Automatic medical decision for diagnosis of infectious diseases based on artificial intelligence approaches

Abstract

Infectious diseases present complex diagnostic challenges due to the overlapping clini- cal caused by diverse pathogens. Meningitis, in particular, remains a sig-nificant global health concern due to its high morbidity and mortality, especially when diagnosis and treatment are delayed. Traditional diagnostic methods often involve in-vasive procedures and extensive laboratory testing, which can be time-consuming and resource-intensive. This Ph.D. research investigates the integration of artificial intelli-gence (AI) into the diagnostic process, aiming to enhance accuracy, speed, and inter-pretability through the use of explainable AI (XAI) techniques.The first phase of this study examines cerebrospinal fluid (CSF) biomarker vari-ations across different age groups—children, adults, and the elderly—within various types of meningitis. By analyzing these patterns, we aim to improve the understand-ing of diagnostic and clinical and their implications for treatment strategies. This analysis establishes a foundational understanding of how biomarkers behave in different populations and infection contexts. Our next contribution focuses on diagnosing multiple meningitis types using ensem- ble models and SHapley Additive exPlanations (SHAP) to interpret feature . Using data from Setif Hospital (Algeria) and Brazil’s SINAN database, we validated our findings across diverse populations. Extreme Gradient Boosting achieved strong performance (accuracy: 0.90, AUROC: 0.94, F1-score: 0.98). SHAP revealed distinct biomarker profiles such as elevated in meningococcal, high in tuberculous, and dominance in H. influenzae , along with clin-ically relevant diagnostic patterns. These results highlight the model’s ability to distin- bacterial, viral, and pathogen-specific meningitis, increasing trust in AI-driven diagnostics. Our third develops specialized models for meningococcal meningitis, emphasizing local explainability for precise diagnosis. We tested several models on 934 cases, with gradient boosting performing best (accuracy: 0.88, AUROC: 0.93, F1-score: 0.87). Using XAI tools like ELI5 and LIME, we provided local explanations that highlighted key diagnostic factors, including Neisseria meningitidis presence, CSF WBC count, patient age, and neutrophil levels. These insights support clinical trust by aligning model with medical reasoning.