Predictive models for hospitalization and mortality in dengue using SINAN data: study protocol for development, temporal validation, and performance evaluation

Technical Summary: Predictive Models for Hospitalization and Mortality in Dengue Using SINAN Data

Problem Statement
Dengue fever imposes a significant clinical and organizational burden on healthcare systems, particularly during epidemics where high case volumes strain triage, hospitalization decisions, and patient monitoring. While a growing body of literature addresses dengue prediction, existing studies often suffer from methodological heterogeneity, insufficiently detailed analytical designs, and a lack of rigorous validation. A critical gap exists in the conflation of distinct clinical outcomes; many studies treat hospitalization, severe disease, and mortality as interchangeable or part of a single continuum. However, these outcomes occur at different stages of the care pathway: the decision to admit a patient relies on information available at initial assessment, whereas mortality risk assessment occurs after admission for patients already in a severe state. Furthermore, published models frequently lack calibration metrics, suffer from high risk of bias (particularly regarding missing data handling and small event-to-predictor ratios), and rarely undergo temporal validation to test stability across epidemic cycles.

Methodology
This study protocol outlines a retrospective analysis using secondary data from the Brazilian Notifiable Diseases Information System (SINAN) covering the period from 2017 to 2025. The study is designed in accordance with the TRIPOD+AI guidelines for predictive modeling and utilizes the Reusable Analytical Pipelines for Infectious Diseases (RAPID) framework.

• Study Design and Populations: Two independent analytical populations are defined to address distinct clinical questions:

  1. Hospitalization Model: Includes all reported dengue cases (confirmed by laboratory or clinical-epidemiological criteria) to predict the outcome of hospitalization. Predictors are restricted to data available up to the point of notification or initial assessment.
  2. Mortality Model: Includes only patients hospitalized for dengue to predict in-hospital mortality. Predictors are restricted to data available up to the point of admission.
  • Exclusion Criteria: Records with missing outcomes, missing dates of symptom onset/notification, or incompatible date sequences are excluded. Discarded cases and inconclusive classifications are also excluded.

• Data Processing and Feature Engineering:

  • Cleaning: Duplicate notifications are handled via key-based deduplication. Variables with incompatible units or coding across years are harmonized.
  • Missing Data: The mechanism of missingness is assessed prior to imputation. Multiple Imputation by Chained Equations (MICE) is the primary strategy, with median/mode imputation used for sensitivity analysis.
  • Preprocessing: Categorical variables are one-hot encoded; continuous variables are standardized. Temporal variables (month, epidemiological week) are transformed using sine and cosine functions to preserve seasonality.
  • Feature Selection: A combination of filter approaches and embedded methods (e.g., LASSO, Elastic Net) is employed. Wrapper strategies are used for algorithms lacking internal selection.

• Modeling Strategy:

  • Algorithms: The study compares regularized regression models against supervised machine learning algorithms, including tree-based ensembles (Random Forest) and gradient boosting methods. A stacking ensemble is evaluated as an exploratory analysis.
  • Class Imbalance: Addressed primarily through class weights inversely proportional to outcome frequency to preserve the interpretation of predicted probabilities as absolute risks. Resampling techniques are restricted to sensitivity analyses within training folds only.
  • Validation: The primary validation strategy is temporal, training on earlier years and testing on subsequent periods within the 2017–2025 window to simulate prospective use and detect temporal drift. Geographic validation is explored if data structure permits.

• Evaluation Metrics:

  • Discrimination: AUROC, AUC-PR, sensitivity, specificity, PPV, NPV, and F1-score.
  • Calibration: Calibration-in-the-large, calibration slope, Brier score, and calibration belts.
  • Clinical Utility: Decision curve analysis.
  • Interpretability: SHAP (SHapley Additive exPlanations) analyses for tree-based models.

Key Contributions

• Outcome Separation: The protocol explicitly decouples hospitalization and mortality modeling, ensuring that predictors are temporally appropriate for the specific decision point (pre-admission vs. post-admission).
• Rigorous Validation: By prioritizing temporal validation over simple random splitting, the study aims to assess the stability of model performance across different epidemic cycles and serotype shifts.
• Comprehensive Evaluation: The study moves beyond discrimination metrics (AUROC) to emphasize calibration and clinical utility, acknowledging that a highly discriminative but poorly calibrated model can lead to inappropriate clinical decisions.
• Reproducibility: The pipeline is implemented as open-source code within the RAPID framework, including scripts to rebuild the dataset from public SINAN files, ensuring full reproducibility without access to restricted data.

Results
As this document is a study protocol, it does not report empirical results, model performance metrics, or specific findings regarding the predictive power of the models. The paper details the plan for generating these results, including the specific algorithms, validation strategies, and metrics to be used in the final analysis.

Significance and Claims
The authors claim that this protocol addresses critical methodological gaps in dengue prediction literature by:

1. Enhancing Clinical Consistency: By separating outcomes, the resulting models are better suited for specific clinical tasks (triage vs. monitoring deterioration) rather than offering a generic "severity" score.
2. Improving Reliability: Through strict adherence to TRIPOD+AI and PROBAST+AI standards, the study aims to produce models with lower risk of bias, particularly regarding missing data and overfitting.
3. Ensuring Temporal Stability: The temporal validation approach is designed to test whether models remain effective as epidemic characteristics change, a factor often overlooked in previous studies.
4. Providing a Reference Framework: By documenting the pipeline via RAPID and sharing open-source code, the study aims to serve as a reusable template for predictive modeling in other infectious diseases and surveillance contexts.

The authors explicitly state that they do not assume the automatic superiority of complex machine learning methods over logistic regression; the final model selection will be driven by validated performance, calibration, parsimony, and interpretability rather than marginal gains in discrimination.