Discriminating Between Fallers and Non-Fallers Using Kinematic Data from the Heel2Toe™ Wearable Sensor

1. Introduction

There is considerable clinical and research interest in risk factors for falls [1,2]) and fall risk assessment. Based on these risk factors, a number of fall risk algorithms have been developed, put into use, and systematicallyx summarized. [3,4,5]

The methods to identify risk factors and develop these algorithms have varied. One method has been to identify fallers, usually through self-reports of falls in the past, and estimate the extent to which current values on a set of candidate variables discriminate between fallers and non-fallers. [6,7,8,9,10] This method assumes participant recall is accurate and that the fall itself did not change the values on the variables of interest, the chicken-egg phenomenon.

A second and by far the strongest method is to assemble a representative sample of the target population, assess all at study entry, and follow participants into the future to ascertain the fall episodes for all and then link variables at study entry to fall occurrence or occurrences. A cohort study provides the most accurate estimates to develop an algorithm when there is complete follow-up of the cohort. [11,12,13,14,15]

A third method is, within a representative cohort, to identify those who fall (cases) and match these cases to one or more controls who have not fallen and look at historical data to identify factors predicting fall status, a case-control study [16,17]

The candidate variables fall into two groups: intrinsic (related to the person), and extrinsic, (related to the environment). [18,19] Within the intrinsic group, a further distinction is made between modifiable and non-modifiable variables and within the non-modifiable there are factors that pre-dispose someone to fall like medications which might alter blood pressure or vigilance and factors related to the circumstances of the fall. In a 2022 systematic review of 31 studies involving 70,868 community seniors the strongest intrinsic risk factors for falls were mainly non-modifiable factors including dementia (2.01, 95% CI: 1.41-2.86), age (1.15, 95% CI: 1.09-1.22), female gender (1.52, 95% CI: 1.27-1.81), fear of falling (2.82, 95% CI: 1.68-4.74), history of falls (3.22, 95% CI: 1.98-5.23), and vision unclear (1.56, 95% CI: 1.29-1.89). The strongest modifiable factors were depression (1.23, 95% CI: 1.10-1.37), and balance disorder (3.00, 95% CI: 2.05-4.39). [1]

Of the modifiable factors contributing to falls, the greatest interest is in those related to gait pattern as most falls occur when someone is upright and walking, [7,20,21,22,23,24] and balance and gait are amenable to intervention. [25,26,27,28]

Many features of gait, which is the manner of walking, have been implicated in falls and these are classified as temporo-spatial, kinetic, and kinematic gait parameters. Temporo-spatial gait parameters refer to measures of distance, time, or distance by time such as step and stride length, cadence, gait speed, step and stride time, time and duration of swing and stance phases of gait. [29] Kinematic parameters relate to range of motion at different joints and body segments during walking, and kinetic parameters relate to ground reaction forces generated during walking. [30,31] Kinematic and kinetic data are generated from technologically advanced systems such as motion capture systems, instrumented walkways, and force plates. [32] However, with the increasing number and sophistication of wearable sensors, it is now possible to measure these gait quality parameters outside of the laboratory without the expertise, cost, and infrastructure associated with these complex systems and then relate these data to health events, such as falls.

Several temporo-spatial variables have been related to falls including gait speed, stride length, and single leg stance [11] and these metrics are readily available from wearable sensors. Kinematic parameters are less easily obtained but the widespread availabiity of wearable sensors had made the acquistion of kinematic data feasible. [33] The kinematic parameters that have been linked to falls have been measured during the different phases of the gait cycle (heel-strike, foot-flat, push-off, swing). These include metrics related to variability in time spent in the different phases of gait, leg velocities, ankle angle velocities, and ankle angles. [34,35,36,37,38,39]

These studies, and others deriving metrics or non-linear dynamic measures from sensor outputs from different body segments, [40] provide a list of potential kinematic variables that discriminate between non fallers but do not provide an algorithm to predict an individual’s risk of falling. This is an important parameter because reducing fall risk can be a targeted outcome similarily to cardiovascular risk. [41,42,43]

