Robust Method for Mid-Activity Tracking and Evaluation of Ankle Health Post-Injury

Abstract

Objective:

To present a robust methodology for evaluating ankle health during ambulation using a wearable device.

Methods:

We developed a novel data capture system that leverages changes within the ankle during ambulation for real-time tracking of bioimpedance. The novel analysis compares the range of reactance at 5kHz to the range of reactance at 100kHz; which removes the technique’s previous reliance on a known baseline. To aid in interpretation of the measurements, we developed a quantitative simulation model based on a literature review of the effects on joint bioimpedance of variations in edematous fluid volume, muscle fiber tears, and blood flow changes.

Results:

The results of the simulation predicted a significant difference in the ratio of the range of the reactance from 5kHz to 100 kHz between the healthy and injured ankles. These results were validated in 15 subjects - with 11 healthy ankles and 7 injured ankles measured. The injured subjects had lateral ankle sprains 2-4 weeks prior to the measurement. The analysis technique differentiated between the healthy and the injured population (p<<0.01), and a correlation (R=0.8) with a static protocol previously validated for its sensitivity to edema.

Conclusion:

The technology presented can detect variations in ankle edema and structural integrity of ankles, and thus could provide valuable feedback to clinicians and patients during the rehabilitation of an ankle injury.

Significance:

This technology could lead to better-informed decision making regarding a patient’s readiness to return to activity and / or tailoring rehabilitation activities to an individual’s changing needs.

I. Introduction

A total of 23,000 ankle sprains occur per day in the United States – 91% of which are lateral ankle sprains – making it the most common sports-related musculoskeletal injury [1–6]. After the first sprain, a patient is much more likely to reinjure the ankle and some patients may even experience long-term disability [7–10]. An ankle sprain is initially evaluated based on the presence and level of edema and limitations to the joint’s range of motion. Both measures are qualitative, subjective, and rely on a healthcare worker’s expertise. Imaging studies are often used to diagnose the injury by revealing structural abnormalities or ligament tears, but these studies are expensive, time-consuming, may expose the patient to radiation, and require an expert to interpret the findings [11, 12]. Physical examination alone has a diagnostic sensitivity of 96% and specificity of 84% [13].

Musculoskeletal injuries have characteristically long recovery times. After diagnosis, with appropriate medical interventions, a patient enters a period of recovery and rehabilitation. During this rehabilitative period, repeat clinical visits or imaging studies are impractical. Ideally, wearable technologies could be used to provide constant feedback to patients during this period. However, the only currently available technologies for quantifying ankle health status focus on the joint’s range of motion, most commonly assessed using inertial measurements [14]. These range of motion measurements do not fully capture the physiological changes occurring in a healing joint. Additional technologies need to be developed to provide precise, actionable feedback to the patient for optimizing their rehabilitation without relying solely on clinician input.

Bioimpedance analysis (BIA) holds merit for inclusion in a musculoskeletal rehabilitation monitoring device. BIA is a non-invasive method for assessing the composition of tissue [15, 16]. To measure BIA, a small electrical current is injected into the tissue of interest. The potential drop in the current as it traverses the tissue corresponds to the impedance of that tissue. BIA is frequency-dependent – low-frequency currents primarily flow through the extracellular fluid paths as they are unable to penetrate cellular membranes. Higher frequency currents are able to penetrate cells, and thus flow through both the extracellular and intracellular paths. In the literature, the Fricke-Morse circuit model is often used to describe the bioimpedance of tissue [17]. This model consists of a resistor (R_e) parallel to a capacitor (C_x) in series with a resistor (R_i). The external resistance (R_e) represents the extracellular content of the tissue, and the capacitor (C_x) in series with the resistor (R_i) represents the capacitance of the cell’s membrane and its intracellular contents.

