Background & Aims

Care providers are challenged by the high prevalence, high complexity, and the need for biopsychosocial approaches for primary chronic pain (PCP). Physical activity (PA) behavior is an important influential factor in chronic pain. However, caregivers have difficulty in assessing PA behavior and associations between subjective (e.g. questionnaires) and objective measures (e.g. accelerometers) are inconsistent. No obvious differences are found in total PA in chronic pain versus healthy controls, except for activity distribution. This emphasizes the need for detailed assessment of PA patterns, that is, the structure of low to vigorous physical activity and sedentary behavior over time. No studies exist with validation of intensity algorithms on real live data. The aim of this study is to develop and validate a machine learning algorithm to convert accelerometer data to detailed time series of daily activity intensity in PCP patients (CPP).

Methods

The CNN+LSTM machine learning model (MLM) was trained and tested with data from 30 healthy participants (HP) and tested in the second stage with data from 21 CPP awaiting interdisciplinary multimodal pain treatment (IMPT). Each participant performed 22 daily life activities in bouts of 80 seconds with a wrist-worn tri-axial accelerometer with a sampling frequency of 12.5 Hz. Each bout was labelled as sedentary, low, moderate or vigorous intensity according to metabolic equivalent of tasks of <1.5, 1.5-3.0, 3.0 to 6 and ?6. The MLM was trained with the labeled accelerometer data from 27 HP with a moving window of 12 seconds and 50% overlap. Data from 3 HP were set aside for testing the model. Results of training and testing with raw and balanced data were compared. Secondly, the MLM was tested with data from 21 CPP. The MLM was applied to real-life timeseries of accelerometer data from CPP. Barcoding was used to visualize and validate the results.

Results

The MLM is accurate in classifying activity intensity in HP data with overall accuracy of 89% and accuracy in the four intensity classes ranging from 71% to 91%. Precision ranges from 0.68 to 0.93. The most classification confusion appears between the classes of low and moderate intensity. The model performs better after balancing with overall accuracy of 89%, ranging from 79% to 93% for the four classes and precision 0.81 to 0.99. Preliminary testing of the model with CPP-data revealed similar values. Applying the model to accelerometer timeseries produces plausible timeseries of activity intensity when compared with diary data. Plotting the timeseries as barcodes yields a clear view of activity patterns that are intuitively to understand.

Conclusions

A CNN+LSTM MLM is able to accurately and precisely classify data from a single wrist-worn tri-axial accelerometer in four classes of activity intensity. The MLM is applicable, accurate and precise in subjects with and without primary chronic pain. When comparing timeseries of classified activity intensities with diaries the model output it can be concluded that the MLM might have external validity.
Coming research will focus on further investigating external validity of the model output on real-live accelerometer timeseries, testing outcome measures that might represent different activity patterns, determining discriminant validity (CPP versus HP) and associations with psychosocial factors and responsiveness.

References

Ng W, Slater H, Starcevich C, Wright A, Mitchell T, Beales D. Barriers and enablers influencing healthcare professionals’ adoption of a biopsychosocial approach to musculoskeletal pain: a systematic review and qualitative evidence synthesis. Pain. 2021;162(8):2154-2185. doi:10.1097/j.pain.0000000000002217

Treede RD, Rief W, Barke A, et al. A classification of chronic pain for ICD-11. Pain. 2015;156(6):1003-1007. doi:10.1097/j.pain.0000000000000160

Ridgers, N. D., Denniss, E., Burnett, A. J., Salmon, J., & Verswijveren, S. J. J. M. (2023). Defining and reporting activity patterns: a modified Delphi study. International Journal of Behavioral Nutrition and Physical Activity, 20(1). https://doi.org/10.1186/s12966-023-01482-6

Huijnen IPJ, Verbunt JA, Roelofs J, Goossens M, Peters M. The disabling role of fluctuations in physical activity in patients with chronic low back pain. European Journal of Pain. 2009;13(10):1076-1079. doi:10.1016/j.ejpain.2008.12.008

Almeida Mendes de, M., da Silva, I. C. M., Ramires, V. V., Reichert, F. F., Martins, R. C., & Tomasi, E. (2018). Calibration of raw accelerometer data to measure physical activity: A systematic review. In Gait and Posture (Vol. 61, pp. 98–110). Elsevier B.V. https://doi.org/10.1016/j.gaitpost.2017.12.028

Presenting Author

Annet Doomen

Poster Authors

Annet Doomen

MSc

University of Applied Sciences, research group Lifestyle and Health, Utrecht, The Netherlands

Lead Author

E-mail

Xiaowen Song MSc

University of Applied Sciences, research group Lifestyle and Health, Utrecht, The Netherlands

Lead Author

E-mail

Michiel Punt Dr

University of Applied Sciences, research group Lifestyle and Health, Utrecht, The Netherlands

Lead Author

E-mail

Richard Felius

University of Applied Sciences, research group Lifestyle and Health, Utrecht, The Netherlands

Lead Author

Rob Smeets

Maastricht University, Department of Rehabilitation Medicine, Research School CAPHRI, Maastricht, The Netherlands

Lead Author

E-mail

Albère Köke Dr

Maastricht University, Department of Rehabilitation Medicine, Research School CAPHRI, Maastricht, The Netherlands

Lead Author

E-mail

Harriet Wittink Prof Dr

University of Applied Sciences, research group Lifestyle and Health, Utrecht, The Netherlands

Lead Author

E-mail

Topics

  • Assessment and Diagnosis

Download Poster