Angular velocites of the ankle during the gait cycle are highly relevant metrics as they relate most closely to causes of falls, trips, slips, and mis-steps. [8,44] A new sensor, the Heel2Toe™ sensor, developed by PhysioBiometrics Inc., was developed specifically to measure ankle angular velocities during walking as human bipedal gait is characterized by a heel-to-toe pattern [45] and losing that pattern has been associated with falls. [46,47] The Heel2Toe sensor, shown in Figure 1, is one of a new generation of smart wearables that, in addition to providing rich data to accurately measure gait, also provides feedback for an optimal gait pattern, providing therapy in every step. [48]

The purpose of this study is to estimate the extent to which measures of ankle angular velocity during walking are associated with falls. This information would be used to quantify fall risk among gait vulnerable people and assess the extent to which interventions to improve gait quality would reduce this risk.

2. Materials and Methods

This is a comparative study of ankle angular velocity measures between people who have or have not experienced a fall in the past year.

2.1. Source of the Data

The data available for this study came from experimental use of the Heel2Toe sensor in a variety of settings including demonstrations and clinical research studies. Data collected during demonstrations of the sensor were not originally acquired for research purposes and therefore permission to use these data for research was granted by the Institutional Review Board of McGill University (A09-M51-23B) as no personal data was collected that could be used to identify the individual. All the clinical research studies using the sensor had ethical approval from their local research review boards and no identifying information was available to the research team associated with this project.

2.2. Sample

Only participants where it was possible to identify their age, disability status, and fall status were included. Excluded were participants with less than 10 steps during a walking session.

2.3. Measurement

Age was classified as 75 years and older or not; disability status was a binary variable based on diagnosis or use of a walking aid, and fall status (yes or no in the past year) was ascertained as part of the clinical research studies and also queried during demonstrations for people with disabilities. Data on sex was not consistently available. From the available descriptive information, 5 strata were formed. One stratum comprised people <75, without a disabilty, and a non-faller. This stratum served as a reference group. Another stratum comprised the fallers. The other three strata were comprised of non-fallers who were either older, with a disability, or both.

The Heel2Toe sensor attaches to the side of shoe and detects motion of the ankle during the gait cycle. The hardware for the Heel2Toe sensor consists of a microprocessor (MCU) running at 150 million instructions per second, an accelerometer, gyroscope, and magnetometer (collectively called an Inertial Measurement Unit (IMU)), each with 3 degrees of freedom running at 100 Hertz, or samples per second.

Metrics available from the Heel2Toe sensor for each step in a walking session are: angular velocity (AV) during heel strike, push off, and foot swing. These metrics are averaged over steps taken during each walking session and coefficient of variation (CV) is also calculated. Also available from the sensor is the time for each part of the gait cycle that can be used to calculate cadence and single leg stance.

Angular velocity of ankle rotation is extracted from the gyroscope (Z axis), clockwise movement is recorded as negative values (the more negative indicates higher velocity), movement anti-clockwise is recorded as positive values. Movement of the foot from initial heelstrike to footflat is a clockwise movement as is foot flat to push-off so both heel strike AV and push-off AV have negative values. To clear the foot during swing phase, the foot must rotate counter-clockwise and hence AV for foot clearance has a positive value.

2.4. Statistical Analysis

Each walking session comprises multiple steps, therefore, each gait metric was presented as the mean and standard deviation (SD) over all steps which was used to derive the CV over the steps of the walking session. The sample was described using the mean and the SD of the mean, and the mean and SD of the CV (presented as absolute value) as well as the range. Spearman correlations were conducted for all pairs of kinematic variables. A linear model was used to regress temporo-spatial and kinematic variable on the variable representing the 5 strata with the stratum representing non-fallers, <75 years, and without disability as the referent group

Logistic regression was used to identify the temporo-spatial and kinematic variables associated with the binary variable, fall status. A full model was tested first comprising all nine variables: three for each of AV of heel strike, push-off, and foot swing; ;three for the CV of these variables; and cadence, step time (proxy for stride length) and single leg stance. As all of these explanatory variables were continuous a natural cubic spline function with 3 degrees of freedom was used to model any potential non-linear effects.

