Fear of falling in community-dwelling older adults: What their gait acceleration pattern reveals

Abstract

Background:

Fear of Falling (FOF) is common among community-dwelling older adults and is associated with increased fall-risk, reduced activity, and gait modifications.

Objective:

In this cross-sectional study, we examined the relationships between FOF and gait quality.

Methods:

Older adults (N=232; age 77±6; 65 % females) reported FOF by a single yes/no question. Gait quality was quantified as (1) harmonic ratio (smoothness) and other time-frequency spatiotemporal variables from triaxial accelerometry (Vertical-V, Mediolateral-ML, Anterior-Posterior -AP) during six-minute walk; (2) gait speed, step-time CoV (variability), and walk-ratio (step-length/cadence) on a 4-m instrumented walkway. Mann Whitney U-tests and Random forest classifier compared gait between those with and without FOF. Selected gait variables were used to build Support Vector Machine (SVM) classifier and performance was evaluated using AUC-ROC.

Results:

Individuals with FOF had slower gait speed (103.66 ± 17.09 vs. 110.07 ± 14.83 cm/s), greater step time CoV (4.17 ± 1.66 vs. 3.72 ± 1.24 %), smaller walk-ratio (0.53 ± 0.08 vs. 0.56 ± 0.07 cm/steps/minute), smaller standard deviation V (0.15 ± 0.06 vs. 0.18 ± 0.09 m/s²), and smaller harmonic-ratio V (2.14 ± 0.73 vs. 2.38 ± 0.58), all p<.01. Linear SVM yielded an AUC-ROC of 67 % on test dataset, coefficient values being gait speed (−0.19), standard deviation V (−0.23), walk-ratio (−0.36), and smoothness V (−0.38) describing associations with presence of FOF.

Conclusion:

Older adults with FOF have reduced gait speed, acceleration adaptability, walk-ratio, and smoothness. Disrupted gait patterns during fear of falling could provide insights into psychosocial distress in older adults. Longitudinal studies are warranted.

1. Introduction

Approximately one in four U.S. residents aged over 65 report falling each year while fall-related emergency room visits and deaths from falls in older adults increased between 2007 and 2016 [1]. Fall risk is often thought to be the result of degrading physical ability; however, there has been increasing interest in the role of psychological factors, such as fear of falling (FOF). The combination of FOF with a fall history and reduced gait speed predicted future falls and injury [2]. It is important to note however, that the relationship between FOF and fall history is inconsistent in the literature, and that the perceived and physiological fall risk do not always correspond. For example, some studies found no association between FOF and recurrent falls [3], while in other studies, perceived and physiological fall risk were shown to be independent predictors of future falls [4,5]. Additionally, FOF is associated with activity restriction [6], lower quality of life [7], and reduced functional ability [8]. FOF is independently associated with bodily stiffening and inefficient use of attentional resources that modify balance and gait while increasing risk of falling [9]. Thus, factors beyond having a fall history need to be investigated in relation to falls and FOF.

Increased gait variability [10] and decreased gait speed [11] are two factors associated with increased fall risk, which is associated with FOF [12,13] independent of history of falling [14]. Measurements of gait variability obtained from instrumented walkway can have clinical and fall risk implications [10] which seem to be most important in those with a normal gait speed [15]. Accelerometry obtained metrics of gait are an important addition to these measures, as they are generalizable to global patterns of gait kinetics and can be extended to overground walking in everyday environments [16]. Accelerometry from an accelerometer secured to the trunk, provides measures of trunk movement in three directions, the mediolateral (ML), anteroposterior (AP), and vertical (V) directions. Trunk accelerometry measures are useful for detecting the movements of the body at the center of balance. Through these measures, past research has found an association between FOF and less smooth trunk oscillation in the ML and AP directions [17]. Multi-directional harmonic ratios also have use in predicting future falls in older adults [18], even in those with no history of falls [19]. Similarly, other accelerometry derived metrics, such as peak frequency [20] and wavelet entropy [21] have been associated with fall risk; however, the utility of these metrics in understanding FOF is not clearly understood. Further exploration into utility of measures such as statistical features, information theory, frequency, and time-frequency measures extracted from accelerometry measures have shown that many of these features provide unique and complementary information in discriminating healthy and abnormal gait [22,23]. While slower gait speed may be related to several negative health outcomes, disruptive accelerometry gait patterns may be unique to specific functional deficits. The role of gait measures in identification of fear of falling is not known. Statistics is generally associated with drawing inferences from data, whereas machine learning is more concerned with finding generalizable predictive patterns [24]. In this analysis, we leverage both methods – traditional statistical tests as well as the supervised machine learning to better our understanding of fear of falling in older adults.

