A Feasibility Test on Preventing PRMDs Based on Deep Learning

So-Hyun Park, Sun-Young Ihm, Aziz Nasridinov, Young-Ho ParkAAAI 2019 · DOI

Introduction

Playing the piano at a professional level requires significant physical exertion, often comparable to athletic performance. Consequently, many pianists suffer from playing-related musculoskeletal disorders (PRMDs) stemming from improper posture or stiff movements. While medical experts have long identified "incorrect" posture as a primary cause of chronic injury, traditional monitoring methods have relied on intrusive wearable sensors or physical markers. These hardware-centric approaches are often expensive, cumbersome, and fail to capture the natural variety of real-world playing environments. This research explores a non-invasive, scalable alternative: using computer vision and deep learning to detect injurious postures directly from standard video footage.

Methodology

The researchers framed posture detection as a video classification task. They evaluated four distinct deep learning architectures to determine the most effective approach for identifying healthy versus injurious playing positions.

Overview of the deep learning framework for posture classification.

The models tested included:

  1. 2D CNN (Inception V3): A standard image classification model fine-tuned on individual video frames.
  2. C3D (Convolutional 3D): A model utilizing 3D filters to capture spatial and temporal features simultaneously.
  3. LRCN (Long-term Recurrent Convolutional Network): A hybrid model that feeds spatial features from a 2D CNN into an LSTM network to process temporal sequences.
  4. 2D CNN + MLP: A simplified version of the LRCN that replaces the LSTM with a Multi-layer Perceptron for final classification.

The primary goal was to classify video clips into two categories: Correct (posture that prevents injury) and Incorrect (posture likely to cause PRMDs, such as "turtle neck").

The Classical Piano Performance Posture (C3P) Dataset

A major contribution of this study is the creation of the Classical Piano Performance Posture (C3P) dataset. Unlike previous datasets recorded in controlled laboratory settings, C3P consists of 102 video clips sourced from YouTube. This ensures a high degree of environmental variety, including diverse lighting conditions, camera angles, and partial occlusions.

The dataset features 15 different players performing 17 unique music pieces. To ensure clinical accuracy, all postures were labeled by a specialist with over a decade of experience in piano education. The videos are standardized at 720 × 576 resolution and 25 fps, with an average clip length of 7.5 seconds.

Key Insights

The study revealed several critical findings regarding the application of AI to ergonomic monitoring:

  • Spatial Dominance: The 2D CNN + MLP model outperformed more complex temporal models like C3D and LRCN. This suggests that the visual "signature" of poor posture (e.g., a hunched neck) is often sufficiently captured in individual frames, and complex temporal modeling may introduce unnecessary noise on smaller datasets.
  • The Power of Transfer Learning: Utilizing a pretrained Inception V3 model allowed the researchers to achieve high accuracy despite the relatively small size of the specialized C3P dataset.
  • Expert Validation: The reliance on a veteran piano educator for ground-truth labeling ensures that the model learns distinctions that are pedagogically and clinically relevant.

Results

The feasibility test demonstrated that deep learning is highly effective at identifying injurious postures. The 2D CNN + MLP architecture emerged as the top performer, achieving an average accuracy of 87.71%.

Performance comparison of different deep learning architectures across multiple iterations.

The results indicate that a combination of robust spatial feature extraction and a straightforward classifier provides the most reliable performance for this specific task.

Conclusion and Future Work

While these results are promising, the authors categorize this work as a feasibility test. Future research will focus on expanding the C3P dataset to improve model robustness. Additionally, moving beyond binary classification to identify specific types of postural errors (e.g., shoulder tension vs. hand positioning) would provide more granular feedback. Ultimately, the goal is to develop a real-time monitoring system that can provide immediate alerts to pianists during practice sessions, helping to prevent chronic injuries before they develop.

Summary updated

BibTeX
@inproceedings{park2019feasibility,
  author = {So-Hyun Park and Sun-Young Ihm and Aziz Nasridinov and Young-Ho Park},
  title = {{A Feasibility Test on Preventing PRMDs Based on Deep Learning}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.1609/aaai.v33i01.330110005},
}
Figure