Objective: We aimed to apply a deep learning neural network directly to video of finger tapping, without human-defined measures or features, for a computer to learn its own patterns that distinguish people with idiopathic Parkinson’s disease (PD) from controls.

Background: The core movement sign of PD is bradykinesia. A classic test for this is finger tapping, in which a clinician observes a person repetitively tap finger and thumb together. This requires an expert eye, a scarce resource, and even experts show considerable variability and inaccuracy. Previous technology applied to finger tapping has been limited to one-dimensional measures of tapping, with specific researcher-defined features derived from those measures.

Method: 152 smartphone videos of 10s finger tapping were collected from 40 people with PD and 37 healthy controls. We down-sampled pixel dimensions and videos were split into 1 second clips. A 3D convolutional neural network was trained on these clips.

Results: For discriminating PD from controls, our model showed training accuracy 0.91, and test accuracy 0.69, with test precision 0.73, test recall 0.76 and test AUROC 0.76. In addition, we report class activation maps for the five most predictive features to show the spatial and temporal parts of each video that the network focuses attention to make a prediction, including an apparent dropping thumb movement distinct for PD.

Conclusion: A deep learning neural network can be applied directly to standard video of finger tapping, to distinguish PD from controls, without a requirement to extract a one-dimensional signal from the video, or pre-define tapping features.

« Back to 2022 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/deep-learning-to-distinguish-parkinsons-from-controls-in-video-without-human-defined-measures/