Predicting physical activity intensity using raw accelerometer signals in manual wheelchair users with spinal cord injury

• Published in: Spinal cord 2022-02, Vol.60 (2), p.149-156
• Main Authors: Shwetar, Yousif, Huang, Zijian, Veerubhotla, Akhila, Knezevic, Steven, Hong, EunKyoung, Spungen, Ann M., Ding, Dan
• Format: Article
• Language: English
• Subjects: 692/699/75; 692/700/2817; Accelerometers; Accelerometry - methods; Anatomy; Biomedical and Life Sciences; Biomedicine; Correlation coefficient; Correlation coefficients; Cross-Sectional Studies; Exercise; Exercise - physiology; Human Physiology; Humans; Neurochemistry; Neuropsychology; Neurosciences; Performance prediction; Physical activity; Research facilities; Research institutions; Spinal cord injuries; Spinal Cord Injuries - diagnosis; Wheelchairs; Wrist
• ISSN: 1362-4393; 1476-5624
• DOI: 10.1038/s41393-021-00728-z

Study design Cross-sectional validation study. Objectives The performance of previously published physical activity (PA) intensity cutoff thresholds based on proprietary ActiGraph counts for manual wheelchair users (MWUs) with spinal cord injury (SCI) was initially evaluated using an out-of-sample dataset of 60 individuals with SCI. Two types of PA intensity classification models based on raw accelerometer signals were developed and evaluated. Setting Research institutions in Pittsburgh PA, Birmingham AL, and Bronx NY. Methods Data were collected from 60 MWUs with SCI who followed a structured activity protocol while wearing an ActiGraph activity monitor on their dominant wrist and portable metabolic cart which measured criterion PA intensity. Data was used to assess published models as well as develop and assess custom models using recall, specificity, precision, as well as normalized Mathew’s correlation coefficient (nMCC). Results All the models performed well for predicting sedentary vs non-sedentary activity, yielding an nMCC of 0.87–0.90. However, all models demonstrated inadequate performance for predicting moderate to vigorous PA (MVPA) with an nMCC of 0.76–0.82. Conclusions The mean absolute deviation (MAD) cutoff threshold yielded the best performance for predicting sedentary vs non-sedentary PA and may be used for tracking daily sedentary activity. None of the models displayed strong performance for MVPA vs non-MVPA. Future studies should investigate combining physiological measures with accelerometry to yield better prediction accuracies for MVPA.