Single frequency BIA (SFBIA) is a type of bioimpedance measure of tissue that typically uses only a 50 kHz current frequency, which is considered the cutoff frequency of biological tissue, or the frequency at which the current is passing through both the extracellular and intracellular paths [18]. In the literature, SFBIA has been used for evaluating the health of muscles and joints using two methods: by comparing the acutely injured muscle / joint to the contralateral healthy muscle / joint or by monitoring an injured joint longitudinally over the course of rehabilitation [19, 20]. In the context of joint injury, SFBIA can potentially quantify edema, but there are several limitations, including (1) the inability to classify the source of bioimpedance changes (i.e., edema versus structural damage versus blood flow), (2) a need for normalization of the bioimpedance for proper interpretation, and (3) variability in the electrode placement and tissue placement, which can greatly alter the signal. To implement SFBIA on an injured patient, the signal from the injured joint is normalized against the contralateral joint or against its baseline signal prior to injury [20, 21]. Either way, this normalization requires an increased number of sensors or an increased duration of wearing them, which limits the portability and ease of use of the system. Additionally, SFBIA is highly sensitive to electrode placement and the position of the tissue volmne, which limits the recordings to laboratory or clinical settings while the patient is immobile. Ideally, with a novel, more portable implementation of bioimpedance analysis, a patient with an injured joint could monitor their rehabilitation over time and the system could provide an early warning sign before reinjury occurs.

In our recent work, to overcome some of the limitations of SFBIA, we developed a BIA method for quantifying joint edema based on positional changes of the ankle joint. This technique compares the effects that changes in ankle position have on the range of the resistance measured at 5kHz against the range of the resistance measured at 100kHz [22]. Changes in ankle position alter the distribution of extracellular fluid within the joint cavity which affects extracellular resistance (R_e from the Fricke-Morse model) - the primary path for low-frequency currents. This method was validated in a human cadaver model (where it demonstrated 20-mL resolution) and on a cohort of human subjects [22]. In that work, the 5kHz signal was compared against the 100kHz signal, which made the need for a healthy contralateral comparison to normalize the findings unnecessary, and reduced the dependency of signal quality on electrode placement. That work mitigated several of the limitations of BIA technologies and presented a wearable form-factor for these sensors. However, that technique still required patients to perform a specific set of movements, could only detect ankle swelling (not structural defects), and could not be recorded during ambulation - all of which limit the capabilities of this system for providing the level of quantitative feedback necessary to best aid in a patient’s rehabilitation.

In this paper, we make several key changes to our technology to address these limitations. Specifically, in this new implementation of our wearable BIA monitoring system, hardware and firmware adjustments enable accurate measurement of the ankle joint’s bioimpedance and position for robust tracking of changes in the joint’s edema during ambulation. We also quantify the relationship between our bioimpedance measurements measured during human subjects’ activities of daily living and the structure of the underlying collagen fibers using a software model. By demonstrating such a relationship, this work suggests that wearable BIA measurements can provide quantitative, clinically relevant information related to a patient’s injury recovery status. Such feedback will enable more personalized and optimized rehabilitative efforts, whether during initial / follow-up visits to the clinic or while the patient simply engages in their normal, everyday activities. We envision that tins type of longitudinal, rapid feedback could alert a patient to when they are approaching a threshold of reinjury, providing the opportunity to readjust their rehabilitative efforts and prevent repeat clinical visits.

II. System and Experimental Design

A. Signal Analysis

The Fricke-Morse circuit model has three components—R_e, R_i, and C_x —which together describe the bioimpedance of tissue. Estimating these values requires impedance measurements at multiple frequencies and a non-linear least squares-based algorithm. Substantial measurement time and computational power are required – both of which are unfavorable for implementation of BIA in a wearable system designed to provide real-time feedback to the user. To circumvent the requirements presented in the Fricke-Morse model estimation, we devised a simple and robust method for assessing the underlying biological phenomenon within the ankle. Our method compares the changes in the reactance at two distinct frequencies recorded while the subject performs a task that stresses the joint.

