Effects of Gradual Spatial and Temporal Cues Provided by Synchronized Walking Avatar on Elderly Gait

Featured Application

The applications of this study are significant for elderly gait training and potential gait rehabilitation. AR technology can provide real-time feedback and guided exercises to help improve balance and walking patterns, reducing fall risk among the elderly. Additionally, AR-based rehabilitation can aid in post-stroke recovery and injury rehabilitation by offering personalized feedback and adaptive exercises to enhance gait stability and coordination.

Abstract

Aging often leads to elderly gait characterized by slower speeds, shorter strides, and increased cycle; improving gait can significantly enhance the quality of life. Early gait training can help reduce gait impairment later on. Augmented reality (AR) technologies have shown promise in gait training, providing real-time feedback and guided exercises to improve walking patterns and gait parameters. The aim of this study was to observe the effects of gradual spatial and temporal cues provided by a synchronized walking avatar on the gait of elderly participants. This experiment involved 19 participants aged over 70 years, who walked while interacting with a synchronized walking avatar that provided audiovisual spatial and temporal cues. Spatial cueing and temporal cueing were provided through distance changes and phase difference changes, respectively. The WalkMate AR system was used to synchronize the avatar’s walking cycle with the participants’, delivering auditory cues matched to foot contacts. This study assessed the immediate and carry-over effects of changes in distance and phase difference on stride length, cycle time, and gait speed. The results indicate that gradual spatial and temporal cueing significantly influences elderly gait parameters, with potential applications in gait rehabilitation and training.

1. Introduction

As people age, their gait experiences some level of impairment [1]. Elderly gait is commonly characterized by slower speeds, shorter strides, and increased double-support stance phase [2,3,4,5,6]. Research indicates that individuals aged 70 and above often exhibit these pronounced gait changes [7,8,9]. These disturbances to gait can often lead to a reduction in quality of life and limit mobility [10]. Additionally, the risk of gait disorders increase with age, and neurological gait disorders such as Parkinson’s Disease become more common [3,11,12]. With the world experiencing an aging population problem, the number of persons over 65 is expected to increase. High habitual activity levels and early interventions can help improve gait and reduce gait impairment associated with age [12,13].

In recent years, elderly gait training using augmented reality (AR) has shown very promising results. Virtual reality (VR) and AR when applied to gait training can lead to significant improvements in balance, gait speed, and overall mobility among older adults [14,15,16,17]. Moreover, research also suggests that holographic displays and AR can further aid in gait coordination in the elderly [16,18]. The application of AR with real-time feedback has also been explored for its effects on gait function in gait-impaired individuals, highlighting the versatility of AR technologies in addressing various mobility challenges [19].

In addition to these advancements, the integration of avatars in AR for gait training presents a novel approach to enhancing rehabilitation outcomes [20,21,22]. By using a synchronized walking avatar, users can receive spatial and temporal cues that mimic natural human movement, thereby promoting better gait patterns and reducing the risk of falls. Booth et al. (2019) emphasized the validity and usability of an avatar-based biofeedback system for gait training, indicating the feasibility of using avatars to enhance gait training with minimal setup time [23]. This underscores the practicality of incorporating avatar-based biofeedback in gait rehabilitation programs to improve gait performance efficiently [23]. Additionally, Alyami and Nessler (2021) highlighted the advantages of intentionally synchronizing gait with an avatar in gait training [24]. They suggested that this method can help individuals adjust to gait asymmetries and enhance motor function. By using synchronized avatars, gait training can be customized to meet individual needs, thereby improving the effectiveness of rehabilitation programs.

In our previous study [25], we explored the effects of a synchronized walking avatar on the gait of young healthy participants. The study used instantaneous distance changes between the participant and avatar as a visual spatial cue to the participants. The results showed that instantaneous increases and decreases in the distance led to similar increases and decreases, respectively, in the stride length of the user irrespective of the changes in phase difference. Simultaneously, the avatar and participant phase difference changes were provided as temporal cues. Similarly, the increases and decreases in the phase difference had proportional changes in the cycle time of the participant. The study showed that despite simultaneous spatial and temporal cueing, spatial and temporal gait parameters could be targeted largely independently. However, the efficacy of this approach has not been evaluated on elderly participants, and cognitive and physical differences in age could alter the effects [26,27]. Additionally, while the previous method for human-phase estimation was sufficient for studies involving healthy young individuals with minimal gait variability, it may not be adequate for elderly gait. A more robust phase estimation and synchronization module may be necessary to effectively handle the synchronization between the avatar and elderly users.

