Early Warning Surveillance System for Hospital Acquired Infections

Early Warning Surveillance System for Hospital Acquired Infections

Background:

Hospital-acquired infections (HAI) are a significant problem for patients, hospital managers and Infection Control teams, who often deal with it by means of subjective and emotional judgment, especially when the day-to-day numbers of cases are relatively small. As intervention is usually costly there is a need for objective statistical methods to help decide when to intervene.

Objective:

I investigated a variant of the quality control procedure known as the Resetting Sequential Probability Ratio Test (RSPRT) as an approach that is promising because of the absence of subjective setting of upper and lower control limits. It is based on sequential likelihood methodology and works by comparing past rates of occurrence against more recent rates. It has been little used mainly because of the less than intuitive meaning of two of its parameters, α and β, which are related (but not equivalent) to types I and II errors.

Methods:

Stochastic simulations of the Poisson process were used to investigate the ability of RSPRT to give early warnings to detect ‘out of control' increases and decreases in incidence occurrence of infection relative to ‘in control' expected occurrence rates spanning from 0.5 to 100 per unit time. The time axis was left dimensionless so that the results can be generalized to any kind of relevant time unit.

Results:

I carried out 10,000 simulations for each of 920 scenarios, each representing a particular ‘in control' mean rate of occurrence and a relative ‘out of control' rate of increase. Analysis of the parameter landscape reveals that the most useful RSPRT parameter set lies in the range 0 ≤ α ≤ 0.1 and 0 ≤ β ≤ 0.1. Setting parameters to α=0.01 and β=0.01 yields a process with useful properties. This is especially true when alarms signals are flagged on condition that these take place at a time within the 5th centile of the expected distribution of the ‘in-control' process because this helps reduce the number of false positives alarms substantially. Figure: example of alarms triggered via conditional-RSPRT for increases (red star symbol) and decreases (blue face symbol) in Pseudomonas aeruginosa colonization on a UCLH ward (black bars – new cases, grey bars – recurring cases):

Conclusions:

Conditional RSPRT is a useful early warning tool for the early detection of HAI outbreaks requiring minimal setting of arbitrary parameters and resulting in fewer false positive alarms than unconditional RSPRT. In these respects it is superior to some popular early warning methods such as CUSUM (a special case of RSPRT) and the scan statistic. I recommend using this method as a relative (rather than absolute) indication of a problem for further investigation, which could then lead to useful early infection control intervention.