1). The Reactance of the Tissue

Researchers have studied the effect of muscular injuries on bioimpedance at the tissue and muscle fiber level. At the tissue level, Nescolarde et. al determined that longitudinal changes in reactance – measured at 50 kHz – is an optimal metric for representing muscle tissue status since it is associated with cell density and soft tissue integrity [21]. At the muscle fiber level, Sanchez et al researched the effect of myotonic dystrophy on the Fricke-Morse circuit components using a mouse model [23, 24]. They attributed the changes in capacitance (C_x) to cell membrane integrity, and intracellular resistance (R_i) to cell inflammation. In the case of injured muscle fibers, they observed a decrease in R_e(~30%) and C_x(~40%) and an increase in R_i(~35%). On the cellular level, Dodde et. al analyzed the effects of applying pressure to porcine cells of the spleen on the Fricke-Morse circuit components [25]. They concluded that applying forces up to 50% of the cell membrane’s strength lead to a non-significant increase in R_e and R_i and decrease in C_x. However, beyond the cell membrane’s strength (i.e. >50%) the cell raptures, and there is a significant decrease in R_e and a significant increase R_i and C_x due to the migration of intracellular fluids to the extracellular space. Building upon this previous research, our team have used the Fricke-Morse circuit model to simulate the effects of edema, collagen fiber tears and blood flow on the reactance of the tissue. The model uses resistive and capacitive values for the Fricke-Morse circuit components in the ankle joint’s impedance space from the literature and our previous work as a baseline [22, 26].

2). The ratio of X_{5 kHz} to X_{100 kHz}

Using that model, two frequencies of electrical current are used to measure BIA. As mentioned, the low-frequency, 5 kHz current will primarily travel through the extracellular space since it is unable to pass through cellular membranes, and the high-frequency, 100 kHz current will take a more direct path using both intracellular and extracellular pathways. When these currents are applied to the ankle, they primarily flow through muscle fibers and blood vessels since the bones have low conductance and there is minimal fatty tissue and static extracellular fluid [27, 28]. Section II.A.1 presented key parameters for modeling current flow. Using that research, we can simulate the effects of blood flow, edema, and collagen fiber tears on the reactance of the ankle joint. In a normal physiologic state during sustained activity, blood flow to the region increases to meet the increasing metabolic demand of the muscles (Figure 1d). When the ankle is sprained, collagen fibers may tear and edema increases (as depicted in Figure 1c and 1e, respectively).

Figure 1.

(a) The anatomy of the ankle joint showing the lack of fatty tissues or static liquids. (b) A sample blood vessel and muscle fiber which are the primary path for the current applied to measure the bioimpedance. (c) A muscle fiber tear, showing the migration of intracellular fluids to the extracellular space surrounding. (d) An increase in the red blood cell count and glucose due to sustained muscle activity. (e) An increase in edema due to muscle inflammation. (f) Logarithmic scale of the ratio of the change in the low frequency reactance to the high frequency reactance due simulating the effect of blood flow, edema, collagen fiber tear, and collagen fiber tear accompanied with edema in the ankle joint on the baseline ankle impedance of 14 healthy subjects.

In an uninjured state, when blood flow is augmented during sustained activity, the increased number of red blood cells increases the number of intracellular pathways for current to travel through, which then decreases R_i. This change also increases the extracellular resistance and decreases the tissue capacitance, as reported by Abdelbaset et. al [29]. In another study by Gheorge et.al, they reported the effect of changes in blood viscosity on the Fricke-Morse to be around 1.5% in R_e, 5% in R_i and 10% in C_x [30].

In an injured state, tenocytes tear, and concomitant collagen fibers rapture, spilling intracellular contents into the extracellular space. This decreases R_e (since there is now more extracellular fluid) and increases R_i (since there are fewer pathways through intact cells). Sanchez et al demonstrated that intact myocyte membranes maintain the system’s capacitance, therefore tom muscle fibers will also lead to a drop in C_x [23]. With injury there will also be an increase in extracellular edema, which further decreases R_e since there is an increased volume of the conductive extracellular fluid. In their study, they reported a 30% decrease in R_e, 35% increase in R_i and a 40% decrease in C_x 24 hours after the injury.