The purpose of this study is to evaluate the immediate and carry-over effects of spatial and temporal cues provided by a synchronized walking avatar on elderly gait. Spatial cues will be provided through distance changes between the participant and the avatar. Temporal cues will be provided through participant and avatar phase difference changes. Additionally, we will look at the immediate effect on gait speed, which was not explored in the previous experiment. In addition to observing the immediate effect, a sub-goal will be to observe any carry-over learning effect by observing the gait without the avatar before and after cueing. This would provide information as to the efficacy of the system as a potential gait training tool. To help improve the carry-over effect, we will use gradual changes in the distance and phase difference instead of the instantaneous ones used in the previous study. The gradual changes will give the participants time to get accustomed to the changes and settle in a comfortable gait. To overcome the limitations of the previous avatar system [25], a new approach to the synchronization module will also be implemented to allow for more accurate phase estimation of elderly gait. To summarize, our contributions include developing a new avatar synchronization method and exploring both the immediate and carry-over effects of this synchronized avatar system on elderly individuals, a population that has not been previously evaluated in this context.

2. Methods

2.1. Subjects

A total of 19 participants were recruited from a Human Resources Center in Japan according to the selection criteria. The summary of the participants’ data is shown in

Table 1. Participant information.

Category	Details
Number of participants	19
Sex	13 males, 6 females
Age (years)	$74.16\pm 2.90$
Height (m)	$1.64\pm 0.06$
Weight (kg)	$61.15\pm 8.26$
Inclusion criteria	(1) Over 70 years old (2) Can walk 200 m without assistance (3) No uncorrected visual, auditory, or other impairment that could have affected avatar perception or walking

. The criteria for inclusion were (1) over 70 years old, (2) could walk 200 m with no assistance or difficulty, and (3) had no uncorrected visual, auditory, or other impairment that could have affected their perception of the avatar and their walking. A total of 13 males and 6 females aged $74.16\pm 2.90$ years met these criteria and participated in the experiment. Their heights averaged $1.64\pm 0.06$ m and their weights $61.15\pm 8.26$ kg. The participants self-reported no visual, auditory, or locomotive impairment that could have affected the results. The experiment was conducted with the approval of the Research Ethics Review Committee of the Tokyo Institute of Technology, and written informed consent was obtained from all participants.

2.2. WalkMate AR System

The system uses a synchronized walking avatar and audio cues in AR to deliver spatial and temporal cues to human participants for gait guidance. The flow of information was implemented as shown in Figure 1. The phase and frequency of the avatar was first synchronized to the gait of the participant, and auditory cues matching the foot contacts were delivered. The system features 3 modules that handle real-time human-phase estimation, gait phase synchronization, and cue presentation. The first two modules were used to synchronize the avatar and participant gait cycles.

2.2.1. Real-Time Human-Phase Estimation

To provide more accurate human-phase estimation for elderly gait, which has higher variability than young healthy gait, a new real-time phase estimation module was developed. Head acceleration from the IMU in the HMD is used as input to the module.

By using the HMD’s IMU, we can estimate the neck trajectory, as shown in Figure 2. In this study, the neck position was estimated from the IMU sensor of the HoloLens 2 on the head, and the foot contact timing was detected based on the trajectory of this neck position. The center of rotation of the neck, $p_{NECK}$, was calculated by using the following equation:

$$p_{NECK}=p_{CAM}+R_{CAM}\times b_{NECK}$$

(1)

Specifically, the center of rotation of the neck ($p_{NECK}$) was calculated from the camera position ($p_{CAM}$) and the camera orientation ($R_{CAM}$). Note that the distance from the head to the neck ($b_{NECK}$) is a constant.

From the trajectory, we can see that foot contact occurs at local minima. If we consider the vertical displacement, Y, and calculate the derivative, $\dot{Y}$, such that

$$\dot{Y}={\displaystyle \frac{\Delta Y}{\Delta t}},$$

(2)

we can detect foot contact when Y is at a minimum and $\dot{Y}$ is 0. To estimate the phase, we first normalize Y and $\dot{Y}$ for both to be on the same scale [−1, 1] by using

$$X_{NORM}=-1+2{\displaystyle \frac{X_i-X_{MIN}}{X_{MAX}-X_{MIN}}},$$

where X represents the parameter we are normalizing.

By plotting the normalized values of Y and $\dot{Y}$ (as seen in Figure 2) we can see foot contact when Y is at minimum, −1, and $\dot{Y}$ is 0.

The stride phase, ${\theta }_Y$, can then be calculated by using the arctan of Y and $\dot{Y}$.

$${\theta }_{y_i}=-{tan}^{-1}{\displaystyle \frac{{\dot{Y}}_i}{Y_i}}$$

(3)

The value is negated to change the direction of the phase change to match the head trajectory vertical displacement, which first decreases when taking a step.

