The second phase of the ACIER study was designed to evaluate the effectiveness of the recently proposed ACI method in discriminating gait patterns between younger and older adults and to investigate its responsiveness to cognitive load during walking tasks compared to other linear and nonlinear gait metrics. Our findings support the primary hypothesis, demonstrating a specific sensitivity of the ACI compared to other gait metrics derived from the lumbar accelerometer, when participants walked in synchrony with a metronome compared to normal walking. In addition, the study aimed to assess the ability of ACI to characterize age-related differences in gait patterns. Consistent with the secondary hypothesis, ACI, along with step and stride regularity, demonstrated a significant ability to discriminate between older and younger adults. These findings indicate that ACI may serve as a sensitive marker for detecting changes in gait control and cognitive demands during walking, which can be useful in predicting fall risk in older individuals.
Building on the responsiveness of the ACI to metronome walking observed in older adults (Phase I [36]), this study confirms this effect in young adults (Table 2–4 and Fig. 7). In addition, we demonstrate a significant correlation between the ACI and the reference standard (scaling exponent of the stride interval time series) in young adults (see Supplementary Material). These findings are consistent with previous research showing similar responsiveness to voluntary synchronization and correlations with the reference standard (DFA) in both treadmill [27, 37, 50, 55] and overground [56] walking studies. Descriptive statistics (Table 2–3) revealed a trend for younger individuals to have smaller differences in ACI-AP (-14%) and ACI-V (-15%) between normal and metronome walking compared to older adults (-27% and − 26%, respectively). This age-related discrepancy is consistent with the observed differences in scaling exponents (Table 2–3, young: -40%, older: -47%). This finding may parallel observations in treadmill walking, where lower scaling exponents are associated with tasks requiring tighter gait synchronization and potentially higher cognitive demands [34]. Similarly, in our study, synchronizing gait with a metronome might require a greater cognitive load for older adults compared to younger participants. However, the regression analyses considering within-subject correlations (multiple regressions, mixed effect models) did not yield a significant condition-by-group interaction for either ACI-AP (p = 0.27), ACI-V (p = 0.16), or ACI-N (p = 0.06) (Supplementary Material, ANOVA F-tests). A possible explanation for this lack of interaction effect could be insufficient statistical power, as interaction terms may require four times the sample size for detection compared to main effects [57]. Future adequately powered studies specifically designed to investigate this potential age effect are warranted.
Our novel contribution lies in the specific sensitivity of the ACI to metronome walking compared to other gait variability metrics used in unsupervised settings. Specifically, walking in synchrony with an isochronous metronome tuned as the preferred SF did not affect gait regularity as assessed by ACF with the lumbar accelerometer. Similarly, other gait metrics, including RMS ratio and LDS, remained unaffected by metronome walking. However, we observed a small (+ 4%) but significant increase in movement intensity (RMS) during this condition, even after accounting for potential within-participant correlations and adjusting for age group and speed using mixed effects models (Fig. 7). Although a type I error cannot be entirely ruled out, we can consider the following explanation: it is possible that participants may have unconsciously adjusted their gait pattern by striking their heels more forcefully to synchronize with the metronome, resulting in higher peak accelerations and, as a result, a slightly higher RMS value.
Our validation study used an indoor circuit, which offered the advantage of directly measuring preferred walking speed through timed travel. Walking speed is a valuable health biomarker [58]. Age-related conditions can limit gait speed in older adults. These conditions include sarcopenia (muscle loss [59]), reduced cardiorespiratory fitness [60], increased energetic cost [61], mitochondrial dysfunction [62], joint stiffness [63] or impaired sensory feedbacks [64]. As a result, older adults typically have a slower preferred walking speed than younger adults, which was confirmed in our study. Indeed, we found that older participants had an average speed of 1.27 m/s (Table 2–3), which is consistent with the reference value of 1.30 m/s reported for highly functional older individuals (mean age 79 years) in a recent large-scale study with a comparable socio-cultural context (Germany, [65]). In contrast, the walking speed of the younger participants was 1.43 m/s, which is slightly higher than the reference value of 1.37 m/s for adults between 20 and 40 years of age calculated from the data of a meta-analysis including 23,000 individuals measured worldwide [66]. Overall, there was an 11% relative difference between young and older participants. This difference was highly significant in both descriptive statistics (standardized ES: -0.77) and in the multiple regression analysis adjusted for walking condition (regression coefficient b = -0.167 m/s, t-test p < 0.0001, Table 4 and Supplementary Material).
