Clinical measures.

Four clinical measures, which are commonly used in research and clinical practice with older adults, were used to evaluate functional mobility. Participants completed the following protocols in a randomised order: Timed Up and Go (TUG) [24], stair test [25], chair stand test [26], and Tinetti Performance Oriented Mobility Assessment (POMA) [27]. For each of the clinical measures, participants completed one familiarisation trial followed by two testing trials [28], with a one-minute rest period between trials [29].

Balance measures.

Multiple static posturography protocols were performed during a single visit to the laboratory using the NeuroCom VSR SPORT force platform integrated with the Balance Manager software (NeuroCom International, Inc., Clackamas, Oregon, USA). Balance performance was quantified during five testing protocols: Modified Clinical Test of Sensory Interaction on Balance (feet apart and narrow stance) [30, 31], unilateral stance [32], limits of stability [33], and weight bearing squat [9]. Throughout all measurements, a screen was positioned at eye level, 1.5 m in front of the force plate [34]. When required, a high-density foam pad was used to create a compliant surface. Participants completed testing barefoot, facing the screen with hands on hips. A one-minute rest period was provided between trials, and a two-minute rest period was provided between protocols [35].

Gait measures.

Kinetic and kinematic data were simultaneously collected during walking trials along a 10-meter indoor walkway at two gait speeds (self-selected usual and maximal). Participants were asked to walk at a comfortable everyday pace for the usual gait speed (UGS) trials and as quickly but safely as possible for the maximal gait speed (MGS) trials [10]. For both speeds, participants completed five familiarisation trials followed by five testing trials, with two minutes rest between trials. To determine gait speed, timing gates (Witty; Microgate, Bolzano, Italy) mounted on tripods were positioned at approximately hip height, five metres apart in the centre of the walkway. GRF data were acquired from three force platforms (Kistler Instruments Ltd., Winterthur, Switzerland) sampling at 1,000 Hz. Data were acquired using the BioWare software (version 5.3.1.7; Kistler Instruments Ltd., Winterthur, Switzerland), configured to record for 20 seconds per trial. Trials were deemed acceptable when both feet contacted separate force plates, without participants noticeably altering their gait style or targeting the plates.

Two-dimensional video data were collected from one high-speed camera (Fastec TS3; Fastec, San Diego, CA, USA) placed perpendicular to the walkway and eight metres from the centre of the force platforms. The camera settings included a frame rate of 100 Hz, shutter speed of 1/1,000 s, resolution of 1,280 x 1,024 pixels, and f-stop of 2.0. Before testing, tape markers were placed on the hip (lateral aspect of the greater trochanter), knee (midpoint between the lateral convexities of the femur and tibia), and ankle (lateral malleolus) joints on the right-hand side of the body to aid digitising reliability during kinematic analysis. Participants completed all trials in their own footwear. A reference frame was constructed using four metal poles placed in the sagittal plane in the centre of the walkway. This was recorded and later used for calibration.

Body composition measures.

Dual energy X-ray absorptiometry (DXA) scans (Lunar iDXA with enCORE software version 15.0; GE Healthcare, Madison, WI, USA) were performed to assess body composition, bone mineral density (BMD), and hip structure. All scans were performed by the same trained operator. Participants were asked to arrive fasted, in a euhydrated state, and having participated in no vigorous exercise for 12 hours preceding the scans [36]. Participants were also asked to void their bladder before the scans [37]. The scanning mode was automatically determined by the enCORE software based on body size; all participants in this study were scanned using the standard mode (estimated body thickness 0.16 m to 0.25 m).

Total and regional body composition measures were derived from one total body scan, lasting approximately seven minutes [38]. Participants adopted a supine position, aligned with the centre line on the DXA scanning table, with the head positioned at the horizontal line at the top of the scanning bed. Proximal femur BMD and structural geometry were evaluated from one left femur scan [39]. For this, participants remained in a supine position in the centre of the scanning table. For any participants with history of a left hip replacement, the right femur was scanned instead (n = 2). On completion of the DXA scans, participants had the opportunity to eat and drink before moving onto the strength measurements.

Strength measures.

