P 57 Markerless, video-based kinematic Tremor Analysis: machine vision in movement disorders. (May 2022)

• Record Type: Journal Article
• Authors: Rönn, A.J.; Friedrich, M.; Palmisano, C.; Rhodes, E.; Peach, R.; Reich, M.M.
• Is Part Of: Clinical neurophysiology. Volume 137(2022)
• Page Start: e48
• Publication Date: 2022-05
• Subjects: Neurophysiology -- Periodicals; Electroencephalography -- Periodicals; Electromyography -- Periodicals; Neurology -- Periodicals; 612.8
• DOI: 10.1016/j.clinph.2022.01.088
• Languages: English
• ISSNs: 1388-2457

Abstract

Abstract : Introduction: Clinimetrics of movement disorders relies on scales such as the Fahn-Tolosa-Marin Tremor Rating Scale for Essential Tremor (ET). However, these rating scales are prone to rater bias. Quantitative measurements of tremor kinematics, which are proxy measures of clinical outcome, can be achieved with gold standard tremor analyses using reflective markers and several special cameras. Due to their resource-intensiveness, broad availability is lacking. We therefore asked whether new developments in the field of machine vision can close the gap between granular clinimetrics and simplicity. Patients & Methods: Using the open-source software DeepLabCut (DLC), we tuned a recurrent convolutional neural network (RCNN) to track anatomic landmarks of the upper body by annotating a total of 1.878 frames from 234 videos of 51 ET patients who had undergone deep brain stimulation for ET and 10 healthy subjects. In a retrospective subcohort of 43 patients, tremor amplitudes were derived using the RCNN predictions. We probed the influence of DBS condition on tremor kinematics and expert annotated tremor rating scores (TRS) using repeated measures (rm-)ANOVA and Wilcoxon signed-ranked tests. Then, we assessed correlation of tremor metrics with TRS. To prospectively validate the method, we calculated the correlation of tremor metrics derived from both the video-based RCNN and the marker-based gold standard approach in 7 patients. The videos for validation were not used for training and were unknown to the network. To test for precision, we assessed relative errors and equivalence of both methods using two one sided T-Tests (TOST). Results: The RCNN achieved a mean Euclidean train/test error of 3.56/10.74 pixels corresponding to the average pixel size of a fingertip. In the retrospective analysis, DBS exerted a significant effect on the mean tremor amplitudes (Friedman test: χ2(3)=45.3, p<.001) and TRS (right: W=7.0, p<.001; left: W=37.5, p<.001). RCNN-derived maximum amplitudes correlated strongly with TRS (right: rs=.789, p<.001; left: rs=.640, p<.001). In the prospective analysis, mean amplitudes derived from both methods were strongly correlated (DBS off: rs=.982, p<.001; DBS on: rs=.824, p<.001). Relative errors of RCNN-derived mean amplitudes from goldstandard measurements did not significantly deviate from zero (p>.06). TOST revealed RCNN-derived measurements to be partially (DBS off: upper bound T(13)=0.5, p=.7; lower bound T(13)=3.9, p<.001) and fully (DBS on: T(13)=-15.2, p<.001; T(13)=14.52, p<.001) located within 90% CI equivalence bounds. Conclusion: We demonstrate the utility of video-based markerless tremor analysis using an RCNN for both retrospective and prospective settings. Our method allows sufficiently accurate measurements of tremor characteristics to assess the severity of the disease and to map individual disease progression.