Finally, the human phase, ${\theta }_h$, can then be calculated by using the following equation:

$${\theta }_{h_i}={\theta }_{h_{i-1}}+{\displaystyle \frac{\Delta {\theta }_y}{2}}mod2\pi$$

(4)

Note that $\Delta {\theta }_y$ is halved to have the values go from 0 to $2\pi$ instead of $4\pi$, which would be required for the 2 foot contacts in one stride.

Estimations and foot contact timings were validated by using hip- and ankle-attached IMUs (WALK-MATE Viewer®; WALK-MATE LAB., Tokyo, Japan) [28].

2.2.2. Gait Phase Synchronization

The WalkMate model [29] was used to synchronize the avatar’s walking cycle, ${\theta }_m$, to that of the participant, ${\theta }_h$. The module features two sub-modules for synchronization and phase difference control.

The first sub-module obtains the system phase based on the following equation:

$${\dot{\theta }}_m={\omega }_m+K_{m}sin({\theta }_h-{\theta }_m),$$

(5)

where ${\theta }_m$ represents the avatar walking rhythm phase and ${\omega }_m$ denotes its natural frequency. $K_m$ refers to the coupling constant. ${\theta }_m$ can then be closely matched to the human phase, ${\theta }_h$, to achieve synchrony.

The second sub-module is used for phase difference control:

$${\dot{\omega }}_m=-\mu sin(\Delta {\theta }_d-\Delta {\theta }_m).$$

(6)

where $\mu$ is the control gain of the module. Phase difference changes between the participant and avatar/system are used to provide temporal cues [25,30].

2.2.3. Avatar and Auditory Cue Presentation

The avatar phase is set to match the system phase provided by the WalkMate model; then, the synchronized avatar is presented as shown in Figure 3. Gradual increases and decreases in the distance between the avatar and participant were used to provide spatial cueing to the participant. The default starting distance was set at 5 m, as shown in Figure 3Middle. The final distance after the Spatial Decrease condition was 2 m in front of the user, as shown in Figure 3Left. Finally, the avatar was presented at 8 m after the increase of 3 m in the Spatial Increase condition, as shown in Figure 3Right.

Audio cues were also provided, synchronized to the foot contact timing of the avatar. By increasing or decreasing the phase difference between the participant and the avatar, temporal cueing guidance was provided. Phase difference changes between the user and avatar are reflected by advances or delays to the steps of the avatar (and auditory cues). Based on previous studies, we assumed that changes in the stride length are directly proportional to changes in the distance and that changes in the cycle time of the user are proportional to changes in the phase difference.

The avatar model and motion were created by using Mixamo (Adobe). The model and motion data were imported into Unity 3D, where the WalkMate model was implemented and used to control the synchronization with the user.

2.3. Experimental Protocol

Participants were instructed to walk 1.5 laps of a 65 m × 2.5 m corridor while wearing an head-mounted display (HMD). Each trial was split into 5 sections within 3 phases separated by turns at the ends of the corridor, as illustrated in Figure 4.

• Phase 1—Base section: No avatar or auditory cue is presented to the participant as they walk down the corridor wearing the HMD.
• Phase 2—conditioned: The avatar appears after the first turn and synchronizes its steps with the participant, playing auditory cues in sync with the foot contacts of the avatar. In the middle of this phase, there is a gradual 12 s change in either the distance or phase difference between the participant and avatar. This phase can be further split into 3 sections:

  (a)

  Before—The time between the first turn and the beginning of the gradual change.

  (b)

  Change—The gradual 12 s change in either the distance or phase difference between the participant and avatar.

  (c)

  After—The time following the end of the gradual change and the beginning of the second turn.

• Phase 3—Post section: Following the second turn, similar to the Base section, the participant walks with no avatar nor auditory cues presented while wearing the headset until the end of the corridor.

Distance changes provided the spatial cues and consisted of gradual increases or decreases of 3 m between the participant and the avatar, where the initial distance was 5 m, resulting in final distances of 8 m and 2 m, respectively. Similarly, target-phase changes provided the temporal cueing and consisted of a ${\displaystyle \frac{\pi }{4}}$-rad difference from an initial phase difference of 0 rad. Increases and decreases in the two changing parameters resulted in four distinct conditions that were tested as follows:

