Generative embedding of sparse data with a tabular foundation model for dengue anticipatory action: a machine learning approach

Background Early outbreak detection often depends on complex, data-intensive models that have limited operational use in sparse surveillance settings. We developed a domain-mechanistic generative embedding that converts case counts and rainfall into a structured representation of dengue transmission for early epidemic-onset detection. Methods We constructed a 132-feature generative embedding from sparse dengue case and rainfall data. A tabular foundation model was evaluated using leave-one-year-out validation with paired cluster-bootstrap uncertainty intervals across 17 Philippine regions and eight dengue-endemic countries. Performance was benchmarked against raw input columns and catch22 time-series features. Findings Raw case and rainfall columns provided weak discrimination for dengue outbreak onset, with AUROC ranging from 0.56 to 0.70. The generative embedding improved prediction to AUROC 0.77 across countries and 0.89 across regions, corresponding to gains of +0.205 and +0.183 over raw columns, respectively, with paired cluster-bootstrap p[≤]0.006. Calibration error remained low at both regional and country scales, with expected calibration error of 0.067 and 0.149, respectively. Predictability was strongest in highly seasonal settings, including Philippine Type I regions, Mexico, Brazil, and the Philippines, whereas year-round transmission or opposing coastal rainfall regimes produced weaker performance. Country estimates based on only one or two retained epidemic seasons were unstable. Interpretation Under sparse surveillance conditions, the predictive capacity of a tabular foundation model depended strongly on the representation supplied to it. A generative embedding of climate and epidemiological dynamics translated limited case and rainfall inputs into actionable early-warning signals, with accuracy scaling according to local seasonal structure. These findings support mechanism-grounded embeddings as a practical route for extending prospective dengue outbreak surveillance in data-limited settings, especially at regional scales where calibration and deployment are most appropriate.