All possible subsets analysis was used to choose the most parsimonous model. This modeling strategy yields a series of nested models which were tested using a likelihood ratio test. From the first model of nine variables all possible 8-variable models were tested against the full model and the variable with the least contribution to the model eliminated. This process was repeated with all n_var-1 models until all non-contrtibuting variables were eliminated. To this model, age category was added . Disability status was also considered but was strongly concordant with age. Once the shape of the function was known, ORs (95% CI) were produced for each part of the curve. All analyses were conducted using SAS version 9.4 and R version 4.3.1.

2.5. Power

As this study took advantage of existing data, power for a fixed sample size is the relevant estimate to consider. There were 68 fallers and 319 non-fallers, even with the conservative rule-of-thumb of 10 events per variable, this study has adequate power for testing 6 to 7 predictors. [49]

3. Results

3.1. Characteristics of the Sample

The sample comprised data on 387 participants who were assessed using the Heel2toe sensor, 68 (17.6%) of the participants self-reported falling in the past year. Of the sample, 240 (62.0%) were <75 years of age and 147 (38.0%) were 75+ years; 148 (39.2%) were classified as having a disability: 29 with Parkinson’s Disease, 12 with Multiple Sclerosis, 33 with orthopaedic conditions, 18 with soft tissue sarcoma; 53 needed a walking aid; and 2 had other disability situations. Table 1 presents temporo-spatial and kinematic characteristics of the sample according to five strata. In comparison to the reference strata (n=205) of non-fallers, <75 years, and without disability, the other samples all had less optimal values. The variables that differed between the strata and the reference sample are bolded and indicated by the inclusion of a p-value when it was <0.05. The variables that consistently distinguished all other strata from the referent strata were metrics related to AV.

Table 2 presents Spearman correlation coefficients among the 9 kinematic gait parameters tested in this study for their relationship with fall status. As these variables a part of the same construct, correlations of 0.8 or more are considered strong, and 0.5 to 0.8 are considered moderate. These are highlighted in the deepest blue. In lighter shades of blue are correlations of 0.4 and correlations of 0.3. The strongest correlations are among the measures of AV.

The results of the all-possible-subsets regression are not presented because the non-linear relationships required a model with a 3-df spline rendering the regression co-efficents uninterpretable. Three variables emerged from the regression model: Heel strike AV, age category, and cadence.

Figure 2 shows the shape of the relationship between heel strike AV on the x-axis and probability of a fall expressed as a percent on the y-axis using a logistic model with a natural cubic spline with 3 df and adjusting for cadence and age category. The dots at the bottom of the graph indicate the values of non-fallers along the x-axis, and those at the top of graph indicate the locations of the fallers. The graph clearly shows the non-linear relationship between heel strike AV and probability of falls. The shaded areas around the line show the 95% CI. As can be seen there is more confidence when values of heel strike AV are more than -450^o/sec. Also evident is that the relationship changes considerably when values of heel strike AV approach -200 ^o/sec. To illustrate, the distribution of heel strike AV was divided into two segments, -450 to -200 and -200 to -50. There was a non-significant relationship between faller status and heel strike AV with values at the lower end, the more optimal values (OR per 1 ^o/sec: 1.003; 95% CI: 0.997 – 1.009). At less optimal values between -200 and -50 ^o/sec, there was a strong association (OR per 1^o/sec: 1.026; 95% CI: 1.01 – 1.04). These values appear small because the units are per 1^o/sec. If the regression parameters are estimatead per 50^o/sec. the OR are 1.17 and 3.45 for the two segments, respectively. For the less optimal segment (-200 to -50 ^o/sec), for every 50 degrees that heel strike is less optimal, the odds of being a faller is increased by a factor of 3.45. Both models showed good predictive strength c-statistics >0.728 for people with strong heel strike and 0.768 for people with weak heel strike.