Many gait factors have been investigated independently in relation to FOF; however, little research has investigated the full range of spatiotemporal and accelerometry based gait characteristics together. Exploring combined gait metrics and their associations with FOF could prove useful in establishing a hierarchy of features that are most relevant to understanding FOF. Given the basis of FOF, we hypothesize that in addition to a slower gait speed, accelerometry measures of timing aspects and smoothness will add to our understanding of FOF. The innovation in this approach is the use of both statistical as well as data-driven machine-learning methods.

2. Methods

2.1. General study design and participants

Data were collected at the baseline visit of Program for Improving Mobility in Aging (PRIMA), a randomized exercise trial meant to compare the effects of interventions on mobility, activity, and participation in older adults [25]. Participants were recruited from areas around Pittsburgh, USA. All data were collected prior to the COVID-19 pandemic. PRIMA enrolled 249 community-dwelling, older adults (65 % female, mean age 77±6). Eligibility for the intervention included having an age of at least 65 and a usual gait speed between 0.6 m/s and 1.2 m/s at time of enrollment. To participate in the study, individuals had to be ambulatory without an assistive device or the assistance of another person. Additionally, participants were ineligible if they had dementia or any other neurological disorder and physical disabilities or if they had medical conditions that made testing or participation in an exercise program unsafe. These included older adults with arthritis, lumbar stenosis, peripheral arterial disease, chronic obstructive pulmonary disease, fixed or fused lower extremity joins such a hip, knee, or ankle, history of stroke. Participants with any progressive movement disorder such as multiple sclerosis or Parkinson’s disease were also excluded. A complete list of exclusion criteria were previously published [25]. In these analyses, we included 232 participants with valid baseline accelerometry data from the lower back. Due to technical issues during data collection and processing, 7 % of the data (17 subjects) could not be utilized. All study activities were reviewed and approved by the University of Pittsburgh Institutional Review Board. Participants provided written informed consent prior to completion of study activities.

2.2. Fall history, fear of falling, and other health characteristics

Fear of falling was measured by the question “Are you afraid of falling?” with answer options of either ‘yes’ or ‘no’. Falls in the past year were self-reported by the question “Have you had a fall in the past year?” with responses of either ‘yes’ or ‘no’. Cognitive health was measured as the Modified Mini-Mental State exam (3MS; score between 0-100 with higher scores indicating better overall cognitive function) [26]. Comorbidities were self-reported and measured using the Duke Comorbidity Index taking into consideration cardiac, neurological, musculoskeletal, visual, and pulmonary conditions as well as depression, sleep, pain, diabetes, and cancer [27] (categorized and summed to create a score ranging from 0 to 8). The Geriatric Depression Scale (GDS) was recorded as a score ranging from 0 to 15 with higher scores indicating more depressive symptoms [28]. Walking confidence was reported by the Modified Gait Efficacy Scale (mGES; score between 10-100 with higher scores indicating greater confidence) [29]. Self-reported mobility in the past month was reported as the Life Space Assessment score (LSA; score between 0-120 with greater scores corresponding to better overall mobility) [30].

2.3. Gait Analysis

Measures for gait performance were obtained through two modalities. 1) walking along a 4-m long instrumented walkway (Zeno Walkway Protokinetics) and 2) an accelerometer (GT3X Actigraph LLC) secured on the lower back.

2.3.1. Gait measures from instrumented walkway

Participants began walking approximately two meters from the start of the mat, stopped approximately two meters past the end of the mat, and completed six passes at their self-selected usual gait speed. The reported gait speed is the subject’s average gait speed over six passes. During these six passes, 24 total steps are generally available for reliable extraction of these measures [31]. PKMAS software (Protokinetics, Havertwon, PA) was used to derive final set of gait variability measures i.e., coefficient of variation of step length, step time, and stride width [31]. Besides these measures, walk ratio (step length/cadence) was also computed from PKMAS. Walk ratio is considered an index of neuro-motor control [32].