• Distance decrease (Spatial Decrease, $S_D$): This condition features a 3 m decrease in the distance between the avatar and the user from 5 m to 2 m over a 12 s period. The expected outcome is a decrease in the stride length of the user.
• Distance increase (Spatial Increase, $S_I$): This condition uses a gradual increase of 3 m in the distance between the avatar and the user from 5 m to 8 m over a 12 s period. The expected outcome is an increase in the stride length of the user.
• Phase difference decrease (Phase Decrease, $P_D$): This condition modifies the phase difference between the user and avatar. A gradual decrease in phase difference from 0 rad to $-{\displaystyle \frac{\pi }{4}}$ is used as the cue to the user. This difference can be seen in the avatar’s step timing being slightly advanced with respect to the user’s. The expected result is a decrease in the cycle time of the user’s gait.
• Phase difference increase (Phase Increase $P_I$): This condition modifies the phase difference between the user and the avatar. A gradual increase in phase difference from 0 rad to ${\displaystyle \frac{\pi }{4}}$ is used as the temporal cue to the user. This difference can be seen in the avatar’s step timing being slightly delayed with respect to the user’s. The expected result is an increase in the cycle time of the user’s gait.

For each participant, each condition was randomly conducted twice. The order was counterbalanced among the participants. Two practice trials were conducted at the beginning of the experiment. The participants used the system, walking with the avatar, but there was no change in the distance or phase difference. These practice trials allowed the participants to get accustomed to using the system and walking with the avatar. Following the practice trials, two straight-line (65 m) baseline trials without the HMD were conducted to record normal gait values. After two practice trials, during which the participants grew accustomed to the system, two 65m baseline measurement trials with no HMD were conducted. Finally, 8 trials for analysis were split into two sets and conducted. Each set contained all four conditions. The first set order was randomized, and the second was counterbalanced to reduce the effect of the trial order. Following each trial, participants were given a 2 to 3 min break to rest and prepare for the next trial. After every four trials, a 5 min break was given. The total experiment lasted 2 h, including practice trials with a combined total of 40 min rest across all trials.

2.4. Hardware

The WalkMate AR system utilized four hardware components divided into two sub-systems: gait measurement and actuating sub-systems.

The gait measurement and tracking was performed by using hip- and ankle-attached IMUs (WALK-MATE Viewer®; WALK-MATE LAB., Tokyo, Japan) [28]. The recorded ankle accelerometer and gyroscopic readings were used to estimate the gait parameters for evaluation and analysis [31]. The real-time monitoring of the gait parameters was facilitated by the controller, an Android Tablet (LAVIE Tab T11; NEC, Tokyo, Japan), running the Walk-Mate Viewer Pro application. The controller was actively being monitored and carried by the experimenter to ensure the safety of the participants.

The actuating sub-system included the AR HMD (Microsoft HoloLens 2 Mixed Reality HMD with WiFi-5; Microsoft, Redmond, WA, USA) and a laptop controller (Latitude 7280; Dell Inc., Round Rock, TX, USA) connected via TCP server running on the HMD and client running on the laptop. The HMD, weighing 566 g, featured its own built-in IMU and spatial surround sound speakers. The laptop was running a custom TCP client and an interface using the Flask library written in Python to send commands to the HMD for changing conditions and starting and ending the trials. The complete information flow presented in Figure 1 was run on the HMD, and commands for cueing conditions were sent via TCP by the experimenter.

2.5. Data Analysis

By using a method proposed by Mao et al., the ankle accelerometer and gyroscopic readings were used to estimate the gait parameters (stride length, cycle time, and gait speed) for evaluation and analysis [31]. Each trial was split into sections, as shown in Figure 4. The gait readings (stride length, cycle time, and speed) for each trial were split into these sections (Base, Before, After, and Post) and grouped by condition (SI, SD, PI, and PD). The sections represent the average of the values, as shown in Figure 4, and the descriptions are as follows:

• Base refers to the average of the values in the first phase of the trial before the first turn and with no avatar showing. Ten strides at the start and end of the phase were omitted to eliminate acceleration and deceleration readings. This was used as the baseline and control value for comparison for the learning effect.
• Before refers to the average values of 10 strides after the first turn but before the gradual cue. The 10 strides immediately after the turn were omitted not to include the acceleration values.
• After refers to the average values of 10 strides before the second turn but after the gradual cue. The last 10 strides before the turn were omitted to eliminate the deceleration readings.
• Post refers to the average of the values in the final section of the trial after the second turn and with no avatar showing. The 10 strides at the start and end of the phase were omitted to eliminate the acceleration and deceleration readings.

The gradual change lasting 12 s in either the distance or phase difference between the participant and the avatar occurred approximately 20 m after the first turn for each of the four conditions. The first 10 and last 10 strides in each trial were removed from analysis, as they featured unstable and high variability gait readings caused by the acceleration and deceleration at the start and end of the trial.