3). Simulation Model

To study the effect of edema, collagen fiber tear and blood flow on the localized reactance of the ankle at 5kHz and 100kHz, we devised a numerical analysis simulation model. The model utilizes the following equation to calculate the impedance of the Fricke-Morse model at a specific frequency (ω).

$$R+jX=\frac{R_e(R_i+\frac{1}{j\omega C_x})}{R_e+R_i+\frac{1}{j\omega C_x}}$$ (1)

The simulation model outputs the ratio of the changes in the reactance at 5kHz and 100kHz due to changes in the Fricke-Morse components from baseline. The baseline impedances used are the bioimpedance spectroscopy data collected from our previous study. Specifically, we used non-linear least squares to estimate the values of the Fricke-Morse circuit components from the bioimpedance spectroscopy measurements of 14 healthy ankles. The estimated Fricke-Morse parameters are presented in Appendix. For each baseline ankle joint bioimpedance, we simulated the effects of blood flow, edema, collagen fiber tear and collagen fiber tear accompanied with edema on the ankle joint using percentage changes reported in the literature for each of these phenomena as mentioned earlier and shown in Figure 1.f.

In Figure 1f, we present the results of our simulation of BIA in the ankle. When comparing the ratio of change in reactance at 5 kHz and at 100 kHz we found a significant difference between healthy and injured ankles as shown in Figure 1f. This result is consistent with our expectations based on the impact of the described pathophysiologic changes during activity and injury, and with the earlier findings of Freeborn et. al in their study on the effects of biceps muscle fatigue on bioimpedance spectroscopy [31, 32]. The findings of this simulation encourage further research and hardware development into the BIA phenomenon and its clinical uses.

4). Limitations of Simulation Model

In Figure 1.f, we also present the effect of varying the intensity in the changes in the Fricke-Morse parameters due to edema, collagen fiber tear, and edema and collagen fiber tear combined. In the case of increasing edema, reducing R_e further led to a change in the reactance at 5kHz and 100kHz at a similar rate making the ratio of the changes in the reactance at 5kHz to the change in the reactance at 100kHz a viable method for the detection of edema, but not quantifying it.

5). Differentiating Between Edema and Muscle Tear

After the acute phase of an injury, the tissues enter a phase of rebuilding which includes a reduction in edema and increasing collagen fiber strength. The level of edema and strength of the reforming fiber are indicative of the progress of the rehabilitation and the probability of reinjury [7]. In Figure 1f, the ratio of the change in the reactance at 5kHz versus at 100 kHz successfully differentiates healthy from injured ankles, but differentiating swelling from collagen fiber strength is a different challenge. In our previous work, we studied the effect that changes in ankle position have on low (5 kHz) and high-frequency (100 kHz) resistance measurements. We showed that these positional changes move the extracellular fluids around the joint altering the extracellular resistance (R_e) which in turn predominantly impacts the low-frequency resistance (extracellular fluid dependent) compared to the high-frequency resistance [22]. Although, in this work we are using the changes in the reactance, instead of the resistance, at low and high frequency for our analysis, the concept of extracellular fluids shifting in the tissue due to the pressure from the joint structure caused by its rotation still applies here.

In this work, we measure the bioimpedance of the ankle joint while subjects are walking using our wearable sensing hardware. Normally, the instantaneous changes observed in the ankle’s bioimpedance during a subject’s gait cycle are possibly due to the tendons and ligaments, and varying blood flow in the current path. However, if there exists edematous extracellular fluid, the pressure from the joint structure would shift the fluid around it changing the extracellular resistance (R_e) in the current path. We hypothesize that in the presence of joint edema, there will be an increase in the instantaneous change per step in the low frequency reactance measurements compared to the lower frequency measurements as shown in Figure 1.f. Nevertheless, these positional changes should not compound throughout the duration of a walking session. Rather, we hypothesize that the baseline changes in bioimpedance from the beginning of a walking session are mainly due to micro-damage to any recently injured tendons, ligaments, and tissue coupled with low-level edema. To test these hypotheses, we developed two metrics for comparing the range of the reactance at 5kHz to the range of the reactance at 100kHz: (1) per step (${\widehat{h}}_{\alpha }$), and (2) per walking session (>200 steps) (β).

