Data-Adaptive Multivariate Control Charts for Routine Health Monitoring

Adaptive Non Adaptive Pe rc en ta ge Raw 7-Day Subtraction Addititive Multiplicative Exp. Smoothing OBJECTIVE This paper investigates the use of data-adaptive multivariate statistical process control (MSPC) charts for outbreak detection using real-world syndromic data. The widely used EARS [1] methods and other adaptive implementations assume implicitly that nonstationarity and/or the lack of historic data preclude the conventional Phase I/Phase II approach of SPC. This work examines that assumption formally by evaluating and comparing the false alarm rates and sensitivity of adaptive and non-adaptive MSPC charts applied to simulated outbreaks injected into both deseasonalized and raw data. BACKGROUND Classical process control charts have been used effectively for industrial quality control for several decades and have many potential applications in hospital surveillance [2]. In this field, the processes examined are assumed to be normal with a constant mean and variance and thus amenable to a 2-stage approach: first, baseline data are analyzed to derive statistical properties expressed as chart parameters, and then the chart is prospectively applied. In sharp contrast, the myriad of processes responsible for creating syndromic data—human behavior, immune physiology, the health care system, to name just a few—often exhibit explicit trends, cyclical behaviors, and statistical properties that change over time. As researchers begin to apply multivariate charts to health surveillance [3], both preconditioning of the data and adaptation of the algorithms become essential. METHODS This investigation, using data preconditioning, algorithm adaptation, and performance analysis, examined authentic syndromic time series derived from three data sources from each of five US cities using both respiratory and gastrointestinal data counts. For data preconditioning, we applied several local deseasonalization techniques—exponential smoothing, subtraction of the count from seven days ago [4] and additive and multiplicative versions of the classic “ratio-to-moving-average” method—to remove seasonal effects that degrade chart performance. To adapt the algorithms to a changing mean and covariance, we implemented each with outlier removal and a moving four week baseline with a 2-day guardband for estimating chart parameters. The adapted MSPC charts included multivariate versions of a cumulative sum, exponentially-weighted moving average (M-