Retrospective time series analysis of veterinary laboratory data: Preparing a historical baseline for cluster detection in syndromic surveillance

Abstract

The practice of disease surveillance has shifted in the last two decades towards the introduction of systems capable of early detection of disease. Modern biosurveillance systems explore different sources of pre-diagnostic data, such as patient's chief complaint upon emergency visit or laboratory test orders. These sources of data can provide more rapid detection than traditional surveillance based on case confirmation, but are less specific, and therefore their use poses challenges related to the presence of background noise and unlabelled temporal aberrations in historical data. The overall goal of this study was to carry out retrospective analysis using three years of laboratory test submissions to the Animal Health Laboratory in the province of Ontario, Canada, in order to prepare the data for use in syndromic surveillance. Daily cases were grouped into syndromes and counts for each syndrome were monitored on a daily basis when medians were higher than one case per day, and weekly otherwise. Poisson regression accounting for day-of-week and month was able to capture the day-of-week effect with minimal influence from temporal aberrations. Applying Poisson regression in an iterative manner, that removed data points above the predicted 95th percentile of daily counts, allowed for the removal of these aberrations in the absence of labelled outbreaks, while maintaining the day-of-week effect that was present in the original data. This resulted in the construction of time series that represent the baseline patterns over the past three years, free of temporal aberrations. The final method was thus able to remove temporal aberrations while keeping the original explainable effects in the data, did not need a training period free of aberrations, had minimal adjustment to the aberrations present in the raw data, and did not require labelled outbreaks. Moreover, it was readily applicable to the weekly data by substituting Poisson regression with moving 95th percentiles.

Introduction

Surveillance has shifted in the last two decades towards systems capable of early detection of disease (Shmueli and Burkom, 2010). Modern biosurveillance systems are designed to take advantage of data assumed to contain signatures of healthcare-seeking behaviours, which are not as specific as diagnosis, but allow for more rapid detection, and can be aggregated as syndromes. Surveillance based on these types of data is therefore referred to as syndromic surveillance (Centers for Disease Control and Prevention (CDC), 2006). A recent review of syndromic surveillance initiatives in veterinary medicine (Dórea et al., 2011) indicated that opportunistic data sources are difficult to find in animal surveillance due to the scarcity of computerized, automatically collected data.

The secondary use of clinical animal data, whether computerized or not, also relies on the voluntary participation of veterinarians and/or producers. One alternative to relying on data shared voluntarily is the exploitation of automatically collected laboratory submission data (Stone, 2007). Laboratory test results have been analyzed retrospectively to detect temporal clustering of bacterial pathogens in public health (Dessau and Steenberg, 1993, Hutwagner et al., 1997, Widdowson et al., 2003) and veterinary medicine (Carpenter, 2002, Zhang et al., 2005). The use of submission data, however, more properly fits the purposes of syndromic surveillance, as test requests are available earlier, though provide less specificity, than test results. Despite having lower population coverage than clinical data, laboratory data are generally stored in computerized systems, and have been available over relatively lengthy periods of time, meaning that historical analyses are usually possible.

When historical computerized data are available, a key challenge involves the construction of outbreak-free baselines, as any outbreaks will typically not be labelled, nor will their shape and magnitude be known (Shmueli and Burkom, 2010). Detection of abnormal behaviour in prospective analysis is based on either modelling and removing expected background (model-driven methods) or comparing profiles to similar data from unaffected populations (data-driven methods) (Yahav and Shmueli, 2007, Shmueli and Burkom, 2010). In both cases, a baseline free of outbreaks is necessary: in the former case to create models of expected behaviour, and in the latter to serve as a comparison to the data being tested. Historical data can provide a baseline for temporal aberration detection algorithms, but data quality and influence of past outbreaks are challenges to overcome when determining ‘typical’ background behaviour against which the presence of abnormalities can be investigated (Shmueli and Burkom, 2010).

The overall goal of this study was to carry out retrospective analysis using three years of laboratory test submissions, related to health events in cattle, made to the Animal Health Laboratory in the province of Ontario, Canada. These historical data were analyzed for their potential use in syndromic surveillance. The retrospective analysis had two specific objectives. The first was to conduct time series analysis in order to discover explainable patterns in the data, such as day-of-week or seasonal effects as well as global trends. The second objective was to identify a procedure that could adequately describe the “normal behaviour” for each syndrome, separating the background behaviour from temporal aberrations present in the historical laboratory test request data.