Design and Evaluation of a Wearable Biofeedback Training System to Reduce Trip-Related Falls During Level Walking

Trip-related falls are a major concern, especially for older adults and individuals with gait impairments as they can lead to serious injuries, hospitalizations, and negatively impacting the quality of life. A low minimum toe clearance (MTC) can be a predictor of tripping risk, and thus, increasing the MTC is a possible way to reduce trip-related falls. In this article, a wearable system is proposed that can measure the MTC in real time using two time-of-flight (ToF) sensors on the shoe and provide auditory biofeedback using a piezo buzzer. Ten healthy female adults were recruited to walk in four conditions: baseline, biofeedback, short retention (same day), and long retention (next day) to design and validate our gait training tool. Average MTC values were compared pre-feedback, post-feedback, and post-retention, and our analysis revealed significant differences between the feedback and retention sessions for the training system. Therefore, the proposed system has the potential to be used as a wearable training system to minimize tripping risks in older adults and gait-impaired populations.

independence, immobilization, and depression [3]. According to a study by Blake et al. [4], over 50% of falls in older adults were caused by trips. Tripping occurs when the foot makes contact with objects or the ground unexpectedly [5]. One of the main reasons for tripping is the reduction in minimum toe clearance (MTC) during walking [6]. MTC is defined as the distance between the ground and forward-swinging foot during the mid-swing phase when the foot is at its maximum speed [7]. Older adults and individuals with gait impairments have a higher likelihood of tripping due to the low MTC that results from foot drop or shuffling gait [8], [9], [10]. Low MTC has been considered a predictor of tripping risk and studies have found that increasing the MTC can reduce the risk of tripping [11]. It has also been proven that providing real-time feedback about an individual's gait parameters is effective in changing foot trajectory control and reducing tripping probability [11].
Gait training plays a prominent role for individuals who want to improve or restore their gait pattern due to gait disorders or aging [12]. Common gait training strategies include treadmill training, muscle strengthening, neurodevelopmental techniques, and intensive mobility exercises [13]. Traditionally, gait training is conducted in a healthcare setting, such as a laboratory or clinic, by a physical therapist who will visually observe various aspects of the patient's gait and outline an appropriate intervention plan to improve the patient's gait performance over time [12], [14]. This method of gait training has different limitations such as: 1) the assessments depend on the knowledge and experience of the physical therapist, which can result in subjective assessments that may vary across multiple visits, and 2) the equipment used to conduct the gait training, such as motion capture, force plates, and treadmills, can be cumbersome and expensive, restricting the location of training to only laboratory-based settings [12], [14]. In recent years, technological advancements have emphasized the effectiveness of using wearable sensors for gait training in conjunction with conventional methods or on their own. Wearable sensors fulfill the need for a low-cost, portable, nonobtrusive, and nonrestrictive device that provides quantitative, objective, and repeatable results in both laboratory-and home-based environments [14]. The sensors are placed on the body at areas of interest to monitor an individual's body movements or vital signs [15]. The most commonly used sensors in gait analysis include accelerometers, gyroscopes, inertial measurement units (IMUs), goniometers, force sensors, and electromyography [15].
Biofeedback is becoming more widely used with wearable sensors for gait training [16]. Biofeedback is the process of delivering real-time information about an individual's physiological or biomechanical parameters [17], [18]. The biofeedback process can be broken down into five components: mode, content, frequency, timing, and amplitude [19]. Mode refers to the mechanism used to convey biofeedback to the user and can be delivered in visual, auditory, tactile, or combined formats [19], [20], [21]. Content represents the type of biological signal measured from the participant and can refer to how the movement was performed or the result of the movement [19]. Frequency addresses the number of times the measured signal event occurs and can be constant, reduced, or fading throughout the session [19]. Timing indicates when the feedback occurs in relation to the execution of the movements and can happen concurrent to the movement or after the movement has been completed [19]. Amplitudedescribes the intensity of the biofeedback signal and can happen concurrent to the execution of movement or after the movement has been completed. In addition, biofeedback can be used as a positive reinforcement to encourage a specific action to occur more often or as a negative reinforcement to decrease the number of times a certain behavior takes place [19]. Providing biofeedback allows a person to understand how their body works and helps them gain better control over specific body functions or movements to improve their health concerns [17].
