Joint assessment of temporal segmentation, time unit and detection algorithms in syndromic surveillance

The choice of the aggregation that defines the temporal unit of epidemiological surveillance is part of the more theoretical framework of the modifiable temporal unit problem (MTUP). It has been demonstrated that this choice influences temporal cluster detection and may lead to false-positive results and poor estimation of regression model parameters. In syndromic surveillance (SyS), despite the choice of which temporal aggregation to use being crucial, it has not yet been addressed in the literature. In most SyS systems, this choice is driven by the frequency of the data collection and/or human resources available, although neither the temporal unit's influence on the performance of anomaly detection algorithms nor on the efficiency of the SyS are known.The main objective of our study was to analyze the influence of the temporal aggregation unit on the performances of SyS detection algorithms used routinely, according to the characteristics of specific syndromes and outbreaks. Simulating daily time series of various syndromes, we tested three different time series aggregation methods. For each of four anomaly detection algorithms and their variants, we calculated seven performance indicators and multi-criteria scores to guide epidemiologists in their choice of which temporal aggregation of surveillance to use. From 19,200 analyzed time series, we observed an effect of temporal aggregation on the performance of the detection algorithms tested. Results also showed that the time aggregation unit was linked to the detection algorithm used, and that strong aggregation-algorithm interactions need to be taken into account when deciding on which aggregation-algorithm pair to use. Using theoretical data, our study also showed that no one ideal aggregation-algorithm pair exists for all contexts when deciding on which temporal unit of surveillance to use, and that the choice depends on several parameters.Our results can help public health practitioners choose the most appropriate time series aggregation and algorithm according to their specific needs. Finally, the present work enabled us to develop recommendations for a One Health project where the same time aggregation type and detection method could be used for both human and animal syndromic surveillance data.