Accelerometry-Based Self-Reported Fatigue Detection During Repeat Bouts of Sub-Maximal Fatiguing Exercise
Date of Award
Master of Science (MS)
Sports and Exercise Analytics
Successful sensor-based fatigue detection is an essential component of remote physical rehabilitation and activity monitoring healthcare systems. Several studies have investigated accelerometry use to detect the presence of neuromuscular fatigue pre- and post-extended bouts of exertion. Similarly, others have investigated the use of a plethora of other sensors to detect neuromuscular fatigue throughout fatiguing exercise. However, there are significant gaps in the literature concerning accelerometry-based neuromuscular fatigue detection throughout fatiguing exercise and the detection of perceived fatigue. Thus, the principal aim of this study was to develop a procedure capable of detecting perceived fatigue using data collected via wearable accelerometers during four fatiguing exercises. A secondary aim was to determine whether misclassification rates vary with proximity to a fatigue threshold. A group of seven healthy men and women each completed a protocol consisting of repeated sets of activity across seated exercises. After each set, fatigue was measured on an 11-point visual analogue scale (VAS-F). Subsequently, each set was manually segmented and paired with its corresponding VAS-F score. Features were extracted using several python-based feature engineering packages, and thereafter, a range of features were selected via ANOVA and maximum-relevance minimum-redundancy feature selection. Furthermore, two different cross-validation strategies were examined in addition to ten variations of VAS-F binning for classification targets. Classification algorithms were then used to detect bins of VAS-F scores. Lastly, the misclassification rates in the ancillary investigation were examined via a Cochran-Armitage test for trend. The most predictive model distinguished sets with a VAS-F score ≥7 from sets with a lower score with an accuracy of 89.6%. Similarly, with thresholds of VAS-F score ≥6 and VAS-F score ≥5, accuracies of 84.7% and 81.5%, respectively, were seen. In about a third of instances, the Cochran-Armitage tests showed significant trends for misclassification of false positives and negatives as a function of distance to the fatigue threshold. To conclude, we successfully predicted perceived fatigue from wearable accelerometry during fatiguing exercise. These procedures will help facilitate the construction of holistic remote healthcare systems that rely on sensor-based physiological state detection.
Available for download on Thursday, March 06, 2025