The four main types of biofeedback that are used for gait training include: 1) EMG biofeedback, which focuses on specific muscle activation; 2) kinematic biofeedback, which pays attention to segment movements or joint angles during the swing phase of the gait cycle; 3) kinetic biofeedback, which concentrates on forces exerted during the stance phase of the gait cycle; and 4) spatiotemporal biofeedback, which targets gait parameters related to space and time such as step length and cadence [22]. A mapping review conducted in 2018 investigated biofeedback systems developed for gait retraining and reached the conclusion that visual feedback was the most common feedback modality, kinematic gait parameters were the most common biofeedback gait parameter, and pressure sensors fixed to the feet or feet insoles were the most common wearable sensors used [16], [23]. Researchers have investigated several biofeedback gait training strategies and modalities to determine which is the most effective and practical for their target population [17]. For instance, Byl et al. [24] used visual biofeedback with wireless pressure and motion sensors to improve the balance, strength, and flexibility in Parkinson's disease (PD) and chronic stroke patients. Kim et al. [25] used auditory biofeedback with a pneumatic pressure insole to improve the gait variables, dynamic balance, and activities of daily living for patients with stroke. Escamilla-Nunez et al. [17] developed a vibrotactile biofeedback system using vibrating motors to evaluate temporal gait symmetry and speed.
Auditory feedback is becoming more prevalent in recent years due to its discreet and portable nature. Previous studies have compared visual and auditory biofeedback for motor response and found that auditory inputs are processed faster than visual inputs [26]. It was also found that the brain gradually reduced its dependence on auditory biofeedback over time, while visual biofeedback had a sustained reliance to visual cues [27]. Due to these advantages, several researchers have been incorporating auditory biofeedback for gait training applications. Casamassima et al. [28] used real-time auditory feedback in the form of vocal messages to encourage PD patients during motor therapy. Mazilu et al. [29] developed a wearable system that plays auditory rhythmic patterns that adapt to the user's walking speed to assist PD patients when they experience freezing of gait. Khoo et al. [30] created a realtime feedback system to actively correct gait asymmetry for poststroke survivors by indicating the swing and stance time using both auditory and electrotactile feedback. The results from the previously mentioned articles all concluded that auditory feedback was successful and effective in improving the abnormal gait of the targeted populations [28], [29], [30].
To date, little research has been conducted that focuses specifically on increasing the MTC to prevent trip-related falls. A study conducted by Miyake et al. [31] proposed a gait training robot and MTC prediction algorithm that helps individuals increase their MTC by applying intermittent force through a cable-driven system. The cable-driven system was able to switch between assistive and nonassistive modes depending on if the MTC became lower than the mean and was successful in encouraging participants to increase their low MTC during and after assistance [31]. Pathak et al. [32] focused on using vibration embedded in shoe insoles to decrease MTC variability while walking. This system presented a simple, compact and portal device; however, it did not specifically study increasing the MTC height in relation to tripping risk [32]. Tirosh et al. [11] proposed a visual biofeedback system to modify the MTC of young, healthy adults while walking on a treadmill. Participants were asked to walk on a treadmill while looking at their projected toe trajectory on a real-time overhead display and attempt to maintain their MTC within their participant-specific threshold range for each step [11]. After the biofeedback condition, participants also performed a retention trial where they were instructed to walk with the learned "higher MTC pattern" without the biofeedback present [11]. The short-and the longterm MTC variability were analyzed and found to be low in the baseline study, higher in the biofeedback condition, and even greater in the retention condition [11]. In different studies, Begg et al. [33] and Nagano et al. [34] tested the same visual biofeedback system on older adults and poststroke patients to determine the system's feasibility on gait-impaired populations and confirmed that the system could successfully increase the MTC above the baseline for both populations.
