Related Experiment Video
Updated: Sep 18, 2025
08:05
Design and Analysis for Fall Detection System Simplification
Published on: April 6, 2020
10.8K
Benchmarking Accelerometer and CNN-Based Vision Systems for Sleep Posture Classification in Healthcare Applications
Minh Long Hoang1, Guido Matrella1, Dalila Giannetto1
1Department of Engineering and Architecture, University of Parma, 43124 Parma, Italy.
Sensors (Basel, Switzerland)
|June 27, 2025
Summary
Accelerometer-based sleep position recognition significantly outperforms image-based methods, achieving over 99.8% accuracy for diagnosing health conditions. This robust approach is ideal for real-time, privacy-sensitive healthcare monitoring.
Area of Science:
- Biomedical Engineering
- Machine Learning in Healthcare
- Wearable Sensor Technology
Background:
- Sleep position monitoring is vital for diagnosing conditions like sleep apnea and pressure ulcers.
- Accurate body posture data aids clinical insights and intelligent healthcare system development.
- Existing methods require comparison for optimal clinical application.
Purpose of the Study:
- To comparatively analyze sleep position recognition using image-based deep learning and accelerometer-based classification.
- To evaluate the performance of Visual Geometry Group 16 (VGG16) and feedforward neural networks for sleep posture detection.
- To determine the most accurate and robust method for clinical and in-home healthcare monitoring.
Main Methods:
- Image-based recognition utilized a fine-tuned VGG16 convolutional neural network with data augmentation (rotation, reflection, scaling, translation).
- Accelerometer-based classification employed a feedforward neural network trained on extracted features (signal sum, standard deviation, max, spike count).
- Both methods classified five positions: prone, supine, right side, left side, and wake up.
Main Results:
- The image-based VGG16 model achieved 93.49% accuracy, with perfect precision/recall for 'right side' and 'wakeup'.
- The accelerometer-based method exceeded 99.8% accuracy across most positions, demonstrating superior performance.
- 'Wake up' detection was highly accurate due to the absence of physiological movement signals.
Conclusions:
- Accelerometer-based classification offers higher precision and robustness compared to image-based deep learning for sleep position recognition.
- The accelerometer approach is particularly suitable for real-time and privacy-sensitive healthcare monitoring applications.
- Results provide insights for selecting appropriate sleep monitoring technology in clinical, in-home, or embedded systems.