In our effort to establish an analytical framework for categorizing gait metrics based on their correlation with observed walking speed, we anticipated that both movement intensity (RMS of lumbar acceleration) and SF would fall into the speed-surrogate category. Unsupervised gait analysis employing a single lumbar accelerometer relies on simplified gait models (e.g., inverted pendulum) or speed-surrogate metrics, such as SF or RMS, to estimate walking speed (Table 1). Inverted pendulum models are typically derived through a double integration procedure of the trunk acceleration to estimate step lengths [67], in conjunction with anthropometric measurements. It is noteworthy that RMS captures the average amplitude of the signal, which is somewhat analogous to integrating the signal. In experimental studies involving subjects walking across a wide spectrum of speeds, a curvilinear relationship has been therefore observed between acceleration RMS and walking speed [68–70]. Our findings corroborate this relationship, demonstrating that the RMS of acceleration and preferred walking speed are positively correlated (correlation coefficients: 0.73 to 0.79, Supplementary Figures). This correlation holds true for both older and younger participants. Consequently, individuals who spontaneously walk at slower speeds also tend to exhibit lower RMS values. Our results confirm therefore that older participants exhibit lower RMS values compared to their younger counterparts. Specifically, during normal walking, the RMS values were 0.38 for older participants versus 0.46 for younger participants (absolute difference: -0.08, relative difference: -17%). Similarly, during metronome walking, the corresponding values were 0.40 and 0.47, respectively (-0.07, -15%). The regression model shows a coefficient of -0.084 for age group adjusted for walking conditions (Table 4). When the regression model is further adjusted for walking speed, the age group effect vanishes (Fig. 7), as logically expected given the strong RMS-speed relationship. However, while the RMS of trunk acceleration holds promise as a surrogate of a direct measure of walking speed, it is essential to consider its sensitivity to factors beyond velocity. Notably, RMS can be influenced by external conditions, including the nature of the walking surface [71] and the slope [72, 73]. This sensitivity to environmental variables may complicate the use of RMS as a robust proxy for walking speed in unsupervised conditions.
As a second speed-surrogate metric, SF, which is biomechanically related to walking speed as the product of cadence and step length, also showed a systematic correlation with walking speed (supplementary figures). However, the strength of this association across groups and conditions (r = 0.43 to 0.69) was weaker than that observed for RMS (r = 0.73 to 0.79). In contrast to RMS, SF did not differ significantly between young and older participants (1.89 vs. 1.90), as confirmed by the multiple regression analyses (Table 4 and Fig. 7). This finding is consistent with the expected pattern in which older adults typically have shorter step lengths compared to younger adults [2], resulting in a similar cadence despite their reduced preferred walking speed (Table 2). Although SF may serve as a surrogate for walking speed, its utility in characterizing age-related gait changes in unsupervised settings seems therefore limited. However, more research is needed to further evaluate the usefulness of SF in assessing fall risk.
To attenuate the speed sensitivity of RMS, the RMS ratio has been proposed as a metric for gait quality assessment [45]. This approach aimed to normalize the mediolateral acceleration RMS by the global RMS of the vector magnitude, potentially revealing gait instability in the frontal plane. While several studies have investigated this hypothesis, the results have been mostly inconclusive [74–76]. In our large study including 100 subjects of various ages, we found no significant change of RMS ratio across the lifespan [40]. Here, too, we observed that RMS ratio was not different between young and older participants (Table 2–4 and Fig. 7). Although the concept of RMS normalization initially appeared promising, the use of the RMS ratio to characterize age-related gait changes seems to have limited utility.