Normalized values were applied to decrease the impact of participant height on the trends [32] and were calculated by using the following equation:

$${Norm}_X=X_i/(h_i/\overline{H}),$$

(7)

where ${Norm}_X$ represents the normalized value of X, which represents cycle time, speed, or stride length. $X_i$ represents the average value for participant i, $h_i$ represents the height of participant i, and $\overline{H}$ represents the average height of all participants. After the averages were calculated for each trial, they were grouped into the four conditions and averaged to provide their respective values for comparison. Comparison between the Before and After values were used to reveal the immediate effect of the cues. Additionally, the comparison between the Base and Post values shows the carry-over or learned effect of the cue.

Change ratio values were used to observe the differences in changes between the conditions. Change ratio values were calculated by using the following equation:

$$ChangeRatio={\displaystyle \frac{\mathit{After}\mathit{Value}-\mathit{Before}\mathit{Value}}{\mathit{Before}\mathit{Value}}}$$

(8)

The change ratios were used for comparisons between the Base and Post sections (Base–Post) and comparisons between the Before and After sections (Before–After).

Repeated measures analyses of variance (ANOVAs) were conducted by using a significance factor of 0.05. Python and its statsmodels library were employed to analyze differences before and after the changes. Two independent factors were tested: conditions (SI, SD, PI, and PD) and sections (Base, Before, After, and Post). Multiple comparison analyses were conducted by using the Benjamini–Hochberg correction method with a significance level of 0.05.

3. Results

3.1. Normalized Gait Values before and after Cueing

3.1.1. Stride Length

Figure 5 shows the normalized stride length values for each condition. The repeated measures ANOVAs revealed a significant effect of the condition ($F=27.4; p<0.0001$) and an additional interaction effect of the sections ($F=5.5739; p<0.0001$).

The SI condition showed the expected increase in the stride length following the gradual increase in the distance between the participant and the avatar. The multiple comparison analysis with the Benjamini–Hochberg procedure showed significant differences between the Base and Post values ($t\left(36\right)=4.51; p=0.0012$) and the Before and After values ($t\left(36\right)=4.67; p=0.0009$). Notably, all pairs of sections revealed significant differences except the Base and Before sections. The After value taken after the change was significantly larger than other sections in the SI condition, indicating an immediate learning effect.

The SD condition showed a significant decrease in the stride length in the After section compared with the Before section ($t\left(36\right)=6.36; p=0.0008$) in the stride length following the gradual decrease in the distance between the participant and the avatar. The reading immediately after the change was significantly different from all other sections, revealing an immediate effect. There was also a noticeable but non-significant decrease in the Post versus Base values ($t\left(36\right)=1.43; p=0.25$).

The temporal conditions, PI and PD, both showed similar increases in the Post values compared with the Base values. However, the PI condition showed an additional significant increase between the Before and After cue values ($t\left(36\right)=2.47; p=0.047$).

3.1.2. Cycle Time

The normalized cycle time values for each condition are shown in Figure 6. The repeated measures ANOVA revealed a significant effect of the conditions ($F=12.3$; $p<0.0001$) and an additional interaction effect of the sections ($F=17.1; p<0.0001$).

The PD condition was the only condition to show significant decreases in the After values compared with the Before values ($t\left(36\right)=6.56; p=0.0002$). However, there were insignificant decreases between the Base and Post values ($t\left(36\right)=0.62; p=0.609$). Though the PI condition showed significant increases in the Post values compared with the Base values ($t\left(36\right)=3.55; p=0.009$), it also showed insignificant increases in the After values compared with the Before values ($t\left(36\right)=1.95; p=0.09$).

Surprisingly, the SD condition showed significant increases in both the Base–Post ($t\left(36\right)=5.42; p<0.001$) and the Before–After ($t\left(36\right)=5.81; p<0.001$) comparisons. Additionally, all conditions showed increases in the Before values compared with the Base values, with significant differences in the PI ($t\left(36\right)=3.44; p=0.01$) and SD ($t\left(36\right)=4.64; p=0.001$) conditions. Both PI and SD shared similar changes, as well as higher variability, in the After section.

The SI condition did not show any significant differences among the sections. However, it did show a noticeable decrease in the After vs. Before sections ($t\left(36\right)=2.03; p=0.08$). There was no noticeable change in the Post–Base comparison.

3.1.3. Speed

In Figure 7, we see the normalized gait speed values for each condition. The SI ($t\left(36\right)=3.69; p=0.0056$) and PD ($t\left(36\right)=8.1; p<0.0001$) conditions showed similar significant increases in gait speed between the Before and After values. However, only the SI condition showed significant increases in the Base–Post comparison ($t\left(36\right)=2.99$; $p=0.018$).

