Jonas Dorn, Caroline Perraudin, Vittorio Paolo Illiano, Oliver Sander
Novartis Pharma AG, Basel
Objective
This work aims to study the relationship between the data collected by wrist-worn activity sensors during an unsupervised scheduled activity, the 5x sit-to-stand test (5xSTS), and electronic patient reported pain and stiffness outcomes (ePROs). It further investigates the feasibility and analyzability of the 5xSTS test.
Data
- 45 subjects participated in the study. Of those, 30 were arthritis patients (18 with rheumatoid arthritis, 10 with osteoarthritis, 2 with psoriatic arthritis), and 15 healthy volunteers
- Subjects were provided with a smartphone application to capture daily measures of pain/stiffness (ePRO) and with a wrist-worn activity sensor for a period of 4 weeks (Actigraph Link)
- Participating subjects were followed for 4 weeks. Unsupervised sit-to-stand tests were performed Mondays/Wednesdays/Fridays after getting up.
Methods
Data processing and 5xSTS detection
Raw accelerometer Actigraph data (sampled at 30 Hz) were post-processed in an automated pipeline including several correction steps (battery depletion), calibration (standardize acceleration during rest state), and transformation (Cartesian to spherical).
5xSTS time windows were extracted from the post-processed acceleration data in several selection stages. Selection was done by time of day (5xSTS performed in the morning), angle of arm crossed in front of chest, alternating elevation patterns, and specific acceleration patterns. The search for acceleration patterns is performed using regular expressions of consecutive strings of low/medium/high accelerations of varying lengths.
Correlation of 5xSTS duration and ePROs
Correlations between 5xSTS and daily ePRO responses on pain and stiffness were evaluated by linear mixed effects regression with the ePRO responses as dependent variables. Varying intercept as well as varying intercept & slope models were tested. The mixed effects regression approach allows accounting for correlations between observations from the same subjects, as well as assessing the contribution of additional covariates.
Validation
The overall data set was split into development, test, and validation set. Tuning of the detection procedure as well as the linear mixed effects modeling were performed on the development set and then checked against the other independent sets.
Results – Compliance
Compliance is bimodal (average 53%). No differences in compliance were found between healthy volunteers and patients. Most subjects with low compliance were highly compliant until they quit. Subjects performed the 5xSTS in a consistent and per-protocol manner in both supervised and unsupervised settings.
Results – 5xSTS « ePRO
Linear mixed-effects models are able to take into account individual patients’ differences in pain threshold and fitness. Prediction discrepancy diagnostics do not show evident bias. The 5xSTS duration correlates with subjects’ self-reported pain/stiffness, particularly with the responses given in the morning. Estimated coefficients are robust throughout the analysis.
Conclusions
- Unsupervised scheduled activities like the 5xSTS test are feasible in studies with acceptable compliance even with no strong reminders or real-time checks
- A semi-automatic workflow facilitates detection, data extraction and analysis from raw sensor data
- Device calibration: Even medical grade sensors need to be re-calibrated (consumer-grade sensor data may not allow re-calibration)
Reference: PAGE 27 (2018) Abstr 8487 [www.page-meeting.org/?abstract=8487]
Poster: Methodology - Other topics