Bayesian networks for predicting clinical outcomes in COVID-19 patients: A retrospective study in a resource-limited setting

Background: The COVID-19 pandemic has highlighted the critical need for robust, interpretable predictive models to guide clinical decision-making for hospitalized patients, particularly in resource-limited settings. While machine learning approaches have achieved high predictive performance, many lack the transparency required for clinical adoption. Bayesian networks provide a rigorous mathematical framework for modeling medical uncertainty and complex causal relationships while maintaining clinical interpretability. Objective: To develop and validate an interpretable Bayesian network model for predicting clinical outcomes in hospitalized COVID-19 patients, including severity, complications, and mortality, and to compare its performance with existing predictive approaches. The model specifically addresses the needs of resource-limited settings where both interpretability and performance are critical. Methods: This retrospective cohort study analyzed 124 hospitalized COVID-19 patients at Tanambao I University Hospital, Madagascar (March 2020–March 2022). Given limited genomic sequencing capacity during the study period, specific SARS-CoV-2 variants infecting individual patients could not be confirmed. After descriptive and statistical implicative analysis, we constructed a Bayesian network integrating predictive variables organized in three levels: input variables (demographic, vital signs, biological, radiological), intermediate variables (clinical scores), and target variables (severity, complications, evolution, death). Parameter learning used maximum likelihood estimation with Laplace smoothing. Model performance was evaluated using stratified 10-fold cross-validation and compared against logistic regression, random forest, and support vector machine approaches. Results: The Bayesian network model exhibited strong diagnostic performance, with an AUC of 0.95 (95% CI: 0.91–0.99) for death prediction, 0.94 (95% CI: 0.90–0.98) for severe outcomes, and 0.93 (95% CI: 0.89–0.97) for unfavorable progression. Stratified 10-fold cross-validation yielded a mean accuracy of 0.85±0.03. The model outperformed logistic regression (AUC 0.89), random forest (AUC 0.91), and SVM (AUC 0.87) while maintaining superior interpretability. Sensitivity analysis identified qSOFA score (mutual information: 0.342), SpO2 levels (0.298), and respiratory distress (0.276) as the most influential variables. Robustness testing showed prediction stability under parameter perturbations of ±15%, with variation remaining below 8%. Conclusions: Bayesian networks constitute a promising tool for COVID-19 outcome prediction in resource-limited settings, combining competitive predictive performance with essential clinical interpretability. The probabilistic approach enables rigorous uncertainty quantification critical for medical decision-making. While the model achieved excellent performance on our cohort, external validation across different populations and SARS-CoV-2 variants is needed to establish broader generalizability.