The SD condition showed significant decreases in both the Base–Post ($t\left(36\right)=4.16$; p = 0.0024) and Before–After ($t\left(36\right)=6.34; p<0.0001$) comparisons. Interestingly, the PI condition did not share similar decreases, but only a significant decrease between the Base and Before values.

3.2. Change Ratio

To observe the difference in both immediate (Before–After comparison) and carry-over (Base–Post comparison) effects amongst the conditions, we look at the change ratio. The change ratio values were calculated by using the normalized gait values and Equation (8).

The Base–Post comparison looks at the change in values from the Base section to the Post section. In these two sections, though the HMD is being worn, no avatar is shown. The Before–After comparison looks at the change in values from immediately before the gradual change in distance or phase difference to the values after the change but before the second turn. In both of these sections, the avatar is shown; however, the Before section has the initial phase difference (0 rad) and distance (5 m), while the After shows the avatar with the updated distance or phase difference depending on the condition.

3.2.1. Stride Length

In Figure 8, we see the change values for stride length for the Base–Post and Before–After comparisons. The repeated measures ANOVA revealed a significant effect of the condition on the value ($F=26.2; p<0.0001$). Multiple comparison testing with Benjamini–Hochberg correction revealed significant differences between all pairs of conditions, except SI and PI ($t\left(74\right)=1.85; p=0.08$) and PI and PD ($t\left(74\right)=1.46; p=0.15$). Additionally, the change values for Base–Post were lower in magnitude than those for Before–After, which indicates a weaker carry-over effect.

As expected, the SI condition showed increases in both the Base–Post and Before–After change values, 6.2% and 2.6%, respectively. Additionally, there was a significant difference between the two comparisons in the SI condition ($t\left(36\right)=3.47; p=0.006$). The PI and PD conditions both had similar increases for both Base–Post and Before–After change values. However, they were not as large and had a larger standard deviation. Additionally, they had no significant differences between the two comparisons.

The SD condition showed the expected decreases in both the Base–Post and Before–After change values. This is the only condition with decreases in the change values. There was also a notable significant difference between the Base–Post and Before–After change values ($t\left(36\right)=5.36; p=0.0003$). The SD Before–After values had the largest difference amongst the conditions, a decrease of 12.6%, while the SD Base–Post values had the smallest one, with a decrease of only 0.8%.

3.2.2. Cycle Time

In Figure 9, the change ratio values of cycle time for the Base–Post and Before–After comparisons can be seen. The repeated measures ANOVA revealed a significant effect of the condition on the value ($F=19.82; p<0.0001$). Notably, only the SI and PD pair ($t\left(74\right)=1.9; p=0.6$) did not show significant differences in the multiple comparison analysis.

Additionally, only the PI condition did not have any significant differences between the Base–Post and Before–After values. However, as expected, the PI condition showed increases in both the Base–Post and Before–After change values, 2.7% and 5%, respectively.

The SD condition showed similar trends to the PI condition, with increases in both the Base–Post and Before–After change values. An 11% increase was seen in the Before–After analysis compared with the significantly lower ($t\left(36\right)=5.23; p=0.0002$) 3% increase in the Base–Post values.

The PD and SI conditions showed very similar trends, both with decreases in the Before–After comparison and almost no change in the Base–Post comparison. In the PD condition, the changes for Base–Post and Before–After were 0.2% and 5%, respectively. The SI condition showed 0.2% and 3% decreases, respectively. Significant differences between the Before–After and Base–Post values were found, with $t\left(36\right)=7.2$ ($p<0.0001$) in the PD condition and $t\left(36\right)=2.38$ ($p=0.04$) in the SI condition.

3.2.3. Speed

Figure 10 further shows similarities between the PD and SI conditions ($t\left(74\right)=1.9$; $p=0.061$). Both showed increases in gait speed for both comparisons. Additionally, both showed significant differences between the Before–After and Base–Post values ($t\left(36\right)=7.77$ ($p<0.0001$) and $t\left(36\right)=2.92$ ($p=0.013$)), respectively. The SI condition showed a 9.7% increase in gait speed in the After section compared with the Before section. The PD condition showed a similar 8% increase in the same comparison. Both showed a significantly lower increase of approximately 2% in the Post section compared with the Base section.

The SD condition showed decreases in both comparisons. The Before–After comparison showed a 20% decrease in gait speed and a significantly lower ($t\left(36\right)=5.78; p<0.0001$) 3% decrease in the Base–Post comparison.

However, the PI condition showed almost no change in the Before–After comparison (<0.1%) and only a small decrease in the Base–Post comparison (0.8%).