Table 3 presents the predicted probability of a fall for people with heel strike AV stronger (more negative) than 200^o/sec and for people with with heel strike AV weaker (closter to 0) than 200^o/sec, according to age category. The regression equations for each of these four strata are also presented. Fall risk decreased with stronger heel strike and increased with weaker heel strike; people aged 75 or older had a greater risk of falls at any degree of heel strike.

While cadence was associated with faller status overall, its effect when the data set was stratified by magnitude of heel strike AV, it was not significant and its effect on the predicted probability estimates was neglible. The lowest risk of falling was observed for cadences between 80 and 100 steps per minute, nadir 90 steps per minute.

Figure 3a and Figure 3b show the gait pattern of two individuals extracted from the data from the z-axis of the gyroscope of the Heel2Toe™ sensor positioned on the outside of the right shoe just under the lateral malleolus; 3a shows a person with an optimal gait pattern and a very low risk of fall and 3b shows a person with a very poor gait pattern and a high risk of falls.

4. Discussion

The results of this analysis indicate that starting the gait cycle with a strong heel strike has the potential to reduce fall risk. One reason for this is that a strong heel strike sets up the other phases of the gait, in particular optimal foot clearance. [44] Foot scuffing has been implicated in falls but the strategy to increase foot clearance is to emphasize a heel-to- gait pattern. [46,50]

AV at push off has previously been found to be associated with falls in a small sample of 163 seniors of mean age 82.6 years. [36] But few other studies have measured AV at the ankle. Some studies measured the degree of dorsi-flexion at the ankle during initial contact. One study measured AV but the sensor was placed mid shank representing leg movement rather than ankle movement. [37,39] A number of systematic reviews describe the potential utility of sensor derived kinematic data from sensors placed on the lower limb or on the lower back for the assessment of fall risk, but no current fall risk prediction algorithms have bee proposed. [40,51,52]

Our values on gait parameters are somewhat different than those reported by Beauchet at al. for a healthy sample of people 65+ years from two consortiums (n=954). Using the GAITrite® system, mean stride time for people 75+ was 1.15 seconds compared with >1.30 seconds in our samples (see Table 1). Similarly, Beauchet et al. reported single leg stance values of 0.41 seconds whereas in our sample with the value over the full sample was lower, approximately 0.31 seconds. This likely reflects that 50% of our sample had a disabling health condition or a history of falls.

This study is limited in that fall status was modeled and not fall incidence. Thus, this algorithm needs to undergo further testing. In addition, the data came from data accumulated from research projects where the sensor was used and hence the amount of data available was varied. Future studies, could also consider testing the algorithm for different walking durations or distances.

This study is limited in that fall status was modeled and not fall incidence. This our algorithm needs to undergo further testing. In addition, the data came from data accumulated from research projects where the sensor was used and hence the amount of data available was varied. The sample was also not selected to be representative of any particular segment of the population.

The results presented here open up additional avenues of scientific enquiry and implementation. The Heel2Toe™ sensor was developed primarily as a gait training device as its main functionality is to provide auditory feedback for a good step, one in which the AV at heel strike surpasses a specific threshold. This reward-feedback loop promotes neuroplasticity likely through the dopamine system. [48] While there are many effective exercise programs targeting balance [53], few target the elements needed for an optimal gait pattern: knowledge, strength and mobility of muscles and joints of the lower limb, core strength, all facets of balance (static, dynamic, anticipatory, and reactive), and coached training. [54,55] Paradoxically, indoor or outdoor walking programs have low evidence for efficacy and could even be harmful [55] and we hypothesize this negative effect could be because a sub-optimal gait pattern can increase scuffing, tripping, or mis-stepping, particularly as the person fatigues. [56] In addition, the role of sarcopenia in gait failure could be mediated through optimization of macro- and micronutrients known to be needed for muscle power and coordination. [57] Preventing falls is a complex goal and needs a multi-modal approach as each element is likely to increase the effectiveness of the other elements.

5. Conclusions

Average angular velocity at heel strike over a walking session of a minimum of 10 steps was the key kinematic variable discriminating between fallers and non-fallers and together with age (≥75 vs. <75 years) was used to propose an algorithm to estimate individual fall risk. Further validation of this risk model is needed.