7‑day dengue forecasts with 9.8% MAPE using epidemiology‑guided machine learning
HAWKEYE Disease Intelligence fuses epidemiology, weather, demographics, and satellite context to forecast dengue with 7‑day MAPE of 9.8% for Dhaka while revealing a strong urban metabolism link via nightlights, enabling policy actions up to two weeks in advance.
The system integrates 30+ datasets with rigorous preprocessing and delivers nowcasts, risk maps, and intervention windows as API outputs for health operations centers.
Provide city‑scale dengue incidence forecasts and risk alerts with weekly cadence, operationalized for Dhaka (2022–2025), and designed for replication across districts through standardized data adapters and harmonized feature schemas.
The deployment pairs statistical forecasting with causal signal analysis to keep models interpretable and resilient to shifting data quality across municipal sources.
Weekly dengue case counts and population‑normalized rates compiled for Dhaka wards and aggregates.
Daily temperature, humidity, rainfall, and pressure used as lagged regressors for vector ecology dynamics.
VIIRS nightlights radiance as an urban activity proxy and mobility surrogate; satellite basemaps for optional cartographic context.
Ward boundaries and population density layers for normalization and targeting.
1,105 temporal records, 36 fused features, 0.16% missing data after harmonization.
Ingestion and QA: Standardize sources, fix calendar alignments, forward‑fill short gaps under strict caps, and verify monotonic ward aggregations for totals.
Feature engineering: Create 7–28‑day lags for temperature, humidity, and rainfall, plus rolling means and anomalies; aggregate nightlights monthly and align to epidemiological weeks.
Forecasting layer: Prophet‑based univariate core with exogenous regressors where supported, backed by gradient‑boosted tabular models for sensitivity and ablation analysis.
Key causal relationships guiding the forecasting system:
Pearson correlation with 14-day lag:
$$ \text{Corr}\big(T_{t-14},\,\text{Dengue}_{t}\big)=0.324 $$
$$ r=0.88 \text{ between nightlights and GDP, informing contact patterns} $$
36 features with low missingness enable robust analysis:
$$ N=1105 \text{ records, } 0.16\% \text{ missing data} $$
Forecast core: Prophet with weekly seasonality, holiday effects where available, and tuned changepoint prior; exogenous covariates engineered as lags and anomalies.
Policy window up to 14 days to stage spraying, larval source reduction, and ward staffing changes.
$$ \text{MAPE} = \frac{100}{N}\sum_{t=1}^{N}\left|\frac{y_{t}-\hat{y}_{t}}{y_{t}}\right| $$
Primary evaluation at 7-day horizon
$$ \text{Corr}\big(T_{t-14},\,\text{Dengue}_{t}\big)=0.324 $$
Validating thermal lead against cases
$$ g_{t} = \frac{y_{t}-y_{t-1}}{y_{t-1}} $$
Used for surge detection and thresholded alerting
df["temp_lag_14"] = df["temp"].shift(14)
df["hum_lag_14"] = df["humidity"].shift(14)
df["rain_lag_14"] = df["rain"].shift(14)
df["nl_month"] = nightlights.resample("W").mean().reindex(df.index, method="ffill")
features = ["temp_lag_14","hum_lag_14","rain_lag_14","nl_month"]m = Prophet(weekly_seasonality=True, yearly_seasonality=True)
for col in features:
m.add_regressor(col, standardize=True)
fit = m.fit(df.rename(columns={"cases":"y","date":"ds"}))
forecast = m.predict(df_future) # 7-day horizonX = df[features].shift(1).dropna()
y = df["cases"].loc[X.index]
model = xgb.XGBRegressor(max_depth=6, n_estimators=400, learning_rate=0.05)
model.fit(X, y)
yhat = model.predict(df_future[features])7‑day MAPE $$=9.8\%$$ on Dhaka, meeting decision‑grade thresholds for weekly operational planning.
Statistically significant temperature→dengue link at 14‑day lag $$ r=0.324 $$ with $$ N=1105 $$, $$ p\ll 0.001 $$.
Fused urban dataset exhibits 0.16% missingness, enabling stable training and fewer imputations.
36 features with low missingness enable robust cross‑validation and ablation checks.
Convert noisy weather streams into biologically meaningful regressors while preserving interpretability.
Prophet + boosted trees provides both transparent baselines and non‑linear capture for robustness under changing behaviors.
Low missingness and consistent ward delineations reduce label noise and improve week‑ahead stability.
Batch ETL refreshes sources weekly and publishes forecasts and risks to the Urban Intelligence API and Vercel dashboards.
Case backlog dumps and testing policy shifts can transiently break stationarity; outlier detection and robust losses mitigate but do not remove risk.
Mobility restrictions or blackouts alter transmission dynamics; models require rapid recalibration using recent windows.
Elasticities and optimal lags vary by microclimate; city‑specific validation is required when scaling nationally.
Fuse mosquito trap counts and vector suitabilities for mechanistic hybrid models.
Graph‑based spillover modeling between wards using mobility or road‑network priors.
Alerting when thermal/rainfall precursors cross city‑specific thresholds to trigger pre‑emptive operations.