PhD Pinboard: Monitoring Continuous Joint Load with Wearables

This PhD Pinboard article by Jimmy Caroupapoullé explores his research into continuous load monitoring with wearable technology and smart devices.

I am Jimmy Caroupapoullé, a visiting researcher at the University of Southampton where I completed my PhD at the Faculty of Engineering.

My doctoral research focused on developing a methodology to evaluate how day-to-day physical activity correlates with disease activity in individuals diagnosed with Rheumatoid Arthritis (RA) and/or Osteoarthritis (OA). I am a telecom engineer by background and developed a platform to continuously monitor biomechanical forces – specifically load rate [1] which refers to the speed at which force is applied to the body - using Inertial Measurement Units (IMUs) captured by smartphones.

While my PhD focused on clinical populations, I am particularly interested in extending this work to individuals with physically demanding occupations, including tactical [2] and elite athletes, where excessive or poorly managed mechanical loading may contribute to early onset of OA.

Why does this research matter?

The spotlight on elite athletes generally focuses on their on-pitch performance, which is now well monitored and understood. Yet, athletes spend the majority of their time off the pitch; recovering, travelling and managing daily life where load exposure is rarely quantified.

Wearable technology offers an opportunity to passively capture this off-pitch loading, providing a more holistic picture of mechanical stress, recovery, and long-term joint health. The challenge, however, is that the metrics commonly reported by consumer devices, such as step count, calories burnt, or activity minutes, do not directly reflect musculoskeletal load and are often difficult for clinicians or practitioners to interpret.

My PhD explored whether raw sensor data from smartphones and wearables could be transformed into clinically and biomechanically meaningful measures of joint loading, bridging the gap between laboratory biomechanics and real-world data. 

My PhD Story

When I started my PhD, I assumed that OA and RA were conditions mainly affecting older adults and that “more physical activity” was always beneficial. Conversations with patients and clinicians challenged this view and revealed a known gap in practice: reliance on patient-reported symptoms provides only a partial, and often biased picture of disease activity. In addition, referrals to specialists can lead to significant delays before being diagnosed.

RA and OA significantly differ but both manifests through pain and stiffness in the joints:

• RA is an autoimmune disease, typically with symmetrical joint pain (i.e. in both knees), that is diagnosed through blood tests.

• OA is a disease caused by the wear and tear of the joint.

Evidence shows that excessive or poorly distributed mechanical loading can be harmful, and OA is more prevalent in elite athletes and individuals with physically demanding occupations [3]. However, sub-optimal mechanical loading might prevent the development of bone strength (osteogenesis) leading towards osteoporosis (OP) where bones become weak and brittle [13]. Imaging techniques like MRI and X-Ray imaging provide a snapshot of structural deterioration but cannot capture the cumulative loading patterns that may contribute to disease onset and progression-particularly in OA.

Therefore, measuring accumulated joint loading, in a continuous pattern and free-living conditions, would be of interest to those potentially susceptible to RA, OA or OP.

How is Mechanical Joint Loading Measured?

Multiple methods exist to measure joint loading and each offer benefits, with increased measurement accuracy often addiing practicality challenges. Therefore, a trade-off - as Jo has previously discussed here - must be considered depending on use case. Figure 1 below illustrates three established approaches for estimating lower limb loading during locomotion:

(1) Implant-based strain measurement: Mechanical strain is measured directly using implants and expressed in micro-strain (με), with the estimated load (E) computed as a calibration-weighted sum of individual strain signal. 

(2) External load measurement: Ground reaction forces are recorded using force plates or treadmill-integrated sensors.

(3) Inertial measurement-based load estimation: Body-worn tri-axial accelerometers measure accelerations along the x, y, and z axes; the resultant load force is estimated from the vector magnitude of acceleration scaled by body mass.

Figure 1: Schematic overview of methods used to estimate lower-limb loading during locomotion.

