Machine learning scoring reveals increased frequency of falls proximal to death in Drosophila melanogaster

Abstract

Falls are a significant cause of human disability and death. Risk factors include normal aging, neurodegenerative disease, and sarcopenia. Drosophila melanogaster is a powerful model for study of normal aging and for modeling human neurodegenerative disease. Aging-associated defects in Drosophila climbing ability have been observed to be associated with falls, and immobility due to a fall is implicated as one cause of death in old flies. An automated method for quantifying Drosophila falls might facilitate the study of causative factors and possible interventions. Here, machine learning methods were developed to identify Drosophila falls in video recordings of 2D movement trajectories. The study employed existing video of aged flies as they approached death, and young flies subjected to lethal dehydration/starvation stress. Approximately 9,000 frames of video were manually annotated using open-source tools and used as the training set for You Only Look Once (YOLOv4) software. The software was tested on specific hours within a 22 hour video that was originally manually-annotated for number of falls per hour and corresponding timestamps. The model predictions were evaluated against the manually-annotated ground truth, revealing a strong correlation between the predicted and actual falls. The frequency of falls per hour increased dramatically 2-4 hours prior to death caused by dehydration/starvation stress, whereas extended periods of increased falls were observed in aged flies prior to death. This automated method effectively quantifies falls in video data without observer bias, providing a robust tool for future studies aimed at understanding causative factors and testing potential interventions.