Integrating google trends and hybrid statistical-machine learning models for dengue surveillance in an inland vietnamese province: a 94-month evaluation (2013-2021) with search-signal anomaly screening

Clinical Snapshot

Key Findings

• P Value: Not explicitly reported, but improvement statistically non-significant

• Effect Size: 11.77% RMSE reduction with GTI integration; XGBoost R² = 0.861 vs baseline R² = 0.823

• Primary Outcome: Dengue case prediction accuracy using Google Trends integration

• Nnt Or Sensitivity: Cross-correlation peaked at lag 0, indicating contemporaneous nowcasting capability rather than predictive forecasting

• Confidence Interval: 95% CI for RMSE improvement crossed zero

Clinical Application

Moderate - requires technical expertise for model implementation and ongoing Google Trends data integration Limited direct applicability to Australian dengue surveillance due to different epidemiological patterns, healthcare infrastructure, and population search behaviours. May have relevance for northern Queensland dengue monitoring during outbreak periods Public health surveillance systems in dengue-endemic regions with adequate internet penetration and local language search data

Abstract

Dengue transmission in inland Southeast Asia shows strong seasonality and short-term surges that challenge timely public-health response. We assessed whether province-level Google search activity can support short-horizon dengue monitoring in Dong Nai, Vietnam. Monthly reported dengue cases from July 2013 to April 2021 were aligned with the Google Trends index (GTI) for "Sốt xuất huyết". Models were evaluated using leakage-free rolling-origin expanding-window cross-validation with 3-month validation blocks. We compared a negative binomial (NB) autoregressive baseline, NB models incorporating GTI, tree-based machine-learning models using the same covariates, and a simple NB-random-forest ensemble. Predictive performance was assessed using root mean squared error (RMSE), mean absolute error (MAE), R2, and discrimination for 95th-percentile outbreak exceedance. Cross-correlation peaked at lag 0, indicating that GTI functioned primarily as a contemporaneous nowcasting signal rather than a long-lead predictor. NB models incorporating GTI showed competitive calibration, while boosted tree models achieved the lowest point-estimate errors; XGBoost achieved RMSE ≈ 40.95 and R2 ≈ 0.861 compared with RMSE ≈ 45.23 and R2 ≈ 0.823 for NB with GTI. Adding GTI to the autoregressive NB baseline reduced RMSE by 11.77%, but the 95% confidence interval crossed zero, indicating modest and statistically non-significant incremental gain. Search-signal anomaly screening suggested that GTI spikes largely coincided with epidemic peaks and did not clearly inflate out-of-sample errors. Overall, GTI provides operational value for dengue nowcasting and alert triage in inland provinces, while future systems should integrate climate, vector, mobility, and media-monitoring data.

References

1. 1.Tuan, D. A., & Uyen, P. V. N. (2026). Integrating google trends and hybrid statistical-machine learning models for dengue surveillance in an inland vietnamese province: a 94-month evaluation (2013-2021) with search-signal anomaly screening. International Journal of Biometeorology. https://doi.org/10.1371/journal.pntd.0008056