Identifying Patterns of Physical Activity in Musculoskeletal Disease

  • Katy Weihrich

Student thesis: Phd


We aimed to find a novel approach for diagnosing functionality in patients with rheumatic and musculoskeletal diseases (RMDs). Accelerometer sensors embedded in consumer devices, such as Fitbits and smartwatches, open new possibilities for continuously tracking patients' symptoms. This can be achieved by developing apps that enable active tracking of functions by requesting questionnaires to be filled out routinely and by developing algorithms that can passively take accelerometer data and analyse patient activity patterns. The Knee Osteoarthritis: Linking Activity and Pain (KOALAP) study provides us with an example of such a potential diagnostic tool. Our goal was to develop a gait detection algorithm (GDA) to separate walking behaviours from non-walking behaviours continuously recorded by accelerometer sensors located on the wrist in free-living conditions. Our GDA performed with 97.20 % accuracy on a self-recorded, labelled dataset (SRDS). However, the translatability of this accuracy onto other datasets needs to be further explored, since the data was only recorded on one young and healthy participant. Furthermore, we aimed to extract step parameters to track a patient's physical function and developed a step parameter extraction algorithm (SPE) for this purpose. We identified the following parameters of interest: Walking Episode Length in minutes, Acceleration Range of gait cycles in m/sec^2, Step Rate in minutes, Symmetry of Step Time in seconds between the left and right step cycle and Variance of the Acceleration Range in m/sec^2. To assist the development of the algorithms and to examine the impact of different walking behaviours on accelerometer data, we used the SRDS. We identified two major ways in which the wrist location of the accelerometer influences the record accelerometer data. Walking can be divided into behaviours with free arm swing and stiff behaviours where the arm is fixed in relation to the body. Stiff behaviours tend to be noisier but also resemble data collected on the leg more closely. Swing behaviours tend to exhibit larger acceleration ranges. Applying our algorithms on the KOALAP dataset we found that we were unable to relate patients' step parameter values directionality to their pain reports. We did, however, find occurrences of significant correlations with some pain scores depending on the participant. We conclude that the step parameter we extracted could help track patients' performance over time, however, the effectiveness and the parameters used would need to be adjusted to the individual patient. To investigate the usefulness of the step parameter further, we applied our algorithms on the accelerometer dataset of the UK BioBank study. The parameters mean and standard deviation differed only slightly between groups of participants with different RMDs and selected control groups. However, conducting a paired t-test between RMD patients and their matched controls revealed a significant difference for most step parameters with mostly small effect sizes. We conclude that age, sex and Townsend scores (which we used to match the control groups) have a larger effect on individual participant step parameters than their RMD. However, if these patients' characteristics are accounted for we do find our step parameters to reflect patients' functionality in large datasets.
Date of Award1 Aug 2022
Original languageEnglish
Awarding Institution
  • The University of Manchester
SupervisorMark Lunt (Supervisor), William Dixon (Supervisor) & Suzanne Verstappen (Supervisor)


  • Smartwatches
  • eHealth
  • Step Parameters
  • Accelerometers
  • Rheumatic and Musculoskeletal Diseases
  • Gait Detection
  • Physical Functioning
  • Osteoarthritis

Cite this