Video-based diagnosis support system for pianists with Musician’s dystonia

T. Oku, Shinichi Furuya, André Dong Lee, Eckart AltenmüllerFrontiers in Neurology 2024 · DOI

Abstract

These findings demonstrate the effectiveness of the present technique in aiding in the accurate diagnosis of Musician's dystonia.

Introduction

Musician’s dystonia (MD) is a career-threatening, task-specific movement disorder affecting approximately 1–2% of professional musicians. It typically manifests as involuntary finger curling or "sticking" during complex performances. Diagnosis is notoriously difficult for general neurologists because the symptoms often only appear during specific, high-level musical tasks. Traditional objective diagnostic tools require expensive MIDI-equipped instruments and specialized software, creating a barrier to early detection. There is a critical need for a low-cost, accessible "Red Flag" screening tool that utilizes standard video equipment.

Methods

Researchers developed a diagnostic system using a standard USB camera and unsupervised machine learning. The approach consists of two primary layers:

1. Anomaly Detection via Convolutional Autoencoder

The system utilizes MediaPipe to extract 3D coordinates for 21 anatomical landmarks on the hand from video footage. A convolutional autoencoder was trained exclusively on "normal" hand kinematics from 20 healthy professional pianists. The model learns to compress and reconstruct healthy movement patterns. When presented with dystonic movement, the reconstruction error increases significantly, serving as a diagnostic indicator.

2. Biomechanical Assessment

In addition to video analysis, custom force sensors measured finger strength, agility, and independence to identify physiological markers beyond the keyboard.

Experimental Design

The study involved 36 professional pianists (20 healthy controls and 16 diagnosed with MD). Participants performed 11 specific pianistic tasks designed to trigger symptoms, including triad chords and scale-like passages.

Example of a triad chord task used for diagnostic screening.
Example of a scale-like task with specific fingering requirements.

Results

The system demonstrated high diagnostic accuracy, particularly during tasks that require complex hand spans or maneuvers.

  • Task Sensitivity: The model was most effective during triad chords and arpeggios involving black keys. These tasks often involve the "thumb-under" maneuver, which is highly sensitive to dystonic interference.
  • Accuracy: For these sensitive tasks, the algorithm achieved 91% sensitivity and 95% specificity using only video data.
  • Biomechanical Findings: MD patients exhibited significantly higher asymmetry in finger independence and degraded force exertion in their affected hands compared to healthy controls, even in non-musical contexts.

Discussion

The use of unsupervised learning addresses the challenge of data scarcity in rare disorders like MD. By training only on healthy data, the model is highly scalable. However, limitations remain, including a demographic mismatch between the healthy control and MD groups. Future work will focus on developing a "severity score" to track rehabilitation progress and integrating explainable AI techniques to identify specific joint abnormalities.

Conclusion

This video-based approach provides a viable, low-cost screening tool for Musician’s dystonia, potentially enabling earlier diagnosis and intervention for affected professionals.

Summary updated

Abstract

These findings demonstrate the effectiveness of the present technique in aiding in the accurate diagnosis of Musician's dystonia.
BibTeX
@inproceedings{oku2024video,
  author = {T. Oku and Shinichi Furuya and André Dong Lee and Eckart Altenmüller},
  title = {{Video-based diagnosis support system for pianists with Musician’s dystonia}},
  booktitle = {Frontiers in Neurology},
  year = {2024},
  doi = {10.3389/fneur.2024.1409962},
}
Figure