Detecting infection-related mortality using dynamical statistical indicators of high-resolution activity time series

Predicting transitions between health, disease, and death across biological systems remains an important challenge with significant implications for both ecological management and medical intervention. Although the principles underlying these transitions are increasingly recognized, accurate and tractable dynamical indicators of health-to-disease transitions remain rare, especially at the level of individual organisms. Here we show that machine learning models using slowing down indicators and permutation entropy can effectively discriminate between groups of individual Drosophila melanogaster that either live or die following experimental bacterial infection. By analysing high-resolution time-series of locomotor activity data from infected fruit flies, we find that individual dynamical indicators, such as the mean, variance, autocorrelation, and permutation entropy, did not differ markedly between flies that survived and those that died during the experiment. However, when these indicators were used to train a Random Forest model, the classifier performed well (AUC = 0.85), demonstrating an accuracy of 81.82% in discriminating between infected flies that would die from infection and those that would survive. Our findings show that combining these simple statistical indicators with machine learning enhances the ability to predict health deterioration in the Drosophila model. This integrated approach not only supports the feasibility of using slowing down indicators and permutation entropy in real-time health monitoring but also provides a framework for applying these methods to the deterioration of health in individuals in a variety of ecological and clinical environments.