Fig. 6: Application of a disease-specific convolutional neural network for postural tremor analysis across cohorts.
From: Validation and application of computer vision algorithms for video-based tremor analysis
a–c In the prospective cohort, DLC-RCNN-derived amplitude measurements are strongly correlated to clinical scores (ρ = 0.72, p < 0.001) and motion capture (ρ = 0.88, p < 0.001). Mean absolute error is 2.55 mm (95% CI [−2.11, 7.29]) with no systematic relationship to measurement magnitudes. d, e DLC-RCNN frequency measurements are moderately correlated to accelerometer (r = 0.44, p < 0.05) with a mean absolute error of −0.69 Hz [95% CI −0.93, 0.44]. f In the retrospective cohort, DLC-RCNN-derived postural tremor amplitudes show a moderate correlation to clinical scores (ρ = 0.72, p < 0.001). DLC-RCNN however failed to capture and measure kinetic tremor in both cohorts.
