Objective: Using a deep learning inspired approach, we developed a tool to compute an objective score of bradykinesia.

Background: Bradykinesia is defined as a motor slowness and is associated with decrement of the amplitude and/or the speed of movement. Bradykinesia is a key parkinsonian feature yet subjectively assessed by the MDS-UPDRS score making reproducible measurements and follow-up challenging.

Method: A large database of videos showing parkinsonian patients performing MDS-UPDRS protocols has been acquired in a Movement Disorder unit. We applied a detection algorithm based on the existing DeepLabCut [1] software to detect 21 different and characteristic points of the hand on a 2-d projection. Another deep learning approach is then used to transpose this 2-d projection on a 3-d hand model, leading to a full 3-d geometrical description of the 21 points as a function of time.

Results: We analyzed separately all three tests of upper limb bradykinesia as described in the MDS-UPDRS. Firstly, for the “finger tapping” protocol, we computed the geometrical distance between the tips of the index and the thumb, leading to a precise detection of the amplitude, speed and acceleration of the tapping. Then, for the “hand movements”, we analyzed the speed and amplitude at which the patient performed successive fist openings and closings, to have a precise estimation of its evolution over time. Finally, the “pronation-supination movements of hands” are assessed by the wrist rotation angle, computed thanks to the 3-d position of several key hand joints.

Conclusion: This tool appears to be useful in several ways. First of all, the precise measurement of bradykinesia leads to a better ranking of its severity and a more accurate assessment of its response to treatment. Moreover, this tool might confirm bradykinesia suspected at an early stage since it allows objective measures compared to clinical examination. A better assessment of parkinsonian patients might be useful to include patients earlier in clinical trials (for instance, testing neuroprotective agents) and to perform an objective clinical follow-up. It could also be useful to develop it as a tool, allowing physicians (including neurologists and non neurologists) to analyze movement and make a diagnosis in difficult cases.

In conclusion, we developed a tool to make bradykinesia analysis objective and more precise. Further improvements and studies will be necessary to define the full extent of this approach.

References: [1] Mathis, A. et al. DeepLabCut: markerless pose estimation of user-defined body parts with deep learning. Nature Neuroscience 21, 1281 (2018).

« Back to MDS Virtual Congress 2020

MDS Abstracts - https://www.mdsabstracts.org/abstract/movement-disorders-analysis-using-a-deep-learning-approach/