Measuring the strain on the lower limbs using implanted sensors is impractical and invasive. Lab-based assessment using force plates remain the gold standard for gait and loading assessments. Their limited accessibility from a continuous perspective motivated our investigation on smart wearables and whether smartphone sensors could provide meaningful measures of load forces.

Data Collection

My project involved patients recruited within the NHS for clinical research. Early work focused on collecting prospective data and identifying key protocol challenges. Participants were provided with a smartphone and a smartwatch for monitoring in real world setting (free living conditions).

Such smartphones and wearables are fundamentally miniature electronic systems designed for continuous, non-invasive data collection on the body. The core of any wearable is its sensor suite, which transforms physical phenomena into electrical signals. Recent years have seen the rise of composite scores (e.g., metabolic or sleep scores) and the introduction of new sensors, such as Electrocardiogram (EKG). Note that most activity classification (walking, biking, car) uses Global Positioning System (GPS) which uses satellite signals to track geographical location and is therefore, less effective indoors [5].

Raw sensor data (e.g., tri-axial acceleration values or light absorption intensity) is unusable on its own. Therefore, sophisticated and often proprietary algorithms are used to clean, filter, and convert this raw data into user-friendly metrics, such as steps, calories, and proprietary scores.

Figure 2 illustrates the smartphone coordinate system and the range of inertial sensors available through the Android operating system. These include data from the accelerometer, gravity sensor, gyroscope (calibrated and uncalibrated), linear acceleration, rotation vector, and step counter. For each sensor type, the corresponding units and axis-specific measurements (e.g., acceleration, angular velocity, or rotation components) are listed below as provided by the Android system.

Figure 2: Smartphone sensors coordinate system. Left: Definition of the smartphone-fixed coordinate system. Right: Summary of sensor event data available from smartphone inertial sensors (adapted from Google documentation, 2018).

Device placement on the body strongly influences measurement quality. Smartphones may be carried inconsistently but are often positioned closer to the body´s centre of gravity of the body, providing more representative data. Smartwatches are worn at the wrist, which is ideal to detect arm swing movements but limits their ability to represent lower limb load. Manufacturers have proprietary algorithms to estimate step counts from the accelerometry data.

Step count is obtained by filtering the peak-to-peak value of the accelerometer signal and counting the peak values. While popular, step count does not accurately reflect musculoskeletal load. Whereas, load rate examines the entire signal, including intensity and pattern. This however, is significantly more battery-consuming.

At the time of the study, only step count and calorie count were available in an aggregated daily format. Calories and Metabolic Equivalent of Task (MET) estimate energy expenditure by comparing oxygen consumption at rest with oxygen use during activity [9]. A consensus study into OA and Physical Activity (PA) agreed on using METs as a way to harmonise measurement of PA variables amongst cohorts but raised the need for specific measure of joint loading in OA [6].

To overcome the limitations of metrics from commercially available devices, we developed custom smartphone applications (RApp & OApp) to continuously collect and process raw accelerator data from smart devices. This is then used to estimate load forces during daily activities, which also enables comparison with implant-based and/or force-plate measurements of load.

Key Findings from Longitudinal Monitoring

My PhD spanned five years (2018 to 2022), including the COVID-19 pandemic, during which time we evaluated the feasibility of using our custom smartphone apps to continuously monitor load rate across multiple contexts, including:

1. Validation against anti-gravity treadmill and in free living conditions with non-diagnosed individuals [10,11].

2. Application in knee OA clinical trial (ACTIVe-OA study) as an outcome monitoring tool to detect within-participants changes that aligned with clinical changes (knee injection) [7].

3. Monitoring occupation-related physical activity across light to heavy manual work. We observed loading differed by workload category (sedentary to light manual to heavy manual) demonstrating sensitivity to real-world exposure [8].

A sustained increase in load rate relative to an individual’s baseline suggests elevated mechanical exposure and could prompt load management actions, such as modified training, targeted rehab, or temporary task modification. Sudden extreme spikes (we observed cases of >500× baseline in individual traces) may indicate acute events or high-risk episodes that warrant clinician review. A practical interpretation for athletes could be engaging in progressive loading from the off-season back into pre-season.

