Predicting Dengue Clinical Severity in Eastern Sudan's 2023 Outbreak: A Comparative Analysis of Statistical and Machine Learning Models Using Routine Surveillance Data
Objectives: This study characterized clinical severity patterns during the 2023 dengue outbreak in Al Qadarif State, Eastern Sudan; identified associated factors; compared statistical and machine-learning models; and explored patterns of clinical manifestations using routine surveillance data. Methods: This retrospective observational study included 2,552 dengue case records. The primary outcome was a pragmatic, study-specific composite clinical severity indicator, defined as the documented presence of bleeding, severe bleeding, loss of consciousness, convulsions, or thrombocytopenia.Analyses included descriptive statistics, multivariable logistic regression, random forest, XGBoost, multiple correspondence analysis, and hierarchical clustering. Model performance was assessed using a stratified hold-out test set, with uncertainty estimated from 1,000 bootstrap resamples. Results: Overall, 182 patients (7.13%) met the study-specific composite clinical severity indicator.Thirty-two deaths were recorded, giving a case-fatality rate of 1.25%. Mortality was higher among patients with severity markers than among those without them (3.85% vs. 1.05%; relative risk=3.65). Skin rash showed the strongest adjusted association with the composite severity indicator (adjusted OR=6.26, 95% CI: 3.15–12.45; P<0.001), followed by symptom-onset timing (adjusted OR=4.65, 95% CI: 2.41–8.98; P<0.001). XGBoost achieved the numerically highest AUROC (0.765, 95% CI: 0.691–0.837) and the lowest Brier score (0.161, 95% CI: 0.145–0.176). Multiple correspondence analysis suggested hemorrhagic and neurological/systemic symptom patterns. Conclusion: The study-specific composite clinical severity indicator was uncommon but was associated with a higher risk of death. Skin rash and symptom-onset timing showed the strongest adjusted associations with the study-specific composite clinical severity indicator. Predictive performance was moderate, and the findings require confirmation using independent datasets with more detailed clinical and laboratory information.