2.3.2. Gait measures from accelerometry

For the gait signal data, a tri-axial body-worn accelerometer was placed at the lower back at the L3 vertebral level. Data were collected during a six-minute walking test [33] in which the participant was instructed to walk as far as they could in the six minutes allotted. If the participant needed to sit down, the test concluded. These accelerometry signals were collected at a sampling rate of 100 Hz for 202 subjects. Due to technical issues, the data for 30 subjects (12.93 %) were collected at 30 Hz and were then up-sampled to 100 Hz with no compromises in the detection of heel strikes or toe-offs [23,34].

Measures from four different signal domains in each of the ML, V, and AP directions were extracted, which included: 1) Statistical features-standard deviation of signal amplitudes (signal amplitude variability) and cross-correlations between axes (signal symmetry and similarity); 2) Frequency measures - harmonic ratio (smoothness), peak frequency (power), centroid frequencies, and bandwidth (range); 3) Information theory measures -Lempel Ziv complexity and entropy rate (regularity); 4) Time-frequency measures - wavelet entropy. Many of these features were described in a prior review of acceleration signal derived measures of gait [35] to further explore the relationships outlined in that review. (See Table 2 for the list of gait metrics calculated for the instrumented walkway and Accelerometry walking.)

Table 2.

Association of fear of falling to gait variables derived from instrumented walkway and tri-axial accelerometer in older adults.

Gait Variable	Total (N=232)	Fear of falling YES (n=96)	Fear of falling NO (n=136)	p
Instrumented walkway measures
**Gait Speed (cm/s)	107.6 ± 16.1	103.7 ± 17.1	110.1 ± 14.8	0.006
**Step Time CoV (%)	4.0 ± 1.5	4.2 ± 1.7	3.7 ± 1.2	0.034
^Step Length CoV (%)	5.1 ± 2.2	5.4 ± 2.4	4.9 ± 1.9	0.090
Stride Width CoV	40.7 ± 74.6	50.7 ± 69.7	43.6 ± 30.5	0.151
**Walk Ratio (cm/steps/minute)	0.5 ± 0.1	0.5 ± 0.1	0.6 ± 0.1	0.009
Accelerometry measures
Statistical measures
^Standard deviation ML (m/s2)	0.13 ± 0.05	0.12 ± 0.04	0.14 ± 0.05	0.093
**Standard deviation V (m/s2)	0.17 ± 0.07	0.15 ± 0.06	0.18 ± 0.09	0.006
*Standard deviation AP (m/s2)	0.18 ± 0.07	0.17 ± 0.06	0.19 ± 0.08	0.023
^Cross-correlation ML-V	0.20 ± 0.08	0.22 ± 0.09	0.20 ± 0.08	0.120
*Cross-correlation ML-AP	0.20 ± 0.09	0.22 ± 0.09	0.19 ± 0.08	0.032
^Cross-correlation AP-V	0.55 ± 0.16	0.53 ± 0.18	0.57 ± 0.15	0.058
Frequency measures
Peak frequency ML (Hz)	1.9 ± 1.7	1.9 ± 1.6	2.0 ± 1.7	0.253
Peak frequency V (Hz)	1.7 ± 0.9	1.8 ± 1.2	1.7 ± 0.6	0.188
Peak frequency AP (Hz)	1.7 ± 0.7	1.8 ± 0.7	1.6 ± 0.7	0.498
^Bandwidth ML (Hz)	3.3 ± 0.7	3.3 ± 0.7	3.4 ± 0.8	0.119
Bandwidth V (Hz)	3.3 ± 0.8	3.3 ± 0.8	3.3 ± 0.9	0.315
Bandwidth AP (Hz)	3.1 ± 0.8	3.1 ± 0.8	3.2 ± 0.8	0.217
^Centroid ML (Hz)	5.0 ± 1.2	4.8 ± 1.1	5.1 ± 1.3	0.063
Centroid V (Hz)	4.0 ± 1.3	4.0 ± 1.3	3.8 ± 1.3	0.419
Centroid AP (Hz)	3.3 ± 0.9	3.4 ± 0.9	3.3 ± 1.0	0.299
^Harmonic ratio ML	0.6 ± 0.3	0.6 ± 0.3	0.6 ± 0.3	0.083
**Harmonic ratio V	2.3 ± 0.7	2.1 ± 0.7	2.4 ± 0.6	0.001
*Harmonic ratio AP	2.8 ± 0.8	2.7 ± 0.8	2.9 ± 0.8	0.044
Information Theoretic measures
Entropy rate ML	0.9 ± 0.06	0.9 ± 0.06	0.9 ± 0.07	0.480
Entropy rate V	0.8 ± 0.08	0.8 ± 0.08	0.8 ± 0.08	0.379
Entropy Rate AP	0.9 ± 0.06	0.9 ± 0.07	0.9 ± 0.06	0.351
Lempel-Ziv ML	0.4 ± 0.09	0.4 ± 0.08	0.4 ± 0.10	0.339
Lempel-Ziv V	0.5 ± 0.09	0.5 ± 0.09	0.5 ± 0.10	0.307
Lempel-Ziv AP	0.4 ± 0.09	0.4 ± 0.08	0.4 ± 0.10	0.410
Time-frequency measures
Wavelet Entropy ML	2.1 ± 0.5	2.1 ± 0.5	2.1 ± 0.5	0.237
*Wavelet Entropy V	1.7 ± 0.6	1.7 ± 0.6	1.6 ± 0.6	0.031
^Wavelet Entropy AP	1.6 ± 0.6	1.6 ± 0.6	1.5 ± 0.7	0.067

