimated distance by the BWM may be improved by introducing a more linear walking route, i.e. limit abrupt turning. We quantified numerous spatio-temporal gait characteristics to highlight the added benefit of a BWM. While this task has been recommended by the NIH Toolbox for assessing overall cardiovascular endurance, the inclusion of a BWM allows researchers to assess the gait characteristics (over a suitable period) (Galna et al 2013, Lord et al 2013a) which have been shown to be sensitive to age (Lord et al 2013b) and pathology (Rochester et al 2012) and could prove useful for future testing. Moreover, the quantified gait characteristics presented here are similar to other studies that used an instrumented walkway (Senden et al 2009, Hollman et al 2011, Lord et al 2013b, Kobsar et al 2014b). However, notable difference are observed for the variability of all gait characteristics which is similar to other work (Kobsar et al 2014a). This can be attributed to the functional differences between systems (pressure sensor versus continuous tracking of a body and its acceleration through space) and drift due to integration (Hartmann et al 2009). However, this approach offers another important dimension to the endurance task which will be investigated in future HAP work within LiveWell.

This is the first study to use the wavelet algorithm to quantify repeated PT as it has been specifically designed for individual sit-to-stand and stand-to-sit transitions only. The mean times had a significantly shorter duration (faster) for the BWM compared to manual recordings (p < 0.000). This could be due to observer variability in recording the start and finish times of the transitions which can be difficult to achieve due to participant performance. Other factors that may impact on calculations include BWM time to complete repeated PT, taken as the time between first and last peaks, figure 2(b). Methodologically the correct timing sequence for a single transition is the time between successive peak/trough or trough/peak ×2 (Bidargaddi et al 2007). Multiplication by the correction factor of 2 was excluded here due to the nature of the task, i.e. a continuous sequence of sit-to-stand-to-sit. The adaptation of the algorithm to include a correction <2 may help refine the slight timing anomaly, which appears systematic in nature, figure 3(a). To test this hypothesis, we calculated a new correction factor for future use with this algorithm during repeated PT. Based on the percentage difference between both methods we computed and subsequently recommend a value of 1.16. When applied to BWM values and compared to those from manual recordings, no significant differences were found (p ≥ 0.603), figure 3(b).

Zoom In Zoom Out Reset image size

Furthermore, the shorter duration PT estimates by the BWM compared to manually recorded times can be attributed to algorithm functionality due to BWM location and definition of PT (Godfrey et al 2014). This is because the algorithm is best suited to estimation of the dominant vertical rise (DVR) transition whereas momentum transfer is quantified by the visual observation of the researcher and that which is normally adopted by OHP (Scarborough et al 2007). However, due to its mechanics the DVR may be a good proxy for assessing lower extremity strength within this task but remains untested (Godfrey et al 2014). Change of sensor location (chest) to suit the momentum transfer strategy is likely to improve agreement but impact negatively on the ability to holistically and accurately quantify the subdomain of physical capability outlined in this study (Godfrey et al 2014).

There were no significant differences between times quantified by the BWM and observer although the BWM had (generally) shorter durations in comparison to manual recordings. This could be explained by the algorithm adopted to quantify PT (exclusion of the correction factor) and the PT strategy adopted (see discussion above, section 4.4). However, this is the first time that the wavelet algorithm has been used to quantify the TUG test and it performed well in recognising the PT at the beginning/end of the task as well as suppressing the walk and turning components of a_V, figure 2(c). Slight differences by the BWM can be attributed to the nature of the composite task (sit-to-stand, walk, turn, walk, turn and stand-to-sit), which may negatively impact on the peak detection accuracy and hence timing. Segmenting the task into its various components with the summation of their individually timed sections may improve results, though this is only attainable with the use of a gyroscope (Salarian et al 2010).