Probabilistic Forecasting of Monthly Dengue Cases Using Epidemiological and Climate Signals: A BiLSTM–Naive Bayes Model Versus Mechanistic and Count-Model Baselines
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Reliable short-term forecasts can help urban health systems anticipate dengue surges and allocate resources. We assembled monthly dengue case counts for Freetown, Sierra Leone (2015–2025), and compared four probabilistic model families under a leakage-safe, rolling-origin protocol at 1–3-month horizons: a negative-binomial generalized linear model (NB-GLM), a negative-binomial INGARCH, a mechanistic renewal model with NB observations, and a Bidirectional LSTM with a negative-binomial output (BiLSTM-NB). All models used the same seasonal harmonics and autoregressive lags; “light” climate inputs (rainfall, temperature, relative humidity) were limited to ≤ 3 features and lagged to reflect real-time availability. Evaluation used proper mean log score, coverage, median width of 50% and 90% predictive intervals, randomized PIT histograms, and Diebold–Mariano tests with Newey–West standard errors. For the main comparison, we aligned forecasts on common (issue, target) pairs per horizon (n=33). No single approach dominated across horizons. At 1–2 months, the INGARCH-NB and NB-GLM frequently achieved the highest mean log scores with near-nominal coverage. At 3 months, a calibrated BiLSTM-NB ensemble more often yielded the best mean log score and narrower central intervals without marked undercoverage. Diebold–Mariano tests indicated statistically significant improvements for INGARCH-NB over mechanistic renewal at shorter horizons and for BiLSTM-NB over mechanistic and GLM baselines at 3 months. PIT profiles were closer to uniform for the top-performing model at each horizon, supporting probabilistic calibration.These results suggest a horizon-specific toolkit for operational dengue forecasting: lean count models for 1–2 months ahead and calibrated deep learners for 3 months, all under consistent leakage controls. The framework and artifacts (per-issue forecasts, aligned indices, and code) provide a transparent baseline for future work in similar urban settings.
Author summary
We conducted this study to help public health teams in Freetown, Sierra Leone, plan clinical capacity and vector control using short-term dengue forecasts they can trust. Many forecasting approaches exist, but they are rarely compared under the same, leakage-safe conditions. We assembled monthly dengue case data (2015–2025) and created consistent seasonal and autoregressive features for all models, using only a light, real-time-feasible set of climate inputs. We then compared four model families: a negative-binomial generalized linear model, an INGARCH count model, a mechanistic renewal model, and a Bidirectional LSTM with a negative-binomial output. Using an expanding-window, rolling-origin evaluation at 1-3 month horizons, we assessed proper scoring, interval coverage and width, PIT histograms, and Diebold–Mariano tests on aligned targets. We found that no single method dominated across horizons: simpler count models performed strongly at 1-2 months, while a calibrated BiLSTM provided the best performance more often at 3 months. These findings suggest a horizon-specific toolkit can improve operational dengue forecasting in similar urban settings.