Although the findings from these studies verified that the proposed devices could be effective in modifying the toe trajectory, there exist some gaps in previous literature. For example, the studies that investigated MTC in regard to tripping risk were all conducted on a treadmill [11], [33], [34]. It has been found that the biomechanics of overground walking differs from treadmill walking, and consequently, the results from these types of studies cannot be directly compared [7]. One of the differences between treadmill and overground walking is that the stance limb travels back when walking on a treadmill, while the MTC event takes place on the opposite leg [35]. In contrast, the stance limb is stationary when walking on the ground [35]. Another difference is that the stride length and stride time are shorter during treadmill walking, which has been associated with a lower MTC [36], [37]. Another limitation specifically for [31] and [11] is the use of large, cumbersome, and expensive equipment to implement the visual biofeedback system and the cable-driven force application system. Having a setup with a treadmill, overhead display, motion capture equipment, or cable-driven system would have to be conducted in a laboratory setting, limiting this system's capacity to be used for at-home rehabilitation.
In our previous study [38], we compared the proposed wearable system against the reference motion capture system and concluded that time-of-flight (ToF) sensors were able to detect the MTC with high accuracy without the use of additional hardware or complex noise removal algorithms. As a result, we wanted to extend our investigation further to evaluate whether the proposed system could be effective as a biofeedback gait training tool. While considerable research has been conducted to improve gait using biofeedback strategies, to the best of our knowledge, no study has attempted to develop an auditory biofeedback training system using ToF sensors to measure the MTC in real-time on overground walking. Therefore, the goals of our study are: 1) to design a wearable system that can perform real-time MTC detection using ToF sensors and use auditory feedback to increase the MTC and 2) to evaluate the effectiveness of the proposed system as a gait training tool. The study was performed on young, healthy adults with no gait impairments to evaluate the feasibility of the biofeedback system. The participant-specific biofeedback threshold strategy is discussed and the MTC data are compared pre-feedback, post-feedback, and post-retention to assess whether the system could effectively increase the MTC height in training conditions.

A. System Design
The biofeedback system consisted of two VL53L0X laser ToF sensors (SensorDots, Melbourne, VIC, Australia) mounted at the lateral edge of the toe and heel for the left and right shoe (two sensors per foot), as shown in Fig. 1(a). The ToF sensors provide distance measurements ranging from 30 to 1200 mm with a 25 • field of view by calculating the time it takes for a laser to reflect off surrounding surfaces and return to the sensor. The ToF sensors were used to detect key foot clearance parameters such as the MTC, maximum heel clearance (MHC), first maximum toe clearance (MX1), and second maximum toe clearance (MX2) [38]. The sensors were attached to the running shoes at least 50 mm from the ground using custom 3-D-printed mounts to ensure that the entire distance range of the sensors was measured while walking. The Teensy 3.6 microcontroller (PJRC, Sherwood, OR, USA) was used to communicate with the sensors, and the data were stored on an secure digital card (SD) at a sampling rate of 33 Hz. The system was powered by four 1.5-V alkaline batteries. A piezo buzzer (Adafruit Industries, New York, NY, USA) provided auditory biofeedback and produced beeping tones at 3 kHz. The microcontroller, buzzer, batteries, and electronic components were placed in a box attached to the participant's waist, as shown in Fig. 1(b). The box was connected to the ToF sensors on the running shoes by using separate ribbon cables that were 2.7 m long. The ribbon cables were taped to the participant's legs during testing.