4. Discussion

This study aimed to explore the effects of spatial and temporal cues through gradual changes in a synchronized walking avatar on elderly gait. Changes in the gait parameters of stride length, cycle time, and gait speed were used to observe the effects. Normalized values were used to observe the significant changes among the different sections (Base, Before, After, and Post). The differences between the Before and After sections reveal the immediate effect of the condition, and the differences between the Base and Post sections reveal the carry-over effect.

4.1. Stride Length

Similar to previous studies, we observed that the spatial cues, the SI and SD conditions, had a significant effect on the stride lengths of the participants. The significant difference between the Before and After stride length values in both the SI and SD conditions indicates an immediate main effect of the spatial cues on the stride length. The increases and decreases in distance, though gradual and uninstructed, led to similar increases and decreases in the stride length as was shown in the previous study with healthy young participants [25]. The stride length in the SD condition after the change was significantly shorter than before the change. It was the largest change in stride length amongst the conditions, potentially because the participants feared passing through the avatar and made an effort to reduce their strides even though not instructed to do so.

Additionally, though not explored in the previous study, we observed a learning carry-over effect in the SI condition with a significant increase in the Post section compared with the Base section. Though there was a slight decrease in the SD condition, the difference was not significant, which could have been due to the shorter strides being more unstable, so the participants reverted to their stable and comfortable stride length [33,34].

The temporal conditions, PI and PD, did not have the same effect strength as the spatial cues on the stride length. This is perhaps due to the temporal cues being mainly auditory and the spatial cues being mainly visual; the latter tend to have dominance in spatial guidance and are more effective in triggering gait adjustments [35,36]. The significant difference found between the Before and After sections in the PI condition could be attributed to the relationship between gait parameters; so, the increase in cycle time caused an increase in stride length [37].

4.2. Cycle Time

The temporal conditions, PI and PD, both showed the expected immediate changes in cycle time. The PI condition showed increases in the Post section compared with the Base section and in the After section compared with the Before section. Additionally, the SD condition showed similar increases in the Cycle time as the PI condition. These increases could be attributed to the relationship between gait parameters. The decrease in stride length in the SD condition could have led to an increase in the cycle time [37,38].

Both the SI and PD condition showed a decrease in cycle time after the cueing compared with before the cueing. This is consistent with the previous study [25], where the phase difference decreases led to decreases in the cycle time of the user. Although the SI and PD conditions showed the immediate effect of the cues with changes in the Before−After comparison, the carry-over learning effect was negligible. This could have been due to the turn between the After and Post sections. Previous research suggest that turning causes a reduction in spatial gait parameters but an increase in temporal parameters such as cycle time [39,40]. This can be seen in Figure 6, where all the Before values, taken following the first turn, were slightly larger than the Base values, taken before the first turn. This increase in cycle time following the turn could have negated the decrease caused by the SI and PD cues.

4.3. Gait Speed

There were two main effects seen on the gait speed in the different conditions. The first effect was shown by the SI and PD conditions. Both showed increases in the Base–Post and Before–After comparisons. However, the change in the Base–Post comparison was significantly smaller than that in the Before–After comparison. This also can be partially attributed to the turn as stated in the previous research study, which showed that gait generally slows during and following a turn [40]. Additionally, this reduced carry-over learning effect could have been due to the disappearance of the avatar following the turn. Previous research suggests that the effects of the cueing diminish when the cue is removed [41,42]. Despite these limitations, the SI condition was able to elicit significant changes in the gait in the Post section compared with the Base section. This suggests that the SI condition may be effective in providing gait training for elderly individuals to help them increase stride length and gait speed.

The second effect was found in the SD condition, with decreases in speed in both the Base–Post and Before–After comparisons. Similar to the previous conditions, the combination of the turns and disappearance of the avatar led to the significantly smaller carry-over effect. The SD condition had the largest change in speed among the four conditions. Perhaps, the cue was the most noticeable as the avatar gradually became much closer from a more public space to social or potentially personal space. The participants may have tried to slow their gait to keep the avatar from getting closer and prevent themselves from walking into the avatar. Despite significant effects with the SI condition, some participants verbally indicated that they did not notice the increased distance but were always aware when the distance was reduced.

Lastly, the PI condition did not show any effect on the gait speed. This was again the result of the relationship between the gait parameters [37]. The PI condition had an increase of approximately $4\%$ in stride length before and after the cue and an increase of approximately $5\%$ in cycle time. These similar increases in both led to almost no change in the speed. Perhaps, the PI condition would not be effective in gait training to increase speed but could be useful in shuffling gait which has short cycle times and stride lengths.

