Do changes in feature selection parameters influence the classification of knee rehabilitation exercises when using body worn accelerometer data?

Abstract

Results:The inter-rater agreements for MOAKS data (Figures 2.A.1, 2.B.1 & 2.C.1) were 1) for raw grades: 0.69 kappa and 0.78 balanced accuracy (patella), 0.71 kappa and 0.83 balanced accuracy (femur), 0.72 kappa and 0.79 balanced accuracy (tibia), 2) for 2 classes: 0.68 kappa and 0.86 balanced accuracy (patella), 0.69 kappa and 0.88 balanced accuracy (femur), 0.69 kappa and 0.87 balanced accuracy (tibia) and 3) for 3 classes: 0.70 kappa and 0.85 balanced accuracy (patella), 0.69 kappa and 0.86 balanced accuracy (femur), and 0.73 kappa and 0.87 balanced accuracy (tibia). Harmonized MOAKS and mWORMS with a heuristic approach (Figures 2.A.2, 2.B.2 & 2.C.2) achieved 1) for 2 classes, a balanced accuracy of: 0.83 for patella, 0.79 for femur, and 0.71 for tibia, and 2) for 3 classes: 0.79 balanced accuracy for no lesion class, 0.72 for PT class, and 0.87 for FT class (patella), 0.84 balanced accuracy for no lesion class, 0.74 for PT class, and 0.85 for FT class (femur), and 0.82 balanced accuracy for no lesion class, 0.68 for PT class, and 0.88 for FT class (tibia). Harmonized MOAKS and mWORMS with a machine learning approach (Figures 2.A.3, 2.B.3 & 2.C.3) achieved 1) for 2 classes, a balanced accuracy of: 0.80 for patella, 0.79 for femur, 0.77 for tibia, and 2) for 3 classes: 0.82 balanced accuracy for no lesion class, 0.72 for PT class, and 0.84 for FT class (patella), 0.81 balanced accuracy for no lesion class, 0.66 for PT class, and 0.8 for FT class (femur) and 0.8 balanced accuracy for no lesion class, 0.62 for PT class, and 0.8 for FT class (tibia). Conclusions: Harmonization of SQ assessments combines data graded with different scaling systems into a single homogenized source. A simplified two or three class scale reduces variability and improves the reliability of annotations for deep learning models. Given sufficiently balanced data, a machine learning approach allows for a more balanced sensitivity and specificity, and is less susceptible to inserting bias in labeled data. Harmonization performance is similar to inter-rater variability, and that machine learning techniques can be used to harmonize multiple data sources.