A proof-of-concept study from Jamalpur, Bangladesh, reports that a Tree-Augmented Naive Bayes (TAN) model, trained only on routine complete blood count (CBC) parameters, could detect dengue with high sensitivity but limited specificity. The study, published in Discover Viruses, analysed a publicly available, fully de-identified dataset of 1,523 patients tested for dengue and other febrile illnesses at Jamalpur 250-Bedded General Hospital during the 2024 outbreak season. The authors emphasise that the work is a feasibility study at a single center, not a clinically deployable diagnostic tool, and that prospective multi-center validation is required before any real-world application.
The motivation is practical. Dengue shares early symptoms with malaria, chikungunya, typhoid and other febrile illnesses that co-circulate during monsoon seasons, and confirmatory tests such as RT-PCR or NS1 antigen tests are often unavailable or slow in rural and peripheral facilities. Bangladesh’s 2023 outbreak registered 321,179 cases and 1,705 deaths, the most lethal on record according to the paper, exposing gaps in early diagnosis. The researchers therefore asked whether routinely collected hematological markers, already measured in almost every febrile patient, could support an interpretable screening and decision-support tool.
The dataset contained 19 hematological and demographic variables per patient, including hemoglobin, white blood cell count, differential counts of neutrophils, lymphocytes, monocytes and eosinophils, platelet count, hematocrit, red cell indices such as MCV, MCH and MCHC, platelet indices including MPV, PDW and PCT, plus age and gender. Of the 1,523 laboratory-confirmed records collected between February and September 2024, 1,042 (68.4 percent) were dengue-positive and 481 (31.6 percent) were negative. Females made up 52.8 percent of the cohort, and patient ages ranged from 5 to 78 years. No missing entries were identified, so no imputation was needed.
Because Bayesian network models require categorical data, the researchers discretized continuous variables into clinically meaningful bins such as Low, Normal and High, using reference ranges and gender-specific intervals where appropriate, for example for hemoglobin and hematocrit. The Boruta feature selection algorithm, a wrapper method built on Random Forest that compares each variable’s importance against randomized shadow features, then confirmed nine predictors: platelet crit (PCT), neutrophils, platelet count, red blood cell count, eosinophils, total WBC count, RDW-CV, MCHC and hematocrit. Age, although univariately significant, was excluded because its contribution appeared mediated through the stronger hematological markers.
The TAN classifier extends Naive Bayes by allowing each attribute one additional attribute as a parent, capturing dependencies that plain Naive Bayes assumes away. The tree structure is learned from conditional mutual information between attribute pairs given the class variable. In the learned network, hematocrit and red blood cell count occupied central positions, consistent with hemoconcentration, while PCT linked to eosinophils, MCHC and platelet count, and neutrophils connected with total WBC count, reflecting leukocyte dynamics. The authors argue this directed acyclic graph offers structural interpretability that post-hoc explanation methods such as SHAP or LIME cannot match, letting clinicians trace the probabilistic logic of a prediction.
All six models compared, TAN, Naive Bayes, Support Vector Machine, Logistic Regression, Linear Discriminant Analysis and K-Nearest Neighbour, were evaluated with tenfold cross-validation on the nine selected features. Every model achieved high sensitivity, ranging from 0.936 to 0.984, but low specificity, ranging from 0.183 to 0.331, a pattern the authors attribute to the biological overlap between early dengue hematological profiles and other endemic febrile illnesses, aggravated by the 68.4 percent prevalence of the positive class. Precision was relatively stable across models at 0.723 to 0.752.
TAN recorded the best balanced performance: sensitivity of 0.961, precision of 0.752, balanced accuracy of 0.637, an F1-score of 0.844 and an AUC-ROC of 0.671. Naive Bayes followed closely with an AUC of 0.666 and balanced accuracy of 0.633, while SVM, Logistic Regression and LDA clustered around AUC values of 0.658 to 0.663. KNN maximised sensitivity at 0.984 but had the lowest specificity, 0.183, and the lowest balanced accuracy, 0.583. A Friedman test across the ten folds found a significant overall difference among the six models, but the Nemenyi post-hoc test showed TAN differed significantly only from KNN; its differences from NB, LR, LDA and SVM were not statistically significant.
The authors are explicit that TAN should therefore be viewed as a competitive and interpretable alternative rather than a statistically superior model. The moderate AUC of 0.671 and the low specificity of 0.314 indicate a limited ability to distinguish dengue from other febrile illnesses using blood counts alone, particularly in high-prevalence settings. They frame the tool’s value as screening, where the priority is avoiding missed true dengue cases rather than confidently excluding non-dengue conditions, and note that univariate associations in the data, such as thrombocytopenia below 150,000 per microliter being linked with dengue positivity at 76.0 percent versus 62.9 percent, align with known dengue pathophysiology of platelet depletion and plasma leakage.
The limitations section documents several constraints. The single-center dataset may not represent the demographic patterns or serotype variation of dengue across Bangladesh. Discretization of continuous variables may have discarded subtle clinical detail. No resampling techniques were applied to address class imbalance, which may have further depressed specificity. The study examined binary detection rather than severity grading, included no longitudinal follow-up of disease course, and omitted tree-based methods because the authors deliberately focused on interpretable probabilistic models. A direct comparison with clinical judgment or rapid diagnostic tests was not feasible with the available data.
Future directions outlined in the paper include multi-center, multi-regional validation, cost-sensitive learning or resampling methods such as SMOTE to improve specificity, dynamic Bayesian networks using time-series CBC data to model infection progression, models retaining continuous variables, and the incorporation of accessible clinical symptoms or simple biochemical markers. Extension to severity prediction, distinguishing mild from severe dengue or estimating the risk of dengue hemorrhagic fever or dengue shock syndrome, is proposed as a potential benefit for triage and resource allocation in low-resource endemic areas. For now, the study’s contribution, as the authors state, is demonstrating clinical interpretability and feasibility rather than generalizable deployment.
Subject of Research: A proof-of-concept Bayesian network model for early dengue detection using routine hematological parameters in Bangladesh.
Article Title: Tree-Augmented Naive Bayes for interpretable early dengue detection using routine hematological parameters as a proof-of-concept case study from resource-limited Bangladesh
Article References: Jahan, I., Shraboni, M. S. A., Tonni, M., Ahmed, S., Parvin, M., & Akther, S. (2026). Tree-Augmented Naive Bayes for interpretable early dengue detection using routine hematological parameters as a proof-of-concept case study from resource-limited Bangladesh. Discover Viruses, 3(1), Article 10. https://doi.org/10.1007/s44370-026-00046-z
Image Credits: AI Generated
DOI: 10.1007/s44370-026-00046-z
Keywords: dengue, Tree-Augmented Naive Bayes, Bayesian network, complete blood count, machine learning, Bangladesh, screening tool, interpretability, class imbalance, resource-limited settings, cross-validation, hematological parameters
News Source: Ophelia Keating. (October 29, 2026). Interpretable Bayesian Model Spots Dengue from Routine Blood Counts in Bangladesh. Scienmag.
