Machine Learning-Driven Localization of Infection Sources in the Human Cardiovascular System

In vivo localization of infection sources is essential for effective diagnosis and targeted disease treatment. In this work, we leverage machine learning models to associate the temporal dynamics of biomarkers detected at static gateway positions with different infection source locations. In particular, we introduce a simulation that models infection sources, the release of biomarkers, and their decay as they flow through the bloodstream. From this, we extract time-series biomarker data with varying decay rates to capture temporal patterns from different infection sources at specific gateway positions. We then train a stacked ensemble model using LightGBM and BernoulliNB to analyze biomarker time-series data for classification. Our results reveal that higher biomarker degradation rates significantly reduce the localization accuracy by limiting the biomarker signal detected at the gateways. A fivefold increase in decay rate lowers the mean cross-validation accuracy from 92to 66. This effect is more pronounced for infection sources located farther from the gateways, e.g., the kidneys. Due to the longer distance, more biomarkers degrade before reaching the wrist-located gateways, leading to a substantial decline in classification performance.