CoV= Coefficient of Variation, ML= Mediolateral axis, V= Vertical axis, AP=Anterior Posterior axis. All Nonsignificant after Bonferroni correction (p= 0.05/32 ~ .002). Uncorrected significant levels

^**<.01,

^*<.05,

^{^}<.10

2.4. Statistical analysis

The appropriate Mann-Whitney U or Chi-square test used to determine mean differences in demographics and health characteristics for those with compared to without FOF. We used the Mann-Whitney U test to examine differences in gait by FOF status; this accounted for inconsistencies in normality for some gait variables. We applied the Bonferroni correction to account for multiple comparisons.

2.5. Machine learning modelling

Data adaptive approaches such as the random forest and kernel-based methods like support vector machine learning make fewer structural assumptions on the functional form of the relationship between predictors and the outcome, particularly when complicated interactions among the predictors exist.

2.5.1. Random forest modelling

Random Forest classifier models were built to rank gait variables associated with FOF. A random forest analysis was used to create a generalizable predictive model [36], and analyze the importance of included variables [37]. The predictive ability of random forest declines with larger numbers of predictors, so pre-filtering and selection of variables that were of meaningful importance and without strong cross-correlations was done to enrich the random forest [38]. Variable selection for the random forest modelling was conducted in two steps. First, all variables with a Mann-Whitney U test of mean difference by FOF status p < 0.1 were selected. Second, Spearman correlations to define inter-variable correlations of these selected gait variables. For variables with inter-variable correlations of ρ < 0.6, only one of variable pair was selected; the variable with greater discriminatory power identified as a lower p in the Mann U analyses of mean differences.

Variables included in the analysis were ranked in their relative importance through use of a Gini index, also known as Gini impurity [39]. This process identifies which variables drive the predictive decision of the model, classified as relative importance. We built a random forest classifier (bootstrap = True, maximum tree depth = 5, maximum variables = ’sqrt’, minimum samples in a leaf = 2, minimum samples to split = 2, number of estimators = 1000). Python 3.8 was used for machine learning modeling. The 5-fold cross-validation area under the curve (AUC) is reported.

2.5.2. Support vector machine modelling

Next, overlapping significant gait variables between both Mann Whitney analysis and Random Forest Variables were identified. These most consistently important variables were included for use in the creation of a Support Vector Machine (SVM) model. SVM is a popular learning-based classifier which selects a multi-dimensional classifying hyperplane such that the margin between classes is maximized [40]. Given input training vectors ${\text{x}}_{\text{i}}\in {\text{R}}^{\text{P}}$, i=1,2,…,n in two classes and a vector $\text{y}\in {\{1,-1\}}^{\text{n}}$, our goal is to find coefficient, $\text{w}\in {\text{R}}^{\text{p}}$ and residual, $\text{b}\in \text{R}$ such that the prediction given by $\text{sign}\left({\text{w}}^{\text{T}}\phi (\text{x})+\text{b}\right)$ is correct for most samples [41]. In a linear SVM, $\text{φ}(\text{x})=\text{x}$.