B. Software Model Development

To compute ${\widehat{h}}_{\alpha }$ and β, it is necessary to split the reactance measured at the two frequencies based on each subjects’ steps as shown in Figure 2. This windowing uses the inertial measurement unit (IMU) employed on our custom hardware. The IMU captures the angular velocity of the foot, which is used to determine if the subject is moving. This is performed by taking 3-second windows of the angular velocity and convolving those values on themselves to compute the energy of that window’s angular velocity (ω[t]) as shown in the following equation.

Figure 2. Data Analysis Workflow for Determining Presence of Edema and disruption to Structural Integrity of the Ankle.

(a) The data acquisition system placed on the subject’s leg with the necessary current and voltage electrodes placed distally and proximally to the ankle joint and the IMU placed on the foot. (b) Sample data of a representative injured subject’s X-axis angular velocity, Z-axis acceleration, and reactance measured at 5kHz and 100 kHz. (c) A magnified view of the sample data showing how the data windows are created and used in splitting to split the reactance data into vectors per step. (d) The reactance vectors per step are used in the model to detect edema and collagen fiber tear in the ankle joint.

$$E_{\omega }[t]={\underset{t=n}{\overset{n+3}{\Sigma }}|\omega [t]|}^2(rad/s^2)$$ (2)

That energy is compared against an experimentally determined threshold of 10,000. If the energy is higher than that threshold, the peaks of the Z-axis (lateral) acceleration signal from the IMU are used to identify the heel-strikes which mark the beginning of each step as shown in Figure 2c. Each peak needs to be at least 350ms from the previous peak and beyond a certain threshold (1g) to remove errors from irregularities in the signal. Since the bioimpedance is sampled at a lower frequency than the IMU, the start and end of each step’s bioimpedance window is identified by finding the absolute minimum time difference between the time of the heel strike and the time of the bioimpedance measurements. These data are then used in the model presented in Figure 2d.

After splitting the reactance measured to per step arrays as shown in Figure 2d, the range and mean of each step’s reactance is calculated. ${\widehat{h}}_{\alpha }$ is calculated by taking the ratio of the range per step of the reactance at 5kHz to the range per step of the reactance at 100kHz. To calculate β at any step (s), the range of the mean of the reactance per step from the start of the walking session to step s is calculated. The ratio of this range at 5kHz to 100kHz is then taken to calculate β as shown in Figure 2d.

C. Hardware and Firmware

To detect edema and structural damage in the ankle joint while the subject is performing their daily activities, we modified our previous hardware and firmware. For the hardware, as discussed, adding an IMU to the peripheral cable allows position and movement of the limb to be measured during the gait cycle. The IMU LSM6DS3 (ST, Geneva, Switzerland) was selected for its accuracy, low power consumption and ease-of-interfacing. For the firmware, an interrupt-based architecture was implemented where the sensors’ (Bioimpedance and IMU) state machines are updated every 1ms and the data are saved onto a double-buffer to ensure a constant sampling rate from the sensors. The buffered data are saved onto an SD-card which is processed offline using Python. Relevant system properties are shown in Table I.

TABLE 1.

System Properties

Parameter	Value
Battery life	12 hours
Size	5.2x3.8x1.8 cm
Weight	32 g
Bioimpedance:
Frequencies:	5 kHz, 100 kHz
Sampling Rate	10 Hz
Resolution	0.2 Ω
Inertial Measurement Unit
Number of Axis	6
Sampling Rate	50 Hz

D. Data Collection Protocol for Method's Evaluation

To validate the hypothesis that ${\widehat{h}}_{\alpha }$ can detect edema in the ankle, we recorded data from 15 subjects during ambulation and asked them to perform the BIA ankle positional protocol as described in our previous work and depicted in Figure 3c once per hour [22]. The data collection protocol was approved by the Georgia Institute of Technology Institutional Review Board, and all subjects provided written informed consent before participating in the study. This positional protocol was previously shown to correlate with edema. In this study, we sought to determine if ${\widehat{h}}_{\alpha }$ measured during walking also correlated with edema. We also sought to test if β (the difference in impedance after a continuous walking session) can differentiate between healthy and injured populations.

