筋活動パターンから行動状態を解読する深層学習

Accurate assessment of animal behavior is limited by manual video observation. Here, we demonstrate that deep learning-based analysis of multi-site electromyograms enables classification of behavioral states. Electromyograms were recorded from five muscles in the limbs and neck of mice alongside video monitoring to establish ground-truth labels (i.e., walking, grooming, rearing). A custom convolutional neural network was trained on electromyogram segments. Our model achieved robust classification accuracy, demonstrating the effective detection of behavioral patterns. Our electromyogram-based model provides an objective, automated, and scalable framework for behavioral classification, which can be seamlessly integrated with video monitoring workflows.