$${\min }_{w,b,\zeta }{w}^{T}w+C{\sum }_{i=1}^n{\zeta }_i$$ (1)

subject to ${\text{y}}_{\text{i}}\left({\text{w}}^{\text{T}}\text{φ}\left({\text{x}}_{\text{i}}\right)+\text{b}\right)\ge 1-{\zeta }_i,\varsigma \ge 0,\text{i}=1,2,\dots \dots ,\text{n}$

In Eq. (1), maximizing the margin $\left(\parallel w{\parallel }^2={\text{ww}}^{\text{T}}\right)$ and incurring a penalty when a sample is misclassified or within the margin boundary, a hyperplane separating the two classes is found. Since a perfectly separable hyperplane can be difficult to find, some samples are allowed to be at a distance ς_i from their correct margin boundary, penalty term C controls the strength of this penalty acting as an inverse regularization parameter.

For n=232 participants with each participant belonging to either of the two classes y = fear of falling (Yes and No), we aim to find w, corresponding to the top gait variables as selected by previously mentioned random forest method. Penalty term, C selected empirically to be 10. This analysis used an SVM to determine magnitude of variable coefficients and final classification of FOF. To build this classifier, the original dataset was randomly split into 2 non-intersecting subsets. The training set consists of n=185 (80 %) of participants. The test set included the remaining n=47 (20 %) participants and was used to evaluate the performance of the final classifier. The proportion of participants with and without FOF are maintained in both sets to avoid bias. Sensitivity, specificity, and area under the ROC curve were evaluated.

3. Results

Measurements were obtained from 232 participants (151 female) of mean(SD) age 77(6). Characteristics of the study sample stratified by FOF are presented in Table 1. A total of 96 participants (41 %) reported FOF and 70 participants (30 %) reported having fallen in the past year. Women were more likely to report FOF than men (Table 1). BMI was higher and reported gait efficacy was lower in those with FOF.

Table 1.

Demographics and health characteristics of older adults with (n=96) and without fear of falling (n=136). Mann Whitney and chi-square

Variable	Total (N=232)	Fear of falling YES (n=96)	Fear of Falling NO (n=136)	P
Age (years)	77.5 ± 6.6	77.6 ± 6.5	77.0 ± 6.8	0.306
Sex (females)	152(65 %)	72 (75 %)	80 (59 %)	0.016
BMI (kg/m2)	28.5 ± 5.8	29.2 ± 5.4	27.8 ± 5.2	0.024
Race (White)	207 (89 %)	84 (87 %)	123 (90 %)	0.619
Education (> 12 years)	187 (80 %)	80 (83 %)	107 (79 %)	0.475
Modified Mini-Mental State Exam (3MS) (0-100)	95.6 ± 7.0	96.1 ± 4.0	95.9 ± 4.0	0.385
Geriatric Depression Scale (GDS) (0-15)	1.05 ± 1.35	1.43 ± 1.51	0.89 ± 1.31	0.002
Comorbidity index (0-8)	2.88 ± 1.25	3.04 ± 1.13	2.74 ± 1.31	0.071
Fallen in the past year (Yes)	70 (30 %)	35 (36 %)	35 (26 %)	0.108
Modified Gait Efficacy Scale (0-100)	85 ± 13	79 ± 13	89 ± 11	<.001
Life-Space Assessment Score (0-120)	74.7 ± 18.6	72.6 ± 18.8	86.6 ± 18.7	0.111

3.1. Mean differences in gait for individuals with and without FOF

Of the spatiotemporal gait variables, gait speed was slower, step-time variability (step time CoV) greater and the walk ratio lower in persons with compared to without FOF (Table 2).

From the measures derived from accelerometery during the 6-minute walk, the variables that differed between those with and without FOF were: one statistical measure, standard deviation of accelerations V was less and one frequency measure, harmonic ratio V was lower in the presence of FOF (p<.01). No information-theoretic variable or time-frequency measures were associated with FOF.

3.2. Modeling fear of falling: gait variable selection

For use in pre-filtering and selection of variables for machine learning modeling, the bivariate correlations of the gait variables compared between groups in Table 2 is shown in Fig. 1a.

Fig. 1.

