Vibration-Based Non-Contact Activity Classification for Home Cage Monitoring Using a Tuned-Beam IMU Sensing Device.

Area of Science:

• Animal Behavior
• Biosensing Technology
• Machine Learning in Biology

Background:

• Accurate monitoring of laboratory animal behavior is crucial for research.
• Non-invasive, continuous monitoring methods are needed for home cage environments.
• Existing methods may be intrusive or limited in scope.

Purpose of the Study:

• To develop and validate a vibration-based, non-contact method for classifying mouse physical activity.
• To enhance signal detection using a novel tuned-beam sensing device.
• To implement a machine learning model for accurate activity classification.

Main Methods:

• A tuned-beam sensing device integrated with a six-axis inertial measurement unit (IMU) was developed to capture cage vibrations.
• Multi-level discrete wavelet transformation (MLDWT) was used for feature extraction from vibration signals.
• A convolutional neural network-long short-term memory (CNN-LSTM) model classified activities based on extracted features.
• Ground truth data was acquired using a camera-based system (EthoVision XT).

Main Results:

• The system achieved 86.81% accuracy and an average F1 score of 0.798 for classifying activities including Resting, Stationary Activity, Walking, Activity in Feeder, and Drinking over a 9-second timeframe.
• The tuned-beam sensor improved the signal-to-noise ratio (SNR) by 20 to 40 times.
• The method demonstrated reliable reproduction of long-term behavioral patterns like sleep and acclimatization, comparable to the reference method.

Conclusions:

• The proposed vibration-based monitoring system offers a low-cost, non-contact solution for accurate mouse activity classification in home cages.
• This technology enables effective long-term behavioral monitoring in husbandry settings.
• The method has the potential to advance research by providing detailed insights into animal welfare and behavior.