Explainable Symptom-Based Monkeypox Detection using Bayesian Networks: A Probabilistic Clinical Decision Support Framework

Abstract

The accuracy of clinical decision-support systems is not sufficient: in fact, they should be eminently interpretable to be utilized in accurate and reliable diagnosis of emerging infectious diseases, such as monkeypox. Despite excellent predictive performance (e.g. 89.3 per cent in a prior study) reported by state-of-the-art machine-learning models (such as Light Gradient Boosting Machine (LGBM) ensembles), their use of post-hoc Explainable AI (XAI) systems, such as SHAP or LIME is intolerable to clinical transparency as to the causal reasoning behind their selection 

The study project will focus on creating a diagnostic model that can be intrinsically interpreted and apply Bayesian Networks (BNs) to diagnose monkeypox by using symptoms. The goal is two-fold: the diagnostic performance of the BN should be evaluated, but more importantly, its usefulness in terms of causal inference as compared to the performance metrics produced by the developed LGBM ensemble. 

Data employed in the study is openly available symptom-based monkey pox data, a binary data, i.e., whether or not the patient has fever, rash, and lymphadenopathy. The Bayesian Network structure was manually designed using domain-specific clinical knowledge derived from Monkeypox symptomology reported in the literature. A naïve Bayesian structure was adopted in which the disease node acts as the parent variable and symptom nodes represent conditionally dependent clinical manifestations.

The BN model had competitive predictive accuracy ( Accuracy ≈ 87.5 - AUC ≈ 0.93). This is a slightly poorer performance than the one achieved by the LGBM ensemble (89.3 00 percent) proposed by Setegn and Dejene, but it nonetheless demonstrates that intrinsic transparency does not mean that there is a major reduction in accuracy. Notably, the DAG of the BN presented a clear simplified set of probabilistic relationships and causal relationships among the symptoms, including fever, headache, lymphadenopathy, and the Monkeypox node that provides a clear and traceable diagnostic pathway and avoids the use of complex external XAI mechanisms. 
Bayesian Networks therefore would be a superior paradigm of high-stakes clinical decision support in the diagnostics of infectious diseases offering competitive accuracy and intrinsic transparency and causal interpretability. This method is a paradigm shift as compared to the retrospective explanation of correlation-based predictions (post-hoc XAI) by the deliberate modelling of the known medical reasoning, which creates a greater confidence in real-world applications in the triage and clinical environments.