Strength assessments were performed at the trunk, knee and ankle for both limbs using an isokinetic dynamometer (System 4 PRO; Biodex Corp., Shirley, New York, USA). For the knee and ankle measurements, the order of testing (joint, type of contraction and speed) was randomised. The rotational axis of the dynamometer was visually aligned with a line traversing the femoral epicondyles at the knee joint centre, and the resistance pad was positioned securely on the tibia superior to the medial malleolus [40]. The testing thigh, contralateral limb, trunk and pelvis were stabilised throughout the protocol using Velcro straps. Peak torque and rate of torque development (RTD) were measured during maximal isometric knee extension trials, with participants completing three submaximal trials before three maximal test trials. Each contraction was performed for five seconds, with a 30 second rest period between trials [41]. Maximal isokinetic concentric joint torques were assessed at the knee and ankle for both flexion and extension at angular velocities of 60°/s and 120°/s [42, 43]. Participants completed five submaximal trials followed by three maximal test trials. Following these measurements, concentric trunk flexion and extension data under isokinetic conditions were collected at angular velocities of 20°/s and 45°/s [44] using the Biodex Dual Position Back Extension/Flexion attachment (Biodex Corp., Shirley, New York, USA) attached to the dynamometer. The fixed axis of the dynamometer was aligned level with the anterior superior iliac spines [45]. Participants performed five submaximal warm up trials before five maximal test trials. For the isokinetic trials, a one-minute rest period was provided between sets.

Clinical measures data.

For each of the clinical measures (except gait speed), the average performance across both test trials was calculated and used within the analysis [46]. In total, ten variables were obtained from the clinical data protocols (Table 2). Gait speed data were collected alongside the other gait variables but included within the clinical measures dataset, given their ease of measurement and common use in clinical and community settings. Further details are provided in the gait data section below.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Variables included within the clinical measures, balance, gait, strength and body composition data packages for the single-domain analyses.

https://doi.org/10.1371/journal.pone.0293729.t002

Balance data.

The balance data were exported from the Balance Manager software, and any additional processing was completed using Microsoft Excel. In total, 52 variables were obtained from these protocols (Table 2). To quantify the visual contribution to balance during the firm and foam trials, the Romberg ratio (eyes closed sway velocity/eyes open sway velocity) was calculated [47]. To quantify the somatosensory contributions to balance during the eyes open and eyes closed trials, the somatosensory ratio (sway velocity foam/sway velocity firm) was calculated [48].

To determine inter-leg symmetry during the unilateral stance test and weight-bearing squat, the symmetry angle [49] was calculated for the eyes open and eyes closed trials, using the arctan function of the ratio of mean values measured from the left and right limbs [50]. As the focus of this analysis was to evaluate the magnitude of asymmetry rather than the direction, the absolute values were reported.

Gait data.

GRF data processing was completed using BioWare software (version 5.3.1.7; Kistler Instruments Ltd., Winterthur, Switzerland), and the video files were analysed using Simi Motion analysis (version 9.2.3; SIMI Reality Motion Systems, Munich, Germany). Overall, 99 variables (Table 2) were processed from the gait data and included in further analysis. Five successful steps for both gait speeds were identified per participant and included within the analysis. For each participant, the five steps analysed were from the same limb. The GRF data were filtered with a second-order low pass Butterworth filter, with a 50 Hz cut-off frequency [51]. UGS and MGS (m/s) were determined for each trial using the distance walked and ambulation time measured from the timing gates. Gait speed reserve was calculated as the ratio between MGS and UGS [52], and was used to quantify the capacity to increase walking speed when needed (as noted above, UGS, MGS and gait speed reserve were analysed in the clinical measures dataset).

Gait variability was measured using the median absolute deviation (MAD) using the five steps for each variable. The MAD provides a robust estimate of variability, which is less sensitive to outliers and artificial inflation compared with other measures such as the coefficient of variation [53]. To allow for comparisons between groups and variables, the MAD scores for each variable were reported as a percentage of the original median value [54]. The percentage MAD scores were calculated for all gait variables at both gait speeds, apart from step width and change in horizontal velocity.

Strength data.

