A New Approach to Monitoring Dengue Activity

Madoff, Lawrence C.; Fisman, David N.; Kass-Hout, Taha

As with any prediction-oriented surveillance tool, a major concern relates to model ''over-fitting'' such that the prediction model performs well in the dataset that was used in its creation but fails to work well in the ''real world''. Reassuringly the authors divided their data into a derivation set and a testing set (or ''holdout set'' as they call it), with the former used for model construction. As can be seen in the table and figure they present, their derived models perform extremely well in both sets in all countries, in the derivation set as expected, but also in the testing set. Perhaps less straightforward is the authors' decision to ''smooth out'' unusual spikes in search volumes in candidate queries; as demonstrated by the influenza example above, extreme surges in public interest in a disease can cause surges in query volumes, as can surges in interest related particular subject that is unrelated to the disease under surveillance but shares attributes that would be the subject of searches. By smoothing search volumes, the authors may have incorporated into their models terms that have the potential to ''misbehave'' in the future. For example, one imagines that if a novel (and frightening) new hemorrhagic fever unrelated to dengue emerges in one of these countries in coming years, one would imagine that the correlation between the search term ''haemorrhagic fever'' and dengue volumes would decline. As we don't have access to the precise query terms that were included in each country-specific model, it is difficult to know whether or not the terms included in the model would be vulnerable to such effects. The authors note that the expanding range of a clinically similar illness (Chikungunya) may confound the utility as well [28] .