Objective: To investigate whether computer vision algorithms can detect bradykinesia using standard smartphone video alone.

Background: Bradykinesia is the core clinical feature of Parkinson’s disease is bradykinesia.  However the detection of bradykinesia relies on a subjective visual judgement made by the clinician.  Inter-rater reliability and diagnostic accuracy are moderate at best.  Existing objective methods to detect or measure bradykinesia require either a specific piece of hardware, or they require patient interaction with an app or website.  Thus, none are in routine clinical use.

Method: 70 videos of finger tapping were recorded using a standard integrated smartphone camera (40 patient hands, 30 control hands).  Computing algorithms were used to automatically separate hand from background and then measure the movement of pixels representing the hand.  The pixel movement signals were then further processed to extract features likely to be related to finger-tap frequency, amplitude and rhythm (e.g. fast fourier transform, jitter and peak-to-peak variability).  The videos were also blindly rated by two movement disorder specialists.

Results: Using support vector machine analysis, we predict the presence of bradykinesia with accuracy of 0.74 (sensitivity: 0.85, specificity: 0.65) and the presence of a Parkinson’s diagnosis with accuracy of 0.67 (sensitivity = 0.60, specificity = 0.77).

Conclusion: Although our results are preliminary and the data set small, they show an accuracy similar to that recorded by blinded human clinical experts, and our contactless method requires no special equipment, and no patient interaction with an app or website.

« Back to 2019 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/computing-algorithms-can-detect-bradykinesia-from-smartphone-videos/