Our results show a significant difference in step and stride regularity between age groups, as measured by the ACF analysis. The ES for normal walking was large (-0.97 and − 0.91, Table 2), which was further supported by the results of the multiple regression analyses (Table 4 and Fig. 7). This is consistent with previous research highlighting similar age-related reductions in regularity [77, 78]. In particular, the previous work by Auvinet et al. [79], which characterized gait patterns across different age groups, reported a substantial difference in stride regularity (standardized ES = -1.22) between the youngest (20–29 years) and oldest (> 70 years) cohorts, although no significant change was found across the lifespan. This observed change in acceleration signal regularity with age suggests that a convergence of musculoskeletal and neurological deficits may limit the ability to maintain consistent gait patterns across strides. Note also that a correlation was observed between ACI and ACF measures in the older group specifically (supplementary material, r = 0.58 to 0.68). This supports the hypothesis that both gait regularity and ACI are concomitant signs of gait quality degradation in older adults. Regarding the categorization of step and stride regularity according to their associations with walking speed, they can be considered as mixed-metrics. Indeed, while only weak correlations were observed in the young group (r = 0.0 to 0.36, Supplementary Figures), stronger associations were observed in the older group (r = 0.42 to 0.58). Finally, ACF analysis shares an advantage with the LDS and ACI methods in that it does not rely on gait event detection to assess gait variability. This is particularly beneficial when analyzing gait in variable environments, where frequent transient events can lead to irregular steps and prevent clear gait cycle delineation.
LDS is a popular nonlinear gait variability metric that reflects gait robustness to perturbations and is considered a predictor of fall risk [15, 49]. Unlike the ACI, where higher values indicate better gait quality, higher LDS results correspond to greater instability and a degraded gait pattern [12, 27]. Therefore, it is expected that older adults, who have a higher risk of falling, will exhibit higher short-term divergent exponents compared to younger individuals. This assumption is partially supported by observational studies. However, definitive conclusions are hindered by methodological inconsistencies in measurement methods, algorithm implementations, and experimental designs [14]. Our 2015 study investigated changes in gait stability between the ages of 20 and 70 in 100 participants using the same LDS calculation as this work. Our model predicted a 13% higher LDS-ML at age 75 than at age 25. Previous studies have also reported varying relative differences between younger and older cohorts: +40% (graphical estimation, Kang & Dingwell 2009 [80]), + 10% (Bruijn et al., 2014 [81]), + 7% (Hamacher et al. 2015 [82]), + 6% (non-significant, Bizovska et al., 2018 [83]), + 5% (non-significant, Lin et al., 2015 [84]), or + 2% (non-significant, Franz et al. 2015 [31]). The observed relative difference in the present study was + 14% during normal walking and + 8% during metronome walking (Table 2–3), which is consistent with previous research. However, the null hypothesis of no difference between age groups could not be rejected due to large confidence intervals (Table 4 and Fig. 7). Further research is necessary to specifically determine the discriminatory power of the LDS in identifying individuals at risk of falls compared to those who are not.
Our findings (Table 2–3, Fig. 7) strongly support the hypothesis that the ACI can distinguish age related changes in the gait patterns. This new discovery has not direct precedent in the literature, given that ACI has only been recently proposed as a gait metrics for replacing DFA in gait analysis under free-living conditions [27, 36, 37]. Based on the mixed-effects models outcomes (Table 4), we can predict a relative change of -24% between the young and older populations. To put this in perspective, we found a contrast of -18% between healthy adults and adults suffering from chronic pain of lower limbs using ACI (still referred to as LDS-L in this 2017 study [44]) in unsupervised gait analysis. This comparison is instructive because chronic pain patients may use greater voluntary control of limb movements to adopt a less painful gait (antalgic gait [85]), possibly similar to the increased cognitive interference observed in older adults.
