Figure 3.

Comparison of the Performance of Models Using Records from Different Devices, Processing Methods, and Pulling Methods

(A–C) Comparison of AUROCs between models using: (A) either accelerometer and gyroscope data; (B) the sum of the squared values of x, y, z axes and 3D data as input; (C) different pulling methods: on record level (by record), on individual level using average prediction across records of each individual (by mean), and on individual level using maximum prediction across all records of each individual (by max) in 5-fold cross-validations.

(D–F) Pairwise comparison between AUROCs of models mentioned in (A) to (C) in bootstrapping. No significant difference between accelerometer and gyroscope input data was observed, while more consistent improvement was observed by 3D input than summed square and by maximal prediction at the individual level than prediction at record level or the mean prediction at the individual level (p < 1 × 10⁻⁶).