In total, 70 strength variables (Table 2) were obtained from the different protocols and included in further analysis. The peak torque trials (sampled at 100 Hz) were exported from the Biodex software for processing using in-house algorithms written in MATLAB (The MathWorks Inc., Natick, Massachusetts, USA). Flexion/extension strength ratios were calculated for the trunk, knee and ankle by taking the quotients between the isokinetic flexor peak torque and isokinetic extensor peak torque. These were calculated for all testing speeds and were determined for the left and right limbs during the ankle and knee trials. For the RTD trials, the torque signal from the dynamometer was sampled at 2000 Hz using a Biopac MP150 data acquisition system integrated with the AcqKnowledge software (version 4.4; Biopac Systems Inc., Santa Barbara, CA, USA). The signal was exported and processed offline using custom written algorithms in MATLAB (The MathWorks Inc., Natick, Massachusetts, USA). The signal was filtered using a second-order low-pass Butterworth filter, with a cut-off frequency of 150 Hz [55]. RTD was defined as the slope of the torque-time curve (△torque/△time) for three time intervals from the onset of the contraction [56]. The onset was determined as the point where the torque signal reached 4 Nm above baseline [57]. These time periods were chosen to represent the rapid muscle responses (≤200 ms) needed to prevent falling when regaining balance following a trip or slip incident [58]. To determine inter-leg symmetry for the peak torque values, the symmetry angle (%) was calculated using the methods described previously.

Body composition data.

All data from both DXA scans were analysed using the enCORE software (version 15.0; GE Healthcare, Madison, WI). For the body composition analysis, regions of interest and cut points were automatically determined by the enCORE software. A total of 14 body composition variables, nine BMD variables, six hip structure variables, and five MQ variables were obtained and included in further analysis (Table 2). Several variables were estimated from the total body scan and additional regions of interest were manually segmented to estimate lean tissue mass in the upper and lower leg for both limbs.

MQ was defined as muscle strength per kilogram of lean tissue mass [59]. Upper leg MQ was calculated using the following equations: isometric knee extension peak torque/upper leg lean tissue mass, isokinetic (60°/s) knee extension peak torque/upper leg lean tissue mass, and isokinetic (60°/s) combined peak torque (knee extension + flexion)/upper leg lean tissue mass [59]. Lower leg MQ was defined using the following equations: isokinetic (60°/s) plantar flexion peak torque/lower leg lean tissue mass, and isokinetic (60°/s) combined peak torque (plantar flexion + dorsiflexion)/lower leg lean tissue mass [60]. These indices were defined for the dominant and non-dominant limbs, based on the highest torque measured during the left and right trials. In addition, a composite measure of MQ was defined for each of the five indices, which was an average measure of the dominant and non-dominant limbs. Inter-leg symmetry was calculated for each of the MQ indices using the symmetry angle (%) as described previously.

Single-domain analysis.

The initial single-domain results were derived using traditional univariate techniques before applying a machine learning multivariate approach. Firstly, two-sample t-tests and effect sizes (Cohen’s d) were used to determine differences between fallers and non-fallers with statistical significance set at p<0.05. For categorical variables, group differences and effect sizes were determined using chi-squared tests and Cramer’s V. Pearson’s correlation coefficients were also used to quantify the relationships between variables, with r values of 0.10, 0.30 and 0.50 representing small, moderate and large associations, respectively [62]. The absolute value of the correlation coefficient was also used to provide an indication of multicollinearity between variables, with r values >0.50 suggesting high collinearity [63]. Bartlett’s test of sphericity was used to assess the redundancy in the data using the correlation matrix [64]. This test determined whether the correlation matrix was significantly different from an identity matrix (i.e. a matrix with ones along the diagonal and zeros for all other entries). From Bartlett’s test, p<0.05 indicated that there was redundancy in the data and that the data were not orthogonal (i.e. uncorrelated).

From a machine learning perspective, random forest analysis (based on 500 trees) [65]) and leave-one-variable-out (LOVO) partial least squares correlation analysis (PLSCA; as described by Weaving et al. 2019 [16]) were employed to identify the important variables that best discriminated between fallers and non-fallers. Variables were identified as important if they were above the cut-off ‘elbow’ on the variable importance scree plots produced using both techniques. A mixed-methods approach, using the Jenks natural breaks algorithm [66] with a subjective validation, was used to determine the optimal cut-off point for important variables. Following this, the important variables selected from both techniques were combined into two refined datasets (comprising a smaller number of variables). Using the refined datasets, classification models were constructed to differentiate between fallers and non-fallers, using PLSCA [67], random forest [65], and logistic regression techniques. Further details about the random forest, PLSCA and logistic regression techniques are presented in S1 Appendix.

