Prediction of the infecting organism in peritoneal dialysis patients with acute peritonitis using interpretable Tsetlin Machines

The analysis of complex biomedical datasets is becoming central to understanding disease mechanisms, aiding risk stratification and guiding patient management. However, the utility of computational methods is often constrained by their lack of interpretability and accessibility for non-experts, which is particularly relevant in clinically critical areas where rapid initiation of targeted therapies is key. To define diagnostically relevant immune signatures in peritoneal dialysis patients presenting with acute peritonitis, we analysed a comprehensive array of cellular and soluble parameters in cloudy peritoneal effluents. Utilising Tsetlin Machines (TMs), a logic-based machine learning approach, we identified pathogen-specific immune fingerprints for different bacterial groups, each characterised by unique biomarker combinations. Unlike traditional ‘black box’ machine learning models such as artificial neural networks, TMs identified clear, logical rules in the dataset that pointed towards distinctly nuanced immune responses to different types of bacterial infection. This demonstrates unambiguously that even when infecting the same anatomical location and causing clinically indistinguishable symptoms, each type of pathogens interacts in a specific way with the body’s immune system. Importantly, these immune signatures could be easily visualised to facilitate their interpretation, thereby not only enhancing diagnostic accuracy but also potentially allowing for rapid, accurate and transparent decision-making based on the patient’s immune profile. This unique diagnostic capacity of TMs could help deliver clear and actionable insights such as early patient risk stratification and support early and informed treatment choices in advance of conventional microbiological culture results, thus guiding antibiotic stewardship and contributing to improved patient outcomes. All underlying tools for the present analysis are available at https://github.com/anatoliy-gorbenko/biomarkers-visualization. The anonymised patient data underlying this article will be shared on reasonable request to the corresponding authors.