Variable selection via correlation analyses and random forest variable importance chart (A) heatmap showing spearman correlations between gait variables Red =Positive correlations Blue = Negative correlations, Darker represents greater coefficient (B) Gini-index based variable importance bar plot obtained via random forest analyses

The only correlations to exceed the inclusion threshold of ρ>0.6 were the standard deviation in the ML with standard deviation in the V (ρ = 0.60) and AP (ρ= 0.76) directions. The standard deviation of V was selected for inclusion in modeling because from Mann Whitney analysis (Table 2), the p-value corresponding to standard deviation of V was lowest i.e., having more discriminatory power compared to ML and AP directions. Harmonic ratio in the AP and V directions were also strongly cross-correlated (ρ = 0.61) and preference was given to the V direction (lower Mann Whitney p value) while AP was excluded. After prefiltering and selection, gait speed, step-time covariation, step-length covariation, walk ratio, standard deviation V, cross-correlation ML-V, ML-AP, and AP-V, bandwidth ML, centroid ML, harmonic ratio ML and V, and wavelet entropy V and AP were included in the random forest analysis. The random forest model (Fig. 1b) indicated that harmonic ratio V, gait speed, bandwidth ML, walk ratio, and standard deviation V were the five consistently most important gait variables for predicting the presence of FOF, respectively, with AUC of 0.63.

The Mann Whitney indicated that gait speed, step time CoV, walk ratio, standard deviation V, and harmonic ratio V were five variables indicating significant difference between FOF and Not FOF, p<.01 (Table 2). Thus, from Mann Whitney and Random Forest analyses, four overlapping significant gait variables included gait speed, walk ratio, standard deviation V, and harmonic ratio V. These variables were therefore used in the creation of the SVM model to predict FOF.

3.3. Modeling Fear of falling: Support Vector Machine (SVM) model

The SVM model was used to recognize the presence of FOF based on the previously identified four significant variables. An ROC curve (Fig. 2) was generated to illustrate predictive ability of the model for the dichotomous FOF.

Fig. 2.

Receiver Operating Curve for the Support Vector Machine model to classify individuals with and without fear of falling

The training set was able to discriminate those with and without FOF with moderate ability with an AUC of 0.65. For the test set, the model showed an AUC of 0.67. The coefficient weights in increasing order were −0.19 (gait speed), −0.23 (standard deviation V), −0.36 (walk ratio), and −0.38 (harmonic ratio V). The training set resulted in a specificity of 0.59 and a sensitivity of 0.59. The test model resulted in a specificity of 0.64 and sensitivity of 0.65.

4. Discussion

Fear of falling (FOF) was associated with gait measures from the instrumented walkway (gait speed and walk ratio) and accelerometry of the trunk (standard deviation and harmonic ratio, both in the vertical direction), which is consistent with results reported in previous research [8–11,15,16,42]. Our findings indicate that disrupted gait characteristics such as walk-ratio and harmonic ratio may be considered as signature patterns of FOF; which may be a potentially reversible psychogenic gait disorder [43,44].

The slower walking speed of individuals with fear of falling compared to those without may reflect cautious gait due to concerns with falling. Changes in gait variables besides walking speed, that is, smaller walk ratio, reduced smoothness, and reduced adaptability could be due to anxiety or some other underlying co-morbidity. A smaller walk-ratio was found in individuals with FOF; smaller walk-ratios have been suggested to be indicative of poor balance control or impaired central control of gait [45,46]. A recent study in fact revealed that smaller walk-ratio to be independently associated with falls, especially in older adults with no deterioration of gait speed [47]. This emphasizes the importance of ‘how’ an older person walks in addition to the speed of walking. Rota and colleagues [32] found the walk ratio to distinguish those with Multiple Sclerosis (MS) from heathy age-matched persons without MS, all who walked at the same speed, indicating walk ratio to be an index of neuromotor control. The walk ratio represents a unique coupling of step length with stepping rate. For healthy adults walking, this coupling of length and rate persists across a range of walking speeds. Our findings illustrate that in the presence of FOF, disordered timing and coordination of walking occurs. In our study, harmonic ratio in the vertical direction was found to be the most discriminatory gait variable for identifying individuals with fear of falling with statistical as well as machine-learning based analysis. Poorer harmonic ratios reflect less smooth walking and have been associated with an increased fall-risk [48,49]. Harmonic ratios have been associated with presence of Parkinson’s disease [50] and functional disability [51].