The receiver operating characteristics (ROC) area under the curve (AUC) was used throughout as a metric of diagnostic accuracy, providing an index of discriminatory ability [68]. An AUC value of 0.50 represented no discriminatory ability, with values of 0.70 to 0.79 considered acceptable, 0.80 to 0.89 considered excellent, and ≥0.90 considered outstanding [68]. An AUC value of one represents perfect discrimination between groups.

Multi-domain analysis.

Overall, 51 variables were included in the further multi-domain analysis. These variables were identified as important by the random forest and LOVO PLSCA during the five single-domain analyses. When compiling the multi-domain data package, some participants had missing values. Therefore, for the multi-domain analysis it was necessary to impute the missing data so that all observations were included (n = 60; fallers = 21, non-fallers = 39), thus enabling the use of the same consistent machine learning strategy and minimising the loss of valuable information that could have been beneficial for discrimination between groups. As such, any missing values were imputed using the Probabilistic Principal Component Analysis (PPCA) technique [69].

Analysis was undertaken using the same multivariate machine learning approach described for the single domains. Following traditional univariate analyses, random forest and LOVO PLSCA were employed to identify the important variables which best discriminated between fallers and non-fallers, as well as highlighting variables of less importance. Subsequently, the important variables selected from both techniques (i.e. above the cut-off elbow on the variable importance plots) were combined into one final refined dataset. Using this refined dataset, classification models were constructed to differentiate between fallers and non-fallers, using PLSCA, random forest and logistic regression techniques.

To determine the general applicability of the classification models (i.e. to test how each model performed on unseen data) and to check that the models were not overfitting the data, cross validation was performed. For the random forest models, the inherent out-of-bag cross validation methods were used (meaning that a separate cross validation was not needed) [70]. Following pilot work, leave-one-out cross validation (LOOCV) was deemed most appropriate for the logistic regression models [71]. This is supported by previous work which has suggested that LOOCV performs well with small sample sizes and produces comparable results to 10-fold cross validation methods [72]. Further details about the cross validation are presented in S1 Appendix.

In addition, and for the first time in this area, PLSCA was used to quantify the strength of the relationships between the various domains. Whilst univariate correlation analyses were performed to quantify the associations between individual variables in each data package, PLSCA has the advantage that it can quantify the strength of the relationships between multiple groups of variables in different domains. This was done by applying PLSCA bilaterally between the domains. The amount of shared information was determined by calculating the singular value inertia, with greater values representing more shared information and stronger relationships between the single-domain data packages [73].

Balance domain.

It is noticeable that the balance variables identified as important were associated with the more challenging protocols (e.g. limits of stability and unilateral stance) with these tasks playing an important role in many activities of daily living (e.g. walking, turning and reaching). These findings corroborate those of others who have reported that fallers exhibited greater instability compared with non-fallers when standing on one limb [74] and on foam with eyes closed [9]. However, the importance of several limits of stability variables is in contrast to some studies [31, 33, 75]. Fallers seemed to have a greater reliance on somatosensory inputs compared with non-fallers; this is a promising finding because this variable is easy to calculate if sway velocity data is being collected on a firm and foam surface. Melzer et al. (2004) [31] suggested that balance control in narrow stance is a useful tool for discriminating between fallers and non-fallers. As such, it was surprising that no variables were selected as important from the narrow stance protocols. It is important to note that instability increased for both groups during the narrow stance trials highlighting that these trials were challenging for both groups and may be unsuitable for falls discrimination purposes in older women.

Gait domain.