B. Biofeedback Strategy
The biofeedback strategy used for the training system involved asking the participants to walk at normal speed with the metronome for ten steps to set a customized participantspecific threshold for each foot. During our pilot testing, several thresholds were tested to determine which would be most suitable to train participants who did not have any gait impairments. We found that thresholds higher than 1.25 resulted in the participant over exaggerating their foot clearance and deviating greatly from their normal gait pattern. On the contrary,   thresholds less than 1.25 were too similar to the participant's existing MTC pattern, and as a result, the biofeedback was rarely activated and the participant was not experiencing any gait training. However, it is worth mentioning that this parameter can be customized for each user according to their mobility issues and requirements. The initial ten baseline MTC values were collected and the threshold was set to 1.25 times the average baseline to train participants to raise their feet higher than their usual MTC. Auditory biofeedback was activated if the participant's MTC on either foot was less than the foot's personalized threshold. Participants were encouraged to raise their feet in future strides to increase their MTC.

C. Experimental Protocol
Ten healthy female participants (n = 10) aged 31.1 ± 12.1 years (mean ± SD), with an average height of 165.1 ± 7.5 cm and a weight of 60 ± 6.9 kg (see Table I), were recruited for the study. The participants did not have gait-related impairments, musculoskeletal conditions, or neurological disorders. All participants reported that their dominant limb was their right foot.
The study was held at the Challenging Environment Assessment Laboratory (CEAL), KITE Research Institute-Toronto Rehab (TRI)-University Health Network (UHN). The University Health Network Research Ethics Board approved the study. Informed written consent was provided by all participants before joining the study.
The walking trials took place at level ground on the path shown in Fig. 2(a). Participants wore the biofeedback system while performing the experiment tasks, as shown in Fig. 2(b). For the experiment, participants were asked to walk at the normal speed set by the metronome (90 BPM) in four conditions: baseline, biofeedback, short retention, and long retention. The baseline condition lasted for 3 min and the threshold was calculated using the baseline MTC data. For the biofeedback condition, participants were asked to walk for 15 min. Participants were given instructions on how to raise their feet, specifically to increase their toe height during the swing phase of the gait cycle. After the biofeedback condition, they were required to take a 10-min break and then asked to perform the "short retention" condition, where they walked normally for 10 min. They were also requested to return the following day and perform the "long retention" condition where they walked for 10 min in the exact same condition.

D. Data Analysis
The MTC values were detected in real time using an algorithm developed in our previous study [32]. This algorithm used the ToF data measured at the toe and heel to estimate the MTC. The ToF signal was corrected by calibrating the data using the least-square (LS) method [33] and compensating the foot angle in real time using Arami's method [34]. In this article, the data were collected from both the left foot and the right foot. The total number of steps from all participants and all conditions was 14 925 from the left foot and 14 611 from the right foot. The first and last steps were excluded for each trial.
JMP Pro Software (Statistical Discovery, SAS, Cary, NC, USA) was used to perform the statistical analysis. Nonparametric analysis of variance (ANOVA) (Kruskal-Wallis tests) was used because the MTC data were not normally distributed. A one-way repeated measures ANOVA was performed with condition (baseline, biofeedback, short retention, and long retention) as the fixed factor. The Kruskal-Wallis nonparametric multiple comparisons test was used to determine specific statistical differences between pairs of means. The statistical significance was considered when p < 0.05. Table II represents the overall improvements in the mean and median MTC values for each condition compared to the baseline data of the training system. The mean MTC increased by 199% and 328% for the BF condition, 254% and 413% for the short retention, and 274% and 439% for the long retention for the left foot and the right foot, respectively, considering all ten participants. The one-factor ANOVA analysis revealed that the mean MTC was significantly different for all four conditions ( p < 0.0001) for both feet.

III. RESULTS AND DISCUSSION
The mean MTC and SD of each participant and condition for the training system are provided in Table III. As expected, the average MTC was greater in the biofeedback and retention conditions than the baseline. It is also observed that the MTC variability, indicated by the SD values, increased considerably with the biofeedback and retention conditions.
The box plots shown in Fig. 3 demonstrate that the median MTC values increased significantly during the biofeedback and retention conditions. There was larger MTC variability in the biofeedback and retention conditions, highlighted by the larger interquartile boxes and the greater number of outliers. The range of the MTC values was also larger for these conditions, indicated through the length of the whiskers. Another observation was that the median MTC was almost similar for both feet for all participants and conditions, except for participant 3.