Figure 3 below shows 24-hour raw accelerometer traces from a smartphone and smartwatch. Periods of high signal variability reflect active use, while flat regions indicate passive monitoring. This figure demonstrates the feasibility of continuous collection and the impact of device placement on signal characteristics. It also highlights between-device differences, with smartphones (bottom figure) producing signals that better represent lower-limb loading with watches (top figure) picking up strong-arm swing and step peaks but are less representative of lower-limb force distribution.

Figure 3: Smartphone and Smartwatch extract of Raw accelerometer data over 24 hours.

Although user behaviour introduces variability between-individuals, long-term monitoring in a real-world setting allows for meaningful within-individual differences comparisons over time. Physical activity patterns changed dramatically in response to fluctuations in disease activity. For example, Figure 4 below illustrates how estimated lower limb loading differed across occupational workload categories, demonstrating the sensitivity of the approach to real workload exposure. The median load rate and variance increase with workload category, and outliers reveal high-exposure events.

Figure 4: Estimates of lower-limb load by workload levels. The boxplots show the median (green line), first and third quartiles (box), min and max (whiskers) including outliers of the estimated lower-limb loads (kg*m/s3) for each respective study participant grouped by workload levels; (A) sedentary (N = 2), (B) light manual (N = 16) and (C) heavy manual (N = 1).

While sample sizes are small in some categories (e.g., heavy manual N=1 in this dataset), the figure shows the method’s ability to discriminate exposure between groups. In one case, load rates exceeded baseline values by more than 500 times. Although real-world data remains limited, such outliers may reflect microfractures or structural stress contributing to OA progression. Larger cohorts are needed to define clinical thresholds.

Final Thoughts

Wearables have the potential to bridge the gap between subjective and objective biomechanics measurements, particularly for conditions driven by cumulative loading. Even without specialist expertise, meaningful insights can be extracted, but clinical relevance must be defined collaboratively with clinicians to ensure these tools are adopted at a global scale.

Our ongoing work focuses on translating these measurements into actionable insights, particularly for physiotherapists and clinicians managing OA and RA. We are currently exploring how continuous load monitoring can support personalised rehabilitation, injury prevention, and long-term joint health, both in clinical populations and high-performance sport.

References

[1] Verheul, J., Nedergaard, N. J., Vanrenterghem, J., & Robinson, M. A. (2020). Measuring biomechanical loads in team sports – from lab to field. Science and Medicine in Football, 4(3), 246–252. https://doi.org/10.1080/24733938.2019.1709654

[2] Wise, Sean R. MD, RMSK, FAAFP; Trigg, Steven D. MD. Optimizing Health, Wellness, and Performance of the Tactical Athlete. Current Sports Medicine Reports 19(2):p 70-75, February 2020. | DOI: 10.1249/JSR.0000000000000684

[3] Canetti, E., Schram, B., Orr, R., Knapik, J., & Pope, R. (2020). Risk factors for development of lower limb osteoarthritis in physically demanding occupations: A systematic review and meta-analysis.. Applied ergonomics, 86, 103097 . https://doi.org/10.1016/j.apergo.2020.103097.

[4] Migliorini, F., Vecchio, G., Pintore, A., Oliva, F., & Maffulli, N. (2022). The Influence of Athletes’ Age in the Onset of Osteoarthritis: A Systematic Review. Sports Medicine and Arthroscopy Review, 30, 97 - 101. https://doi.org/10.1097/jsa.0000000000000345.

[5] Martin, B., Addona, V., Wolfson, J., Adomavicius, G., & Fan, Y. (2017). Methods for Real-Time Prediction of the Mode of Travel Using Smartphone-Based GPS and Accelerometer Data. Sensors (Basel, Switzerland), 17. https://doi.org/10.3390/s17092058.