These gait features associated with FOF illustrate alterations in the timing, coordination, consistency, and flow of walking. Similar effects on timing and consistency in performance of well-learned, nearly automatic human motor skills, such as speech and swallowing have been noted in association with psychological stress, anxiety and depression [52–55]. Disruptions in time domain features such as number of pauses and duration of speech [52,53], and spectral domain features such as fundamental frequency in speech [53,54] have shown to be associated with psychosocial conditions such as stress, anxiety and depression. As suggested for speech timing disruptions associated with psychological disorders, these disrupted gait characteristics may prove to be useful as a performance indicator of underlying anxiety and depression [54]. Machine learning methods can be applied to build generalizable robust gait features based models, that can perform diagnosis of FOF as well as prognosis of an intervention to treat the FOF. Machine learning can analyze large amounts of data and turn that information into functional tools that can assist both doctors and patients. [56]. Moreover, machine learning could be helpful in decision making around interventions, particularly those involving a personalized medicine approach.

FOF is a complex and multi-factorial problem. Age-related neurological changes, psychological health, and environmental characteristics may contribute to FOF and gait changes independently. Thus, functional decline, restricted mobility, depressive symptoms, and decreased quality of life are all closely related to FOF and the causal relationship between them needs further assessment [57,58]. It should be noted that the risk of falling is prevalent both indoors and outdoors and sensor technology, like the one used in our study, can be used in both settings. Examining gait characteristics derived in laboratory-controlled settings in relation to health characteristics such as FOF is useful. Such analyses can also help in understanding the range of walking signal responses that individuals display in everyday life.

4.1. Public health implications

Gait assessment in clinical settings is a growing area of research. Our research emphasizes the need for monitoring fear of falling as an outcome, with gait metrics indicative of overall health of an individual beside their balance and mobility aspects. There is still social stigmatization with FOF in older adults, which makes them less mobile, further leading to muscle loss, balance issues, and increased fall-risk. Gait variables in different directions may be associated with different health and mobility related outcomes, which is informative and can uniquely guide interventions. For example, in this analysis of fear of falling, vertical direction was found to be of particular significance. Physical therapy and exercise intervention strategies targeted at gait dimensions particularly speed, walk-ratio, smoothness in vertical direction, and acceleration adaptability in vertical direction may help in overcoming FOF. For example, cognitive physical therapy improved gait speed during 6-minute walk in older adults that had gait problems as a consequence of the FOF [59]. It is important to assess whether FOF changes gait in ways that make people less stable or whether gait changes make people more likely to develop FOF. Longitudinal studies could confirm the temporal direction of these associations. Moreover, it would be beneficial to also investigate whether FOF-related gait patterns also occur in disease cohorts such as Parkinson’s. Among persons with diseases known to disrupt the timing and coordination of walking (e.g. Parkinson’s Disease, Multiple Sclerosis), the FOF-related gait patterns described could be recognized amidst the disordered walking. This could help in examining whether FOF has a unique recognizable gait pattern or it is one of several neurological behavioral influences on the well-learned motor skill of walking.

4.2. Limitations and future work

We used a yes/no i.e., binary outcome of self-reported fear of falling. A fall efficacy scale, with continuous scoring format may give more details to understand FOF. We obtained moderate sensitivity and specificity scores (of the order of ~0.60) during machine-learning related analyses; this may be due to this less detailed representation of FOF, i.e., a binary Yes/No response. More categories emphasizing degree of FOF could be used. From our cross-sectional study design, no insight as to causality of FOF can be made. It is also important to account for the context in which one is fearful of falling as well as having any fall history [60]. In our sample, fall history in the past year was not found to be associated with FOF, and we had limited sample size to perform stratified analyses of individuals with and without FOF and having and not having a history of falling. Another aspect for future studies is to assess the use of walking aids and its impact on relationship among FOF, falls, and gait; this study was done with participants who were ambulatory without an assistive device. We measured gait in the laboratory and future studies should evaluate these measures in natural environments. For example, gait assessed for dynamic balance control under complex conditions such as dual-task, mimicking free-living conditions may reveal stronger associations with FOF [61,62].

5. Conclusion

Older adults with fear of falling were found to have a poorer gait quality compared with those without fear of falling – slower walking speed, lower smoothness, lower walk-ratio, and a lower acceleration adaptability. To conclude, future studies are warranted to examine the causality of relationship between fear of falling and gait characteristics, and to explore underlying mechanisms relating physiological signals such as gait with psychosocial conditions such as anxiety. Perhaps a specific combination of the gait metrics could prove to be a ‘signature’ of fear of falling and of psychological distress. Machine learning methods need further exploration and can prove to be useful in the context of gait assessment and fear of falling.