Figure 3. Recording setup and 8-hour recording protocol timeline.

(a) The wearable data acquisition is placed on the subject’s leg. (b) The overall recording protocol took 8 hours with the 5-minute positional protocol as depicted in (c) being performed every hour.

Figure 3 depicts the overall testing protocol. The modified system is placed on the subject’s ankle as shown in Figure 3a. Red dot gel electrodes (3M, Saint Paul, MN) are used for bioimpedance measurements. The electrode snaps and IMU are secured using Kinesio tape (Kinesio, Albuquerque, NM) to further secure them and dampen the forces from movement. With the recording setup in place, the subjects performed the 5-minute positional protocol shown in Figure 3c. The subjects then perform their normal daily activities while performing the 5-minute positional protocol every hour as depicted in Figure 3b.

The participants were recruited via word of mouth by either the engineering research staff or the Georgia Institute of Technology athletic trainer of the study team. The sensors were outfitted in the lab or the athletic center in the early morning to reduce any residual effect from prior movement on the data. The subjects were then instructed to go about their daily activities for eight hours. After the eight hours of data collection, the subjects returned to the lab or the athletic center for the device to be removed.

The study was performed on 15 subjects between the ages of 18 and 30. Of these subjects, 7 had an ankle injury in the two to four weeks prior to data collection. Data from the contralateral healthy ankle were also collected from 3 of the 7 injured subjects on a separate day. On the remaining 8 healthy subjects, data were collected from their dominant ankle due to its higher chance of injury [2]. In total, 11 healthy ankles and 7 injured ankles were recorded. Four of the injured ankles were diagnosed by a medical professional and the rest were self-reported. The injuries were grades 1-2 lateral ankle sprains. The data were analyzed offline using Python.

E. Statistical Analysis

To test the ability of the methods presented to sperate between healthy and injured groups, we first tested the data for normality using Wilk-Shapiro test. Since one of the data groups (healthy ${\widehat{h}}_{\alpha }$) failed the normality test, we used Wilcoxon rank-sum test where a p-value less than 0.05 is considered significant. For each score (${\widehat{h}}_{\alpha }$ and β). Cohen’s (d) effect size between the healthy and injured group was also calculated where an effective size higher than 1.4 is considered large effect [33]. An example of the variables extracted from the model in Figure 2 is presented in Figure 4. We also used Pearson correlation tests to show the correlation of ${\widehat{h}}_{\alpha }$ to the static protocol.

Figure 4. Method for Comparing the Full Walking Session to 5 Minute protocol.

(a) The range of change in the reactance measured at 5kHz and 100kHz during a continuous walking session is used to calculate β. (b) The range of the change in the reactance measured at 5kHz and 100kHz per step is used to calculate ${\widehat{h}}_{\alpha }$. (c) The mean ${\widehat{h}}_{\alpha }$ of the last 10 steps is correlated to the ratio of the range of change in the reactance measured at 5kHz and 100kHz using Pearson’s correlation

III. Results and Discussion

A. Ratio of the Ranges of Reactance Per Step (${\widehat{h}}_{\alpha }$)

To test our hypothesis that ${\widehat{h}}_{\alpha }$ is sensitive to detecting edema, we correlated the ratio of the range of the reactance at 5 kHz to the range at 100 kHz from Figure 4a against those values found during the 5-minute, static positional protocol from Figure 4b as shown in Figure 4c. To ensure that no residual edema or muscle tear from prior walking sessions skewed the result, data from the first substantial continuous walking session were used. In this context, a “substantial” session is considered one in which the subject walks for more than 200 steps with a maximum of pause of 1 minute between successive steps. For all 15 subjects, the mean of ${\widehat{h}}_{\alpha }$ was calculated from the last 10 steps of the first walking session and tested for correlation with the 5-minute positional protocol performed immediately after that session for all 15 subjects as shown in Figure 4c. This comparison yielded a Pearson’s conelation coefficient of 0.8 as shown in Figure 5c. This supports the hypothesis that ${\widehat{h}}_{\alpha }$ is sensitive to edema, since the positional protocol was previously shown to correlate with edema levels [22]. The results of ${\widehat{h}}_{\alpha }$ and the ratio of the range of X_5kHz to X_100kHz from the static protocol for the injured and healthy group are presented in Figures 5 (a, b, c) and tested for statistical significance (p = 0.0021) and Cohen’s effect size of 1.6. This indicates the method’s ability to differentiate between healthy and injured ankles. There is also similarity between the injured and healthy population ${\widehat{h}}_{\alpha }$ scores in Figure 5b and the results of the simulation model for changes in edema and blood flow from Figure 1f.