The analysis of the gait measures indicated that a combination of spatiotemporal, kinematic, GRF, and variability variables can distinguish between fallers and non-fallers with an outstanding degree of accuracy (AUC≥0.90). The gait variables identified as being good discriminators were taken from both the UGS and MGS trials which supports the use of these conditions for screening purposes [52]. Overall, the important variables suggest that fallers adopted a more cautious gait strategy compared with non-fallers, with fallers exhibiting shorter steps, greater knee flexion at toe-off, a longer braking phase duration, and a more pronounced double support strategy. These observations may be indicative of reduced dynamic balance ability [76] and weight acceptance ability [77] in fallers. For some of the gait measures, fallers demonstrated increased variability, suggesting that they walked with an inconsistent gait pattern [78]. On the other hand, fallers demonstrated lower variability for some variables, suggesting that a degree of variability is necessary for maintaining dynamic balance [79], although this may demonstrate the availability of fewer strategies to deal with gait instability and perturbations [80]. Several gait variables appeared redundant when discriminating between groups, such as mid-stance knee angle (an estimation of foot clearance during swing) and propulsive impulse. Considering these variables would seem important for the navigation of obstacles and control of speed during the gait cycle [81], it may be that both groups exhibited age-related declines (independent of falls history) which reduced the sensitivity of these variables for falls discrimination.

Strength domain.

Several peak torque variables, namely knee flexion, dorsiflexion and plantar flexion, were identified as important discriminators which is in agreement with previous investigations that have reported lower maximal strength in fallers at the knee and ankle [34, 40]. Considering the key role of these muscles during activities of daily living [46], the reduced strength capacity of fallers likely contributed to the balance [82] and gait differences [83] observed between the two groups. Three asymmetry variables were identified as important discriminators, although an inconsistent pattern was found across different muscle groups and contraction types [84]. Whilst more research is needed to fully understand the patterns of strength asymmetry in elderly older women, the measurement of asymmetry should be considered in research and clinical practice. The variable importance analysis highlighted some variables that were not important in discriminating between the two groups. These included knee extension peak torque, knee extensor RTD and trunk strength. These findings likely reflect the contrasting research that exists regarding the discriminatory ability of RTD [40, 85] and trunk strength [86, 87] that may result from differences in the measurement protocols (i.e. muscle group, contraction type) that have been adopted. Nevertheless, the present findings highlight that maximal strength variables for the knee flexors, dorsiflexors and plantar flexors may be more important than maximal or rapid strength variables for the knee extensors or trunk muscles for inclusion in falls screening protocols.

Clinical measures.

A combination of variables measured during the TUG, POMA and gait speed protocols appeared useful when differentiating between groups. Both the TUG and POMA incorporate a range of movements used during daily living, which may explain why these clinical measures were identified as important. Although previous studies have also reported significant differences between fallers and non-fallers for TUG time [88, 89], the optimal cut-off threshold (7.85 s) in this work was quicker than most of the previous literature and values used in clinical settings [90]. This suggests that quicker cut-off thresholds may be necessary to improve the classification accuracy of the TUG. A novel aspect of this study was the inclusion of gait speed reserve [52], with fallers demonstrating a greater capacity to increase walking speed relative to their UGS. Interestingly, performance during the chair stand and stair tests were not important discriminators despite their similarities with activities of daily living [91]. From this, it may be suggested that these measures are not chosen for discrimination purposes in this population.

Body composition domain.

The results showed that MQ appears to be an important discriminator between fallers and non-fallers. The findings concur with the few studies that have analysed this variable and reported fallers as having poorer MQ compared with non-fallers [92]. MQ has previously been associated with gait and functional performance in older adults [60, 93], and in this way, lower MQ in fallers may have contributed to poorer performance in gait and clinical measures variables. Total fat mass was also selected as an important discriminator, with fallers demonstrating higher fat mass compared with non-fallers as observed previously [94]. Increased body fat may be associated with greater intramuscular fat infiltration which can impair muscle function leading to declines in muscle strength [95]; however, DXA scans are unable to detect fat infiltration in muscle [96]. Although the use of DXA is not without its limitations [29], it is routinely used in this population, and the findings demonstrate the discriminatory sensitivity of segmental MQ measures which are easily obtained from these scans. In terms of BMD, the femoral strength index was selected as an important discriminator, with fallers demonstrating lower values for the this compared with non-fallers, suggesting poorer bone strength and an increased risk of hip fracture from a fall on the greater trochanter [97]. Given that the femoral strength index is a composite variable which integrates BMD and structural parameters and is adjusted for body size [98], it may provide more insight than individual measures of bone structure and geometry alone, which were not identified as important.