The MTC distributions of all participants for the baseline, biofeedback, and retention conditions are shown in Fig. 4. The biofeedback and retention distributions are highly skewed to the right due to the presence of large extreme MTC values in these conditions. This right-skewed distribution is related to the nature of our problem where there is a defined lower limit when walking since the MTC values cannot be lower than zero, but there is no defined upper limit. Fig. 5 shows the examples of the MTC fluctuations for participants 1, 2, and 7 for the baseline, biofeedback, and retention conditions. Generally, it was observed that the baseline data are fairly consistent for all participants; however, the MTC values and variability are significantly larger in both short and long retention conditions. Participants 1 and 2 were able to successfully maintain their MTC above the personalized threshold for the biofeedback condition, but participant 7 had a difficult time at the beginning of the trial learning how to raise each foot above their threshold. This resulted in the large amount of variation seen in their left foot data for the biofeedback condition. However, after being trained by the biofeedback, participant 7 was able to learn how to successfully raise both feet above the threshold for the rest of the trial, which is clearly evident in the data. All participants tended to overshoot the MTC in the retention conditions since there is no upper threshold or boundary. Some participants showed trends where there would be several smaller declines in the data, followed by a sudden spike in the MTC, as seen in participant 1's short retention left foot data. This is because the participant suddenly remembered to increase their MTC throughout the trial. An overall gradual decline was also observed in the MTC during the retention conditions for all participants. This could be due to fatigue by the end of the trial. Another reason could be that the participants were more aware of their MTC at the beginning of the trial and exaggerated their MTC. However, over time they started returning to their normal MTC values.

IV. LIMITATIONS AND FUTURE WORK
For this study, female participants without gait impairments were recruited to evaluate the feasibility of the proposed training system. Due to the current design of the wearable system, we were limited to recruiting participants that fit into the shoe size (women's size 8). The small sample size limited the generalizable conclusions drawn from the study observations. In the future, the system will be redesigned into a simple wearable that can be mounted to all types of footwear. Future studies will have a larger sample size that includes different genders, age groups, and gait impairments to test the effectiveness of the training system on a representative population. Additional gait parameters will also be collected and analyzed such as the step length and step pace to study its impact on MTC height, balance, and risk of falling.
Although using one buzzer at one beeping tone was effective in increasing the MTC, the information provided was vague, and some participants found it challenging to understand how to raise their feet effectively when instructions were not given. Participants would raise their MHC or MX2 instead of their MTC, which would result in an unnatural gait pattern. The ambiguity of the biofeedback caused participants to overshoot their MTC, resulting in greater MTC variability. A future study will be performed to investigate whether additional feedback would be effective in preventing participants from lifting their foot too high and minimizing MTC variability.
Different beep tone frequencies could also differentiate which foot is experiencing a low MTC. A multisensory approach might be more beneficial than separate modes of feedback, such as vibrotactile feedback.
Finally, the feedback training and retention sessions should be explored further to determine the actual effectiveness of the biofeedback over different periods of time. With additional feedback and retention sessions, the feedback can be gradually removed over time to ensure that the user does not become dependent on the feedback and has truly learned the new gait pattern.
V. CONCLUSION A novel wearable system was proposed that measures the MTC in real time using ToF sensors and provides auditory biofeedback to increase the MTC. The purpose of the study was to determine whether the system could be used for training applications with older adults or gait-impaired individuals to minimize tripping risks that could lead to falls. Biofeedback with personalized thresholds was used to successfully warn the users to walk with an increased MTC for feedback and retention sessions. However, greater MTC variability was also observed during the feedback and retention conditions. Since ToF sensors have not been used in previous biofeedback training systems, the results of this study also proved that ToF sensors could be a useful device for biofeedback systems to improve gait. The proposed system has a high potential to be used in various applications, such as with assistive wearable technologies or at-home rehabilitation training, to contribute to the ultimate goal of reducing trip-related falls.