Comparison of monocular video-based methods for measuring the amplitude of hand tremor

Accurate assessment of hand tremor is critical to help diagnose and monitor multiple neurological conditions. In routine practice, clinicians subjectively estimate tremor frequency and amplitude, but this has poor reliability. Recently, objective methods for assessing hand tremor using computer vision approaches have been proposed. This work compares the accuracy of three such approaches. Tremor amplitude, in pixels, is first estimated from smartphone videos by extracting a pose skeleton and then estimating the extreme points of motion. Conversion to metric amplitude requires scene-depth information. Three approaches are investigated: (a) on-device LiDAR (b) deep-learning prediction and (c) a non-dimensional estimate of amplitude, normalising against the size of the hand. 120 videos of hands simulating rest, postural, and kinetic tremor were recorded from 20 healthy participants. Gold-standard measurements of amplitude were obtained using a within-scene ruler. We conducted Bland-Altman analysis for each approach, where bias was calculated as gold-standard - estimate. In sub-group analysis, we reported results for each tremor test and by skin tone. The bias was -0.01 cm (95% limits of agreement = ±0.78 cm) using method (a). The bias was -0.91 cm (95% limits of agreement of ±1.88 cm) for method (b). There was no statistically significant difference in amplitude accuracy, depending on skin colour as assessed using the Monk Skin Tone scale (p=0.32). Method (c) was highly correlated (r=0.977) with the gold-standard. In conclusion, depth estimates from a state-of-the-art neural network are too inaccurate for quantifying hand tremor amplitude. In contrast, LiDAR depth measurements allow hand tremor amplitude to be estimated accurately for simulated tremor.