In the initial phase of the ACIER study, which focused exclusively on older adults, our analysis showed that the ACI measured in the frontal plane (ACI-ML) was inadequate for assessing gait correlation structure [36], while both ACI-AP and ACI-V could be used interchangeably. Additionally, we hypothesized that ACI-N, as an orientation-free metric, may be advantageous in unsupervised gait analysis scenarios where maintaining a consistent sensor position is difficult. However, adding the results from the young adult cohort leads to a more nuanced interpretation. ACI-AP showed increased sensitivity to age-related changes in gait patterns compared to ACI-V and ACI-N, as shown in Fig. 7. Furthermore, ACI-AP appears to be a speed-independent metric, whereas ACI-V and ACI-N are categorized as mixed metrics that exhibit a correlation with walking speed that is stronger in older adults (Supplementary Figures). While the pronounced sensitivity of the ACI-AP requires further investigation, it is plausible to suggest that accelerations within the sagittal plane—aligned with the displacement of the body—may contain information for the regulation, and thus for correlation structure, of stride length and duration. These insights are likely to be less pronounced for accelerations that occur perpendicular to the direction of displacement.
In contrast to the observed age-related differences in ACI, scaling exponents derived from foot acceleration data using DFA did not show significant age group differences in the mixed model analysis (p = 0.06, Supplementary Material, t-test). This lack of significance may be due to limitations of the DFA method, as highlighted in our companion article regarding its potential inaccuracy at short measurement times [36]. Although the observed standardized difference between age groups in normal walking (effect size = -0.19, Table 2) lacks statistical significance, it aligns with the findings of the above-mentioned meta-analysis (ES= -0.20, combining eight studies [13]). This suggests that ACI (ES=-0.77, Table 2) may be more effective than DFA in detecting age-related changes in gait patterns, particularly in shorter walking bouts or in unsupervised settings. However, further independent studies are needed to consolidate this observation.
Overall, our findings reinforce the notion that reliance on a single metric may not adequately capture the complexity of age-related decline in walking abilities. A comprehensive suite of both linear and non-linear gait metrics derived from acceleration signals appears to be essential to identify the various age-related adaptations in gait patterns. We propose that the ACI-AP, considered as a speed-independent metric, could reveal subtle changes in the automated control of gait—a modification not necessarily associated with reduced walking speed. At the same time, metrics of step and stride regularity, which we categorize as mixed-metric, could reveal difficulties in maintaining a consistent gait pattern from one stride to the next. This inconsistency could be caused by a reduced lower limb strength, potentially leading in parallel to a reduced preferred walking speed.
In conclusion, ACI shows a unique sensitivity to metronome walking in both young and older adults. This suggests that ACI specifically captures changes in gait control associated with increased cognitive demands during synchronized walking. Second, the ACI appears to be a valuable tool for discriminating age-related differences in gait patterns. This finding is consistent with the observed differences in step and stride regularity, further highlighting the potential of ACI as a complementary marker of gait quality decline in older populations, and thus as a tool for identifying older adults at risk for falls. In addition, due to the relative ease of measuring ACI, it could be used to evaluate practical interventions, such as in the recent clinical trial aimed at restoring gait automaticity in older adults that began in parallel with the ACIER study [86]. The broader implications of these findings go beyond fall risk assessment. The sensitivity of ACI to cognitive load during gait opens new avenues for investigating the intricate interplay between cognitive function and motor control of human locomotion, which may help to gain deeper insights into the mechanisms underlying gait disorders.
Building on these promising findings, the next phase of the ACIER study will focus on the clinical utility of the ACI for fall risk assessment. We will use a retrospective approach to differentiate between participants who have recently fallen and those who have not. This will involve the analysis of gait data collected in the first phase of the ACIER study and the comparison of ACI scores between these two groups of older individuals. Subsequently, the final phase will use a prospective approach with a longitudinal design and survival analysis. This involves following older participants over a two-year period to assess the association of ACI and other gait measures with time to first and second falls.
By addressing these future directions, the ACIER study can significantly contribute to the development of novel and reliable gait assessment tools using readily available wearable sensors, ultimately aiding in fall prevention strategies, and improving mobility monitoring in older adults.