4.4. Synchronized Walking Avatar and Audio Cues on Gait Guidance

The system developed in this study demonstrated the potential to effectively guide the participants’ gait by using a combination of a synchronized walking avatar and auditory cues. The avatar, which was modeled after healthy gait patterns, provided a consistent and natural visual reference for the participants. This visual cue may have activated the mirror neuron system, a neural mechanism that allows individuals to learn and refine motor skills through observation and imitation [43,44]. By observing the avatar’s movements from behind, participants were likely able to internalize and mimic the gait pattern, thereby improving their own walking technique [45]. This is in contrast with the study by Sangani et al., which did not show significant changes in step length when viewing the avatar from behind [21]. Once participants started walking with the avatar as a dyad, the changes in distance between the two prompted a change in their stride length [46]. Despite not being instructed to maintain the distance, the participants likely modified their stride lengths to keep a constant distance from the avatar [47].

The visual cues in the form of the avatar and the activation of the mirror neuron system also showed significant effects on the speed of the participants. Other studies using rhythmical visual stimuli, such as Naik et al.’s, revealed stride length guidance but no significant changes in speed as found in this study [48]. Similarly, other studies using rhythmical visual stimuli found significant effects on stride length but no significant changes in gait speed as found in this study [49]. These effects could be attributed to reduced attentional demand with subconscious reactions through the mirror neuron system when using the avatar compared with the increased attentional demand found with other visual cues [50,51]. Although the use of avatars can reduce attentional demand, other studies using avatars, such as Kannape et al.’s, found increased cognitive demand limiting the effects when participants tried to synchronize their steps to the avatar’s [52]. In this study, by using the avatar synchronized to the user, we reduced the cognitive load for the user. Meerhoff et al. suggested that if the avatar’s movements do not closely match the user’s pace or rhythm, it may disrupt the user’s natural gait cycle and lead to increased variability and reduced effects [53]. The synchronization system developed in this study for elderly gait showed significant effects on the gait parameters. The system showed similar significant effects on the elderly as our previous study on healthy young participants [25]. This is in contrast with previous studies’ expectations that age would affect the outcome of the avatar-based method [54].

The synchronized auditory cues acted as rhythmic signals that complemented the visual feedback from the avatar, making it easier for participants to maintain a consistent pace and rhythm. Interactive and mutual synchronization between the avatar’s and participant’s rhythms allowed for temporal guidance of the participant’s gait [55]. The effectiveness of this interactive synchronization supports previous research studies, such as Muto et al. (2012) and Miyake (2009), both of which showed temporal gait guidance through interpersonal mutual synchronization between system and user [55,56]. The effectiveness of rhythmic cues can decrease if they are not properly aligned with the individual’s natural pace or if they become too cognitively demanding [57]. The mutual synchronization of the system and participant reduced the increased attentional demand, gait variability, and reduced stability shown in previous studies [58,59].

Overall, the participants verbally reported that the synchronized visual and auditory feedback made it easier to adjust their gait, particularly in maintaining a steady tempo and achieving a balanced stride. The real-time synchronized system offered personalized gait guidance by continuously adapting to each participant’s unique walking pattern. By providing real-time feedback, the system ensured that the guidance was closely aligned with the individual’s specific gait characteristics, enhancing the effectiveness of gait correction and rehabilitation. Furthermore, the study observed improvements in gait parameters such as step length, cycle time, and walking speed, indicating that the combined use of visual and auditory cues could contribute more efficiently to gait rehabilitation. These findings suggest that incorporating such a system into rehabilitation protocols may lead to better outcomes for elderly individuals undergoing gait rehabilitation. Moreover, this approach has the potential to improve the overall quality of life for elderly patients, enabling them to maintain greater independence and engage more fully in daily activities.

5. Conclusions

In conclusion, the study demonstrates that gradual spatial and temporal cues provided by a synchronized walking avatar can significantly influence the gait parameters of elderly participants. The results show that both immediate and carry-over effects can be seen, which suggests potential for gait training applications. In particular, the SI and PD conditions, or potentially their combination, can be used to improve common gait impairment in elderly gait (shorter strides, longer cycle time, and slower speed). Additionally, the PI condition can be potentially useful in shuffling gait to combat the shorter strides and short cycle times. The carry-over effects were reduced by turns following the change. Future research should evaluate the long-term impacts of the system and potential integration with other gait training or rehabilitation systems. The system’s effectiveness as a gait guidance and rehabilitation tool should be assessed in future studies by comparing a control group and a treatment group over a defined period of time. Additionally, the effect of performing turns and the avatar on the carry-over effect should be studied further.

Overall, avatar-based AR gait training offers a promising method for improving quality of life and mobility in the elderly. By leveraging advanced technology to provide personalized and engaging rehabilitation, this approach holds significant potential for widespread application in enhancing the independence and well-being of older adults.