Figure 5.

(a) Plot showing h_α vs steps for all subjects and (b) a scatter plot of the mean of h_α for the last ten steps for the healthy and injured groups showing a statistically significant p-value. (c) Plot of the mean h_α at the last ten steps in a continuous walking session, correlated the output of the 5-minute protocol done after with a Pearson correlation coefficient of 0.8. (d) Plot showing β vs steps for all subjects and (e) a scatter plot of the β at the last step of a continuous walking session showing a statistically significant p-value.

B. Ratio of the Ranges of Reactance Per Walking Session (β)

While the step-by-step analysis demonstrated that BIA at these two frequencies could detect edema, it provided little information about the integrity of the tissue. During a substantial period of walking, micro-tears are expected to occur in the ligaments, tendons, and connective tissue [34]. To better understand the impact that this degradation of tissue integrity has on the reactance measured using BIA, we calculated the range of reactance recording during a continuous walking period as shown in Figure 4a. Significant differences in this range between the injured and healthy groups when taking the β from the last step of the continuous walking session was found (p<<0.001) and Cohen’s d effect size of 1.96 as shown in Figure 5 (d, e)). To control for inter-subject inter-session variability (particularly in the number of steps per walking session), the β is also calculated at the 200^th step of the first substantial walking session for all subjects and a significant difference between the injured and the healthy groups was found (p <<0.01). There is also similarity between the β scores for the healthy and injured population shown in Figure 5e and results of the simulation model for collagen fiber tear and blood flow from Figure 1f.

C. Zero Crossings Data Analysis

During the typical gait cycle, the ankle’s direction of rotation changes (thus, its angular velocity = 0) four times: (1) at a neutral position (i.e., with the foot and shank at or near 90° to each other) just prior to heel-strike, (2) in a slightly plantarflexed state once the foot is flat on the ground just after heel-strike, (3) in a dorsiflexed state just prior to heel-rise, and (4) in a plantarflexed state just prior to toe-off leading into swing phase. These same joint configurations are performed in the static positional protocol, providing points of comparison between the dynamic (walking) and static (positional) tasks[22]. The software model presented was tested using only the data closest to the zero crossings by choosing the bioimpedance measurements that had the absolute minimum time difference from the time of the zero-crossings. For the ${\widehat{h}}_{\alpha }$ calculated using the zero crossings data, the Speannan’s correlation coefficient with the ratio of the range of reactance at 5kHz to the range of reactance at 100 kHz from the 5-minute protocol is 0.63. The calculated p-value is p<<0.01 for the separation between the healthy and injury group. The difference in the correlation score using all bioimpedance measurements (as presented in Section III.A and B) and the measurements closest to the zero crossings may be due to the relatively low sampling rate of the bioimpedance or due to a delayed response for the impedance from the changes in the ankle position caused by the loading of the joint at these positions. In some cases, the nearest bioimpedance measurement to the zero crossing was up to 50ms away. For the β score using the data closest to the zero crossings, the p-value is « 0.01.

Using only the data at the angular velocity’s zero crossing would enable significant reduction in the bioimpedance samples needed and hence the power consumption, as the bioimpedance can be measured only at the zero-crossings using an interrupt-based approach. This would allow the software model to be fully implemented on an embedded processor in a wearable-form factor.

