How far ahead of the case count do public data forecast a large dengue season? A retrospective simulation by Brazilian state
Year: 2026
Type: article
Abstract: Background: the number of Brazilian states with a high-incidence dengue season (300 or more probable cases per 100,000) ranged from 3 to 19 between 2014/15 and 2024/25. It is not known how far ahead of the simple case count public data allow forecasting this. Methods: retrospective simulation in the 27 states and 11 seasons (297 pairs), each state with its own season, at each of the first 43 weeks, partially rebuilding what was available on each date. Eight predictors (count projection, recent growth, off-season cases, mild winter, thermal suitability, El Niño and, by amendment, cases in the capital and in the interior, each compared with its own history) were tested alone and in pairs, in penalized logistic regressions anchored on the state's history; criterion M chooses, each season, the model with the lowest error in earlier seasons. Lead time is the number of weeks between accuracy stably passing 75% and the count of probable cases making half of the high seasons evident. The main analysis uses the final classification of cases; a sensitivity analysis uses all notifications, as an alternative approximation of what was available. Additional analyses, defined after the first results: a fixed alert rule, a three-state rule with the projection band and a test on held-out seasons. Results: the count projection passed 75% accuracy at week 10, 25 weeks before evidence from the count (95% CI 14 to 30; 11 under the strict rule), and 18 weeks (12 under the strict rule) with the count less dependent on the final classification, in seasons 2016/17 to 2020/21. The interior improved the projection in every version (incremental BSS 0.06 to 0.14), more than off-season cases in most analyses and with a very close value in one of them; the capital added no clear gain to the projection. Criterion M, whose final composition is projection plus interior, had 18 weeks (CI 11 to 29; 10 under the strict rule) in the main analysis and 11 (5 under the strict rule) in the version less dependent on the final classification. The fixed pair projection plus interior had the lowest error and the longest lead time under the strict rule (13 weeks), and was the one chosen when selection used only 2011/12 to 2018/19, but those held-out seasons had already been examined. With a probability alert rule, all models alerted before the count in most high seasons, with false alarms in 37% to 57% of non-high seasons; requiring the whole projection band above the threshold cut false alarms to 0% to 2%, but the median lead time to 0 to 4 weeks, and declared low, in some week, 10% to 25% of the seasons that ended high. Conclusion: in this retrospective simulation, the count projection anticipates evidence from the count; with criterion M, whose final composition is the projection and the cases in the interior (in the retrospective forecasts, M changed models over the seasons), a high season can be forecast 18 weeks ahead of the simple case count in the main analysis, but the lead time falls with the count less dependent on the final classification, and false alarms are frequent. Confirmation depends on the prospective check.
Author: Wistuba Melo da Cunha, Willian
Language: English
Field: Medicine
Domain: Health Sciences
Open Access status: green
OpenAlex ID: W7225595329