[6] Gates, L. S., Leyland, K. M., Sheard, S., Jackson, K., Kelly, P., Callahan, L. F., Pate, R., Roos, E. M., Ainsworth, B., Cooper, C., Foster, C., Newton, J. L., Batt, M. E. & Arden, N. K. 2017. Physical activity and osteoarthritis: a consensus study to harmonise self-reporting methods of physical activity across international cohorts. Rheumatology International, 37, 469-478.

[7] Yu SP, Ferreira ML, Duong V, Caroupapoullé J, Arden NK, Bennell KL, Hunter DJ. Responsiveness of an activity tracker as a measurement tool in a knee osteoarthritis clinical trial (ACTIVe-OA study). Ann Phys Rehabil Med. 2022 Sep;65(5):101619. doi: 10.1016/j.rehab.2021.101619. Epub 2022 Feb 14. PMID: 34879312.

[8] Wang, X., Perry, T.A., Caroupapoullé, J. et al. Monitoring work-related physical activity and estimating lower-limb loading: a proof-of-concept study. BMC Musculoskelet Disord 22, 552 (2021). https://doi.org/10.1186/s12891-021-04409-z

[9] Ainsworth, B. E., Haskell, W. L., Whitt, M. C., Irwin, M. L., Swartz, A. M., Strath, S. J., O’brien, W. L., Bassett, D. R., Schmitz, K. H., Emplaincourt, P. O., Jacobs, D. R. & Leon, A. S. 2000. Compendium of Physical Activities: an update of activity codes and MET intensities. Medicine & Science in Sports & Exercise, 32, S498-S516.

[10] Nazirizadeh, S. 2018. Load Rate Estimates on the Lower Limb Joints Using the Accelerometer of Wearable During Physical Activity. Doctor of Philosophy, University of Southampton.

[11] Nazirizadeh S, Stokes M, Arden NK, Forrester AI. Validity of load rate estimation using accelerometers during physical activity on an anti-gravity treadmill. Journal of Rehabilitation and Assistive Technologies Engineering. 2021;8. doi:10.1177/2055668320929551

[12] Kanis, J. A., Delmas, P., Burckhardt, P., Cooper, C. & Torgerson, D. 1997. Guidelines for diagnosis and management of osteoporosis. Osteoporosis International, 7, 390-406.

FAQs

Do I need a smartwatch to measure load rate?

No. Both smartphones and smartwatches contain the necessary components and sensors, but the location on the body affects the relevance of the measurements. A phone is not always carried by the user, unlike a watch, but it is typically kept closer to the centre of gravity of the user, which is more representative of the load in the lower limbs.

What is the difference between load rate and step counts?

Step count is obtained by filtering the peak-to-peak value of the accelerometer signal and counting the peak values. Smartwatches provide higher peaks, and manufacturers (e.g., Fitbit, Apple) have proprietary algorithms optimised to estimate step counts. Load rate examines the entire signal, including intensity and pattern, which is significantly more battery-consuming.

What is the difference between this monitoring methodology and the activity reminder that I get on my watch?

Smartphones and watches monitor the level of inactivity to estimate sedentary behaviour. These devices do not continuously record the entire sensor signal, as it affects the storage and battery performance. Activity classification (running, walking) considers repetitive sensor patterns, cross-checked with location data and speed. Note that these are also extracted using a cellular network signal (the distance of the device to the closest cellular antennas), not only GPS positioning.

How can data from smart devices be used to help with Rheumatoid Arthritis and Osteoarthritis (OA) diagnosis and prevention?

Anyone preparing to run a marathon knows to warm-up accordingly. Yet, it is rarely thought about by the general population for more everyday activities (e.g., before gardening or kneeling for 5 hours). Data obtained from smart wearables can be used to monitor and quantify the long term impact of physical activity on MSK conditions. Continuous joint load monitoring using wearables offers a practical way to understand how everyday activity - and not just training - shapes injury risk and long-term joint health.