Our study underscores the potential of ML models, particularly gradient boosting algorithms, to outperform traditional warning signs in predicting severe dengue. This improved predictive ability could transform clinical decision-making, enabling earlier and more accurate identification of high-risk patients, thereby improving outcomes in dengue-endemic regions like Puerto Rico.

Hemoconcentration, days post symptom onset, and leukopenia emerged as the most important features across multiple ML models, aligning with their known relevance in dengue prognosis [47–49]. Hemoconcentration, which reflects plasma leakage through an increased red blood cell concentration, was consistently highlighted as a top predictor by CatBoost, XGBoost, LightGBM, and several other algorithms. Days post symptom onset is a crucial temporal marker, likely capturing the dynamic nature of disease as cases often progress to the critical phase of dengue (when severe disease occurs) 3–7 days after symptom onset. Leukopenia, or low white blood cell count, often reflects the body’s response to viral infections, including dengue. In addition, pale skin, age group, and clinical warning signs such as restlessness, abdominal pain, and persistent vomiting were identified as key predictors of severe dengue progression. The variability in feature importance across models emphasizes the complexity of severe dengue prediction, highlighting the need for tailored approaches that account for both individual patient characteristics and disease progression.

Our analysis highlights the strong predictive performance of gradient boosting algorithms— CatBoost, XGBoost, and LightGBM—with AUC values above 94%, reflecting their ability to capture complex, non-linear patterns in clinical data [35, 50–52]. Despite this high predictive accuracy, the interpretability of these models remains a limitation in clinical settings, where transparency in decision-making is critical for trust and practical use [53–55]. Compared to simpler models like LR, which offers straightforward interpretations of how each variable influences severe dengue risk, gradient boosting algorithms can be challenging to interpret. This trade-off between high performance and interpretability suggests that ML models may be most useful as supplementary tools for alerting clinicians to high-risk cases, rather than as standalone decision aids.

In contrast, simpler models like Naïve Bayes, Decision Trees, and KNN showed weaker correlations and lower AUC values, reflecting their limitations in capturing data complexity. The ensemble model, achieving the highest AUC of 97.7%, demonstrates the added value of combining multiple ML algorithms to enhance predictive accuracy, with high sensitivity and specificity, making it particularly useful in resource-limited settings [56, 57]. Additionally, the exclusion of immune status and serotype data had minimal impact on model performance, indicating these variables are not essential for accurate prediction in this context. The high NPV across models suggests that ML tools can still be valuable for identifying low-risk patients who may not require intensive monitoring. In these cases, the ML model’s recommendation could serve as an early discharge or outpatient management decision-support mechanism, further optimizing healthcare resource allocation.

The post-hoc variable reduction analysis demonstrated that a streamlined set of 20 variables achieved strong predictive accuracy (AUC of 96.5%), close to the full 40-variable model. This reduced set offers a balance between interpretability and performance, making it more practical for clinical application, particularly in settings with limited diagnostic resources. By focusing on essential predictors—such as hemoconcentration, days post onset, leukopenia, and key symptoms—this approach prioritizes feasibility and interpretability, even if it means a modest sacrifice in predictive power. For added clinical utility, LR could complement ML approaches by further refining and validating the reduced variable set with interpretable odds ratios, enabling clinicians to apply these findings more confidently in practice [58].

Traditional warning signs showed both strengths and limitations. Abdominal pain and restlessness were the most sensitive indicators, consistent with other studies [6, 9], yet their low specificities limit their utility. Conversely, markers like hepatomegaly and hemoconcentration had high specificity but low sensitivity. Combining multiple warning signs improved specificity while maintaining high sensitivity, yielding the highest AUC among tested combinations. In contrast, gradient boosting ML algorithms offered a more balanced approach with high sensitivity and specificity, crucial for accurate risk stratification in clinical settings. Our findings align with a recent study where an 8-gene XGBoost model outperformed clinical warning signs, significantly improving negative predictive power and demonstrating strong generalizability across patient cohorts [54]. Although our models focus on demographic and clinical features, the integration of gene expression data presents an intriguing avenue for future exploration [54, 55].

This study has several limitations. First, these ML models need to be re-fitted to different variables and populations to ensure accuracy across various settings. Second, the relatively small dataset increases the risk of overfitting, potentially affecting the models’ robustness and generalizability. Third, the prevalence of DENV-1 cases from the 2012-2013 outbreak may limit the models’ applicability to other periods, regions, populations, age groups, or serotypes. Fourth, the models were developed using data from the SEDSS, where inclusion criteria required febrile illness, potentially limiting generalizability to broader populations. Fifth, due to limited sample sizes for serotypes and immune statuses, we could not conduct ML analyses for these subgroups; future studies with larger datasets are needed to validate findings. Sixth, there is a potential limitation related to the inclusion of false negatives—SEDSS cases who may have later presented to a non-SEDSS facility with severe disease—although this is considered unlikely given typical healthcare-seeking behaviors. Seventh, the SEDSS data’s robustness may not accurately reflect real-world conditions, where datasets are often sparse, contain free-text fields, or have incomplete information, potentially affecting model performance. Eighth, dengue and severe dengue are often underdiagnosed and underreported, which could impact model results; however, our findings likely represent a conservative estimate when accounting for underreporting. Finally, implementing ML models in clinical practice may require computational resources and infrastructure not available in all settings, particularly in low-resource environments where dengue is endemic.