Evaluating the incremental value of Google Trends for provincial dengue surveillance in Vietnam: a medical informatics time-series study.

Abstract

Internet search data are increasingly used to complement infectious disease surveillance, but their operational value depends on whether they add decision-relevant information beyond routine case reports. To evaluate Google Trends as an auxiliary digital signal for weekly dengue surveillance in Ba Ria-Vung Tau province, Vietnam, focusing on signal availability, temporal alignment, incremental predictive value, and delayed-reporting use cases. We analysed 209 consecutive weekly observations from 30 December 2018 to 25 December 2022, including reported dengue cases and the Google Trends Index (GTI) for the Vietnamese term "sốt xuất huyết" (dengue fever). After lag construction, 206 weeks were available for modelling. We characterized signal sparsity and temporal alignment, compared one-week-ahead persistence, autoregressive (AR), and AR + GTI forecasts, and evaluated chronological hold-out performance, expanding-window rolling-origin validation, simulated one- and two-week reporting delays, sensitivity to GTI representation and seasonality, and SHAP-based feature attribution. Weekly dengue cases ranged from 6 to 913, whereas GTI was sparse, with 131/209 weeks (62.7 %) equal to zero. Raw GTI and dengue counts showed strong co-movement, but after removing seasonal and first-order autoregressive structure, residual correlations were small (one-week GTI lead r = 0.153). Under full surveillance, the primary AR + raw-GTI model increased hold-out error versus AR-only (MAE 72.52 vs 63.70; RMSE 110.58 vs 92.25), although selected alternative GTI transformations produced numerical gains. A simulated one-week reporting delay showed numerical improvement with GTI (MAE 74.33 vs 80.04), but the two-week-delay results were mixed. SHAP attribution confirmed that previous-week dengue incidence dominated the primary AR + GTI model. GTI reflected dengue-related public attention but did not demonstrate consistent or robust incremental predictive value beyond timely routine case surveillance. Selected exploratory specifications suggested possible value when recent case reports were delayed, supporting a conditional role for GTI as an auxiliary signal rather than a stand-alone disease proxy or automatic alert trigger.