Deep learning of Parkinson's movement from video, without human-defined measures

Area of Science:

• Neurology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Parkinson's disease (PD) diagnosis relies on bradykinesia, often assessed via finger tapping.
• Current finger tapping assessments require expert observation, which is scarce and prone to variability.
• Existing technological approaches use limited, researcher-defined features from tapping signals.

Purpose of the Study:

• To apply a deep learning neural network directly to finger tapping videos for PD classification.
• To assess the accuracy of AI in distinguishing idiopathic PD from controls.
• To visualize the learned features of the deep learning model.

Main Methods:

• Collected 152 smartphone videos of finger tapping from 40 PD patients and 37 controls.
• Processed videos by down-sampling and splitting into 1-second clips.
• Trained a 3D convolutional neural network on the video clips.

Main Results:

• The deep learning model achieved a test accuracy of 0.69 for discriminating PD from controls.
• The model demonstrated a test precision of 0.73 and test recall of 0.76.
• Class activation maps identified distinct spatial and temporal features, including a unique thumb movement in PD patients.

Conclusions:

• Deep learning can directly analyze finger tapping videos to differentiate PD from controls.
• This AI method bypasses the need for manual feature extraction or expert interpretation.
• The study presents a novel, technology-driven approach for PD assessment.