Tree-Augmented Naive Bayes for interpretable early dengue detection using routine hematological parameters as a proof-of-concept case study from resource-limited Bangladesh

Dengue fever is a significant community health problem in tropical regions such as Bangladesh, where it is difficult to diagnose early because it has similar symptoms to other febrile illnesses. Confirmatory tests are frequently non-existent in resource-limited settings, increasing the necessity for tools that utilize common clinical data as a means of relieving healthcare load. This paper presents a proof-of-concept case study using the Tree-Augmented Naive Bayes (TAN) Bayesian network to detect dengue using complete blood count (CBC) features. We analysed a publicly available, fully de-identified dataset from a single center in Jamalpur, Bangladesh, comprising 1523 patients tested for dengue and other febrile illnesses, with 19 hematological and demographic variables. Probabilistic modeling and clinical alignment involved the discretization of variables. The Boruta algorithm was used to choose nine significant predictors. TAN has been compared against Naive Bayes, Support Vector Machine, Logistic Regression, Linear Discriminant Analysis, and K-Nearest Neighbour using tenfold cross-validation. TAN achieved high sensitivity (0.961) but relatively low specificity (0.314), with moderate balanced accuracy (0.637) and AUC-ROC (0.671), indicating its suitability as a screening and decision-support tool rather than a definitive diagnostic model. Its directed acyclic graph expresses relationships among such variables as platelet count, hematocrit, and leukocyte differentials and provides better probabilistic information than independence-based models. Key technical limitations of this proof-of-concept study include single-center data, discretization of continuous variables, class imbalance, and lack of external validation, all of which are documented in the manuscript’s Limitations section. TAN showed strong sensitivity, although its specificity was limited, making it more suitable for use in unbalanced or uncertain diagnostic settings. In conclusion, this proof-of-concept case study demonstrates that TAN is an interpretable and high-sensitivity dengue detection algorithm that uses available CBC data within a single resource-limited setting. The primary contribution is clinical interpretability and feasibility, not generalizable deployment. Prospective multi-center validation is required before clinical application.