Exploring parameter optimisation in machine learning algorithms for locomotor task discrimination using wearable sensors

The accurate identification of locomotion states from wearable sensor data using machine learning relies heavily on carefully selecting algorithm parameters, which remains a challenging task. This study systematically optimised key parameters-including window length, sampling frequency, temporal resolution, overlapping value, and normalisation effects-to enhance the accuracy of machine learning models for distinguishing different locomotor tasks. Our study was conducted on participants (N = 35, 19 ♂10 ♀, 27.4 ± 26.5 years, 1.74 ± 0.8 m, 71.5 ± 11.3 kg) who wore accelerometers on the sacrum, thighs and shanks. Principal component and discriminant function analyses were applied to acceleration data from three locomotor tasks: self-selected slow, normal and fast walking. The parameters explored for the optimisation of the algorithm were accelerometer window length, sampling frequency, spectral temporal resolution, overlapping value, and accelerometer amplitude normalisation effects. Unnormalised data, with longer feature window lengths and decreasing temporal resolutions, yielded the highest quality discrimination. Setting the sampling rate to 40 Hz and overlapping value to 66% provided optimal discrimination. Baseline results highlight that the sacrum is the best-performing location, yet optimal (longer) window lengths, and optimal (shorter) temporal resolutions change the best-performing sensor attachment location to the shanks. Specific values of parameters were found to be optimal for our study, and these results can guide manufacturers, engineers, and researchers in designing wearable devices and machine learning algorithms that more effectively identify locomotor tasks. Practitioners and clinicians may also use these findings to select appropriate tools or methodologies tailored to their specific research or clinical objectives.