Distributed data processing for public health surveillance

Lazarus, Ross; Yih, Katherine; Platt, Richard

While it is possible to centrally collate and process deidentified records, there is a potential problem with statistical inference if multiple records from the same individual are not distinguished. This problem arises because many statistical analysis techniques applicable to surveillance, such as Generalised Linear Mixed Models [4] (GLMM), depend on the assumption that observations are statistically independent. Inference based on this assumption using ambulatory care encounter data will likely be biased if the model cannot distinguish observations from multiple encounters during a single course of illness from a single individual patient. Although the extent of this bias has not been quantified, the problem is clearly illustrated by real data. In more than half of the individuals with multiple lower respiratory syndrome encounters over a four year period from one large ambulatory care practice, a second encounter with the same syndrome was noted less than 21 days after the first encounter [1] . Our approach to this problem of statistical independence is to aggregate multiple encounters from a single individual into "episodes" of illness, and is described in more detail below. Reliably automating this aggregation requires that every patient's records be uniquely identifiable.