D. Limitations of Human Subject Study

In this study, there was no standard protocol for the type of tasks performed through the day or specific number of steps that each subject was asked to do. There was also no stratification for the injured subjects and the severity of their injury. These limitations made it not possible to quantify the severity of the injury using our technology. Moreover, this lack of standardization limited our ability to study the directionality of the changes in reactance.

E. Implications of Findings on Feasibility of Wearable Joint Healthy Monitoring

The work presented in this paper provides a robust method for capturing and analyzing bioimpedance in the ankle. Previous work in this field typically prescribed a set of controlled exercises to ensure accurate measurements. In our work, we designed a solution for adapting this technology into a wearable form factor that leveraged the impact of ambulation on the signals rather than mitigating them. Additionally, this novel method of signal interpretation requires minimal algorithmic and computational complexity making it suitable for being embedded into a miniaturized, wearable system. The results of this algorithm are presented in Sections III. A, B, and C and demonstrate that this technique is capable of real-time detection of edema and tissue integrity changes in the ankle during activities of daily living. This usability improvement and enhanced algorithm enables real-time, joint health status updates for the wearer. These notifications, if properly utilized could greatly aid in personalizing clinical rehabilitation efforts.

IV. Conclusion And Future Work

In this paper, we present for the first time a system that performs BIA in the context of detecting edema and structural integrity in the ankle joint while a subject is walking. In our previous work, these measurements were only possible in a static, controlled enviromnent while performing a set of standardized movements and were limited to detecting edema only. The system fuses IMU and bioimpedance signals to evaluate and detect edema and muscle tears, which are crucial for physicians to understand the recovery progress and the subject’s ability to return to activity. The analysis technique for interpreting BIA while walking was first validated in a simulation model using parameters and results from the literature and a human subjects study. The simulation results held promise, so 15 subjects both with and without injury were recorded while walking. Two metrics were developed to describe the inter-step range of reactance and the intra-walking session range of reactance. Both metrics statistically separated the injured ankles from the healthy ankles (p<<0.01). The inter-step analysis was shown to correlate with edema. The intra-session range of reactance was thought to also include the effects of microtears on the tissue.

Future work will include working with physicians to help us calibrate our model to provide better feedback about the status of the ankle’s health. Future work will also include in-lab standardized experiments and state of the art imaging systems to further correlate and quantify the results of our model to the physical phenomena observed allowing us to study the effects on the direction of the data not only its magnitude. It will also include the studying of the effect of the joint’s loading state on the fluid content (i.e. compression while bearing weight vs floating while swinging freely). Overall, this work presents a substantial step toward adapting bioimpedance into a portable form-factor and encourages future research into its uses in constantly monitoring injury, pathologic changes, and rehabilitation.

V. Appendix

In this section we present the algorithm used to calculate the values of the Fricke-Morse components used as baseline in our simulation model. In our previous work, bioimpedance spectroscopy data were collected over the frequency range from 5kHz to 100kHz with a resolution of 371 Hz for a total of 256 bioimpedance measurements per sweep. Using Python, we created a function that calculates real component the Fricke-Morse impedance for any given R_e, R_i and C_x values at the same frequencies from our bioimpedance spectroscopy data [30]. This function along with the data collected from the ankle joint from our previous work are inputted into SciPy’s curve fit function to estimate the Fricke-Morse circuit components for each ankle. Figure 6 shows an example bioimpedance spectroscopy of a healthy ankle joint and the estimated Fricke-Morse impedance using our algorithm. Table II shows the Fricke-Morse component values for 14 healthy ankles. The bioimpedance values presented are similar to the values reported by King et al. [26]

Figure 6.

An example of localized bioimpedance spectroscopy data and its associated Frick-Morse estimated parameters. (R_e~100, Ω R_i~520Ω, C_x = 10nF)

TABLE 2.

Ankle Fricke-Morse Circuit Component Values

Re(Ω)	Ri(Ω)	Cx (nF)
102	520	9.6
107	633	8
160	940	5.4
163	1178	3.3
115	543	9.2
108	770	5.2
113	877	5.4
152	676	7.7
127	746	6.5
100	869	5
128	1022	5.5
108	765	6
137	577	10
101	511	9.6